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

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

Version: ~ [ linux-5.4-rc3 ] ~ [ linux-5.3.6 ] ~ [ linux-5.2.21 ] ~ [ linux-5.1.21 ] ~ [ linux-5.0.21 ] ~ [ linux-4.20.17 ] ~ [ linux-4.19.79 ] ~ [ linux-4.18.20 ] ~ [ linux-4.17.19 ] ~ [ linux-4.16.18 ] ~ [ linux-4.15.18 ] ~ [ linux-4.14.149 ] ~ [ linux-4.13.16 ] ~ [ linux-4.12.14 ] ~ [ linux-4.11.12 ] ~ [ linux-4.10.17 ] ~ [ linux-4.9.196 ] ~ [ linux-4.8.17 ] ~ [ linux-4.7.10 ] ~ [ linux-4.6.7 ] ~ [ linux-4.5.7 ] ~ [ linux-4.4.196 ] ~ [ 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-3.16.75 ] ~ [ 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->vm_file,
2336                                                         vma_pages(vma));
2337         up_write(&task->mm->mmap_sem);
2338 
2339         /*
2340          * keep track of user level virtual address
2341          */
2342         ctx->ctx_smpl_vaddr = (void *)vma->vm_start;
2343         *(unsigned long *)user_vaddr = vma->vm_start;
2344 
2345         return 0;
2346 
2347 error:
2348         kmem_cache_free(vm_area_cachep, vma);
2349 error_kmem:
2350         pfm_rvfree(smpl_buf, size);
2351 
2352         return -ENOMEM;
2353 }
2354 
2355 /*
2356  * XXX: do something better here
2357  */
2358 static int
2359 pfm_bad_permissions(struct task_struct *task)
2360 {
2361         const struct cred *tcred;
2362         kuid_t uid = current_uid();
2363         kgid_t gid = current_gid();
2364         int ret;
2365 
2366         rcu_read_lock();
2367         tcred = __task_cred(task);
2368 
2369         /* inspired by ptrace_attach() */
2370         DPRINT(("cur: uid=%d gid=%d task: euid=%d suid=%d uid=%d egid=%d sgid=%d\n",
2371                 from_kuid(&init_user_ns, uid),
2372                 from_kgid(&init_user_ns, gid),
2373                 from_kuid(&init_user_ns, tcred->euid),
2374                 from_kuid(&init_user_ns, tcred->suid),
2375                 from_kuid(&init_user_ns, tcred->uid),
2376                 from_kgid(&init_user_ns, tcred->egid),
2377                 from_kgid(&init_user_ns, tcred->sgid)));
2378 
2379         ret = ((!uid_eq(uid, tcred->euid))
2380                || (!uid_eq(uid, tcred->suid))
2381                || (!uid_eq(uid, tcred->uid))
2382                || (!gid_eq(gid, tcred->egid))
2383                || (!gid_eq(gid, tcred->sgid))
2384                || (!gid_eq(gid, tcred->gid))) && !capable(CAP_SYS_PTRACE);
2385 
2386         rcu_read_unlock();
2387         return ret;
2388 }
2389 
2390 static int
2391 pfarg_is_sane(struct task_struct *task, pfarg_context_t *pfx)
2392 {
2393         int ctx_flags;
2394 
2395         /* valid signal */
2396 
2397         ctx_flags = pfx->ctx_flags;
2398 
2399         if (ctx_flags & PFM_FL_SYSTEM_WIDE) {
2400 
2401                 /*
2402                  * cannot block in this mode
2403                  */
2404                 if (ctx_flags & PFM_FL_NOTIFY_BLOCK) {
2405                         DPRINT(("cannot use blocking mode when in system wide monitoring\n"));
2406                         return -EINVAL;
2407                 }
2408         } else {
2409         }
2410         /* probably more to add here */
2411 
2412         return 0;
2413 }
2414 
2415 static int
2416 pfm_setup_buffer_fmt(struct task_struct *task, struct file *filp, pfm_context_t *ctx, unsigned int ctx_flags,
2417                      unsigned int cpu, pfarg_context_t *arg)
2418 {
2419         pfm_buffer_fmt_t *fmt = NULL;
2420         unsigned long size = 0UL;
2421         void *uaddr = NULL;
2422         void *fmt_arg = NULL;
2423         int ret = 0;
2424 #define PFM_CTXARG_BUF_ARG(a)   (pfm_buffer_fmt_t *)(a+1)
2425 
2426         /* invoke and lock buffer format, if found */
2427         fmt = pfm_find_buffer_fmt(arg->ctx_smpl_buf_id);
2428         if (fmt == NULL) {
2429                 DPRINT(("[%d] cannot find buffer format\n", task_pid_nr(task)));
2430                 return -EINVAL;
2431         }
2432 
2433         /*
2434          * buffer argument MUST be contiguous to pfarg_context_t
2435          */
2436         if (fmt->fmt_arg_size) fmt_arg = PFM_CTXARG_BUF_ARG(arg);
2437 
2438         ret = pfm_buf_fmt_validate(fmt, task, ctx_flags, cpu, fmt_arg);
2439 
2440         DPRINT(("[%d] after validate(0x%x,%d,%p)=%d\n", task_pid_nr(task), ctx_flags, cpu, fmt_arg, ret));
2441 
2442         if (ret) goto error;
2443 
2444         /* link buffer format and context */
2445         ctx->ctx_buf_fmt = fmt;
2446         ctx->ctx_fl_is_sampling = 1; /* assume record() is defined */
2447 
2448         /*
2449          * check if buffer format wants to use perfmon buffer allocation/mapping service
2450          */
2451         ret = pfm_buf_fmt_getsize(fmt, task, ctx_flags, cpu, fmt_arg, &size);
2452         if (ret) goto error;
2453 
2454         if (size) {
2455                 /*
2456                  * buffer is always remapped into the caller's address space
2457                  */
2458                 ret = pfm_smpl_buffer_alloc(current, filp, ctx, size, &uaddr);
2459                 if (ret) goto error;
2460 
2461                 /* keep track of user address of buffer */
2462                 arg->ctx_smpl_vaddr = uaddr;
2463         }
2464         ret = pfm_buf_fmt_init(fmt, task, ctx->ctx_smpl_hdr, ctx_flags, cpu, fmt_arg);
2465 
2466 error:
2467         return ret;
2468 }
2469 
2470 static void
2471 pfm_reset_pmu_state(pfm_context_t *ctx)
2472 {
2473         int i;
2474 
2475         /*
2476          * install reset values for PMC.
2477          */
2478         for (i=1; PMC_IS_LAST(i) == 0; i++) {
2479                 if (PMC_IS_IMPL(i) == 0) continue;
2480                 ctx->ctx_pmcs[i] = PMC_DFL_VAL(i);
2481                 DPRINT(("pmc[%d]=0x%lx\n", i, ctx->ctx_pmcs[i]));
2482         }
2483         /*
2484          * PMD registers are set to 0UL when the context in memset()
2485          */
2486 
2487         /*
2488          * On context switched restore, we must restore ALL pmc and ALL pmd even
2489          * when they are not actively used by the task. In UP, the incoming process
2490          * may otherwise pick up left over PMC, PMD state from the previous process.
2491          * As opposed to PMD, stale PMC can cause harm to the incoming
2492          * process because they may change what is being measured.
2493          * Therefore, we must systematically reinstall the entire
2494          * PMC state. In SMP, the same thing is possible on the
2495          * same CPU but also on between 2 CPUs.
2496          *
2497          * The problem with PMD is information leaking especially
2498          * to user level when psr.sp=0
2499          *
2500          * There is unfortunately no easy way to avoid this problem
2501          * on either UP or SMP. This definitively slows down the
2502          * pfm_load_regs() function.
2503          */
2504 
2505          /*
2506           * bitmask of all PMCs accessible to this context
2507           *
2508           * PMC0 is treated differently.
2509           */
2510         ctx->ctx_all_pmcs[0] = pmu_conf->impl_pmcs[0] & ~0x1;
2511 
2512         /*
2513          * bitmask of all PMDs that are accessible to this context
2514          */
2515         ctx->ctx_all_pmds[0] = pmu_conf->impl_pmds[0];
2516 
2517         DPRINT(("<%d> all_pmcs=0x%lx all_pmds=0x%lx\n", ctx->ctx_fd, ctx->ctx_all_pmcs[0],ctx->ctx_all_pmds[0]));
2518 
2519         /*
2520          * useful in case of re-enable after disable
2521          */
2522         ctx->ctx_used_ibrs[0] = 0UL;
2523         ctx->ctx_used_dbrs[0] = 0UL;
2524 }
2525 
2526 static int
2527 pfm_ctx_getsize(void *arg, size_t *sz)
2528 {
2529         pfarg_context_t *req = (pfarg_context_t *)arg;
2530         pfm_buffer_fmt_t *fmt;
2531 
2532         *sz = 0;
2533 
2534         if (!pfm_uuid_cmp(req->ctx_smpl_buf_id, pfm_null_uuid)) return 0;
2535 
2536         fmt = pfm_find_buffer_fmt(req->ctx_smpl_buf_id);
2537         if (fmt == NULL) {
2538                 DPRINT(("cannot find buffer format\n"));
2539                 return -EINVAL;
2540         }
2541         /* get just enough to copy in user parameters */
2542         *sz = fmt->fmt_arg_size;
2543         DPRINT(("arg_size=%lu\n", *sz));
2544 
2545         return 0;
2546 }
2547 
2548 
2549 
2550 /*
2551  * cannot attach if :
2552  *      - kernel task
2553  *      - task not owned by caller
2554  *      - task incompatible with context mode
2555  */
2556 static int
2557 pfm_task_incompatible(pfm_context_t *ctx, struct task_struct *task)
2558 {
2559         /*
2560          * no kernel task or task not owner by caller
2561          */
2562         if (task->mm == NULL) {
2563                 DPRINT(("task [%d] has not memory context (kernel thread)\n", task_pid_nr(task)));
2564                 return -EPERM;
2565         }
2566         if (pfm_bad_permissions(task)) {
2567                 DPRINT(("no permission to attach to  [%d]\n", task_pid_nr(task)));
2568                 return -EPERM;
2569         }
2570         /*
2571          * cannot block in self-monitoring mode
2572          */
2573         if (CTX_OVFL_NOBLOCK(ctx) == 0 && task == current) {
2574                 DPRINT(("cannot load a blocking context on self for [%d]\n", task_pid_nr(task)));
2575                 return -EINVAL;
2576         }
2577 
2578         if (task->exit_state == EXIT_ZOMBIE) {
2579                 DPRINT(("cannot attach to  zombie task [%d]\n", task_pid_nr(task)));
2580                 return -EBUSY;
2581         }
2582 
2583         /*
2584          * always ok for self
2585          */
2586         if (task == current) return 0;
2587 
2588         if (!task_is_stopped_or_traced(task)) {
2589                 DPRINT(("cannot attach to non-stopped task [%d] state=%ld\n", task_pid_nr(task), task->state));
2590                 return -EBUSY;
2591         }
2592         /*
2593          * make sure the task is off any CPU
2594          */
2595         wait_task_inactive(task, 0);
2596 
2597         /* more to come... */
2598 
2599         return 0;
2600 }
2601 
2602 static int
2603 pfm_get_task(pfm_context_t *ctx, pid_t pid, struct task_struct **task)
2604 {
2605         struct task_struct *p = current;
2606         int ret;
2607 
2608         /* XXX: need to add more checks here */
2609         if (pid < 2) return -EPERM;
2610 
2611         if (pid != task_pid_vnr(current)) {
2612 
2613                 read_lock(&tasklist_lock);
2614 
2615                 p = find_task_by_vpid(pid);
2616 
2617                 /* make sure task cannot go away while we operate on it */
2618                 if (p) get_task_struct(p);
2619 
2620                 read_unlock(&tasklist_lock);
2621 
2622                 if (p == NULL) return -ESRCH;
2623         }
2624 
2625         ret = pfm_task_incompatible(ctx, p);
2626         if (ret == 0) {
2627                 *task = p;
2628         } else if (p != current) {
2629                 pfm_put_task(p);
2630         }
2631         return ret;
2632 }
2633 
2634 
2635 
2636 static int
2637 pfm_context_create(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
2638 {
2639         pfarg_context_t *req = (pfarg_context_t *)arg;
2640         struct file *filp;
2641         struct path path;
2642         int ctx_flags;
2643         int fd;
2644         int ret;
2645 
2646         /* let's check the arguments first */
2647         ret = pfarg_is_sane(current, req);
2648         if (ret < 0)
2649                 return ret;
2650 
2651         ctx_flags = req->ctx_flags;
2652 
2653         ret = -ENOMEM;
2654 
2655         fd = get_unused_fd_flags(0);
2656         if (fd < 0)
2657                 return fd;
2658 
2659         ctx = pfm_context_alloc(ctx_flags);
2660         if (!ctx)
2661                 goto error;
2662 
2663         filp = pfm_alloc_file(ctx);
2664         if (IS_ERR(filp)) {
2665                 ret = PTR_ERR(filp);
2666                 goto error_file;
2667         }
2668 
2669         req->ctx_fd = ctx->ctx_fd = fd;
2670 
2671         /*
2672          * does the user want to sample?
2673          */
2674         if (pfm_uuid_cmp(req->ctx_smpl_buf_id, pfm_null_uuid)) {
2675                 ret = pfm_setup_buffer_fmt(current, filp, ctx, ctx_flags, 0, req);
2676                 if (ret)
2677                         goto buffer_error;
2678         }
2679 
2680         DPRINT(("ctx=%p flags=0x%x system=%d notify_block=%d excl_idle=%d no_msg=%d ctx_fd=%d\n",
2681                 ctx,
2682                 ctx_flags,
2683                 ctx->ctx_fl_system,
2684                 ctx->ctx_fl_block,
2685                 ctx->ctx_fl_excl_idle,
2686                 ctx->ctx_fl_no_msg,
2687                 ctx->ctx_fd));
2688 
2689         /*
2690          * initialize soft PMU state
2691          */
2692         pfm_reset_pmu_state(ctx);
2693 
2694         fd_install(fd, filp);
2695 
2696         return 0;
2697 
2698 buffer_error:
2699         path = filp->f_path;
2700         put_filp(filp);
2701         path_put(&path);
2702 
2703         if (ctx->ctx_buf_fmt) {
2704                 pfm_buf_fmt_exit(ctx->ctx_buf_fmt, current, NULL, regs);
2705         }
2706 error_file:
2707         pfm_context_free(ctx);
2708 
2709 error:
2710         put_unused_fd(fd);
2711         return ret;
2712 }
2713 
2714 static inline unsigned long
2715 pfm_new_counter_value (pfm_counter_t *reg, int is_long_reset)
2716 {
2717         unsigned long val = is_long_reset ? reg->long_reset : reg->short_reset;
2718         unsigned long new_seed, old_seed = reg->seed, mask = reg->mask;
2719         extern unsigned long carta_random32 (unsigned long seed);
2720 
2721         if (reg->flags & PFM_REGFL_RANDOM) {
2722                 new_seed = carta_random32(old_seed);
2723                 val -= (old_seed & mask);       /* counter values are negative numbers! */
2724                 if ((mask >> 32) != 0)
2725                         /* construct a full 64-bit random value: */
2726                         new_seed |= carta_random32(old_seed >> 32) << 32;
2727                 reg->seed = new_seed;
2728         }
2729         reg->lval = val;
2730         return val;
2731 }
2732 
2733 static void
2734 pfm_reset_regs_masked(pfm_context_t *ctx, unsigned long *ovfl_regs, int is_long_reset)
2735 {
2736         unsigned long mask = ovfl_regs[0];
2737         unsigned long reset_others = 0UL;
2738         unsigned long val;
2739         int i;
2740 
2741         /*
2742          * now restore reset value on sampling overflowed counters
2743          */
2744         mask >>= PMU_FIRST_COUNTER;
2745         for(i = PMU_FIRST_COUNTER; mask; i++, mask >>= 1) {
2746 
2747                 if ((mask & 0x1UL) == 0UL) continue;
2748 
2749                 ctx->ctx_pmds[i].val = val = pfm_new_counter_value(ctx->ctx_pmds+ i, is_long_reset);
2750                 reset_others        |= ctx->ctx_pmds[i].reset_pmds[0];
2751 
2752                 DPRINT_ovfl((" %s reset ctx_pmds[%d]=%lx\n", is_long_reset ? "long" : "short", i, val));
2753         }
2754 
2755         /*
2756          * Now take care of resetting the other registers
2757          */
2758         for(i = 0; reset_others; i++, reset_others >>= 1) {
2759 
2760                 if ((reset_others & 0x1) == 0) continue;
2761 
2762                 ctx->ctx_pmds[i].val = val = pfm_new_counter_value(ctx->ctx_pmds + i, is_long_reset);
2763 
2764                 DPRINT_ovfl(("%s reset_others pmd[%d]=%lx\n",
2765                           is_long_reset ? "long" : "short", i, val));
2766         }
2767 }
2768 
2769 static void
2770 pfm_reset_regs(pfm_context_t *ctx, unsigned long *ovfl_regs, int is_long_reset)
2771 {
2772         unsigned long mask = ovfl_regs[0];
2773         unsigned long reset_others = 0UL;
2774         unsigned long val;
2775         int i;
2776 
2777         DPRINT_ovfl(("ovfl_regs=0x%lx is_long_reset=%d\n", ovfl_regs[0], is_long_reset));
2778 
2779         if (ctx->ctx_state == PFM_CTX_MASKED) {
2780                 pfm_reset_regs_masked(ctx, ovfl_regs, is_long_reset);
2781                 return;
2782         }
2783 
2784         /*
2785          * now restore reset value on sampling overflowed counters
2786          */
2787         mask >>= PMU_FIRST_COUNTER;
2788         for(i = PMU_FIRST_COUNTER; mask; i++, mask >>= 1) {
2789 
2790                 if ((mask & 0x1UL) == 0UL) continue;
2791 
2792                 val           = pfm_new_counter_value(ctx->ctx_pmds+ i, is_long_reset);
2793                 reset_others |= ctx->ctx_pmds[i].reset_pmds[0];
2794 
2795                 DPRINT_ovfl((" %s reset ctx_pmds[%d]=%lx\n", is_long_reset ? "long" : "short", i, val));
2796 
2797                 pfm_write_soft_counter(ctx, i, val);
2798         }
2799 
2800         /*
2801          * Now take care of resetting the other registers
2802          */
2803         for(i = 0; reset_others; i++, reset_others >>= 1) {
2804 
2805                 if ((reset_others & 0x1) == 0) continue;
2806 
2807                 val = pfm_new_counter_value(ctx->ctx_pmds + i, is_long_reset);
2808 
2809                 if (PMD_IS_COUNTING(i)) {
2810                         pfm_write_soft_counter(ctx, i, val);
2811                 } else {
2812                         ia64_set_pmd(i, val);
2813                 }
2814                 DPRINT_ovfl(("%s reset_others pmd[%d]=%lx\n",
2815                           is_long_reset ? "long" : "short", i, val));
2816         }
2817         ia64_srlz_d();
2818 }
2819 
2820 static int
2821 pfm_write_pmcs(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
2822 {
2823         struct task_struct *task;
2824         pfarg_reg_t *req = (pfarg_reg_t *)arg;
2825         unsigned long value, pmc_pm;
2826         unsigned long smpl_pmds, reset_pmds, impl_pmds;
2827         unsigned int cnum, reg_flags, flags, pmc_type;
2828         int i, can_access_pmu = 0, is_loaded, is_system, expert_mode;
2829         int is_monitor, is_counting, state;
2830         int ret = -EINVAL;
2831         pfm_reg_check_t wr_func;
2832 #define PFM_CHECK_PMC_PM(x, y, z) ((x)->ctx_fl_system ^ PMC_PM(y, z))
2833 
2834         state     = ctx->ctx_state;
2835         is_loaded = state == PFM_CTX_LOADED ? 1 : 0;
2836         is_system = ctx->ctx_fl_system;
2837         task      = ctx->ctx_task;
2838         impl_pmds = pmu_conf->impl_pmds[0];
2839 
2840         if (state == PFM_CTX_ZOMBIE) return -EINVAL;
2841 
2842         if (is_loaded) {
2843                 /*
2844                  * In system wide and when the context is loaded, access can only happen
2845                  * when the caller is running on the CPU being monitored by the session.
