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

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

Version: ~ [ linux-5.12-rc5 ] ~ [ linux-5.11.11 ] ~ [ linux-5.10.27 ] ~ [ linux-5.9.16 ] ~ [ linux-5.8.18 ] ~ [ linux-5.7.19 ] ~ [ linux-5.6.19 ] ~ [ linux-5.5.19 ] ~ [ linux-5.4.109 ] ~ [ linux-5.3.18 ] ~ [ linux-5.2.21 ] ~ [ linux-5.1.21 ] ~ [ linux-5.0.21 ] ~ [ linux-4.20.17 ] ~ [ linux-4.19.184 ] ~ [ linux-4.18.20 ] ~ [ linux-4.17.19 ] ~ [ linux-4.16.18 ] ~ [ linux-4.15.18 ] ~ [ linux-4.14.228 ] ~ [ linux-4.13.16 ] ~ [ linux-4.12.14 ] ~ [ linux-4.11.12 ] ~ [ linux-4.10.17 ] ~ [ linux-4.9.264 ] ~ [ linux-4.8.17 ] ~ [ linux-4.7.10 ] ~ [ linux-4.6.7 ] ~ [ linux-4.5.7 ] ~ [ linux-4.4.264 ] ~ [ linux-4.3.6 ] ~ [ linux-4.2.8 ] ~ [ linux-4.1.52 ] ~ [ linux-4.0.9 ] ~ [ linux-3.18.140 ] ~ [ linux-3.16.85 ] ~ [ linux-3.14.79 ] ~ [ linux-3.12.74 ] ~ [ linux-3.10.108 ] ~ [ 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 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 ctl_table pfm_sysctl_dir[] = {
556         {
557                 .procname       = "perfmon",
558                 .mode           = 0555,
559                 .child          = pfm_ctl_table,
560         },
561         {}
562 };
563 static 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 int
2149 pfm_no_open(struct inode *irrelevant, struct file *dontcare)
2150 {
2151         DPRINT(("pfm_no_open called\n"));
2152         return -ENXIO;
2153 }
2154 
2155 
2156 
2157 static const struct file_operations pfm_file_ops = {
2158         .llseek         = no_llseek,
2159         .read           = pfm_read,
2160         .write          = pfm_write,
2161         .poll           = pfm_poll,
2162         .unlocked_ioctl = pfm_ioctl,
2163         .open           = pfm_no_open,  /* special open code to disallow open via /proc */
2164         .fasync         = pfm_fasync,
2165         .release        = pfm_close,
2166         .flush          = pfm_flush
2167 };
2168 
2169 static char *pfmfs_dname(struct dentry *dentry, char *buffer, int buflen)
2170 {
2171         return dynamic_dname(dentry, buffer, buflen, "pfm:[%lu]",
2172                              dentry->d_inode->i_ino);
2173 }
2174 
2175 static const struct dentry_operations pfmfs_dentry_operations = {
2176         .d_delete = always_delete_dentry,
2177         .d_dname = pfmfs_dname,
2178 };
2179 
2180 
2181 static struct file *
2182 pfm_alloc_file(pfm_context_t *ctx)
2183 {
2184         struct file *file;
2185         struct inode *inode;
2186         struct path path;
2187         struct qstr this = { .name = "" };
2188 
2189         /*
2190          * allocate a new inode
2191          */
2192         inode = new_inode(pfmfs_mnt->mnt_sb);
2193         if (!inode)
2194                 return ERR_PTR(-ENOMEM);
2195 
2196         DPRINT(("new inode ino=%ld @%p\n", inode->i_ino, inode));
2197 
2198         inode->i_mode = S_IFCHR|S_IRUGO;
2199         inode->i_uid  = current_fsuid();
2200         inode->i_gid  = current_fsgid();
2201 
2202         /*
2203          * allocate a new dcache entry
2204          */
2205         path.dentry = d_alloc(pfmfs_mnt->mnt_root, &this);
2206         if (!path.dentry) {
2207                 iput(inode);
2208                 return ERR_PTR(-ENOMEM);
2209         }
2210         path.mnt = mntget(pfmfs_mnt);
2211 
2212         d_add(path.dentry, inode);
2213 
2214         file = alloc_file(&path, FMODE_READ, &pfm_file_ops);
2215         if (IS_ERR(file)) {
2216                 path_put(&path);
2217                 return file;
2218         }
2219 
2220         file->f_flags = O_RDONLY;
2221         file->private_data = ctx;
2222 
2223         return file;
2224 }
2225 
2226 static int
2227 pfm_remap_buffer(struct vm_area_struct *vma, unsigned long buf, unsigned long addr, unsigned long size)
2228 {
2229         DPRINT(("CPU%d buf=0x%lx addr=0x%lx size=%ld\n", smp_processor_id(), buf, addr, size));
2230 
2231         while (size > 0) {
2232                 unsigned long pfn = ia64_tpa(buf) >> PAGE_SHIFT;
2233 
2234 
2235                 if (remap_pfn_range(vma, addr, pfn, PAGE_SIZE, PAGE_READONLY))
2236                         return -ENOMEM;
2237 
2238                 addr  += PAGE_SIZE;
2239                 buf   += PAGE_SIZE;
2240                 size  -= PAGE_SIZE;
2241         }
2242         return 0;
2243 }
2244 
2245 /*
2246  * allocate a sampling buffer and remaps it into the user address space of the task
2247  */
2248 static int
2249 pfm_smpl_buffer_alloc(struct task_struct *task, struct file *filp, pfm_context_t *ctx, unsigned long rsize, void **user_vaddr)
2250 {
2251         struct mm_struct *mm = task->mm;
2252         struct vm_area_struct *vma = NULL;
2253         unsigned long size;
2254         void *smpl_buf;
2255 
2256 
2257         /*
2258          * the fixed header + requested size and align to page boundary
2259          */
2260         size = PAGE_ALIGN(rsize);
2261 
2262         DPRINT(("sampling buffer rsize=%lu size=%lu bytes\n", rsize, size));
2263 
2264         /*
2265          * check requested size to avoid Denial-of-service attacks
2266          * XXX: may have to refine this test
2267          * Check against address space limit.
2268          *
2269          * if ((mm->total_vm << PAGE_SHIFT) + len> task->rlim[RLIMIT_AS].rlim_cur)
2270          *      return -ENOMEM;
2271          */
2272         if (size > task_rlimit(task, RLIMIT_MEMLOCK))
2273                 return -ENOMEM;
2274 
2275         /*
2276          * We do the easy to undo allocations first.
2277          *
2278          * pfm_rvmalloc(), clears the buffer, so there is no leak
2279          */
2280         smpl_buf = pfm_rvmalloc(size);
2281         if (smpl_buf == NULL) {
2282                 DPRINT(("Can't allocate sampling buffer\n"));
2283                 return -ENOMEM;
2284         }
2285 
2286         DPRINT(("smpl_buf @%p\n", smpl_buf));
2287 
2288         /* allocate vma */
2289         vma = kmem_cache_zalloc(vm_area_cachep, GFP_KERNEL);
2290         if (!vma) {
2291                 DPRINT(("Cannot allocate vma\n"));
2292                 goto error_kmem;
2293         }
2294         INIT_LIST_HEAD(&vma->anon_vma_chain);
2295 
2296         /*
2297          * partially initialize the vma for the sampling buffer
2298          */
2299         vma->vm_mm           = mm;
2300         vma->vm_file         = get_file(filp);
2301         vma->vm_flags        = VM_READ|VM_MAYREAD|VM_DONTEXPAND|VM_DONTDUMP;
2302         vma->vm_page_prot    = PAGE_READONLY; /* XXX may need to change */
2303 
2304         /*
2305          * Now we have everything we need and we can initialize
2306          * and connect all the data structures
2307          */
2308 
2309         ctx->ctx_smpl_hdr   = smpl_buf;
2310         ctx->ctx_smpl_size  = size; /* aligned size */
2311 
2312         /*
2313          * Let's do the difficult operations next.
2314          *
2315          * now we atomically find some area in the address space and
2316          * remap the buffer in it.
2317          */
2318         down_write(&task->mm->mmap_sem);
2319 
2320         /* find some free area in address space, must have mmap sem held */
2321         vma->vm_start = get_unmapped_area(NULL, 0, size, 0, MAP_PRIVATE|MAP_ANONYMOUS);
2322         if (IS_ERR_VALUE(vma->vm_start)) {
2323                 DPRINT(("Cannot find unmapped area for size %ld\n", size));
2324                 up_write(&task->mm->mmap_sem);
2325                 goto error;
2326         }
2327         vma->vm_end = vma->vm_start + size;
2328         vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT;
2329 
2330         DPRINT(("aligned size=%ld, hdr=%p mapped @0x%lx\n", size, ctx->ctx_smpl_hdr, vma->vm_start));
2331 
2332         /* can only be applied to current task, need to have the mm semaphore held when called */
2333         if (pfm_remap_buffer(vma, (unsigned long)smpl_buf, vma->vm_start, size)) {
2334                 DPRINT(("Can't remap buffer\n"));
2335                 up_write(&task->mm->mmap_sem);
2336                 goto error;
2337         }
2338 
2339         /*
2340          * now insert the vma in the vm list for the process, must be
2341          * done with mmap lock held
2342          */
2343         insert_vm_struct(mm, vma);
2344 
2345         vm_stat_account(vma->vm_mm, vma->vm_flags, vma->vm_file,
2346                                                         vma_pages(vma));
2347         up_write(&task->mm->mmap_sem);
2348 
2349         /*
2350          * keep track of user level virtual address
2351          */
2352         ctx->ctx_smpl_vaddr = (void *)vma->vm_start;
2353         *(unsigned long *)user_vaddr = vma->vm_start;
2354 
2355         return 0;
2356 
2357 error:
2358         kmem_cache_free(vm_area_cachep, vma);
2359 error_kmem:
2360         pfm_rvfree(smpl_buf, size);
2361 
2362         return -ENOMEM;
2363 }
2364 
2365 /*
2366  * XXX: do something better here
2367  */
2368 static int
2369 pfm_bad_permissions(struct task_struct *task)
2370 {
2371         const struct cred *tcred;
2372         kuid_t uid = current_uid();
2373         kgid_t gid = current_gid();
2374         int ret;
2375 
2376         rcu_read_lock();
2377         tcred = __task_cred(task);
2378 
2379         /* inspired by ptrace_attach() */
2380         DPRINT(("cur: uid=%d gid=%d task: euid=%d suid=%d uid=%d egid=%d sgid=%d\n",
2381                 from_kuid(&init_user_ns, uid),
2382                 from_kgid(&init_user_ns, gid),
2383                 from_kuid(&init_user_ns, tcred->euid),
2384                 from_kuid(&init_user_ns, tcred->suid),
2385                 from_kuid(&init_user_ns, tcred->uid),
2386                 from_kgid(&init_user_ns, tcred->egid),
2387                 from_kgid(&init_user_ns, tcred->sgid)));
2388 
2389         ret = ((!uid_eq(uid, tcred->euid))
2390                || (!uid_eq(uid, tcred->suid))
2391                || (!uid_eq(uid, tcred->uid))
2392                || (!gid_eq(gid, tcred->egid))
2393                || (!gid_eq(gid, tcred->sgid))
2394                || (!gid_eq(gid, tcred->gid))) && !capable(CAP_SYS_PTRACE);
2395 
2396         rcu_read_unlock();
2397         return ret;
2398 }
2399 
2400 static int
2401 pfarg_is_sane(struct task_struct *task, pfarg_context_t *pfx)
2402 {
2403         int ctx_flags;
2404 
2405         /* valid signal */
2406 
2407         ctx_flags = pfx->ctx_flags;
2408 
2409         if (ctx_flags & PFM_FL_SYSTEM_WIDE) {
2410 
2411                 /*
2412                  * cannot block in this mode
2413                  */
2414                 if (ctx_flags & PFM_FL_NOTIFY_BLOCK) {
2415                         DPRINT(("cannot use blocking mode when in system wide monitoring\n"));
2416                         return -EINVAL;
2417                 }
2418         } else {
2419         }
2420         /* probably more to add here */
2421 
2422         return 0;
2423 }
2424 
2425 static int
2426 pfm_setup_buffer_fmt(struct task_struct *task, struct file *filp, pfm_context_t *ctx, unsigned int ctx_flags,
2427                      unsigned int cpu, pfarg_context_t *arg)
2428 {
2429         pfm_buffer_fmt_t *fmt = NULL;
2430         unsigned long size = 0UL;
2431         void *uaddr = NULL;
2432         void *fmt_arg = NULL;
2433         int ret = 0;
2434 #define PFM_CTXARG_BUF_ARG(a)   (pfm_buffer_fmt_t *)(a+1)
2435 
2436         /* invoke and lock buffer format, if found */
2437         fmt = pfm_find_buffer_fmt(arg->ctx_smpl_buf_id);
2438         if (fmt == NULL) {
2439                 DPRINT(("[%d] cannot find buffer format\n", task_pid_nr(task)));
2440                 return -EINVAL;
2441         }
2442 
2443         /*
2444          * buffer argument MUST be contiguous to pfarg_context_t
2445          */
2446         if (fmt->fmt_arg_size) fmt_arg = PFM_CTXARG_BUF_ARG(arg);
2447 
2448         ret = pfm_buf_fmt_validate(fmt, task, ctx_flags, cpu, fmt_arg);
2449 
2450         DPRINT(("[%d] after validate(0x%x,%d,%p)=%d\n", task_pid_nr(task), ctx_flags, cpu, fmt_arg, ret));
2451 
2452         if (ret) goto error;
2453 
2454         /* link buffer format and context */
2455         ctx->ctx_buf_fmt = fmt;
2456         ctx->ctx_fl_is_sampling = 1; /* assume record() is defined */
2457 
2458         /*
2459          * check if buffer format wants to use perfmon buffer allocation/mapping service
2460          */
2461         ret = pfm_buf_fmt_getsize(fmt, task, ctx_flags, cpu, fmt_arg, &size);
2462         if (ret) goto error;
2463 
2464         if (size) {
2465                 /*
2466                  * buffer is always remapped into the caller's address space
2467                  */
2468                 ret = pfm_smpl_buffer_alloc(current, filp, ctx, size, &uaddr);
2469                 if (ret) goto error;
2470 
2471                 /* keep track of user address of buffer */
2472                 arg->ctx_smpl_vaddr = uaddr;
2473         }
2474         ret = pfm_buf_fmt_init(fmt, task, ctx->ctx_smpl_hdr, ctx_flags, cpu, fmt_arg);
2475 
2476 error:
2477         return ret;
2478 }
2479 
2480 static void
2481 pfm_reset_pmu_state(pfm_context_t *ctx)
2482 {
2483         int i;
2484 
2485         /*
2486          * install reset values for PMC.
