~ [ source navigation ] ~ [ diff markup ] ~ [ identifier search ] ~

TOMOYO Linux Cross Reference
Linux/arch/ia64/kernel/perfmon.c

Version: ~ [ linux-5.3-rc5 ] ~ [ linux-5.2.9 ] ~ [ linux-5.1.21 ] ~ [ linux-5.0.21 ] ~ [ linux-4.20.17 ] ~ [ linux-4.19.67 ] ~ [ linux-4.18.20 ] ~ [ linux-4.17.19 ] ~ [ linux-4.16.18 ] ~ [ linux-4.15.18 ] ~ [ linux-4.14.139 ] ~ [ linux-4.13.16 ] ~ [ linux-4.12.14 ] ~ [ linux-4.11.12 ] ~ [ linux-4.10.17 ] ~ [ linux-4.9.189 ] ~ [ linux-4.8.17 ] ~ [ linux-4.7.10 ] ~ [ linux-4.6.7 ] ~ [ linux-4.5.7 ] ~ [ linux-4.4.189 ] ~ [ linux-4.3.6 ] ~ [ linux-4.2.8 ] ~ [ linux-4.1.52 ] ~ [ linux-4.0.9 ] ~ [ linux-3.19.8 ] ~ [ linux-3.18.140 ] ~ [ linux-3.17.8 ] ~ [ linux-3.16.72 ] ~ [ linux-3.15.10 ] ~ [ linux-3.14.79 ] ~ [ linux-3.13.11 ] ~ [ linux-3.12.74 ] ~ [ linux-3.11.10 ] ~ [ linux-3.10.108 ] ~ [ linux-3.9.11 ] ~ [ linux-3.8.13 ] ~ [ linux-3.7.10 ] ~ [ linux-3.6.11 ] ~ [ linux-3.5.7 ] ~ [ linux-3.4.113 ] ~ [ linux-3.3.8 ] ~ [ linux-3.2.102 ] ~ [ linux-3.1.10 ] ~ [ linux-3.0.101 ] ~ [ linux-2.6.32.71 ] ~ [ linux-2.6.0 ] ~ [ linux-2.4.37.11 ] ~ [ unix-v6-master ] ~ [ ccs-tools-1.8.5 ] ~ [ policy-sample ] ~
Architecture: ~ [ i386 ] ~ [ alpha ] ~ [ m68k ] ~ [ mips ] ~ [ ppc ] ~ [ sparc ] ~ [ sparc64 ] ~

  1 /*
  2  * This file implements the perfmon-2 subsystem which is used
  3  * to program the IA-64 Performance Monitoring Unit (PMU).
  4  *
  5  * The initial version of perfmon.c was written by
  6  * Ganesh Venkitachalam, IBM Corp.
  7  *
  8  * Then it was modified for perfmon-1.x by Stephane Eranian and
  9  * David Mosberger, Hewlett Packard Co.
 10  *
 11  * Version Perfmon-2.x is a rewrite of perfmon-1.x
 12  * by Stephane Eranian, Hewlett Packard Co.
 13  *
 14  * Copyright (C) 1999-2005  Hewlett Packard Co
 15  *               Stephane Eranian <eranian@hpl.hp.com>
 16  *               David Mosberger-Tang <davidm@hpl.hp.com>
 17  *
 18  * More information about perfmon available at:
 19  *      http://www.hpl.hp.com/research/linux/perfmon
 20  */
 21 
 22 #include <linux/module.h>
 23 #include <linux/kernel.h>
 24 #include <linux/sched.h>
 25 #include <linux/interrupt.h>
 26 #include <linux/proc_fs.h>
 27 #include <linux/seq_file.h>
 28 #include <linux/init.h>
 29 #include <linux/vmalloc.h>
 30 #include <linux/mm.h>
 31 #include <linux/sysctl.h>
 32 #include <linux/list.h>
 33 #include <linux/file.h>
 34 #include <linux/poll.h>
 35 #include <linux/vfs.h>
 36 #include <linux/smp.h>
 37 #include <linux/pagemap.h>
 38 #include <linux/mount.h>
 39 #include <linux/bitops.h>
 40 #include <linux/capability.h>
 41 #include <linux/rcupdate.h>
 42 #include <linux/completion.h>
 43 #include <linux/tracehook.h>
 44 #include <linux/slab.h>
 45 #include <linux/cpu.h>
 46 
 47 #include <asm/errno.h>
 48 #include <asm/intrinsics.h>
 49 #include <asm/page.h>
 50 #include <asm/perfmon.h>
 51 #include <asm/processor.h>
 52 #include <asm/signal.h>
 53 #include <asm/uaccess.h>
 54 #include <asm/delay.h>
 55 
 56 #ifdef CONFIG_PERFMON
 57 /*
 58  * perfmon context state
 59  */
 60 #define PFM_CTX_UNLOADED        1       /* context is not loaded onto any task */
 61 #define PFM_CTX_LOADED          2       /* context is loaded onto a task */
 62 #define PFM_CTX_MASKED          3       /* context is loaded but monitoring is masked due to overflow */
 63 #define PFM_CTX_ZOMBIE          4       /* owner of the context is closing it */
 64 
 65 #define PFM_INVALID_ACTIVATION  (~0UL)
 66 
 67 #define PFM_NUM_PMC_REGS        64      /* PMC save area for ctxsw */
 68 #define PFM_NUM_PMD_REGS        64      /* PMD save area for ctxsw */
 69 
 70 /*
 71  * depth of message queue
 72  */
 73 #define PFM_MAX_MSGS            32
 74 #define PFM_CTXQ_EMPTY(g)       ((g)->ctx_msgq_head == (g)->ctx_msgq_tail)
 75 
 76 /*
 77  * type of a PMU register (bitmask).
 78  * bitmask structure:
 79  *      bit0   : register implemented
 80  *      bit1   : end marker
 81  *      bit2-3 : reserved
 82  *      bit4   : pmc has pmc.pm
 83  *      bit5   : pmc controls a counter (has pmc.oi), pmd is used as counter
 84  *      bit6-7 : register type
 85  *      bit8-31: reserved
 86  */
 87 #define PFM_REG_NOTIMPL         0x0 /* not implemented at all */
 88 #define PFM_REG_IMPL            0x1 /* register implemented */
 89 #define PFM_REG_END             0x2 /* end marker */
 90 #define PFM_REG_MONITOR         (0x1<<4|PFM_REG_IMPL) /* a PMC with a pmc.pm field only */
 91 #define PFM_REG_COUNTING        (0x2<<4|PFM_REG_MONITOR) /* a monitor + pmc.oi+ PMD used as a counter */
 92 #define PFM_REG_CONTROL         (0x4<<4|PFM_REG_IMPL) /* PMU control register */
 93 #define PFM_REG_CONFIG          (0x8<<4|PFM_REG_IMPL) /* configuration register */
 94 #define PFM_REG_BUFFER          (0xc<<4|PFM_REG_IMPL) /* PMD used as buffer */
 95 
 96 #define PMC_IS_LAST(i)  (pmu_conf->pmc_desc[i].type & PFM_REG_END)
 97 #define PMD_IS_LAST(i)  (pmu_conf->pmd_desc[i].type & PFM_REG_END)
 98 
 99 #define PMC_OVFL_NOTIFY(ctx, i) ((ctx)->ctx_pmds[i].flags &  PFM_REGFL_OVFL_NOTIFY)
100 
101 /* i assumed unsigned */
102 #define PMC_IS_IMPL(i)    (i< PMU_MAX_PMCS && (pmu_conf->pmc_desc[i].type & PFM_REG_IMPL))
103 #define PMD_IS_IMPL(i)    (i< PMU_MAX_PMDS && (pmu_conf->pmd_desc[i].type & PFM_REG_IMPL))
104 
105 /* XXX: these assume that register i is implemented */
106 #define PMD_IS_COUNTING(i) ((pmu_conf->pmd_desc[i].type & PFM_REG_COUNTING) == PFM_REG_COUNTING)
107 #define PMC_IS_COUNTING(i) ((pmu_conf->pmc_desc[i].type & PFM_REG_COUNTING) == PFM_REG_COUNTING)
108 #define PMC_IS_MONITOR(i)  ((pmu_conf->pmc_desc[i].type & PFM_REG_MONITOR)  == PFM_REG_MONITOR)
109 #define PMC_IS_CONTROL(i)  ((pmu_conf->pmc_desc[i].type & PFM_REG_CONTROL)  == PFM_REG_CONTROL)
110 
111 #define PMC_DFL_VAL(i)     pmu_conf->pmc_desc[i].default_value
112 #define PMC_RSVD_MASK(i)   pmu_conf->pmc_desc[i].reserved_mask
113 #define PMD_PMD_DEP(i)     pmu_conf->pmd_desc[i].dep_pmd[0]
114 #define PMC_PMD_DEP(i)     pmu_conf->pmc_desc[i].dep_pmd[0]
115 
116 #define PFM_NUM_IBRS      IA64_NUM_DBG_REGS
117 #define PFM_NUM_DBRS      IA64_NUM_DBG_REGS
118 
119 #define CTX_OVFL_NOBLOCK(c)     ((c)->ctx_fl_block == 0)
120 #define CTX_HAS_SMPL(c)         ((c)->ctx_fl_is_sampling)
121 #define PFM_CTX_TASK(h)         (h)->ctx_task
122 
123 #define PMU_PMC_OI              5 /* position of pmc.oi bit */
124 
125 /* XXX: does not support more than 64 PMDs */
126 #define CTX_USED_PMD(ctx, mask) (ctx)->ctx_used_pmds[0] |= (mask)
127 #define CTX_IS_USED_PMD(ctx, c) (((ctx)->ctx_used_pmds[0] & (1UL << (c))) != 0UL)
128 
129 #define CTX_USED_MONITOR(ctx, mask) (ctx)->ctx_used_monitors[0] |= (mask)
130 
131 #define CTX_USED_IBR(ctx,n)     (ctx)->ctx_used_ibrs[(n)>>6] |= 1UL<< ((n) % 64)
132 #define CTX_USED_DBR(ctx,n)     (ctx)->ctx_used_dbrs[(n)>>6] |= 1UL<< ((n) % 64)
133 #define CTX_USES_DBREGS(ctx)    (((pfm_context_t *)(ctx))->ctx_fl_using_dbreg==1)
134 #define PFM_CODE_RR     0       /* requesting code range restriction */
135 #define PFM_DATA_RR     1       /* requestion data range restriction */
136 
137 #define PFM_CPUINFO_CLEAR(v)    pfm_get_cpu_var(pfm_syst_info) &= ~(v)
138 #define PFM_CPUINFO_SET(v)      pfm_get_cpu_var(pfm_syst_info) |= (v)
139 #define PFM_CPUINFO_GET()       pfm_get_cpu_var(pfm_syst_info)
140 
141 #define RDEP(x) (1UL<<(x))
142 
143 /*
144  * context protection macros
145  * in SMP:
146  *      - we need to protect against CPU concurrency (spin_lock)
147  *      - we need to protect against PMU overflow interrupts (local_irq_disable)
148  * in UP:
149  *      - we need to protect against PMU overflow interrupts (local_irq_disable)
150  *
151  * spin_lock_irqsave()/spin_unlock_irqrestore():
152  *      in SMP: local_irq_disable + spin_lock
153  *      in UP : local_irq_disable
154  *
155  * spin_lock()/spin_lock():
156  *      in UP : removed automatically
157  *      in SMP: protect against context accesses from other CPU. interrupts
158  *              are not masked. This is useful for the PMU interrupt handler
159  *              because we know we will not get PMU concurrency in that code.
160  */
161 #define PROTECT_CTX(c, f) \
162         do {  \
163                 DPRINT(("spinlock_irq_save ctx %p by [%d]\n", c, task_pid_nr(current))); \
164                 spin_lock_irqsave(&(c)->ctx_lock, f); \
165                 DPRINT(("spinlocked ctx %p  by [%d]\n", c, task_pid_nr(current))); \
166         } while(0)
167 
168 #define UNPROTECT_CTX(c, f) \
169         do { \
170                 DPRINT(("spinlock_irq_restore ctx %p by [%d]\n", c, task_pid_nr(current))); \
171                 spin_unlock_irqrestore(&(c)->ctx_lock, f); \
172         } while(0)
173 
174 #define PROTECT_CTX_NOPRINT(c, f) \
175         do {  \
176                 spin_lock_irqsave(&(c)->ctx_lock, f); \
177         } while(0)
178 
179 
180 #define UNPROTECT_CTX_NOPRINT(c, f) \
181         do { \
182                 spin_unlock_irqrestore(&(c)->ctx_lock, f); \
183         } while(0)
184 
185 
186 #define PROTECT_CTX_NOIRQ(c) \
187         do {  \
188                 spin_lock(&(c)->ctx_lock); \
189         } while(0)
190 
191 #define UNPROTECT_CTX_NOIRQ(c) \
192         do { \
193                 spin_unlock(&(c)->ctx_lock); \
194         } while(0)
195 
196 
197 #ifdef CONFIG_SMP
198 
199 #define GET_ACTIVATION()        pfm_get_cpu_var(pmu_activation_number)
200 #define INC_ACTIVATION()        pfm_get_cpu_var(pmu_activation_number)++
201 #define SET_ACTIVATION(c)       (c)->ctx_last_activation = GET_ACTIVATION()
202 
203 #else /* !CONFIG_SMP */
204 #define SET_ACTIVATION(t)       do {} while(0)
205 #define GET_ACTIVATION(t)       do {} while(0)
206 #define INC_ACTIVATION(t)       do {} while(0)
207 #endif /* CONFIG_SMP */
208 
209 #define SET_PMU_OWNER(t, c)     do { pfm_get_cpu_var(pmu_owner) = (t); pfm_get_cpu_var(pmu_ctx) = (c); } while(0)
210 #define GET_PMU_OWNER()         pfm_get_cpu_var(pmu_owner)
211 #define GET_PMU_CTX()           pfm_get_cpu_var(pmu_ctx)
212 
213 #define LOCK_PFS(g)             spin_lock_irqsave(&pfm_sessions.pfs_lock, g)
214 #define UNLOCK_PFS(g)           spin_unlock_irqrestore(&pfm_sessions.pfs_lock, g)
215 
216 #define PFM_REG_RETFLAG_SET(flags, val) do { flags &= ~PFM_REG_RETFL_MASK; flags |= (val); } while(0)
217 
218 /*
219  * cmp0 must be the value of pmc0
220  */
221 #define PMC0_HAS_OVFL(cmp0)  (cmp0 & ~0x1UL)
222 
223 #define PFMFS_MAGIC 0xa0b4d889
224 
225 /*
226  * debugging
227  */
228 #define PFM_DEBUGGING 1
229 #ifdef PFM_DEBUGGING
230 #define DPRINT(a) \
231         do { \
232                 if (unlikely(pfm_sysctl.debug >0)) { printk("%s.%d: CPU%d [%d] ", __func__, __LINE__, smp_processor_id(), task_pid_nr(current)); printk a; } \
233         } while (0)
234 
235 #define DPRINT_ovfl(a) \
236         do { \
237                 if (unlikely(pfm_sysctl.debug > 0 && pfm_sysctl.debug_ovfl >0)) { printk("%s.%d: CPU%d [%d] ", __func__, __LINE__, smp_processor_id(), task_pid_nr(current)); printk a; } \
238         } while (0)
239 #endif
240 
241 /*
242  * 64-bit software counter structure
243  *
244  * the next_reset_type is applied to the next call to pfm_reset_regs()
245  */
246 typedef struct {
247         unsigned long   val;            /* virtual 64bit counter value */
248         unsigned long   lval;           /* last reset value */
249         unsigned long   long_reset;     /* reset value on sampling overflow */
250         unsigned long   short_reset;    /* reset value on overflow */
251         unsigned long   reset_pmds[4];  /* which other pmds to reset when this counter overflows */
252         unsigned long   smpl_pmds[4];   /* which pmds are accessed when counter overflow */
253         unsigned long   seed;           /* seed for random-number generator */
254         unsigned long   mask;           /* mask for random-number generator */
255         unsigned int    flags;          /* notify/do not notify */
256         unsigned long   eventid;        /* overflow event identifier */
257 } pfm_counter_t;
258 
259 /*
260  * context flags
261  */
262 typedef struct {
263         unsigned int block:1;           /* when 1, task will blocked on user notifications */
264         unsigned int system:1;          /* do system wide monitoring */
265         unsigned int using_dbreg:1;     /* using range restrictions (debug registers) */
266         unsigned int is_sampling:1;     /* true if using a custom format */
267         unsigned int excl_idle:1;       /* exclude idle task in system wide session */
268         unsigned int going_zombie:1;    /* context is zombie (MASKED+blocking) */
269         unsigned int trap_reason:2;     /* reason for going into pfm_handle_work() */
270         unsigned int no_msg:1;          /* no message sent on overflow */
271         unsigned int can_restart:1;     /* allowed to issue a PFM_RESTART */
272         unsigned int reserved:22;
273 } pfm_context_flags_t;
274 
275 #define PFM_TRAP_REASON_NONE            0x0     /* default value */
276 #define PFM_TRAP_REASON_BLOCK           0x1     /* we need to block on overflow */
277 #define PFM_TRAP_REASON_RESET           0x2     /* we need to reset PMDs */
278 
279 
280 /*
281  * perfmon context: encapsulates all the state of a monitoring session
282  */
283 
284 typedef struct pfm_context {
285         spinlock_t              ctx_lock;               /* context protection */
286 
287         pfm_context_flags_t     ctx_flags;              /* bitmask of flags  (block reason incl.) */
288         unsigned int            ctx_state;              /* state: active/inactive (no bitfield) */
289 
290         struct task_struct      *ctx_task;              /* task to which context is attached */
291 
292         unsigned long           ctx_ovfl_regs[4];       /* which registers overflowed (notification) */
293 
294         struct completion       ctx_restart_done;       /* use for blocking notification mode */
295 
296         unsigned long           ctx_used_pmds[4];       /* bitmask of PMD used            */
297         unsigned long           ctx_all_pmds[4];        /* bitmask of all accessible PMDs */
298         unsigned long           ctx_reload_pmds[4];     /* bitmask of force reload PMD on ctxsw in */
299 
300         unsigned long           ctx_all_pmcs[4];        /* bitmask of all accessible PMCs */
301         unsigned long           ctx_reload_pmcs[4];     /* bitmask of force reload PMC on ctxsw in */
302         unsigned long           ctx_used_monitors[4];   /* bitmask of monitor PMC being used */
303 
304         unsigned long           ctx_pmcs[PFM_NUM_PMC_REGS];     /*  saved copies of PMC values */
305 
306         unsigned int            ctx_used_ibrs[1];               /* bitmask of used IBR (speedup ctxsw in) */
307         unsigned int            ctx_used_dbrs[1];               /* bitmask of used DBR (speedup ctxsw in) */
308         unsigned long           ctx_dbrs[IA64_NUM_DBG_REGS];    /* DBR values (cache) when not loaded */
309         unsigned long           ctx_ibrs[IA64_NUM_DBG_REGS];    /* IBR values (cache) when not loaded */
310 
311         pfm_counter_t           ctx_pmds[PFM_NUM_PMD_REGS]; /* software state for PMDS */
312 
313         unsigned long           th_pmcs[PFM_NUM_PMC_REGS];      /* PMC thread save state */
314         unsigned long           th_pmds[PFM_NUM_PMD_REGS];      /* PMD thread save state */
315 
316         unsigned long           ctx_saved_psr_up;       /* only contains psr.up value */
317 
318         unsigned long           ctx_last_activation;    /* context last activation number for last_cpu */
319         unsigned int            ctx_last_cpu;           /* CPU id of current or last CPU used (SMP only) */
320         unsigned int            ctx_cpu;                /* cpu to which perfmon is applied (system wide) */
321 
322         int                     ctx_fd;                 /* file descriptor used my this context */
323         pfm_ovfl_arg_t          ctx_ovfl_arg;           /* argument to custom buffer format handler */
324 
325         pfm_buffer_fmt_t        *ctx_buf_fmt;           /* buffer format callbacks */
326         void                    *ctx_smpl_hdr;          /* points to sampling buffer header kernel vaddr */
327         unsigned long           ctx_smpl_size;          /* size of sampling buffer */
328         void                    *ctx_smpl_vaddr;        /* user level virtual address of smpl buffer */
329 
330         wait_queue_head_t       ctx_msgq_wait;
331         pfm_msg_t               ctx_msgq[PFM_MAX_MSGS];
332         int                     ctx_msgq_head;
333         int                     ctx_msgq_tail;
334         struct fasync_struct    *ctx_async_queue;
335 
336         wait_queue_head_t       ctx_zombieq;            /* termination cleanup wait queue */
337 } pfm_context_t;
338 
339 /*
340  * magic number used to verify that structure is really
341  * a perfmon context
342  */
343 #define PFM_IS_FILE(f)          ((f)->f_op == &pfm_file_ops)
344 
345 #define PFM_GET_CTX(t)          ((pfm_context_t *)(t)->thread.pfm_context)
346 
347 #ifdef CONFIG_SMP
348 #define SET_LAST_CPU(ctx, v)    (ctx)->ctx_last_cpu = (v)
349 #define GET_LAST_CPU(ctx)       (ctx)->ctx_last_cpu
350 #else
351 #define SET_LAST_CPU(ctx, v)    do {} while(0)
352 #define GET_LAST_CPU(ctx)       do {} while(0)
353 #endif
354 
355 
356 #define ctx_fl_block            ctx_flags.block
357 #define ctx_fl_system           ctx_flags.system
358 #define ctx_fl_using_dbreg      ctx_flags.using_dbreg
359 #define ctx_fl_is_sampling      ctx_flags.is_sampling
360 #define ctx_fl_excl_idle        ctx_flags.excl_idle
361 #define ctx_fl_going_zombie     ctx_flags.going_zombie
362 #define ctx_fl_trap_reason      ctx_flags.trap_reason
363 #define ctx_fl_no_msg           ctx_flags.no_msg
364 #define ctx_fl_can_restart      ctx_flags.can_restart
365 
366 #define PFM_SET_WORK_PENDING(t, v)      do { (t)->thread.pfm_needs_checking = v; } while(0);
367 #define PFM_GET_WORK_PENDING(t)         (t)->thread.pfm_needs_checking
368 
369 /*
370  * global information about all sessions
371  * mostly used to synchronize between system wide and per-process
372  */
373 typedef struct {
374         spinlock_t              pfs_lock;                  /* lock the structure */
375 
376         unsigned int            pfs_task_sessions;         /* number of per task sessions */
377         unsigned int            pfs_sys_sessions;          /* number of per system wide sessions */
378         unsigned int            pfs_sys_use_dbregs;        /* incremented when a system wide session uses debug regs */
379         unsigned int            pfs_ptrace_use_dbregs;     /* incremented when a process uses debug regs */
380         struct task_struct      *pfs_sys_session[NR_CPUS]; /* point to task owning a system-wide session */
381 } pfm_session_t;
382 
383 /*
384  * information about a PMC or PMD.
385  * dep_pmd[]: a bitmask of dependent PMD registers
386  * dep_pmc[]: a bitmask of dependent PMC registers
387  */
388 typedef int (*pfm_reg_check_t)(struct task_struct *task, pfm_context_t *ctx, unsigned int cnum, unsigned long *val, struct pt_regs *regs);
389 typedef struct {
390         unsigned int            type;
391         int                     pm_pos;
392         unsigned long           default_value;  /* power-on default value */
393         unsigned long           reserved_mask;  /* bitmask of reserved bits */
394         pfm_reg_check_t         read_check;
395         pfm_reg_check_t         write_check;
396         unsigned long           dep_pmd[4];
397         unsigned long           dep_pmc[4];
398 } pfm_reg_desc_t;
399 
400 /* assume cnum is a valid monitor */
401 #define PMC_PM(cnum, val)       (((val) >> (pmu_conf->pmc_desc[cnum].pm_pos)) & 0x1)
402 
403 /*
404  * This structure is initialized at boot time and contains
405  * a description of the PMU main characteristics.