2846                  * It does not have to be the owner (ctx_task) of the context per se.
2847                  */
2848                 if (is_system && ctx->ctx_cpu != smp_processor_id()) {
2849                         DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
2850                         return -EBUSY;
2851                 }
2852                 can_access_pmu = GET_PMU_OWNER() == task || is_system ? 1 : 0;
2853         }
2854         expert_mode = pfm_sysctl.expert_mode; 
2855 
2856         for (i = 0; i < count; i++, req++) {
2857 
2858                 cnum       = req->reg_num;
2859                 reg_flags  = req->reg_flags;
2860                 value      = req->reg_value;
2861                 smpl_pmds  = req->reg_smpl_pmds[0];
2862                 reset_pmds = req->reg_reset_pmds[0];
2863                 flags      = 0;
2864 
2865 
2866                 if (cnum >= PMU_MAX_PMCS) {
2867                         DPRINT(("pmc%u is invalid\n", cnum));
2868                         goto error;
2869                 }
2870 
2871                 pmc_type   = pmu_conf->pmc_desc[cnum].type;
2872                 pmc_pm     = (value >> pmu_conf->pmc_desc[cnum].pm_pos) & 0x1;
2873                 is_counting = (pmc_type & PFM_REG_COUNTING) == PFM_REG_COUNTING ? 1 : 0;
2874                 is_monitor  = (pmc_type & PFM_REG_MONITOR) == PFM_REG_MONITOR ? 1 : 0;
2875 
2876                 /*
2877                  * we reject all non implemented PMC as well
2878                  * as attempts to modify PMC[0-3] which are used
2879                  * as status registers by the PMU
2880                  */
2881                 if ((pmc_type & PFM_REG_IMPL) == 0 || (pmc_type & PFM_REG_CONTROL) == PFM_REG_CONTROL) {
2882                         DPRINT(("pmc%u is unimplemented or no-access pmc_type=%x\n", cnum, pmc_type));
2883                         goto error;
2884                 }
2885                 wr_func = pmu_conf->pmc_desc[cnum].write_check;
2886                 /*
2887                  * If the PMC is a monitor, then if the value is not the default:
2888                  *      - system-wide session: PMCx.pm=1 (privileged monitor)
2889                  *      - per-task           : PMCx.pm=0 (user monitor)
2890                  */
2891                 if (is_monitor && value != PMC_DFL_VAL(cnum) && is_system ^ pmc_pm) {
2892                         DPRINT(("pmc%u pmc_pm=%lu is_system=%d\n",
2893                                 cnum,
2894                                 pmc_pm,
2895                                 is_system));
2896                         goto error;
2897                 }
2898 
2899                 if (is_counting) {
2900                         /*
2901                          * enforce generation of overflow interrupt. Necessary on all
2902                          * CPUs.
2903                          */
2904                         value |= 1 << PMU_PMC_OI;
2905 
2906                         if (reg_flags & PFM_REGFL_OVFL_NOTIFY) {
2907                                 flags |= PFM_REGFL_OVFL_NOTIFY;
2908                         }
2909 
2910                         if (reg_flags & PFM_REGFL_RANDOM) flags |= PFM_REGFL_RANDOM;
2911 
2912                         /* verify validity of smpl_pmds */
2913                         if ((smpl_pmds & impl_pmds) != smpl_pmds) {
2914                                 DPRINT(("invalid smpl_pmds 0x%lx for pmc%u\n", smpl_pmds, cnum));
2915                                 goto error;
2916                         }
2917 
2918                         /* verify validity of reset_pmds */
2919                         if ((reset_pmds & impl_pmds) != reset_pmds) {
2920                                 DPRINT(("invalid reset_pmds 0x%lx for pmc%u\n", reset_pmds, cnum));
2921                                 goto error;
2922                         }
2923                 } else {
2924                         if (reg_flags & (PFM_REGFL_OVFL_NOTIFY|PFM_REGFL_RANDOM)) {
2925                                 DPRINT(("cannot set ovfl_notify or random on pmc%u\n", cnum));
2926                                 goto error;
2927                         }
2928                         /* eventid on non-counting monitors are ignored */
2929                 }
2930 
2931                 /*
2932                  * execute write checker, if any
2933                  */
2934                 if (likely(expert_mode == 0 && wr_func)) {
2935                         ret = (*wr_func)(task, ctx, cnum, &value, regs);
2936                         if (ret) goto error;
2937                         ret = -EINVAL;
2938                 }
2939 
2940                 /*
2941                  * no error on this register
2942                  */
2943                 PFM_REG_RETFLAG_SET(req->reg_flags, 0);
2944 
2945                 /*
2946                  * Now we commit the changes to the software state
2947                  */
2948 
2949                 /*
2950                  * update overflow information
2951                  */
2952                 if (is_counting) {
2953                         /*
2954                          * full flag update each time a register is programmed
2955                          */
2956                         ctx->ctx_pmds[cnum].flags = flags;
2957 
2958                         ctx->ctx_pmds[cnum].reset_pmds[0] = reset_pmds;
2959                         ctx->ctx_pmds[cnum].smpl_pmds[0]  = smpl_pmds;
2960                         ctx->ctx_pmds[cnum].eventid       = req->reg_smpl_eventid;
2961 
2962                         /*
2963                          * Mark all PMDS to be accessed as used.
2964                          *
2965                          * We do not keep track of PMC because we have to
2966                          * systematically restore ALL of them.
2967                          *
2968                          * We do not update the used_monitors mask, because
2969                          * if we have not programmed them, then will be in
2970                          * a quiescent state, therefore we will not need to
2971                          * mask/restore then when context is MASKED.
2972                          */
2973                         CTX_USED_PMD(ctx, reset_pmds);
2974                         CTX_USED_PMD(ctx, smpl_pmds);
2975                         /*
2976                          * make sure we do not try to reset on
2977                          * restart because we have established new values
2978                          */
2979                         if (state == PFM_CTX_MASKED) ctx->ctx_ovfl_regs[0] &= ~1UL << cnum;
2980                 }
2981                 /*
2982                  * Needed in case the user does not initialize the equivalent
2983                  * PMD. Clearing is done indirectly via pfm_reset_pmu_state() so there is no
2984                  * possible leak here.
2985                  */
2986                 CTX_USED_PMD(ctx, pmu_conf->pmc_desc[cnum].dep_pmd[0]);
2987 
2988                 /*
2989                  * keep track of the monitor PMC that we are using.
2990                  * we save the value of the pmc in ctx_pmcs[] and if
2991                  * the monitoring is not stopped for the context we also
2992                  * place it in the saved state area so that it will be
2993                  * picked up later by the context switch code.
2994                  *
2995                  * The value in ctx_pmcs[] can only be changed in pfm_write_pmcs().
2996                  *
2997                  * The value in th_pmcs[] may be modified on overflow, i.e.,  when
2998                  * monitoring needs to be stopped.
2999                  */
3000                 if (is_monitor) CTX_USED_MONITOR(ctx, 1UL << cnum);
3001 
3002                 /*
3003                  * update context state
3004                  */
3005                 ctx->ctx_pmcs[cnum] = value;
3006 
3007                 if (is_loaded) {
3008                         /*
3009                          * write thread state
3010                          */
3011                         if (is_system == 0) ctx->th_pmcs[cnum] = value;
3012 
3013                         /*
3014                          * write hardware register if we can
3015                          */
3016                         if (can_access_pmu) {
3017                                 ia64_set_pmc(cnum, value);
3018                         }
3019 #ifdef CONFIG_SMP
3020                         else {
3021                                 /*
3022                                  * per-task SMP only here
3023                                  *
3024                                  * we are guaranteed that the task is not running on the other CPU,
3025                                  * we indicate that this PMD will need to be reloaded if the task
3026                                  * is rescheduled on the CPU it ran last on.
3027                                  */
3028                                 ctx->ctx_reload_pmcs[0] |= 1UL << cnum;
3029                         }
3030 #endif
3031                 }
3032 
3033                 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",
3034                           cnum,
3035                           value,
3036                           is_loaded,
3037                           can_access_pmu,
3038                           flags,
3039                           ctx->ctx_all_pmcs[0],
3040                           ctx->ctx_used_pmds[0],
3041                           ctx->ctx_pmds[cnum].eventid,
3042                           smpl_pmds,
3043                           reset_pmds,
3044                           ctx->ctx_reload_pmcs[0],
3045                           ctx->ctx_used_monitors[0],
3046                           ctx->ctx_ovfl_regs[0]));
3047         }
3048 
3049         /*
3050          * make sure the changes are visible
3051          */
3052         if (can_access_pmu) ia64_srlz_d();
3053 
3054         return 0;
3055 error:
3056         PFM_REG_RETFLAG_SET(req->reg_flags, PFM_REG_RETFL_EINVAL);
3057         return ret;
3058 }
3059 
3060 static int
3061 pfm_write_pmds(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3062 {
3063         struct task_struct *task;
3064         pfarg_reg_t *req = (pfarg_reg_t *)arg;
3065         unsigned long value, hw_value, ovfl_mask;
3066         unsigned int cnum;
3067         int i, can_access_pmu = 0, state;
3068         int is_counting, is_loaded, is_system, expert_mode;
3069         int ret = -EINVAL;
3070         pfm_reg_check_t wr_func;
3071 
3072 
3073         state     = ctx->ctx_state;
3074         is_loaded = state == PFM_CTX_LOADED ? 1 : 0;
3075         is_system = ctx->ctx_fl_system;
3076         ovfl_mask = pmu_conf->ovfl_val;
3077         task      = ctx->ctx_task;
3078 
3079         if (unlikely(state == PFM_CTX_ZOMBIE)) return -EINVAL;
3080 
3081         /*
3082          * on both UP and SMP, we can only write to the PMC when the task is
3083          * the owner of the local PMU.
3084          */
3085         if (likely(is_loaded)) {
3086                 /*
3087                  * In system wide and when the context is loaded, access can only happen
3088                  * when the caller is running on the CPU being monitored by the session.
3089                  * It does not have to be the owner (ctx_task) of the context per se.
3090                  */
3091                 if (unlikely(is_system && ctx->ctx_cpu != smp_processor_id())) {
3092                         DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
3093                         return -EBUSY;
3094                 }
3095                 can_access_pmu = GET_PMU_OWNER() == task || is_system ? 1 : 0;
3096         }
3097         expert_mode = pfm_sysctl.expert_mode; 
3098 
3099         for (i = 0; i < count; i++, req++) {
3100 
3101                 cnum  = req->reg_num;
3102                 value = req->reg_value;
3103 
3104                 if (!PMD_IS_IMPL(cnum)) {
3105                         DPRINT(("pmd[%u] is unimplemented or invalid\n", cnum));
3106                         goto abort_mission;
3107                 }
3108                 is_counting = PMD_IS_COUNTING(cnum);
3109                 wr_func     = pmu_conf->pmd_desc[cnum].write_check;
3110 
3111                 /*
3112                  * execute write checker, if any
3113                  */
3114                 if (unlikely(expert_mode == 0 && wr_func)) {
3115                         unsigned long v = value;
3116 
3117                         ret = (*wr_func)(task, ctx, cnum, &v, regs);
3118                         if (ret) goto abort_mission;
3119 
3120                         value = v;
3121                         ret   = -EINVAL;
3122                 }
3123 
3124                 /*
3125                  * no error on this register
3126                  */
3127                 PFM_REG_RETFLAG_SET(req->reg_flags, 0);
3128 
3129                 /*
3130                  * now commit changes to software state
3131                  */
3132                 hw_value = value;
3133 
3134                 /*
3135                  * update virtualized (64bits) counter
3136                  */
3137                 if (is_counting) {
3138                         /*
3139                          * write context state
3140                          */
3141                         ctx->ctx_pmds[cnum].lval = value;
3142 
3143                         /*
3144                          * when context is load we use the split value
3145                          */
3146                         if (is_loaded) {
3147                                 hw_value = value &  ovfl_mask;
3148                                 value    = value & ~ovfl_mask;
3149                         }
3150                 }
3151                 /*
3152                  * update reset values (not just for counters)
3153                  */
3154                 ctx->ctx_pmds[cnum].long_reset  = req->reg_long_reset;
3155                 ctx->ctx_pmds[cnum].short_reset = req->reg_short_reset;
3156 
3157                 /*
3158                  * update randomization parameters (not just for counters)
3159                  */
3160                 ctx->ctx_pmds[cnum].seed = req->reg_random_seed;
3161                 ctx->ctx_pmds[cnum].mask = req->reg_random_mask;
3162 
3163                 /*
3164                  * update context value
3165                  */
3166                 ctx->ctx_pmds[cnum].val  = value;
3167 
3168                 /*
3169                  * Keep track of what we use
3170                  *
3171                  * We do not keep track of PMC because we have to
3172                  * systematically restore ALL of them.
3173                  */
3174                 CTX_USED_PMD(ctx, PMD_PMD_DEP(cnum));
3175 
3176                 /*
3177                  * mark this PMD register used as well
3178                  */
3179                 CTX_USED_PMD(ctx, RDEP(cnum));
3180 
3181                 /*
3182                  * make sure we do not try to reset on
3183                  * restart because we have established new values
3184                  */
3185                 if (is_counting && state == PFM_CTX_MASKED) {
3186                         ctx->ctx_ovfl_regs[0] &= ~1UL << cnum;
3187                 }
3188 
3189                 if (is_loaded) {
3190                         /*
3191                          * write thread state
3192                          */
3193                         if (is_system == 0) ctx->th_pmds[cnum] = hw_value;
3194 
3195                         /*
3196                          * write hardware register if we can
3197                          */
3198                         if (can_access_pmu) {
3199                                 ia64_set_pmd(cnum, hw_value);
3200                         } else {
3201 #ifdef CONFIG_SMP
3202                                 /*
3203                                  * we are guaranteed that the task is not running on the other CPU,
3204                                  * we indicate that this PMD will need to be reloaded if the task
3205                                  * is rescheduled on the CPU it ran last on.
3206                                  */
3207                                 ctx->ctx_reload_pmds[0] |= 1UL << cnum;
3208 #endif
3209                         }
3210                 }
3211 
3212                 DPRINT(("pmd[%u]=0x%lx ld=%d apmu=%d, hw_value=0x%lx ctx_pmd=0x%lx  short_reset=0x%lx "
3213                           "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",
3214                         cnum,
3215                         value,
3216                         is_loaded,
3217                         can_access_pmu,
3218                         hw_value,
3219                         ctx->ctx_pmds[cnum].val,
3220                         ctx->ctx_pmds[cnum].short_reset,
3221                         ctx->ctx_pmds[cnum].long_reset,
3222                         PMC_OVFL_NOTIFY(ctx, cnum) ? 'Y':'N',
3223                         ctx->ctx_pmds[cnum].seed,
3224                         ctx->ctx_pmds[cnum].mask,
3225                         ctx->ctx_used_pmds[0],
3226                         ctx->ctx_pmds[cnum].reset_pmds[0],
3227                         ctx->ctx_reload_pmds[0],
3228                         ctx->ctx_all_pmds[0],
3229                         ctx->ctx_ovfl_regs[0]));
3230         }
3231 
3232         /*
3233          * make changes visible
3234          */
3235         if (can_access_pmu) ia64_srlz_d();
3236 
3237         return 0;
3238 
3239 abort_mission:
3240         /*
3241          * for now, we have only one possibility for error
3242          */
3243         PFM_REG_RETFLAG_SET(req->reg_flags, PFM_REG_RETFL_EINVAL);
3244         return ret;
3245 }
3246 
3247 /*
3248  * By the way of PROTECT_CONTEXT(), interrupts are masked while we are in this function.