2487          */
2488         for (i=1; PMC_IS_LAST(i) == 0; i++) {
2489                 if (PMC_IS_IMPL(i) == 0) continue;
2490                 ctx->ctx_pmcs[i] = PMC_DFL_VAL(i);
2491                 DPRINT(("pmc[%d]=0x%lx\n", i, ctx->ctx_pmcs[i]));
2492         }
2493         /*
2494          * PMD registers are set to 0UL when the context in memset()
2495          */
2496 
2497         /*
2498          * On context switched restore, we must restore ALL pmc and ALL pmd even
2499          * when they are not actively used by the task. In UP, the incoming process
2500          * may otherwise pick up left over PMC, PMD state from the previous process.
2501          * As opposed to PMD, stale PMC can cause harm to the incoming
2502          * process because they may change what is being measured.
2503          * Therefore, we must systematically reinstall the entire
2504          * PMC state. In SMP, the same thing is possible on the
2505          * same CPU but also on between 2 CPUs.
2506          *
2507          * The problem with PMD is information leaking especially
2508          * to user level when psr.sp=0
2509          *
2510          * There is unfortunately no easy way to avoid this problem
2511          * on either UP or SMP. This definitively slows down the
2512          * pfm_load_regs() function.
2513          */
2514 
2515          /*
2516           * bitmask of all PMCs accessible to this context
2517           *
2518           * PMC0 is treated differently.
2519           */
2520         ctx->ctx_all_pmcs[0] = pmu_conf->impl_pmcs[0] & ~0x1;
2521 
2522         /*
2523          * bitmask of all PMDs that are accessible to this context
2524          */
2525         ctx->ctx_all_pmds[0] = pmu_conf->impl_pmds[0];
2526 
2527         DPRINT(("<%d> all_pmcs=0x%lx all_pmds=0x%lx\n", ctx->ctx_fd, ctx->ctx_all_pmcs[0],ctx->ctx_all_pmds[0]));
2528 
2529         /*
2530          * useful in case of re-enable after disable
2531          */
2532         ctx->ctx_used_ibrs[0] = 0UL;
2533         ctx->ctx_used_dbrs[0] = 0UL;
2534 }
2535 
2536 static int
2537 pfm_ctx_getsize(void *arg, size_t *sz)
2538 {
2539         pfarg_context_t *req = (pfarg_context_t *)arg;
2540         pfm_buffer_fmt_t *fmt;
2541 
2542         *sz = 0;
2543 
2544         if (!pfm_uuid_cmp(req->ctx_smpl_buf_id, pfm_null_uuid)) return 0;
2545 
2546         fmt = pfm_find_buffer_fmt(req->ctx_smpl_buf_id);
2547         if (fmt == NULL) {
2548                 DPRINT(("cannot find buffer format\n"));
2549                 return -EINVAL;
2550         }
2551         /* get just enough to copy in user parameters */
2552         *sz = fmt->fmt_arg_size;
2553         DPRINT(("arg_size=%lu\n", *sz));
2554 
2555         return 0;
2556 }
2557 
2558 
2559 
2560 /*
2561  * cannot attach if :
2562  *      - kernel task
2563  *      - task not owned by caller
2564  *      - task incompatible with context mode
2565  */
2566 static int
2567 pfm_task_incompatible(pfm_context_t *ctx, struct task_struct *task)
2568 {
2569         /*
2570          * no kernel task or task not owner by caller
2571          */
2572         if (task->mm == NULL) {
2573                 DPRINT(("task [%d] has not memory context (kernel thread)\n", task_pid_nr(task)));
2574                 return -EPERM;
2575         }
2576         if (pfm_bad_permissions(task)) {
2577                 DPRINT(("no permission to attach to  [%d]\n", task_pid_nr(task)));
2578                 return -EPERM;
2579         }
2580         /*
2581          * cannot block in self-monitoring mode
2582          */
2583         if (CTX_OVFL_NOBLOCK(ctx) == 0 && task == current) {
2584                 DPRINT(("cannot load a blocking context on self for [%d]\n", task_pid_nr(task)));
2585                 return -EINVAL;
2586         }
2587 
2588         if (task->exit_state == EXIT_ZOMBIE) {
2589                 DPRINT(("cannot attach to  zombie task [%d]\n", task_pid_nr(task)));
2590                 return -EBUSY;
2591         }
2592 
2593         /*
2594          * always ok for self
2595          */
2596         if (task == current) return 0;
2597 
2598         if (!task_is_stopped_or_traced(task)) {
2599                 DPRINT(("cannot attach to non-stopped task [%d] state=%ld\n", task_pid_nr(task), task->state));
2600                 return -EBUSY;
2601         }
2602         /*
2603          * make sure the task is off any CPU
2604          */
2605         wait_task_inactive(task, 0);
2606 
2607         /* more to come... */
2608 
2609         return 0;
2610 }
2611 
2612 static int
2613 pfm_get_task(pfm_context_t *ctx, pid_t pid, struct task_struct **task)
2614 {
2615         struct task_struct *p = current;
2616         int ret;
2617 
2618         /* XXX: need to add more checks here */
2619         if (pid < 2) return -EPERM;
2620 
2621         if (pid != task_pid_vnr(current)) {
2622 
2623                 read_lock(&tasklist_lock);
2624 
2625                 p = find_task_by_vpid(pid);
2626 
2627                 /* make sure task cannot go away while we operate on it */
2628                 if (p) get_task_struct(p);
2629 
2630                 read_unlock(&tasklist_lock);
2631 
2632                 if (p == NULL) return -ESRCH;
2633         }
2634 
2635         ret = pfm_task_incompatible(ctx, p);
2636         if (ret == 0) {
2637                 *task = p;
2638         } else if (p != current) {
2639                 pfm_put_task(p);
2640         }
2641         return ret;
2642 }
2643 
2644 
2645 
2646 static int
2647 pfm_context_create(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
2648 {
2649         pfarg_context_t *req = (pfarg_context_t *)arg;
2650         struct file *filp;
2651         struct path path;
2652         int ctx_flags;
2653         int fd;
2654         int ret;
2655 
2656         /* let's check the arguments first */
2657         ret = pfarg_is_sane(current, req);
2658         if (ret < 0)
2659                 return ret;
2660 
2661         ctx_flags = req->ctx_flags;
2662 
2663         ret = -ENOMEM;
2664 
2665         fd = get_unused_fd();
2666         if (fd < 0)
2667                 return fd;
2668 
2669         ctx = pfm_context_alloc(ctx_flags);
2670         if (!ctx)
2671                 goto error;
2672 
2673         filp = pfm_alloc_file(ctx);
2674         if (IS_ERR(filp)) {
2675                 ret = PTR_ERR(filp);
2676                 goto error_file;
2677         }
2678 
2679         req->ctx_fd = ctx->ctx_fd = fd;
2680 
2681         /*
2682          * does the user want to sample?
2683          */
2684         if (pfm_uuid_cmp(req->ctx_smpl_buf_id, pfm_null_uuid)) {
2685                 ret = pfm_setup_buffer_fmt(current, filp, ctx, ctx_flags, 0, req);
2686                 if (ret)
2687                         goto buffer_error;
2688         }
2689 
2690         DPRINT(("ctx=%p flags=0x%x system=%d notify_block=%d excl_idle=%d no_msg=%d ctx_fd=%d\n",
2691                 ctx,
2692                 ctx_flags,
2693                 ctx->ctx_fl_system,
2694                 ctx->ctx_fl_block,
2695                 ctx->ctx_fl_excl_idle,
2696                 ctx->ctx_fl_no_msg,
2697                 ctx->ctx_fd));
2698 
2699         /*
2700          * initialize soft PMU state
2701          */
2702         pfm_reset_pmu_state(ctx);
2703 
2704         fd_install(fd, filp);
2705 
2706         return 0;
2707 
2708 buffer_error:
2709         path = filp->f_path;
2710         put_filp(filp);
2711         path_put(&path);
2712 
2713         if (ctx->ctx_buf_fmt) {
2714                 pfm_buf_fmt_exit(ctx->ctx_buf_fmt, current, NULL, regs);
2715         }
2716 error_file:
2717         pfm_context_free(ctx);
2718 
2719 error:
2720         put_unused_fd(fd);
2721         return ret;
2722 }
2723 
2724 static inline unsigned long
2725 pfm_new_counter_value (pfm_counter_t *reg, int is_long_reset)
2726 {
2727         unsigned long val = is_long_reset ? reg->long_reset : reg->short_reset;
2728         unsigned long new_seed, old_seed = reg->seed, mask = reg->mask;
2729         extern unsigned long carta_random32 (unsigned long seed);
2730 
2731         if (reg->flags & PFM_REGFL_RANDOM) {
2732                 new_seed = carta_random32(old_seed);
2733                 val -= (old_seed & mask);       /* counter values are negative numbers! */
2734                 if ((mask >> 32) != 0)
2735                         /* construct a full 64-bit random value: */
2736                         new_seed |= carta_random32(old_seed >> 32) << 32;
2737                 reg->seed = new_seed;
2738         }
2739         reg->lval = val;
2740         return val;
2741 }
2742 
2743 static void
2744 pfm_reset_regs_masked(pfm_context_t *ctx, unsigned long *ovfl_regs, int is_long_reset)
2745 {
2746         unsigned long mask = ovfl_regs[0];
2747         unsigned long reset_others = 0UL;
2748         unsigned long val;
2749         int i;
2750 
2751         /*
2752          * now restore reset value on sampling overflowed counters
2753          */
2754         mask >>= PMU_FIRST_COUNTER;
2755         for(i = PMU_FIRST_COUNTER; mask; i++, mask >>= 1) {
2756 
2757                 if ((mask & 0x1UL) == 0UL) continue;
2758 
2759                 ctx->ctx_pmds[i].val = val = pfm_new_counter_value(ctx->ctx_pmds+ i, is_long_reset);
2760                 reset_others        |= ctx->ctx_pmds[i].reset_pmds[0];
2761 
2762                 DPRINT_ovfl((" %s reset ctx_pmds[%d]=%lx\n", is_long_reset ? "long" : "short", i, val));
2763         }
2764 
2765         /*
2766          * Now take care of resetting the other registers
2767          */
2768         for(i = 0; reset_others; i++, reset_others >>= 1) {
2769 
2770                 if ((reset_others & 0x1) == 0) continue;
2771 
2772                 ctx->ctx_pmds[i].val = val = pfm_new_counter_value(ctx->ctx_pmds + i, is_long_reset);
2773 
2774                 DPRINT_ovfl(("%s reset_others pmd[%d]=%lx\n",
2775                           is_long_reset ? "long" : "short", i, val));
2776         }
2777 }
2778 
2779 static void
2780 pfm_reset_regs(pfm_context_t *ctx, unsigned long *ovfl_regs, int is_long_reset)
2781 {
2782         unsigned long mask = ovfl_regs[0];
2783         unsigned long reset_others = 0UL;
2784         unsigned long val;
2785         int i;
2786 
2787         DPRINT_ovfl(("ovfl_regs=0x%lx is_long_reset=%d\n", ovfl_regs[0], is_long_reset));
2788 
2789         if (ctx->ctx_state == PFM_CTX_MASKED) {
2790                 pfm_reset_regs_masked(ctx, ovfl_regs, is_long_reset);
2791                 return;
2792         }
2793 
2794         /*
2795          * now restore reset value on sampling overflowed counters
2796          */
2797         mask >>= PMU_FIRST_COUNTER;
2798         for(i = PMU_FIRST_COUNTER; mask; i++, mask >>= 1) {
2799 
2800                 if ((mask & 0x1UL) == 0UL) continue;
2801 
2802                 val           = pfm_new_counter_value(ctx->ctx_pmds+ i, is_long_reset);
2803                 reset_others |= ctx->ctx_pmds[i].reset_pmds[0];
2804 
2805                 DPRINT_ovfl((" %s reset ctx_pmds[%d]=%lx\n", is_long_reset ? "long" : "short", i, val));
2806 
2807                 pfm_write_soft_counter(ctx, i, val);
2808         }
2809 
2810         /*
2811          * Now take care of resetting the other registers
2812          */
2813         for(i = 0; reset_others; i++, reset_others >>= 1) {
2814 
2815                 if ((reset_others & 0x1) == 0) continue;
2816 
2817                 val = pfm_new_counter_value(ctx->ctx_pmds + i, is_long_reset);
2818 
2819                 if (PMD_IS_COUNTING(i)) {
2820                         pfm_write_soft_counter(ctx, i, val);
2821                 } else {
2822                         ia64_set_pmd(i, val);
2823                 }
2824                 DPRINT_ovfl(("%s reset_others pmd[%d]=%lx\n",
2825                           is_long_reset ? "long" : "short", i, val));
2826         }
2827         ia64_srlz_d();
2828 }
2829 
2830 static int
2831 pfm_write_pmcs(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
2832 {
2833         struct task_struct *task;
2834         pfarg_reg_t *req = (pfarg_reg_t *)arg;
2835         unsigned long value, pmc_pm;
2836         unsigned long smpl_pmds, reset_pmds, impl_pmds;
2837         unsigned int cnum, reg_flags, flags, pmc_type;
2838         int i, can_access_pmu = 0, is_loaded, is_system, expert_mode;
2839         int is_monitor, is_counting, state;
2840         int ret = -EINVAL;
2841         pfm_reg_check_t wr_func;
2842 #define PFM_CHECK_PMC_PM(x, y, z) ((x)->ctx_fl_system ^ PMC_PM(y, z))
2843 
2844         state     = ctx->ctx_state;
2845         is_loaded = state == PFM_CTX_LOADED ? 1 : 0;
2846         is_system = ctx->ctx_fl_system;
2847         task      = ctx->ctx_task;
2848         impl_pmds = pmu_conf->impl_pmds[0];
2849 
2850         if (state == PFM_CTX_ZOMBIE) return -EINVAL;
2851 
2852         if (is_loaded) {
2853                 /*
2854                  * In system wide and when the context is loaded, access can only happen
2855                  * when the caller is running on the CPU being monitored by the session.