406  *
407  * If the probe function is defined, detection is based
408  * on its return value: 
409  *      - 0 means recognized PMU
410  *      - anything else means not supported
411  * When the probe function is not defined, then the pmu_family field
412  * is used and it must match the host CPU family such that:
413  *      - cpu->family & config->pmu_family != 0
414  */
415 typedef struct {
416         unsigned long  ovfl_val;        /* overflow value for counters */
417 
418         pfm_reg_desc_t *pmc_desc;       /* detailed PMC register dependencies descriptions */
419         pfm_reg_desc_t *pmd_desc;       /* detailed PMD register dependencies descriptions */
420 
421         unsigned int   num_pmcs;        /* number of PMCS: computed at init time */
422         unsigned int   num_pmds;        /* number of PMDS: computed at init time */
423         unsigned long  impl_pmcs[4];    /* bitmask of implemented PMCS */
424         unsigned long  impl_pmds[4];    /* bitmask of implemented PMDS */
425 
426         char          *pmu_name;        /* PMU family name */
427         unsigned int  pmu_family;       /* cpuid family pattern used to identify pmu */
428         unsigned int  flags;            /* pmu specific flags */
429         unsigned int  num_ibrs;         /* number of IBRS: computed at init time */
430         unsigned int  num_dbrs;         /* number of DBRS: computed at init time */
431         unsigned int  num_counters;     /* PMC/PMD counting pairs : computed at init time */
432         int           (*probe)(void);   /* customized probe routine */
433         unsigned int  use_rr_dbregs:1;  /* set if debug registers used for range restriction */
434 } pmu_config_t;
435 /*
436  * PMU specific flags
437  */
438 #define PFM_PMU_IRQ_RESEND      1       /* PMU needs explicit IRQ resend */
439 
440 /*
441  * debug register related type definitions
442  */
443 typedef struct {
444         unsigned long ibr_mask:56;
445         unsigned long ibr_plm:4;
446         unsigned long ibr_ig:3;
447         unsigned long ibr_x:1;
448 } ibr_mask_reg_t;
449 
450 typedef struct {
451         unsigned long dbr_mask:56;
452         unsigned long dbr_plm:4;
453         unsigned long dbr_ig:2;
454         unsigned long dbr_w:1;
455         unsigned long dbr_r:1;
456 } dbr_mask_reg_t;
457 
458 typedef union {
459         unsigned long  val;
460         ibr_mask_reg_t ibr;
461         dbr_mask_reg_t dbr;
462 } dbreg_t;
463 
464 
465 /*
466  * perfmon command descriptions
467  */
468 typedef struct {
469         int             (*cmd_func)(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs);
470         char            *cmd_name;
471         int             cmd_flags;
472         unsigned int    cmd_narg;
473         size_t          cmd_argsize;
474         int             (*cmd_getsize)(void *arg, size_t *sz);
475 } pfm_cmd_desc_t;
476 
477 #define PFM_CMD_FD              0x01    /* command requires a file descriptor */
478 #define PFM_CMD_ARG_READ        0x02    /* command must read argument(s) */
479 #define PFM_CMD_ARG_RW          0x04    /* command must read/write argument(s) */
480 #define PFM_CMD_STOP            0x08    /* command does not work on zombie context */
481 
482 
483 #define PFM_CMD_NAME(cmd)       pfm_cmd_tab[(cmd)].cmd_name
484 #define PFM_CMD_READ_ARG(cmd)   (pfm_cmd_tab[(cmd)].cmd_flags & PFM_CMD_ARG_READ)
485 #define PFM_CMD_RW_ARG(cmd)     (pfm_cmd_tab[(cmd)].cmd_flags & PFM_CMD_ARG_RW)
486 #define PFM_CMD_USE_FD(cmd)     (pfm_cmd_tab[(cmd)].cmd_flags & PFM_CMD_FD)
487 #define PFM_CMD_STOPPED(cmd)    (pfm_cmd_tab[(cmd)].cmd_flags & PFM_CMD_STOP)
488 
489 #define PFM_CMD_ARG_MANY        -1 /* cannot be zero */
490 
491 typedef struct {
492         unsigned long pfm_spurious_ovfl_intr_count;     /* keep track of spurious ovfl interrupts */
493         unsigned long pfm_replay_ovfl_intr_count;       /* keep track of replayed ovfl interrupts */
494         unsigned long pfm_ovfl_intr_count;              /* keep track of ovfl interrupts */
495         unsigned long pfm_ovfl_intr_cycles;             /* cycles spent processing ovfl interrupts */
496         unsigned long pfm_ovfl_intr_cycles_min;         /* min cycles spent processing ovfl interrupts */
497         unsigned long pfm_ovfl_intr_cycles_max;         /* max cycles spent processing ovfl interrupts */
498         unsigned long pfm_smpl_handler_calls;
499         unsigned long pfm_smpl_handler_cycles;
500         char pad[SMP_CACHE_BYTES] ____cacheline_aligned;
501 } pfm_stats_t;
502 
503 /*
504  * perfmon internal variables
505  */
506 static pfm_stats_t              pfm_stats[NR_CPUS];
507 static pfm_session_t            pfm_sessions;   /* global sessions information */
508 
509 static DEFINE_SPINLOCK(pfm_alt_install_check);
510 static pfm_intr_handler_desc_t  *pfm_alt_intr_handler;
511 
512 static struct proc_dir_entry    *perfmon_dir;
513 static pfm_uuid_t               pfm_null_uuid = {0,};
514 
515 static spinlock_t               pfm_buffer_fmt_lock;
516 static LIST_HEAD(pfm_buffer_fmt_list);
517 
518 static pmu_config_t             *pmu_conf;
519 
520 /* sysctl() controls */
521 pfm_sysctl_t pfm_sysctl;
522 EXPORT_SYMBOL(pfm_sysctl);
523 
524 static struct ctl_table pfm_ctl_table[] = {
525         {
526                 .procname       = "debug",
527                 .data           = &pfm_sysctl.debug,
528                 .maxlen         = sizeof(int),
529                 .mode           = 0666,
530                 .proc_handler   = proc_dointvec,
531         },
532         {
533                 .procname       = "debug_ovfl",
534                 .data           = &pfm_sysctl.debug_ovfl,
535                 .maxlen         = sizeof(int),
536                 .mode           = 0666,
537                 .proc_handler   = proc_dointvec,
538         },
539         {
540                 .procname       = "fastctxsw",
541                 .data           = &pfm_sysctl.fastctxsw,
542                 .maxlen         = sizeof(int),
543                 .mode           = 0600,
544                 .proc_handler   = proc_dointvec,
545         },
546         {
547                 .procname       = "expert_mode",
548                 .data           = &pfm_sysctl.expert_mode,
549                 .maxlen         = sizeof(int),
550                 .mode           = 0600,
551                 .proc_handler   = proc_dointvec,
552         },
553         {}
554 };
555 static struct ctl_table pfm_sysctl_dir[] = {
556         {
557                 .procname       = "perfmon",
558                 .mode           = 0555,
559                 .child          = pfm_ctl_table,
560         },
561         {}
562 };
563 static struct ctl_table pfm_sysctl_root[] = {
564         {
565                 .procname       = "kernel",
566                 .mode           = 0555,
567                 .child          = pfm_sysctl_dir,
568         },
569         {}
570 };
571 static struct ctl_table_header *pfm_sysctl_header;
572 
573 static int pfm_context_unload(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs);
574 
575 #define pfm_get_cpu_var(v)              __ia64_per_cpu_var(v)
576 #define pfm_get_cpu_data(a,b)           per_cpu(a, b)
577 
578 static inline void
579 pfm_put_task(struct task_struct *task)
580 {
581         if (task != current) put_task_struct(task);
582 }
583 
584 static inline void
585 pfm_reserve_page(unsigned long a)
586 {
587         SetPageReserved(vmalloc_to_page((void *)a));
588 }
589 static inline void
590 pfm_unreserve_page(unsigned long a)
591 {
592         ClearPageReserved(vmalloc_to_page((void*)a));
593 }
594 
595 static inline unsigned long
596 pfm_protect_ctx_ctxsw(pfm_context_t *x)
597 {
598         spin_lock(&(x)->ctx_lock);
599         return 0UL;
600 }
601 
602 static inline void
603 pfm_unprotect_ctx_ctxsw(pfm_context_t *x, unsigned long f)
604 {
605         spin_unlock(&(x)->ctx_lock);
606 }
607 
608 /* forward declaration */
609 static const struct dentry_operations pfmfs_dentry_operations;
610 
611 static struct dentry *
612 pfmfs_mount(struct file_system_type *fs_type, int flags, const char *dev_name, void *data)
613 {
614         return mount_pseudo(fs_type, "pfm:", NULL, &pfmfs_dentry_operations,
615                         PFMFS_MAGIC);
616 }
617 
618 static struct file_system_type pfm_fs_type = {
619         .name     = "pfmfs",
620         .mount    = pfmfs_mount,
621         .kill_sb  = kill_anon_super,
622 };
623 MODULE_ALIAS_FS("pfmfs");
624 
625 DEFINE_PER_CPU(unsigned long, pfm_syst_info);
626 DEFINE_PER_CPU(struct task_struct *, pmu_owner);
627 DEFINE_PER_CPU(pfm_context_t  *, pmu_ctx);
628 DEFINE_PER_CPU(unsigned long, pmu_activation_number);
629 EXPORT_PER_CPU_SYMBOL_GPL(pfm_syst_info);
630 
631 
632 /* forward declaration */
633 static const struct file_operations pfm_file_ops;
634 
635 /*
636  * forward declarations
637  */
638 #ifndef CONFIG_SMP
639 static void pfm_lazy_save_regs (struct task_struct *ta);
640 #endif
641 
642 void dump_pmu_state(const char *);
643 static int pfm_write_ibr_dbr(int mode, pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs);
644 
645 #include "perfmon_itanium.h"
646 #include "perfmon_mckinley.h"
647 #include "perfmon_montecito.h"
648 #include "perfmon_generic.h"
649 
650 static pmu_config_t *pmu_confs[]={
651         &pmu_conf_mont,
652         &pmu_conf_mck,
653         &pmu_conf_ita,
654         &pmu_conf_gen, /* must be last */
655         NULL
656 };
657 
658 
659 static int pfm_end_notify_user(pfm_context_t *ctx);
660 
661 static inline void
662 pfm_clear_psr_pp(void)
663 {
664         ia64_rsm(IA64_PSR_PP);
665         ia64_srlz_i();
666 }
667 
668 static inline void
669 pfm_set_psr_pp(void)
670 {
671         ia64_ssm(IA64_PSR_PP);
672         ia64_srlz_i();
673 }
674 
675 static inline void
676 pfm_clear_psr_up(void)
677 {
678         ia64_rsm(IA64_PSR_UP);
679         ia64_srlz_i();
680 }
681 
682 static inline void
683 pfm_set_psr_up(void)
684 {
685         ia64_ssm(IA64_PSR_UP);
686         ia64_srlz_i();
687 }
688 
689 static inline unsigned long
690 pfm_get_psr(void)
691 {
692         unsigned long tmp;
693         tmp = ia64_getreg(_IA64_REG_PSR);
694         ia64_srlz_i();
695         return tmp;
696 }
697 
698 static inline void
699 pfm_set_psr_l(unsigned long val)
700 {
701         ia64_setreg(_IA64_REG_PSR_L, val);
702         ia64_srlz_i();
703 }
704 
705 static inline void
706 pfm_freeze_pmu(void)
707 {
708         ia64_set_pmc(0,1UL);
709         ia64_srlz_d();
710 }
711 
712 static inline void
713 pfm_unfreeze_pmu(void)
714 {
715         ia64_set_pmc(0,0UL);
716         ia64_srlz_d();
717 }
718 
719 static inline void
720 pfm_restore_ibrs(unsigned long *ibrs, unsigned int nibrs)
721 {
722         int i;
723 
724         for (i=0; i < nibrs; i++) {
725                 ia64_set_ibr(i, ibrs[i]);
726                 ia64_dv_serialize_instruction();
727         }
728         ia64_srlz_i();
729 }
730 
731 static inline void
732 pfm_restore_dbrs(unsigned long *dbrs, unsigned int ndbrs)
733 {
734         int i;
735 
736         for (i=0; i < ndbrs; i++) {
737                 ia64_set_dbr(i, dbrs[i]);
738                 ia64_dv_serialize_data();
739         }
740         ia64_srlz_d();
741 }
742 
743 /*
744  * PMD[i] must be a counter. no check is made
745  */
746 static inline unsigned long
747 pfm_read_soft_counter(pfm_context_t *ctx, int i)
748 {
749         return ctx->ctx_pmds[i].val + (ia64_get_pmd(i) & pmu_conf->ovfl_val);
750 }
751 
752 /*
753  * PMD[i] must be a counter. no check is made
754  */
755 static inline void
756 pfm_write_soft_counter(pfm_context_t *ctx, int i, unsigned long val)
757 {
758         unsigned long ovfl_val = pmu_conf->ovfl_val;
759 
760         ctx->ctx_pmds[i].val = val  & ~ovfl_val;
761         /*
762          * writing to unimplemented part is ignore, so we do not need to
763          * mask off top part
764          */
765         ia64_set_pmd(i, val & ovfl_val);
766 }
767 
768 static pfm_msg_t *
769 pfm_get_new_msg(pfm_context_t *ctx)
770 {
771         int idx, next;
772 
773         next = (ctx->ctx_msgq_tail+1) % PFM_MAX_MSGS;
774 
775         DPRINT(("ctx_fd=%p head=%d tail=%d\n", ctx, ctx->ctx_msgq_head, ctx->ctx_msgq_tail));
776         if (next == ctx->ctx_msgq_head) return NULL;
777 
778         idx =   ctx->ctx_msgq_tail;
779         ctx->ctx_msgq_tail = next;
780 
781         DPRINT(("ctx=%p head=%d tail=%d msg=%d\n", ctx, ctx->ctx_msgq_head, ctx->ctx_msgq_tail, idx));
782 
783         return ctx->ctx_msgq+idx;
784 }
785 
786 static pfm_msg_t *
787 pfm_get_next_msg(pfm_context_t *ctx)
788 {
789         pfm_msg_t *msg;
790 
791         DPRINT(("ctx=%p head=%d tail=%d\n", ctx, ctx->ctx_msgq_head, ctx->ctx_msgq_tail));
792 
793         if (PFM_CTXQ_EMPTY(ctx)) return NULL;
794 
795         /*
796          * get oldest message
797          */
798         msg = ctx->ctx_msgq+ctx->ctx_msgq_head;
799 
800         /*
801          * and move forward
802          */
803         ctx->ctx_msgq_head = (ctx->ctx_msgq_head+1) % PFM_MAX_MSGS;
804 
805         DPRINT(("ctx=%p head=%d tail=%d type=%d\n", ctx, ctx->ctx_msgq_head, ctx->ctx_msgq_tail, msg->pfm_gen_msg.msg_type));
806 
807         return msg;
808 }
809 
810 static void
811 pfm_reset_msgq(pfm_context_t *ctx)
812 {
813         ctx->ctx_msgq_head = ctx->ctx_msgq_tail = 0;
814         DPRINT(("ctx=%p msgq reset\n", ctx));
815 }
816 
817 static void *
818 pfm_rvmalloc(unsigned long size)
819 {
820         void *mem;
821         unsigned long addr;
822 
823         size = PAGE_ALIGN(size);
824         mem  = vzalloc(size);
825         if (mem) {
826                 //printk("perfmon: CPU%d pfm_rvmalloc(%ld)=%p\n", smp_processor_id(), size, mem);
827                 addr = (unsigned long)mem;
828                 while (size > 0) {
829                         pfm_reserve_page(addr);
830                         addr+=PAGE_SIZE;
831                         size-=PAGE_SIZE;
832                 }
833         }
834         return mem;
835 }
836 
837 static void
838 pfm_rvfree(void *mem, unsigned long size)
839 {
840         unsigned long addr;
841 
842         if (mem) {
843                 DPRINT(("freeing physical buffer @%p size=%lu\n", mem, size));
844                 addr = (unsigned long) mem;
845                 while ((long) size > 0) {
846                         pfm_unreserve_page(addr);
847                         addr+=PAGE_SIZE;
848                         size-=PAGE_SIZE;
849                 }
850                 vfree(mem);
851         }
852         return;
853 }
854 
855 static pfm_context_t *
856 pfm_context_alloc(int ctx_flags)
857 {
858         pfm_context_t *ctx;
859 
860         /* 
861          * allocate context descriptor 
862          * must be able to free with interrupts disabled
863          */
864         ctx = kzalloc(sizeof(pfm_context_t), GFP_KERNEL);
865         if (ctx) {
866                 DPRINT(("alloc ctx @%p\n", ctx));
867 
868                 /*
869                  * init context protection lock
870                  */
871                 spin_lock_init(&ctx->ctx_lock);
872 
873                 /*
874                  * context is unloaded
875                  */
876                 ctx->ctx_state = PFM_CTX_UNLOADED;
877 
878                 /*
879                  * initialization of context's flags
880                  */
881                 ctx->ctx_fl_block       = (ctx_flags & PFM_FL_NOTIFY_BLOCK) ? 1 : 0;
882                 ctx->ctx_fl_system      = (ctx_flags & PFM_FL_SYSTEM_WIDE) ? 1: 0;
883                 ctx->ctx_fl_no_msg      = (ctx_flags & PFM_FL_OVFL_NO_MSG) ? 1: 0;
884                 /*
885                  * will move to set properties
886                  * ctx->ctx_fl_excl_idle   = (ctx_flags & PFM_FL_EXCL_IDLE) ? 1: 0;
887                  */
888 
889                 /*
890                  * init restart semaphore to locked
891                  */
892                 init_completion(&ctx->ctx_restart_done);
893 
894                 /*
895                  * activation is used in SMP only
896                  */
897                 ctx->ctx_last_activation = PFM_INVALID_ACTIVATION;
898                 SET_LAST_CPU(ctx, -1);
899 
900                 /*
901                  * initialize notification message queue
902                  */
903                 ctx->ctx_msgq_head = ctx->ctx_msgq_tail = 0;
904                 init_waitqueue_head(&ctx->ctx_msgq_wait);
905                 init_waitqueue_head(&ctx->ctx_zombieq);
906 
907         }
908         return ctx;
909 }
910 
911 static void
912 pfm_context_free(pfm_context_t *ctx)
913 {
914         if (ctx) {
915                 DPRINT(("free ctx @%p\n", ctx));
916                 kfree(ctx);
917         }
918 }
919 
920 static void
921 pfm_mask_monitoring(struct task_struct *task)
922 {
923         pfm_context_t *ctx = PFM_GET_CTX(task);
924         unsigned long mask, val, ovfl_mask;
925         int i;
926 
927         DPRINT_ovfl(("masking monitoring for [%d]\n", task_pid_nr(task)));
928 
929         ovfl_mask = pmu_conf->ovfl_val;
930         /*
931          * monitoring can only be masked as a result of a valid
932          * counter overflow. In UP, it means that the PMU still
933          * has an owner. Note that the owner can be different
934          * from the current task. However the PMU state belongs
935          * to the owner.
936          * In SMP, a valid overflow only happens when task is
937          * current. Therefore if we come here, we know that
938          * the PMU state belongs to the current task, therefore
939          * we can access the live registers.
940          *
941          * So in both cases, the live register contains the owner's
942          * state. We can ONLY touch the PMU registers and NOT the PSR.
943          *
944          * As a consequence to this call, the ctx->th_pmds[] array
945          * contains stale information which must be ignored
946          * when context is reloaded AND monitoring is active (see
947          * pfm_restart).
948          */
949         mask = ctx->ctx_used_pmds[0];
950         for (i = 0; mask; i++, mask>>=1) {
951                 /* skip non used pmds */
952                 if ((mask & 0x1) == 0) continue;
953                 val = ia64_get_pmd(i);
954 
955                 if (PMD_IS_COUNTING(i)) {
956                         /*
957                          * we rebuild the full 64 bit value of the counter
958                          */
959                         ctx->ctx_pmds[i].val += (val & ovfl_mask);
960                 } else {
961                         ctx->ctx_pmds[i].val = val;
962                 }
963                 DPRINT_ovfl(("pmd[%d]=0x%lx hw_pmd=0x%lx\n",
964                         i,
965                         ctx->ctx_pmds[i].val,
966                         val & ovfl_mask));
967         }
968         /*
969          * mask monitoring by setting the privilege level to 0
970          * we cannot use psr.pp/psr.up for this, it is controlled by
971          * the user
972          *
973          * if task is current, modify actual registers, otherwise modify
974          * thread save state, i.e., what will be restored in pfm_load_regs()
975          */
976         mask = ctx->ctx_used_monitors[0] >> PMU_FIRST_COUNTER;
977         for(i= PMU_FIRST_COUNTER; mask; i++, mask>>=1) {
978                 if ((mask & 0x1) == 0UL) continue;
979                 ia64_set_pmc(i, ctx->th_pmcs[i] & ~0xfUL);
980                 ctx->th_pmcs[i] &= ~0xfUL;
981                 DPRINT_ovfl(("pmc[%d]=0x%lx\n", i, ctx->th_pmcs[i]));
982         }
983         /*
984          * make all of this visible
985          */
986         ia64_srlz_d();
987 }
988 
989 /*
990  * must always be done with task == current
991  *
992  * context must be in MASKED state when calling
993  */
994 static void
995 pfm_restore_monitoring(struct task_struct *task)
996 {
997         pfm_context_t *ctx = PFM_GET_CTX(task);
998         unsigned long mask, ovfl_mask;
999         unsigned long psr, val;
1000         int i, is_system;
1001 
1002         is_system = ctx->ctx_fl_system;
1003         ovfl_mask = pmu_conf->ovfl_val;
1004 
1005         if (task != current) {
1006                 printk(KERN_ERR "perfmon.%d: invalid task[%d] current[%d]\n", __LINE__, task_pid_nr(task), task_pid_nr(current));
1007                 return;
1008         }
1009         if (ctx->ctx_state != PFM_CTX_MASKED) {
1010                 printk(KERN_ERR "perfmon.%d: task[%d] current[%d] invalid state=%d\n", __LINE__,
1011                         task_pid_nr(task), task_pid_nr(current), ctx->ctx_state);
1012                 return;
1013         }
1014         psr = pfm_get_psr();
1015         /*
1016          * monitoring is masked via the PMC.
1017          * As we restore their value, we do not want each counter to
1018          * restart right away. We stop monitoring using the PSR,
1019          * restore the PMC (and PMD) and then re-establish the psr
1020          * as it was. Note that there can be no pending overflow at
1021          * this point, because monitoring was MASKED.
1022          *
1023          * system-wide session are pinned and self-monitoring
1024          */
1025         if (is_system && (PFM_CPUINFO_GET() & PFM_CPUINFO_DCR_PP)) {
1026                 /* disable dcr pp */
1027                 ia64_setreg(_IA64_REG_CR_DCR, ia64_getreg(_IA64_REG_CR_DCR) & ~IA64_DCR_PP);
1028                 pfm_clear_psr_pp();
1029         } else {
1030                 pfm_clear_psr_up();
1031         }
1032         /*
1033          * first, we restore the PMD
1034          */
1035         mask = ctx->ctx_used_pmds[0];
1036         for (i = 0; mask; i++, mask>>=1) {
1037                 /* skip non used pmds */
1038                 if ((mask & 0x1) == 0) continue;
1039 
1040                 if (PMD_IS_COUNTING(i)) {
1041                         /*
1042                          * we split the 64bit value according to
1043                          * counter width
1044                          */
1045                         val = ctx->ctx_pmds[i].val & ovfl_mask;
1046                         ctx->ctx_pmds[i].val &= ~ovfl_mask;
1047                 } else {
1048                         val = ctx->ctx_pmds[i].val;
1049                 }
1050                 ia64_set_pmd(i, val);
1051 
1052                 DPRINT(("pmd[%d]=0x%lx hw_pmd=0x%lx\n",
1053                         i,
1054                         ctx->ctx_pmds[i].val,
1055                         val));
1056         }
1057         /*
1058          * restore the PMCs
1059          */
1060         mask = ctx->ctx_used_monitors[0] >> PMU_FIRST_COUNTER;
1061         for(i= PMU_FIRST_COUNTER; mask; i++, mask>>=1) {
1062                 if ((mask & 0x1) == 0UL) continue;
1063                 ctx->th_pmcs[i] = ctx->ctx_pmcs[i];
1064                 ia64_set_pmc(i, ctx->th_pmcs[i]);
1065                 DPRINT(("[%d] pmc[%d]=0x%lx\n",
1066                                         task_pid_nr(task), i, ctx->th_pmcs[i]));
1067         }
1068         ia64_srlz_d();
1069 
1070         /*
1071          * must restore DBR/IBR because could be modified while masked
1072          * XXX: need to optimize 
1073          */
1074         if (ctx->ctx_fl_using_dbreg) {
1075                 pfm_restore_ibrs(ctx->ctx_ibrs, pmu_conf->num_ibrs);
1076                 pfm_restore_dbrs(ctx->ctx_dbrs, pmu_conf->num_dbrs);
1077         }
1078 
1079         /*
1080          * now restore PSR
1081          */
1082         if (is_system && (PFM_CPUINFO_GET() & PFM_CPUINFO_DCR_PP)) {
1083                 /* enable dcr pp */
1084                 ia64_setreg(_IA64_REG_CR_DCR, ia64_getreg(_IA64_REG_CR_DCR) | IA64_DCR_PP);
1085                 ia64_srlz_i();
1086         }
1087         pfm_set_psr_l(psr);
1088 }
1089 
1090 static inline void
1091 pfm_save_pmds(unsigned long *pmds, unsigned long mask)
1092 {
1093         int i;
1094 
1095         ia64_srlz_d();
1096 
1097         for (i=0; mask; i++, mask>>=1) {
1098                 if (mask & 0x1) pmds[i] = ia64_get_pmd(i);
1099         }
1100 }
1101 
1102 /*
1103  * reload from thread state (used for ctxw only)
1104  */
1105 static inline void
1106 pfm_restore_pmds(unsigned long *pmds, unsigned long mask)
1107 {
1108         int i;
1109         unsigned long val, ovfl_val = pmu_conf->ovfl_val;
1110 
1111         for (i=0; mask; i++, mask>>=1) {
1112                 if ((mask & 0x1) == 0) continue;
1113                 val = PMD_IS_COUNTING(i) ? pmds[i] & ovfl_val : pmds[i];
1114                 ia64_set_pmd(i, val);
1115         }
1116         ia64_srlz_d();
1117 }
1118 
1119 /*
1120  * propagate PMD from context to thread-state
1121  */
1122 static inline void
1123 pfm_copy_pmds(struct task_struct *task, pfm_context_t *ctx)
1124 {
1125         unsigned long ovfl_val = pmu_conf->ovfl_val;
1126         unsigned long mask = ctx->ctx_all_pmds[0];
1127         unsigned long val;
1128         int i;
1129 
1130         DPRINT(("mask=0x%lx\n", mask));
1131 
1132         for (i=0; mask; i++, mask>>=1) {
1133 
1134                 val = ctx->ctx_pmds[i].val;
1135 
1136                 /*
1137                  * We break up the 64 bit value into 2 pieces
1138                  * the lower bits go to the machine state in the
1139                  * thread (will be reloaded on ctxsw in).
1140                  * The upper part stays in the soft-counter.
1141                  */
1142                 if (PMD_IS_COUNTING(i)) {
1143                         ctx->ctx_pmds[i].val = val & ~ovfl_val;
1144                          val &= ovfl_val;
1145                 }
1146                 ctx->th_pmds[i] = val;
1147 
1148                 DPRINT(("pmd[%d]=0x%lx soft_val=0x%lx\n",
1149                         i,
1150                         ctx->th_pmds[i],
1151                         ctx->ctx_pmds[i].val));
1152         }
1153 }
1154 
1155 /*
1156  * propagate PMC from context to thread-state
1157  */
1158 static inline void
1159 pfm_copy_pmcs(struct task_struct *task, pfm_context_t *ctx)
1160 {
1161         unsigned long mask = ctx->ctx_all_pmcs[0];
1162         int i;
1163 
1164         DPRINT(("mask=0x%lx\n", mask));
1165 
1166         for (i=0; mask; i++, mask>>=1) {
1167                 /* masking 0 with ovfl_val yields 0 */
1168                 ctx->th_pmcs[i] = ctx->ctx_pmcs[i];
1169                 DPRINT(("pmc[%d]=0x%lx\n", i, ctx->th_pmcs[i]));
1170         }
1171 }
1172 
1173 
1174 
1175 static inline void
1176 pfm_restore_pmcs(unsigned long *pmcs, unsigned long mask)
1177 {
1178         int i;
1179 
1180         for (i=0; mask; i++, mask>>=1) {
1181                 if ((mask & 0x1) == 0) continue;
1182                 ia64_set_pmc(i, pmcs[i]);
1183         }
1184         ia64_srlz_d();
1185 }
1186 
1187 static inline int
1188 pfm_uuid_cmp(pfm_uuid_t a, pfm_uuid_t b)
1189 {
1190         return memcmp(a, b, sizeof(pfm_uuid_t));
1191 }
1192 
1193 static inline int
1194 pfm_buf_fmt_exit(pfm_buffer_fmt_t *fmt, struct task_struct *task, void *buf, struct pt_regs *regs)
1195 {
1196         int ret = 0;
1197         if (fmt->fmt_exit) ret = (*fmt->fmt_exit)(task, buf, regs);
1198         return ret;
1199 }
1200 
1201 static inline int
1202 pfm_buf_fmt_getsize(pfm_buffer_fmt_t *fmt, struct task_struct *task, unsigned int flags, int cpu, void *arg, unsigned long *size)
1203 {
1204         int ret = 0;
1205         if (fmt->fmt_getsize) ret = (*fmt->fmt_getsize)(task, flags, cpu, arg, size);
1206         return ret;
1207 }
1208 
1209 
1210 static inline int
1211 pfm_buf_fmt_validate(pfm_buffer_fmt_t *fmt, struct task_struct *task, unsigned int flags,
1212                      int cpu, void *arg)
1213 {
1214         int ret = 0;
1215         if (fmt->fmt_validate) ret = (*fmt->fmt_validate)(task, flags, cpu, arg);
1216         return ret;
1217 }
1218 
1219 static inline int
1220 pfm_buf_fmt_init(pfm_buffer_fmt_t *fmt, struct task_struct *task, void *buf, unsigned int flags,
1221                      int cpu, void *arg)
1222 {
1223         int ret = 0;
1224         if (fmt->fmt_init) ret = (*fmt->fmt_init)(task, buf, flags, cpu, arg);
1225         return ret;
1226 }
1227 
1228 static inline int
1229 pfm_buf_fmt_restart(pfm_buffer_fmt_t *fmt, struct task_struct *task, pfm_ovfl_ctrl_t *ctrl, void *buf, struct pt_regs *regs)
1230 {
1231         int ret = 0;
1232         if (fmt->fmt_restart) ret = (*fmt->fmt_restart)(task, ctrl, buf, regs);
1233         return ret;
1234 }
1235 
1236 static inline int
1237 pfm_buf_fmt_restart_active(pfm_buffer_fmt_t *fmt, struct task_struct *task, pfm_ovfl_ctrl_t *ctrl, void *buf, struct pt_regs *regs)
1238 {
1239         int ret = 0;
1240         if (fmt->fmt_restart_active) ret = (*fmt->fmt_restart_active)(task, ctrl, buf, regs);
1241         return ret;
1242 }
1243 
1244 static pfm_buffer_fmt_t *
1245 __pfm_find_buffer_fmt(pfm_uuid_t uuid)
1246 {
1247         struct list_head * pos;
1248         pfm_buffer_fmt_t * entry;
1249 
1250         list_for_each(pos, &pfm_buffer_fmt_list) {
1251                 entry = list_entry(pos, pfm_buffer_fmt_t, fmt_list);
1252                 if (pfm_uuid_cmp(uuid, entry->fmt_uuid) == 0)
1253                         return entry;
1254         }
1255         return NULL;
1256 }
1257  
1258 /*
1259  * find a buffer format based on its uuid
1260  */
1261 static pfm_buffer_fmt_t *
1262 pfm_find_buffer_fmt(pfm_uuid_t uuid)
1263 {
1264         pfm_buffer_fmt_t * fmt;
1265         spin_lock(&pfm_buffer_fmt_lock);
1266         fmt = __pfm_find_buffer_fmt(uuid);
1267         spin_unlock(&pfm_buffer_fmt_lock);
1268         return fmt;
1269 }
1270  
1271 int
1272 pfm_register_buffer_fmt(pfm_buffer_fmt_t *fmt)
1273 {
1274         int ret = 0;
1275 
1276         /* some sanity checks */
1277         if (fmt == NULL || fmt->fmt_name == NULL) return -EINVAL;
1278 
1279         /* we need at least a handler */
1280         if (fmt->fmt_handler == NULL) return -EINVAL;
1281 
1282         /*
1283          * XXX: need check validity of fmt_arg_size
1284          */
1285 
1286         spin_lock(&pfm_buffer_fmt_lock);
1287 
1288         if (__pfm_find_buffer_fmt(fmt->fmt_uuid)) {
1289                 printk(KERN_ERR "perfmon: duplicate sampling format: %s\n", fmt->fmt_name);
1290                 ret = -EBUSY;
1291                 goto out;
1292         } 
1293         list_add(&fmt->fmt_list, &pfm_buffer_fmt_list);
1294         printk(KERN_INFO "perfmon: added sampling format %s\n", fmt->fmt_name);
1295 
1296 out:
1297         spin_unlock(&pfm_buffer_fmt_lock);
1298         return ret;
1299 }
1300 EXPORT_SYMBOL(pfm_register_buffer_fmt);
1301 
1302 int
1303 pfm_unregister_buffer_fmt(pfm_uuid_t uuid)
1304 {
1305         pfm_buffer_fmt_t *fmt;
1306         int ret = 0;
1307 
1308         spin_lock(&pfm_buffer_fmt_lock);
1309 
1310         fmt = __pfm_find_buffer_fmt(uuid);
1311         if (!fmt) {
1312                 printk(KERN_ERR "perfmon: cannot unregister format, not found\n");
1313                 ret = -EINVAL;
1314                 goto out;
1315         }
1316         list_del_init(&fmt->fmt_list);
1317         printk(KERN_INFO "perfmon: removed sampling format: %s\n", fmt->fmt_name);
1318 
1319 out:
1320         spin_unlock(&pfm_buffer_fmt_lock);
1321         return ret;
1322 
1323 }
1324 EXPORT_SYMBOL(pfm_unregister_buffer_fmt);
1325 
1326 static int
1327 pfm_reserve_session(struct task_struct *task, int is_syswide, unsigned int cpu)
1328 {
1329         unsigned long flags;
1330         /*
1331          * validity checks on cpu_mask have been done upstream
1332          */
1333         LOCK_PFS(flags);
1334 
1335         DPRINT(("in sys_sessions=%u task_sessions=%u dbregs=%u syswide=%d cpu=%u\n",
1336                 pfm_sessions.pfs_sys_sessions,
1337                 pfm_sessions.pfs_task_sessions,
1338                 pfm_sessions.pfs_sys_use_dbregs,
1339                 is_syswide,
1340                 cpu));
1341 
1342         if (is_syswide) {
1343                 /*
1344                  * cannot mix system wide and per-task sessions
1345                  */
1346                 if (pfm_sessions.pfs_task_sessions > 0UL) {
1347                         DPRINT(("system wide not possible, %u conflicting task_sessions\n",
1348                                 pfm_sessions.pfs_task_sessions));
1349                         goto abort;
1350                 }
1351 
1352                 if (pfm_sessions.pfs_sys_session[cpu]) goto error_conflict;
1353 
1354                 DPRINT(("reserving system wide session on CPU%u currently on CPU%u\n", cpu, smp_processor_id()));
1355 
1356                 pfm_sessions.pfs_sys_session[cpu] = task;
1357 
1358                 pfm_sessions.pfs_sys_sessions++ ;
1359 
1360         } else {
1361                 if (pfm_sessions.pfs_sys_sessions) goto abort;
1362                 pfm_sessions.pfs_task_sessions++;
1363         }
1364 
1365         DPRINT(("out sys_sessions=%u task_sessions=%u dbregs=%u syswide=%d cpu=%u\n",
1366                 pfm_sessions.pfs_sys_sessions,
1367                 pfm_sessions.pfs_task_sessions,
1368                 pfm_sessions.pfs_sys_use_dbregs,
1369                 is_syswide,
1370                 cpu));
1371 
1372         /*
1373          * Force idle() into poll mode
1374          */
1375         cpu_idle_poll_ctrl(true);
1376 
1377         UNLOCK_PFS(flags);
1378 
1379         return 0;
1380 
1381 error_conflict:
1382         DPRINT(("system wide not possible, conflicting session [%d] on CPU%d\n",
1383                 task_pid_nr(pfm_sessions.pfs_sys_session[cpu]),
1384                 cpu));
1385 abort:
1386         UNLOCK_PFS(flags);
1387 
1388         return -EBUSY;
1389 
1390 }
1391 
1392 static int
1393 pfm_unreserve_session(pfm_context_t *ctx, int is_syswide, unsigned int cpu)
1394 {
1395         unsigned long flags;
1396         /*
1397          * validity checks on cpu_mask have been done upstream
1398          */
1399         LOCK_PFS(flags);
1400 
1401         DPRINT(("in sys_sessions=%u task_sessions=%u dbregs=%u syswide=%d cpu=%u\n",
1402                 pfm_sessions.pfs_sys_sessions,
1403                 pfm_sessions.pfs_task_sessions,
1404                 pfm_sessions.pfs_sys_use_dbregs,
1405                 is_syswide,
1406                 cpu));
1407 
1408 
1409         if (is_syswide) {
1410                 pfm_sessions.pfs_sys_session[cpu] = NULL;
1411                 /*
1412                  * would not work with perfmon+more than one bit in cpu_mask
1413                  */
1414                 if (ctx && ctx->ctx_fl_using_dbreg) {
1415                         if (pfm_sessions.pfs_sys_use_dbregs == 0) {
1416                                 printk(KERN_ERR "perfmon: invalid release for ctx %p sys_use_dbregs=0\n", ctx);
1417                         } else {
1418                                 pfm_sessions.pfs_sys_use_dbregs--;
1419                         }
1420                 }
1421                 pfm_sessions.pfs_sys_sessions--;
1422         } else {
1423                 pfm_sessions.pfs_task_sessions--;
1424         }
1425         DPRINT(("out sys_sessions=%u task_sessions=%u dbregs=%u syswide=%d cpu=%u\n",
1426                 pfm_sessions.pfs_sys_sessions,
1427                 pfm_sessions.pfs_task_sessions,
1428                 pfm_sessions.pfs_sys_use_dbregs,
1429                 is_syswide,
1430                 cpu));
1431 
1432         /* Undo forced polling. Last session reenables pal_halt */
1433         cpu_idle_poll_ctrl(false);
1434 
1435         UNLOCK_PFS(flags);
1436 
1437         return 0;
1438 }
1439 
1440 /*
1441  * removes virtual mapping of the sampling buffer.
1442  * IMPORTANT: cannot be called with interrupts disable, e.g. inside
1443  * a PROTECT_CTX() section.