3249  * Therefore we know, we do not have to worry about the PMU overflow interrupt. If an
3250  * interrupt is delivered during the call, it will be kept pending until we leave, making
3251  * it appears as if it had been generated at the UNPROTECT_CONTEXT(). At least we are
3252  * guaranteed to return consistent data to the user, it may simply be old. It is not
3253  * trivial to treat the overflow while inside the call because you may end up in
3254  * some module sampling buffer code causing deadlocks.
3255  */
3256 static int
3257 pfm_read_pmds(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3258 {
3259         struct task_struct *task;
3260         unsigned long val = 0UL, lval, ovfl_mask, sval;
3261         pfarg_reg_t *req = (pfarg_reg_t *)arg;
3262         unsigned int cnum, reg_flags = 0;
3263         int i, can_access_pmu = 0, state;
3264         int is_loaded, is_system, is_counting, expert_mode;
3265         int ret = -EINVAL;
3266         pfm_reg_check_t rd_func;
3267 
3268         /*
3269          * access is possible when loaded only for
3270          * self-monitoring tasks or in UP mode
3271          */
3272 
3273         state     = ctx->ctx_state;
3274         is_loaded = state == PFM_CTX_LOADED ? 1 : 0;
3275         is_system = ctx->ctx_fl_system;
3276         ovfl_mask = pmu_conf->ovfl_val;
3277         task      = ctx->ctx_task;
3278 
3279         if (state == PFM_CTX_ZOMBIE) return -EINVAL;
3280 
3281         if (likely(is_loaded)) {
3282                 /*
3283                  * In system wide and when the context is loaded, access can only happen
3284                  * when the caller is running on the CPU being monitored by the session.
3285                  * It does not have to be the owner (ctx_task) of the context per se.
3286                  */
3287                 if (unlikely(is_system && ctx->ctx_cpu != smp_processor_id())) {
3288                         DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
3289                         return -EBUSY;
3290                 }
3291                 /*
3292                  * this can be true when not self-monitoring only in UP
3293                  */
3294                 can_access_pmu = GET_PMU_OWNER() == task || is_system ? 1 : 0;
3295 
3296                 if (can_access_pmu) ia64_srlz_d();
3297         }
3298         expert_mode = pfm_sysctl.expert_mode; 
3299 
3300         DPRINT(("ld=%d apmu=%d ctx_state=%d\n",
3301                 is_loaded,
3302                 can_access_pmu,
3303                 state));
3304 
3305         /*
3306          * on both UP and SMP, we can only read the PMD from the hardware register when
3307          * the task is the owner of the local PMU.
3308          */
3309 
3310         for (i = 0; i < count; i++, req++) {
3311 
3312                 cnum        = req->reg_num;
3313                 reg_flags   = req->reg_flags;
3314 
3315                 if (unlikely(!PMD_IS_IMPL(cnum))) goto error;
3316                 /*
3317                  * we can only read the register that we use. That includes
3318                  * the one we explicitly initialize AND the one we want included
3319                  * in the sampling buffer (smpl_regs).
3320                  *
3321                  * Having this restriction allows optimization in the ctxsw routine
3322                  * without compromising security (leaks)
3323                  */
3324                 if (unlikely(!CTX_IS_USED_PMD(ctx, cnum))) goto error;
3325 
3326                 sval        = ctx->ctx_pmds[cnum].val;
3327                 lval        = ctx->ctx_pmds[cnum].lval;
3328                 is_counting = PMD_IS_COUNTING(cnum);
3329 
3330                 /*
3331                  * If the task is not the current one, then we check if the
3332                  * PMU state is still in the local live register due to lazy ctxsw.
3333                  * If true, then we read directly from the registers.
3334                  */
3335                 if (can_access_pmu){
3336                         val = ia64_get_pmd(cnum);
3337                 } else {
3338                         /*
3339                          * context has been saved
3340                          * if context is zombie, then task does not exist anymore.
3341                          * In this case, we use the full value saved in the context (pfm_flush_regs()).
3342                          */
3343                         val = is_loaded ? ctx->th_pmds[cnum] : 0UL;
3344                 }
3345                 rd_func = pmu_conf->pmd_desc[cnum].read_check;
3346 
3347                 if (is_counting) {
3348                         /*
3349                          * XXX: need to check for overflow when loaded
3350                          */
3351                         val &= ovfl_mask;
3352                         val += sval;
3353                 }
3354 
3355                 /*
3356                  * execute read checker, if any
3357                  */
3358                 if (unlikely(expert_mode == 0 && rd_func)) {
3359                         unsigned long v = val;
3360                         ret = (*rd_func)(ctx->ctx_task, ctx, cnum, &v, regs);
3361                         if (ret) goto error;
3362                         val = v;
3363                         ret = -EINVAL;
3364                 }
3365 
3366                 PFM_REG_RETFLAG_SET(reg_flags, 0);
3367 
3368                 DPRINT(("pmd[%u]=0x%lx\n", cnum, val));
3369 
3370                 /*
3371                  * update register return value, abort all if problem during copy.
3372                  * we only modify the reg_flags field. no check mode is fine because
3373                  * access has been verified upfront in sys_perfmonctl().
3374                  */
3375                 req->reg_value            = val;
3376                 req->reg_flags            = reg_flags;
3377                 req->reg_last_reset_val   = lval;
3378         }
3379 
3380         return 0;
3381 
3382 error:
3383         PFM_REG_RETFLAG_SET(req->reg_flags, PFM_REG_RETFL_EINVAL);
3384         return ret;
3385 }
3386 
3387 int
3388 pfm_mod_write_pmcs(struct task_struct *task, void *req, unsigned int nreq, struct pt_regs *regs)
3389 {
3390         pfm_context_t *ctx;
3391 
3392         if (req == NULL) return -EINVAL;
3393 
3394         ctx = GET_PMU_CTX();
3395 
3396         if (ctx == NULL) return -EINVAL;
3397 
3398         /*
3399          * for now limit to current task, which is enough when calling
3400          * from overflow handler
3401          */
3402         if (task != current && ctx->ctx_fl_system == 0) return -EBUSY;
3403 
3404         return pfm_write_pmcs(ctx, req, nreq, regs);
3405 }
3406 EXPORT_SYMBOL(pfm_mod_write_pmcs);
3407 
3408 int
3409 pfm_mod_read_pmds(struct task_struct *task, void *req, unsigned int nreq, struct pt_regs *regs)
3410 {
3411         pfm_context_t *ctx;
3412 
3413         if (req == NULL) return -EINVAL;
3414 
3415         ctx = GET_PMU_CTX();
3416 
3417         if (ctx == NULL) return -EINVAL;
3418 
3419         /*
3420          * for now limit to current task, which is enough when calling
3421          * from overflow handler
3422          */
3423         if (task != current && ctx->ctx_fl_system == 0) return -EBUSY;
3424 
3425         return pfm_read_pmds(ctx, req, nreq, regs);
3426 }
3427 EXPORT_SYMBOL(pfm_mod_read_pmds);
3428 
3429 /*
3430  * Only call this function when a process it trying to
3431  * write the debug registers (reading is always allowed)
3432  */
3433 int
3434 pfm_use_debug_registers(struct task_struct *task)
3435 {
3436         pfm_context_t *ctx = task->thread.pfm_context;
3437         unsigned long flags;
3438         int ret = 0;
3439 
3440         if (pmu_conf->use_rr_dbregs == 0) return 0;
3441 
3442         DPRINT(("called for [%d]\n", task_pid_nr(task)));
3443 
3444         /*
3445          * do it only once
3446          */
3447         if (task->thread.flags & IA64_THREAD_DBG_VALID) return 0;
3448 
3449         /*
3450          * Even on SMP, we do not need to use an atomic here because
3451          * the only way in is via ptrace() and this is possible only when the
3452          * process is stopped. Even in the case where the ctxsw out is not totally
3453          * completed by the time we come here, there is no way the 'stopped' process
3454          * could be in the middle of fiddling with the pfm_write_ibr_dbr() routine.
3455          * So this is always safe.
3456          */
3457         if (ctx && ctx->ctx_fl_using_dbreg == 1) return -1;
3458 
3459         LOCK_PFS(flags);
3460 
3461         /*
3462          * We cannot allow setting breakpoints when system wide monitoring
3463          * sessions are using the debug registers.
3464          */
3465         if (pfm_sessions.pfs_sys_use_dbregs> 0)
3466                 ret = -1;
3467         else
3468                 pfm_sessions.pfs_ptrace_use_dbregs++;
3469 
3470         DPRINT(("ptrace_use_dbregs=%u  sys_use_dbregs=%u by [%d] ret = %d\n",
3471                   pfm_sessions.pfs_ptrace_use_dbregs,
3472                   pfm_sessions.pfs_sys_use_dbregs,
3473                   task_pid_nr(task), ret));
3474 
3475         UNLOCK_PFS(flags);
3476 
3477         return ret;
3478 }
3479 
3480 /*
3481  * This function is called for every task that exits with the
3482  * IA64_THREAD_DBG_VALID set. This indicates a task which was
3483  * able to use the debug registers for debugging purposes via
3484  * ptrace(). Therefore we know it was not using them for
3485  * performance monitoring, so we only decrement the number
3486  * of "ptraced" debug register users to keep the count up to date
3487  */
3488 int
3489 pfm_release_debug_registers(struct task_struct *task)
3490 {
3491         unsigned long flags;
3492         int ret;
3493 
3494         if (pmu_conf->use_rr_dbregs == 0) return 0;
3495 
3496         LOCK_PFS(flags);
3497         if (pfm_sessions.pfs_ptrace_use_dbregs == 0) {
3498                 printk(KERN_ERR "perfmon: invalid release for [%d] ptrace_use_dbregs=0\n", task_pid_nr(task));
3499                 ret = -1;
3500         }  else {
3501                 pfm_sessions.pfs_ptrace_use_dbregs--;
3502                 ret = 0;
3503         }
3504         UNLOCK_PFS(flags);
3505 
3506         return ret;
3507 }
3508 
3509 static int
3510 pfm_restart(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3511 {
3512         struct task_struct *task;
3513         pfm_buffer_fmt_t *fmt;
3514         pfm_ovfl_ctrl_t rst_ctrl;
3515         int state, is_system;
3516         int ret = 0;
3517 
3518         state     = ctx->ctx_state;
3519         fmt       = ctx->ctx_buf_fmt;
3520         is_system = ctx->ctx_fl_system;
3521         task      = PFM_CTX_TASK(ctx);
3522 
3523         switch(state) {
3524                 case PFM_CTX_MASKED:
3525                         break;
3526                 case PFM_CTX_LOADED: 
3527                         if (CTX_HAS_SMPL(ctx) && fmt->fmt_restart_active) break;
3528                         /* fall through */
3529                 case PFM_CTX_UNLOADED:
3530                 case PFM_CTX_ZOMBIE:
3531                         DPRINT(("invalid state=%d\n", state));
3532                         return -EBUSY;
3533                 default:
3534                         DPRINT(("state=%d, cannot operate (no active_restart handler)\n", state));
3535                         return -EINVAL;
3536         }
3537 
3538         /*
3539          * In system wide and when the context is loaded, access can only happen
3540          * when the caller is running on the CPU being monitored by the session.
3541          * It does not have to be the owner (ctx_task) of the context per se.
3542          */
3543         if (is_system && ctx->ctx_cpu != smp_processor_id()) {
3544                 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
3545                 return -EBUSY;
3546         }
3547 
3548         /* sanity check */
3549         if (unlikely(task == NULL)) {
3550                 printk(KERN_ERR "perfmon: [%d] pfm_restart no task\n", task_pid_nr(current));
3551                 return -EINVAL;
3552         }
3553 
3554         if (task == current || is_system) {
3555 
3556                 fmt = ctx->ctx_buf_fmt;
3557 
3558                 DPRINT(("restarting self %d ovfl=0x%lx\n",
3559                         task_pid_nr(task),
3560                         ctx->ctx_ovfl_regs[0]));
3561 
3562                 if (CTX_HAS_SMPL(ctx)) {
3563 
3564                         prefetch(ctx->ctx_smpl_hdr);
3565 
3566                         rst_ctrl.bits.mask_monitoring = 0;
3567                         rst_ctrl.bits.reset_ovfl_pmds = 0;
3568 
3569                         if (state == PFM_CTX_LOADED)
3570                                 ret = pfm_buf_fmt_restart_active(fmt, task, &rst_ctrl, ctx->ctx_smpl_hdr, regs);
3571                         else
3572                                 ret = pfm_buf_fmt_restart(fmt, task, &rst_ctrl, ctx->ctx_smpl_hdr, regs);
3573                 } else {
3574                         rst_ctrl.bits.mask_monitoring = 0;
3575                         rst_ctrl.bits.reset_ovfl_pmds = 1;
3576                 }
3577 
3578                 if (ret == 0) {
3579                         if (rst_ctrl.bits.reset_ovfl_pmds)
3580                                 pfm_reset_regs(ctx, ctx->ctx_ovfl_regs, PFM_PMD_LONG_RESET);
3581 
3582                         if (rst_ctrl.bits.mask_monitoring == 0) {
3583                                 DPRINT(("resuming monitoring for [%d]\n", task_pid_nr(task)));
3584 
3585                                 if (state == PFM_CTX_MASKED) pfm_restore_monitoring(task);
3586                         } else {
3587                                 DPRINT(("keeping monitoring stopped for [%d]\n", task_pid_nr(task)));
3588 
3589                                 // cannot use pfm_stop_monitoring(task, regs);
3590                         }
3591                 }
3592                 /*
3593                  * clear overflowed PMD mask to remove any stale information
3594                  */
3595                 ctx->ctx_ovfl_regs[0] = 0UL;
3596 
3597                 /*
3598                  * back to LOADED state
3599                  */
3600                 ctx->ctx_state = PFM_CTX_LOADED;
3601 
3602                 /*
3603                  * XXX: not really useful for self monitoring
3604                  */
3605                 ctx->ctx_fl_can_restart = 0;
3606 
3607                 return 0;
3608         }
3609 
3610         /* 
3611          * restart another task
3612          */
3613 
3614         /*
3615          * When PFM_CTX_MASKED, we cannot issue a restart before the previous 
3616          * one is seen by the task.
3617          */
3618         if (state == PFM_CTX_MASKED) {
3619                 if (ctx->ctx_fl_can_restart == 0) return -EINVAL;
3620                 /*
3621                  * will prevent subsequent restart before this one is
3622                  * seen by other task
3623                  */
3624                 ctx->ctx_fl_can_restart = 0;
3625         }
3626 
3627         /*
3628          * if blocking, then post the semaphore is PFM_CTX_MASKED, i.e.
3629          * the task is blocked or on its way to block. That's the normal
3630          * restart path. If the monitoring is not masked, then the task
3631          * can be actively monitoring and we cannot directly intervene.
3632          * Therefore we use the trap mechanism to catch the task and
3633          * force it to reset the buffer/reset PMDs.
3634          *
3635          * if non-blocking, then we ensure that the task will go into
3636          * pfm_handle_work() before returning to user mode.
3637          *
3638          * We cannot explicitly reset another task, it MUST always
3639          * be done by the task itself. This works for system wide because
3640          * the tool that is controlling the session is logically doing 
3641          * "self-monitoring".
3642          */
3643         if (CTX_OVFL_NOBLOCK(ctx) == 0 && state == PFM_CTX_MASKED) {
3644                 DPRINT(("unblocking [%d]\n", task_pid_nr(task)));
3645                 complete(&ctx->ctx_restart_done);
3646         } else {
3647                 DPRINT(("[%d] armed exit trap\n", task_pid_nr(task)));
3648 
3649                 ctx->ctx_fl_trap_reason = PFM_TRAP_REASON_RESET;
3650 
3651                 PFM_SET_WORK_PENDING(task, 1);
3652 
3653                 set_notify_resume(task);
3654 
3655                 /*
3656                  * XXX: send reschedule if task runs on another CPU
3657                  */
3658         }
3659         return 0;
3660 }
3661 
3662 static int
3663 pfm_debug(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3664 {
3665         unsigned int m = *(unsigned int *)arg;
3666 
3667         pfm_sysctl.debug = m == 0 ? 0 : 1;
3668 
3669         printk(KERN_INFO "perfmon debugging %s (timing reset)\n", pfm_sysctl.debug ? "on" : "off");
3670 
3671         if (m == 0) {
3672                 memset(pfm_stats, 0, sizeof(pfm_stats));
3673                 for(m=0; m < NR_CPUS; m++) pfm_stats[m].pfm_ovfl_intr_cycles_min = ~0UL;
3674         }
3675         return 0;
3676 }
3677 
3678 /*
3679  * arg can be NULL and count can be zero for this function
3680  */
3681 static int
3682 pfm_write_ibr_dbr(int mode, pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3683 {
3684         struct thread_struct *thread = NULL;
3685         struct task_struct *task;
3686         pfarg_dbreg_t *req = (pfarg_dbreg_t *)arg;
3687         unsigned long flags;
3688         dbreg_t dbreg;
3689         unsigned int rnum;
3690         int first_time;
3691         int ret = 0, state;
3692         int i, can_access_pmu = 0;
3693         int is_system, is_loaded;
3694 
3695         if (pmu_conf->use_rr_dbregs == 0) return -EINVAL;
3696 
3697         state     = ctx->ctx_state;
3698         is_loaded = state == PFM_CTX_LOADED ? 1 : 0;
3699         is_system = ctx->ctx_fl_system;
3700         task      = ctx->ctx_task;
3701 
3702         if (state == PFM_CTX_ZOMBIE) return -EINVAL;
3703 
3704         /*
3705          * on both UP and SMP, we can only write to the PMC when the task is
3706          * the owner of the local PMU.