2856                  * It does not have to be the owner (ctx_task) of the context per se.
2857                  */
2858                 if (is_system && ctx->ctx_cpu != smp_processor_id()) {
2859                         DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
2860                         return -EBUSY;
2861                 }
2862                 can_access_pmu = GET_PMU_OWNER() == task || is_system ? 1 : 0;
2863         }
2864         expert_mode = pfm_sysctl.expert_mode; 
2865 
2866         for (i = 0; i < count; i++, req++) {
2867 
2868                 cnum       = req->reg_num;
2869                 reg_flags  = req->reg_flags;
2870                 value      = req->reg_value;
2871                 smpl_pmds  = req->reg_smpl_pmds[0];
2872                 reset_pmds = req->reg_reset_pmds[0];
2873                 flags      = 0;
2874 
2875 
2876                 if (cnum >= PMU_MAX_PMCS) {
2877                         DPRINT(("pmc%u is invalid\n", cnum));
2878                         goto error;
2879                 }
2880 
2881                 pmc_type   = pmu_conf->pmc_desc[cnum].type;
2882                 pmc_pm     = (value >> pmu_conf->pmc_desc[cnum].pm_pos) & 0x1;
2883                 is_counting = (pmc_type & PFM_REG_COUNTING) == PFM_REG_COUNTING ? 1 : 0;
2884                 is_monitor  = (pmc_type & PFM_REG_MONITOR) == PFM_REG_MONITOR ? 1 : 0;
2885 
2886                 /*
2887                  * we reject all non implemented PMC as well
2888                  * as attempts to modify PMC[0-3] which are used
2889                  * as status registers by the PMU
2890                  */
2891                 if ((pmc_type & PFM_REG_IMPL) == 0 || (pmc_type & PFM_REG_CONTROL) == PFM_REG_CONTROL) {
2892                         DPRINT(("pmc%u is unimplemented or no-access pmc_type=%x\n", cnum, pmc_type));
2893                         goto error;
2894                 }
2895                 wr_func = pmu_conf->pmc_desc[cnum].write_check;
2896                 /*
2897                  * If the PMC is a monitor, then if the value is not the default:
2898                  *      - system-wide session: PMCx.pm=1 (privileged monitor)
2899                  *      - per-task           : PMCx.pm=0 (user monitor)
2900                  */
2901                 if (is_monitor && value != PMC_DFL_VAL(cnum) && is_system ^ pmc_pm) {
2902                         DPRINT(("pmc%u pmc_pm=%lu is_system=%d\n",
2903                                 cnum,
2904                                 pmc_pm,
2905                                 is_system));
2906                         goto error;
2907                 }
2908 
2909                 if (is_counting) {
2910                         /*
2911                          * enforce generation of overflow interrupt. Necessary on all
2912                          * CPUs.
2913                          */
2914                         value |= 1 << PMU_PMC_OI;
2915 
2916                         if (reg_flags & PFM_REGFL_OVFL_NOTIFY) {
2917                                 flags |= PFM_REGFL_OVFL_NOTIFY;
2918                         }
2919 
2920                         if (reg_flags & PFM_REGFL_RANDOM) flags |= PFM_REGFL_RANDOM;
2921 
2922                         /* verify validity of smpl_pmds */
2923                         if ((smpl_pmds & impl_pmds) != smpl_pmds) {
2924                                 DPRINT(("invalid smpl_pmds 0x%lx for pmc%u\n", smpl_pmds, cnum));
2925                                 goto error;
2926                         }
2927 
2928                         /* verify validity of reset_pmds */
2929                         if ((reset_pmds & impl_pmds) != reset_pmds) {
2930                                 DPRINT(("invalid reset_pmds 0x%lx for pmc%u\n", reset_pmds, cnum));
2931                                 goto error;
2932                         }
2933                 } else {
2934                         if (reg_flags & (PFM_REGFL_OVFL_NOTIFY|PFM_REGFL_RANDOM)) {
2935                                 DPRINT(("cannot set ovfl_notify or random on pmc%u\n", cnum));
2936                                 goto error;
2937                         }
2938                         /* eventid on non-counting monitors are ignored */
2939                 }
2940 
2941                 /*
2942                  * execute write checker, if any
2943                  */
2944                 if (likely(expert_mode == 0 && wr_func)) {
2945                         ret = (*wr_func)(task, ctx, cnum, &value, regs);
2946                         if (ret) goto error;
2947                         ret = -EINVAL;
2948                 }
2949 
2950                 /*
2951                  * no error on this register
2952                  */
2953                 PFM_REG_RETFLAG_SET(req->reg_flags, 0);
2954 
2955                 /*
2956                  * Now we commit the changes to the software state
2957                  */
2958 
2959                 /*
2960                  * update overflow information
2961                  */
2962                 if (is_counting) {
2963                         /*
2964                          * full flag update each time a register is programmed
2965                          */
2966                         ctx->ctx_pmds[cnum].flags = flags;
2967 
2968                         ctx->ctx_pmds[cnum].reset_pmds[0] = reset_pmds;
2969                         ctx->ctx_pmds[cnum].smpl_pmds[0]  = smpl_pmds;
2970                         ctx->ctx_pmds[cnum].eventid       = req->reg_smpl_eventid;
2971 
2972                         /*
2973                          * Mark all PMDS to be accessed as used.
2974                          *
2975                          * We do not keep track of PMC because we have to
2976                          * systematically restore ALL of them.
2977                          *
2978                          * We do not update the used_monitors mask, because
2979                          * if we have not programmed them, then will be in
2980                          * a quiescent state, therefore we will not need to
2981                          * mask/restore then when context is MASKED.
2982                          */
2983                         CTX_USED_PMD(ctx, reset_pmds);
2984                         CTX_USED_PMD(ctx, smpl_pmds);
2985                         /*
2986                          * make sure we do not try to reset on
2987                          * restart because we have established new values
2988                          */
2989                         if (state == PFM_CTX_MASKED) ctx->ctx_ovfl_regs[0] &= ~1UL << cnum;
2990                 }
2991                 /*
2992                  * Needed in case the user does not initialize the equivalent
2993                  * PMD. Clearing is done indirectly via pfm_reset_pmu_state() so there is no
2994                  * possible leak here.
2995                  */
2996                 CTX_USED_PMD(ctx, pmu_conf->pmc_desc[cnum].dep_pmd[0]);
2997 
2998                 /*
2999                  * keep track of the monitor PMC that we are using.
3000                  * we save the value of the pmc in ctx_pmcs[] and if
3001                  * the monitoring is not stopped for the context we also
3002                  * place it in the saved state area so that it will be
3003                  * picked up later by the context switch code.
3004                  *
3005                  * The value in ctx_pmcs[] can only be changed in pfm_write_pmcs().
3006                  *
3007                  * The value in th_pmcs[] may be modified on overflow, i.e.,  when
3008                  * monitoring needs to be stopped.
3009                  */
3010                 if (is_monitor) CTX_USED_MONITOR(ctx, 1UL << cnum);
3011 
3012                 /*
3013                  * update context state
3014                  */
3015                 ctx->ctx_pmcs[cnum] = value;
3016 
3017                 if (is_loaded) {
3018                         /*
3019                          * write thread state
3020                          */
3021                         if (is_system == 0) ctx->th_pmcs[cnum] = value;
3022 
3023                         /*
3024                          * write hardware register if we can
3025                          */
3026                         if (can_access_pmu) {
3027                                 ia64_set_pmc(cnum, value);
3028                         }
3029 #ifdef CONFIG_SMP
3030                         else {
3031                                 /*
3032                                  * per-task SMP only here
3033                                  *
3034                                  * we are guaranteed that the task is not running on the other CPU,
3035                                  * we indicate that this PMD will need to be reloaded if the task
3036                                  * is rescheduled on the CPU it ran last on.
3037                                  */
3038                                 ctx->ctx_reload_pmcs[0] |= 1UL << cnum;
3039                         }
3040 #endif
3041                 }
3042 
3043                 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",
3044                           cnum,
3045                           value,
3046                           is_loaded,
3047                           can_access_pmu,
3048                           flags,
3049                           ctx->ctx_all_pmcs[0],
3050                           ctx->ctx_used_pmds[0],
3051                           ctx->ctx_pmds[cnum].eventid,
3052                           smpl_pmds,
3053                           reset_pmds,
3054                           ctx->ctx_reload_pmcs[0],
3055                           ctx->ctx_used_monitors[0],
3056                           ctx->ctx_ovfl_regs[0]));
3057         }
3058 
3059         /*
3060          * make sure the changes are visible
3061          */
3062         if (can_access_pmu) ia64_srlz_d();
3063 
3064         return 0;
3065 error:
3066         PFM_REG_RETFLAG_SET(req->reg_flags, PFM_REG_RETFL_EINVAL);
3067         return ret;
3068 }
3069 
3070 static int
3071 pfm_write_pmds(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3072 {
3073         struct task_struct *task;
3074         pfarg_reg_t *req = (pfarg_reg_t *)arg;
3075         unsigned long value, hw_value, ovfl_mask;
3076         unsigned int cnum;
3077         int i, can_access_pmu = 0, state;
3078         int is_counting, is_loaded, is_system, expert_mode;
3079         int ret = -EINVAL;
3080         pfm_reg_check_t wr_func;
3081 
3082 
3083         state     = ctx->ctx_state;
3084         is_loaded = state == PFM_CTX_LOADED ? 1 : 0;
3085         is_system = ctx->ctx_fl_system;
3086         ovfl_mask = pmu_conf->ovfl_val;
3087         task      = ctx->ctx_task;
3088 
3089         if (unlikely(state == PFM_CTX_ZOMBIE)) return -EINVAL;
3090 
3091         /*
3092          * on both UP and SMP, we can only write to the PMC when the task is
3093          * the owner of the local PMU.
3094          */
3095         if (likely(is_loaded)) {
3096                 /*
3097                  * In system wide and when the context is loaded, access can only happen
3098                  * when the caller is running on the CPU being monitored by the session.
3099                  * It does not have to be the owner (ctx_task) of the context per se.
3100                  */
3101                 if (unlikely(is_system && ctx->ctx_cpu != smp_processor_id())) {
3102                         DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
3103                         return -EBUSY;
3104                 }
3105                 can_access_pmu = GET_PMU_OWNER() == task || is_system ? 1 : 0;
3106         }
3107         expert_mode = pfm_sysctl.expert_mode; 
3108 
3109         for (i = 0; i < count; i++, req++) {
3110 
3111                 cnum  = req->reg_num;
3112                 value = req->reg_value;
3113 
3114                 if (!PMD_IS_IMPL(cnum)) {
3115                         DPRINT(("pmd[%u] is unimplemented or invalid\n", cnum));
3116                         goto abort_mission;
3117                 }
3118                 is_counting = PMD_IS_COUNTING(cnum);
3119                 wr_func     = pmu_conf->pmd_desc[cnum].write_check;
3120 
3121                 /*
3122                  * execute write checker, if any
3123                  */
3124                 if (unlikely(expert_mode == 0 && wr_func)) {
3125                         unsigned long v = value;
3126 
3127                         ret = (*wr_func)(task, ctx, cnum, &v, regs);
3128                         if (ret) goto abort_mission;
3129 
3130                         value = v;
3131                         ret   = -EINVAL;
3132                 }
3133 
3134                 /*
3135                  * no error on this register
3136                  */
3137                 PFM_REG_RETFLAG_SET(req->reg_flags, 0);
3138 
3139                 /*
3140                  * now commit changes to software state
3141                  */
3142                 hw_value = value;
3143 
3144                 /*
3145                  * update virtualized (64bits) counter
3146                  */
3147                 if (is_counting) {
3148                         /*
3149                          * write context state
3150                          */
3151                         ctx->ctx_pmds[cnum].lval = value;
3152 
3153                         /*
3154                          * when context is load we use the split value
3155                          */
3156                         if (is_loaded) {
3157                                 hw_value = value &  ovfl_mask;
3158                                 value    = value & ~ovfl_mask;
3159                         }
3160                 }
3161                 /*
3162                  * update reset values (not just for counters)
3163                  */
3164                 ctx->ctx_pmds[cnum].long_reset  = req->reg_long_reset;
3165                 ctx->ctx_pmds[cnum].short_reset = req->reg_short_reset;
3166 
3167                 /*
3168                  * update randomization parameters (not just for counters)
3169                  */
3170                 ctx->ctx_pmds[cnum].seed = req->reg_random_seed;
3171                 ctx->ctx_pmds[cnum].mask = req->reg_random_mask;
3172 
3173                 /*
3174                  * update context value
3175                  */
3176                 ctx->ctx_pmds[cnum].val  = value;
3177 
3178                 /*
3179                  * Keep track of what we use
3180                  *
3181                  * We do not keep track of PMC because we have to
3182                  * systematically restore ALL of them.