1444  */
1445 static int
1446 pfm_remove_smpl_mapping(void *vaddr, unsigned long size)
1447 {
1448         struct task_struct *task = current;
1449         int r;
1450 
1451         /* sanity checks */
1452         if (task->mm == NULL || size == 0UL || vaddr == NULL) {
1453                 printk(KERN_ERR "perfmon: pfm_remove_smpl_mapping [%d] invalid context mm=%p\n", task_pid_nr(task), task->mm);
1454                 return -EINVAL;
1455         }
1456 
1457         DPRINT(("smpl_vaddr=%p size=%lu\n", vaddr, size));
1458 
1459         /*
1460          * does the actual unmapping
1461          */
1462         r = vm_munmap((unsigned long)vaddr, size);
1463 
1464         if (r !=0) {
1465                 printk(KERN_ERR "perfmon: [%d] unable to unmap sampling buffer @%p size=%lu\n", task_pid_nr(task), vaddr, size);
1466         }
1467 
1468         DPRINT(("do_unmap(%p, %lu)=%d\n", vaddr, size, r));
1469 
1470         return 0;
1471 }
1472 
1473 /*
1474  * free actual physical storage used by sampling buffer
1475  */
1476 #if 0
1477 static int
1478 pfm_free_smpl_buffer(pfm_context_t *ctx)
1479 {
1480         pfm_buffer_fmt_t *fmt;
1481 
1482         if (ctx->ctx_smpl_hdr == NULL) goto invalid_free;
1483 
1484         /*
1485          * we won't use the buffer format anymore
1486          */
1487         fmt = ctx->ctx_buf_fmt;
1488 
1489         DPRINT(("sampling buffer @%p size %lu vaddr=%p\n",
1490                 ctx->ctx_smpl_hdr,
1491                 ctx->ctx_smpl_size,
1492                 ctx->ctx_smpl_vaddr));
1493 
1494         pfm_buf_fmt_exit(fmt, current, NULL, NULL);
1495 
1496         /*
1497          * free the buffer
1498          */
1499         pfm_rvfree(ctx->ctx_smpl_hdr, ctx->ctx_smpl_size);
1500 
1501         ctx->ctx_smpl_hdr  = NULL;
1502         ctx->ctx_smpl_size = 0UL;
1503 
1504         return 0;
1505 
1506 invalid_free:
1507         printk(KERN_ERR "perfmon: pfm_free_smpl_buffer [%d] no buffer\n", task_pid_nr(current));
1508         return -EINVAL;
1509 }
1510 #endif
1511 
1512 static inline void
1513 pfm_exit_smpl_buffer(pfm_buffer_fmt_t *fmt)
1514 {
1515         if (fmt == NULL) return;
1516 
1517         pfm_buf_fmt_exit(fmt, current, NULL, NULL);
1518 
1519 }
1520 
1521 /*
1522  * pfmfs should _never_ be mounted by userland - too much of security hassle,
1523  * no real gain from having the whole whorehouse mounted. So we don't need
1524  * any operations on the root directory. However, we need a non-trivial
1525  * d_name - pfm: will go nicely and kill the special-casing in procfs.
1526  */
1527 static struct vfsmount *pfmfs_mnt __read_mostly;
1528 
1529 static int __init
1530 init_pfm_fs(void)
1531 {
1532         int err = register_filesystem(&pfm_fs_type);
1533         if (!err) {
1534                 pfmfs_mnt = kern_mount(&pfm_fs_type);
1535                 err = PTR_ERR(pfmfs_mnt);
1536                 if (IS_ERR(pfmfs_mnt))
1537                         unregister_filesystem(&pfm_fs_type);
1538                 else
1539                         err = 0;
1540         }
1541         return err;
1542 }
1543 
1544 static ssize_t
1545 pfm_read(struct file *filp, char __user *buf, size_t size, loff_t *ppos)
1546 {
1547         pfm_context_t *ctx;
1548         pfm_msg_t *msg;
1549         ssize_t ret;
1550         unsigned long flags;
1551         DECLARE_WAITQUEUE(wait, current);
1552         if (PFM_IS_FILE(filp) == 0) {
1553                 printk(KERN_ERR "perfmon: pfm_poll: bad magic [%d]\n", task_pid_nr(current));
1554                 return -EINVAL;
1555         }
1556 
1557         ctx = filp->private_data;
1558         if (ctx == NULL) {
1559                 printk(KERN_ERR "perfmon: pfm_read: NULL ctx [%d]\n", task_pid_nr(current));
1560                 return -EINVAL;
1561         }
1562 
1563         /*
1564          * check even when there is no message
1565          */
1566         if (size < sizeof(pfm_msg_t)) {
1567                 DPRINT(("message is too small ctx=%p (>=%ld)\n", ctx, sizeof(pfm_msg_t)));
1568                 return -EINVAL;
1569         }
1570 
1571         PROTECT_CTX(ctx, flags);
1572 
1573         /*
1574          * put ourselves on the wait queue
1575          */
1576         add_wait_queue(&ctx->ctx_msgq_wait, &wait);
1577 
1578 
1579         for(;;) {
1580                 /*
1581                  * check wait queue
1582                  */
1583 
1584                 set_current_state(TASK_INTERRUPTIBLE);
1585 
1586                 DPRINT(("head=%d tail=%d\n", ctx->ctx_msgq_head, ctx->ctx_msgq_tail));
1587 
1588                 ret = 0;
1589                 if(PFM_CTXQ_EMPTY(ctx) == 0) break;
1590 
1591                 UNPROTECT_CTX(ctx, flags);
1592 
1593                 /*
1594                  * check non-blocking read
1595                  */
1596                 ret = -EAGAIN;
1597                 if(filp->f_flags & O_NONBLOCK) break;
1598 
1599                 /*
1600                  * check pending signals
1601                  */
1602                 if(signal_pending(current)) {
1603                         ret = -EINTR;
1604                         break;
1605                 }
1606                 /*
1607                  * no message, so wait
1608                  */
1609                 schedule();
1610 
1611                 PROTECT_CTX(ctx, flags);
1612         }
1613         DPRINT(("[%d] back to running ret=%ld\n", task_pid_nr(current), ret));
1614         set_current_state(TASK_RUNNING);
1615         remove_wait_queue(&ctx->ctx_msgq_wait, &wait);
1616 
1617         if (ret < 0) goto abort;
1618 
1619         ret = -EINVAL;
1620         msg = pfm_get_next_msg(ctx);
1621         if (msg == NULL) {
1622                 printk(KERN_ERR "perfmon: pfm_read no msg for ctx=%p [%d]\n", ctx, task_pid_nr(current));
1623                 goto abort_locked;
1624         }
1625 
1626         DPRINT(("fd=%d type=%d\n", msg->pfm_gen_msg.msg_ctx_fd, msg->pfm_gen_msg.msg_type));
1627 
1628         ret = -EFAULT;
1629         if(copy_to_user(buf, msg, sizeof(pfm_msg_t)) == 0) ret = sizeof(pfm_msg_t);
1630 
1631 abort_locked:
1632         UNPROTECT_CTX(ctx, flags);
1633 abort:
1634         return ret;
1635 }
1636 
1637 static ssize_t
1638 pfm_write(struct file *file, const char __user *ubuf,
1639                           size_t size, loff_t *ppos)
1640 {
1641         DPRINT(("pfm_write called\n"));
1642         return -EINVAL;
1643 }
1644 
1645 static unsigned int
1646 pfm_poll(struct file *filp, poll_table * wait)
1647 {
1648         pfm_context_t *ctx;
1649         unsigned long flags;
1650         unsigned int mask = 0;
1651 
1652         if (PFM_IS_FILE(filp) == 0) {
1653                 printk(KERN_ERR "perfmon: pfm_poll: bad magic [%d]\n", task_pid_nr(current));
1654                 return 0;
1655         }
1656 
1657         ctx = filp->private_data;
1658         if (ctx == NULL) {
1659                 printk(KERN_ERR "perfmon: pfm_poll: NULL ctx [%d]\n", task_pid_nr(current));
1660                 return 0;
1661         }
1662 
1663 
1664         DPRINT(("pfm_poll ctx_fd=%d before poll_wait\n", ctx->ctx_fd));
1665 
1666         poll_wait(filp, &ctx->ctx_msgq_wait, wait);
1667 
1668         PROTECT_CTX(ctx, flags);
1669 
1670         if (PFM_CTXQ_EMPTY(ctx) == 0)
1671                 mask =  POLLIN | POLLRDNORM;
1672 
1673         UNPROTECT_CTX(ctx, flags);
1674 
1675         DPRINT(("pfm_poll ctx_fd=%d mask=0x%x\n", ctx->ctx_fd, mask));
1676 
1677         return mask;
1678 }
1679 
1680 static long
1681 pfm_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
1682 {
1683         DPRINT(("pfm_ioctl called\n"));
1684         return -EINVAL;
1685 }
1686 
1687 /*
1688  * interrupt cannot be masked when coming here
1689  */
1690 static inline int
1691 pfm_do_fasync(int fd, struct file *filp, pfm_context_t *ctx, int on)
1692 {
1693         int ret;
1694 
1695         ret = fasync_helper (fd, filp, on, &ctx->ctx_async_queue);
1696 
1697         DPRINT(("pfm_fasync called by [%d] on ctx_fd=%d on=%d async_queue=%p ret=%d\n",
1698                 task_pid_nr(current),
1699                 fd,
1700                 on,
1701                 ctx->ctx_async_queue, ret));
1702 
1703         return ret;
1704 }
1705 
1706 static int
1707 pfm_fasync(int fd, struct file *filp, int on)
1708 {
1709         pfm_context_t *ctx;
1710         int ret;
1711 
1712         if (PFM_IS_FILE(filp) == 0) {
1713                 printk(KERN_ERR "perfmon: pfm_fasync bad magic [%d]\n", task_pid_nr(current));
1714                 return -EBADF;
1715         }
1716 
1717         ctx = filp->private_data;
1718         if (ctx == NULL) {
1719                 printk(KERN_ERR "perfmon: pfm_fasync NULL ctx [%d]\n", task_pid_nr(current));
1720                 return -EBADF;
1721         }
1722         /*
1723          * we cannot mask interrupts during this call because this may
1724          * may go to sleep if memory is not readily avalaible.
1725          *
1726          * We are protected from the conetxt disappearing by the get_fd()/put_fd()
1727          * done in caller. Serialization of this function is ensured by caller.
1728          */
1729         ret = pfm_do_fasync(fd, filp, ctx, on);
1730 
1731 
1732         DPRINT(("pfm_fasync called on ctx_fd=%d on=%d async_queue=%p ret=%d\n",
1733                 fd,
1734                 on,
1735                 ctx->ctx_async_queue, ret));
1736 
1737         return ret;
1738 }
1739 
1740 #ifdef CONFIG_SMP
1741 /*
1742  * this function is exclusively called from pfm_close().
1743  * The context is not protected at that time, nor are interrupts
1744  * on the remote CPU. That's necessary to avoid deadlocks.
1745  */
1746 static void
1747 pfm_syswide_force_stop(void *info)
1748 {
1749         pfm_context_t   *ctx = (pfm_context_t *)info;
1750         struct pt_regs *regs = task_pt_regs(current);
1751         struct task_struct *owner;
1752         unsigned long flags;
1753         int ret;
1754 
1755         if (ctx->ctx_cpu != smp_processor_id()) {
1756                 printk(KERN_ERR "perfmon: pfm_syswide_force_stop for CPU%d  but on CPU%d\n",
1757                         ctx->ctx_cpu,
1758                         smp_processor_id());
1759                 return;
1760         }
1761         owner = GET_PMU_OWNER();
1762         if (owner != ctx->ctx_task) {
1763                 printk(KERN_ERR "perfmon: pfm_syswide_force_stop CPU%d unexpected owner [%d] instead of [%d]\n",
1764                         smp_processor_id(),
1765                         task_pid_nr(owner), task_pid_nr(ctx->ctx_task));
1766                 return;
1767         }
1768         if (GET_PMU_CTX() != ctx) {
1769                 printk(KERN_ERR "perfmon: pfm_syswide_force_stop CPU%d unexpected ctx %p instead of %p\n",
1770                         smp_processor_id(),
1771                         GET_PMU_CTX(), ctx);
1772                 return;
1773         }
1774 
1775         DPRINT(("on CPU%d forcing system wide stop for [%d]\n", smp_processor_id(), task_pid_nr(ctx->ctx_task)));
1776         /*
1777          * the context is already protected in pfm_close(), we simply
1778          * need to mask interrupts to avoid a PMU interrupt race on
1779          * this CPU
1780          */
1781         local_irq_save(flags);
1782 
1783         ret = pfm_context_unload(ctx, NULL, 0, regs);
1784         if (ret) {
1785                 DPRINT(("context_unload returned %d\n", ret));
1786         }
1787 
1788         /*
1789          * unmask interrupts, PMU interrupts are now spurious here
1790          */
1791         local_irq_restore(flags);
1792 }
1793 
1794 static void
1795 pfm_syswide_cleanup_other_cpu(pfm_context_t *ctx)
1796 {
1797         int ret;
1798 
1799         DPRINT(("calling CPU%d for cleanup\n", ctx->ctx_cpu));
1800         ret = smp_call_function_single(ctx->ctx_cpu, pfm_syswide_force_stop, ctx, 1);
1801         DPRINT(("called CPU%d for cleanup ret=%d\n", ctx->ctx_cpu, ret));
1802 }
1803 #endif /* CONFIG_SMP */
1804 
1805 /*
1806  * called for each close(). Partially free resources.
1807  * When caller is self-monitoring, the context is unloaded.
1808  */
1809 static int
1810 pfm_flush(struct file *filp, fl_owner_t id)
1811 {
1812         pfm_context_t *ctx;
1813         struct task_struct *task;
1814         struct pt_regs *regs;
1815         unsigned long flags;
1816         unsigned long smpl_buf_size = 0UL;
1817         void *smpl_buf_vaddr = NULL;
1818         int state, is_system;
1819 
1820         if (PFM_IS_FILE(filp) == 0) {
1821                 DPRINT(("bad magic for\n"));
1822                 return -EBADF;
1823         }
1824 
1825         ctx = filp->private_data;
1826         if (ctx == NULL) {
1827                 printk(KERN_ERR "perfmon: pfm_flush: NULL ctx [%d]\n", task_pid_nr(current));
1828                 return -EBADF;
1829         }
1830 
1831         /*
1832          * remove our file from the async queue, if we use this mode.
1833          * This can be done without the context being protected. We come
1834          * here when the context has become unreachable by other tasks.
1835          *
1836          * We may still have active monitoring at this point and we may
1837          * end up in pfm_overflow_handler(). However, fasync_helper()
1838          * operates with interrupts disabled and it cleans up the
1839          * queue. If the PMU handler is called prior to entering
1840          * fasync_helper() then it will send a signal. If it is
1841          * invoked after, it will find an empty queue and no
1842          * signal will be sent. In both case, we are safe
1843          */
1844         PROTECT_CTX(ctx, flags);
1845 
1846         state     = ctx->ctx_state;
1847         is_system = ctx->ctx_fl_system;
1848 
1849         task = PFM_CTX_TASK(ctx);
1850         regs = task_pt_regs(task);
1851 
1852         DPRINT(("ctx_state=%d is_current=%d\n",
1853                 state,
1854                 task == current ? 1 : 0));
1855 
1856         /*
1857          * if state == UNLOADED, then task is NULL
1858          */
1859 
1860         /*
1861          * we must stop and unload because we are losing access to the context.
1862          */
1863         if (task == current) {
1864 #ifdef CONFIG_SMP
1865                 /*
1866                  * the task IS the owner but it migrated to another CPU: that's bad
1867                  * but we must handle this cleanly. Unfortunately, the kernel does
1868                  * not provide a mechanism to block migration (while the context is loaded).
1869                  *
1870                  * We need to release the resource on the ORIGINAL cpu.
1871                  */
1872                 if (is_system && ctx->ctx_cpu != smp_processor_id()) {
1873 
1874                         DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
1875                         /*
1876                          * keep context protected but unmask interrupt for IPI
1877                          */
1878                         local_irq_restore(flags);
1879 
1880                         pfm_syswide_cleanup_other_cpu(ctx);
1881 
1882                         /*
1883                          * restore interrupt masking
1884                          */
1885                         local_irq_save(flags);
1886 
1887                         /*
1888                          * context is unloaded at this point
1889                          */
1890                 } else
1891 #endif /* CONFIG_SMP */
1892                 {
1893 
1894                         DPRINT(("forcing unload\n"));
1895                         /*
1896                         * stop and unload, returning with state UNLOADED
1897                         * and session unreserved.
1898                         */
1899                         pfm_context_unload(ctx, NULL, 0, regs);
1900 
1901                         DPRINT(("ctx_state=%d\n", ctx->ctx_state));
1902                 }
1903         }
1904 
1905         /*
1906          * remove virtual mapping, if any, for the calling task.
1907          * cannot reset ctx field until last user is calling close().
1908          *
1909          * ctx_smpl_vaddr must never be cleared because it is needed
1910          * by every task with access to the context
1911          *
1912          * When called from do_exit(), the mm context is gone already, therefore
1913          * mm is NULL, i.e., the VMA is already gone  and we do not have to
1914          * do anything here
1915          */
1916         if (ctx->ctx_smpl_vaddr && current->mm) {
1917                 smpl_buf_vaddr = ctx->ctx_smpl_vaddr;
1918                 smpl_buf_size  = ctx->ctx_smpl_size;
1919         }
1920 
1921         UNPROTECT_CTX(ctx, flags);
1922 
1923         /*
1924          * if there was a mapping, then we systematically remove it
1925          * at this point. Cannot be done inside critical section
1926          * because some VM function reenables interrupts.
1927          *
1928          */
1929         if (smpl_buf_vaddr) pfm_remove_smpl_mapping(smpl_buf_vaddr, smpl_buf_size);
1930 
1931         return 0;
1932 }
1933 /*
1934  * called either on explicit close() or from exit_files(). 
1935  * Only the LAST user of the file gets to this point, i.e., it is
1936  * called only ONCE.
1937  *
1938  * IMPORTANT: we get called ONLY when the refcnt on the file gets to zero 
1939  * (fput()),i.e, last task to access the file. Nobody else can access the 
1940  * file at this point.
1941  *
1942  * When called from exit_files(), the VMA has been freed because exit_mm()
1943  * is executed before exit_files().
1944  *
1945  * When called from exit_files(), the current task is not yet ZOMBIE but we
1946  * flush the PMU state to the context. 
1947  */
1948 static int
1949 pfm_close(struct inode *inode, struct file *filp)
1950 {
1951         pfm_context_t *ctx;
1952         struct task_struct *task;
1953         struct pt_regs *regs;
1954         DECLARE_WAITQUEUE(wait, current);
1955         unsigned long flags;
1956         unsigned long smpl_buf_size = 0UL;
1957         void *smpl_buf_addr = NULL;
1958         int free_possible = 1;
1959         int state, is_system;
1960 
1961         DPRINT(("pfm_close called private=%p\n", filp->private_data));
1962 
1963         if (PFM_IS_FILE(filp) == 0) {
1964                 DPRINT(("bad magic\n"));
1965                 return -EBADF;
1966         }
1967         
1968         ctx = filp->private_data;
1969         if (ctx == NULL) {
1970                 printk(KERN_ERR "perfmon: pfm_close: NULL ctx [%d]\n", task_pid_nr(current));
1971                 return -EBADF;
1972         }
1973 
1974         PROTECT_CTX(ctx, flags);
1975 
1976         state     = ctx->ctx_state;
1977         is_system = ctx->ctx_fl_system;
1978 
1979         task = PFM_CTX_TASK(ctx);
1980         regs = task_pt_regs(task);
1981 
1982         DPRINT(("ctx_state=%d is_current=%d\n", 
1983                 state,
1984                 task == current ? 1 : 0));
1985 
1986         /*
1987          * if task == current, then pfm_flush() unloaded the context
1988          */
1989         if (state == PFM_CTX_UNLOADED) goto doit;
1990 
1991         /*
1992          * context is loaded/masked and task != current, we need to
1993          * either force an unload or go zombie
1994          */
1995 
1996         /*
1997          * The task is currently blocked or will block after an overflow.
1998          * we must force it to wakeup to get out of the
1999          * MASKED state and transition to the unloaded state by itself.
2000          *
2001          * This situation is only possible for per-task mode
2002          */
2003         if (state == PFM_CTX_MASKED && CTX_OVFL_NOBLOCK(ctx) == 0) {
2004 
2005                 /*
2006                  * set a "partial" zombie state to be checked
2007                  * upon return from down() in pfm_handle_work().
2008                  *
2009                  * We cannot use the ZOMBIE state, because it is checked
2010                  * by pfm_load_regs() which is called upon wakeup from down().
2011                  * In such case, it would free the context and then we would
2012                  * return to pfm_handle_work() which would access the
2013                  * stale context. Instead, we set a flag invisible to pfm_load_regs()
2014                  * but visible to pfm_handle_work().
2015                  *
2016                  * For some window of time, we have a zombie context with
2017                  * ctx_state = MASKED  and not ZOMBIE
2018                  */
2019                 ctx->ctx_fl_going_zombie = 1;
2020 
2021                 /*
2022                  * force task to wake up from MASKED state
2023                  */
2024                 complete(&ctx->ctx_restart_done);
2025 
2026                 DPRINT(("waking up ctx_state=%d\n", state));
2027 
2028                 /*
2029                  * put ourself to sleep waiting for the other
2030                  * task to report completion
2031                  *
2032                  * the context is protected by mutex, therefore there
2033                  * is no risk of being notified of completion before
2034                  * begin actually on the waitq.
2035                  */
2036                 set_current_state(TASK_INTERRUPTIBLE);
2037                 add_wait_queue(&ctx->ctx_zombieq, &wait);
2038 
2039                 UNPROTECT_CTX(ctx, flags);
2040 
2041                 /*
2042                  * XXX: check for signals :
2043                  *      - ok for explicit close
2044                  *      - not ok when coming from exit_files()
2045                  */
2046                 schedule();
2047 
2048 
2049                 PROTECT_CTX(ctx, flags);
2050 
2051 
2052                 remove_wait_queue(&ctx->ctx_zombieq, &wait);
2053                 set_current_state(TASK_RUNNING);
2054 
2055                 /*
2056                  * context is unloaded at this point
2057                  */
2058                 DPRINT(("after zombie wakeup ctx_state=%d for\n", state));
2059         }
2060         else if (task != current) {
2061 #ifdef CONFIG_SMP
2062                 /*
2063                  * switch context to zombie state
2064                  */
2065                 ctx->ctx_state = PFM_CTX_ZOMBIE;
2066 
2067                 DPRINT(("zombie ctx for [%d]\n", task_pid_nr(task)));
2068                 /*
2069                  * cannot free the context on the spot. deferred until
2070                  * the task notices the ZOMBIE state
2071                  */
2072                 free_possible = 0;
2073 #else
2074                 pfm_context_unload(ctx, NULL, 0, regs);
2075 #endif
2076         }
2077 
2078 doit:
2079         /* reload state, may have changed during  opening of critical section */
2080         state = ctx->ctx_state;
2081 
2082         /*
2083          * the context is still attached to a task (possibly current)
2084          * we cannot destroy it right now
2085          */
2086 
2087         /*
2088          * we must free the sampling buffer right here because
2089          * we cannot rely on it being cleaned up later by the
2090          * monitored task. It is not possible to free vmalloc'ed
2091          * memory in pfm_load_regs(). Instead, we remove the buffer
2092          * now. should there be subsequent PMU overflow originally
2093          * meant for sampling, the will be converted to spurious
2094          * and that's fine because the monitoring tools is gone anyway.
2095          */
2096         if (ctx->ctx_smpl_hdr) {
2097                 smpl_buf_addr = ctx->ctx_smpl_hdr;
2098                 smpl_buf_size = ctx->ctx_smpl_size;
2099                 /* no more sampling */
2100                 ctx->ctx_smpl_hdr = NULL;
2101                 ctx->ctx_fl_is_sampling = 0;
2102         }
2103 
2104         DPRINT(("ctx_state=%d free_possible=%d addr=%p size=%lu\n",
2105                 state,
2106                 free_possible,
2107                 smpl_buf_addr,
2108                 smpl_buf_size));
2109 
2110         if (smpl_buf_addr) pfm_exit_smpl_buffer(ctx->ctx_buf_fmt);
2111 
2112         /*
2113          * UNLOADED that the session has already been unreserved.
2114          */
2115         if (state == PFM_CTX_ZOMBIE) {
2116                 pfm_unreserve_session(ctx, ctx->ctx_fl_system , ctx->ctx_cpu);
2117         }
2118 
2119         /*
2120          * disconnect file descriptor from context must be done
2121          * before we unlock.
2122          */
2123         filp->private_data = NULL;
2124 
2125         /*
2126          * if we free on the spot, the context is now completely unreachable
2127          * from the callers side. The monitored task side is also cut, so we
2128          * can freely cut.
2129          *
2130          * If we have a deferred free, only the caller side is disconnected.
2131          */
2132         UNPROTECT_CTX(ctx, flags);
2133 
2134         /*
2135          * All memory free operations (especially for vmalloc'ed memory)
2136          * MUST be done with interrupts ENABLED.
2137          */
2138         if (smpl_buf_addr)  pfm_rvfree(smpl_buf_addr, smpl_buf_size);
2139 
2140         /*
2141          * return the memory used by the context
2142          */
2143         if (free_possible) pfm_context_free(ctx);
2144 
2145         return 0;
2146 }
2147 
2148 static const struct file_operations pfm_file_ops = {
2149         .llseek         = no_llseek,
2150         .read           = pfm_read,
2151         .write          = pfm_write,
2152         .poll           = pfm_poll,
2153         .unlocked_ioctl = pfm_ioctl,
2154         .fasync         = pfm_fasync,
2155         .release        = pfm_close,
2156         .flush          = pfm_flush
2157 };
2158 
2159 static char *pfmfs_dname(struct dentry *dentry, char *buffer, int buflen)
2160 {
2161         return dynamic_dname(dentry, buffer, buflen, "pfm:[%lu]",
2162                              d_inode(dentry)->i_ino);
2163 }
2164 
2165 static const struct dentry_operations pfmfs_dentry_operations = {
2166         .d_delete = always_delete_dentry,
2167         .d_dname = pfmfs_dname,
2168 };
2169 
2170 
2171 static struct file *
2172 pfm_alloc_file(pfm_context_t *ctx)
2173 {
2174         struct file *file;
2175         struct inode *inode;
2176         struct path path;
2177         struct qstr this = { .name = "" };
2178 
2179         /*
2180          * allocate a new inode
2181          */
2182         inode = new_inode(pfmfs_mnt->mnt_sb);
2183         if (!inode)
2184                 return ERR_PTR(-ENOMEM);
2185 
2186         DPRINT(("new inode ino=%ld @%p\n", inode->i_ino, inode));
2187 
2188         inode->i_mode = S_IFCHR|S_IRUGO;
2189         inode->i_uid  = current_fsuid();
2190         inode->i_gid  = current_fsgid();
2191 
2192         /*
2193          * allocate a new dcache entry
2194          */
2195         path.dentry = d_alloc(pfmfs_mnt->mnt_root, &this);
2196         if (!path.dentry) {
2197                 iput(inode);
2198                 return ERR_PTR(-ENOMEM);
2199         }
2200         path.mnt = mntget(pfmfs_mnt);
2201 
2202         d_add(path.dentry, inode);
2203 
2204         file = alloc_file(&path, FMODE_READ, &pfm_file_ops);
2205         if (IS_ERR(file)) {
2206                 path_put(&path);
2207                 return file;
2208         }
2209 
2210         file->f_flags = O_RDONLY;
2211         file->private_data = ctx;
2212 
2213         return file;
2214 }
2215 
2216 static int
2217 pfm_remap_buffer(struct vm_area_struct *vma, unsigned long buf, unsigned long addr, unsigned long size)
2218 {
2219         DPRINT(("CPU%d buf=0x%lx addr=0x%lx size=%ld\n", smp_processor_id(), buf, addr, size));
2220 
2221         while (size > 0) {
2222                 unsigned long pfn = ia64_tpa(buf) >> PAGE_SHIFT;
2223 
2224 
2225                 if (remap_pfn_range(vma, addr, pfn, PAGE_SIZE, PAGE_READONLY))
2226                         return -ENOMEM;
2227 
2228                 addr  += PAGE_SIZE;
2229                 buf   += PAGE_SIZE;
2230                 size  -= PAGE_SIZE;
2231         }
2232         return 0;
2233 }
2234 
2235 /*
2236  * allocate a sampling buffer and remaps it into the user address space of the task
2237  */
2238 static int
2239 pfm_smpl_buffer_alloc(struct task_struct *task, struct file *filp, pfm_context_t *ctx, unsigned long rsize, void **user_vaddr)
2240 {
2241         struct mm_struct *mm = task->mm;
2242         struct vm_area_struct *vma = NULL;
2243         unsigned long size;
2244         void *smpl_buf;
2245 
2246 
2247         /*
2248          * the fixed header + requested size and align to page boundary
2249          */
2250         size = PAGE_ALIGN(rsize);
2251 
2252         DPRINT(("sampling buffer rsize=%lu size=%lu bytes\n", rsize, size));
2253 
2254         /*
2255          * check requested size to avoid Denial-of-service attacks
2256          * XXX: may have to refine this test
2257          * Check against address space limit.
2258          *
2259          * if ((mm->total_vm << PAGE_SHIFT) + len> task->rlim[RLIMIT_AS].rlim_cur)
2260          *      return -ENOMEM;
2261          */
2262         if (size > task_rlimit(task, RLIMIT_MEMLOCK))
2263                 return -ENOMEM;
2264 
2265         /*
2266          * We do the easy to undo allocations first.
2267          *
2268          * pfm_rvmalloc(), clears the buffer, so there is no leak
2269          */
2270         smpl_buf = pfm_rvmalloc(size);
2271         if (smpl_buf == NULL) {
2272                 DPRINT(("Can't allocate sampling buffer\n"));
2273                 return -ENOMEM;
2274         }
2275 
2276         DPRINT(("smpl_buf @%p\n", smpl_buf));
2277 
2278         /* allocate vma */
2279         vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
2280         if (!vma) {
2281                 DPRINT(("Cannot allocate vma\n"));
2282                 goto error_kmem;
2283         }
2284         INIT_LIST_HEAD(&vma->anon_vma_chain);
2285 
2286         /*
2287          * partially initialize the vma for the sampling buffer
2288          */
2289         vma->vm_mm           = mm;
2290         vma->vm_file         = get_file(filp);
2291         vma->vm_flags        = VM_READ|VM_MAYREAD|VM_DONTEXPAND|VM_DONTDUMP;
2292         vma->vm_page_prot    = PAGE_READONLY; /* XXX may need to change */
2293 
2294         /*
2295          * Now we have everything we need and we can initialize
2296          * and connect all the data structures
2297          */
2298 
2299         ctx->ctx_smpl_hdr   = smpl_buf;
2300         ctx->ctx_smpl_size  = size; /* aligned size */
2301 
2302         /*
2303          * Let's do the difficult operations next.
2304          *
2305          * now we atomically find some area in the address space and
2306          * remap the buffer in it.