3707          */
3708         if (is_loaded) {
3709                 thread = &task->thread;
3710                 /*
3711                  * In system wide and when the context is loaded, access can only happen
3712                  * when the caller is running on the CPU being monitored by the session.
3713                  * It does not have to be the owner (ctx_task) of the context per se.
3714                  */
3715                 if (unlikely(is_system && ctx->ctx_cpu != smp_processor_id())) {
3716                         DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
3717                         return -EBUSY;
3718                 }
3719                 can_access_pmu = GET_PMU_OWNER() == task || is_system ? 1 : 0;
3720         }
3721 
3722         /*
3723          * we do not need to check for ipsr.db because we do clear ibr.x, dbr.r, and dbr.w
3724          * ensuring that no real breakpoint can be installed via this call.
3725          *
3726          * IMPORTANT: regs can be NULL in this function
3727          */
3728 
3729         first_time = ctx->ctx_fl_using_dbreg == 0;
3730 
3731         /*
3732          * don't bother if we are loaded and task is being debugged
3733          */
3734         if (is_loaded && (thread->flags & IA64_THREAD_DBG_VALID) != 0) {
3735                 DPRINT(("debug registers already in use for [%d]\n", task_pid_nr(task)));
3736                 return -EBUSY;
3737         }
3738 
3739         /*
3740          * check for debug registers in system wide mode
3741          *
3742          * If though a check is done in pfm_context_load(),
3743          * we must repeat it here, in case the registers are
3744          * written after the context is loaded
3745          */
3746         if (is_loaded) {
3747                 LOCK_PFS(flags);
3748 
3749                 if (first_time && is_system) {
3750                         if (pfm_sessions.pfs_ptrace_use_dbregs)
3751                                 ret = -EBUSY;
3752                         else
3753                                 pfm_sessions.pfs_sys_use_dbregs++;
3754                 }
3755                 UNLOCK_PFS(flags);
3756         }
3757 
3758         if (ret != 0) return ret;
3759 
3760         /*
3761          * mark ourself as user of the debug registers for
3762          * perfmon purposes.
3763          */
3764         ctx->ctx_fl_using_dbreg = 1;
3765 
3766         /*
3767          * clear hardware registers to make sure we don't
3768          * pick up stale state.
3769          *
3770          * for a system wide session, we do not use
3771          * thread.dbr, thread.ibr because this process
3772          * never leaves the current CPU and the state
3773          * is shared by all processes running on it
3774          */
3775         if (first_time && can_access_pmu) {
3776                 DPRINT(("[%d] clearing ibrs, dbrs\n", task_pid_nr(task)));
3777                 for (i=0; i < pmu_conf->num_ibrs; i++) {
3778                         ia64_set_ibr(i, 0UL);
3779                         ia64_dv_serialize_instruction();
3780                 }
3781                 ia64_srlz_i();
3782                 for (i=0; i < pmu_conf->num_dbrs; i++) {
3783                         ia64_set_dbr(i, 0UL);
3784                         ia64_dv_serialize_data();
3785                 }
3786                 ia64_srlz_d();
3787         }
3788 
3789         /*
3790          * Now install the values into the registers
3791          */
3792         for (i = 0; i < count; i++, req++) {
3793 
3794                 rnum      = req->dbreg_num;
3795                 dbreg.val = req->dbreg_value;
3796 
3797                 ret = -EINVAL;
3798 
3799                 if ((mode == PFM_CODE_RR && rnum >= PFM_NUM_IBRS) || ((mode == PFM_DATA_RR) && rnum >= PFM_NUM_DBRS)) {
3800                         DPRINT(("invalid register %u val=0x%lx mode=%d i=%d count=%d\n",
3801                                   rnum, dbreg.val, mode, i, count));
3802 
3803                         goto abort_mission;
3804                 }
3805 
3806                 /*
3807                  * make sure we do not install enabled breakpoint
3808                  */
3809                 if (rnum & 0x1) {
3810                         if (mode == PFM_CODE_RR)
3811                                 dbreg.ibr.ibr_x = 0;
3812                         else
3813                                 dbreg.dbr.dbr_r = dbreg.dbr.dbr_w = 0;
3814                 }
3815 
3816                 PFM_REG_RETFLAG_SET(req->dbreg_flags, 0);
3817 
3818                 /*
3819                  * Debug registers, just like PMC, can only be modified
3820                  * by a kernel call. Moreover, perfmon() access to those
3821                  * registers are centralized in this routine. The hardware
3822                  * does not modify the value of these registers, therefore,
3823                  * if we save them as they are written, we can avoid having
3824                  * to save them on context switch out. This is made possible
3825                  * by the fact that when perfmon uses debug registers, ptrace()
3826                  * won't be able to modify them concurrently.
3827                  */
3828                 if (mode == PFM_CODE_RR) {
3829                         CTX_USED_IBR(ctx, rnum);
3830 
3831                         if (can_access_pmu) {
3832                                 ia64_set_ibr(rnum, dbreg.val);
3833                                 ia64_dv_serialize_instruction();
3834                         }
3835 
3836                         ctx->ctx_ibrs[rnum] = dbreg.val;
3837 
3838                         DPRINT(("write ibr%u=0x%lx used_ibrs=0x%x ld=%d apmu=%d\n",
3839                                 rnum, dbreg.val, ctx->ctx_used_ibrs[0], is_loaded, can_access_pmu));
3840                 } else {
3841                         CTX_USED_DBR(ctx, rnum);
3842 
3843                         if (can_access_pmu) {
3844                                 ia64_set_dbr(rnum, dbreg.val);
3845                                 ia64_dv_serialize_data();
3846                         }
3847                         ctx->ctx_dbrs[rnum] = dbreg.val;
3848 
3849                         DPRINT(("write dbr%u=0x%lx used_dbrs=0x%x ld=%d apmu=%d\n",
3850                                 rnum, dbreg.val, ctx->ctx_used_dbrs[0], is_loaded, can_access_pmu));
3851                 }
3852         }
3853 
3854         return 0;
3855 
3856 abort_mission:
3857         /*
3858          * in case it was our first attempt, we undo the global modifications
3859          */
3860         if (first_time) {
3861                 LOCK_PFS(flags);
3862                 if (ctx->ctx_fl_system) {
3863                         pfm_sessions.pfs_sys_use_dbregs--;
3864                 }
3865                 UNLOCK_PFS(flags);
3866                 ctx->ctx_fl_using_dbreg = 0;
3867         }
3868         /*
3869          * install error return flag
3870          */
3871         PFM_REG_RETFLAG_SET(req->dbreg_flags, PFM_REG_RETFL_EINVAL);
3872 
3873         return ret;
3874 }
3875 
3876 static int
3877 pfm_write_ibrs(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3878 {
3879         return pfm_write_ibr_dbr(PFM_CODE_RR, ctx, arg, count, regs);
3880 }
3881 
3882 static int
3883 pfm_write_dbrs(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3884 {
3885         return pfm_write_ibr_dbr(PFM_DATA_RR, ctx, arg, count, regs);
3886 }
3887 
3888 int
3889 pfm_mod_write_ibrs(struct task_struct *task, void *req, unsigned int nreq, struct pt_regs *regs)
3890 {
3891         pfm_context_t *ctx;
3892 
3893         if (req == NULL) return -EINVAL;
3894 
3895         ctx = GET_PMU_CTX();
3896 
3897         if (ctx == NULL) return -EINVAL;
3898 
3899         /*
3900          * for now limit to current task, which is enough when calling
3901          * from overflow handler
3902          */
3903         if (task != current && ctx->ctx_fl_system == 0) return -EBUSY;
3904 
3905         return pfm_write_ibrs(ctx, req, nreq, regs);
3906 }
3907 EXPORT_SYMBOL(pfm_mod_write_ibrs);
3908 
3909 int
3910 pfm_mod_write_dbrs(struct task_struct *task, void *req, unsigned int nreq, struct pt_regs *regs)
3911 {
3912         pfm_context_t *ctx;
3913 
3914         if (req == NULL) return -EINVAL;
3915 
3916         ctx = GET_PMU_CTX();
3917 
3918         if (ctx == NULL) return -EINVAL;
3919 
3920         /*
3921          * for now limit to current task, which is enough when calling
3922          * from overflow handler
3923          */
3924         if (task != current && ctx->ctx_fl_system == 0) return -EBUSY;
3925 
3926         return pfm_write_dbrs(ctx, req, nreq, regs);
3927 }
3928 EXPORT_SYMBOL(pfm_mod_write_dbrs);
3929 
3930 
3931 static int
3932 pfm_get_features(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3933 {
3934         pfarg_features_t *req = (pfarg_features_t *)arg;
3935 
3936         req->ft_version = PFM_VERSION;
3937         return 0;
3938 }
3939 
3940 static int
3941 pfm_stop(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3942 {
3943         struct pt_regs *tregs;
3944         struct task_struct *task = PFM_CTX_TASK(ctx);
3945         int state, is_system;
3946 
3947         state     = ctx->ctx_state;
3948         is_system = ctx->ctx_fl_system;
3949 
3950         /*
3951          * context must be attached to issue the stop command (includes LOADED,MASKED,ZOMBIE)
3952          */
3953         if (state == PFM_CTX_UNLOADED) return -EINVAL;
3954 
3955         /*
3956          * In system wide and when the context is loaded, access can only happen
3957          * when the caller is running on the CPU being monitored by the session.
3958          * It does not have to be the owner (ctx_task) of the context per se.
3959          */
3960         if (is_system && ctx->ctx_cpu != smp_processor_id()) {
3961                 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
3962                 return -EBUSY;
3963         }
3964         DPRINT(("task [%d] ctx_state=%d is_system=%d\n",
3965                 task_pid_nr(PFM_CTX_TASK(ctx)),
3966                 state,
3967                 is_system));
3968         /*
3969          * in system mode, we need to update the PMU directly
3970          * and the user level state of the caller, which may not
3971          * necessarily be the creator of the context.
3972          */
3973         if (is_system) {
3974                 /*
3975                  * Update local PMU first
3976                  *
3977                  * disable dcr pp
3978                  */
3979                 ia64_setreg(_IA64_REG_CR_DCR, ia64_getreg(_IA64_REG_CR_DCR) & ~IA64_DCR_PP);
3980                 ia64_srlz_i();
3981 
3982                 /*
3983                  * update local cpuinfo
3984                  */
3985                 PFM_CPUINFO_CLEAR(PFM_CPUINFO_DCR_PP);
3986 
3987                 /*
3988                  * stop monitoring, does srlz.i
3989                  */
3990                 pfm_clear_psr_pp();
3991 
3992                 /*
3993                  * stop monitoring in the caller
3994                  */
3995                 ia64_psr(regs)->pp = 0;
3996 
3997                 return 0;
3998         }
3999         /*
4000          * per-task mode
4001          */
4002 
4003         if (task == current) {
4004                 /* stop monitoring  at kernel level */
4005                 pfm_clear_psr_up();
4006 
4007                 /*
4008                  * stop monitoring at the user level
4009                  */
4010                 ia64_psr(regs)->up = 0;
4011         } else {
4012                 tregs = task_pt_regs(task);
4013 
4014                 /*
4015                  * stop monitoring at the user level
4016                  */
4017                 ia64_psr(tregs)->up = 0;
4018 
4019                 /*
4020                  * monitoring disabled in kernel at next reschedule
4021                  */
4022                 ctx->ctx_saved_psr_up = 0;
4023                 DPRINT(("task=[%d]\n", task_pid_nr(task)));
4024         }
4025         return 0;
4026 }
4027 
4028 
4029 static int
4030 pfm_start(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
4031 {
4032         struct pt_regs *tregs;
4033         int state, is_system;
4034 
4035         state     = ctx->ctx_state;
4036         is_system = ctx->ctx_fl_system;
4037 
4038         if (state != PFM_CTX_LOADED) return -EINVAL;
4039 
4040         /*
4041          * In system wide and when the context is loaded, access can only happen
4042          * when the caller is running on the CPU being monitored by the session.
4043          * It does not have to be the owner (ctx_task) of the context per se.
4044          */
4045         if (is_system && ctx->ctx_cpu != smp_processor_id()) {
4046                 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
4047                 return -EBUSY;
4048         }
4049 
4050         /*
4051          * in system mode, we need to update the PMU directly
4052          * and the user level state of the caller, which may not
4053          * necessarily be the creator of the context.
4054          */
4055         if (is_system) {
4056 
4057                 /*
4058                  * set user level psr.pp for the caller
4059                  */
4060                 ia64_psr(regs)->pp = 1;
4061 
4062                 /*
4063                  * now update the local PMU and cpuinfo
4064                  */
4065                 PFM_CPUINFO_SET(PFM_CPUINFO_DCR_PP);
4066 
4067                 /*
4068                  * start monitoring at kernel level
4069                  */
4070                 pfm_set_psr_pp();
4071 
4072                 /* enable dcr pp */
4073                 ia64_setreg(_IA64_REG_CR_DCR, ia64_getreg(_IA64_REG_CR_DCR) | IA64_DCR_PP);
4074                 ia64_srlz_i();
4075 
4076                 return 0;
4077         }
4078 
4079         /*
4080          * per-process mode
4081          */
4082 
4083         if (ctx->ctx_task == current) {
4084 
4085                 /* start monitoring at kernel level */
4086                 pfm_set_psr_up();
4087 
4088                 /*
4089                  * activate monitoring at user level
4090                  */
4091                 ia64_psr(regs)->up = 1;
4092 
4093         } else {
4094                 tregs = task_pt_regs(ctx->ctx_task);
4095 
4096                 /*
4097                  * start monitoring at the kernel level the next
4098                  * time the task is scheduled
4099                  */
4100                 ctx->ctx_saved_psr_up = IA64_PSR_UP;
4101 
4102                 /*
4103                  * activate monitoring at user level
4104                  */
4105                 ia64_psr(tregs)->up = 1;
4106         }
4107         return 0;
4108 }
4109 
4110 static int
4111 pfm_get_pmc_reset(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
4112 {
4113         pfarg_reg_t *req = (pfarg_reg_t *)arg;
4114         unsigned int cnum;
4115         int i;
4116         int ret = -EINVAL;
4117 
4118         for (i = 0; i < count; i++, req++) {
4119 
4120                 cnum = req->reg_num;
4121 
4122                 if (!PMC_IS_IMPL(cnum)) goto abort_mission;
4123 
4124                 req->reg_value = PMC_DFL_VAL(cnum);
4125 
4126                 PFM_REG_RETFLAG_SET(req->reg_flags, 0);
4127 
4128                 DPRINT(("pmc_reset_val pmc[%u]=0x%lx\n", cnum, req->reg_value));
4129         }
4130         return 0;
4131 
4132 abort_mission:
4133         PFM_REG_RETFLAG_SET(req->reg_flags, PFM_REG_RETFL_EINVAL);
4134         return ret;
4135 }
4136 
4137 static int
4138 pfm_check_task_exist(pfm_context_t *ctx)
4139 {
4140         struct task_struct *g, *t;
4141         int ret = -ESRCH;
4142 
4143         read_lock(&tasklist_lock);
4144 
4145         do_each_thread (g, t) {
4146                 if (t->thread.pfm_context == ctx) {
4147                         ret = 0;
4148                         goto out;
4149                 }
4150         } while_each_thread (g, t);
4151 out:
4152         read_unlock(&tasklist_lock);
4153 
4154         DPRINT(("pfm_check_task_exist: ret=%d ctx=%p\n", ret, ctx));
4155 
4156         return ret;
4157 }
4158 
4159 static int
4160 pfm_context_load(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
4161 {
4162         struct task_struct *task;
4163         struct thread_struct *thread;
4164         struct pfm_context_t *old;
4165         unsigned long flags;
4166 #ifndef CONFIG_SMP
4167         struct task_struct *owner_task = NULL;
4168 #endif
4169         pfarg_load_t *req = (pfarg_load_t *)arg;
4170         unsigned long *pmcs_source, *pmds_source;
4171         int the_cpu;
4172         int ret = 0;
4173         int state, is_system, set_dbregs = 0;
4174 
4175         state     = ctx->ctx_state;
4176         is_system = ctx->ctx_fl_system;
4177         /*
4178          * can only load from unloaded or terminated state
4179          */
4180         if (state != PFM_CTX_UNLOADED) {
4181                 DPRINT(("cannot load to [%d], invalid ctx_state=%d\n",
4182                         req->load_pid,
4183                         ctx->ctx_state));
4184                 return -EBUSY;
4185         }
4186 
4187         DPRINT(("load_pid [%d] using_dbreg=%d\n", req->load_pid, ctx->ctx_fl_using_dbreg));
4188 
4189         if (CTX_OVFL_NOBLOCK(ctx) == 0 && req->load_pid == current->pid) {
4190                 DPRINT(("cannot use blocking mode on self\n"));
4191                 return -EINVAL;
4192         }
4193 
4194         ret = pfm_get_task(ctx, req->load_pid, &task);
4195         if (ret) {
4196                 DPRINT(("load_pid [%d] get_task=%d\n", req->load_pid, ret));
4197                 return ret;
4198         }
4199 
4200         ret = -EINVAL;
4201 
4202         /*
4203          * system wide is self monitoring only
4204          */
4205         if (is_system && task != current) {
4206                 DPRINT(("system wide is self monitoring only load_pid=%d\n",
4207                         req->load_pid));
4208                 goto error;
4209         }
4210 
4211         thread = &task->thread;
4212 
4213         ret = 0;
4214         /*
4215          * cannot load a context which is using range restrictions,
4216          * into a task that is being debugged.