3183                  */
3184                 CTX_USED_PMD(ctx, PMD_PMD_DEP(cnum));
3185 
3186                 /*
3187                  * mark this PMD register used as well
3188                  */
3189                 CTX_USED_PMD(ctx, RDEP(cnum));
3190 
3191                 /*
3192                  * make sure we do not try to reset on
3193                  * restart because we have established new values
3194                  */
3195                 if (is_counting && state == PFM_CTX_MASKED) {
3196                         ctx->ctx_ovfl_regs[0] &= ~1UL << cnum;
3197                 }
3198 
3199                 if (is_loaded) {
3200                         /*
3201                          * write thread state
3202                          */
3203                         if (is_system == 0) ctx->th_pmds[cnum] = hw_value;
3204 
3205                         /*
3206                          * write hardware register if we can
3207                          */
3208                         if (can_access_pmu) {
3209                                 ia64_set_pmd(cnum, hw_value);
3210                         } else {
3211 #ifdef CONFIG_SMP
3212                                 /*
3213                                  * we are guaranteed that the task is not running on the other CPU,
3214                                  * we indicate that this PMD will need to be reloaded if the task
3215                                  * is rescheduled on the CPU it ran last on.
3216                                  */
3217                                 ctx->ctx_reload_pmds[0] |= 1UL << cnum;
3218 #endif
3219                         }
3220                 }
3221 
3222                 DPRINT(("pmd[%u]=0x%lx ld=%d apmu=%d, hw_value=0x%lx ctx_pmd=0x%lx  short_reset=0x%lx "
3223                           "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",
3224                         cnum,
3225                         value,
3226                         is_loaded,
3227                         can_access_pmu,
3228                         hw_value,
3229                         ctx->ctx_pmds[cnum].val,
3230                         ctx->ctx_pmds[cnum].short_reset,
3231                         ctx->ctx_pmds[cnum].long_reset,
3232                         PMC_OVFL_NOTIFY(ctx, cnum) ? 'Y':'N',
3233                         ctx->ctx_pmds[cnum].seed,
3234                         ctx->ctx_pmds[cnum].mask,
3235                         ctx->ctx_used_pmds[0],
3236                         ctx->ctx_pmds[cnum].reset_pmds[0],
3237                         ctx->ctx_reload_pmds[0],
3238                         ctx->ctx_all_pmds[0],
3239                         ctx->ctx_ovfl_regs[0]));
3240         }
3241 
3242         /*
3243          * make changes visible
3244          */
3245         if (can_access_pmu) ia64_srlz_d();
3246 
3247         return 0;
3248 
3249 abort_mission:
3250         /*
3251          * for now, we have only one possibility for error
3252          */
3253         PFM_REG_RETFLAG_SET(req->reg_flags, PFM_REG_RETFL_EINVAL);
3254         return ret;
3255 }
3256 
3257 /*
3258  * By the way of PROTECT_CONTEXT(), interrupts are masked while we are in this function.
3259  * Therefore we know, we do not have to worry about the PMU overflow interrupt. If an
3260  * interrupt is delivered during the call, it will be kept pending until we leave, making
3261  * it appears as if it had been generated at the UNPROTECT_CONTEXT(). At least we are
3262  * guaranteed to return consistent data to the user, it may simply be old. It is not
3263  * trivial to treat the overflow while inside the call because you may end up in
3264  * some module sampling buffer code causing deadlocks.
3265  */
3266 static int
3267 pfm_read_pmds(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3268 {
3269         struct task_struct *task;
3270         unsigned long val = 0UL, lval, ovfl_mask, sval;
3271         pfarg_reg_t *req = (pfarg_reg_t *)arg;
3272         unsigned int cnum, reg_flags = 0;
3273         int i, can_access_pmu = 0, state;
3274         int is_loaded, is_system, is_counting, expert_mode;
3275         int ret = -EINVAL;
3276         pfm_reg_check_t rd_func;
3277 
3278         /*
3279          * access is possible when loaded only for
3280          * self-monitoring tasks or in UP mode
3281          */
3282 
3283         state     = ctx->ctx_state;
3284         is_loaded = state == PFM_CTX_LOADED ? 1 : 0;
3285         is_system = ctx->ctx_fl_system;
3286         ovfl_mask = pmu_conf->ovfl_val;
3287         task      = ctx->ctx_task;
3288 
3289         if (state == PFM_CTX_ZOMBIE) return -EINVAL;
3290 
3291         if (likely(is_loaded)) {
3292                 /*
3293                  * In system wide and when the context is loaded, access can only happen
3294                  * when the caller is running on the CPU being monitored by the session.
3295                  * It does not have to be the owner (ctx_task) of the context per se.
3296                  */
3297                 if (unlikely(is_system && ctx->ctx_cpu != smp_processor_id())) {
3298                         DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
3299                         return -EBUSY;
3300                 }
3301                 /*
3302                  * this can be true when not self-monitoring only in UP
3303                  */
3304                 can_access_pmu = GET_PMU_OWNER() == task || is_system ? 1 : 0;
3305 
3306                 if (can_access_pmu) ia64_srlz_d();
3307         }
3308         expert_mode = pfm_sysctl.expert_mode; 
3309 
3310         DPRINT(("ld=%d apmu=%d ctx_state=%d\n",
3311                 is_loaded,
3312                 can_access_pmu,
3313                 state));
3314 
3315         /*
3316          * on both UP and SMP, we can only read the PMD from the hardware register when
3317          * the task is the owner of the local PMU.
3318          */
3319 
3320         for (i = 0; i < count; i++, req++) {
3321 
3322                 cnum        = req->reg_num;
3323                 reg_flags   = req->reg_flags;
3324 
3325                 if (unlikely(!PMD_IS_IMPL(cnum))) goto error;
3326                 /*
3327                  * we can only read the register that we use. That includes
3328                  * the one we explicitly initialize AND the one we want included
3329                  * in the sampling buffer (smpl_regs).
3330                  *
3331                  * Having this restriction allows optimization in the ctxsw routine
3332                  * without compromising security (leaks)
3333                  */
3334                 if (unlikely(!CTX_IS_USED_PMD(ctx, cnum))) goto error;
3335 
3336                 sval        = ctx->ctx_pmds[cnum].val;
3337                 lval        = ctx->ctx_pmds[cnum].lval;
3338                 is_counting = PMD_IS_COUNTING(cnum);
3339 
3340                 /*
3341                  * If the task is not the current one, then we check if the
3342                  * PMU state is still in the local live register due to lazy ctxsw.
3343                  * If true, then we read directly from the registers.
3344                  */
3345                 if (can_access_pmu){
3346                         val = ia64_get_pmd(cnum);
3347                 } else {
3348                         /*
3349                          * context has been saved
3350                          * if context is zombie, then task does not exist anymore.
3351                          * In this case, we use the full value saved in the context (pfm_flush_regs()).
3352                          */
3353                         val = is_loaded ? ctx->th_pmds[cnum] : 0UL;
3354                 }
3355                 rd_func = pmu_conf->pmd_desc[cnum].read_check;
3356 
3357                 if (is_counting) {
3358                         /*
3359                          * XXX: need to check for overflow when loaded
3360                          */
3361                         val &= ovfl_mask;
3362                         val += sval;
3363                 }
3364 
3365                 /*
3366                  * execute read checker, if any
3367                  */
3368                 if (unlikely(expert_mode == 0 && rd_func)) {
3369                         unsigned long v = val;
3370                         ret = (*rd_func)(ctx->ctx_task, ctx, cnum, &v, regs);
3371                         if (ret) goto error;
3372                         val = v;
3373                         ret = -EINVAL;
3374                 }
3375 
3376                 PFM_REG_RETFLAG_SET(reg_flags, 0);
3377 
3378                 DPRINT(("pmd[%u]=0x%lx\n", cnum, val));
3379 
3380                 /*
3381                  * update register return value, abort all if problem during copy.
3382                  * we only modify the reg_flags field. no check mode is fine because
3383                  * access has been verified upfront in sys_perfmonctl().
3384                  */
3385                 req->reg_value            = val;
3386                 req->reg_flags            = reg_flags;
3387                 req->reg_last_reset_val   = lval;
3388         }
3389 
3390         return 0;
3391 
3392 error:
3393         PFM_REG_RETFLAG_SET(req->reg_flags, PFM_REG_RETFL_EINVAL);
3394         return ret;
3395 }
3396 
3397 int
3398 pfm_mod_write_pmcs(struct task_struct *task, void *req, unsigned int nreq, struct pt_regs *regs)
3399 {
3400         pfm_context_t *ctx;
3401 
3402         if (req == NULL) return -EINVAL;
3403 
3404         ctx = GET_PMU_CTX();
3405 
3406         if (ctx == NULL) return -EINVAL;
3407 
3408         /*
3409          * for now limit to current task, which is enough when calling
3410          * from overflow handler
3411          */
3412         if (task != current && ctx->ctx_fl_system == 0) return -EBUSY;
3413 
3414         return pfm_write_pmcs(ctx, req, nreq, regs);
3415 }
3416 EXPORT_SYMBOL(pfm_mod_write_pmcs);
3417 
3418 int
3419 pfm_mod_read_pmds(struct task_struct *task, void *req, unsigned int nreq, struct pt_regs *regs)
3420 {
3421         pfm_context_t *ctx;
3422 
3423         if (req == NULL) return -EINVAL;
3424 
3425         ctx = GET_PMU_CTX();
3426 
3427         if (ctx == NULL) return -EINVAL;
3428 
3429         /*
3430          * for now limit to current task, which is enough when calling
3431          * from overflow handler
3432          */
3433         if (task != current && ctx->ctx_fl_system == 0) return -EBUSY;
3434 
3435         return pfm_read_pmds(ctx, req, nreq, regs);
3436 }
3437 EXPORT_SYMBOL(pfm_mod_read_pmds);
3438 
3439 /*
3440  * Only call this function when a process it trying to
3441  * write the debug registers (reading is always allowed)
3442  */
3443 int
3444 pfm_use_debug_registers(struct task_struct *task)
3445 {
3446         pfm_context_t *ctx = task->thread.pfm_context;
3447         unsigned long flags;
3448         int ret = 0;
3449 
3450         if (pmu_conf->use_rr_dbregs == 0) return 0;
3451 
3452         DPRINT(("called for [%d]\n", task_pid_nr(task)));
3453 
3454         /*
3455          * do it only once
3456          */
3457         if (task->thread.flags & IA64_THREAD_DBG_VALID) return 0;
3458 
3459         /*
3460          * Even on SMP, we do not need to use an atomic here because
3461          * the only way in is via ptrace() and this is possible only when the
3462          * process is stopped. Even in the case where the ctxsw out is not totally
3463          * completed by the time we come here, there is no way the 'stopped' process
3464          * could be in the middle of fiddling with the pfm_write_ibr_dbr() routine.
3465          * So this is always safe.
3466          */
3467         if (ctx && ctx->ctx_fl_using_dbreg == 1) return -1;
3468 
3469         LOCK_PFS(flags);
3470 
3471         /*
3472          * We cannot allow setting breakpoints when system wide monitoring
3473          * sessions are using the debug registers.
3474          */
3475         if (pfm_sessions.pfs_sys_use_dbregs> 0)
3476                 ret = -1;
3477         else
3478                 pfm_sessions.pfs_ptrace_use_dbregs++;
3479 
3480         DPRINT(("ptrace_use_dbregs=%u  sys_use_dbregs=%u by [%d] ret = %d\n",
3481                   pfm_sessions.pfs_ptrace_use_dbregs,
3482                   pfm_sessions.pfs_sys_use_dbregs,
3483                   task_pid_nr(task), ret));
3484 
3485         UNLOCK_PFS(flags);
3486 
3487         return ret;
3488 }
3489 
3490 /*
3491  * This function is called for every task that exits with the
3492  * IA64_THREAD_DBG_VALID set. This indicates a task which was
3493  * able to use the debug registers for debugging purposes via
3494  * ptrace(). Therefore we know it was not using them for
3495  * performance monitoring, so we only decrement the number
3496  * of "ptraced" debug register users to keep the count up to date
3497  */
3498 int
3499 pfm_release_debug_registers(struct task_struct *task)
3500 {
3501         unsigned long flags;
3502         int ret;
3503 
3504         if (pmu_conf->use_rr_dbregs == 0) return 0;
3505 
3506         LOCK_PFS(flags);
3507         if (pfm_sessions.pfs_ptrace_use_dbregs == 0) {
3508                 printk(KERN_ERR "perfmon: invalid release for [%d] ptrace_use_dbregs=0\n", task_pid_nr(task));
3509                 ret = -1;
3510         }  else {
3511                 pfm_sessions.pfs_ptrace_use_dbregs--;
3512                 ret = 0;
3513         }
3514         UNLOCK_PFS(flags);
3515 
3516         return ret;
3517 }
3518 
3519 static int
3520 pfm_restart(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3521 {
3522         struct task_struct *task;
3523         pfm_buffer_fmt_t *fmt;
3524         pfm_ovfl_ctrl_t rst_ctrl;
3525         int state, is_system;
3526         int ret = 0;
3527 
3528         state     = ctx->ctx_state;
3529         fmt       = ctx->ctx_buf_fmt;
3530         is_system = ctx->ctx_fl_system;
3531         task      = PFM_CTX_TASK(ctx);
3532 
3533         switch(state) {
3534                 case PFM_CTX_MASKED:
3535                         break;
3536                 case PFM_CTX_LOADED: 
3537                         if (CTX_HAS_SMPL(ctx) && fmt->fmt_restart_active) break;
3538                         /* fall through */
3539                 case PFM_CTX_UNLOADED:
3540                 case PFM_CTX_ZOMBIE:
3541                         DPRINT(("invalid state=%d\n", state));
3542                         return -EBUSY;
3543                 default:
3544                         DPRINT(("state=%d, cannot operate (no active_restart handler)\n", state));
3545                         return -EINVAL;
3546         }
3547 
3548         /*
3549          * In system wide and when the context is loaded, access can only happen
3550          * when the caller is running on the CPU being monitored by the session.