2307          */
2308         down_write(&task->mm->mmap_sem);
2309 
2310         /* find some free area in address space, must have mmap sem held */
2311         vma->vm_start = get_unmapped_area(NULL, 0, size, 0, MAP_PRIVATE|MAP_ANONYMOUS);
2312         if (IS_ERR_VALUE(vma->vm_start)) {
2313                 DPRINT(("Cannot find unmapped area for size %ld\n", size));
2314                 up_write(&task->mm->mmap_sem);
2315                 goto error;
2316         }
2317         vma->vm_end = vma->vm_start + size;
2318         vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT;
2319 
2320         DPRINT(("aligned size=%ld, hdr=%p mapped @0x%lx\n", size, ctx->ctx_smpl_hdr, vma->vm_start));
2321 
2322         /* can only be applied to current task, need to have the mm semaphore held when called */
2323         if (pfm_remap_buffer(vma, (unsigned long)smpl_buf, vma->vm_start, size)) {
2324                 DPRINT(("Can't remap buffer\n"));
2325                 up_write(&task->mm->mmap_sem);
2326                 goto error;
2327         }
2328 
2329         /*
2330          * now insert the vma in the vm list for the process, must be
2331          * done with mmap lock held
2332          */
2333         insert_vm_struct(mm, vma);
2334 
2335         vm_stat_account(vma->vm_mm, vma->vm_flags, vma_pages(vma));
2336         up_write(&task->mm->mmap_sem);
2337 
2338         /*
2339          * keep track of user level virtual address
2340          */
2341         ctx->ctx_smpl_vaddr = (void *)vma->vm_start;
2342         *(unsigned long *)user_vaddr = vma->vm_start;
2343 
2344         return 0;
2345 
2346 error:
2347         kmem_cache_free(vm_area_cachep, vma);
2348 error_kmem:
2349         pfm_rvfree(smpl_buf, size);
2350 
2351         return -ENOMEM;
2352 }
2353 
2354 /*
2355  * XXX: do something better here
2356  */
2357 static int
2358 pfm_bad_permissions(struct task_struct *task)
2359 {
2360         const struct cred *tcred;
2361         kuid_t uid = current_uid();
2362         kgid_t gid = current_gid();
2363         int ret;
2364 
2365         rcu_read_lock();
2366         tcred = __task_cred(task);
2367 
2368         /* inspired by ptrace_attach() */
2369         DPRINT(("cur: uid=%d gid=%d task: euid=%d suid=%d uid=%d egid=%d sgid=%d\n",
2370                 from_kuid(&init_user_ns, uid),
2371                 from_kgid(&init_user_ns, gid),
2372                 from_kuid(&init_user_ns, tcred->euid),
2373                 from_kuid(&init_user_ns, tcred->suid),
2374                 from_kuid(&init_user_ns, tcred->uid),
2375                 from_kgid(&init_user_ns, tcred->egid),
2376                 from_kgid(&init_user_ns, tcred->sgid)));
2377 
2378         ret = ((!uid_eq(uid, tcred->euid))
2379                || (!uid_eq(uid, tcred->suid))
2380                || (!uid_eq(uid, tcred->uid))
2381                || (!gid_eq(gid, tcred->egid))
2382                || (!gid_eq(gid, tcred->sgid))
2383                || (!gid_eq(gid, tcred->gid))) && !capable(CAP_SYS_PTRACE);
2384 
2385         rcu_read_unlock();
2386         return ret;
2387 }
2388 
2389 static int
2390 pfarg_is_sane(struct task_struct *task, pfarg_context_t *pfx)
2391 {
2392         int ctx_flags;
2393 
2394         /* valid signal */
2395 
2396         ctx_flags = pfx->ctx_flags;
2397 
2398         if (ctx_flags & PFM_FL_SYSTEM_WIDE) {
2399 
2400                 /*
2401                  * cannot block in this mode
2402                  */
2403                 if (ctx_flags & PFM_FL_NOTIFY_BLOCK) {
2404                         DPRINT(("cannot use blocking mode when in system wide monitoring\n"));
2405                         return -EINVAL;
2406                 }
2407         } else {
2408         }
2409         /* probably more to add here */
2410 
2411         return 0;
2412 }
2413 
2414 static int
2415 pfm_setup_buffer_fmt(struct task_struct *task, struct file *filp, pfm_context_t *ctx, unsigned int ctx_flags,
2416                      unsigned int cpu, pfarg_context_t *arg)
2417 {
2418         pfm_buffer_fmt_t *fmt = NULL;
2419         unsigned long size = 0UL;
2420         void *uaddr = NULL;
2421         void *fmt_arg = NULL;
2422         int ret = 0;
2423 #define PFM_CTXARG_BUF_ARG(a)   (pfm_buffer_fmt_t *)(a+1)
2424 
2425         /* invoke and lock buffer format, if found */
2426         fmt = pfm_find_buffer_fmt(arg->ctx_smpl_buf_id);
2427         if (fmt == NULL) {
2428                 DPRINT(("[%d] cannot find buffer format\n", task_pid_nr(task)));
2429                 return -EINVAL;
2430         }
2431 
2432         /*
2433          * buffer argument MUST be contiguous to pfarg_context_t
2434          */
2435         if (fmt->fmt_arg_size) fmt_arg = PFM_CTXARG_BUF_ARG(arg);
2436 
2437         ret = pfm_buf_fmt_validate(fmt, task, ctx_flags, cpu, fmt_arg);
2438 
2439         DPRINT(("[%d] after validate(0x%x,%d,%p)=%d\n", task_pid_nr(task), ctx_flags, cpu, fmt_arg, ret));
2440 
2441         if (ret) goto error;
2442 
2443         /* link buffer format and context */
2444         ctx->ctx_buf_fmt = fmt;
2445         ctx->ctx_fl_is_sampling = 1; /* assume record() is defined */
2446 
2447         /*
2448          * check if buffer format wants to use perfmon buffer allocation/mapping service
2449          */
2450         ret = pfm_buf_fmt_getsize(fmt, task, ctx_flags, cpu, fmt_arg, &size);
2451         if (ret) goto error;
2452 
2453         if (size) {
2454                 /*
2455                  * buffer is always remapped into the caller's address space
2456                  */
2457                 ret = pfm_smpl_buffer_alloc(current, filp, ctx, size, &uaddr);
2458                 if (ret) goto error;
2459 
2460                 /* keep track of user address of buffer */
2461                 arg->ctx_smpl_vaddr = uaddr;
2462         }
2463         ret = pfm_buf_fmt_init(fmt, task, ctx->ctx_smpl_hdr, ctx_flags, cpu, fmt_arg);
2464 
2465 error:
2466         return ret;
2467 }
2468 
2469 static void
2470 pfm_reset_pmu_state(pfm_context_t *ctx)
2471 {
2472         int i;
2473 
2474         /*
2475          * install reset values for PMC.
2476          */
2477         for (i=1; PMC_IS_LAST(i) == 0; i++) {
2478                 if (PMC_IS_IMPL(i) == 0) continue;
2479                 ctx->ctx_pmcs[i] = PMC_DFL_VAL(i);
2480                 DPRINT(("pmc[%d]=0x%lx\n", i, ctx->ctx_pmcs[i]));
2481         }
2482         /*
2483          * PMD registers are set to 0UL when the context in memset()
2484          */
2485 
2486         /*
2487          * On context switched restore, we must restore ALL pmc and ALL pmd even
2488          * when they are not actively used by the task. In UP, the incoming process
2489          * may otherwise pick up left over PMC, PMD state from the previous process.
2490          * As opposed to PMD, stale PMC can cause harm to the incoming
2491          * process because they may change what is being measured.
2492          * Therefore, we must systematically reinstall the entire
2493          * PMC state. In SMP, the same thing is possible on the
2494          * same CPU but also on between 2 CPUs.
2495          *
2496          * The problem with PMD is information leaking especially
2497          * to user level when psr.sp=0
2498          *
2499          * There is unfortunately no easy way to avoid this problem
2500          * on either UP or SMP. This definitively slows down the
2501          * pfm_load_regs() function.
2502          */
2503 
2504          /*
2505           * bitmask of all PMCs accessible to this context
2506           *
2507           * PMC0 is treated differently.
2508           */
2509         ctx->ctx_all_pmcs[0] = pmu_conf->impl_pmcs[0] & ~0x1;
2510 
2511         /*
2512          * bitmask of all PMDs that are accessible to this context
2513          */
2514         ctx->ctx_all_pmds[0] = pmu_conf->impl_pmds[0];
2515 
2516         DPRINT(("<%d> all_pmcs=0x%lx all_pmds=0x%lx\n", ctx->ctx_fd, ctx->ctx_all_pmcs[0],ctx->ctx_all_pmds[0]));
2517 
2518         /*
2519          * useful in case of re-enable after disable
2520          */
2521         ctx->ctx_used_ibrs[0] = 0UL;
2522         ctx->ctx_used_dbrs[0] = 0UL;
2523 }
2524 
2525 static int
2526 pfm_ctx_getsize(void *arg, size_t *sz)
2527 {
2528         pfarg_context_t *req = (pfarg_context_t *)arg;
2529         pfm_buffer_fmt_t *fmt;
2530 
2531         *sz = 0;
2532 
2533         if (!pfm_uuid_cmp(req->ctx_smpl_buf_id, pfm_null_uuid)) return 0;
2534 
2535         fmt = pfm_find_buffer_fmt(req->ctx_smpl_buf_id);
2536         if (fmt == NULL) {
2537                 DPRINT(("cannot find buffer format\n"));
2538                 return -EINVAL;
2539         }
2540         /* get just enough to copy in user parameters */
2541         *sz = fmt->fmt_arg_size;
2542         DPRINT(("arg_size=%lu\n", *sz));
2543 
2544         return 0;
2545 }
2546 
2547 
2548 
2549 /*
2550  * cannot attach if :
2551  *      - kernel task
2552  *      - task not owned by caller
2553  *      - task incompatible with context mode
2554  */
2555 static int
2556 pfm_task_incompatible(pfm_context_t *ctx, struct task_struct *task)
2557 {
2558         /*
2559          * no kernel task or task not owner by caller
2560          */
2561         if (task->mm == NULL) {
2562                 DPRINT(("task [%d] has not memory context (kernel thread)\n", task_pid_nr(task)));
2563                 return -EPERM;
2564         }
2565         if (pfm_bad_permissions(task)) {
2566                 DPRINT(("no permission to attach to  [%d]\n", task_pid_nr(task)));
2567                 return -EPERM;
2568         }
2569         /*
2570          * cannot block in self-monitoring mode
2571          */
2572         if (CTX_OVFL_NOBLOCK(ctx) == 0 && task == current) {
2573                 DPRINT(("cannot load a blocking context on self for [%d]\n", task_pid_nr(task)));
2574                 return -EINVAL;
2575         }
2576 
2577         if (task->exit_state == EXIT_ZOMBIE) {
2578                 DPRINT(("cannot attach to  zombie task [%d]\n", task_pid_nr(task)));
2579                 return -EBUSY;
2580         }
2581 
2582         /*
2583          * always ok for self
2584          */
2585         if (task == current) return 0;
2586 
2587         if (!task_is_stopped_or_traced(task)) {
2588                 DPRINT(("cannot attach to non-stopped task [%d] state=%ld\n", task_pid_nr(task), task->state));
2589                 return -EBUSY;
2590         }
2591         /*
2592          * make sure the task is off any CPU
2593          */
2594         wait_task_inactive(task, 0);
2595 
2596         /* more to come... */
2597 
2598         return 0;
2599 }
2600 
2601 static int
2602 pfm_get_task(pfm_context_t *ctx, pid_t pid, struct task_struct **task)
2603 {
2604         struct task_struct *p = current;
2605         int ret;
2606 
2607         /* XXX: need to add more checks here */
2608         if (pid < 2) return -EPERM;
2609 
2610         if (pid != task_pid_vnr(current)) {
2611 
2612                 read_lock(&tasklist_lock);
2613 
2614                 p = find_task_by_vpid(pid);
2615 
2616                 /* make sure task cannot go away while we operate on it */
2617                 if (p) get_task_struct(p);
2618 
2619                 read_unlock(&tasklist_lock);
2620 
2621                 if (p == NULL) return -ESRCH;
2622         }
2623 
2624         ret = pfm_task_incompatible(ctx, p);
2625         if (ret == 0) {
2626                 *task = p;
2627         } else if (p != current) {
2628                 pfm_put_task(p);
2629         }
2630         return ret;
2631 }
2632 
2633 
2634 
2635 static int
2636 pfm_context_create(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
2637 {
2638         pfarg_context_t *req = (pfarg_context_t *)arg;
2639         struct file *filp;
2640         struct path path;
2641         int ctx_flags;
2642         int fd;
2643         int ret;
2644 
2645         /* let's check the arguments first */
2646         ret = pfarg_is_sane(current, req);
2647         if (ret < 0)
2648                 return ret;
2649 
2650         ctx_flags = req->ctx_flags;
2651 
2652         ret = -ENOMEM;
2653 
2654         fd = get_unused_fd_flags(0);
2655         if (fd < 0)
2656                 return fd;
2657 
2658         ctx = pfm_context_alloc(ctx_flags);
2659         if (!ctx)
2660                 goto error;
2661 
2662         filp = pfm_alloc_file(ctx);
2663         if (IS_ERR(filp)) {
2664                 ret = PTR_ERR(filp);
2665                 goto error_file;
2666         }
2667 
2668         req->ctx_fd = ctx->ctx_fd = fd;
2669 
2670         /*
2671          * does the user want to sample?
2672          */
2673         if (pfm_uuid_cmp(req->ctx_smpl_buf_id, pfm_null_uuid)) {
2674                 ret = pfm_setup_buffer_fmt(current, filp, ctx, ctx_flags, 0, req);
2675                 if (ret)
2676                         goto buffer_error;
2677         }
2678 
2679         DPRINT(("ctx=%p flags=0x%x system=%d notify_block=%d excl_idle=%d no_msg=%d ctx_fd=%d\n",
2680                 ctx,
2681                 ctx_flags,
2682                 ctx->ctx_fl_system,
2683                 ctx->ctx_fl_block,
2684                 ctx->ctx_fl_excl_idle,
2685                 ctx->ctx_fl_no_msg,
2686                 ctx->ctx_fd));
2687 
2688         /*
2689          * initialize soft PMU state
2690          */
2691         pfm_reset_pmu_state(ctx);
2692 
2693         fd_install(fd, filp);
2694 
2695         return 0;
2696 
2697 buffer_error:
2698         path = filp->f_path;
2699         put_filp(filp);
2700         path_put(&path);
2701 
2702         if (ctx->ctx_buf_fmt) {
2703                 pfm_buf_fmt_exit(ctx->ctx_buf_fmt, current, NULL, regs);
2704         }
2705 error_file:
2706         pfm_context_free(ctx);
2707 
2708 error:
2709         put_unused_fd(fd);
2710         return ret;
2711 }
2712 
2713 static inline unsigned long
2714 pfm_new_counter_value (pfm_counter_t *reg, int is_long_reset)
2715 {
2716         unsigned long val = is_long_reset ? reg->long_reset : reg->short_reset;
2717         unsigned long new_seed, old_seed = reg->seed, mask = reg->mask;
2718         extern unsigned long carta_random32 (unsigned long seed);
2719 
2720         if (reg->flags & PFM_REGFL_RANDOM) {
2721                 new_seed = carta_random32(old_seed);
2722                 val -= (old_seed & mask);       /* counter values are negative numbers! */
2723                 if ((mask >> 32) != 0)
2724                         /* construct a full 64-bit random value: */
2725                         new_seed |= carta_random32(old_seed >> 32) << 32;
2726                 reg->seed = new_seed;
2727         }
2728         reg->lval = val;
2729         return val;
2730 }
2731 
2732 static void
2733 pfm_reset_regs_masked(pfm_context_t *ctx, unsigned long *ovfl_regs, int is_long_reset)
2734 {
2735         unsigned long mask = ovfl_regs[0];
2736         unsigned long reset_others = 0UL;
2737         unsigned long val;
2738         int i;
2739 
2740         /*
2741          * now restore reset value on sampling overflowed counters
2742          */
2743         mask >>= PMU_FIRST_COUNTER;
2744         for(i = PMU_FIRST_COUNTER; mask; i++, mask >>= 1) {
2745 
2746                 if ((mask & 0x1UL) == 0UL) continue;
2747 
2748                 ctx->ctx_pmds[i].val = val = pfm_new_counter_value(ctx->ctx_pmds+ i, is_long_reset);
2749                 reset_others        |= ctx->ctx_pmds[i].reset_pmds[0];
2750 
2751                 DPRINT_ovfl((" %s reset ctx_pmds[%d]=%lx\n", is_long_reset ? "long" : "short", i, val));
2752         }
2753 
2754         /*
2755          * Now take care of resetting the other registers
2756          */
2757         for(i = 0; reset_others; i++, reset_others >>= 1) {
2758 
2759                 if ((reset_others & 0x1) == 0) continue;
2760 
2761                 ctx->ctx_pmds[i].val = val = pfm_new_counter_value(ctx->ctx_pmds + i, is_long_reset);
2762 
2763                 DPRINT_ovfl(("%s reset_others pmd[%d]=%lx\n",
2764                           is_long_reset ? "long" : "short", i, val));
2765         }
2766 }
2767 
2768 static void
2769 pfm_reset_regs(pfm_context_t *ctx, unsigned long *ovfl_regs, int is_long_reset)
2770 {
2771         unsigned long mask = ovfl_regs[0];
2772         unsigned long reset_others = 0UL;
2773         unsigned long val;
2774         int i;
2775 
2776         DPRINT_ovfl(("ovfl_regs=0x%lx is_long_reset=%d\n", ovfl_regs[0], is_long_reset));
2777 
2778         if (ctx->ctx_state == PFM_CTX_MASKED) {
2779                 pfm_reset_regs_masked(ctx, ovfl_regs, is_long_reset);
2780                 return;
2781         }
2782 
2783         /*
2784          * now restore reset value on sampling overflowed counters
2785          */
2786         mask >>= PMU_FIRST_COUNTER;
2787         for(i = PMU_FIRST_COUNTER; mask; i++, mask >>= 1) {
2788 
2789                 if ((mask & 0x1UL) == 0UL) continue;
2790 
2791                 val           = pfm_new_counter_value(ctx->ctx_pmds+ i, is_long_reset);
2792                 reset_others |= ctx->ctx_pmds[i].reset_pmds[0];
2793 
2794                 DPRINT_ovfl((" %s reset ctx_pmds[%d]=%lx\n", is_long_reset ? "long" : "short", i, val));
2795 
2796                 pfm_write_soft_counter(ctx, i, val);
2797         }
2798 
2799         /*
2800          * Now take care of resetting the other registers
2801          */
2802         for(i = 0; reset_others; i++, reset_others >>= 1) {
2803 
2804                 if ((reset_others & 0x1) == 0) continue;
2805 
2806                 val = pfm_new_counter_value(ctx->ctx_pmds + i, is_long_reset);
2807 
2808                 if (PMD_IS_COUNTING(i)) {
2809                         pfm_write_soft_counter(ctx, i, val);
2810                 } else {
2811                         ia64_set_pmd(i, val);
2812                 }
2813                 DPRINT_ovfl(("%s reset_others pmd[%d]=%lx\n",
2814                           is_long_reset ? "long" : "short", i, val));
2815         }
2816         ia64_srlz_d();
2817 }
2818 
2819 static int
2820 pfm_write_pmcs(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
2821 {
2822         struct task_struct *task;
2823         pfarg_reg_t *req = (pfarg_reg_t *)arg;
2824         unsigned long value, pmc_pm;
2825         unsigned long smpl_pmds, reset_pmds, impl_pmds;
2826         unsigned int cnum, reg_flags, flags, pmc_type;
2827         int i, can_access_pmu = 0, is_loaded, is_system, expert_mode;
2828         int is_monitor, is_counting, state;
2829         int ret = -EINVAL;
2830         pfm_reg_check_t wr_func;
2831 #define PFM_CHECK_PMC_PM(x, y, z) ((x)->ctx_fl_system ^ PMC_PM(y, z))
2832 
2833         state     = ctx->ctx_state;
2834         is_loaded = state == PFM_CTX_LOADED ? 1 : 0;
2835         is_system = ctx->ctx_fl_system;
2836         task      = ctx->ctx_task;
2837         impl_pmds = pmu_conf->impl_pmds[0];
2838 
2839         if (state == PFM_CTX_ZOMBIE) return -EINVAL;
2840 
2841         if (is_loaded) {
2842                 /*
2843                  * In system wide and when the context is loaded, access can only happen
2844                  * when the caller is running on the CPU being monitored by the session.
2845                  * It does not have to be the owner (ctx_task) of the context per se.
2846                  */
2847                 if (is_system && ctx->ctx_cpu != smp_processor_id()) {
2848                         DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
2849                         return -EBUSY;
2850                 }
2851                 can_access_pmu = GET_PMU_OWNER() == task || is_system ? 1 : 0;
2852         }
2853         expert_mode = pfm_sysctl.expert_mode; 
2854 
2855         for (i = 0; i < count; i++, req++) {
2856 
2857                 cnum       = req->reg_num;
2858                 reg_flags  = req->reg_flags;
2859                 value      = req->reg_value;
2860                 smpl_pmds  = req->reg_smpl_pmds[0];
2861                 reset_pmds = req->reg_reset_pmds[0];
2862                 flags      = 0;
2863 
2864 
2865                 if (cnum >= PMU_MAX_PMCS) {
2866                         DPRINT(("pmc%u is invalid\n", cnum));
2867                         goto error;
2868                 }
2869 
2870                 pmc_type   = pmu_conf->pmc_desc[cnum].type;
2871                 pmc_pm     = (value >> pmu_conf->pmc_desc[cnum].pm_pos) & 0x1;
2872                 is_counting = (pmc_type & PFM_REG_COUNTING) == PFM_REG_COUNTING ? 1 : 0;
2873                 is_monitor  = (pmc_type & PFM_REG_MONITOR) == PFM_REG_MONITOR ? 1 : 0;
2874 
2875                 /*
2876                  * we reject all non implemented PMC as well
2877                  * as attempts to modify PMC[0-3] which are used
2878                  * as status registers by the PMU
2879                  */
2880                 if ((pmc_type & PFM_REG_IMPL) == 0 || (pmc_type & PFM_REG_CONTROL) == PFM_REG_CONTROL) {
2881                         DPRINT(("pmc%u is unimplemented or no-access pmc_type=%x\n", cnum, pmc_type));
2882                         goto error;
2883                 }
2884                 wr_func = pmu_conf->pmc_desc[cnum].write_check;
2885                 /*
2886                  * If the PMC is a monitor, then if the value is not the default:
2887                  *      - system-wide session: PMCx.pm=1 (privileged monitor)
2888                  *      - per-task           : PMCx.pm=0 (user monitor)
2889                  */
2890                 if (is_monitor && value != PMC_DFL_VAL(cnum) && is_system ^ pmc_pm) {
2891                         DPRINT(("pmc%u pmc_pm=%lu is_system=%d\n",
2892                                 cnum,
2893                                 pmc_pm,
2894                                 is_system));
2895                         goto error;
2896                 }
2897 
2898                 if (is_counting) {
2899                         /*
2900                          * enforce generation of overflow interrupt. Necessary on all
2901                          * CPUs.
2902                          */
2903                         value |= 1 << PMU_PMC_OI;
2904 
2905                         if (reg_flags & PFM_REGFL_OVFL_NOTIFY) {
2906                                 flags |= PFM_REGFL_OVFL_NOTIFY;
2907                         }
2908 
2909                         if (reg_flags & PFM_REGFL_RANDOM) flags |= PFM_REGFL_RANDOM;
2910 
2911                         /* verify validity of smpl_pmds */
2912                         if ((smpl_pmds & impl_pmds) != smpl_pmds) {
2913                                 DPRINT(("invalid smpl_pmds 0x%lx for pmc%u\n", smpl_pmds, cnum));
2914                                 goto error;
2915                         }
2916 
2917                         /* verify validity of reset_pmds */
2918                         if ((reset_pmds & impl_pmds) != reset_pmds) {
2919                                 DPRINT(("invalid reset_pmds 0x%lx for pmc%u\n", reset_pmds, cnum));
2920                                 goto error;
2921                         }
2922                 } else {
2923                         if (reg_flags & (PFM_REGFL_OVFL_NOTIFY|PFM_REGFL_RANDOM)) {
2924                                 DPRINT(("cannot set ovfl_notify or random on pmc%u\n", cnum));
2925                                 goto error;
2926                         }
2927                         /* eventid on non-counting monitors are ignored */
2928                 }
2929 
2930                 /*
2931                  * execute write checker, if any
2932                  */
2933                 if (likely(expert_mode == 0 && wr_func)) {
2934                         ret = (*wr_func)(task, ctx, cnum, &value, regs);
2935                         if (ret) goto error;
2936                         ret = -EINVAL;
2937                 }
2938 
2939                 /*
2940                  * no error on this register
2941                  */
2942                 PFM_REG_RETFLAG_SET(req->reg_flags, 0);
2943 
2944                 /*
2945                  * Now we commit the changes to the software state
2946                  */
2947 
2948                 /*
2949                  * update overflow information
2950                  */
2951                 if (is_counting) {
2952                         /*
2953                          * full flag update each time a register is programmed
2954                          */
2955                         ctx->ctx_pmds[cnum].flags = flags;
2956 
2957                         ctx->ctx_pmds[cnum].reset_pmds[0] = reset_pmds;
2958                         ctx->ctx_pmds[cnum].smpl_pmds[0]  = smpl_pmds;
2959                         ctx->ctx_pmds[cnum].eventid       = req->reg_smpl_eventid;
2960 
2961                         /*
2962                          * Mark all PMDS to be accessed as used.
2963                          *
2964                          * We do not keep track of PMC because we have to
2965                          * systematically restore ALL of them.
2966                          *
2967                          * We do not update the used_monitors mask, because
2968                          * if we have not programmed them, then will be in
2969                          * a quiescent state, therefore we will not need to
2970                          * mask/restore then when context is MASKED.
2971                          */
2972                         CTX_USED_PMD(ctx, reset_pmds);
2973                         CTX_USED_PMD(ctx, smpl_pmds);
2974                         /*
2975                          * make sure we do not try to reset on
2976                          * restart because we have established new values
2977                          */
2978                         if (state == PFM_CTX_MASKED) ctx->ctx_ovfl_regs[0] &= ~1UL << cnum;
2979                 }
2980                 /*
2981                  * Needed in case the user does not initialize the equivalent
2982                  * PMD. Clearing is done indirectly via pfm_reset_pmu_state() so there is no
2983                  * possible leak here.
2984                  */
2985                 CTX_USED_PMD(ctx, pmu_conf->pmc_desc[cnum].dep_pmd[0]);
2986 
2987                 /*
2988                  * keep track of the monitor PMC that we are using.
2989                  * we save the value of the pmc in ctx_pmcs[] and if
2990                  * the monitoring is not stopped for the context we also
2991                  * place it in the saved state area so that it will be
2992                  * picked up later by the context switch code.
2993                  *
2994                  * The value in ctx_pmcs[] can only be changed in pfm_write_pmcs().
2995                  *
2996                  * The value in th_pmcs[] may be modified on overflow, i.e.,  when
2997                  * monitoring needs to be stopped.
2998                  */
2999                 if (is_monitor) CTX_USED_MONITOR(ctx, 1UL << cnum);
3000 
3001                 /*
3002                  * update context state
3003                  */
3004                 ctx->ctx_pmcs[cnum] = value;
3005 
3006                 if (is_loaded) {
3007                         /*
3008                          * write thread state
3009                          */
3010                         if (is_system == 0) ctx->th_pmcs[cnum] = value;
3011 
3012                         /*
3013                          * write hardware register if we can
3014                          */
3015                         if (can_access_pmu) {
3016                                 ia64_set_pmc(cnum, value);
3017                         }
3018 #ifdef CONFIG_SMP
3019                         else {
3020                                 /*
3021                                  * per-task SMP only here
3022                                  *
3023                                  * we are guaranteed that the task is not running on the other CPU,
3024                                  * we indicate that this PMD will need to be reloaded if the task
3025                                  * is rescheduled on the CPU it ran last on.
3026                                  */
3027                                 ctx->ctx_reload_pmcs[0] |= 1UL << cnum;
3028                         }
3029 #endif
3030                 }
3031 
3032                 DPRINT(("pmc[%u]=0x%lx ld=%d apmu=%d flags=0x%x all_pmcs=0x%lx used_pmds=0x%lx eventid=%ld smpl_pmds=0x%lx reset_pmds=0x%lx reloads_pmcs=0x%lx used_monitors=0x%lx ovfl_regs=0x%lx\n",
3033                           cnum,
3034                           value,
3035                           is_loaded,
3036                           can_access_pmu,
3037                           flags,
3038                           ctx->ctx_all_pmcs[0],
3039                           ctx->ctx_used_pmds[0],
3040                           ctx->ctx_pmds[cnum].eventid,
3041                           smpl_pmds,
3042                           reset_pmds,
3043                           ctx->ctx_reload_pmcs[0],
3044                           ctx->ctx_used_monitors[0],
3045                           ctx->ctx_ovfl_regs[0]));
3046         }
3047 
3048         /*
3049          * make sure the changes are visible
3050          */
3051         if (can_access_pmu) ia64_srlz_d();
3052 
3053         return 0;
3054 error:
3055         PFM_REG_RETFLAG_SET(req->reg_flags, PFM_REG_RETFL_EINVAL);
3056         return ret;
3057 }
3058 
3059 static int
3060 pfm_write_pmds(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3061 {
3062         struct task_struct *task;
3063         pfarg_reg_t *req = (pfarg_reg_t *)arg;
3064         unsigned long value, hw_value, ovfl_mask;
3065         unsigned int cnum;
3066         int i, can_access_pmu = 0, state;
3067         int is_counting, is_loaded, is_system, expert_mode;
3068         int ret = -EINVAL;
3069         pfm_reg_check_t wr_func;
3070 
3071 
3072         state     = ctx->ctx_state;
3073         is_loaded = state == PFM_CTX_LOADED ? 1 : 0;
3074         is_system = ctx->ctx_fl_system;
3075         ovfl_mask = pmu_conf->ovfl_val;
3076         task      = ctx->ctx_task;
3077 
3078         if (unlikely(state == PFM_CTX_ZOMBIE)) return -EINVAL;
3079 
3080         /*
3081          * on both UP and SMP, we can only write to the PMC when the task is
3082          * the owner of the local PMU.
3083          */
3084         if (likely(is_loaded)) {
3085                 /*
3086                  * In system wide and when the context is loaded, access can only happen
3087                  * when the caller is running on the CPU being monitored by the session.
3088                  * It does not have to be the owner (ctx_task) of the context per se.
3089                  */
3090                 if (unlikely(is_system && ctx->ctx_cpu != smp_processor_id())) {
3091                         DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
3092                         return -EBUSY;
3093                 }
3094                 can_access_pmu = GET_PMU_OWNER() == task || is_system ? 1 : 0;
3095         }
3096         expert_mode = pfm_sysctl.expert_mode; 
3097 
3098         for (i = 0; i < count; i++, req++) {
3099 
3100                 cnum  = req->reg_num;
3101                 value = req->reg_value;
3102 
3103                 if (!PMD_IS_IMPL(cnum)) {
3104                         DPRINT(("pmd[%u] is unimplemented or invalid\n", cnum));
3105                         goto abort_mission;
3106                 }
3107                 is_counting = PMD_IS_COUNTING(cnum);
3108                 wr_func     = pmu_conf->pmd_desc[cnum].write_check;
3109 
3110                 /*
3111                  * execute write checker, if any
3112                  */
3113                 if (unlikely(expert_mode == 0 && wr_func)) {
3114                         unsigned long v = value;
3115 
3116                         ret = (*wr_func)(task, ctx, cnum, &v, regs);
3117                         if (ret) goto abort_mission;
3118 
3119                         value = v;
3120                         ret   = -EINVAL;
3121                 }
3122 
3123                 /*
3124                  * no error on this register
3125                  */
3126                 PFM_REG_RETFLAG_SET(req->reg_flags, 0);
3127 
3128                 /*
3129                  * now commit changes to software state
3130                  */
3131                 hw_value = value;
3132 
3133                 /*
3134                  * update virtualized (64bits) counter
3135                  */
3136                 if (is_counting) {
3137                         /*
3138                          * write context state
3139                          */
3140                         ctx->ctx_pmds[cnum].lval = value;
3141 
3142                         /*
3143                          * when context is load we use the split value
3144                          */
3145                         if (is_loaded) {
3146                                 hw_value = value &  ovfl_mask;
3147                                 value    = value & ~ovfl_mask;
3148                         }
3149                 }
3150                 /*
3151                  * update reset values (not just for counters)
3152                  */
3153                 ctx->ctx_pmds[cnum].long_reset  = req->reg_long_reset;
3154                 ctx->ctx_pmds[cnum].short_reset = req->reg_short_reset;
3155 
3156                 /*
3157                  * update randomization parameters (not just for counters)
3158                  */
3159                 ctx->ctx_pmds[cnum].seed = req->reg_random_seed;
3160                 ctx->ctx_pmds[cnum].mask = req->reg_random_mask;
3161 
3162                 /*
3163                  * update context value
3164                  */
3165                 ctx->ctx_pmds[cnum].val  = value;
3166 
3167                 /*
3168                  * Keep track of what we use
3169                  *
3170                  * We do not keep track of PMC because we have to
3171                  * systematically restore ALL of them.