4217          */
4218         if (ctx->ctx_fl_using_dbreg) {
4219                 if (thread->flags & IA64_THREAD_DBG_VALID) {
4220                         ret = -EBUSY;
4221                         DPRINT(("load_pid [%d] task is debugged, cannot load range restrictions\n", req->load_pid));
4222                         goto error;
4223                 }
4224                 LOCK_PFS(flags);
4225 
4226                 if (is_system) {
4227                         if (pfm_sessions.pfs_ptrace_use_dbregs) {
4228                                 DPRINT(("cannot load [%d] dbregs in use\n",
4229                                                         task_pid_nr(task)));
4230                                 ret = -EBUSY;
4231                         } else {
4232                                 pfm_sessions.pfs_sys_use_dbregs++;
4233                                 DPRINT(("load [%d] increased sys_use_dbreg=%u\n", task_pid_nr(task), pfm_sessions.pfs_sys_use_dbregs));
4234                                 set_dbregs = 1;
4235                         }
4236                 }
4237 
4238                 UNLOCK_PFS(flags);
4239 
4240                 if (ret) goto error;
4241         }
4242 
4243         /*
4244          * SMP system-wide monitoring implies self-monitoring.
4245          *
4246          * The programming model expects the task to
4247          * be pinned on a CPU throughout the session.
4248          * Here we take note of the current CPU at the
4249          * time the context is loaded. No call from
4250          * another CPU will be allowed.
4251          *
4252          * The pinning via shed_setaffinity()
4253          * must be done by the calling task prior
4254          * to this call.
4255          *
4256          * systemwide: keep track of CPU this session is supposed to run on
4257          */
4258         the_cpu = ctx->ctx_cpu = smp_processor_id();
4259 
4260         ret = -EBUSY;
4261         /*
4262          * now reserve the session
4263          */
4264         ret = pfm_reserve_session(current, is_system, the_cpu);
4265         if (ret) goto error;
4266 
4267         /*
4268          * task is necessarily stopped at this point.
4269          *
4270          * If the previous context was zombie, then it got removed in
4271          * pfm_save_regs(). Therefore we should not see it here.
4272          * If we see a context, then this is an active context
4273          *
4274          * XXX: needs to be atomic
4275          */
4276         DPRINT(("before cmpxchg() old_ctx=%p new_ctx=%p\n",
4277                 thread->pfm_context, ctx));
4278 
4279         ret = -EBUSY;
4280         old = ia64_cmpxchg(acq, &thread->pfm_context, NULL, ctx, sizeof(pfm_context_t *));
4281         if (old != NULL) {
4282                 DPRINT(("load_pid [%d] already has a context\n", req->load_pid));
4283                 goto error_unres;
4284         }
4285 
4286         pfm_reset_msgq(ctx);
4287 
4288         ctx->ctx_state = PFM_CTX_LOADED;
4289 
4290         /*
4291          * link context to task
4292          */
4293         ctx->ctx_task = task;
4294 
4295         if (is_system) {
4296                 /*
4297                  * we load as stopped
4298                  */
4299                 PFM_CPUINFO_SET(PFM_CPUINFO_SYST_WIDE);
4300                 PFM_CPUINFO_CLEAR(PFM_CPUINFO_DCR_PP);
4301 
4302                 if (ctx->ctx_fl_excl_idle) PFM_CPUINFO_SET(PFM_CPUINFO_EXCL_IDLE);
4303         } else {
4304                 thread->flags |= IA64_THREAD_PM_VALID;
4305         }
4306 
4307         /*
4308          * propagate into thread-state
4309          */
4310         pfm_copy_pmds(task, ctx);
4311         pfm_copy_pmcs(task, ctx);
4312 
4313         pmcs_source = ctx->th_pmcs;
4314         pmds_source = ctx->th_pmds;
4315 
4316         /*
4317          * always the case for system-wide
4318          */
4319         if (task == current) {
4320 
4321                 if (is_system == 0) {
4322 
4323                         /* allow user level control */
4324                         ia64_psr(regs)->sp = 0;
4325                         DPRINT(("clearing psr.sp for [%d]\n", task_pid_nr(task)));
4326 
4327                         SET_LAST_CPU(ctx, smp_processor_id());
4328                         INC_ACTIVATION();
4329                         SET_ACTIVATION(ctx);
4330 #ifndef CONFIG_SMP
4331                         /*
4332                          * push the other task out, if any
4333                          */
4334                         owner_task = GET_PMU_OWNER();
4335                         if (owner_task) pfm_lazy_save_regs(owner_task);
4336 #endif
4337                 }
4338                 /*
4339                  * load all PMD from ctx to PMU (as opposed to thread state)
4340                  * restore all PMC from ctx to PMU
4341                  */
4342                 pfm_restore_pmds(pmds_source, ctx->ctx_all_pmds[0]);
4343                 pfm_restore_pmcs(pmcs_source, ctx->ctx_all_pmcs[0]);
4344 
4345                 ctx->ctx_reload_pmcs[0] = 0UL;
4346                 ctx->ctx_reload_pmds[0] = 0UL;
4347 
4348                 /*
4349                  * guaranteed safe by earlier check against DBG_VALID
4350                  */
4351                 if (ctx->ctx_fl_using_dbreg) {
4352                         pfm_restore_ibrs(ctx->ctx_ibrs, pmu_conf->num_ibrs);
4353                         pfm_restore_dbrs(ctx->ctx_dbrs, pmu_conf->num_dbrs);
4354                 }
4355                 /*
4356                  * set new ownership
4357                  */
4358                 SET_PMU_OWNER(task, ctx);
4359 
4360                 DPRINT(("context loaded on PMU for [%d]\n", task_pid_nr(task)));
4361         } else {
4362                 /*
4363                  * when not current, task MUST be stopped, so this is safe
4364                  */
4365                 regs = task_pt_regs(task);
4366 
4367                 /* force a full reload */
4368                 ctx->ctx_last_activation = PFM_INVALID_ACTIVATION;
4369                 SET_LAST_CPU(ctx, -1);
4370 
4371                 /* initial saved psr (stopped) */
4372                 ctx->ctx_saved_psr_up = 0UL;
4373                 ia64_psr(regs)->up = ia64_psr(regs)->pp = 0;
4374         }
4375 
4376         ret = 0;
4377 
4378 error_unres:
4379         if (ret) pfm_unreserve_session(ctx, ctx->ctx_fl_system, the_cpu);
4380 error:
4381         /*
4382          * we must undo the dbregs setting (for system-wide)
4383          */
4384         if (ret && set_dbregs) {
4385                 LOCK_PFS(flags);
4386                 pfm_sessions.pfs_sys_use_dbregs--;
4387                 UNLOCK_PFS(flags);
4388         }
4389         /*
4390          * release task, there is now a link with the context
4391          */
4392         if (is_system == 0 && task != current) {
4393                 pfm_put_task(task);
4394 
4395                 if (ret == 0) {
4396                         ret = pfm_check_task_exist(ctx);
4397                         if (ret) {
4398                                 ctx->ctx_state = PFM_CTX_UNLOADED;
4399                                 ctx->ctx_task  = NULL;
4400                         }
4401                 }
4402         }
4403         return ret;
4404 }
4405 
4406 /*
4407  * in this function, we do not need to increase the use count
4408  * for the task via get_task_struct(), because we hold the
4409  * context lock. If the task were to disappear while having
4410  * a context attached, it would go through pfm_exit_thread()
4411  * which also grabs the context lock  and would therefore be blocked
4412  * until we are here.
4413  */
4414 static void pfm_flush_pmds(struct task_struct *, pfm_context_t *ctx);
4415 
4416 static int
4417 pfm_context_unload(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
4418 {
4419         struct task_struct *task = PFM_CTX_TASK(ctx);
4420         struct pt_regs *tregs;
4421         int prev_state, is_system;
4422         int ret;
4423 
4424         DPRINT(("ctx_state=%d task [%d]\n", ctx->ctx_state, task ? task_pid_nr(task) : -1));
4425 
4426         prev_state = ctx->ctx_state;
4427         is_system  = ctx->ctx_fl_system;
4428 
4429         /*
4430          * unload only when necessary
4431          */
4432         if (prev_state == PFM_CTX_UNLOADED) {
4433                 DPRINT(("ctx_state=%d, nothing to do\n", prev_state));
4434                 return 0;
4435         }
4436 
4437         /*
4438          * clear psr and dcr bits
4439          */
4440         ret = pfm_stop(ctx, NULL, 0, regs);
4441         if (ret) return ret;
4442 
4443         ctx->ctx_state = PFM_CTX_UNLOADED;
4444 
4445         /*
4446          * in system mode, we need to update the PMU directly
4447          * and the user level state of the caller, which may not
4448          * necessarily be the creator of the context.
4449          */
4450         if (is_system) {
4451 
4452                 /*
4453                  * Update cpuinfo
4454                  *
4455                  * local PMU is taken care of in pfm_stop()
4456                  */
4457                 PFM_CPUINFO_CLEAR(PFM_CPUINFO_SYST_WIDE);
4458                 PFM_CPUINFO_CLEAR(PFM_CPUINFO_EXCL_IDLE);
4459 
4460                 /*
4461                  * save PMDs in context
4462                  * release ownership
4463                  */
4464                 pfm_flush_pmds(current, ctx);
4465 
4466                 /*
4467                  * at this point we are done with the PMU
4468                  * so we can unreserve the resource.
4469                  */
4470                 if (prev_state != PFM_CTX_ZOMBIE) 
4471                         pfm_unreserve_session(ctx, 1 , ctx->ctx_cpu);
4472 
4473                 /*
4474                  * disconnect context from task
4475                  */
4476                 task->thread.pfm_context = NULL;
4477                 /*
4478                  * disconnect task from context
4479                  */
4480                 ctx->ctx_task = NULL;
4481 
4482                 /*
4483                  * There is nothing more to cleanup here.
4484                  */
4485                 return 0;
4486         }
4487 
4488         /*
4489          * per-task mode
4490          */
4491         tregs = task == current ? regs : task_pt_regs(task);
4492 
4493         if (task == current) {
4494                 /*
4495                  * cancel user level control
4496                  */
4497                 ia64_psr(regs)->sp = 1;
4498 
4499                 DPRINT(("setting psr.sp for [%d]\n", task_pid_nr(task)));
4500         }
4501         /*
4502          * save PMDs to context
4503          * release ownership
4504          */
4505         pfm_flush_pmds(task, ctx);
4506 
4507         /*
4508          * at this point we are done with the PMU
4509          * so we can unreserve the resource.
4510          *
4511          * when state was ZOMBIE, we have already unreserved.
4512          */
4513         if (prev_state != PFM_CTX_ZOMBIE) 
4514                 pfm_unreserve_session(ctx, 0 , ctx->ctx_cpu);
4515 
4516         /*
4517          * reset activation counter and psr
4518          */
4519         ctx->ctx_last_activation = PFM_INVALID_ACTIVATION;
4520         SET_LAST_CPU(ctx, -1);
4521 
4522         /*
4523          * PMU state will not be restored
4524          */
4525         task->thread.flags &= ~IA64_THREAD_PM_VALID;
4526 
4527         /*
4528          * break links between context and task
4529          */
4530         task->thread.pfm_context  = NULL;
4531         ctx->ctx_task             = NULL;
4532 
4533         PFM_SET_WORK_PENDING(task, 0);
4534 
4535         ctx->ctx_fl_trap_reason  = PFM_TRAP_REASON_NONE;
4536         ctx->ctx_fl_can_restart  = 0;
4537         ctx->ctx_fl_going_zombie = 0;
4538 
4539         DPRINT(("disconnected [%d] from context\n", task_pid_nr(task)));
4540 
4541         return 0;
4542 }
4543 
4544 
4545 /*
4546  * called only from exit_thread(): task == current
4547  * we come here only if current has a context attached (loaded or masked)
4548  */
4549 void
4550 pfm_exit_thread(struct task_struct *task)
4551 {
4552         pfm_context_t *ctx;
4553         unsigned long flags;
4554         struct pt_regs *regs = task_pt_regs(task);
4555         int ret, state;
4556         int free_ok = 0;
4557 
4558         ctx = PFM_GET_CTX(task);
4559 
4560         PROTECT_CTX(ctx, flags);
4561 
4562         DPRINT(("state=%d task [%d]\n", ctx->ctx_state, task_pid_nr(task)));
4563 
4564         state = ctx->ctx_state;
4565         switch(state) {
4566                 case PFM_CTX_UNLOADED:
4567                         /*
4568                          * only comes to this function if pfm_context is not NULL, i.e., cannot
4569                          * be in unloaded state
4570                          */
4571                         printk(KERN_ERR "perfmon: pfm_exit_thread [%d] ctx unloaded\n", task_pid_nr(task));
4572                         break;
4573                 case PFM_CTX_LOADED:
4574                 case PFM_CTX_MASKED:
4575                         ret = pfm_context_unload(ctx, NULL, 0, regs);
4576                         if (ret) {
4577                                 printk(KERN_ERR "perfmon: pfm_exit_thread [%d] state=%d unload failed %d\n", task_pid_nr(task), state, ret);
4578                         }
4579                         DPRINT(("ctx unloaded for current state was %d\n", state));
4580 
4581                         pfm_end_notify_user(ctx);
4582                         break;
4583                 case PFM_CTX_ZOMBIE:
4584                         ret = pfm_context_unload(ctx, NULL, 0, regs);
4585                         if (ret) {
4586                                 printk(KERN_ERR "perfmon: pfm_exit_thread [%d] state=%d unload failed %d\n", task_pid_nr(task), state, ret);
4587                         }
4588                         free_ok = 1;
4589                         break;
4590                 default:
4591                         printk(KERN_ERR "perfmon: pfm_exit_thread [%d] unexpected state=%d\n", task_pid_nr(task), state);
4592                         break;
4593         }
4594         UNPROTECT_CTX(ctx, flags);
4595 
4596         { u64 psr = pfm_get_psr();
4597           BUG_ON(psr & (IA64_PSR_UP|IA64_PSR_PP));
4598           BUG_ON(GET_PMU_OWNER());
4599           BUG_ON(ia64_psr(regs)->up);
4600           BUG_ON(ia64_psr(regs)->pp);
4601         }
4602 
4603         /*
4604          * All memory free operations (especially for vmalloc'ed memory)
4605          * MUST be done with interrupts ENABLED.
4606          */
4607         if (free_ok) pfm_context_free(ctx);
4608 }
4609 
4610 /*
4611  * functions MUST be listed in the increasing order of their index (see permfon.h)
4612  */
4613 #define PFM_CMD(name, flags, arg_count, arg_type, getsz) { name, #name, flags, arg_count, sizeof(arg_type), getsz }
4614 #define PFM_CMD_S(name, flags) { name, #name, flags, 0, 0, NULL }
4615 #define PFM_CMD_PCLRWS  (PFM_CMD_FD|PFM_CMD_ARG_RW|PFM_CMD_STOP)
4616 #define PFM_CMD_PCLRW   (PFM_CMD_FD|PFM_CMD_ARG_RW)
4617 #define PFM_CMD_NONE    { NULL, "no-cmd", 0, 0, 0, NULL}
4618 
4619 static pfm_cmd_desc_t pfm_cmd_tab[]={
4620 /* 0  */PFM_CMD_NONE,
4621 /* 1  */PFM_CMD(pfm_write_pmcs, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_reg_t, NULL),
4622 /* 2  */PFM_CMD(pfm_write_pmds, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_reg_t, NULL),
4623 /* 3  */PFM_CMD(pfm_read_pmds, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_reg_t, NULL),
4624 /* 4  */PFM_CMD_S(pfm_stop, PFM_CMD_PCLRWS),
4625 /* 5  */PFM_CMD_S(pfm_start, PFM_CMD_PCLRWS),
4626 /* 6  */PFM_CMD_NONE,
4627 /* 7  */PFM_CMD_NONE,
4628 /* 8  */PFM_CMD(pfm_context_create, PFM_CMD_ARG_RW, 1, pfarg_context_t, pfm_ctx_getsize),
4629 /* 9  */PFM_CMD_NONE,
4630 /* 10 */PFM_CMD_S(pfm_restart, PFM_CMD_PCLRW),
4631 /* 11 */PFM_CMD_NONE,
4632 /* 12 */PFM_CMD(pfm_get_features, PFM_CMD_ARG_RW, 1, pfarg_features_t, NULL),
4633 /* 13 */PFM_CMD(pfm_debug, 0, 1, unsigned int, NULL),
4634 /* 14 */PFM_CMD_NONE,
4635 /* 15 */PFM_CMD(pfm_get_pmc_reset, PFM_CMD_ARG_RW, PFM_CMD_ARG_MANY, pfarg_reg_t, NULL),
4636 /* 16 */PFM_CMD(pfm_context_load, PFM_CMD_PCLRWS, 1, pfarg_load_t, NULL),
4637 /* 17 */PFM_CMD_S(pfm_context_unload, PFM_CMD_PCLRWS),
4638 /* 18 */PFM_CMD_NONE,
4639 /* 19 */PFM_CMD_NONE,
4640 /* 20 */PFM_CMD_NONE,
4641 /* 21 */PFM_CMD_NONE,
4642 /* 22 */PFM_CMD_NONE,
4643 /* 23 */PFM_CMD_NONE,
4644 /* 24 */PFM_CMD_NONE,
4645 /* 25 */PFM_CMD_NONE,
4646 /* 26 */PFM_CMD_NONE,
4647 /* 27 */PFM_CMD_NONE,
4648 /* 28 */PFM_CMD_NONE,
4649 /* 29 */PFM_CMD_NONE,
4650 /* 30 */PFM_CMD_NONE,
4651 /* 31 */PFM_CMD_NONE,
4652 /* 32 */PFM_CMD(pfm_write_ibrs, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_dbreg_t, NULL),
4653 /* 33 */PFM_CMD(pfm_write_dbrs, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_dbreg_t, NULL)
4654 };
4655 #define PFM_CMD_COUNT   (sizeof(pfm_cmd_tab)/sizeof(pfm_cmd_desc_t))
4656 
4657 static int
4658 pfm_check_task_state(pfm_context_t *ctx, int cmd, unsigned long flags)
4659 {
4660         struct task_struct *task;
4661         int state, old_state;
4662 
4663 recheck:
4664         state = ctx->ctx_state;
4665         task  = ctx->ctx_task;
4666 
4667         if (task == NULL) {
4668                 DPRINT(("context %d no task, state=%d\n", ctx->ctx_fd, state));
4669                 return 0;
4670         }
4671 
4672         DPRINT(("context %d state=%d [%d] task_state=%ld must_stop=%d\n",
4673                 ctx->ctx_fd,
4674                 state,
4675                 task_pid_nr(task),
4676                 task->state, PFM_CMD_STOPPED(cmd)));
4677 
4678         /*
4679          * self-monitoring always ok.