3551          * It does not have to be the owner (ctx_task) of the context per se.
3552          */
3553         if (is_system && ctx->ctx_cpu != smp_processor_id()) {
3554                 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
3555                 return -EBUSY;
3556         }
3557 
3558         /* sanity check */
3559         if (unlikely(task == NULL)) {
3560                 printk(KERN_ERR "perfmon: [%d] pfm_restart no task\n", task_pid_nr(current));
3561                 return -EINVAL;
3562         }
3563 
3564         if (task == current || is_system) {
3565 
3566                 fmt = ctx->ctx_buf_fmt;
3567 
3568                 DPRINT(("restarting self %d ovfl=0x%lx\n",
3569                         task_pid_nr(task),
3570                         ctx->ctx_ovfl_regs[0]));
3571 
3572                 if (CTX_HAS_SMPL(ctx)) {
3573 
3574                         prefetch(ctx->ctx_smpl_hdr);
3575 
3576                         rst_ctrl.bits.mask_monitoring = 0;
3577                         rst_ctrl.bits.reset_ovfl_pmds = 0;
3578 
3579                         if (state == PFM_CTX_LOADED)
3580                                 ret = pfm_buf_fmt_restart_active(fmt, task, &rst_ctrl, ctx->ctx_smpl_hdr, regs);
3581                         else
3582                                 ret = pfm_buf_fmt_restart(fmt, task, &rst_ctrl, ctx->ctx_smpl_hdr, regs);
3583                 } else {
3584                         rst_ctrl.bits.mask_monitoring = 0;
3585                         rst_ctrl.bits.reset_ovfl_pmds = 1;
3586                 }
3587 
3588                 if (ret == 0) {
3589                         if (rst_ctrl.bits.reset_ovfl_pmds)
3590                                 pfm_reset_regs(ctx, ctx->ctx_ovfl_regs, PFM_PMD_LONG_RESET);
3591 
3592                         if (rst_ctrl.bits.mask_monitoring == 0) {
3593                                 DPRINT(("resuming monitoring for [%d]\n", task_pid_nr(task)));
3594 
3595                                 if (state == PFM_CTX_MASKED) pfm_restore_monitoring(task);
3596                         } else {
3597                                 DPRINT(("keeping monitoring stopped for [%d]\n", task_pid_nr(task)));
3598 
3599                                 // cannot use pfm_stop_monitoring(task, regs);
3600                         }
3601                 }
3602                 /*
3603                  * clear overflowed PMD mask to remove any stale information
3604                  */
3605                 ctx->ctx_ovfl_regs[0] = 0UL;
3606 
3607                 /*
3608                  * back to LOADED state
3609                  */
3610                 ctx->ctx_state = PFM_CTX_LOADED;
3611 
3612                 /*
3613                  * XXX: not really useful for self monitoring
3614                  */
3615                 ctx->ctx_fl_can_restart = 0;
3616 
3617                 return 0;
3618         }
3619 
3620         /* 
3621          * restart another task
3622          */
3623 
3624         /*
3625          * When PFM_CTX_MASKED, we cannot issue a restart before the previous 
3626          * one is seen by the task.
3627          */
3628         if (state == PFM_CTX_MASKED) {
3629                 if (ctx->ctx_fl_can_restart == 0) return -EINVAL;
3630                 /*
3631                  * will prevent subsequent restart before this one is
3632                  * seen by other task
3633                  */
3634                 ctx->ctx_fl_can_restart = 0;
3635         }
3636 
3637         /*
3638          * if blocking, then post the semaphore is PFM_CTX_MASKED, i.e.
3639          * the task is blocked or on its way to block. That's the normal
3640          * restart path. If the monitoring is not masked, then the task
3641          * can be actively monitoring and we cannot directly intervene.
3642          * Therefore we use the trap mechanism to catch the task and
3643          * force it to reset the buffer/reset PMDs.
3644          *
3645          * if non-blocking, then we ensure that the task will go into
3646          * pfm_handle_work() before returning to user mode.
3647          *
3648          * We cannot explicitly reset another task, it MUST always
3649          * be done by the task itself. This works for system wide because
3650          * the tool that is controlling the session is logically doing 
3651          * "self-monitoring".
3652          */
3653         if (CTX_OVFL_NOBLOCK(ctx) == 0 && state == PFM_CTX_MASKED) {
3654                 DPRINT(("unblocking [%d]\n", task_pid_nr(task)));
3655                 complete(&ctx->ctx_restart_done);
3656         } else {
3657                 DPRINT(("[%d] armed exit trap\n", task_pid_nr(task)));
3658 
3659                 ctx->ctx_fl_trap_reason = PFM_TRAP_REASON_RESET;
3660 
3661                 PFM_SET_WORK_PENDING(task, 1);
3662 
3663                 set_notify_resume(task);
3664 
3665                 /*
3666                  * XXX: send reschedule if task runs on another CPU
3667                  */
3668         }
3669         return 0;
3670 }
3671 
3672 static int
3673 pfm_debug(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3674 {
3675         unsigned int m = *(unsigned int *)arg;
3676 
3677         pfm_sysctl.debug = m == 0 ? 0 : 1;
3678 
3679         printk(KERN_INFO "perfmon debugging %s (timing reset)\n", pfm_sysctl.debug ? "on" : "off");
3680 
3681         if (m == 0) {
3682                 memset(pfm_stats, 0, sizeof(pfm_stats));
3683                 for(m=0; m < NR_CPUS; m++) pfm_stats[m].pfm_ovfl_intr_cycles_min = ~0UL;
3684         }
3685         return 0;
3686 }
3687 
3688 /*
3689  * arg can be NULL and count can be zero for this function
3690  */
3691 static int
3692 pfm_write_ibr_dbr(int mode, pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3693 {
3694         struct thread_struct *thread = NULL;
3695         struct task_struct *task;
3696         pfarg_dbreg_t *req = (pfarg_dbreg_t *)arg;
3697         unsigned long flags;
3698         dbreg_t dbreg;
3699         unsigned int rnum;
3700         int first_time;
3701         int ret = 0, state;
3702         int i, can_access_pmu = 0;
3703         int is_system, is_loaded;
3704 
3705         if (pmu_conf->use_rr_dbregs == 0) return -EINVAL;
3706 
3707         state     = ctx->ctx_state;
3708         is_loaded = state == PFM_CTX_LOADED ? 1 : 0;
3709         is_system = ctx->ctx_fl_system;
3710         task      = ctx->ctx_task;
3711 
3712         if (state == PFM_CTX_ZOMBIE) return -EINVAL;
3713 
3714         /*
3715          * on both UP and SMP, we can only write to the PMC when the task is
3716          * the owner of the local PMU.
3717          */
3718         if (is_loaded) {
3719                 thread = &task->thread;
3720                 /*
3721                  * In system wide and when the context is loaded, access can only happen
3722                  * when the caller is running on the CPU being monitored by the session.
3723                  * It does not have to be the owner (ctx_task) of the context per se.
3724                  */
3725                 if (unlikely(is_system && ctx->ctx_cpu != smp_processor_id())) {
3726                         DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
3727                         return -EBUSY;
3728                 }
3729                 can_access_pmu = GET_PMU_OWNER() == task || is_system ? 1 : 0;
3730         }
3731 
3732         /*
3733          * we do not need to check for ipsr.db because we do clear ibr.x, dbr.r, and dbr.w
3734          * ensuring that no real breakpoint can be installed via this call.
3735          *
3736          * IMPORTANT: regs can be NULL in this function
3737          */
3738 
3739         first_time = ctx->ctx_fl_using_dbreg == 0;
3740 
3741         /*
3742          * don't bother if we are loaded and task is being debugged
3743          */
3744         if (is_loaded && (thread->flags & IA64_THREAD_DBG_VALID) != 0) {
3745                 DPRINT(("debug registers already in use for [%d]\n", task_pid_nr(task)));
3746                 return -EBUSY;
3747         }
3748 
3749         /*
3750          * check for debug registers in system wide mode
3751          *
3752          * If though a check is done in pfm_context_load(),
3753          * we must repeat it here, in case the registers are
3754          * written after the context is loaded
3755          */
3756         if (is_loaded) {
3757                 LOCK_PFS(flags);
3758 
3759                 if (first_time && is_system) {
3760                         if (pfm_sessions.pfs_ptrace_use_dbregs)
3761                                 ret = -EBUSY;
3762                         else
3763                                 pfm_sessions.pfs_sys_use_dbregs++;
3764                 }
3765                 UNLOCK_PFS(flags);
3766         }
3767 
3768         if (ret != 0) return ret;
3769 
3770         /*
3771          * mark ourself as user of the debug registers for
3772          * perfmon purposes.
3773          */
3774         ctx->ctx_fl_using_dbreg = 1;
3775 
3776         /*
3777          * clear hardware registers to make sure we don't
3778          * pick up stale state.
3779          *
3780          * for a system wide session, we do not use
3781          * thread.dbr, thread.ibr because this process
3782          * never leaves the current CPU and the state
3783          * is shared by all processes running on it
3784          */
3785         if (first_time && can_access_pmu) {
3786                 DPRINT(("[%d] clearing ibrs, dbrs\n", task_pid_nr(task)));
3787                 for (i=0; i < pmu_conf->num_ibrs; i++) {
3788                         ia64_set_ibr(i, 0UL);
3789                         ia64_dv_serialize_instruction();
3790                 }
3791                 ia64_srlz_i();
3792                 for (i=0; i < pmu_conf->num_dbrs; i++) {
3793                         ia64_set_dbr(i, 0UL);
3794                         ia64_dv_serialize_data();
3795                 }
3796                 ia64_srlz_d();
3797         }
3798 
3799         /*
3800          * Now install the values into the registers
3801          */
3802         for (i = 0; i < count; i++, req++) {
3803 
3804                 rnum      = req->dbreg_num;
3805                 dbreg.val = req->dbreg_value;
3806 
3807                 ret = -EINVAL;
3808 
3809                 if ((mode == PFM_CODE_RR && rnum >= PFM_NUM_IBRS) || ((mode == PFM_DATA_RR) && rnum >= PFM_NUM_DBRS)) {
3810                         DPRINT(("invalid register %u val=0x%lx mode=%d i=%d count=%d\n",
3811                                   rnum, dbreg.val, mode, i, count));
3812 
3813                         goto abort_mission;
3814                 }
3815 
3816                 /*
3817                  * make sure we do not install enabled breakpoint
3818                  */
3819                 if (rnum & 0x1) {
3820                         if (mode == PFM_CODE_RR)
3821                                 dbreg.ibr.ibr_x = 0;
3822                         else
3823                                 dbreg.dbr.dbr_r = dbreg.dbr.dbr_w = 0;
3824                 }
3825 
3826                 PFM_REG_RETFLAG_SET(req->dbreg_flags, 0);
3827 
3828                 /*
3829                  * Debug registers, just like PMC, can only be modified
3830                  * by a kernel call. Moreover, perfmon() access to those
3831                  * registers are centralized in this routine. The hardware
3832                  * does not modify the value of these registers, therefore,
3833                  * if we save them as they are written, we can avoid having
3834                  * to save them on context switch out. This is made possible
3835                  * by the fact that when perfmon uses debug registers, ptrace()
3836                  * won't be able to modify them concurrently.