3172                  */
3173                 CTX_USED_PMD(ctx, PMD_PMD_DEP(cnum));
3174 
3175                 /*
3176                  * mark this PMD register used as well
3177                  */
3178                 CTX_USED_PMD(ctx, RDEP(cnum));
3179 
3180                 /*
3181                  * make sure we do not try to reset on
3182                  * restart because we have established new values
3183                  */
3184                 if (is_counting && state == PFM_CTX_MASKED) {
3185                         ctx->ctx_ovfl_regs[0] &= ~1UL << cnum;
3186                 }
3187 
3188                 if (is_loaded) {
3189                         /*
3190                          * write thread state
3191                          */
3192                         if (is_system == 0) ctx->th_pmds[cnum] = hw_value;
3193 
3194                         /*
3195                          * write hardware register if we can
3196                          */
3197                         if (can_access_pmu) {
3198                                 ia64_set_pmd(cnum, hw_value);
3199                         } else {
3200 #ifdef CONFIG_SMP
3201                                 /*
3202                                  * we are guaranteed that the task is not running on the other CPU,
3203                                  * we indicate that this PMD will need to be reloaded if the task
3204                                  * is rescheduled on the CPU it ran last on.
3205                                  */
3206                                 ctx->ctx_reload_pmds[0] |= 1UL << cnum;
3207 #endif
3208                         }
3209                 }
3210 
3211                 DPRINT(("pmd[%u]=0x%lx ld=%d apmu=%d, hw_value=0x%lx ctx_pmd=0x%lx  short_reset=0x%lx "
3212                           "long_reset=0x%lx notify=%c seed=0x%lx mask=0x%lx used_pmds=0x%lx reset_pmds=0x%lx reload_pmds=0x%lx all_pmds=0x%lx ovfl_regs=0x%lx\n",
3213                         cnum,
3214                         value,
3215                         is_loaded,
3216                         can_access_pmu,
3217                         hw_value,
3218                         ctx->ctx_pmds[cnum].val,
3219                         ctx->ctx_pmds[cnum].short_reset,
3220                         ctx->ctx_pmds[cnum].long_reset,
3221                         PMC_OVFL_NOTIFY(ctx, cnum) ? 'Y':'N',
3222                         ctx->ctx_pmds[cnum].seed,
3223                         ctx->ctx_pmds[cnum].mask,
3224                         ctx->ctx_used_pmds[0],
3225                         ctx->ctx_pmds[cnum].reset_pmds[0],
3226                         ctx->ctx_reload_pmds[0],
3227                         ctx->ctx_all_pmds[0],
3228                         ctx->ctx_ovfl_regs[0]));
3229         }
3230 
3231         /*
3232          * make changes visible
3233          */
3234         if (can_access_pmu) ia64_srlz_d();
3235 
3236         return 0;
3237 
3238 abort_mission:
3239         /*
3240          * for now, we have only one possibility for error
3241          */
3242         PFM_REG_RETFLAG_SET(req->reg_flags, PFM_REG_RETFL_EINVAL);
3243         return ret;
3244 }
3245 
3246 /*
3247  * By the way of PROTECT_CONTEXT(), interrupts are masked while we are in this function.
3248  * Therefore we know, we do not have to worry about the PMU overflow interrupt. If an
3249  * interrupt is delivered during the call, it will be kept pending until we leave, making
3250  * it appears as if it had been generated at the UNPROTECT_CONTEXT(). At least we are
3251  * guaranteed to return consistent data to the user, it may simply be old. It is not
3252  * trivial to treat the overflow while inside the call because you may end up in
3253  * some module sampling buffer code causing deadlocks.
3254  */
3255 static int
3256 pfm_read_pmds(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3257 {
3258         struct task_struct *task;
3259         unsigned long val = 0UL, lval, ovfl_mask, sval;
3260         pfarg_reg_t *req = (pfarg_reg_t *)arg;
3261         unsigned int cnum, reg_flags = 0;
3262         int i, can_access_pmu = 0, state;
3263         int is_loaded, is_system, is_counting, expert_mode;
3264         int ret = -EINVAL;
3265         pfm_reg_check_t rd_func;
3266 
3267         /*
3268          * access is possible when loaded only for
3269          * self-monitoring tasks or in UP mode
3270          */
3271 
3272         state     = ctx->ctx_state;
3273         is_loaded = state == PFM_CTX_LOADED ? 1 : 0;
3274         is_system = ctx->ctx_fl_system;
3275         ovfl_mask = pmu_conf->ovfl_val;
3276         task      = ctx->ctx_task;
3277 
3278         if (state == PFM_CTX_ZOMBIE) return -EINVAL;
3279 
3280         if (likely(is_loaded)) {
3281                 /*
3282                  * In system wide and when the context is loaded, access can only happen
3283                  * when the caller is running on the CPU being monitored by the session.
3284                  * It does not have to be the owner (ctx_task) of the context per se.
3285                  */
3286                 if (unlikely(is_system && ctx->ctx_cpu != smp_processor_id())) {
3287                         DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
3288                         return -EBUSY;
3289                 }
3290                 /*
3291                  * this can be true when not self-monitoring only in UP
3292                  */
3293                 can_access_pmu = GET_PMU_OWNER() == task || is_system ? 1 : 0;
3294 
3295                 if (can_access_pmu) ia64_srlz_d();
3296         }
3297         expert_mode = pfm_sysctl.expert_mode; 
3298 
3299         DPRINT(("ld=%d apmu=%d ctx_state=%d\n",
3300                 is_loaded,
3301                 can_access_pmu,
3302                 state));
3303 
3304         /*
3305          * on both UP and SMP, we can only read the PMD from the hardware register when
3306          * the task is the owner of the local PMU.
3307          */
3308 
3309         for (i = 0; i < count; i++, req++) {
3310 
3311                 cnum        = req->reg_num;
3312                 reg_flags   = req->reg_flags;
3313 
3314                 if (unlikely(!PMD_IS_IMPL(cnum))) goto error;
3315                 /*
3316                  * we can only read the register that we use. That includes
3317                  * the one we explicitly initialize AND the one we want included
3318                  * in the sampling buffer (smpl_regs).
3319                  *
3320                  * Having this restriction allows optimization in the ctxsw routine
3321                  * without compromising security (leaks)
3322                  */
3323                 if (unlikely(!CTX_IS_USED_PMD(ctx, cnum))) goto error;
3324 
3325                 sval        = ctx->ctx_pmds[cnum].val;
3326                 lval        = ctx->ctx_pmds[cnum].lval;
3327                 is_counting = PMD_IS_COUNTING(cnum);
3328 
3329                 /*
3330                  * If the task is not the current one, then we check if the
3331                  * PMU state is still in the local live register due to lazy ctxsw.
3332                  * If true, then we read directly from the registers.
3333                  */
3334                 if (can_access_pmu){
3335                         val = ia64_get_pmd(cnum);
3336                 } else {
3337                         /*
3338                          * context has been saved
3339                          * if context is zombie, then task does not exist anymore.
3340                          * In this case, we use the full value saved in the context (pfm_flush_regs()).
3341                          */
3342                         val = is_loaded ? ctx->th_pmds[cnum] : 0UL;
3343                 }
3344                 rd_func = pmu_conf->pmd_desc[cnum].read_check;
3345 
3346                 if (is_counting) {
3347                         /*
3348                          * XXX: need to check for overflow when loaded
3349                          */
3350                         val &= ovfl_mask;
3351                         val += sval;
3352                 }
3353 
3354                 /*
3355                  * execute read checker, if any
3356                  */
3357                 if (unlikely(expert_mode == 0 && rd_func)) {
3358                         unsigned long v = val;
3359                         ret = (*rd_func)(ctx->ctx_task, ctx, cnum, &v, regs);
3360                         if (ret) goto error;
3361                         val = v;
3362                         ret = -EINVAL;
3363                 }
3364 
3365                 PFM_REG_RETFLAG_SET(reg_flags, 0);
3366 
3367                 DPRINT(("pmd[%u]=0x%lx\n", cnum, val));
3368 
3369                 /*
3370                  * update register return value, abort all if problem during copy.
3371                  * we only modify the reg_flags field. no check mode is fine because
3372                  * access has been verified upfront in sys_perfmonctl().
3373                  */
3374                 req->reg_value            = val;
3375                 req->reg_flags            = reg_flags;
3376                 req->reg_last_reset_val   = lval;
3377         }
3378 
3379         return 0;
3380 
3381 error:
3382         PFM_REG_RETFLAG_SET(req->reg_flags, PFM_REG_RETFL_EINVAL);
3383         return ret;
3384 }
3385 
3386 int
3387 pfm_mod_write_pmcs(struct task_struct *task, void *req, unsigned int nreq, struct pt_regs *regs)
3388 {
3389         pfm_context_t *ctx;
3390 
3391         if (req == NULL) return -EINVAL;
3392 
3393         ctx = GET_PMU_CTX();
3394 
3395         if (ctx == NULL) return -EINVAL;
3396 
3397         /*
3398          * for now limit to current task, which is enough when calling
3399          * from overflow handler
3400          */
3401         if (task != current && ctx->ctx_fl_system == 0) return -EBUSY;
3402 
3403         return pfm_write_pmcs(ctx, req, nreq, regs);
3404 }
3405 EXPORT_SYMBOL(pfm_mod_write_pmcs);
3406 
3407 int
3408 pfm_mod_read_pmds(struct task_struct *task, void *req, unsigned int nreq, struct pt_regs *regs)
3409 {
3410         pfm_context_t *ctx;
3411 
3412         if (req == NULL) return -EINVAL;
3413 
3414         ctx = GET_PMU_CTX();
3415 
3416         if (ctx == NULL) return -EINVAL;
3417 
3418         /*
3419          * for now limit to current task, which is enough when calling
3420          * from overflow handler
3421          */
3422         if (task != current && ctx->ctx_fl_system == 0) return -EBUSY;
3423 
3424         return pfm_read_pmds(ctx, req, nreq, regs);
3425 }
3426 EXPORT_SYMBOL(pfm_mod_read_pmds);
3427 
3428 /*
3429  * Only call this function when a process it trying to
3430  * write the debug registers (reading is always allowed)
3431  */
3432 int
3433 pfm_use_debug_registers(struct task_struct *task)
3434 {
3435         pfm_context_t *ctx = task->thread.pfm_context;
3436         unsigned long flags;
3437         int ret = 0;
3438 
3439         if (pmu_conf->use_rr_dbregs == 0) return 0;
3440 
3441         DPRINT(("called for [%d]\n", task_pid_nr(task)));
3442 
3443         /*
3444          * do it only once
3445          */
3446         if (task->thread.flags & IA64_THREAD_DBG_VALID) return 0;
3447 
3448         /*
3449          * Even on SMP, we do not need to use an atomic here because
3450          * the only way in is via ptrace() and this is possible only when the
3451          * process is stopped. Even in the case where the ctxsw out is not totally
3452          * completed by the time we come here, there is no way the 'stopped' process
3453          * could be in the middle of fiddling with the pfm_write_ibr_dbr() routine.
3454          * So this is always safe.
3455          */
3456         if (ctx && ctx->ctx_fl_using_dbreg == 1) return -1;
3457 
3458         LOCK_PFS(flags);
3459 
3460         /*
3461          * We cannot allow setting breakpoints when system wide monitoring
3462          * sessions are using the debug registers.
3463          */
3464         if (pfm_sessions.pfs_sys_use_dbregs> 0)
3465                 ret = -1;
3466         else
3467                 pfm_sessions.pfs_ptrace_use_dbregs++;
3468 
3469         DPRINT(("ptrace_use_dbregs=%u  sys_use_dbregs=%u by [%d] ret = %d\n",
3470                   pfm_sessions.pfs_ptrace_use_dbregs,
3471                   pfm_sessions.pfs_sys_use_dbregs,
3472                   task_pid_nr(task), ret));
3473 
3474         UNLOCK_PFS(flags);
3475 
3476         return ret;
3477 }
3478 
3479 /*
3480  * This function is called for every task that exits with the
3481  * IA64_THREAD_DBG_VALID set. This indicates a task which was
3482  * able to use the debug registers for debugging purposes via
3483  * ptrace(). Therefore we know it was not using them for
3484  * performance monitoring, so we only decrement the number
3485  * of "ptraced" debug register users to keep the count up to date
3486  */
3487 int
3488 pfm_release_debug_registers(struct task_struct *task)
3489 {
3490         unsigned long flags;
3491         int ret;
3492 
3493         if (pmu_conf->use_rr_dbregs == 0) return 0;
3494 
3495         LOCK_PFS(flags);
3496         if (pfm_sessions.pfs_ptrace_use_dbregs == 0) {
3497                 printk(KERN_ERR "perfmon: invalid release for [%d] ptrace_use_dbregs=0\n", task_pid_nr(task));
3498                 ret = -1;
3499         }  else {
3500                 pfm_sessions.pfs_ptrace_use_dbregs--;
3501                 ret = 0;
3502         }
3503         UNLOCK_PFS(flags);
3504 
3505         return ret;
3506 }
3507 
3508 static int
3509 pfm_restart(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3510 {
3511         struct task_struct *task;
3512         pfm_buffer_fmt_t *fmt;
3513         pfm_ovfl_ctrl_t rst_ctrl;
3514         int state, is_system;
3515         int ret = 0;
3516 
3517         state     = ctx->ctx_state;
3518         fmt       = ctx->ctx_buf_fmt;
3519         is_system = ctx->ctx_fl_system;
3520         task      = PFM_CTX_TASK(ctx);
3521 
3522         switch(state) {
3523                 case PFM_CTX_MASKED:
3524                         break;
3525                 case PFM_CTX_LOADED: 
3526                         if (CTX_HAS_SMPL(ctx) && fmt->fmt_restart_active) break;
3527                         /* fall through */
3528                 case PFM_CTX_UNLOADED:
3529                 case PFM_CTX_ZOMBIE:
3530                         DPRINT(("invalid state=%d\n", state));
3531                         return -EBUSY;
3532                 default:
3533                         DPRINT(("state=%d, cannot operate (no active_restart handler)\n", state));
3534                         return -EINVAL;
3535         }
3536 
3537         /*
3538          * In system wide and when the context is loaded, access can only happen
3539          * when the caller is running on the CPU being monitored by the session.
3540          * It does not have to be the owner (ctx_task) of the context per se.
3541          */
3542         if (is_system && ctx->ctx_cpu != smp_processor_id()) {
3543                 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
3544                 return -EBUSY;
3545         }
3546 
3547         /* sanity check */
3548         if (unlikely(task == NULL)) {
3549                 printk(KERN_ERR "perfmon: [%d] pfm_restart no task\n", task_pid_nr(current));
3550                 return -EINVAL;
3551         }
3552 
3553         if (task == current || is_system) {
3554 
3555                 fmt = ctx->ctx_buf_fmt;
3556 
3557                 DPRINT(("restarting self %d ovfl=0x%lx\n",
3558                         task_pid_nr(task),
3559                         ctx->ctx_ovfl_regs[0]));
3560 
3561                 if (CTX_HAS_SMPL(ctx)) {
3562 
3563                         prefetch(ctx->ctx_smpl_hdr);
3564 
3565                         rst_ctrl.bits.mask_monitoring = 0;
3566                         rst_ctrl.bits.reset_ovfl_pmds = 0;
3567 
3568                         if (state == PFM_CTX_LOADED)
3569                                 ret = pfm_buf_fmt_restart_active(fmt, task, &rst_ctrl, ctx->ctx_smpl_hdr, regs);
3570                         else
3571                                 ret = pfm_buf_fmt_restart(fmt, task, &rst_ctrl, ctx->ctx_smpl_hdr, regs);
3572                 } else {
3573                         rst_ctrl.bits.mask_monitoring = 0;
3574                         rst_ctrl.bits.reset_ovfl_pmds = 1;
3575                 }
3576 
3577                 if (ret == 0) {
3578                         if (rst_ctrl.bits.reset_ovfl_pmds)
3579                                 pfm_reset_regs(ctx, ctx->ctx_ovfl_regs, PFM_PMD_LONG_RESET);
3580 
3581                         if (rst_ctrl.bits.mask_monitoring == 0) {
3582                                 DPRINT(("resuming monitoring for [%d]\n", task_pid_nr(task)));
3583 
3584                                 if (state == PFM_CTX_MASKED) pfm_restore_monitoring(task);
3585                         } else {
3586                                 DPRINT(("keeping monitoring stopped for [%d]\n", task_pid_nr(task)));
3587 
3588                                 // cannot use pfm_stop_monitoring(task, regs);
3589                         }
3590                 }
3591                 /*
3592                  * clear overflowed PMD mask to remove any stale information
3593                  */
3594                 ctx->ctx_ovfl_regs[0] = 0UL;
3595 
3596                 /*
3597                  * back to LOADED state
3598                  */
3599                 ctx->ctx_state = PFM_CTX_LOADED;
3600 
3601                 /*
3602                  * XXX: not really useful for self monitoring
3603                  */
3604                 ctx->ctx_fl_can_restart = 0;
3605 
3606                 return 0;
3607         }
3608 
3609         /* 
3610          * restart another task
3611          */
3612 
3613         /*
3614          * When PFM_CTX_MASKED, we cannot issue a restart before the previous 
3615          * one is seen by the task.
3616          */
3617         if (state == PFM_CTX_MASKED) {
3618                 if (ctx->ctx_fl_can_restart == 0) return -EINVAL;
3619                 /*
3620                  * will prevent subsequent restart before this one is
3621                  * seen by other task
3622                  */
3623                 ctx->ctx_fl_can_restart = 0;
3624         }
3625 
3626         /*
3627          * if blocking, then post the semaphore is PFM_CTX_MASKED, i.e.
3628          * the task is blocked or on its way to block. That's the normal
3629          * restart path. If the monitoring is not masked, then the task
3630          * can be actively monitoring and we cannot directly intervene.
3631          * Therefore we use the trap mechanism to catch the task and
3632          * force it to reset the buffer/reset PMDs.
3633          *
3634          * if non-blocking, then we ensure that the task will go into
3635          * pfm_handle_work() before returning to user mode.
3636          *
3637          * We cannot explicitly reset another task, it MUST always
3638          * be done by the task itself. This works for system wide because
3639          * the tool that is controlling the session is logically doing 
3640          * "self-monitoring".
3641          */
3642         if (CTX_OVFL_NOBLOCK(ctx) == 0 && state == PFM_CTX_MASKED) {
3643                 DPRINT(("unblocking [%d]\n", task_pid_nr(task)));
3644                 complete(&ctx->ctx_restart_done);
3645         } else {
3646                 DPRINT(("[%d] armed exit trap\n", task_pid_nr(task)));
3647 
3648                 ctx->ctx_fl_trap_reason = PFM_TRAP_REASON_RESET;
3649 
3650                 PFM_SET_WORK_PENDING(task, 1);
3651 
3652                 set_notify_resume(task);
3653 
3654                 /*
3655                  * XXX: send reschedule if task runs on another CPU
3656                  */
3657         }
3658         return 0;
3659 }
3660 
3661 static int
3662 pfm_debug(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3663 {
3664         unsigned int m = *(unsigned int *)arg;
3665 
3666         pfm_sysctl.debug = m == 0 ? 0 : 1;
3667 
3668         printk(KERN_INFO "perfmon debugging %s (timing reset)\n", pfm_sysctl.debug ? "on" : "off");
3669 
3670         if (m == 0) {
3671                 memset(pfm_stats, 0, sizeof(pfm_stats));
3672                 for(m=0; m < NR_CPUS; m++) pfm_stats[m].pfm_ovfl_intr_cycles_min = ~0UL;
3673         }
3674         return 0;
3675 }
3676 
3677 /*
3678  * arg can be NULL and count can be zero for this function
3679  */
3680 static int
3681 pfm_write_ibr_dbr(int mode, pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3682 {
3683         struct thread_struct *thread = NULL;
3684         struct task_struct *task;
3685         pfarg_dbreg_t *req = (pfarg_dbreg_t *)arg;
3686         unsigned long flags;
3687         dbreg_t dbreg;
3688         unsigned int rnum;
3689         int first_time;
3690         int ret = 0, state;
3691         int i, can_access_pmu = 0;
3692         int is_system, is_loaded;
3693 
3694         if (pmu_conf->use_rr_dbregs == 0) return -EINVAL;
3695 
3696         state     = ctx->ctx_state;
3697         is_loaded = state == PFM_CTX_LOADED ? 1 : 0;
3698         is_system = ctx->ctx_fl_system;
3699         task      = ctx->ctx_task;
3700 
3701         if (state == PFM_CTX_ZOMBIE) return -EINVAL;
3702 
3703         /*
3704          * on both UP and SMP, we can only write to the PMC when the task is
3705          * the owner of the local PMU.
3706          */
3707         if (is_loaded) {
3708                 thread = &task->thread;
3709                 /*
3710                  * In system wide and when the context is loaded, access can only happen
3711                  * when the caller is running on the CPU being monitored by the session.
3712                  * It does not have to be the owner (ctx_task) of the context per se.
3713                  */
3714                 if (unlikely(is_system && ctx->ctx_cpu != smp_processor_id())) {
3715                         DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
3716                         return -EBUSY;
3717                 }
3718                 can_access_pmu = GET_PMU_OWNER() == task || is_system ? 1 : 0;
3719         }
3720 
3721         /*
3722          * we do not need to check for ipsr.db because we do clear ibr.x, dbr.r, and dbr.w
3723          * ensuring that no real breakpoint can be installed via this call.
3724          *
3725          * IMPORTANT: regs can be NULL in this function
3726          */
3727 
3728         first_time = ctx->ctx_fl_using_dbreg == 0;
3729 
3730         /*
3731          * don't bother if we are loaded and task is being debugged
3732          */
3733         if (is_loaded && (thread->flags & IA64_THREAD_DBG_VALID) != 0) {
3734                 DPRINT(("debug registers already in use for [%d]\n", task_pid_nr(task)));
3735                 return -EBUSY;
3736         }
3737 
3738         /*
3739          * check for debug registers in system wide mode
3740          *
3741          * If though a check is done in pfm_context_load(),
3742          * we must repeat it here, in case the registers are
3743          * written after the context is loaded
3744          */
3745         if (is_loaded) {
3746                 LOCK_PFS(flags);
3747 
3748                 if (first_time && is_system) {
3749                         if (pfm_sessions.pfs_ptrace_use_dbregs)
3750                                 ret = -EBUSY;
3751                         else
3752                                 pfm_sessions.pfs_sys_use_dbregs++;
3753                 }
3754                 UNLOCK_PFS(flags);
3755         }
3756 
3757         if (ret != 0) return ret;
3758 
3759         /*
3760          * mark ourself as user of the debug registers for
3761          * perfmon purposes.
3762          */
3763         ctx->ctx_fl_using_dbreg = 1;
3764 
3765         /*
3766          * clear hardware registers to make sure we don't
3767          * pick up stale state.
3768          *
3769          * for a system wide session, we do not use
3770          * thread.dbr, thread.ibr because this process
3771          * never leaves the current CPU and the state
3772          * is shared by all processes running on it
3773          */
3774         if (first_time && can_access_pmu) {
3775                 DPRINT(("[%d] clearing ibrs, dbrs\n", task_pid_nr(task)));
3776                 for (i=0; i < pmu_conf->num_ibrs; i++) {
3777                         ia64_set_ibr(i, 0UL);
3778                         ia64_dv_serialize_instruction();
3779                 }
3780                 ia64_srlz_i();
3781                 for (i=0; i < pmu_conf->num_dbrs; i++) {
3782                         ia64_set_dbr(i, 0UL);
3783                         ia64_dv_serialize_data();
3784                 }
3785                 ia64_srlz_d();
3786         }
3787 
3788         /*
3789          * Now install the values into the registers
3790          */
3791         for (i = 0; i < count; i++, req++) {
3792 
3793                 rnum      = req->dbreg_num;
3794                 dbreg.val = req->dbreg_value;
3795 
3796                 ret = -EINVAL;
3797 
3798                 if ((mode == PFM_CODE_RR && rnum >= PFM_NUM_IBRS) || ((mode == PFM_DATA_RR) && rnum >= PFM_NUM_DBRS)) {
3799                         DPRINT(("invalid register %u val=0x%lx mode=%d i=%d count=%d\n",
3800                                   rnum, dbreg.val, mode, i, count));
3801 
3802                         goto abort_mission;
3803                 }
3804 
3805                 /*
3806                  * make sure we do not install enabled breakpoint
3807                  */
3808                 if (rnum & 0x1) {
3809                         if (mode == PFM_CODE_RR)
3810                                 dbreg.ibr.ibr_x = 0;
3811                         else
3812                                 dbreg.dbr.dbr_r = dbreg.dbr.dbr_w = 0;
3813                 }
3814 
3815                 PFM_REG_RETFLAG_SET(req->dbreg_flags, 0);
3816 
3817                 /*
3818                  * Debug registers, just like PMC, can only be modified
3819                  * by a kernel call. Moreover, perfmon() access to those
3820                  * registers are centralized in this routine. The hardware
3821                  * does not modify the value of these registers, therefore,
3822                  * if we save them as they are written, we can avoid having
3823                  * to save them on context switch out. This is made possible
3824                  * by the fact that when perfmon uses debug registers, ptrace()
3825                  * won't be able to modify them concurrently.
3826                  */
3827                 if (mode == PFM_CODE_RR) {
3828                         CTX_USED_IBR(ctx, rnum);
3829 
3830                         if (can_access_pmu) {
3831                                 ia64_set_ibr(rnum, dbreg.val);
3832                                 ia64_dv_serialize_instruction();
3833                         }
3834 
3835                         ctx->ctx_ibrs[rnum] = dbreg.val;
3836 
3837                         DPRINT(("write ibr%u=0x%lx used_ibrs=0x%x ld=%d apmu=%d\n",
3838                                 rnum, dbreg.val, ctx->ctx_used_ibrs[0], is_loaded, can_access_pmu));
3839                 } else {
3840                         CTX_USED_DBR(ctx, rnum);
3841 
3842                         if (can_access_pmu) {
3843                                 ia64_set_dbr(rnum, dbreg.val);
3844                                 ia64_dv_serialize_data();
3845                         }
3846                         ctx->ctx_dbrs[rnum] = dbreg.val;
3847 
3848                         DPRINT(("write dbr%u=0x%lx used_dbrs=0x%x ld=%d apmu=%d\n",
3849                                 rnum, dbreg.val, ctx->ctx_used_dbrs[0], is_loaded, can_access_pmu));
3850                 }
3851         }
3852 
3853         return 0;
3854 
3855 abort_mission:
3856         /*
3857          * in case it was our first attempt, we undo the global modifications
3858          */
3859         if (first_time) {
3860                 LOCK_PFS(flags);
3861                 if (ctx->ctx_fl_system) {
3862                         pfm_sessions.pfs_sys_use_dbregs--;
3863                 }
3864                 UNLOCK_PFS(flags);
3865                 ctx->ctx_fl_using_dbreg = 0;
3866         }
3867         /*
3868          * install error return flag
3869          */
3870         PFM_REG_RETFLAG_SET(req->dbreg_flags, PFM_REG_RETFL_EINVAL);
3871 
3872         return ret;
3873 }
3874 
3875 static int
3876 pfm_write_ibrs(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3877 {
3878         return pfm_write_ibr_dbr(PFM_CODE_RR, ctx, arg, count, regs);
3879 }
3880 
3881 static int
3882 pfm_write_dbrs(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3883 {
3884         return pfm_write_ibr_dbr(PFM_DATA_RR, ctx, arg, count, regs);
3885 }
3886 
3887 int
3888 pfm_mod_write_ibrs(struct task_struct *task, void *req, unsigned int nreq, struct pt_regs *regs)
3889 {
3890         pfm_context_t *ctx;
3891 
3892         if (req == NULL) return -EINVAL;
3893 
3894         ctx = GET_PMU_CTX();
3895 
3896         if (ctx == NULL) return -EINVAL;
3897 
3898         /*
3899          * for now limit to current task, which is enough when calling
3900          * from overflow handler
3901          */
3902         if (task != current && ctx->ctx_fl_system == 0) return -EBUSY;
3903 
3904         return pfm_write_ibrs(ctx, req, nreq, regs);
3905 }
3906 EXPORT_SYMBOL(pfm_mod_write_ibrs);
3907 
3908 int
3909 pfm_mod_write_dbrs(struct task_struct *task, void *req, unsigned int nreq, struct pt_regs *regs)
3910 {
3911         pfm_context_t *ctx;
3912 
3913         if (req == NULL) return -EINVAL;
3914 
3915         ctx = GET_PMU_CTX();
3916 
3917         if (ctx == NULL) return -EINVAL;
3918 
3919         /*
3920          * for now limit to current task, which is enough when calling
3921          * from overflow handler
3922          */
3923         if (task != current && ctx->ctx_fl_system == 0) return -EBUSY;
3924 
3925         return pfm_write_dbrs(ctx, req, nreq, regs);
3926 }
3927 EXPORT_SYMBOL(pfm_mod_write_dbrs);
3928 
3929 
3930 static int
3931 pfm_get_features(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3932 {
3933         pfarg_features_t *req = (pfarg_features_t *)arg;
3934 
3935         req->ft_version = PFM_VERSION;
3936         return 0;
3937 }
3938 
3939 static int
3940 pfm_stop(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3941 {
3942         struct pt_regs *tregs;
3943         struct task_struct *task = PFM_CTX_TASK(ctx);
3944         int state, is_system;
3945 
3946         state     = ctx->ctx_state;
3947         is_system = ctx->ctx_fl_system;
3948 
3949         /*
3950          * context must be attached to issue the stop command (includes LOADED,MASKED,ZOMBIE)
3951          */
3952         if (state == PFM_CTX_UNLOADED) return -EINVAL;
3953 
3954         /*
3955          * In system wide and when the context is loaded, access can only happen
3956          * when the caller is running on the CPU being monitored by the session.
3957          * It does not have to be the owner (ctx_task) of the context per se.
3958          */
3959         if (is_system && ctx->ctx_cpu != smp_processor_id()) {
3960                 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
3961                 return -EBUSY;
3962         }
3963         DPRINT(("task [%d] ctx_state=%d is_system=%d\n",
3964                 task_pid_nr(PFM_CTX_TASK(ctx)),
3965                 state,
3966                 is_system));
3967         /*
3968          * in system mode, we need to update the PMU directly
3969          * and the user level state of the caller, which may not
3970          * necessarily be the creator of the context.