4680          *
4681          * for system-wide the caller can either be the creator of the
4682          * context (to one to which the context is attached to) OR
4683          * a task running on the same CPU as the session.
4684          */
4685         if (task == current || ctx->ctx_fl_system) return 0;
4686 
4687         /*
4688          * we are monitoring another thread
4689          */
4690         switch(state) {
4691                 case PFM_CTX_UNLOADED:
4692                         /*
4693                          * if context is UNLOADED we are safe to go
4694                          */
4695                         return 0;
4696                 case PFM_CTX_ZOMBIE:
4697                         /*
4698                          * no command can operate on a zombie context
4699                          */
4700                         DPRINT(("cmd %d state zombie cannot operate on context\n", cmd));
4701                         return -EINVAL;
4702                 case PFM_CTX_MASKED:
4703                         /*
4704                          * PMU state has been saved to software even though
4705                          * the thread may still be running.
4706                          */
4707                         if (cmd != PFM_UNLOAD_CONTEXT) return 0;
4708         }
4709 
4710         /*
4711          * context is LOADED or MASKED. Some commands may need to have 
4712          * the task stopped.
4713          *
4714          * We could lift this restriction for UP but it would mean that
4715          * the user has no guarantee the task would not run between
4716          * two successive calls to perfmonctl(). That's probably OK.
4717          * If this user wants to ensure the task does not run, then
4718          * the task must be stopped.
4719          */
4720         if (PFM_CMD_STOPPED(cmd)) {
4721                 if (!task_is_stopped_or_traced(task)) {
4722                         DPRINT(("[%d] task not in stopped state\n", task_pid_nr(task)));
4723                         return -EBUSY;
4724                 }
4725                 /*
4726                  * task is now stopped, wait for ctxsw out
4727                  *
4728                  * This is an interesting point in the code.
4729                  * We need to unprotect the context because
4730                  * the pfm_save_regs() routines needs to grab
4731                  * the same lock. There are danger in doing
4732                  * this because it leaves a window open for
4733                  * another task to get access to the context
4734                  * and possibly change its state. The one thing
4735                  * that is not possible is for the context to disappear
4736                  * because we are protected by the VFS layer, i.e.,
4737                  * get_fd()/put_fd().
4738                  */
4739                 old_state = state;
4740 
4741                 UNPROTECT_CTX(ctx, flags);
4742 
4743                 wait_task_inactive(task, 0);
4744 
4745                 PROTECT_CTX(ctx, flags);
4746 
4747                 /*
4748                  * we must recheck to verify if state has changed
4749                  */
4750                 if (ctx->ctx_state != old_state) {
4751                         DPRINT(("old_state=%d new_state=%d\n", old_state, ctx->ctx_state));
4752                         goto recheck;
4753                 }
4754         }
4755         return 0;
4756 }
4757 
4758 /*
4759  * system-call entry point (must return long)
4760  */
4761 asmlinkage long
4762 sys_perfmonctl (int fd, int cmd, void __user *arg, int count)
4763 {
4764         struct fd f = {NULL, 0};
4765         pfm_context_t *ctx = NULL;
4766         unsigned long flags = 0UL;
4767         void *args_k = NULL;
4768         long ret; /* will expand int return types */
4769         size_t base_sz, sz, xtra_sz = 0;
4770         int narg, completed_args = 0, call_made = 0, cmd_flags;
4771         int (*func)(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs);
4772         int (*getsize)(void *arg, size_t *sz);
4773 #define PFM_MAX_ARGSIZE 4096
4774 
4775         /*
4776          * reject any call if perfmon was disabled at initialization
4777          */
4778         if (unlikely(pmu_conf == NULL)) return -ENOSYS;
4779 
4780         if (unlikely(cmd < 0 || cmd >= PFM_CMD_COUNT)) {
4781                 DPRINT(("invalid cmd=%d\n", cmd));
4782                 return -EINVAL;
4783         }
4784 
4785         func      = pfm_cmd_tab[cmd].cmd_func;
4786         narg      = pfm_cmd_tab[cmd].cmd_narg;
4787         base_sz   = pfm_cmd_tab[cmd].cmd_argsize;
4788         getsize   = pfm_cmd_tab[cmd].cmd_getsize;
4789         cmd_flags = pfm_cmd_tab[cmd].cmd_flags;
4790 
4791         if (unlikely(func == NULL)) {
4792                 DPRINT(("invalid cmd=%d\n", cmd));
4793                 return -EINVAL;
4794         }
4795 
4796         DPRINT(("cmd=%s idx=%d narg=0x%x argsz=%lu count=%d\n",
4797                 PFM_CMD_NAME(cmd),
4798                 cmd,
4799                 narg,
4800                 base_sz,
4801                 count));
4802 
4803         /*
4804          * check if number of arguments matches what the command expects
4805          */
4806         if (unlikely((narg == PFM_CMD_ARG_MANY && count <= 0) || (narg > 0 && narg != count)))
4807                 return -EINVAL;
4808 
4809 restart_args:
4810         sz = xtra_sz + base_sz*count;
4811         /*
4812          * limit abuse to min page size
4813          */
4814         if (unlikely(sz > PFM_MAX_ARGSIZE)) {
4815                 printk(KERN_ERR "perfmon: [%d] argument too big %lu\n", task_pid_nr(current), sz);
4816                 return -E2BIG;
4817         }
4818 
4819         /*
4820          * allocate default-sized argument buffer
4821          */
4822         if (likely(count && args_k == NULL)) {
4823                 args_k = kmalloc(PFM_MAX_ARGSIZE, GFP_KERNEL);
4824                 if (args_k == NULL) return -ENOMEM;
4825         }
4826 
4827         ret = -EFAULT;
4828 
4829         /*
4830          * copy arguments
4831          *
4832          * assume sz = 0 for command without parameters
4833          */
4834         if (sz && copy_from_user(args_k, arg, sz)) {
4835                 DPRINT(("cannot copy_from_user %lu bytes @%p\n", sz, arg));
4836                 goto error_args;
4837         }
4838 
4839         /*
4840          * check if command supports extra parameters
4841          */
4842         if (completed_args == 0 && getsize) {
4843                 /*
4844                  * get extra parameters size (based on main argument)
4845                  */
4846                 ret = (*getsize)(args_k, &xtra_sz);
4847                 if (ret) goto error_args;
4848 
4849                 completed_args = 1;
4850 
4851                 DPRINT(("restart_args sz=%lu xtra_sz=%lu\n", sz, xtra_sz));
4852 
4853                 /* retry if necessary */
4854                 if (likely(xtra_sz)) goto restart_args;
4855         }
4856 
4857         if (unlikely((cmd_flags & PFM_CMD_FD) == 0)) goto skip_fd;
4858 
4859         ret = -EBADF;
4860 
4861         f = fdget(fd);
4862         if (unlikely(f.file == NULL)) {
4863                 DPRINT(("invalid fd %d\n", fd));
4864                 goto error_args;
4865         }
4866         if (unlikely(PFM_IS_FILE(f.file) == 0)) {
4867                 DPRINT(("fd %d not related to perfmon\n", fd));
4868                 goto error_args;
4869         }
4870 
4871         ctx = f.file->private_data;
4872         if (unlikely(ctx == NULL)) {
4873                 DPRINT(("no context for fd %d\n", fd));
4874                 goto error_args;
4875         }
4876         prefetch(&ctx->ctx_state);
4877 
4878         PROTECT_CTX(ctx, flags);
4879 
4880         /*
4881          * check task is stopped
4882          */
4883         ret = pfm_check_task_state(ctx, cmd, flags);
4884         if (unlikely(ret)) goto abort_locked;
4885 
4886 skip_fd:
4887         ret = (*func)(ctx, args_k, count, task_pt_regs(current));
4888 
4889         call_made = 1;
4890 
4891 abort_locked:
4892         if (likely(ctx)) {
4893                 DPRINT(("context unlocked\n"));
4894                 UNPROTECT_CTX(ctx, flags);
4895         }
4896 
4897         /* copy argument back to user, if needed */
4898         if (call_made && PFM_CMD_RW_ARG(cmd) && copy_to_user(arg, args_k, base_sz*count)) ret = -EFAULT;
4899 
4900 error_args:
4901         if (f.file)
4902                 fdput(f);
4903 
4904         kfree(args_k);
4905 
4906         DPRINT(("cmd=%s ret=%ld\n", PFM_CMD_NAME(cmd), ret));
4907 
4908         return ret;
4909 }
4910 
4911 static void
4912 pfm_resume_after_ovfl(pfm_context_t *ctx, unsigned long ovfl_regs, struct pt_regs *regs)
4913 {
4914         pfm_buffer_fmt_t *fmt = ctx->ctx_buf_fmt;
4915         pfm_ovfl_ctrl_t rst_ctrl;
4916         int state;
4917         int ret = 0;
4918 
4919         state = ctx->ctx_state;
4920         /*
4921          * Unlock sampling buffer and reset index atomically
4922          * XXX: not really needed when blocking
4923          */
4924         if (CTX_HAS_SMPL(ctx)) {
4925 
4926                 rst_ctrl.bits.mask_monitoring = 0;
4927                 rst_ctrl.bits.reset_ovfl_pmds = 0;
4928 
4929                 if (state == PFM_CTX_LOADED)
4930                         ret = pfm_buf_fmt_restart_active(fmt, current, &rst_ctrl, ctx->ctx_smpl_hdr, regs);
4931                 else
4932                         ret = pfm_buf_fmt_restart(fmt, current, &rst_ctrl, ctx->ctx_smpl_hdr, regs);
4933         } else {
4934                 rst_ctrl.bits.mask_monitoring = 0;
4935                 rst_ctrl.bits.reset_ovfl_pmds = 1;
4936         }
4937 
4938         if (ret == 0) {
4939                 if (rst_ctrl.bits.reset_ovfl_pmds) {
4940                         pfm_reset_regs(ctx, &ovfl_regs, PFM_PMD_LONG_RESET);
4941                 }
4942                 if (rst_ctrl.bits.mask_monitoring == 0) {
4943                         DPRINT(("resuming monitoring\n"));
4944                         if (ctx->ctx_state == PFM_CTX_MASKED) pfm_restore_monitoring(current);
4945                 } else {
4946                         DPRINT(("stopping monitoring\n"));
4947                         //pfm_stop_monitoring(current, regs);
4948                 }
4949                 ctx->ctx_state = PFM_CTX_LOADED;
4950         }
4951 }
4952 
4953 /*
4954  * context MUST BE LOCKED when calling
4955  * can only be called for current
4956  */
4957 static void
4958 pfm_context_force_terminate(pfm_context_t *ctx, struct pt_regs *regs)
4959 {
4960         int ret;
4961 
4962         DPRINT(("entering for [%d]\n", task_pid_nr(current)));
4963 
4964         ret = pfm_context_unload(ctx, NULL, 0, regs);
4965         if (ret) {
4966                 printk(KERN_ERR "pfm_context_force_terminate: [%d] unloaded failed with %d\n", task_pid_nr(current), ret);
4967         }
4968 
4969         /*
4970          * and wakeup controlling task, indicating we are now disconnected
4971          */
4972         wake_up_interruptible(&ctx->ctx_zombieq);
4973 
4974         /*
4975          * given that context is still locked, the controlling
4976          * task will only get access when we return from
4977          * pfm_handle_work().
4978          */
4979 }
4980 
4981 static int pfm_ovfl_notify_user(pfm_context_t *ctx, unsigned long ovfl_pmds);
4982 
4983  /*
4984   * pfm_handle_work() can be called with interrupts enabled
4985   * (TIF_NEED_RESCHED) or disabled. The down_interruptible
4986   * call may sleep, therefore we must re-enable interrupts
4987   * to avoid deadlocks. It is safe to do so because this function
4988   * is called ONLY when returning to user level (pUStk=1), in which case
4989   * there is no risk of kernel stack overflow due to deep
4990   * interrupt nesting.
4991   */
4992 void
4993 pfm_handle_work(void)
4994 {
4995         pfm_context_t *ctx;
4996         struct pt_regs *regs;
4997         unsigned long flags, dummy_flags;
4998         unsigned long ovfl_regs;
4999         unsigned int reason;
5000         int ret;
5001 
5002         ctx = PFM_GET_CTX(current);
5003         if (ctx == NULL) {
5004                 printk(KERN_ERR "perfmon: [%d] has no PFM context\n",
5005                         task_pid_nr(current));
5006                 return;
5007         }
5008 
5009         PROTECT_CTX(ctx, flags);
5010 
5011         PFM_SET_WORK_PENDING(current, 0);
5012 
5013         regs = task_pt_regs(current);
5014 
5015         /*
5016          * extract reason for being here and clear
5017          */
5018         reason = ctx->ctx_fl_trap_reason;
5019         ctx->ctx_fl_trap_reason = PFM_TRAP_REASON_NONE;
5020         ovfl_regs = ctx->ctx_ovfl_regs[0];
5021 
5022         DPRINT(("reason=%d state=%d\n", reason, ctx->ctx_state));
5023 
5024         /*
5025          * must be done before we check for simple-reset mode
5026          */
5027         if (ctx->ctx_fl_going_zombie || ctx->ctx_state == PFM_CTX_ZOMBIE)
5028                 goto do_zombie;
5029 
5030         //if (CTX_OVFL_NOBLOCK(ctx)) goto skip_blocking;
5031         if (reason == PFM_TRAP_REASON_RESET)
5032                 goto skip_blocking;
5033 
5034         /*
5035          * restore interrupt mask to what it was on entry.
5036          * Could be enabled/diasbled.
5037          */
5038         UNPROTECT_CTX(ctx, flags);
5039 
5040         /*
5041          * force interrupt enable because of down_interruptible()
5042          */
5043         local_irq_enable();
5044 
5045         DPRINT(("before block sleeping\n"));
5046 
5047         /*
5048          * may go through without blocking on SMP systems
5049          * if restart has been received already by the time we call down()
5050          */
5051         ret = wait_for_completion_interruptible(&ctx->ctx_restart_done);
5052 
5053         DPRINT(("after block sleeping ret=%d\n", ret));
5054 
5055         /*
5056          * lock context and mask interrupts again
5057          * We save flags into a dummy because we may have
5058          * altered interrupts mask compared to entry in this
5059          * function.
5060          */
5061         PROTECT_CTX(ctx, dummy_flags);
5062 
5063         /*
5064          * we need to read the ovfl_regs only after wake-up
5065          * because we may have had pfm_write_pmds() in between
5066          * and that can changed PMD values and therefore 
5067          * ovfl_regs is reset for these new PMD values.