3837                  */
3838                 if (mode == PFM_CODE_RR) {
3839                         CTX_USED_IBR(ctx, rnum);
3840 
3841                         if (can_access_pmu) {
3842                                 ia64_set_ibr(rnum, dbreg.val);
3843                                 ia64_dv_serialize_instruction();
3844                         }
3845 
3846                         ctx->ctx_ibrs[rnum] = dbreg.val;
3847 
3848                         DPRINT(("write ibr%u=0x%lx used_ibrs=0x%x ld=%d apmu=%d\n",
3849                                 rnum, dbreg.val, ctx->ctx_used_ibrs[0], is_loaded, can_access_pmu));
3850                 } else {
3851                         CTX_USED_DBR(ctx, rnum);
3852 
3853                         if (can_access_pmu) {
3854                                 ia64_set_dbr(rnum, dbreg.val);
3855                                 ia64_dv_serialize_data();
3856                         }
3857                         ctx->ctx_dbrs[rnum] = dbreg.val;
3858 
3859                         DPRINT(("write dbr%u=0x%lx used_dbrs=0x%x ld=%d apmu=%d\n",
3860                                 rnum, dbreg.val, ctx->ctx_used_dbrs[0], is_loaded, can_access_pmu));
3861                 }
3862         }
3863 
3864         return 0;
3865 
3866 abort_mission:
3867         /*
3868          * in case it was our first attempt, we undo the global modifications
3869          */
3870         if (first_time) {
3871                 LOCK_PFS(flags);
3872                 if (ctx->ctx_fl_system) {
3873                         pfm_sessions.pfs_sys_use_dbregs--;
3874                 }
3875                 UNLOCK_PFS(flags);
3876                 ctx->ctx_fl_using_dbreg = 0;
3877         }
3878         /*
3879          * install error return flag
3880          */
3881         PFM_REG_RETFLAG_SET(req->dbreg_flags, PFM_REG_RETFL_EINVAL);
3882 
3883         return ret;
3884 }
3885 
3886 static int
3887 pfm_write_ibrs(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3888 {
3889         return pfm_write_ibr_dbr(PFM_CODE_RR, ctx, arg, count, regs);
3890 }
3891 
3892 static int
3893 pfm_write_dbrs(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3894 {
3895         return pfm_write_ibr_dbr(PFM_DATA_RR, ctx, arg, count, regs);
3896 }
3897 
3898 int
3899 pfm_mod_write_ibrs(struct task_struct *task, void *req, unsigned int nreq, struct pt_regs *regs)
3900 {
3901         pfm_context_t *ctx;
3902 
3903         if (req == NULL) return -EINVAL;
3904 
3905         ctx = GET_PMU_CTX();
3906 
3907         if (ctx == NULL) return -EINVAL;
3908 
3909         /*
3910          * for now limit to current task, which is enough when calling
3911          * from overflow handler
3912          */
3913         if (task != current && ctx->ctx_fl_system == 0) return -EBUSY;
3914 
3915         return pfm_write_ibrs(ctx, req, nreq, regs);
3916 }
3917 EXPORT_SYMBOL(pfm_mod_write_ibrs);
3918 
3919 int
3920 pfm_mod_write_dbrs(struct task_struct *task, void *req, unsigned int nreq, struct pt_regs *regs)
3921 {
3922         pfm_context_t *ctx;
3923 
3924         if (req == NULL) return -EINVAL;
3925 
3926         ctx = GET_PMU_CTX();
3927 
3928         if (ctx == NULL) return -EINVAL;
3929 
3930         /*
3931          * for now limit to current task, which is enough when calling
3932          * from overflow handler
3933          */
3934         if (task != current && ctx->ctx_fl_system == 0) return -EBUSY;
3935 
3936         return pfm_write_dbrs(ctx, req, nreq, regs);
3937 }
3938 EXPORT_SYMBOL(pfm_mod_write_dbrs);
3939 
3940 
3941 static int
3942 pfm_get_features(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3943 {
3944         pfarg_features_t *req = (pfarg_features_t *)arg;
3945 
3946         req->ft_version = PFM_VERSION;
3947         return 0;
3948 }
3949 
3950 static int
3951 pfm_stop(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
3952 {
3953         struct pt_regs *tregs;
3954         struct task_struct *task = PFM_CTX_TASK(ctx);
3955         int state, is_system;
3956 
3957         state     = ctx->ctx_state;
3958         is_system = ctx->ctx_fl_system;
3959 
3960         /*
3961          * context must be attached to issue the stop command (includes LOADED,MASKED,ZOMBIE)
3962          */
3963         if (state == PFM_CTX_UNLOADED) return -EINVAL;
3964 
3965         /*
3966          * In system wide and when the context is loaded, access can only happen
3967          * when the caller is running on the CPU being monitored by the session.
3968          * It does not have to be the owner (ctx_task) of the context per se.
3969          */
3970         if (is_system && ctx->ctx_cpu != smp_processor_id()) {
3971                 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
3972                 return -EBUSY;
3973         }
3974         DPRINT(("task [%d] ctx_state=%d is_system=%d\n",
3975                 task_pid_nr(PFM_CTX_TASK(ctx)),
3976                 state,
3977                 is_system));
3978         /*
3979          * in system mode, we need to update the PMU directly
3980          * and the user level state of the caller, which may not
3981          * necessarily be the creator of the context.
3982          */
3983         if (is_system) {
3984                 /*
3985                  * Update local PMU first
3986                  *
3987                  * disable dcr pp
3988                  */
3989                 ia64_setreg(_IA64_REG_CR_DCR, ia64_getreg(_IA64_REG_CR_DCR) & ~IA64_DCR_PP);
3990                 ia64_srlz_i();
3991 
3992                 /*
3993                  * update local cpuinfo
3994                  */
3995                 PFM_CPUINFO_CLEAR(PFM_CPUINFO_DCR_PP);
3996 
3997                 /*
3998                  * stop monitoring, does srlz.i
3999                  */
4000                 pfm_clear_psr_pp();
4001 
4002                 /*
4003                  * stop monitoring in the caller
4004                  */
4005                 ia64_psr(regs)->pp = 0;
4006 
4007                 return 0;
4008         }
4009         /*
4010          * per-task mode
4011          */
4012 
4013         if (task == current) {
4014                 /* stop monitoring  at kernel level */
4015                 pfm_clear_psr_up();
4016 
4017                 /*
4018                  * stop monitoring at the user level
4019                  */
4020                 ia64_psr(regs)->up = 0;
4021         } else {
4022                 tregs = task_pt_regs(task);
4023 
4024                 /*
4025                  * stop monitoring at the user level
4026                  */
4027                 ia64_psr(tregs)->up = 0;
4028 
4029                 /*
4030                  * monitoring disabled in kernel at next reschedule
4031                  */
4032                 ctx->ctx_saved_psr_up = 0;
4033                 DPRINT(("task=[%d]\n", task_pid_nr(task)));
4034         }
4035         return 0;
4036 }
4037 
4038 
4039 static int
4040 pfm_start(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
4041 {
4042         struct pt_regs *tregs;
4043         int state, is_system;
4044 
4045         state     = ctx->ctx_state;
4046         is_system = ctx->ctx_fl_system;
4047 
4048         if (state != PFM_CTX_LOADED) return -EINVAL;
4049 
4050         /*
4051          * In system wide and when the context is loaded, access can only happen
4052          * when the caller is running on the CPU being monitored by the session.
4053          * It does not have to be the owner (ctx_task) of the context per se.
4054          */
4055         if (is_system && ctx->ctx_cpu != smp_processor_id()) {
4056                 DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
4057                 return -EBUSY;
4058         }
4059 
4060         /*
4061          * in system mode, we need to update the PMU directly
4062          * and the user level state of the caller, which may not
4063          * necessarily be the creator of the context.
4064          */
4065         if (is_system) {
4066 
4067                 /*
4068                  * set user level psr.pp for the caller
4069                  */
4070                 ia64_psr(regs)->pp = 1;
4071 
4072                 /*
4073                  * now update the local PMU and cpuinfo
4074                  */
4075                 PFM_CPUINFO_SET(PFM_CPUINFO_DCR_PP);
4076 
4077                 /*
4078                  * start monitoring at kernel level
4079                  */
4080                 pfm_set_psr_pp();
4081 
4082                 /* enable dcr pp */
4083                 ia64_setreg(_IA64_REG_CR_DCR, ia64_getreg(_IA64_REG_CR_DCR) | IA64_DCR_PP);
4084                 ia64_srlz_i();
4085 
4086                 return 0;
4087         }
4088 
4089         /*
4090          * per-process mode
4091          */
4092 
4093         if (ctx->ctx_task == current) {
4094 
4095                 /* start monitoring at kernel level */
4096                 pfm_set_psr_up();
4097 
4098                 /*
4099                  * activate monitoring at user level
4100                  */
4101                 ia64_psr(regs)->up = 1;
4102 
4103         } else {
4104                 tregs = task_pt_regs(ctx->ctx_task);
4105 
4106                 /*
4107                  * start monitoring at the kernel level the next
4108                  * time the task is scheduled
4109                  */
4110                 ctx->ctx_saved_psr_up = IA64_PSR_UP;
4111 
4112                 /*
4113                  * activate monitoring at user level
4114                  */
4115                 ia64_psr(tregs)->up = 1;
4116         }
4117         return 0;
4118 }
4119 
4120 static int
4121 pfm_get_pmc_reset(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
4122 {
4123         pfarg_reg_t *req = (pfarg_reg_t *)arg;
4124         unsigned int cnum;
4125         int i;
4126         int ret = -EINVAL;
4127 
4128         for (i = 0; i < count; i++, req++) {
4129 
4130                 cnum = req->reg_num;
4131 
4132                 if (!PMC_IS_IMPL(cnum)) goto abort_mission;
4133 
4134                 req->reg_value = PMC_DFL_VAL(cnum);
4135 
4136                 PFM_REG_RETFLAG_SET(req->reg_flags, 0);
4137 
4138                 DPRINT(("pmc_reset_val pmc[%u]=0x%lx\n", cnum, req->reg_value));
4139         }
4140         return 0;
4141 
4142 abort_mission:
4143         PFM_REG_RETFLAG_SET(req->reg_flags, PFM_REG_RETFL_EINVAL);
4144         return ret;
4145 }
4146 
4147 static int
4148 pfm_check_task_exist(pfm_context_t *ctx)
4149 {
4150         struct task_struct *g, *t;
4151         int ret = -ESRCH;
4152 
4153         read_lock(&tasklist_lock);
4154 
4155         do_each_thread (g, t) {
4156                 if (t->thread.pfm_context == ctx) {
4157                         ret = 0;
4158                         goto out;
4159                 }
4160         } while_each_thread (g, t);
4161 out:
4162         read_unlock(&tasklist_lock);
4163 
4164         DPRINT(("pfm_check_task_exist: ret=%d ctx=%p\n", ret, ctx));
4165 
4166         return ret;
4167 }
4168 
4169 static int
4170 pfm_context_load(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
4171 {
4172         struct task_struct *task;
4173         struct thread_struct *thread;
4174         struct pfm_context_t *old;
4175         unsigned long flags;
4176 #ifndef CONFIG_SMP
4177         struct task_struct *owner_task = NULL;
4178 #endif
4179         pfarg_load_t *req = (pfarg_load_t *)arg;
4180         unsigned long *pmcs_source, *pmds_source;
4181         int the_cpu;
4182         int ret = 0;
4183         int state, is_system, set_dbregs = 0;
4184 
4185         state     = ctx->ctx_state;
4186         is_system = ctx->ctx_fl_system;
4187         /*
4188          * can only load from unloaded or terminated state
4189          */
4190         if (state != PFM_CTX_UNLOADED) {
4191                 DPRINT(("cannot load to [%d], invalid ctx_state=%d\n",
4192                         req->load_pid,
4193                         ctx->ctx_state));
4194                 return -EBUSY;
4195         }
4196 
4197         DPRINT(("load_pid [%d] using_dbreg=%d\n", req->load_pid, ctx->ctx_fl_using_dbreg));
4198 
4199         if (CTX_OVFL_NOBLOCK(ctx) == 0 && req->load_pid == current->pid) {
4200                 DPRINT(("cannot use blocking mode on self\n"));
4201                 return -EINVAL;
4202         }
4203 
4204         ret = pfm_get_task(ctx, req->load_pid, &task);
4205         if (ret) {
4206                 DPRINT(("load_pid [%d] get_task=%d\n", req->load_pid, ret));
4207                 return ret;
4208         }
4209 
4210         ret = -EINVAL;
4211 
4212         /*
4213          * system wide is self monitoring only
4214          */
4215         if (is_system && task != current) {
4216                 DPRINT(("system wide is self monitoring only load_pid=%d\n",
4217                         req->load_pid));
4218                 goto error;
4219         }
4220 
4221         thread = &task->thread;
4222 
4223         ret = 0;
4224         /*
4225          * cannot load a context which is using range restrictions,
4226          * into a task that is being debugged.
4227          */
4228         if (ctx->ctx_fl_using_dbreg) {
4229                 if (thread->flags & IA64_THREAD_DBG_VALID) {
4230                         ret = -EBUSY;
4231                         DPRINT(("load_pid [%d] task is debugged, cannot load range restrictions\n", req->load_pid));
4232                         goto error;
4233                 }
4234                 LOCK_PFS(flags);
4235 
4236                 if (is_system) {
4237                         if (pfm_sessions.pfs_ptrace_use_dbregs) {
4238                                 DPRINT(("cannot load [%d] dbregs in use\n",
4239                                                         task_pid_nr(task)));
4240                                 ret = -EBUSY;
4241                         } else {
4242                                 pfm_sessions.pfs_sys_use_dbregs++;
4243                                 DPRINT(("load [%d] increased sys_use_dbreg=%u\n", task_pid_nr(task), pfm_sessions.pfs_sys_use_dbregs));
4244                                 set_dbregs = 1;
4245                         }
4246                 }
4247 
4248                 UNLOCK_PFS(flags);
4249 
4250                 if (ret) goto error;
4251         }
4252 
4253         /*
4254          * SMP system-wide monitoring implies self-monitoring.
4255          *
4256          * The programming model expects the task to
4257          * be pinned on a CPU throughout the session.
4258          * Here we take note of the current CPU at the
4259          * time the context is loaded. No call from
4260          * another CPU will be allowed.
4261          *
4262          * The pinning via shed_setaffinity()
4263          * must be done by the calling task prior
4264          * to this call.
4265          *
4266          * systemwide: keep track of CPU this session is supposed to run on
4267          */
4268         the_cpu = ctx->ctx_cpu = smp_processor_id();
4269 
4270         ret = -EBUSY;
4271         /*
4272          * now reserve the session
4273          */
4274         ret = pfm_reserve_session(current, is_system, the_cpu);
4275         if (ret) goto error;
4276 
4277         /*
4278          * task is necessarily stopped at this point.