3971          */
3972         if (is_system) {
3973                 /*
3974                  * Update local PMU first
3975                  *
3976                  * disable dcr pp
3977                  */
3978                 ia64_setreg(_IA64_REG_CR_DCR, ia64_getreg(_IA64_REG_CR_DCR) & ~IA64_DCR_PP);
3979                 ia64_srlz_i();
3980 
3981                 /*
3982                  * update local cpuinfo
3983                  */
3984                 PFM_CPUINFO_CLEAR(PFM_CPUINFO_DCR_PP);
3985 
3986                 /*
3987                  * stop monitoring, does srlz.i
3988                  */
3989                 pfm_clear_psr_pp();
3990 
3991                 /*
3992                  * stop monitoring in the caller
3993                  */
3994                 ia64_psr(regs)->pp = 0;
3995 
3996                 return 0;
3997         }
3998         /*
3999          * per-task mode
4000          */
4001 
4002         if (task == current) {
4003                 /* stop monitoring  at kernel level */
4004                 pfm_clear_psr_up();
4005 
4006                 /*
4007                  * stop monitoring at the user level
4008                  */
4009                 ia64_psr(regs)->up = 0;
4010         } else {
4011                 tregs = task_pt_regs(task);
4012 
4013                 /*
4014                  * stop monitoring at the user level
4015                  */
4016                 ia64_psr(tregs)->up = 0;
4017 
4018                 /*
4019                  * monitoring disabled in kernel at next reschedule
4020                  */
4021                 ctx->ctx_saved_psr_up = 0;
4022                 DPRINT(("task=[%d]\n", task_pid_nr(task)));
4023         }
4024         return 0;
4025 }
4026 
4027 
4028 static int
4029 pfm_start(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
4030 {
4031         struct pt_regs *tregs;
4032         int state, is_system;
4033 
4034         state     = ctx->ctx_state;
4035         is_system = ctx->ctx_fl_system;
4036 
4037         if (state != PFM_CTX_LOADED) return -EINVAL;
4038 
4039         /*
4040          * In system wide and when the context is loaded, access can only happen
4041          * when the caller is running on the CPU being monitored by the session.
4042          * It does not have to be the owner (ctx_task) of the context per se.
4043          */
4044         if (is_system && ctx->ctx_cpu != smp_processor_id()) {
4045                 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
4046                 return -EBUSY;
4047         }
4048 
4049         /*
4050          * in system mode, we need to update the PMU directly
4051          * and the user level state of the caller, which may not
4052          * necessarily be the creator of the context.
4053          */
4054         if (is_system) {
4055 
4056                 /*
4057                  * set user level psr.pp for the caller
4058                  */
4059                 ia64_psr(regs)->pp = 1;
4060 
4061                 /*
4062                  * now update the local PMU and cpuinfo
4063                  */
4064                 PFM_CPUINFO_SET(PFM_CPUINFO_DCR_PP);
4065 
4066                 /*
4067                  * start monitoring at kernel level
4068                  */
4069                 pfm_set_psr_pp();
4070 
4071                 /* enable dcr pp */
4072                 ia64_setreg(_IA64_REG_CR_DCR, ia64_getreg(_IA64_REG_CR_DCR) | IA64_DCR_PP);
4073                 ia64_srlz_i();
4074 
4075                 return 0;
4076         }
4077 
4078         /*
4079          * per-process mode
4080          */
4081 
4082         if (ctx->ctx_task == current) {
4083 
4084                 /* start monitoring at kernel level */
4085                 pfm_set_psr_up();
4086 
4087                 /*
4088                  * activate monitoring at user level
4089                  */
4090                 ia64_psr(regs)->up = 1;
4091 
4092         } else {
4093                 tregs = task_pt_regs(ctx->ctx_task);
4094 
4095                 /*
4096                  * start monitoring at the kernel level the next
4097                  * time the task is scheduled
4098                  */
4099                 ctx->ctx_saved_psr_up = IA64_PSR_UP;
4100 
4101                 /*
4102                  * activate monitoring at user level
4103                  */
4104                 ia64_psr(tregs)->up = 1;
4105         }
4106         return 0;
4107 }
4108 
4109 static int
4110 pfm_get_pmc_reset(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
4111 {
4112         pfarg_reg_t *req = (pfarg_reg_t *)arg;
4113         unsigned int cnum;
4114         int i;
4115         int ret = -EINVAL;
4116 
4117         for (i = 0; i < count; i++, req++) {
4118 
4119                 cnum = req->reg_num;
4120 
4121                 if (!PMC_IS_IMPL(cnum)) goto abort_mission;
4122 
4123                 req->reg_value = PMC_DFL_VAL(cnum);
4124 
4125                 PFM_REG_RETFLAG_SET(req->reg_flags, 0);
4126 
4127                 DPRINT(("pmc_reset_val pmc[%u]=0x%lx\n", cnum, req->reg_value));
4128         }
4129         return 0;
4130 
4131 abort_mission:
4132         PFM_REG_RETFLAG_SET(req->reg_flags, PFM_REG_RETFL_EINVAL);
4133         return ret;
4134 }
4135 
4136 static int
4137 pfm_check_task_exist(pfm_context_t *ctx)
4138 {
4139         struct task_struct *g, *t;
4140         int ret = -ESRCH;
4141 
4142         read_lock(&tasklist_lock);
4143 
4144         do_each_thread (g, t) {
4145                 if (t->thread.pfm_context == ctx) {
4146                         ret = 0;
4147                         goto out;
4148                 }
4149         } while_each_thread (g, t);
4150 out:
4151         read_unlock(&tasklist_lock);
4152 
4153         DPRINT(("pfm_check_task_exist: ret=%d ctx=%p\n", ret, ctx));
4154 
4155         return ret;
4156 }
4157 
4158 static int
4159 pfm_context_load(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
4160 {
4161         struct task_struct *task;
4162         struct thread_struct *thread;
4163         struct pfm_context_t *old;
4164         unsigned long flags;
4165 #ifndef CONFIG_SMP
4166         struct task_struct *owner_task = NULL;
4167 #endif
4168         pfarg_load_t *req = (pfarg_load_t *)arg;
4169         unsigned long *pmcs_source, *pmds_source;
4170         int the_cpu;
4171         int ret = 0;
4172         int state, is_system, set_dbregs = 0;
4173 
4174         state     = ctx->ctx_state;
4175         is_system = ctx->ctx_fl_system;
4176         /*
4177          * can only load from unloaded or terminated state
4178          */
4179         if (state != PFM_CTX_UNLOADED) {
4180                 DPRINT(("cannot load to [%d], invalid ctx_state=%d\n",
4181                         req->load_pid,
4182                         ctx->ctx_state));
4183                 return -EBUSY;
4184         }
4185 
4186         DPRINT(("load_pid [%d] using_dbreg=%d\n", req->load_pid, ctx->ctx_fl_using_dbreg));
4187 
4188         if (CTX_OVFL_NOBLOCK(ctx) == 0 && req->load_pid == current->pid) {
4189                 DPRINT(("cannot use blocking mode on self\n"));
4190                 return -EINVAL;
4191         }
4192 
4193         ret = pfm_get_task(ctx, req->load_pid, &task);
4194         if (ret) {
4195                 DPRINT(("load_pid [%d] get_task=%d\n", req->load_pid, ret));
4196                 return ret;
4197         }
4198 
4199         ret = -EINVAL;
4200 
4201         /*
4202          * system wide is self monitoring only
4203          */
4204         if (is_system && task != current) {
4205                 DPRINT(("system wide is self monitoring only load_pid=%d\n",
4206                         req->load_pid));
4207                 goto error;
4208         }
4209 
4210         thread = &task->thread;
4211 
4212         ret = 0;
4213         /*
4214          * cannot load a context which is using range restrictions,
4215          * into a task that is being debugged.
4216          */
4217         if (ctx->ctx_fl_using_dbreg) {
4218                 if (thread->flags & IA64_THREAD_DBG_VALID) {
4219                         ret = -EBUSY;
4220                         DPRINT(("load_pid [%d] task is debugged, cannot load range restrictions\n", req->load_pid));
4221                         goto error;
4222                 }
4223                 LOCK_PFS(flags);
4224 
4225                 if (is_system) {
4226                         if (pfm_sessions.pfs_ptrace_use_dbregs) {
4227                                 DPRINT(("cannot load [%d] dbregs in use\n",
4228                                                         task_pid_nr(task)));
4229                                 ret = -EBUSY;
4230                         } else {
4231                                 pfm_sessions.pfs_sys_use_dbregs++;
4232                                 DPRINT(("load [%d] increased sys_use_dbreg=%u\n", task_pid_nr(task), pfm_sessions.pfs_sys_use_dbregs));
4233                                 set_dbregs = 1;
4234                         }
4235                 }
4236 
4237                 UNLOCK_PFS(flags);
4238 
4239                 if (ret) goto error;
4240         }
4241 
4242         /*
4243          * SMP system-wide monitoring implies self-monitoring.
4244          *
4245          * The programming model expects the task to
4246          * be pinned on a CPU throughout the session.
4247          * Here we take note of the current CPU at the
4248          * time the context is loaded. No call from
4249          * another CPU will be allowed.
4250          *
4251          * The pinning via shed_setaffinity()
4252          * must be done by the calling task prior
4253          * to this call.
4254          *
4255          * systemwide: keep track of CPU this session is supposed to run on
4256          */
4257         the_cpu = ctx->ctx_cpu = smp_processor_id();
4258 
4259         ret = -EBUSY;
4260         /*
4261          * now reserve the session
4262          */
4263         ret = pfm_reserve_session(current, is_system, the_cpu);
4264         if (ret) goto error;
4265 
4266         /*
4267          * task is necessarily stopped at this point.
4268          *
4269          * If the previous context was zombie, then it got removed in
4270          * pfm_save_regs(). Therefore we should not see it here.
4271          * If we see a context, then this is an active context
4272          *
4273          * XXX: needs to be atomic
4274          */
4275         DPRINT(("before cmpxchg() old_ctx=%p new_ctx=%p\n",
4276                 thread->pfm_context, ctx));
4277 
4278         ret = -EBUSY;
4279         old = ia64_cmpxchg(acq, &thread->pfm_context, NULL, ctx, sizeof(pfm_context_t *));
4280         if (old != NULL) {
4281                 DPRINT(("load_pid [%d] already has a context\n", req->load_pid));
4282                 goto error_unres;
4283         }
4284 
4285         pfm_reset_msgq(ctx);
4286 
4287         ctx->ctx_state = PFM_CTX_LOADED;
4288 
4289         /*
4290          * link context to task
4291          */
4292         ctx->ctx_task = task;
4293 
4294         if (is_system) {
4295                 /*
4296                  * we load as stopped
4297                  */
4298                 PFM_CPUINFO_SET(PFM_CPUINFO_SYST_WIDE);
4299                 PFM_CPUINFO_CLEAR(PFM_CPUINFO_DCR_PP);
4300 
4301                 if (ctx->ctx_fl_excl_idle) PFM_CPUINFO_SET(PFM_CPUINFO_EXCL_IDLE);
4302         } else {
4303                 thread->flags |= IA64_THREAD_PM_VALID;
4304         }
4305 
4306         /*
4307          * propagate into thread-state
4308          */
4309         pfm_copy_pmds(task, ctx);
4310         pfm_copy_pmcs(task, ctx);
4311 
4312         pmcs_source = ctx->th_pmcs;
4313         pmds_source = ctx->th_pmds;
4314 
4315         /*
4316          * always the case for system-wide
4317          */
4318         if (task == current) {
4319 
4320                 if (is_system == 0) {
4321 
4322                         /* allow user level control */
4323                         ia64_psr(regs)->sp = 0;
4324                         DPRINT(("clearing psr.sp for [%d]\n", task_pid_nr(task)));
4325 
4326                         SET_LAST_CPU(ctx, smp_processor_id());
4327                         INC_ACTIVATION();
4328                         SET_ACTIVATION(ctx);
4329 #ifndef CONFIG_SMP
4330                         /*
4331                          * push the other task out, if any
4332                          */
4333                         owner_task = GET_PMU_OWNER();
4334                         if (owner_task) pfm_lazy_save_regs(owner_task);
4335 #endif
4336                 }
4337                 /*
4338                  * load all PMD from ctx to PMU (as opposed to thread state)
4339                  * restore all PMC from ctx to PMU
4340                  */
4341                 pfm_restore_pmds(pmds_source, ctx->ctx_all_pmds[0]);
4342                 pfm_restore_pmcs(pmcs_source, ctx->ctx_all_pmcs[0]);
4343 
4344                 ctx->ctx_reload_pmcs[0] = 0UL;
4345                 ctx->ctx_reload_pmds[0] = 0UL;
4346 
4347                 /*
4348                  * guaranteed safe by earlier check against DBG_VALID
4349                  */
4350                 if (ctx->ctx_fl_using_dbreg) {
4351                         pfm_restore_ibrs(ctx->ctx_ibrs, pmu_conf->num_ibrs);
4352                         pfm_restore_dbrs(ctx->ctx_dbrs, pmu_conf->num_dbrs);
4353                 }
4354                 /*
4355                  * set new ownership
4356                  */
4357                 SET_PMU_OWNER(task, ctx);
4358 
4359                 DPRINT(("context loaded on PMU for [%d]\n", task_pid_nr(task)));
4360         } else {
4361                 /*
4362                  * when not current, task MUST be stopped, so this is safe
4363                  */
4364                 regs = task_pt_regs(task);
4365 
4366                 /* force a full reload */
4367                 ctx->ctx_last_activation = PFM_INVALID_ACTIVATION;
4368                 SET_LAST_CPU(ctx, -1);
4369 
4370                 /* initial saved psr (stopped) */
4371                 ctx->ctx_saved_psr_up = 0UL;
4372                 ia64_psr(regs)->up = ia64_psr(regs)->pp = 0;
4373         }
4374 
4375         ret = 0;
4376 
4377 error_unres:
4378         if (ret) pfm_unreserve_session(ctx, ctx->ctx_fl_system, the_cpu);
4379 error:
4380         /*
4381          * we must undo the dbregs setting (for system-wide)
4382          */
4383         if (ret && set_dbregs) {
4384                 LOCK_PFS(flags);
4385                 pfm_sessions.pfs_sys_use_dbregs--;
4386                 UNLOCK_PFS(flags);
4387         }
4388         /*
4389          * release task, there is now a link with the context
4390          */
4391         if (is_system == 0 && task != current) {
4392                 pfm_put_task(task);
4393 
4394                 if (ret == 0) {
4395                         ret = pfm_check_task_exist(ctx);
4396                         if (ret) {
4397                                 ctx->ctx_state = PFM_CTX_UNLOADED;
4398                                 ctx->ctx_task  = NULL;
4399                         }
4400                 }
4401         }
4402         return ret;
4403 }
4404 
4405 /*
4406  * in this function, we do not need to increase the use count
4407  * for the task via get_task_struct(), because we hold the
4408  * context lock. If the task were to disappear while having
4409  * a context attached, it would go through pfm_exit_thread()
4410  * which also grabs the context lock  and would therefore be blocked
4411  * until we are here.
4412  */
4413 static void pfm_flush_pmds(struct task_struct *, pfm_context_t *ctx);
4414 
4415 static int
4416 pfm_context_unload(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
4417 {
4418         struct task_struct *task = PFM_CTX_TASK(ctx);
4419         struct pt_regs *tregs;
4420         int prev_state, is_system;
4421         int ret;
4422 
4423         DPRINT(("ctx_state=%d task [%d]\n", ctx->ctx_state, task ? task_pid_nr(task) : -1));
4424 
4425         prev_state = ctx->ctx_state;
4426         is_system  = ctx->ctx_fl_system;
4427 
4428         /*
4429          * unload only when necessary
4430          */
4431         if (prev_state == PFM_CTX_UNLOADED) {
4432                 DPRINT(("ctx_state=%d, nothing to do\n", prev_state));
4433                 return 0;
4434         }
4435 
4436         /*
4437          * clear psr and dcr bits
4438          */
4439         ret = pfm_stop(ctx, NULL, 0, regs);
4440         if (ret) return ret;
4441 
4442         ctx->ctx_state = PFM_CTX_UNLOADED;
4443 
4444         /*
4445          * in system mode, we need to update the PMU directly
4446          * and the user level state of the caller, which may not
4447          * necessarily be the creator of the context.
4448          */
4449         if (is_system) {
4450 
4451                 /*
4452                  * Update cpuinfo
4453                  *
4454                  * local PMU is taken care of in pfm_stop()
4455                  */
4456                 PFM_CPUINFO_CLEAR(PFM_CPUINFO_SYST_WIDE);
4457                 PFM_CPUINFO_CLEAR(PFM_CPUINFO_EXCL_IDLE);
4458 
4459                 /*
4460                  * save PMDs in context
4461                  * release ownership
4462                  */
4463                 pfm_flush_pmds(current, ctx);
4464 
4465                 /*
4466                  * at this point we are done with the PMU
4467                  * so we can unreserve the resource.
4468                  */
4469                 if (prev_state != PFM_CTX_ZOMBIE) 
4470                         pfm_unreserve_session(ctx, 1 , ctx->ctx_cpu);
4471 
4472                 /*
4473                  * disconnect context from task
4474                  */
4475                 task->thread.pfm_context = NULL;
4476                 /*
4477                  * disconnect task from context
4478                  */
4479                 ctx->ctx_task = NULL;
4480 
4481                 /*
4482                  * There is nothing more to cleanup here.
4483                  */
4484                 return 0;
4485         }
4486 
4487         /*
4488          * per-task mode
4489          */
4490         tregs = task == current ? regs : task_pt_regs(task);
4491 
4492         if (task == current) {
4493                 /*
4494                  * cancel user level control
4495                  */
4496                 ia64_psr(regs)->sp = 1;
4497 
4498                 DPRINT(("setting psr.sp for [%d]\n", task_pid_nr(task)));
4499         }
4500         /*
4501          * save PMDs to context
4502          * release ownership
4503          */
4504         pfm_flush_pmds(task, ctx);
4505 
4506         /*
4507          * at this point we are done with the PMU
4508          * so we can unreserve the resource.
4509          *
4510          * when state was ZOMBIE, we have already unreserved.
4511          */
4512         if (prev_state != PFM_CTX_ZOMBIE) 
4513                 pfm_unreserve_session(ctx, 0 , ctx->ctx_cpu);
4514 
4515         /*
4516          * reset activation counter and psr
4517          */
4518         ctx->ctx_last_activation = PFM_INVALID_ACTIVATION;
4519         SET_LAST_CPU(ctx, -1);
4520 
4521         /*
4522          * PMU state will not be restored
4523          */
4524         task->thread.flags &= ~IA64_THREAD_PM_VALID;
4525 
4526         /*
4527          * break links between context and task
4528          */
4529         task->thread.pfm_context  = NULL;
4530         ctx->ctx_task             = NULL;
4531 
4532         PFM_SET_WORK_PENDING(task, 0);
4533 
4534         ctx->ctx_fl_trap_reason  = PFM_TRAP_REASON_NONE;
4535         ctx->ctx_fl_can_restart  = 0;
4536         ctx->ctx_fl_going_zombie = 0;
4537 
4538         DPRINT(("disconnected [%d] from context\n", task_pid_nr(task)));
4539 
4540         return 0;
4541 }
4542 
4543 
4544 /*
4545  * called only from exit_thread()
4546  * we come here only if the task has a context attached (loaded or masked)
4547  */
4548 void
4549 pfm_exit_thread(struct task_struct *task)
4550 {
4551         pfm_context_t *ctx;
4552         unsigned long flags;
4553         struct pt_regs *regs = task_pt_regs(task);
4554         int ret, state;
4555         int free_ok = 0;
4556 
4557         ctx = PFM_GET_CTX(task);
4558 
4559         PROTECT_CTX(ctx, flags);
4560 
4561         DPRINT(("state=%d task [%d]\n", ctx->ctx_state, task_pid_nr(task)));
4562 
4563         state = ctx->ctx_state;
4564         switch(state) {
4565                 case PFM_CTX_UNLOADED:
4566                         /*
4567                          * only comes to this function if pfm_context is not NULL, i.e., cannot
4568                          * be in unloaded state
4569                          */
4570                         printk(KERN_ERR "perfmon: pfm_exit_thread [%d] ctx unloaded\n", task_pid_nr(task));
4571                         break;
4572                 case PFM_CTX_LOADED:
4573                 case PFM_CTX_MASKED:
4574                         ret = pfm_context_unload(ctx, NULL, 0, regs);
4575                         if (ret) {
4576                                 printk(KERN_ERR "perfmon: pfm_exit_thread [%d] state=%d unload failed %d\n", task_pid_nr(task), state, ret);
4577                         }
4578                         DPRINT(("ctx unloaded for current state was %d\n", state));
4579 
4580                         pfm_end_notify_user(ctx);
4581                         break;
4582                 case PFM_CTX_ZOMBIE:
4583                         ret = pfm_context_unload(ctx, NULL, 0, regs);
4584                         if (ret) {
4585                                 printk(KERN_ERR "perfmon: pfm_exit_thread [%d] state=%d unload failed %d\n", task_pid_nr(task), state, ret);
4586                         }
4587                         free_ok = 1;
4588                         break;
4589                 default:
4590                         printk(KERN_ERR "perfmon: pfm_exit_thread [%d] unexpected state=%d\n", task_pid_nr(task), state);
4591                         break;
4592         }
4593         UNPROTECT_CTX(ctx, flags);
4594 
4595         { u64 psr = pfm_get_psr();
4596           BUG_ON(psr & (IA64_PSR_UP|IA64_PSR_PP));
4597           BUG_ON(GET_PMU_OWNER());
4598           BUG_ON(ia64_psr(regs)->up);
4599           BUG_ON(ia64_psr(regs)->pp);
4600         }
4601 
4602         /*
4603          * All memory free operations (especially for vmalloc'ed memory)
4604          * MUST be done with interrupts ENABLED.
4605          */
4606         if (free_ok) pfm_context_free(ctx);
4607 }
4608 
4609 /*
4610  * functions MUST be listed in the increasing order of their index (see permfon.h)
4611  */
4612 #define PFM_CMD(name, flags, arg_count, arg_type, getsz) { name, #name, flags, arg_count, sizeof(arg_type), getsz }
4613 #define PFM_CMD_S(name, flags) { name, #name, flags, 0, 0, NULL }
4614 #define PFM_CMD_PCLRWS  (PFM_CMD_FD|PFM_CMD_ARG_RW|PFM_CMD_STOP)
4615 #define PFM_CMD_PCLRW   (PFM_CMD_FD|PFM_CMD_ARG_RW)
4616 #define PFM_CMD_NONE    { NULL, "no-cmd", 0, 0, 0, NULL}
4617 
4618 static pfm_cmd_desc_t pfm_cmd_tab[]={
4619 /* 0  */PFM_CMD_NONE,
4620 /* 1  */PFM_CMD(pfm_write_pmcs, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_reg_t, NULL),
4621 /* 2  */PFM_CMD(pfm_write_pmds, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_reg_t, NULL),
4622 /* 3  */PFM_CMD(pfm_read_pmds, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_reg_t, NULL),
4623 /* 4  */PFM_CMD_S(pfm_stop, PFM_CMD_PCLRWS),
4624 /* 5  */PFM_CMD_S(pfm_start, PFM_CMD_PCLRWS),
4625 /* 6  */PFM_CMD_NONE,
4626 /* 7  */PFM_CMD_NONE,
4627 /* 8  */PFM_CMD(pfm_context_create, PFM_CMD_ARG_RW, 1, pfarg_context_t, pfm_ctx_getsize),
4628 /* 9  */PFM_CMD_NONE,
4629 /* 10 */PFM_CMD_S(pfm_restart, PFM_CMD_PCLRW),
4630 /* 11 */PFM_CMD_NONE,
4631 /* 12 */PFM_CMD(pfm_get_features, PFM_CMD_ARG_RW, 1, pfarg_features_t, NULL),
4632 /* 13 */PFM_CMD(pfm_debug, 0, 1, unsigned int, NULL),
4633 /* 14 */PFM_CMD_NONE,
4634 /* 15 */PFM_CMD(pfm_get_pmc_reset, PFM_CMD_ARG_RW, PFM_CMD_ARG_MANY, pfarg_reg_t, NULL),
4635 /* 16 */PFM_CMD(pfm_context_load, PFM_CMD_PCLRWS, 1, pfarg_load_t, NULL),
4636 /* 17 */PFM_CMD_S(pfm_context_unload, PFM_CMD_PCLRWS),
4637 /* 18 */PFM_CMD_NONE,
4638 /* 19 */PFM_CMD_NONE,
4639 /* 20 */PFM_CMD_NONE,
4640 /* 21 */PFM_CMD_NONE,
4641 /* 22 */PFM_CMD_NONE,
4642 /* 23 */PFM_CMD_NONE,
4643 /* 24 */PFM_CMD_NONE,
4644 /* 25 */PFM_CMD_NONE,
4645 /* 26 */PFM_CMD_NONE,
4646 /* 27 */PFM_CMD_NONE,
4647 /* 28 */PFM_CMD_NONE,
4648 /* 29 */PFM_CMD_NONE,
4649 /* 30 */PFM_CMD_NONE,
4650 /* 31 */PFM_CMD_NONE,
4651 /* 32 */PFM_CMD(pfm_write_ibrs, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_dbreg_t, NULL),
4652 /* 33 */PFM_CMD(pfm_write_dbrs, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_dbreg_t, NULL)
4653 };
4654 #define PFM_CMD_COUNT   (sizeof(pfm_cmd_tab)/sizeof(pfm_cmd_desc_t))
4655 
4656 static int
4657 pfm_check_task_state(pfm_context_t *ctx, int cmd, unsigned long flags)
4658 {
4659         struct task_struct *task;
4660         int state, old_state;
4661 
4662 recheck:
4663         state = ctx->ctx_state;
4664         task  = ctx->ctx_task;
4665 
4666         if (task == NULL) {
4667                 DPRINT(("context %d no task, state=%d\n", ctx->ctx_fd, state));
4668                 return 0;
4669         }
4670 
4671         DPRINT(("context %d state=%d [%d] task_state=%ld must_stop=%d\n",
4672                 ctx->ctx_fd,
4673                 state,
4674                 task_pid_nr(task),
4675                 task->state, PFM_CMD_STOPPED(cmd)));
4676 
4677         /*
4678          * self-monitoring always ok.
4679          *
4680          * for system-wide the caller can either be the creator of the
4681          * context (to one to which the context is attached to) OR
4682          * a task running on the same CPU as the session.
4683          */
4684         if (task == current || ctx->ctx_fl_system) return 0;
4685 
4686         /*
4687          * we are monitoring another thread
4688          */
4689         switch(state) {
4690                 case PFM_CTX_UNLOADED:
4691                         /*
4692                          * if context is UNLOADED we are safe to go
4693                          */
4694                         return 0;
4695                 case PFM_CTX_ZOMBIE:
4696                         /*
4697                          * no command can operate on a zombie context
4698                          */
4699                         DPRINT(("cmd %d state zombie cannot operate on context\n", cmd));
4700                         return -EINVAL;
4701                 case PFM_CTX_MASKED:
4702                         /*
4703                          * PMU state has been saved to software even though
4704                          * the thread may still be running.
4705                          */
4706                         if (cmd != PFM_UNLOAD_CONTEXT) return 0;
4707         }
4708 
4709         /*
4710          * context is LOADED or MASKED. Some commands may need to have 
4711          * the task stopped.
4712          *
4713          * We could lift this restriction for UP but it would mean that
4714          * the user has no guarantee the task would not run between
4715          * two successive calls to perfmonctl(). That's probably OK.
4716          * If this user wants to ensure the task does not run, then
4717          * the task must be stopped.
4718          */
4719         if (PFM_CMD_STOPPED(cmd)) {
4720                 if (!task_is_stopped_or_traced(task)) {
4721                         DPRINT(("[%d] task not in stopped state\n", task_pid_nr(task)));
4722                         return -EBUSY;
4723                 }
4724                 /*
4725                  * task is now stopped, wait for ctxsw out
4726                  *
4727                  * This is an interesting point in the code.
4728                  * We need to unprotect the context because
4729                  * the pfm_save_regs() routines needs to grab
4730                  * the same lock. There are danger in doing
4731                  * this because it leaves a window open for
4732                  * another task to get access to the context
4733                  * and possibly change its state. The one thing
4734                  * that is not possible is for the context to disappear
4735                  * because we are protected by the VFS layer, i.e.,
4736                  * get_fd()/put_fd().