5068          */
5069         ovfl_regs = ctx->ctx_ovfl_regs[0];
5070 
5071         if (ctx->ctx_fl_going_zombie) {
5072 do_zombie:
5073                 DPRINT(("context is zombie, bailing out\n"));
5074                 pfm_context_force_terminate(ctx, regs);
5075                 goto nothing_to_do;
5076         }
5077         /*
5078          * in case of interruption of down() we don't restart anything
5079          */
5080         if (ret < 0)
5081                 goto nothing_to_do;
5082 
5083 skip_blocking:
5084         pfm_resume_after_ovfl(ctx, ovfl_regs, regs);
5085         ctx->ctx_ovfl_regs[0] = 0UL;
5086 
5087 nothing_to_do:
5088         /*
5089          * restore flags as they were upon entry
5090          */
5091         UNPROTECT_CTX(ctx, flags);
5092 }
5093 
5094 static int
5095 pfm_notify_user(pfm_context_t *ctx, pfm_msg_t *msg)
5096 {
5097         if (ctx->ctx_state == PFM_CTX_ZOMBIE) {
5098                 DPRINT(("ignoring overflow notification, owner is zombie\n"));
5099                 return 0;
5100         }
5101 
5102         DPRINT(("waking up somebody\n"));
5103 
5104         if (msg) wake_up_interruptible(&ctx->ctx_msgq_wait);
5105 
5106         /*
5107          * safe, we are not in intr handler, nor in ctxsw when
5108          * we come here
5109          */
5110         kill_fasync (&ctx->ctx_async_queue, SIGIO, POLL_IN);
5111 
5112         return 0;
5113 }
5114 
5115 static int
5116 pfm_ovfl_notify_user(pfm_context_t *ctx, unsigned long ovfl_pmds)
5117 {
5118         pfm_msg_t *msg = NULL;
5119 
5120         if (ctx->ctx_fl_no_msg == 0) {
5121                 msg = pfm_get_new_msg(ctx);
5122                 if (msg == NULL) {
5123                         printk(KERN_ERR "perfmon: pfm_ovfl_notify_user no more notification msgs\n");
5124                         return -1;
5125                 }
5126 
5127                 msg->pfm_ovfl_msg.msg_type         = PFM_MSG_OVFL;
5128                 msg->pfm_ovfl_msg.msg_ctx_fd       = ctx->ctx_fd;
5129                 msg->pfm_ovfl_msg.msg_active_set   = 0;
5130                 msg->pfm_ovfl_msg.msg_ovfl_pmds[0] = ovfl_pmds;
5131                 msg->pfm_ovfl_msg.msg_ovfl_pmds[1] = 0UL;
5132                 msg->pfm_ovfl_msg.msg_ovfl_pmds[2] = 0UL;
5133                 msg->pfm_ovfl_msg.msg_ovfl_pmds[3] = 0UL;
5134                 msg->pfm_ovfl_msg.msg_tstamp       = 0UL;
5135         }
5136 
5137         DPRINT(("ovfl msg: msg=%p no_msg=%d fd=%d ovfl_pmds=0x%lx\n",
5138                 msg,
5139                 ctx->ctx_fl_no_msg,
5140                 ctx->ctx_fd,
5141                 ovfl_pmds));
5142 
5143         return pfm_notify_user(ctx, msg);
5144 }
5145 
5146 static int
5147 pfm_end_notify_user(pfm_context_t *ctx)
5148 {
5149         pfm_msg_t *msg;
5150 
5151         msg = pfm_get_new_msg(ctx);
5152         if (msg == NULL) {
5153                 printk(KERN_ERR "perfmon: pfm_end_notify_user no more notification msgs\n");
5154                 return -1;
5155         }
5156         /* no leak */
5157         memset(msg, 0, sizeof(*msg));
5158 
5159         msg->pfm_end_msg.msg_type    = PFM_MSG_END;
5160         msg->pfm_end_msg.msg_ctx_fd  = ctx->ctx_fd;
5161         msg->pfm_ovfl_msg.msg_tstamp = 0UL;
5162 
5163         DPRINT(("end msg: msg=%p no_msg=%d ctx_fd=%d\n",
5164                 msg,
5165                 ctx->ctx_fl_no_msg,
5166                 ctx->ctx_fd));
5167 
5168         return pfm_notify_user(ctx, msg);
5169 }
5170 
5171 /*
5172  * main overflow processing routine.
5173  * it can be called from the interrupt path or explicitly during the context switch code
5174  */
5175 static void pfm_overflow_handler(struct task_struct *task, pfm_context_t *ctx,
5176                                 unsigned long pmc0, struct pt_regs *regs)
5177 {
5178         pfm_ovfl_arg_t *ovfl_arg;
5179         unsigned long mask;
5180         unsigned long old_val, ovfl_val, new_val;
5181         unsigned long ovfl_notify = 0UL, ovfl_pmds = 0UL, smpl_pmds = 0UL, reset_pmds;
5182         unsigned long tstamp;
5183         pfm_ovfl_ctrl_t ovfl_ctrl;
5184         unsigned int i, has_smpl;
5185         int must_notify = 0;
5186 
5187         if (unlikely(ctx->ctx_state == PFM_CTX_ZOMBIE)) goto stop_monitoring;
5188 
5189         /*
5190          * sanity test. Should never happen
5191          */
5192         if (unlikely((pmc0 & 0x1) == 0)) goto sanity_check;
5193 
5194         tstamp   = ia64_get_itc();
5195         mask     = pmc0 >> PMU_FIRST_COUNTER;
5196         ovfl_val = pmu_conf->ovfl_val;
5197         has_smpl = CTX_HAS_SMPL(ctx);
5198 
5199         DPRINT_ovfl(("pmc0=0x%lx pid=%d iip=0x%lx, %s "
5200                      "used_pmds=0x%lx\n",
5201                         pmc0,
5202                         task ? task_pid_nr(task): -1,
5203                         (regs ? regs->cr_iip : 0),
5204                         CTX_OVFL_NOBLOCK(ctx) ? "nonblocking" : "blocking",
5205                         ctx->ctx_used_pmds[0]));
5206 
5207 
5208         /*
5209          * first we update the virtual counters
5210          * assume there was a prior ia64_srlz_d() issued
5211          */
5212         for (i = PMU_FIRST_COUNTER; mask ; i++, mask >>= 1) {
5213 
5214                 /* skip pmd which did not overflow */
5215                 if ((mask & 0x1) == 0) continue;
5216 
5217                 /*
5218                  * Note that the pmd is not necessarily 0 at this point as qualified events
5219                  * may have happened before the PMU was frozen. The residual count is not
5220                  * taken into consideration here but will be with any read of the pmd via
5221                  * pfm_read_pmds().
5222                  */
5223                 old_val              = new_val = ctx->ctx_pmds[i].val;
5224                 new_val             += 1 + ovfl_val;
5225                 ctx->ctx_pmds[i].val = new_val;
5226 
5227                 /*
5228                  * check for overflow condition
5229                  */
5230                 if (likely(old_val > new_val)) {
5231                         ovfl_pmds |= 1UL << i;
5232                         if (PMC_OVFL_NOTIFY(ctx, i)) ovfl_notify |= 1UL << i;
5233                 }
5234 
5235                 DPRINT_ovfl(("ctx_pmd[%d].val=0x%lx old_val=0x%lx pmd=0x%lx ovfl_pmds=0x%lx ovfl_notify=0x%lx\n",
5236                         i,
5237                         new_val,
5238                         old_val,
5239                         ia64_get_pmd(i) & ovfl_val,
5240                         ovfl_pmds,
5241                         ovfl_notify));
5242         }
5243 
5244         /*
5245          * there was no 64-bit overflow, nothing else to do
5246          */
5247         if (ovfl_pmds == 0UL) return;
5248 
5249         /* 
5250          * reset all control bits
5251          */
5252         ovfl_ctrl.val = 0;
5253         reset_pmds    = 0UL;
5254 
5255         /*
5256          * if a sampling format module exists, then we "cache" the overflow by 
5257          * calling the module's handler() routine.
5258          */
5259         if (has_smpl) {
5260                 unsigned long start_cycles, end_cycles;
5261                 unsigned long pmd_mask;
5262                 int j, k, ret = 0;
5263                 int this_cpu = smp_processor_id();
5264 
5265                 pmd_mask = ovfl_pmds >> PMU_FIRST_COUNTER;
5266                 ovfl_arg = &ctx->ctx_ovfl_arg;
5267 
5268                 prefetch(ctx->ctx_smpl_hdr);
5269 
5270                 for(i=PMU_FIRST_COUNTER; pmd_mask && ret == 0; i++, pmd_mask >>=1) {
5271 
5272                         mask = 1UL << i;
5273 
5274                         if ((pmd_mask & 0x1) == 0) continue;
5275 
5276                         ovfl_arg->ovfl_pmd      = (unsigned char )i;
5277                         ovfl_arg->ovfl_notify   = ovfl_notify & mask ? 1 : 0;
5278                         ovfl_arg->active_set    = 0;
5279                         ovfl_arg->ovfl_ctrl.val = 0; /* module must fill in all fields */
5280                         ovfl_arg->smpl_pmds[0]  = smpl_pmds = ctx->ctx_pmds[i].smpl_pmds[0];
5281 
5282                         ovfl_arg->pmd_value      = ctx->ctx_pmds[i].val;
5283                         ovfl_arg->pmd_last_reset = ctx->ctx_pmds[i].lval;
5284                         ovfl_arg->pmd_eventid    = ctx->ctx_pmds[i].eventid;
5285 
5286                         /*
5287                          * copy values of pmds of interest. Sampling format may copy them
5288                          * into sampling buffer.
5289                          */
5290                         if (smpl_pmds) {
5291                                 for(j=0, k=0; smpl_pmds; j++, smpl_pmds >>=1) {
5292                                         if ((smpl_pmds & 0x1) == 0) continue;
5293                                         ovfl_arg->smpl_pmds_values[k++] = PMD_IS_COUNTING(j) ?  pfm_read_soft_counter(ctx, j) : ia64_get_pmd(j);
5294                                         DPRINT_ovfl(("smpl_pmd[%d]=pmd%u=0x%lx\n", k-1, j, ovfl_arg->smpl_pmds_values[k-1]));
5295                                 }
5296                         }
5297 
5298                         pfm_stats[this_cpu].pfm_smpl_handler_calls++;
5299 
5300                         start_cycles = ia64_get_itc();
5301 
5302                         /*
5303                          * call custom buffer format record (handler) routine
5304                          */
5305                         ret = (*ctx->ctx_buf_fmt->fmt_handler)(task, ctx->ctx_smpl_hdr, ovfl_arg, regs, tstamp);
5306 
5307                         end_cycles = ia64_get_itc();
5308 
5309                         /*
5310                          * For those controls, we take the union because they have
5311                          * an all or nothing behavior.
5312                          */
5313                         ovfl_ctrl.bits.notify_user     |= ovfl_arg->ovfl_ctrl.bits.notify_user;
5314                         ovfl_ctrl.bits.block_task      |= ovfl_arg->ovfl_ctrl.bits.block_task;
5315                         ovfl_ctrl.bits.mask_monitoring |= ovfl_arg->ovfl_ctrl.bits.mask_monitoring;
5316                         /*
5317                          * build the bitmask of pmds to reset now
5318                          */
5319                         if (ovfl_arg->ovfl_ctrl.bits.reset_ovfl_pmds) reset_pmds |= mask;
5320 
5321                         pfm_stats[this_cpu].pfm_smpl_handler_cycles += end_cycles - start_cycles;
5322                 }
5323                 /*
5324                  * when the module cannot handle the rest of the overflows, we abort right here
5325                  */
5326                 if (ret && pmd_mask) {
5327                         DPRINT(("handler aborts leftover ovfl_pmds=0x%lx\n",
5328                                 pmd_mask<<PMU_FIRST_COUNTER));
5329                 }
5330                 /*
5331                  * remove the pmds we reset now from the set of pmds to reset in pfm_restart()
5332                  */
5333                 ovfl_pmds &= ~reset_pmds;
5334         } else {
5335                 /*
5336                  * when no sampling module is used, then the default
5337                  * is to notify on overflow if requested by user
5338                  */
5339                 ovfl_ctrl.bits.notify_user     = ovfl_notify ? 1 : 0;
5340                 ovfl_ctrl.bits.block_task      = ovfl_notify ? 1 : 0;
5341                 ovfl_ctrl.bits.mask_monitoring = ovfl_notify ? 1 : 0; /* XXX: change for saturation */
5342                 ovfl_ctrl.bits.reset_ovfl_pmds = ovfl_notify ? 0 : 1;
5343                 /*
5344                  * if needed, we reset all overflowed pmds
5345                  */
5346                 if (ovfl_notify == 0) reset_pmds = ovfl_pmds;
5347         }
5348 
5349         DPRINT_ovfl(("ovfl_pmds=0x%lx reset_pmds=0x%lx\n", ovfl_pmds, reset_pmds));
5350 
5351         /*
5352          * reset the requested PMD registers using the short reset values
5353          */
5354         if (reset_pmds) {
5355                 unsigned long bm = reset_pmds;
5356                 pfm_reset_regs(ctx, &bm, PFM_PMD_SHORT_RESET);
5357         }
5358 
5359         if (ovfl_notify && ovfl_ctrl.bits.notify_user) {
5360                 /*
5361                  * keep track of what to reset when unblocking
5362                  */
5363                 ctx->ctx_ovfl_regs[0] = ovfl_pmds;
5364 
5365                 /*
5366                  * check for blocking context 
5367                  */
5368                 if (CTX_OVFL_NOBLOCK(ctx) == 0 && ovfl_ctrl.bits.block_task) {
5369 
5370                         ctx->ctx_fl_trap_reason = PFM_TRAP_REASON_BLOCK;
5371 
5372                         /*
5373                          * set the perfmon specific checking pending work for the task
5374                          */
5375                         PFM_SET_WORK_PENDING(task, 1);
5376 
5377                         /*
5378                          * when coming from ctxsw, current still points to the
5379                          * previous task, therefore we must work with task and not current.
5380                          */
5381                         set_notify_resume(task);
5382                 }
5383                 /*
5384                  * defer until state is changed (shorten spin window). the context is locked
5385                  * anyway, so the signal receiver would come spin for nothing.
5386                  */
5387                 must_notify = 1;
5388         }
5389 
5390         DPRINT_ovfl(("owner [%d] pending=%ld reason=%u ovfl_pmds=0x%lx ovfl_notify=0x%lx masked=%d\n",
5391                         GET_PMU_OWNER() ? task_pid_nr(GET_PMU_OWNER()) : -1,
5392                         PFM_GET_WORK_PENDING(task),
5393                         ctx->ctx_fl_trap_reason,
5394                         ovfl_pmds,
5395                         ovfl_notify,
5396                         ovfl_ctrl.bits.mask_monitoring ? 1 : 0));
5397         /*
5398          * in case monitoring must be stopped, we toggle the psr bits
5399          */
5400         if (ovfl_ctrl.bits.mask_monitoring) {
5401                 pfm_mask_monitoring(task);
5402                 ctx->ctx_state = PFM_CTX_MASKED;
5403                 ctx->ctx_fl_can_restart = 1;
5404         }
5405 
5406         /*
5407          * send notification now
5408          */
5409         if (must_notify) pfm_ovfl_notify_user(ctx, ovfl_notify);
5410 
5411         return;
5412 
5413 sanity_check:
5414         printk(KERN_ERR "perfmon: CPU%d overflow handler [%d] pmc0=0x%lx\n",
5415                         smp_processor_id(),
5416                         task ? task_pid_nr(task) : -1,
5417                         pmc0);
5418         return;
5419 
5420 stop_monitoring:
5421         /*
5422          * in SMP, zombie context is never restored but reclaimed in pfm_load_regs().
5423          * Moreover, zombies are also reclaimed in pfm_save_regs(). Therefore we can
5424          * come here as zombie only if the task is the current task. In which case, we
5425          * can access the PMU  hardware directly.
5426          *
5427          * Note that zombies do have PM_VALID set. So here we do the minimal.
5428          *
5429          * In case the context was zombified it could not be reclaimed at the time
5430          * the monitoring program exited. At this point, the PMU reservation has been
5431          * returned, the sampiing buffer has been freed. We must convert this call
5432          * into a spurious interrupt. However, we must also avoid infinite overflows
5433          * by stopping monitoring for this task. We can only come here for a per-task
5434          * context. All we need to do is to stop monitoring using the psr bits which
5435          * are always task private. By re-enabling secure montioring, we ensure that
5436          * the monitored task will not be able to re-activate monitoring.
5437          * The task will eventually be context switched out, at which point the context
5438          * will be reclaimed (that includes releasing ownership of the PMU).
5439          *
5440          * So there might be a window of time where the number of per-task session is zero
5441          * yet one PMU might have a owner and get at most one overflow interrupt for a zombie
5442          * context. This is safe because if a per-task session comes in, it will push this one
5443          * out and by the virtue on pfm_save_regs(), this one will disappear. If a system wide
5444          * session is force on that CPU, given that we use task pinning, pfm_save_regs() will
5445          * also push our zombie context out.
5446          *
5447          * Overall pretty hairy stuff....