4279          *
4280          * If the previous context was zombie, then it got removed in
4281          * pfm_save_regs(). Therefore we should not see it here.
4282          * If we see a context, then this is an active context
4283          *
4284          * XXX: needs to be atomic
4285          */
4286         DPRINT(("before cmpxchg() old_ctx=%p new_ctx=%p\n",
4287                 thread->pfm_context, ctx));
4288 
4289         ret = -EBUSY;
4290         old = ia64_cmpxchg(acq, &thread->pfm_context, NULL, ctx, sizeof(pfm_context_t *));
4291         if (old != NULL) {
4292                 DPRINT(("load_pid [%d] already has a context\n", req->load_pid));
4293                 goto error_unres;
4294         }
4295 
4296         pfm_reset_msgq(ctx);
4297 
4298         ctx->ctx_state = PFM_CTX_LOADED;
4299 
4300         /*
4301          * link context to task
4302          */
4303         ctx->ctx_task = task;
4304 
4305         if (is_system) {
4306                 /*
4307                  * we load as stopped
4308                  */
4309                 PFM_CPUINFO_SET(PFM_CPUINFO_SYST_WIDE);
4310                 PFM_CPUINFO_CLEAR(PFM_CPUINFO_DCR_PP);
4311 
4312                 if (ctx->ctx_fl_excl_idle) PFM_CPUINFO_SET(PFM_CPUINFO_EXCL_IDLE);
4313         } else {
4314                 thread->flags |= IA64_THREAD_PM_VALID;
4315         }
4316 
4317         /*
4318          * propagate into thread-state
4319          */
4320         pfm_copy_pmds(task, ctx);
4321         pfm_copy_pmcs(task, ctx);
4322 
4323         pmcs_source = ctx->th_pmcs;
4324         pmds_source = ctx->th_pmds;
4325 
4326         /*
4327          * always the case for system-wide
4328          */
4329         if (task == current) {
4330 
4331                 if (is_system == 0) {
4332 
4333                         /* allow user level control */
4334                         ia64_psr(regs)->sp = 0;
4335                         DPRINT(("clearing psr.sp for [%d]\n", task_pid_nr(task)));
4336 
4337                         SET_LAST_CPU(ctx, smp_processor_id());
4338                         INC_ACTIVATION();
4339                         SET_ACTIVATION(ctx);
4340 #ifndef CONFIG_SMP
4341                         /*
4342                          * push the other task out, if any
4343                          */
4344                         owner_task = GET_PMU_OWNER();
4345                         if (owner_task) pfm_lazy_save_regs(owner_task);
4346 #endif
4347                 }
4348                 /*
4349                  * load all PMD from ctx to PMU (as opposed to thread state)
4350                  * restore all PMC from ctx to PMU
4351                  */
4352                 pfm_restore_pmds(pmds_source, ctx->ctx_all_pmds[0]);
4353                 pfm_restore_pmcs(pmcs_source, ctx->ctx_all_pmcs[0]);
4354 
4355                 ctx->ctx_reload_pmcs[0] = 0UL;
4356                 ctx->ctx_reload_pmds[0] = 0UL;
4357 
4358                 /*
4359                  * guaranteed safe by earlier check against DBG_VALID
4360                  */
4361                 if (ctx->ctx_fl_using_dbreg) {
4362                         pfm_restore_ibrs(ctx->ctx_ibrs, pmu_conf->num_ibrs);
4363                         pfm_restore_dbrs(ctx->ctx_dbrs, pmu_conf->num_dbrs);
4364                 }
4365                 /*
4366                  * set new ownership
4367                  */
4368                 SET_PMU_OWNER(task, ctx);
4369 
4370                 DPRINT(("context loaded on PMU for [%d]\n", task_pid_nr(task)));
4371         } else {
4372                 /*
4373                  * when not current, task MUST be stopped, so this is safe
4374                  */
4375                 regs = task_pt_regs(task);
4376 
4377                 /* force a full reload */
4378                 ctx->ctx_last_activation = PFM_INVALID_ACTIVATION;
4379                 SET_LAST_CPU(ctx, -1);
4380 
4381                 /* initial saved psr (stopped) */
4382                 ctx->ctx_saved_psr_up = 0UL;
4383                 ia64_psr(regs)->up = ia64_psr(regs)->pp = 0;
4384         }
4385 
4386         ret = 0;
4387 
4388 error_unres:
4389         if (ret) pfm_unreserve_session(ctx, ctx->ctx_fl_system, the_cpu);
4390 error:
4391         /*
4392          * we must undo the dbregs setting (for system-wide)
4393          */
4394         if (ret && set_dbregs) {
4395                 LOCK_PFS(flags);
4396                 pfm_sessions.pfs_sys_use_dbregs--;
4397                 UNLOCK_PFS(flags);
4398         }
4399         /*
4400          * release task, there is now a link with the context
4401          */
4402         if (is_system == 0 && task != current) {
4403                 pfm_put_task(task);
4404 
4405                 if (ret == 0) {
4406                         ret = pfm_check_task_exist(ctx);
4407                         if (ret) {
4408                                 ctx->ctx_state = PFM_CTX_UNLOADED;
4409                                 ctx->ctx_task  = NULL;
4410                         }
4411                 }
4412         }
4413         return ret;
4414 }
4415 
4416 /*
4417  * in this function, we do not need to increase the use count
4418  * for the task via get_task_struct(), because we hold the
4419  * context lock. If the task were to disappear while having
4420  * a context attached, it would go through pfm_exit_thread()
4421  * which also grabs the context lock  and would therefore be blocked
4422  * until we are here.
4423  */
4424 static void pfm_flush_pmds(struct task_struct *, pfm_context_t *ctx);
4425 
4426 static int
4427 pfm_context_unload(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
4428 {
4429         struct task_struct *task = PFM_CTX_TASK(ctx);
4430         struct pt_regs *tregs;
4431         int prev_state, is_system;
4432         int ret;
4433 
4434         DPRINT(("ctx_state=%d task [%d]\n", ctx->ctx_state, task ? task_pid_nr(task) : -1));
4435 
4436         prev_state = ctx->ctx_state;
4437         is_system  = ctx->ctx_fl_system;
4438 
4439         /*
4440          * unload only when necessary
4441          */
4442         if (prev_state == PFM_CTX_UNLOADED) {
4443                 DPRINT(("ctx_state=%d, nothing to do\n", prev_state));
4444                 return 0;
4445         }
4446 
4447         /*
4448          * clear psr and dcr bits
4449          */
4450         ret = pfm_stop(ctx, NULL, 0, regs);
4451         if (ret) return ret;
4452 
4453         ctx->ctx_state = PFM_CTX_UNLOADED;
4454 
4455         /*
4456          * in system mode, we need to update the PMU directly
4457          * and the user level state of the caller, which may not
4458          * necessarily be the creator of the context.
4459          */
4460         if (is_system) {
4461 
4462                 /*
4463                  * Update cpuinfo
4464                  *
4465                  * local PMU is taken care of in pfm_stop()
4466                  */
4467                 PFM_CPUINFO_CLEAR(PFM_CPUINFO_SYST_WIDE);
4468                 PFM_CPUINFO_CLEAR(PFM_CPUINFO_EXCL_IDLE);
4469 
4470                 /*
4471                  * save PMDs in context
4472                  * release ownership
4473                  */
4474                 pfm_flush_pmds(current, ctx);
4475 
4476                 /*
4477                  * at this point we are done with the PMU
4478                  * so we can unreserve the resource.
4479                  */
4480                 if (prev_state != PFM_CTX_ZOMBIE) 
4481                         pfm_unreserve_session(ctx, 1 , ctx->ctx_cpu);
4482 
4483                 /*
4484                  * disconnect context from task
4485                  */
4486                 task->thread.pfm_context = NULL;
4487                 /*
4488                  * disconnect task from context
4489                  */
4490                 ctx->ctx_task = NULL;
4491 
4492                 /*
4493                  * There is nothing more to cleanup here.
4494                  */
4495                 return 0;
4496         }
4497 
4498         /*
4499          * per-task mode
4500          */
4501         tregs = task == current ? regs : task_pt_regs(task);
4502 
4503         if (task == current) {
4504                 /*
4505                  * cancel user level control
4506                  */
4507                 ia64_psr(regs)->sp = 1;
4508 
4509                 DPRINT(("setting psr.sp for [%d]\n", task_pid_nr(task)));
4510         }
4511         /*
4512          * save PMDs to context
4513          * release ownership
4514          */
4515         pfm_flush_pmds(task, ctx);
4516 
4517         /*
4518          * at this point we are done with the PMU
4519          * so we can unreserve the resource.
4520          *
4521          * when state was ZOMBIE, we have already unreserved.
4522          */
4523         if (prev_state != PFM_CTX_ZOMBIE) 
4524                 pfm_unreserve_session(ctx, 0 , ctx->ctx_cpu);
4525 
4526         /*
4527          * reset activation counter and psr
4528          */
4529         ctx->ctx_last_activation = PFM_INVALID_ACTIVATION;
4530         SET_LAST_CPU(ctx, -1);
4531 
4532         /*
4533          * PMU state will not be restored
4534          */
4535         task->thread.flags &= ~IA64_THREAD_PM_VALID;
4536 
4537         /*
4538          * break links between context and task
4539          */
4540         task->thread.pfm_context  = NULL;
4541         ctx->ctx_task             = NULL;
4542 
4543         PFM_SET_WORK_PENDING(task, 0);
4544 
4545         ctx->ctx_fl_trap_reason  = PFM_TRAP_REASON_NONE;
4546         ctx->ctx_fl_can_restart  = 0;
4547         ctx->ctx_fl_going_zombie = 0;
4548 
4549         DPRINT(("disconnected [%d] from context\n", task_pid_nr(task)));
4550 
4551         return 0;
4552 }
4553 
4554 
4555 /*
4556  * called only from exit_thread(): task == current
4557  * we come here only if current has a context attached (loaded or masked)
4558  */
4559 void
4560 pfm_exit_thread(struct task_struct *task)
4561 {
4562         pfm_context_t *ctx;
4563         unsigned long flags;
4564         struct pt_regs *regs = task_pt_regs(task);
4565         int ret, state;
4566         int free_ok = 0;
4567 
4568         ctx = PFM_GET_CTX(task);
4569 
4570         PROTECT_CTX(ctx, flags);
4571 
4572         DPRINT(("state=%d task [%d]\n", ctx->ctx_state, task_pid_nr(task)));
4573 
4574         state = ctx->ctx_state;
4575         switch(state) {
4576                 case PFM_CTX_UNLOADED:
4577                         /*
4578                          * only comes to this function if pfm_context is not NULL, i.e., cannot
4579                          * be in unloaded state
4580                          */
4581                         printk(KERN_ERR "perfmon: pfm_exit_thread [%d] ctx unloaded\n", task_pid_nr(task));
4582                         break;
4583                 case PFM_CTX_LOADED:
4584                 case PFM_CTX_MASKED:
4585                         ret = pfm_context_unload(ctx, NULL, 0, regs);
4586                         if (ret) {
4587                                 printk(KERN_ERR "perfmon: pfm_exit_thread [%d] state=%d unload failed %d\n", task_pid_nr(task), state, ret);
4588                         }
4589                         DPRINT(("ctx unloaded for current state was %d\n", state));
4590 
4591                         pfm_end_notify_user(ctx);
4592                         break;
4593                 case PFM_CTX_ZOMBIE:
4594                         ret = pfm_context_unload(ctx, NULL, 0, regs);
4595                         if (ret) {
4596                                 printk(KERN_ERR "perfmon: pfm_exit_thread [%d] state=%d unload failed %d\n", task_pid_nr(task), state, ret);
4597                         }
4598                         free_ok = 1;
4599                         break;
4600                 default:
4601                         printk(KERN_ERR "perfmon: pfm_exit_thread [%d] unexpected state=%d\n", task_pid_nr(task), state);
4602                         break;
4603         }
4604         UNPROTECT_CTX(ctx, flags);
4605 
4606         { u64 psr = pfm_get_psr();
4607           BUG_ON(psr & (IA64_PSR_UP|IA64_PSR_PP));
4608           BUG_ON(GET_PMU_OWNER());
4609           BUG_ON(ia64_psr(regs)->up);
4610           BUG_ON(ia64_psr(regs)->pp);
4611         }
4612 
4613         /*
4614          * All memory free operations (especially for vmalloc'ed memory)
4615          * MUST be done with interrupts ENABLED.