4737                  */
4738                 old_state = state;
4739 
4740                 UNPROTECT_CTX(ctx, flags);
4741 
4742                 wait_task_inactive(task, 0);
4743 
4744                 PROTECT_CTX(ctx, flags);
4745 
4746                 /*
4747                  * we must recheck to verify if state has changed
4748                  */
4749                 if (ctx->ctx_state != old_state) {
4750                         DPRINT(("old_state=%d new_state=%d\n", old_state, ctx->ctx_state));
4751                         goto recheck;
4752                 }
4753         }
4754         return 0;
4755 }
4756 
4757 /*
4758  * system-call entry point (must return long)
4759  */
4760 asmlinkage long
4761 sys_perfmonctl (int fd, int cmd, void __user *arg, int count)
4762 {
4763         struct fd f = {NULL, 0};
4764         pfm_context_t *ctx = NULL;
4765         unsigned long flags = 0UL;
4766         void *args_k = NULL;
4767         long ret; /* will expand int return types */
4768         size_t base_sz, sz, xtra_sz = 0;
4769         int narg, completed_args = 0, call_made = 0, cmd_flags;
4770         int (*func)(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs);
4771         int (*getsize)(void *arg, size_t *sz);
4772 #define PFM_MAX_ARGSIZE 4096
4773 
4774         /*
4775          * reject any call if perfmon was disabled at initialization
4776          */
4777         if (unlikely(pmu_conf == NULL)) return -ENOSYS;
4778 
4779         if (unlikely(cmd < 0 || cmd >= PFM_CMD_COUNT)) {
4780                 DPRINT(("invalid cmd=%d\n", cmd));
4781                 return -EINVAL;
4782         }
4783 
4784         func      = pfm_cmd_tab[cmd].cmd_func;
4785         narg      = pfm_cmd_tab[cmd].cmd_narg;
4786         base_sz   = pfm_cmd_tab[cmd].cmd_argsize;
4787         getsize   = pfm_cmd_tab[cmd].cmd_getsize;
4788         cmd_flags = pfm_cmd_tab[cmd].cmd_flags;
4789 
4790         if (unlikely(func == NULL)) {
4791                 DPRINT(("invalid cmd=%d\n", cmd));
4792                 return -EINVAL;
4793         }
4794 
4795         DPRINT(("cmd=%s idx=%d narg=0x%x argsz=%lu count=%d\n",
4796                 PFM_CMD_NAME(cmd),
4797                 cmd,
4798                 narg,
4799                 base_sz,
4800                 count));
4801 
4802         /*
4803          * check if number of arguments matches what the command expects
4804          */
4805         if (unlikely((narg == PFM_CMD_ARG_MANY && count <= 0) || (narg > 0 && narg != count)))
4806                 return -EINVAL;
4807 
4808 restart_args:
4809         sz = xtra_sz + base_sz*count;
4810         /*
4811          * limit abuse to min page size
4812          */
4813         if (unlikely(sz > PFM_MAX_ARGSIZE)) {
4814                 printk(KERN_ERR "perfmon: [%d] argument too big %lu\n", task_pid_nr(current), sz);
4815                 return -E2BIG;
4816         }
4817 
4818         /*
4819          * allocate default-sized argument buffer
4820          */
4821         if (likely(count && args_k == NULL)) {
4822                 args_k = kmalloc(PFM_MAX_ARGSIZE, GFP_KERNEL);
4823                 if (args_k == NULL) return -ENOMEM;
4824         }
4825 
4826         ret = -EFAULT;
4827 
4828         /*
4829          * copy arguments
4830          *
4831          * assume sz = 0 for command without parameters
4832          */
4833         if (sz && copy_from_user(args_k, arg, sz)) {
4834                 DPRINT(("cannot copy_from_user %lu bytes @%p\n", sz, arg));
4835                 goto error_args;
4836         }
4837 
4838         /*
4839          * check if command supports extra parameters
4840          */
4841         if (completed_args == 0 && getsize) {
4842                 /*
4843                  * get extra parameters size (based on main argument)
4844                  */
4845                 ret = (*getsize)(args_k, &xtra_sz);
4846                 if (ret) goto error_args;
4847 
4848                 completed_args = 1;
4849 
4850                 DPRINT(("restart_args sz=%lu xtra_sz=%lu\n", sz, xtra_sz));
4851 
4852                 /* retry if necessary */
4853                 if (likely(xtra_sz)) goto restart_args;
4854         }
4855 
4856         if (unlikely((cmd_flags & PFM_CMD_FD) == 0)) goto skip_fd;
4857 
4858         ret = -EBADF;
4859 
4860         f = fdget(fd);
4861         if (unlikely(f.file == NULL)) {
4862                 DPRINT(("invalid fd %d\n", fd));
4863                 goto error_args;
4864         }
4865         if (unlikely(PFM_IS_FILE(f.file) == 0)) {
4866                 DPRINT(("fd %d not related to perfmon\n", fd));
4867                 goto error_args;
4868         }
4869 
4870         ctx = f.file->private_data;
4871         if (unlikely(ctx == NULL)) {
4872                 DPRINT(("no context for fd %d\n", fd));
4873                 goto error_args;
4874         }
4875         prefetch(&ctx->ctx_state);
4876 
4877         PROTECT_CTX(ctx, flags);
4878 
4879         /*
4880          * check task is stopped
4881          */
4882         ret = pfm_check_task_state(ctx, cmd, flags);
4883         if (unlikely(ret)) goto abort_locked;
4884 
4885 skip_fd:
4886         ret = (*func)(ctx, args_k, count, task_pt_regs(current));
4887 
4888         call_made = 1;
4889 
4890 abort_locked:
4891         if (likely(ctx)) {
4892                 DPRINT(("context unlocked\n"));
4893                 UNPROTECT_CTX(ctx, flags);
4894         }
4895 
4896         /* copy argument back to user, if needed */
4897         if (call_made && PFM_CMD_RW_ARG(cmd) && copy_to_user(arg, args_k, base_sz*count)) ret = -EFAULT;
4898 
4899 error_args:
4900         if (f.file)
4901                 fdput(f);
4902 
4903         kfree(args_k);
4904 
4905         DPRINT(("cmd=%s ret=%ld\n", PFM_CMD_NAME(cmd), ret));
4906 
4907         return ret;
4908 }
4909 
4910 static void
4911 pfm_resume_after_ovfl(pfm_context_t *ctx, unsigned long ovfl_regs, struct pt_regs *regs)
4912 {
4913         pfm_buffer_fmt_t *fmt = ctx->ctx_buf_fmt;
4914         pfm_ovfl_ctrl_t rst_ctrl;
4915         int state;
4916         int ret = 0;
4917 
4918         state = ctx->ctx_state;
4919         /*
4920          * Unlock sampling buffer and reset index atomically
4921          * XXX: not really needed when blocking
4922          */
4923         if (CTX_HAS_SMPL(ctx)) {
4924 
4925                 rst_ctrl.bits.mask_monitoring = 0;
4926                 rst_ctrl.bits.reset_ovfl_pmds = 0;
4927 
4928                 if (state == PFM_CTX_LOADED)
4929                         ret = pfm_buf_fmt_restart_active(fmt, current, &rst_ctrl, ctx->ctx_smpl_hdr, regs);
4930                 else
4931                         ret = pfm_buf_fmt_restart(fmt, current, &rst_ctrl, ctx->ctx_smpl_hdr, regs);
4932         } else {
4933                 rst_ctrl.bits.mask_monitoring = 0;
4934                 rst_ctrl.bits.reset_ovfl_pmds = 1;
4935         }
4936 
4937         if (ret == 0) {
4938                 if (rst_ctrl.bits.reset_ovfl_pmds) {
4939                         pfm_reset_regs(ctx, &ovfl_regs, PFM_PMD_LONG_RESET);
4940                 }
4941                 if (rst_ctrl.bits.mask_monitoring == 0) {
4942                         DPRINT(("resuming monitoring\n"));
4943                         if (ctx->ctx_state == PFM_CTX_MASKED) pfm_restore_monitoring(current);
4944                 } else {
4945                         DPRINT(("stopping monitoring\n"));
4946                         //pfm_stop_monitoring(current, regs);
4947                 }
4948                 ctx->ctx_state = PFM_CTX_LOADED;
4949         }
4950 }
4951 
4952 /*
4953  * context MUST BE LOCKED when calling
4954  * can only be called for current
4955  */
4956 static void
4957 pfm_context_force_terminate(pfm_context_t *ctx, struct pt_regs *regs)
4958 {
4959         int ret;
4960 
4961         DPRINT(("entering for [%d]\n", task_pid_nr(current)));
4962 
4963         ret = pfm_context_unload(ctx, NULL, 0, regs);
4964         if (ret) {
4965                 printk(KERN_ERR "pfm_context_force_terminate: [%d] unloaded failed with %d\n", task_pid_nr(current), ret);
4966         }
4967 
4968         /*
4969          * and wakeup controlling task, indicating we are now disconnected
4970          */
4971         wake_up_interruptible(&ctx->ctx_zombieq);
4972 
4973         /*
4974          * given that context is still locked, the controlling
4975          * task will only get access when we return from
4976          * pfm_handle_work().
4977          */
4978 }
4979 
4980 static int pfm_ovfl_notify_user(pfm_context_t *ctx, unsigned long ovfl_pmds);
4981 
4982  /*
4983   * pfm_handle_work() can be called with interrupts enabled
4984   * (TIF_NEED_RESCHED) or disabled. The down_interruptible
4985   * call may sleep, therefore we must re-enable interrupts
4986   * to avoid deadlocks. It is safe to do so because this function
4987   * is called ONLY when returning to user level (pUStk=1), in which case
4988   * there is no risk of kernel stack overflow due to deep
4989   * interrupt nesting.
4990   */
4991 void
4992 pfm_handle_work(void)
4993 {
4994         pfm_context_t *ctx;
4995         struct pt_regs *regs;
4996         unsigned long flags, dummy_flags;
4997         unsigned long ovfl_regs;
4998         unsigned int reason;
4999         int ret;
5000 
5001         ctx = PFM_GET_CTX(current);
5002         if (ctx == NULL) {
5003                 printk(KERN_ERR "perfmon: [%d] has no PFM context\n",
5004                         task_pid_nr(current));
5005                 return;
5006         }
5007 
5008         PROTECT_CTX(ctx, flags);
5009 
5010         PFM_SET_WORK_PENDING(current, 0);
5011 
5012         regs = task_pt_regs(current);
5013 
5014         /*
5015          * extract reason for being here and clear
5016          */
5017         reason = ctx->ctx_fl_trap_reason;
5018         ctx->ctx_fl_trap_reason = PFM_TRAP_REASON_NONE;
5019         ovfl_regs = ctx->ctx_ovfl_regs[0];
5020 
5021         DPRINT(("reason=%d state=%d\n", reason, ctx->ctx_state));
5022 
5023         /*
5024          * must be done before we check for simple-reset mode
5025          */
5026         if (ctx->ctx_fl_going_zombie || ctx->ctx_state == PFM_CTX_ZOMBIE)
5027                 goto do_zombie;
5028 
5029         //if (CTX_OVFL_NOBLOCK(ctx)) goto skip_blocking;
5030         if (reason == PFM_TRAP_REASON_RESET)
5031                 goto skip_blocking;
5032 
5033         /*
5034          * restore interrupt mask to what it was on entry.
5035          * Could be enabled/diasbled.
5036          */
5037         UNPROTECT_CTX(ctx, flags);
5038 
5039         /*
5040          * force interrupt enable because of down_interruptible()
5041          */
5042         local_irq_enable();
5043 
5044         DPRINT(("before block sleeping\n"));
5045 
5046         /*
5047          * may go through without blocking on SMP systems
5048          * if restart has been received already by the time we call down()
5049          */
5050         ret = wait_for_completion_interruptible(&ctx->ctx_restart_done);
5051 
5052         DPRINT(("after block sleeping ret=%d\n", ret));
5053 
5054         /*
5055          * lock context and mask interrupts again
5056          * We save flags into a dummy because we may have
5057          * altered interrupts mask compared to entry in this
5058          * function.
5059          */
5060         PROTECT_CTX(ctx, dummy_flags);
5061 
5062         /*
5063          * we need to read the ovfl_regs only after wake-up
5064          * because we may have had pfm_write_pmds() in between
5065          * and that can changed PMD values and therefore 
5066          * ovfl_regs is reset for these new PMD values.
5067          */
5068         ovfl_regs = ctx->ctx_ovfl_regs[0];
5069 
5070         if (ctx->ctx_fl_going_zombie) {
5071 do_zombie:
5072                 DPRINT(("context is zombie, bailing out\n"));
5073                 pfm_context_force_terminate(ctx, regs);
5074                 goto nothing_to_do;
5075         }
5076         /*
5077          * in case of interruption of down() we don't restart anything
5078          */
5079         if (ret < 0)
5080                 goto nothing_to_do;
5081 
5082 skip_blocking:
5083         pfm_resume_after_ovfl(ctx, ovfl_regs, regs);
5084         ctx->ctx_ovfl_regs[0] = 0UL;
5085 
5086 nothing_to_do:
5087         /*
5088          * restore flags as they were upon entry
5089          */
5090         UNPROTECT_CTX(ctx, flags);
5091 }
5092 
5093 static int
5094 pfm_notify_user(pfm_context_t *ctx, pfm_msg_t *msg)
5095 {
5096         if (ctx->ctx_state == PFM_CTX_ZOMBIE) {
5097                 DPRINT(("ignoring overflow notification, owner is zombie\n"));
5098                 return 0;
5099         }
5100 
5101         DPRINT(("waking up somebody\n"));
5102 
5103         if (msg) wake_up_interruptible(&ctx->ctx_msgq_wait);
5104 
5105         /*
5106          * safe, we are not in intr handler, nor in ctxsw when
5107          * we come here
5108          */
5109         kill_fasync (&ctx->ctx_async_queue, SIGIO, POLL_IN);
5110 
5111         return 0;
5112 }
5113 
5114 static int
5115 pfm_ovfl_notify_user(pfm_context_t *ctx, unsigned long ovfl_pmds)
5116 {
5117         pfm_msg_t *msg = NULL;
5118 
5119         if (ctx->ctx_fl_no_msg == 0) {
5120                 msg = pfm_get_new_msg(ctx);
5121                 if (msg == NULL) {
5122                         printk(KERN_ERR "perfmon: pfm_ovfl_notify_user no more notification msgs\n");
5123                         return -1;
5124                 }
5125 
5126                 msg->pfm_ovfl_msg.msg_type         = PFM_MSG_OVFL;
5127                 msg->pfm_ovfl_msg.msg_ctx_fd       = ctx->ctx_fd;
5128                 msg->pfm_ovfl_msg.msg_active_set   = 0;
5129                 msg->pfm_ovfl_msg.msg_ovfl_pmds[0] = ovfl_pmds;
5130                 msg->pfm_ovfl_msg.msg_ovfl_pmds[1] = 0UL;
5131                 msg->pfm_ovfl_msg.msg_ovfl_pmds[2] = 0UL;
5132                 msg->pfm_ovfl_msg.msg_ovfl_pmds[3] = 0UL;
5133                 msg->pfm_ovfl_msg.msg_tstamp       = 0UL;
5134         }
5135 
5136         DPRINT(("ovfl msg: msg=%p no_msg=%d fd=%d ovfl_pmds=0x%lx\n",
5137                 msg,
5138                 ctx->ctx_fl_no_msg,
5139                 ctx->ctx_fd,
5140                 ovfl_pmds));
5141 
5142         return pfm_notify_user(ctx, msg);
5143 }
5144 
5145 static int
5146 pfm_end_notify_user(pfm_context_t *ctx)
5147 {
5148         pfm_msg_t *msg;
5149 
5150         msg = pfm_get_new_msg(ctx);
5151         if (msg == NULL) {
5152                 printk(KERN_ERR "perfmon: pfm_end_notify_user no more notification msgs\n");
5153                 return -1;
5154         }
5155         /* no leak */
5156         memset(msg, 0, sizeof(*msg));
5157 
5158         msg->pfm_end_msg.msg_type    = PFM_MSG_END;
5159         msg->pfm_end_msg.msg_ctx_fd  = ctx->ctx_fd;
5160         msg->pfm_ovfl_msg.msg_tstamp = 0UL;
5161 
5162         DPRINT(("end msg: msg=%p no_msg=%d ctx_fd=%d\n",
5163                 msg,
5164                 ctx->ctx_fl_no_msg,
5165                 ctx->ctx_fd));
5166 
5167         return pfm_notify_user(ctx, msg);
5168 }
5169 
5170 /*
5171  * main overflow processing routine.
5172  * it can be called from the interrupt path or explicitly during the context switch code
5173  */
5174 static void pfm_overflow_handler(struct task_struct *task, pfm_context_t *ctx,
5175                                 unsigned long pmc0, struct pt_regs *regs)
5176 {
5177         pfm_ovfl_arg_t *ovfl_arg;
5178         unsigned long mask;
5179         unsigned long old_val, ovfl_val, new_val;
5180         unsigned long ovfl_notify = 0UL, ovfl_pmds = 0UL, smpl_pmds = 0UL, reset_pmds;
5181         unsigned long tstamp;
5182         pfm_ovfl_ctrl_t ovfl_ctrl;
5183         unsigned int i, has_smpl;
5184         int must_notify = 0;
5185 
5186         if (unlikely(ctx->ctx_state == PFM_CTX_ZOMBIE)) goto stop_monitoring;
5187 
5188         /*
5189          * sanity test. Should never happen
5190          */
5191         if (unlikely((pmc0 & 0x1) == 0)) goto sanity_check;
5192 
5193         tstamp   = ia64_get_itc();
5194         mask     = pmc0 >> PMU_FIRST_COUNTER;
5195         ovfl_val = pmu_conf->ovfl_val;
5196         has_smpl = CTX_HAS_SMPL(ctx);
5197 
5198         DPRINT_ovfl(("pmc0=0x%lx pid=%d iip=0x%lx, %s "
5199                      "used_pmds=0x%lx\n",
5200                         pmc0,
5201                         task ? task_pid_nr(task): -1,
5202                         (regs ? regs->cr_iip : 0),
5203                         CTX_OVFL_NOBLOCK(ctx) ? "nonblocking" : "blocking",
5204                         ctx->ctx_used_pmds[0]));
5205 
5206 
5207         /*
5208          * first we update the virtual counters
5209          * assume there was a prior ia64_srlz_d() issued
5210          */
5211         for (i = PMU_FIRST_COUNTER; mask ; i++, mask >>= 1) {
5212 
5213                 /* skip pmd which did not overflow */
5214                 if ((mask & 0x1) == 0) continue;
5215 
5216                 /*
5217                  * Note that the pmd is not necessarily 0 at this point as qualified events
5218                  * may have happened before the PMU was frozen. The residual count is not
5219                  * taken into consideration here but will be with any read of the pmd via
5220                  * pfm_read_pmds().
5221                  */
5222                 old_val              = new_val = ctx->ctx_pmds[i].val;
5223                 new_val             += 1 + ovfl_val;
5224                 ctx->ctx_pmds[i].val = new_val;
5225 
5226                 /*
5227                  * check for overflow condition
5228                  */
5229                 if (likely(old_val > new_val)) {
5230                         ovfl_pmds |= 1UL << i;
5231                         if (PMC_OVFL_NOTIFY(ctx, i)) ovfl_notify |= 1UL << i;
5232                 }
5233 
5234                 DPRINT_ovfl(("ctx_pmd[%d].val=0x%lx old_val=0x%lx pmd=0x%lx ovfl_pmds=0x%lx ovfl_notify=0x%lx\n",
5235                         i,
5236                         new_val,
5237                         old_val,
5238                         ia64_get_pmd(i) & ovfl_val,
5239                         ovfl_pmds,
5240                         ovfl_notify));
5241         }
5242 
5243         /*
5244          * there was no 64-bit overflow, nothing else to do
5245          */
5246         if (ovfl_pmds == 0UL) return;
5247 
5248         /* 
5249          * reset all control bits
5250          */
5251         ovfl_ctrl.val = 0;
5252         reset_pmds    = 0UL;
5253 
5254         /*
5255          * if a sampling format module exists, then we "cache" the overflow by 
5256          * calling the module's handler() routine.
5257          */
5258         if (has_smpl) {
5259                 unsigned long start_cycles, end_cycles;
5260                 unsigned long pmd_mask;
5261                 int j, k, ret = 0;
5262                 int this_cpu = smp_processor_id();
5263 
5264                 pmd_mask = ovfl_pmds >> PMU_FIRST_COUNTER;
5265                 ovfl_arg = &ctx->ctx_ovfl_arg;
5266 
5267                 prefetch(ctx->ctx_smpl_hdr);
5268 
5269                 for(i=PMU_FIRST_COUNTER; pmd_mask && ret == 0; i++, pmd_mask >>=1) {
5270 
5271                         mask = 1UL << i;
5272 
5273                         if ((pmd_mask & 0x1) == 0) continue;
5274 
5275                         ovfl_arg->ovfl_pmd      = (unsigned char )i;
5276                         ovfl_arg->ovfl_notify   = ovfl_notify & mask ? 1 : 0;
5277                         ovfl_arg->active_set    = 0;
5278                         ovfl_arg->ovfl_ctrl.val = 0; /* module must fill in all fields */
5279                         ovfl_arg->smpl_pmds[0]  = smpl_pmds = ctx->ctx_pmds[i].smpl_pmds[0];
5280 
5281                         ovfl_arg->pmd_value      = ctx->ctx_pmds[i].val;
5282                         ovfl_arg->pmd_last_reset = ctx->ctx_pmds[i].lval;
5283                         ovfl_arg->pmd_eventid    = ctx->ctx_pmds[i].eventid;
5284 
5285                         /*
5286                          * copy values of pmds of interest. Sampling format may copy them
5287                          * into sampling buffer.
5288                          */
5289                         if (smpl_pmds) {
5290                                 for(j=0, k=0; smpl_pmds; j++, smpl_pmds >>=1) {
5291                                         if ((smpl_pmds & 0x1) == 0) continue;
5292                                         ovfl_arg->smpl_pmds_values[k++] = PMD_IS_COUNTING(j) ?  pfm_read_soft_counter(ctx, j) : ia64_get_pmd(j);
5293                                         DPRINT_ovfl(("smpl_pmd[%d]=pmd%u=0x%lx\n", k-1, j, ovfl_arg->smpl_pmds_values[k-1]));
5294                                 }
5295                         }
5296 
5297                         pfm_stats[this_cpu].pfm_smpl_handler_calls++;
5298 
5299                         start_cycles = ia64_get_itc();
5300 
5301                         /*
5302                          * call custom buffer format record (handler) routine
5303                          */
5304                         ret = (*ctx->ctx_buf_fmt->fmt_handler)(task, ctx->ctx_smpl_hdr, ovfl_arg, regs, tstamp);
5305 
5306                         end_cycles = ia64_get_itc();
5307 
5308                         /*
5309                          * For those controls, we take the union because they have
5310                          * an all or nothing behavior.
5311                          */
5312                         ovfl_ctrl.bits.notify_user     |= ovfl_arg->ovfl_ctrl.bits.notify_user;
5313                         ovfl_ctrl.bits.block_task      |= ovfl_arg->ovfl_ctrl.bits.block_task;
5314                         ovfl_ctrl.bits.mask_monitoring |= ovfl_arg->ovfl_ctrl.bits.mask_monitoring;
5315                         /*
5316                          * build the bitmask of pmds to reset now
5317                          */
5318                         if (ovfl_arg->ovfl_ctrl.bits.reset_ovfl_pmds) reset_pmds |= mask;
5319 
5320                         pfm_stats[this_cpu].pfm_smpl_handler_cycles += end_cycles - start_cycles;
5321                 }
5322                 /*
5323                  * when the module cannot handle the rest of the overflows, we abort right here
5324                  */
5325                 if (ret && pmd_mask) {
5326                         DPRINT(("handler aborts leftover ovfl_pmds=0x%lx\n",
5327                                 pmd_mask<<PMU_FIRST_COUNTER));
5328                 }
5329                 /*
5330                  * remove the pmds we reset now from the set of pmds to reset in pfm_restart()
5331                  */
5332                 ovfl_pmds &= ~reset_pmds;
5333         } else {
5334                 /*
5335                  * when no sampling module is used, then the default
5336                  * is to notify on overflow if requested by user
5337                  */
5338                 ovfl_ctrl.bits.notify_user     = ovfl_notify ? 1 : 0;
5339                 ovfl_ctrl.bits.block_task      = ovfl_notify ? 1 : 0;
5340                 ovfl_ctrl.bits.mask_monitoring = ovfl_notify ? 1 : 0; /* XXX: change for saturation */
5341                 ovfl_ctrl.bits.reset_ovfl_pmds = ovfl_notify ? 0 : 1;
5342                 /*
5343                  * if needed, we reset all overflowed pmds
5344                  */
5345                 if (ovfl_notify == 0) reset_pmds = ovfl_pmds;
5346         }
5347 
5348         DPRINT_ovfl(("ovfl_pmds=0x%lx reset_pmds=0x%lx\n", ovfl_pmds, reset_pmds));
5349 
5350         /*
5351          * reset the requested PMD registers using the short reset values
5352          */
5353         if (reset_pmds) {
5354                 unsigned long bm = reset_pmds;
5355                 pfm_reset_regs(ctx, &bm, PFM_PMD_SHORT_RESET);
5356         }
5357 
5358         if (ovfl_notify && ovfl_ctrl.bits.notify_user) {
5359                 /*
5360                  * keep track of what to reset when unblocking
5361                  */
5362                 ctx->ctx_ovfl_regs[0] = ovfl_pmds;
5363 
5364                 /*
5365                  * check for blocking context 
5366                  */
5367                 if (CTX_OVFL_NOBLOCK(ctx) == 0 && ovfl_ctrl.bits.block_task) {
5368 
5369                         ctx->ctx_fl_trap_reason = PFM_TRAP_REASON_BLOCK;
5370 
5371                         /*
5372                          * set the perfmon specific checking pending work for the task
5373                          */
5374                         PFM_SET_WORK_PENDING(task, 1);
5375 
5376                         /*
5377                          * when coming from ctxsw, current still points to the
5378                          * previous task, therefore we must work with task and not current.
5379                          */
5380                         set_notify_resume(task);
5381                 }
5382                 /*
5383                  * defer until state is changed (shorten spin window). the context is locked
5384                  * anyway, so the signal receiver would come spin for nothing.
5385                  */
5386                 must_notify = 1;
5387         }
5388 
5389         DPRINT_ovfl(("owner [%d] pending=%ld reason=%u ovfl_pmds=0x%lx ovfl_notify=0x%lx masked=%d\n",
5390                         GET_PMU_OWNER() ? task_pid_nr(GET_PMU_OWNER()) : -1,
5391                         PFM_GET_WORK_PENDING(task),
5392                         ctx->ctx_fl_trap_reason,
5393                         ovfl_pmds,
5394                         ovfl_notify,
5395                         ovfl_ctrl.bits.mask_monitoring ? 1 : 0));
5396         /*
5397          * in case monitoring must be stopped, we toggle the psr bits
5398          */
5399         if (ovfl_ctrl.bits.mask_monitoring) {
5400                 pfm_mask_monitoring(task);
5401                 ctx->ctx_state = PFM_CTX_MASKED;
5402                 ctx->ctx_fl_can_restart = 1;
5403         }
5404 
5405         /*
5406          * send notification now
5407          */
5408         if (must_notify) pfm_ovfl_notify_user(ctx, ovfl_notify);
5409 
5410         return;
5411 
5412 sanity_check:
5413         printk(KERN_ERR "perfmon: CPU%d overflow handler [%d] pmc0=0x%lx\n",
5414                         smp_processor_id(),
5415                         task ? task_pid_nr(task) : -1,
5416                         pmc0);
5417         return;
5418 
5419 stop_monitoring:
5420         /*
5421          * in SMP, zombie context is never restored but reclaimed in pfm_load_regs().
5422          * Moreover, zombies are also reclaimed in pfm_save_regs(). Therefore we can
5423          * come here as zombie only if the task is the current task. In which case, we
5424          * can access the PMU  hardware directly.
5425          *
5426          * Note that zombies do have PM_VALID set. So here we do the minimal.
5427          *
5428          * In case the context was zombified it could not be reclaimed at the time
5429          * the monitoring program exited. At this point, the PMU reservation has been
5430          * returned, the sampiing buffer has been freed. We must convert this call
5431          * into a spurious interrupt. However, we must also avoid infinite overflows
5432          * by stopping monitoring for this task. We can only come here for a per-task
5433          * context. All we need to do is to stop monitoring using the psr bits which
5434          * are always task private. By re-enabling secure montioring, we ensure that
5435          * the monitored task will not be able to re-activate monitoring.
5436          * The task will eventually be context switched out, at which point the context
5437          * will be reclaimed (that includes releasing ownership of the PMU).
5438          *
5439          * So there might be a window of time where the number of per-task session is zero
5440          * yet one PMU might have a owner and get at most one overflow interrupt for a zombie
5441          * context. This is safe because if a per-task session comes in, it will push this one
5442          * out and by the virtue on pfm_save_regs(), this one will disappear. If a system wide
5443          * session is force on that CPU, given that we use task pinning, pfm_save_regs() will
5444          * also push our zombie context out.
5445          *
5446          * Overall pretty hairy stuff....