5448          */
5449         DPRINT(("ctx is zombie for [%d], converted to spurious\n", task ? task_pid_nr(task): -1));
5450         pfm_clear_psr_up();
5451         ia64_psr(regs)->up = 0;
5452         ia64_psr(regs)->sp = 1;
5453         return;
5454 }
5455 
5456 static int
5457 pfm_do_interrupt_handler(void *arg, struct pt_regs *regs)
5458 {
5459         struct task_struct *task;
5460         pfm_context_t *ctx;
5461         unsigned long flags;
5462         u64 pmc0;
5463         int this_cpu = smp_processor_id();
5464         int retval = 0;
5465 
5466         pfm_stats[this_cpu].pfm_ovfl_intr_count++;
5467 
5468         /*
5469          * srlz.d done before arriving here
5470          */
5471         pmc0 = ia64_get_pmc(0);
5472 
5473         task = GET_PMU_OWNER();
5474         ctx  = GET_PMU_CTX();
5475 
5476         /*
5477          * if we have some pending bits set
5478          * assumes : if any PMC0.bit[63-1] is set, then PMC0.fr = 1
5479          */
5480         if (PMC0_HAS_OVFL(pmc0) && task) {
5481                 /*
5482                  * we assume that pmc0.fr is always set here
5483                  */
5484 
5485                 /* sanity check */
5486                 if (!ctx) goto report_spurious1;
5487 
5488                 if (ctx->ctx_fl_system == 0 && (task->thread.flags & IA64_THREAD_PM_VALID) == 0) 
5489                         goto report_spurious2;
5490 
5491                 PROTECT_CTX_NOPRINT(ctx, flags);
5492 
5493                 pfm_overflow_handler(task, ctx, pmc0, regs);
5494 
5495                 UNPROTECT_CTX_NOPRINT(ctx, flags);
5496 
5497         } else {
5498                 pfm_stats[this_cpu].pfm_spurious_ovfl_intr_count++;
5499                 retval = -1;
5500         }
5501         /*
5502          * keep it unfrozen at all times
5503          */
5504         pfm_unfreeze_pmu();
5505 
5506         return retval;
5507 
5508 report_spurious1:
5509         printk(KERN_INFO "perfmon: spurious overflow interrupt on CPU%d: process %d has no PFM context\n",
5510                 this_cpu, task_pid_nr(task));
5511         pfm_unfreeze_pmu();
5512         return -1;
5513 report_spurious2:
5514         printk(KERN_INFO "perfmon: spurious overflow interrupt on CPU%d: process %d, invalid flag\n", 
5515                 this_cpu, 
5516                 task_pid_nr(task));
5517         pfm_unfreeze_pmu();
5518         return -1;
5519 }
5520 
5521 static irqreturn_t
5522 pfm_interrupt_handler(int irq, void *arg)
5523 {
5524         unsigned long start_cycles, total_cycles;
5525         unsigned long min, max;
5526         int this_cpu;
5527         int ret;
5528         struct pt_regs *regs = get_irq_regs();
5529 
5530         this_cpu = get_cpu();
5531         if (likely(!pfm_alt_intr_handler)) {
5532                 min = pfm_stats[this_cpu].pfm_ovfl_intr_cycles_min;
5533                 max = pfm_stats[this_cpu].pfm_ovfl_intr_cycles_max;
5534 
5535                 start_cycles = ia64_get_itc();
5536 
5537                 ret = pfm_do_interrupt_handler(arg, regs);
5538 
5539                 total_cycles = ia64_get_itc();
5540 
5541                 /*
5542                  * don't measure spurious interrupts
5543                  */
5544                 if (likely(ret == 0)) {
5545                         total_cycles -= start_cycles;
5546 
5547                         if (total_cycles < min) pfm_stats[this_cpu].pfm_ovfl_intr_cycles_min = total_cycles;
5548                         if (total_cycles > max) pfm_stats[this_cpu].pfm_ovfl_intr_cycles_max = total_cycles;
5549 
5550                         pfm_stats[this_cpu].pfm_ovfl_intr_cycles += total_cycles;
5551                 }
5552         }
5553         else {
5554                 (*pfm_alt_intr_handler->handler)(irq, arg, regs);
5555         }
5556 
5557         put_cpu();
5558         return IRQ_HANDLED;
5559 }
5560 
5561 /*
5562  * /proc/perfmon interface, for debug only
5563  */
5564 
5565 #define PFM_PROC_SHOW_HEADER    ((void *)(long)nr_cpu_ids+1)
5566 
5567 static void *
5568 pfm_proc_start(struct seq_file *m, loff_t *pos)
5569 {
5570         if (*pos == 0) {
5571                 return PFM_PROC_SHOW_HEADER;
5572         }
5573 
5574         while (*pos <= nr_cpu_ids) {
5575                 if (cpu_online(*pos - 1)) {
5576                         return (void *)*pos;
5577                 }
5578                 ++*pos;
5579         }
5580         return NULL;
5581 }
5582 
5583 static void *
5584 pfm_proc_next(struct seq_file *m, void *v, loff_t *pos)
5585 {
5586         ++*pos;
5587         return pfm_proc_start(m, pos);
5588 }
5589 
5590 static void
5591 pfm_proc_stop(struct seq_file *m, void *v)
5592 {
5593 }
5594 
5595 static void
5596 pfm_proc_show_header(struct seq_file *m)
5597 {
5598         struct list_head * pos;
5599         pfm_buffer_fmt_t * entry;
5600         unsigned long flags;
5601 
5602         seq_printf(m,
5603                 "perfmon version           : %u.%u\n"
5604                 "model                     : %s\n"
5605                 "fastctxsw                 : %s\n"
5606                 "expert mode               : %s\n"
5607                 "ovfl_mask                 : 0x%lx\n"
5608                 "PMU flags                 : 0x%x\n",
5609                 PFM_VERSION_MAJ, PFM_VERSION_MIN,
5610                 pmu_conf->pmu_name,
5611                 pfm_sysctl.fastctxsw > 0 ? "Yes": "No",
5612                 pfm_sysctl.expert_mode > 0 ? "Yes": "No",
5613                 pmu_conf->ovfl_val,
5614                 pmu_conf->flags);
5615 
5616         LOCK_PFS(flags);
5617 
5618         seq_printf(m,
5619                 "proc_sessions             : %u\n"
5620                 "sys_sessions              : %u\n"
5621                 "sys_use_dbregs            : %u\n"
5622                 "ptrace_use_dbregs         : %u\n",
5623                 pfm_sessions.pfs_task_sessions,
5624                 pfm_sessions.pfs_sys_sessions,
5625                 pfm_sessions.pfs_sys_use_dbregs,
5626                 pfm_sessions.pfs_ptrace_use_dbregs);
5627 
5628         UNLOCK_PFS(flags);
5629 
5630         spin_lock(&pfm_buffer_fmt_lock);
5631 
5632         list_for_each(pos, &pfm_buffer_fmt_list) {
5633                 entry = list_entry(pos, pfm_buffer_fmt_t, fmt_list);
5634                 seq_printf(m, "format                    : %16phD %s\n",
5635                            entry->fmt_uuid, entry->fmt_name);
5636         }
5637         spin_unlock(&pfm_buffer_fmt_lock);
5638 
5639 }
5640 
5641 static int
5642 pfm_proc_show(struct seq_file *m, void *v)
5643 {
5644         unsigned long psr;
5645         unsigned int i;
5646         int cpu;
5647 
5648         if (v == PFM_PROC_SHOW_HEADER) {
5649                 pfm_proc_show_header(m);
5650                 return 0;
5651         }
5652 
5653         /* show info for CPU (v - 1) */
5654 
5655         cpu = (long)v - 1;
5656         seq_printf(m,
5657                 "CPU%-2d overflow intrs      : %lu\n"
5658                 "CPU%-2d overflow cycles     : %lu\n"
5659                 "CPU%-2d overflow min        : %lu\n"
5660                 "CPU%-2d overflow max        : %lu\n"
5661                 "CPU%-2d smpl handler calls  : %lu\n"
5662                 "CPU%-2d smpl handler cycles : %lu\n"
5663                 "CPU%-2d spurious intrs      : %lu\n"
5664                 "CPU%-2d replay   intrs      : %lu\n"
5665                 "CPU%-2d syst_wide           : %d\n"
5666                 "CPU%-2d dcr_pp              : %d\n"
5667                 "CPU%-2d exclude idle        : %d\n"
5668                 "CPU%-2d owner               : %d\n"
5669                 "CPU%-2d context             : %p\n"
5670                 "CPU%-2d activations         : %lu\n",
5671                 cpu, pfm_stats[cpu].pfm_ovfl_intr_count,
5672                 cpu, pfm_stats[cpu].pfm_ovfl_intr_cycles,
5673                 cpu, pfm_stats[cpu].pfm_ovfl_intr_cycles_min,
5674                 cpu, pfm_stats[cpu].pfm_ovfl_intr_cycles_max,
5675                 cpu, pfm_stats[cpu].pfm_smpl_handler_calls,
5676                 cpu, pfm_stats[cpu].pfm_smpl_handler_cycles,
5677                 cpu, pfm_stats[cpu].pfm_spurious_ovfl_intr_count,
5678                 cpu, pfm_stats[cpu].pfm_replay_ovfl_intr_count,
5679                 cpu, pfm_get_cpu_data(pfm_syst_info, cpu) & PFM_CPUINFO_SYST_WIDE ? 1 : 0,
5680                 cpu, pfm_get_cpu_data(pfm_syst_info, cpu) & PFM_CPUINFO_DCR_PP ? 1 : 0,
5681                 cpu, pfm_get_cpu_data(pfm_syst_info, cpu) & PFM_CPUINFO_EXCL_IDLE ? 1 : 0,
5682                 cpu, pfm_get_cpu_data(pmu_owner, cpu) ? pfm_get_cpu_data(pmu_owner, cpu)->pid: -1,
5683                 cpu, pfm_get_cpu_data(pmu_ctx, cpu),
5684                 cpu, pfm_get_cpu_data(pmu_activation_number, cpu));
5685 
5686         if (num_online_cpus() == 1 && pfm_sysctl.debug > 0) {
5687 
5688                 psr = pfm_get_psr();
5689 
5690                 ia64_srlz_d();
5691 
5692                 seq_printf(m, 
5693                         "CPU%-2d psr                 : 0x%lx\n"
5694                         "CPU%-2d pmc0                : 0x%lx\n", 
5695                         cpu, psr,
5696                         cpu, ia64_get_pmc(0));
5697 
5698                 for (i=0; PMC_IS_LAST(i) == 0;  i++) {
5699                         if (PMC_IS_COUNTING(i) == 0) continue;
5700                         seq_printf(m, 
5701                                 "CPU%-2d pmc%u                : 0x%lx\n"
5702                                 "CPU%-2d pmd%u                : 0x%lx\n", 
5703                                 cpu, i, ia64_get_pmc(i),
5704                                 cpu, i, ia64_get_pmd(i));
5705                 }
5706         }
5707         return 0;
5708 }
5709 
5710 const struct seq_operations pfm_seq_ops = {
5711         .start =        pfm_proc_start,
5712         .next =         pfm_proc_next,
5713         .stop =         pfm_proc_stop,
5714         .show =         pfm_proc_show
5715 };
5716 
5717 static int
5718 pfm_proc_open(struct inode *inode, struct file *file)
5719 {
5720         return seq_open(file, &pfm_seq_ops);
5721 }
5722 
5723 
5724 /*
5725  * we come here as soon as local_cpu_data->pfm_syst_wide is set. this happens
5726  * during pfm_enable() hence before pfm_start(). We cannot assume monitoring
5727  * is active or inactive based on mode. We must rely on the value in
5728  * local_cpu_data->pfm_syst_info
5729  */
5730 void
5731 pfm_syst_wide_update_task(struct task_struct *task, unsigned long info, int is_ctxswin)
5732 {
5733         struct pt_regs *regs;
5734         unsigned long dcr;
5735         unsigned long dcr_pp;
5736 
5737         dcr_pp = info & PFM_CPUINFO_DCR_PP ? 1 : 0;
5738 
5739         /*
5740          * pid 0 is guaranteed to be the idle task. There is one such task with pid 0
5741          * on every CPU, so we can rely on the pid to identify the idle task.
5742          */
5743         if ((info & PFM_CPUINFO_EXCL_IDLE) == 0 || task->pid) {
5744                 regs = task_pt_regs(task);
5745                 ia64_psr(regs)->pp = is_ctxswin ? dcr_pp : 0;
5746                 return;
5747         }
5748         /*
5749          * if monitoring has started
5750          */
5751         if (dcr_pp) {
5752                 dcr = ia64_getreg(_IA64_REG_CR_DCR);
5753                 /*
5754                  * context switching in?
5755                  */
5756                 if (is_ctxswin) {
5757                         /* mask monitoring for the idle task */
5758                         ia64_setreg(_IA64_REG_CR_DCR, dcr & ~IA64_DCR_PP);
5759                         pfm_clear_psr_pp();
5760                         ia64_srlz_i();
5761                         return;
5762                 }
5763                 /*
5764                  * context switching out
5765                  * restore monitoring for next task
5766                  *
5767                  * Due to inlining this odd if-then-else construction generates
5768                  * better code.
5769                  */
5770                 ia64_setreg(_IA64_REG_CR_DCR, dcr |IA64_DCR_PP);
5771                 pfm_set_psr_pp();
5772                 ia64_srlz_i();
5773         }
5774 }
5775 
5776 #ifdef CONFIG_SMP
5777 
5778 static void
5779 pfm_force_cleanup(pfm_context_t *ctx, struct pt_regs *regs)
5780 {
5781         struct task_struct *task = ctx->ctx_task;
5782 
5783         ia64_psr(regs)->up = 0;
5784         ia64_psr(regs)->sp = 1;
5785 
5786         if (GET_PMU_OWNER() == task) {
5787                 DPRINT(("cleared ownership for [%d]\n",
5788                                         task_pid_nr(ctx->ctx_task)));
5789                 SET_PMU_OWNER(NULL, NULL);
5790         }
5791 
5792         /*
5793          * disconnect the task from the context and vice-versa
5794          */
5795         PFM_SET_WORK_PENDING(task, 0);
5796 
5797         task->thread.pfm_context  = NULL;
5798         task->thread.flags       &= ~IA64_THREAD_PM_VALID;
5799 
5800         DPRINT(("force cleanup for [%d]\n",  task_pid_nr(task)));
5801 }
5802 
5803 
5804 /*
5805  * in 2.6, interrupts are masked when we come here and the runqueue lock is held
5806  */
5807 void
5808 pfm_save_regs(struct task_struct *task)
5809 {
5810         pfm_context_t *ctx;
5811         unsigned long flags;
5812         u64 psr;
5813 
5814 
5815         ctx = PFM_GET_CTX(task);
5816         if (ctx == NULL) return;
5817 
5818         /*
5819          * we always come here with interrupts ALREADY disabled by
5820          * the scheduler. So we simply need to protect against concurrent
5821          * access, not CPU concurrency.
5822          */
5823         flags = pfm_protect_ctx_ctxsw(ctx);
5824 
5825         if (ctx->ctx_state == PFM_CTX_ZOMBIE) {
5826                 struct pt_regs *regs = task_pt_regs(task);
5827 
5828                 pfm_clear_psr_up();
5829 
5830                 pfm_force_cleanup(ctx, regs);
5831 
5832                 BUG_ON(ctx->ctx_smpl_hdr);
5833 
5834                 pfm_unprotect_ctx_ctxsw(ctx, flags);
5835 
5836                 pfm_context_free(ctx);
5837                 return;
5838         }
5839 
5840         /*
5841          * save current PSR: needed because we modify it
5842          */
5843         ia64_srlz_d();
5844         psr = pfm_get_psr();
5845 
5846         BUG_ON(psr & (IA64_PSR_I));
5847 
5848         /*
5849          * stop monitoring:
5850          * This is the last instruction which may generate an overflow
5851          *
5852          * We do not need to set psr.sp because, it is irrelevant in kernel.
5853          * It will be restored from ipsr when going back to user level
5854          */
5855         pfm_clear_psr_up();
5856 
5857         /*
5858          * keep a copy of psr.up (for reload)
5859          */
5860         ctx->ctx_saved_psr_up = psr & IA64_PSR_UP;
5861 
5862         /*
5863          * release ownership of this PMU.
5864          * PM interrupts are masked, so nothing
5865          * can happen.
5866          */
5867         SET_PMU_OWNER(NULL, NULL);
5868 
5869         /*
5870          * we systematically save the PMD as we have no
5871          * guarantee we will be schedule at that same
5872          * CPU again.
5873          */
5874         pfm_save_pmds(ctx->th_pmds, ctx->ctx_used_pmds[0]);
5875 
5876         /*
5877          * save pmc0 ia64_srlz_d() done in pfm_save_pmds()
5878          * we will need it on the restore path to check
5879          * for pending overflow.
5880          */
5881         ctx->th_pmcs[0] = ia64_get_pmc(0);
5882 
5883         /*
5884          * unfreeze PMU if had pending overflows
5885          */
5886         if (ctx->th_pmcs[0] & ~0x1UL) pfm_unfreeze_pmu();
5887 
5888         /*
5889          * finally, allow context access.
5890          * interrupts will still be masked after this call.
5891          */
5892         pfm_unprotect_ctx_ctxsw(ctx, flags);
5893 }
5894 
5895 #else /* !CONFIG_SMP */
5896 void
5897 pfm_save_regs(struct task_struct *task)
5898 {
5899         pfm_context_t *ctx;
5900         u64 psr;
5901 
5902         ctx = PFM_GET_CTX(task);
5903         if (ctx == NULL) return;
5904 
5905         /*
5906          * save current PSR: needed because we modify it
5907          */
5908         psr = pfm_get_psr();
5909 
5910         BUG_ON(psr & (IA64_PSR_I));
5911 
5912         /*
5913          * stop monitoring:
5914          * This is the last instruction which may generate an overflow
5915          *
5916          * We do not need to set psr.sp because, it is irrelevant in kernel.
5917          * It will be restored from ipsr when going back to user level
5918          */
5919         pfm_clear_psr_up();
5920 
5921         /*
5922          * keep a copy of psr.up (for reload)
5923          */
5924         ctx->ctx_saved_psr_up = psr & IA64_PSR_UP;
5925 }
5926 
5927 static void
5928 pfm_lazy_save_regs (struct task_struct *task)
5929 {
5930         pfm_context_t *ctx;
5931         unsigned long flags;
5932 
5933         { u64 psr  = pfm_get_psr();
5934           BUG_ON(psr & IA64_PSR_UP);
5935         }
5936 
5937         ctx = PFM_GET_CTX(task);
5938 
5939         /*
5940          * we need to mask PMU overflow here to
5941          * make sure that we maintain pmc0 until
5942          * we save it. overflow interrupts are
5943          * treated as spurious if there is no
5944          * owner.
5945          *
5946          * XXX: I don't think this is necessary
5947          */
5948         PROTECT_CTX(ctx,flags);
5949 
5950         /*
5951          * release ownership of this PMU.
5952          * must be done before we save the registers.
5953          *
5954          * after this call any PMU interrupt is treated
5955          * as spurious.
5956          */
5957         SET_PMU_OWNER(NULL, NULL);
5958 
5959         /*
5960          * save all the pmds we use
5961          */
5962         pfm_save_pmds(ctx->th_pmds, ctx->ctx_used_pmds[0]);