5447          */
5448         DPRINT(("ctx is zombie for [%d], converted to spurious\n", task ? task_pid_nr(task): -1));
5449         pfm_clear_psr_up();
5450         ia64_psr(regs)->up = 0;
5451         ia64_psr(regs)->sp = 1;
5452         return;
5453 }
5454 
5455 static int
5456 pfm_do_interrupt_handler(void *arg, struct pt_regs *regs)
5457 {
5458         struct task_struct *task;
5459         pfm_context_t *ctx;
5460         unsigned long flags;
5461         u64 pmc0;
5462         int this_cpu = smp_processor_id();
5463         int retval = 0;
5464 
5465         pfm_stats[this_cpu].pfm_ovfl_intr_count++;
5466 
5467         /*
5468          * srlz.d done before arriving here
5469          */
5470         pmc0 = ia64_get_pmc(0);
5471 
5472         task = GET_PMU_OWNER();
5473         ctx  = GET_PMU_CTX();
5474 
5475         /*
5476          * if we have some pending bits set
5477          * assumes : if any PMC0.bit[63-1] is set, then PMC0.fr = 1
5478          */
5479         if (PMC0_HAS_OVFL(pmc0) && task) {
5480                 /*
5481                  * we assume that pmc0.fr is always set here
5482                  */
5483 
5484                 /* sanity check */
5485                 if (!ctx) goto report_spurious1;
5486 
5487                 if (ctx->ctx_fl_system == 0 && (task->thread.flags & IA64_THREAD_PM_VALID) == 0) 
5488                         goto report_spurious2;
5489 
5490                 PROTECT_CTX_NOPRINT(ctx, flags);
5491 
5492                 pfm_overflow_handler(task, ctx, pmc0, regs);
5493 
5494                 UNPROTECT_CTX_NOPRINT(ctx, flags);
5495 
5496         } else {
5497                 pfm_stats[this_cpu].pfm_spurious_ovfl_intr_count++;
5498                 retval = -1;
5499         }
5500         /*
5501          * keep it unfrozen at all times
5502          */
5503         pfm_unfreeze_pmu();
5504 
5505         return retval;
5506 
5507 report_spurious1:
5508         printk(KERN_INFO "perfmon: spurious overflow interrupt on CPU%d: process %d has no PFM context\n",
5509                 this_cpu, task_pid_nr(task));
5510         pfm_unfreeze_pmu();
5511         return -1;
5512 report_spurious2:
5513         printk(KERN_INFO "perfmon: spurious overflow interrupt on CPU%d: process %d, invalid flag\n", 
5514                 this_cpu, 
5515                 task_pid_nr(task));
5516         pfm_unfreeze_pmu();
5517         return -1;
5518 }
5519 
5520 static irqreturn_t
5521 pfm_interrupt_handler(int irq, void *arg)
5522 {
5523         unsigned long start_cycles, total_cycles;
5524         unsigned long min, max;
5525         int this_cpu;
5526         int ret;
5527         struct pt_regs *regs = get_irq_regs();
5528 
5529         this_cpu = get_cpu();
5530         if (likely(!pfm_alt_intr_handler)) {
5531                 min = pfm_stats[this_cpu].pfm_ovfl_intr_cycles_min;
5532                 max = pfm_stats[this_cpu].pfm_ovfl_intr_cycles_max;
5533 
5534                 start_cycles = ia64_get_itc();
5535 
5536                 ret = pfm_do_interrupt_handler(arg, regs);
5537 
5538                 total_cycles = ia64_get_itc();
5539 
5540                 /*
5541                  * don't measure spurious interrupts
5542                  */
5543                 if (likely(ret == 0)) {
5544                         total_cycles -= start_cycles;
5545 
5546                         if (total_cycles < min) pfm_stats[this_cpu].pfm_ovfl_intr_cycles_min = total_cycles;
5547                         if (total_cycles > max) pfm_stats[this_cpu].pfm_ovfl_intr_cycles_max = total_cycles;
5548 
5549                         pfm_stats[this_cpu].pfm_ovfl_intr_cycles += total_cycles;
5550                 }
5551         }
5552         else {
5553                 (*pfm_alt_intr_handler->handler)(irq, arg, regs);
5554         }
5555 
5556         put_cpu();
5557         return IRQ_HANDLED;
5558 }
5559 
5560 /*
5561  * /proc/perfmon interface, for debug only
5562  */
5563 
5564 #define PFM_PROC_SHOW_HEADER    ((void *)(long)nr_cpu_ids+1)
5565 
5566 static void *
5567 pfm_proc_start(struct seq_file *m, loff_t *pos)
5568 {
5569         if (*pos == 0) {
5570                 return PFM_PROC_SHOW_HEADER;
5571         }
5572 
5573         while (*pos <= nr_cpu_ids) {
5574                 if (cpu_online(*pos - 1)) {
5575                         return (void *)*pos;
5576                 }
5577                 ++*pos;
5578         }
5579         return NULL;
5580 }
5581 
5582 static void *
5583 pfm_proc_next(struct seq_file *m, void *v, loff_t *pos)
5584 {
5585         ++*pos;
5586         return pfm_proc_start(m, pos);
5587 }
5588 
5589 static void
5590 pfm_proc_stop(struct seq_file *m, void *v)
5591 {
5592 }
5593 
5594 static void
5595 pfm_proc_show_header(struct seq_file *m)
5596 {
5597         struct list_head * pos;
5598         pfm_buffer_fmt_t * entry;
5599         unsigned long flags;
5600 
5601         seq_printf(m,
5602                 "perfmon version           : %u.%u\n"
5603                 "model                     : %s\n"
5604                 "fastctxsw                 : %s\n"
5605                 "expert mode               : %s\n"
5606                 "ovfl_mask                 : 0x%lx\n"
5607                 "PMU flags                 : 0x%x\n",
5608                 PFM_VERSION_MAJ, PFM_VERSION_MIN,
5609                 pmu_conf->pmu_name,
5610                 pfm_sysctl.fastctxsw > 0 ? "Yes": "No",
5611                 pfm_sysctl.expert_mode > 0 ? "Yes": "No",
5612                 pmu_conf->ovfl_val,
5613                 pmu_conf->flags);
5614 
5615         LOCK_PFS(flags);
5616 
5617         seq_printf(m,
5618                 "proc_sessions             : %u\n"
5619                 "sys_sessions              : %u\n"
5620                 "sys_use_dbregs            : %u\n"
5621                 "ptrace_use_dbregs         : %u\n",
5622                 pfm_sessions.pfs_task_sessions,
5623                 pfm_sessions.pfs_sys_sessions,
5624                 pfm_sessions.pfs_sys_use_dbregs,
5625                 pfm_sessions.pfs_ptrace_use_dbregs);
5626 
5627         UNLOCK_PFS(flags);
5628 
5629         spin_lock(&pfm_buffer_fmt_lock);
5630 
5631         list_for_each(pos, &pfm_buffer_fmt_list) {
5632                 entry = list_entry(pos, pfm_buffer_fmt_t, fmt_list);
5633                 seq_printf(m, "format                    : %16phD %s\n",
5634                            entry->fmt_uuid, entry->fmt_name);
5635         }
5636         spin_unlock(&pfm_buffer_fmt_lock);
5637 
5638 }
5639 
5640 static int
5641 pfm_proc_show(struct seq_file *m, void *v)
5642 {
5643         unsigned long psr;
5644         unsigned int i;
5645         int cpu;
5646 
5647         if (v == PFM_PROC_SHOW_HEADER) {
5648                 pfm_proc_show_header(m);
5649                 return 0;
5650         }
5651 
5652         /* show info for CPU (v - 1) */
5653 
5654         cpu = (long)v - 1;
5655         seq_printf(m,
5656                 "CPU%-2d overflow intrs      : %lu\n"
5657                 "CPU%-2d overflow cycles     : %lu\n"
5658                 "CPU%-2d overflow min        : %lu\n"
5659                 "CPU%-2d overflow max        : %lu\n"
5660                 "CPU%-2d smpl handler calls  : %lu\n"
5661                 "CPU%-2d smpl handler cycles : %lu\n"
5662                 "CPU%-2d spurious intrs      : %lu\n"
5663                 "CPU%-2d replay   intrs      : %lu\n"
5664                 "CPU%-2d syst_wide           : %d\n"
5665                 "CPU%-2d dcr_pp              : %d\n"
5666                 "CPU%-2d exclude idle        : %d\n"
5667                 "CPU%-2d owner               : %d\n"
5668                 "CPU%-2d context             : %p\n"
5669                 "CPU%-2d activations         : %lu\n",
5670                 cpu, pfm_stats[cpu].pfm_ovfl_intr_count,
5671                 cpu, pfm_stats[cpu].pfm_ovfl_intr_cycles,
5672                 cpu, pfm_stats[cpu].pfm_ovfl_intr_cycles_min,
5673                 cpu, pfm_stats[cpu].pfm_ovfl_intr_cycles_max,
5674                 cpu, pfm_stats[cpu].pfm_smpl_handler_calls,
5675                 cpu, pfm_stats[cpu].pfm_smpl_handler_cycles,
5676                 cpu, pfm_stats[cpu].pfm_spurious_ovfl_intr_count,
5677                 cpu, pfm_stats[cpu].pfm_replay_ovfl_intr_count,
5678                 cpu, pfm_get_cpu_data(pfm_syst_info, cpu) & PFM_CPUINFO_SYST_WIDE ? 1 : 0,
5679                 cpu, pfm_get_cpu_data(pfm_syst_info, cpu) & PFM_CPUINFO_DCR_PP ? 1 : 0,
5680                 cpu, pfm_get_cpu_data(pfm_syst_info, cpu) & PFM_CPUINFO_EXCL_IDLE ? 1 : 0,
5681                 cpu, pfm_get_cpu_data(pmu_owner, cpu) ? pfm_get_cpu_data(pmu_owner, cpu)->pid: -1,
5682                 cpu, pfm_get_cpu_data(pmu_ctx, cpu),
5683                 cpu, pfm_get_cpu_data(pmu_activation_number, cpu));
5684 
5685         if (num_online_cpus() == 1 && pfm_sysctl.debug > 0) {
5686 
5687                 psr = pfm_get_psr();
5688 
5689                 ia64_srlz_d();
5690 
5691                 seq_printf(m, 
5692                         "CPU%-2d psr                 : 0x%lx\n"
5693                         "CPU%-2d pmc0                : 0x%lx\n", 
5694                         cpu, psr,
5695                         cpu, ia64_get_pmc(0));
5696 
5697                 for (i=0; PMC_IS_LAST(i) == 0;  i++) {
5698                         if (PMC_IS_COUNTING(i) == 0) continue;
5699                         seq_printf(m, 
5700                                 "CPU%-2d pmc%u                : 0x%lx\n"
5701                                 "CPU%-2d pmd%u                : 0x%lx\n", 
5702                                 cpu, i, ia64_get_pmc(i),
5703                                 cpu, i, ia64_get_pmd(i));
5704                 }
5705         }
5706         return 0;
5707 }
5708 
5709 const struct seq_operations pfm_seq_ops = {
5710         .start =        pfm_proc_start,
5711         .next =         pfm_proc_next,
5712         .stop =         pfm_proc_stop,
5713         .show =         pfm_proc_show
5714 };
5715 
5716 static int
5717 pfm_proc_open(struct inode *inode, struct file *file)
5718 {
5719         return seq_open(file, &pfm_seq_ops);
5720 }
5721 
5722 
5723 /*
5724  * we come here as soon as local_cpu_data->pfm_syst_wide is set. this happens
5725  * during pfm_enable() hence before pfm_start(). We cannot assume monitoring
5726  * is active or inactive based on mode. We must rely on the value in
5727  * local_cpu_data->pfm_syst_info
5728  */
5729 void
5730 pfm_syst_wide_update_task(struct task_struct *task, unsigned long info, int is_ctxswin)
5731 {
5732         struct pt_regs *regs;
5733         unsigned long dcr;
5734         unsigned long dcr_pp;
5735 
5736         dcr_pp = info & PFM_CPUINFO_DCR_PP ? 1 : 0;
5737 
5738         /*
5739          * pid 0 is guaranteed to be the idle task. There is one such task with pid 0
5740          * on every CPU, so we can rely on the pid to identify the idle task.
5741          */
5742         if ((info & PFM_CPUINFO_EXCL_IDLE) == 0 || task->pid) {
5743                 regs = task_pt_regs(task);
5744                 ia64_psr(regs)->pp = is_ctxswin ? dcr_pp : 0;
5745                 return;
5746         }
5747         /*
5748          * if monitoring has started
5749          */
5750         if (dcr_pp) {
5751                 dcr = ia64_getreg(_IA64_REG_CR_DCR);
5752                 /*
5753                  * context switching in?
5754                  */
5755                 if (is_ctxswin) {
5756                         /* mask monitoring for the idle task */
5757                         ia64_setreg(_IA64_REG_CR_DCR, dcr & ~IA64_DCR_PP);
5758                         pfm_clear_psr_pp();
5759                         ia64_srlz_i();
5760                         return;
5761                 }
5762                 /*
5763                  * context switching out
5764                  * restore monitoring for next task
5765                  *
5766                  * Due to inlining this odd if-then-else construction generates
5767                  * better code.
5768                  */
5769                 ia64_setreg(_IA64_REG_CR_DCR, dcr |IA64_DCR_PP);
5770                 pfm_set_psr_pp();
5771                 ia64_srlz_i();
5772         }
5773 }
5774 
5775 #ifdef CONFIG_SMP
5776 
5777 static void
5778 pfm_force_cleanup(pfm_context_t *ctx, struct pt_regs *regs)
5779 {
5780         struct task_struct *task = ctx->ctx_task;
5781 
5782         ia64_psr(regs)->up = 0;
5783         ia64_psr(regs)->sp = 1;
5784 
5785         if (GET_PMU_OWNER() == task) {
5786                 DPRINT(("cleared ownership for [%d]\n",
5787                                         task_pid_nr(ctx->ctx_task)));
5788                 SET_PMU_OWNER(NULL, NULL);
5789         }
5790 
5791         /*
5792          * disconnect the task from the context and vice-versa
5793          */
5794         PFM_SET_WORK_PENDING(task, 0);
5795 
5796         task->thread.pfm_context  = NULL;
5797         task->thread.flags       &= ~IA64_THREAD_PM_VALID;
5798 
5799         DPRINT(("force cleanup for [%d]\n",  task_pid_nr(task)));
5800 }
5801 
5802 
5803 /*
5804  * in 2.6, interrupts are masked when we come here and the runqueue lock is held
5805  */
5806 void
5807 pfm_save_regs(struct task_struct *task)
5808 {
5809         pfm_context_t *ctx;
5810         unsigned long flags;
5811         u64 psr;
5812 
5813 
5814         ctx = PFM_GET_CTX(task);
5815         if (ctx == NULL) return;
5816 
5817         /*
5818          * we always come here with interrupts ALREADY disabled by
5819          * the scheduler. So we simply need to protect against concurrent
5820          * access, not CPU concurrency.
5821          */
5822         flags = pfm_protect_ctx_ctxsw(ctx);
5823 
5824         if (ctx->ctx_state == PFM_CTX_ZOMBIE) {
5825                 struct pt_regs *regs = task_pt_regs(task);
5826 
5827                 pfm_clear_psr_up();
5828 
5829                 pfm_force_cleanup(ctx, regs);
5830 
5831                 BUG_ON(ctx->ctx_smpl_hdr);
5832 
5833                 pfm_unprotect_ctx_ctxsw(ctx, flags);
5834 
5835                 pfm_context_free(ctx);
5836                 return;
5837         }
5838 
5839         /*
5840          * save current PSR: needed because we modify it
5841          */
5842         ia64_srlz_d();
5843         psr = pfm_get_psr();
5844 
5845         BUG_ON(psr & (IA64_PSR_I));
5846 
5847         /*
5848          * stop monitoring:
5849          * This is the last instruction which may generate an overflow
5850          *
5851          * We do not need to set psr.sp because, it is irrelevant in kernel.
5852          * It will be restored from ipsr when going back to user level
5853          */
5854         pfm_clear_psr_up();
5855 
5856         /*
5857          * keep a copy of psr.up (for reload)
5858          */
5859         ctx->ctx_saved_psr_up = psr & IA64_PSR_UP;
5860 
5861         /*
5862          * release ownership of this PMU.
5863          * PM interrupts are masked, so nothing
5864          * can happen.
5865          */
5866         SET_PMU_OWNER(NULL, NULL);
5867 
5868         /*
5869          * we systematically save the PMD as we have no
5870          * guarantee we will be schedule at that same
5871          * CPU again.
5872          */
5873         pfm_save_pmds(ctx->th_pmds, ctx->ctx_used_pmds[0]);
5874 
5875         /*
5876          * save pmc0 ia64_srlz_d() done in pfm_save_pmds()
5877          * we will need it on the restore path to check
5878          * for pending overflow.
5879          */
5880         ctx->th_pmcs[0] = ia64_get_pmc(0);
5881 
5882         /*
5883          * unfreeze PMU if had pending overflows
5884          */
5885         if (ctx->th_pmcs[0] & ~0x1UL) pfm_unfreeze_pmu();
5886 
5887         /*
5888          * finally, allow context access.
5889          * interrupts will still be masked after this call.
5890          */
5891         pfm_unprotect_ctx_ctxsw(ctx, flags);
5892 }
5893 
5894 #else /* !CONFIG_SMP */
5895 void
5896 pfm_save_regs(struct task_struct *task)
5897 {
5898         pfm_context_t *ctx;
5899         u64 psr;
5900 
5901         ctx = PFM_GET_CTX(task);
5902         if (ctx == NULL) return;
5903 
5904         /*
5905          * save current PSR: needed because we modify it
5906          */
5907         psr = pfm_get_psr();
5908 
5909         BUG_ON(psr & (IA64_PSR_I));
5910 
5911         /*
5912          * stop monitoring:
5913          * This is the last instruction which may generate an overflow
5914          *
5915          * We do not need to set psr.sp because, it is irrelevant in kernel.
5916          * It will be restored from ipsr when going back to user level
5917          */
5918         pfm_clear_psr_up();
5919 
5920         /*
5921          * keep a copy of psr.up (for reload)
5922          */
5923         ctx->ctx_saved_psr_up = psr & IA64_PSR_UP;
5924 }
5925 
5926 static void
5927 pfm_lazy_save_regs (struct task_struct *task)
5928 {
5929         pfm_context_t *ctx;
5930         unsigned long flags;
5931 
5932         { u64 psr  = pfm_get_psr();
5933           BUG_ON(psr & IA64_PSR_UP);
5934         }
5935 
5936         ctx = PFM_GET_CTX(task);
5937 
5938         /*
5939          * we need to mask PMU overflow here to
5940          * make sure that we maintain pmc0 until
5941          * we save it. overflow interrupts are
5942          * treated as spurious if there is no
5943          * owner.
5944          *
5945          * XXX: I don't think this is necessary
5946          */
5947         PROTECT_CTX(ctx,flags);
5948 
5949         /*
5950          * release ownership of this PMU.
5951          * must be done before we save the registers.
5952          *
5953          * after this call any PMU interrupt is treated
5954          * as spurious.
5955          */
5956         SET_PMU_OWNER(NULL, NULL);
5957 
5958         /*
5959          * save all the pmds we use
5960          */
5961         pfm_save_pmds(ctx->th_pmds, ctx->ctx_used_pmds[0]);
5962 
5963         /*
5964          * save pmc0 ia64_srlz_d() done in pfm_save_pmds()
5965          * it is needed to check for pended overflow
5966          * on the restore path
5967          */
5968         ctx->th_pmcs[0] = ia64_get_pmc(0);
5969 
5970         /*
5971          * unfreeze PMU if had pending overflows
5972          */
5973         if (ctx->th_pmcs[0] & ~0x1UL) pfm_unfreeze_pmu();
5974 
5975         /*
5976          * now get can unmask PMU interrupts, they will
5977          * be treated as purely spurious and we will not
5978          * lose any information
5979          */
5980         UNPROTECT_CTX(ctx,flags);
5981 }
5982 #endif /* CONFIG_SMP */
5983 
5984 #ifdef CONFIG_SMP
5985 /*
5986  * in 2.6, interrupts are masked when we come here and the runqueue lock is held
5987  */
5988 void
5989 pfm_load_regs (struct task_struct *task)
5990 {
5991         pfm_context_t *ctx;
5992         unsigned long pmc_mask = 0UL, pmd_mask = 0UL;
5993         unsigned long flags;
5994         u64 psr, psr_up;
5995         int need_irq_resend;
5996 
5997         ctx = PFM_GET_CTX(task);
5998         if (unlikely(ctx == NULL)) return;
5999 
6000         BUG_ON(GET_PMU_OWNER());
6001 
6002         /*
6003          * possible on unload
6004          */
6005         if (unlikely((task->thread.flags & IA64_THREAD_PM_VALID) == 0)) return;
6006 
6007         /*
6008          * we always come here with interrupts ALREADY disabled by
6009          * the scheduler. So we simply need to protect against concurrent
6010          * access, not CPU concurrency.
6011          */
6012         flags = pfm_protect_ctx_ctxsw(ctx);
6013         psr   = pfm_get_psr();
6014 
6015         need_irq_resend = pmu_conf->flags & PFM_PMU_IRQ_RESEND;
6016 
6017         BUG_ON(psr & (IA64_PSR_UP|IA64_PSR_PP));
6018         BUG_ON(psr & IA64_PSR_I);
6019 
6020         if (unlikely(ctx->ctx_state == PFM_CTX_ZOMBIE)) {
6021                 struct pt_regs *regs = task_pt_regs(task);
6022 
6023                 BUG_ON(ctx->ctx_smpl_hdr);
6024 
6025                 pfm_force_cleanup(ctx, regs);
6026 
6027                 pfm_unprotect_ctx_ctxsw(ctx, flags);
6028 
6029                 /*
6030                  * this one (kmalloc'ed) is fine with interrupts disabled
6031                  */
6032                 pfm_context_free(ctx);
6033 
6034                 return;
6035         }
6036 
6037         /*
6038          * we restore ALL the debug registers to avoid picking up
6039          * stale state.
6040          */
6041         if (ctx->ctx_fl_using_dbreg) {
6042                 pfm_restore_ibrs(ctx->ctx_ibrs, pmu_conf->num_ibrs);
6043                 pfm_restore_dbrs(ctx->ctx_dbrs, pmu_conf->num_dbrs);
6044         }
6045         /*
6046          * retrieve saved psr.up
6047          */
6048         psr_up = ctx->ctx_saved_psr_up;
6049 
6050         /*
6051          * if we were the last user of the PMU on that CPU,
6052          * then nothing to do except restore psr
6053          */
6054         if (GET_LAST_CPU(ctx) == smp_processor_id() && ctx->ctx_last_activation == GET_ACTIVATION()) {
6055 
6056                 /*
6057                  * retrieve partial reload masks (due to user modifications)
6058                  */
6059                 pmc_mask = ctx->ctx_reload_pmcs[0];
6060                 pmd_mask = ctx->ctx_reload_pmds[0];
6061 
6062         } else {
6063                 /*
6064                  * To avoid leaking information to the user level when psr.sp=0,
6065                  * we must reload ALL implemented pmds (even the ones we don't use).
6066                  * In the kernel we only allow PFM_READ_PMDS on registers which
6067                  * we initialized or requested (sampling) so there is no risk there.
6068                  */
6069                 pmd_mask = pfm_sysctl.fastctxsw ?  ctx->ctx_used_pmds[0] : ctx->ctx_all_pmds[0];
6070 
6071                 /*
6072                  * ALL accessible PMCs are systematically reloaded, unused registers
6073                  * get their default (from pfm_reset_pmu_state()) values to avoid picking
6074                  * up stale configuration.
6075                  *
6076                  * PMC0 is never in the mask. It is always restored separately.
6077                  */
6078                 pmc_mask = ctx->ctx_all_pmcs[0];
6079         }
6080         /*
6081          * when context is MASKED, we will restore PMC with plm=0
6082          * and PMD with stale information, but that's ok, nothing
6083          * will be captured.
6084          *
6085          * XXX: optimize here
6086          */
6087         if (pmd_mask) pfm_restore_pmds(ctx->th_pmds, pmd_mask);
6088         if (pmc_mask) pfm_restore_pmcs(ctx->th_pmcs, pmc_mask);
6089 
6090         /*
6091          * check for pending overflow at the time the state
6092          * was saved.
6093          */
6094         if (unlikely(PMC0_HAS_OVFL(ctx->th_pmcs[0]))) {
6095                 /*
6096                  * reload pmc0 with the overflow information
6097                  * On McKinley PMU, this will trigger a PMU interrupt
6098                  */
6099                 ia64_set_pmc(0, ctx->th_pmcs[0]);
6100                 ia64_srlz_d();
6101                 ctx->th_pmcs[0] = 0UL;
6102 
6103                 /*
6104                  * will replay the PMU interrupt
6105                  */
6106                 if (need_irq_resend) ia64_resend_irq(IA64_PERFMON_VECTOR);
6107 
6108                 pfm_stats[smp_processor_id()].pfm_replay_ovfl_intr_count++;
6109         }
6110 
6111         /*
6112          * we just did a reload, so we reset the partial reload fields
6113          */
6114         ctx->ctx_reload_pmcs[0] = 0UL;
6115         ctx->ctx_reload_pmds[0] = 0UL;
6116 
6117         SET_LAST_CPU(ctx, smp_processor_id());
6118 
6119         /*
6120          * dump activation value for this PMU
6121          */
6122         INC_ACTIVATION();
6123         /*
6124          * record current activation for this context
6125          */
6126         SET_ACTIVATION(ctx);
6127 
6128         /*
6129          * establish new ownership. 
6130          */
6131         SET_PMU_OWNER(task, ctx);
6132 
6133         /*
6134          * restore the psr.up bit. measurement
6135          * is active again.
6136          * no PMU interrupt can happen at this point
6137          * because we still have interrupts disabled.
6138          */
6139         if (likely(psr_up)) pfm_set_psr_up();
6140 
6141         /*
6142          * allow concurrent access to context
6143          */
6144         pfm_unprotect_ctx_ctxsw(ctx, flags);
6145 }
6146 #else /*  !CONFIG_SMP */
6147 /*
6148  * reload PMU state for UP kernels
6149  * in 2.5 we come here with interrupts disabled
6150  */
6151 void
6152 pfm_load_regs (struct task_struct *task)
6153 {
6154         pfm_context_t *ctx;
6155         struct task_struct *owner;
6156         unsigned long pmd_mask, pmc_mask;
6157         u64 psr, psr_up;
6158         int need_irq_resend;
6159 
6160         owner = GET_PMU_OWNER();
6161         ctx   = PFM_GET_CTX(task);
6162         psr   = pfm_get_psr();
6163 
6164         BUG_ON(psr & (IA64_PSR_UP|IA64_PSR_PP));
6165         BUG_ON(psr & IA64_PSR_I);
6166 
6167         /*
6168          * we restore ALL the debug registers to avoid picking up
6169          * stale state.
6170          *
6171          * This must be done even when the task is still the owner
6172          * as the registers may have been modified via ptrace()
6173          * (not perfmon) by the previous task.
6174          */
6175         if (ctx->ctx_fl_using_dbreg) {
6176                 pfm_restore_ibrs(ctx->ctx_ibrs, pmu_conf->num_ibrs);
6177                 pfm_restore_dbrs(ctx->ctx_dbrs, pmu_conf->num_dbrs);
6178         }
6179 
6180         /*
6181          * retrieved saved psr.up
6182          */
6183         psr_up = ctx->ctx_saved_psr_up;
6184         need_irq_resend = pmu_conf->flags & PFM_PMU_IRQ_RESEND;
6185 
6186         /*
6187          * short path, our state is still there, just
6188          * need to restore psr and we go
6189          *
6190          * we do not touch either PMC nor PMD. the psr is not touched
6191          * by the overflow_handler. So we are safe w.r.t. to interrupt
6192          * concurrency even without interrupt masking.
6193          */
6194         if (likely(owner == task)) {
6195                 if (likely(psr_up)) pfm_set_psr_up();
6196                 return;
6197         }
6198 
6199         /*
6200          * someone else is still using the PMU, first push it out and
6201          * then we'll be able to install our stuff !
6202          *
6203          * Upon return, there will be no owner for the current PMU
6204          */
6205         if (owner) pfm_lazy_save_regs(owner);
6206 
6207         /*
6208          * To avoid leaking information to the user level when psr.sp=0,
6209          * we must reload ALL implemented pmds (even the ones we don't use).
6210          * In the kernel we only allow PFM_READ_PMDS on registers which
6211          * we initialized or requested (sampling) so there is no risk there.
6212          */
6213         pmd_mask = pfm_sysctl.fastctxsw ?  ctx->ctx_used_pmds[0] : ctx->ctx_all_pmds[0];
6214 
6215         /*
6216          * ALL accessible PMCs are systematically reloaded, unused registers
6217          * get their default (from pfm_reset_pmu_state()) values to avoid picking
6218          * up stale configuration.
6219          *
6220          * PMC0 is never in the mask. It is always restored separately
6221          */
6222         pmc_mask = ctx->ctx_all_pmcs[0];
6223 
6224         pfm_restore_pmds(ctx->th_pmds, pmd_mask);
6225         pfm_restore_pmcs(ctx->th_pmcs, pmc_mask);
6226 
6227         /*
6228          * check for pending overflow at the time the state
6229          * was saved.
6230          */
6231         if (unlikely(PMC0_HAS_OVFL(ctx->th_pmcs[0]))) {
6232                 /*
6233                  * reload pmc0 with the overflow information
6234                  * On McKinley PMU, this will trigger a PMU interrupt
6235                  */
6236                 ia64_set_pmc(0, ctx->th_pmcs[0]);
6237                 ia64_srlz_d();
6238 
6239                 ctx->th_pmcs[0] = 0UL;
6240 
6241                 /*
6242                  * will replay the PMU interrupt
6243                  */
6244                 if (need_irq_resend) ia64_resend_irq(IA64_PERFMON_VECTOR);
6245 
6246                 pfm_stats[smp_processor_id()].pfm_replay_ovfl_intr_count++;
6247         }
6248 
6249         /*
6250          * establish new ownership. 
6251          */
6252         SET_PMU_OWNER(task, ctx);
6253 
6254         /*
6255          * restore the psr.up bit. measurement
6256          * is active again.
6257          * no PMU interrupt can happen at this point
6258          * because we still have interrupts disabled.
6259          */
6260         if (likely(psr_up)) pfm_set_psr_up();
6261 }
6262 #endif /* CONFIG_SMP */
6263 
6264 /*
6265  * this function assumes monitoring is stopped
6266  */
6267 static void
6268 pfm_flush_pmds(struct task_struct *task, pfm_context_t *ctx)
6269 {
6270         u64 pmc0;
6271         unsigned long mask2, val, pmd_val, ovfl_val;
6272         int i, can_access_pmu = 0;
6273         int is_self;
6274 
6275         /*
6276          * is the caller the task being monitored (or which initiated the
6277          * session for system wide measurements)
6278          */
6279         is_self = ctx->ctx_task == task ? 1 : 0;
6280 
6281         /*
6282          * can access PMU is task is the owner of the PMU state on the current CPU
6283          * or if we are running on the CPU bound to the context in system-wide mode
6284          * (that is not necessarily the task the context is attached to in this mode).
6285          * In system-wide we always have can_access_pmu true because a task running on an
6286          * invalid processor is flagged earlier in the call stack (see pfm_stop).
6287          */
6288         can_access_pmu = (GET_PMU_OWNER() == task) || (ctx->ctx_fl_system && ctx->ctx_cpu == smp_processor_id());
6289         if (can_access_pmu) {
6290                 /*
6291                  * Mark the PMU as not owned
6292                  * This will cause the interrupt handler to do nothing in case an overflow
6293                  * interrupt was in-flight
6294                  * This also guarantees that pmc0 will contain the final state
6295                  * It virtually gives us full control on overflow processing from that point
6296                  * on.
6297                  */
6298                 SET_PMU_OWNER(NULL, NULL);
6299                 DPRINT(("releasing ownership\n"));
6300 
6301                 /*
6302                  * read current overflow status:
6303                  *
6304                  * we are guaranteed to read the final stable state
6305                  */
6306                 ia64_srlz_d();
6307                 pmc0 = ia64_get_pmc(0); /* slow */
6308 
6309                 /*
6310                  * reset freeze bit, overflow status information destroyed
6311                  */
6312                 pfm_unfreeze_pmu();
6313         } else {
6314                 pmc0 = ctx->th_pmcs[0];
6315                 /*
6316                  * clear whatever overflow status bits there were
6317                  */
6318                 ctx->th_pmcs[0] = 0;
6319         }
6320         ovfl_val = pmu_conf->ovfl_val;
6321         /*
6322          * we save all the used pmds
6323          * we take care of overflows for counting PMDs
6324          *
6325          * XXX: sampling situation is not taken into account here
6326          */
6327         mask2 = ctx->ctx_used_pmds[0];
6328 
6329         DPRINT(("is_self=%d ovfl_val=0x%lx mask2=0x%lx\n", is_self, ovfl_val, mask2));
6330 
6331         for (i = 0; mask2; i++, mask2>>=1) {
6332 
6333                 /* skip non used pmds */
6334                 if ((mask2 & 0x1) == 0) continue;
6335 
6336                 /*
6337                  * can access PMU always true in system wide mode
6338                  */
6339                 val = pmd_val = can_access_pmu ? ia64_get_pmd(i) : ctx->th_pmds[i];
6340 
6341                 if (PMD_IS_COUNTING(i)) {
6342                         DPRINT(("[%d] pmd[%d] ctx_pmd=0x%lx hw_pmd=0x%lx\n",
6343                                 task_pid_nr(task),
6344                                 i,
6345                                 ctx->ctx_pmds[i].val,
6346                                 val & ovfl_val));
6347 
6348                         /*
6349                          * we rebuild the full 64 bit value of the counter
6350                          */
6351                         val = ctx->ctx_pmds[i].val + (val & ovfl_val);
6352 
6353                         /*
6354                          * now everything is in ctx_pmds[] and we need
6355                          * to clear the saved context from save_regs() such that
6356                          * pfm_read_pmds() gets the correct value
6357                          */
6358                         pmd_val = 0UL;
6359 
6360                         /*
6361                          * take care of overflow inline
6362                          */
6363                         if (pmc0 & (1UL << i)) {
6364                                 val += 1 + ovfl_val;
6365                                 DPRINT(("[%d] pmd[%d] overflowed\n", task_pid_nr(task), i));
6366                         }
6367                 }
6368 
6369                 DPRINT(("[%d] ctx_pmd[%d]=0x%lx  pmd_val=0x%lx\n", task_pid_nr(task), i, val, pmd_val));
6370 
6371                 if (is_self) ctx->th_pmds[i] = pmd_val;
6372 
6373                 ctx->ctx_pmds[i].val = val;
6374         }
6375 }
6376 
6377 static struct irqaction perfmon_irqaction = {
6378         .handler = pfm_interrupt_handler,
6379         .name    = "perfmon"
6380 };
6381 
6382 static void
6383 pfm_alt_save_pmu_state(void *data)
6384 {
6385         struct pt_regs *regs;
6386 
6387         regs = task_pt_regs(current);
6388 
6389         DPRINT(("called\n"));
6390 
6391         /*
6392          * should not be necessary but
6393          * let's take not risk
6394          */
6395         pfm_clear_psr_up();
6396         pfm_clear_psr_pp();
6397         ia64_psr(regs)->pp = 0;
6398 
6399         /*
6400          * This call is required
6401          * May cause a spurious interrupt on some processors
6402          */
6403         pfm_freeze_pmu();
6404 
6405         ia64_srlz_d();
6406 }
6407 
6408 void
6409 pfm_alt_restore_pmu_state(void *data)
6410 {
6411         struct pt_regs *regs;
6412 
6413         regs = task_pt_regs(current);
6414 
6415         DPRINT(("called\n"));
6416 
6417         /*
6418          * put PMU back in state expected
6419          * by perfmon
6420          */
6421         pfm_clear_psr_up();
6422         pfm_clear_psr_pp();
6423         ia64_psr(regs)->pp = 0;
6424 
6425         /*
6426          * perfmon runs with PMU unfrozen at all times
6427          */
6428         pfm_unfreeze_pmu();
6429 
6430         ia64_srlz_d();
6431 }
6432 
6433 int
6434 pfm_install_alt_pmu_interrupt(pfm_intr_handler_desc_t *hdl)
6435 {
6436         int ret, i;
6437         int reserve_cpu;
6438 
6439         /* some sanity checks */
6440         if (hdl == NULL || hdl->handler == NULL) return -EINVAL;
6441 
6442         /* do the easy test first */
6443         if (pfm_alt_intr_handler) return -EBUSY;
6444 
6445         /* one at a time in the install or remove, just fail the others */
6446         if (!spin_trylock(&pfm_alt_install_check)) {
6447                 return -EBUSY;
6448         }
6449 
6450         /* reserve our session */
6451         for_each_online_cpu(reserve_cpu) {
6452                 ret = pfm_reserve_session(NULL, 1, reserve_cpu);
6453                 if (ret) goto cleanup_reserve;
6454         }
6455 
6456         /* save the current system wide pmu states */
6457         ret = on_each_cpu(pfm_alt_save_pmu_state, NULL, 1);
6458         if (ret) {
6459                 DPRINT(("on_each_cpu() failed: %d\n", ret));
6460                 goto cleanup_reserve;
6461         }
6462 
6463         /* officially change to the alternate interrupt handler */
6464