4616          */
4617         if (free_ok) pfm_context_free(ctx);
4618 }
4619 
4620 /*
4621  * functions MUST be listed in the increasing order of their index (see permfon.h)
4622  */
4623 #define PFM_CMD(name, flags, arg_count, arg_type, getsz) { name, #name, flags, arg_count, sizeof(arg_type), getsz }
4624 #define PFM_CMD_S(name, flags) { name, #name, flags, 0, 0, NULL }
4625 #define PFM_CMD_PCLRWS  (PFM_CMD_FD|PFM_CMD_ARG_RW|PFM_CMD_STOP)
4626 #define PFM_CMD_PCLRW   (PFM_CMD_FD|PFM_CMD_ARG_RW)
4627 #define PFM_CMD_NONE    { NULL, "no-cmd", 0, 0, 0, NULL}
4628 
4629 static pfm_cmd_desc_t pfm_cmd_tab[]={
4630 /* 0  */PFM_CMD_NONE,
4631 /* 1  */PFM_CMD(pfm_write_pmcs, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_reg_t, NULL),
4632 /* 2  */PFM_CMD(pfm_write_pmds, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_reg_t, NULL),
4633 /* 3  */PFM_CMD(pfm_read_pmds, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_reg_t, NULL),
4634 /* 4  */PFM_CMD_S(pfm_stop, PFM_CMD_PCLRWS),
4635 /* 5  */PFM_CMD_S(pfm_start, PFM_CMD_PCLRWS),
4636 /* 6  */PFM_CMD_NONE,
4637 /* 7  */PFM_CMD_NONE,
4638 /* 8  */PFM_CMD(pfm_context_create, PFM_CMD_ARG_RW, 1, pfarg_context_t, pfm_ctx_getsize),
4639 /* 9  */PFM_CMD_NONE,
4640 /* 10 */PFM_CMD_S(pfm_restart, PFM_CMD_PCLRW),
4641 /* 11 */PFM_CMD_NONE,
4642 /* 12 */PFM_CMD(pfm_get_features, PFM_CMD_ARG_RW, 1, pfarg_features_t, NULL),
4643 /* 13 */PFM_CMD(pfm_debug, 0, 1, unsigned int, NULL),
4644 /* 14 */PFM_CMD_NONE,
4645 /* 15 */PFM_CMD(pfm_get_pmc_reset, PFM_CMD_ARG_RW, PFM_CMD_ARG_MANY, pfarg_reg_t, NULL),
4646 /* 16 */PFM_CMD(pfm_context_load, PFM_CMD_PCLRWS, 1, pfarg_load_t, NULL),
4647 /* 17 */PFM_CMD_S(pfm_context_unload, PFM_CMD_PCLRWS),
4648 /* 18 */PFM_CMD_NONE,
4649 /* 19 */PFM_CMD_NONE,
4650 /* 20 */PFM_CMD_NONE,
4651 /* 21 */PFM_CMD_NONE,
4652 /* 22 */PFM_CMD_NONE,
4653 /* 23 */PFM_CMD_NONE,
4654 /* 24 */PFM_CMD_NONE,
4655 /* 25 */PFM_CMD_NONE,
4656 /* 26 */PFM_CMD_NONE,
4657 /* 27 */PFM_CMD_NONE,
4658 /* 28 */PFM_CMD_NONE,
4659 /* 29 */PFM_CMD_NONE,
4660 /* 30 */PFM_CMD_NONE,
4661 /* 31 */PFM_CMD_NONE,
4662 /* 32 */PFM_CMD(pfm_write_ibrs, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_dbreg_t, NULL),
4663 /* 33 */PFM_CMD(pfm_write_dbrs, PFM_CMD_PCLRWS, PFM_CMD_ARG_MANY, pfarg_dbreg_t, NULL)
4664 };
4665 #define PFM_CMD_COUNT   (sizeof(pfm_cmd_tab)/sizeof(pfm_cmd_desc_t))
4666 
4667 static int
4668 pfm_check_task_state(pfm_context_t *ctx, int cmd, unsigned long flags)
4669 {
4670         struct task_struct *task;
4671         int state, old_state;
4672 
4673 recheck:
4674         state = ctx->ctx_state;
4675         task  = ctx->ctx_task;
4676 
4677         if (task == NULL) {
4678                 DPRINT(("context %d no task, state=%d\n", ctx->ctx_fd, state));
4679                 return 0;
4680         }
4681 
4682         DPRINT(("context %d state=%d [%d] task_state=%ld must_stop=%d\n",
4683                 ctx->ctx_fd,
4684                 state,
4685                 task_pid_nr(task),
4686                 task->state, PFM_CMD_STOPPED(cmd)));
4687 
4688         /*
4689          * self-monitoring always ok.
4690          *
4691          * for system-wide the caller can either be the creator of the
4692          * context (to one to which the context is attached to) OR
4693          * a task running on the same CPU as the session.
4694          */
4695         if (task == current || ctx->ctx_fl_system) return 0;
4696 
4697         /*
4698          * we are monitoring another thread
4699          */
4700         switch(state) {
4701                 case PFM_CTX_UNLOADED:
4702                         /*
4703                          * if context is UNLOADED we are safe to go
4704                          */
4705                         return 0;
4706                 case PFM_CTX_ZOMBIE:
4707                         /*
4708                          * no command can operate on a zombie context
4709                          */
4710                         DPRINT(("cmd %d state zombie cannot operate on context\n", cmd));
4711                         return -EINVAL;
4712                 case PFM_CTX_MASKED:
4713                         /*
4714                          * PMU state has been saved to software even though
4715                          * the thread may still be running.
4716                          */
4717                         if (cmd != PFM_UNLOAD_CONTEXT) return 0;
4718         }
4719 
4720         /*
4721          * context is LOADED or MASKED. Some commands may need to have 
4722          * the task stopped.
4723          *
4724          * We could lift this restriction for UP but it would mean that
4725          * the user has no guarantee the task would not run between
4726          * two successive calls to perfmonctl(). That's probably OK.
4727          * If this user wants to ensure the task does not run, then
4728          * the task must be stopped.
4729          */
4730         if (PFM_CMD_STOPPED(cmd)) {
4731                 if (!task_is_stopped_or_traced(task)) {
4732                         DPRINT(("[%d] task not in stopped state\n", task_pid_nr(task)));
4733                         return -EBUSY;
4734                 }
4735                 /*
4736                  * task is now stopped, wait for ctxsw out
4737                  *
4738                  * This is an interesting point in the code.
4739                  * We need to unprotect the context because
4740                  * the pfm_save_regs() routines needs to grab
4741                  * the same lock. There are danger in doing
4742                  * this because it leaves a window open for
4743                  * another task to get access to the context
4744                  * and possibly change its state. The one thing
4745                  * that is not possible is for the context to disappear
4746                  * because we are protected by the VFS layer, i.e.,
4747                  * get_fd()/put_fd().
4748                  */
4749                 old_state = state;
4750 
4751                 UNPROTECT_CTX(ctx, flags);
4752 
4753                 wait_task_inactive(task, 0);
4754 
4755                 PROTECT_CTX(ctx, flags);
4756 
4757                 /*
4758                  * we must recheck to verify if state has changed
4759                  */
4760                 if (ctx->ctx_state != old_state) {
4761                         DPRINT(("old_state=%d new_state=%d\n", old_state, ctx->ctx_state));
4762                         goto recheck;
4763                 }
4764         }
4765         return 0;
4766 }
4767 
4768 /*
4769  * system-call entry point (must return long)
4770  */
4771 asmlinkage long
4772 sys_perfmonctl (int fd, int cmd, void __user *arg, int count)
4773 {
4774         struct fd f = {NULL, 0};
4775         pfm_context_t *ctx = NULL;
4776         unsigned long flags = 0UL;
4777         void *args_k = NULL;
4778         long ret; /* will expand int return types */
4779         size_t base_sz, sz, xtra_sz = 0;
4780         int narg, completed_args = 0, call_made = 0, cmd_flags;
4781         int (*func)(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs);
4782         int (*getsize)(void *arg, size_t *sz);
4783 #define PFM_MAX_ARGSIZE 4096
4784 
4785         /*
4786          * reject any call if perfmon was disabled at initialization
4787          */
4788         if (unlikely(pmu_conf == NULL)) return -ENOSYS;
4789 
4790         if (unlikely(cmd < 0 || cmd >= PFM_CMD_COUNT)) {
4791                 DPRINT(("invalid cmd=%d\n", cmd));
4792                 return -EINVAL;
4793         }
4794 
4795         func      = pfm_cmd_tab[cmd].cmd_func;
4796         narg      = pfm_cmd_tab[cmd].cmd_narg;
4797         base_sz   = pfm_cmd_tab[cmd].cmd_argsize;
4798         getsize   = pfm_cmd_tab[cmd].cmd_getsize;
4799         cmd_flags = pfm_cmd_tab[cmd].cmd_flags;
4800 
4801         if (unlikely(func == NULL)) {
4802                 DPRINT(("invalid cmd=%d\n", cmd));
4803                 return -EINVAL;
4804         }
4805 
4806         DPRINT(("cmd=%s idx=%d narg=0x%x argsz=%lu count=%d\n",
4807                 PFM_CMD_NAME(cmd),
4808                 cmd,
4809                 narg,
4810                 base_sz,
4811                 count));
4812 
4813         /*
4814          * check if number of arguments matches what the command expects
4815          */
4816         if (unlikely((narg == PFM_CMD_ARG_MANY && count <= 0) || (narg > 0 && narg != count)))
4817                 return -EINVAL;
4818 
4819 restart_args:
4820         sz = xtra_sz + base_sz*count;
4821         /*
4822          * limit abuse to min page size
4823          */
4824         if (unlikely(sz > PFM_MAX_ARGSIZE)) {
4825                 printk(KERN_ERR "perfmon: [%d] argument too big %lu\n", task_pid_nr(current), sz);
4826                 return -E2BIG;
4827         }
4828 
4829         /*
4830          * allocate default-sized argument buffer
4831          */
4832         if (likely(count && args_k == NULL)) {
4833                 args_k = kmalloc(PFM_MAX_ARGSIZE, GFP_KERNEL);
4834                 if (args_k == NULL) return -ENOMEM;
4835         }
4836 
4837         ret = -EFAULT;
4838 
4839         /*
4840          * copy arguments
4841          *
4842          * assume sz = 0 for command without parameters
4843          */
4844         if (sz && copy_from_user(args_k, arg, sz)) {
4845                 DPRINT(("cannot copy_from_user %lu bytes @%p\n", sz, arg));
4846                 goto error_args;
4847         }
4848 
4849         /*
4850          * check if command supports extra parameters
4851          */
4852         if (completed_args == 0 && getsize) {
4853                 /*
4854                  * get extra parameters size (based on main argument)
4855                  */
4856                 ret = (*getsize)(args_k, &xtra_sz);
4857                 if (ret) goto error_args;
4858 
4859                 completed_args = 1;
4860 
4861                 DPRINT(("restart_args sz=%lu xtra_sz=%lu\n", sz, xtra_sz));
4862 
4863                 /* retry if necessary */
4864                 if (likely(xtra_sz)) goto restart_args;
4865         }
4866 
4867         if (unlikely((cmd_flags & PFM_CMD_FD) == 0)) goto skip_fd;
4868 
4869         ret = -EBADF;
4870 
4871         f = fdget(fd);
4872         if (unlikely(f.file == NULL)) {
4873                 DPRINT(("invalid fd %d\n", fd));
4874                 goto error_args;
4875         }
4876         if (unlikely(PFM_IS_FILE(f.file) == 0)) {
4877                 DPRINT(("fd %d not related to perfmon\n", fd));
4878                 goto error_args;
4879         }
4880 
4881         ctx = f.file->private_data;
4882         if (unlikely(ctx == NULL)) {
4883                 DPRINT(("no context for fd %d\n", fd));
4884                 goto error_args;
4885         }
4886         prefetch(&ctx->ctx_state);
4887 
4888         PROTECT_CTX(ctx, flags);
4889 
4890         /*
4891          * check task is stopped
4892          */
4893         ret = pfm_check_task_state(ctx, cmd, flags);
4894         if (unlikely(ret)) goto abort_locked;
4895 
4896 skip_fd:
4897         ret = (*func)(ctx, args_k, count, task_pt_regs(current));
4898 
4899         call_made = 1;
4900 
4901 abort_locked:
4902         if (likely(ctx)) {
4903                 DPRINT(("context unlocked\n"));
4904                 UNPROTECT_CTX(ctx, flags);
4905         }
4906 
4907         /* copy argument back to user, if needed */
4908         if (call_made && PFM_CMD_RW_ARG(cmd) && copy_to_user(arg, args_k, base_sz*count)) ret = -EFAULT;
4909 
4910 error_args:
4911         if (f.file)
4912                 fdput(f);
4913 
4914         kfree(args_k);
4915 
4916         DPRINT(("cmd=%s ret=%ld\n", PFM_CMD_NAME(cmd), ret));
4917 
4918         return ret;
4919 }
4920 
4921 static void
4922 pfm_resume_after_ovfl(pfm_context_t *ctx, unsigned long ovfl_regs, struct pt_regs *regs)
4923 {
4924         pfm_buffer_fmt_t *fmt = ctx->ctx_buf_fmt;
4925         pfm_ovfl_ctrl_t rst_ctrl;
4926         int state;
4927         int ret = 0;
4928 
4929         state = ctx->ctx_state;
4930         /*
4931          * Unlock sampling buffer and reset index atomically
4932          * XXX: not really needed when blocking
4933          */
4934         if (CTX_HAS_SMPL(ctx)) {
4935 
4936                 rst_ctrl.bits.mask_monitoring = 0;
4937                 rst_ctrl.bits.reset_ovfl_pmds = 0;
4938 
4939                 if (state == PFM_CTX_LOADED)
4940                         ret = pfm_buf_fmt_restart_active(fmt, current, &rst_ctrl, ctx->ctx_smpl_hdr, regs);
4941                 else
4942                         ret = pfm_buf_fmt_restart(fmt, current, &rst_ctrl, ctx->ctx_smpl_hdr, regs);
4943         } else {
4944                 rst_ctrl.bits.mask_monitoring = 0;
4945                 rst_ctrl.bits.reset_ovfl_pmds = 1;
4946         }
4947 
4948         if (ret == 0) {
4949                 if (rst_ctrl.bits.reset_ovfl_pmds) {
4950                         pfm_reset_regs(ctx, &ovfl_regs, PFM_PMD_LONG_RESET);
4951                 }
4952                 if (rst_ctrl.bits.mask_monitoring == 0) {
4953                         DPRINT(("resuming monitoring\n"));
4954                         if (ctx->ctx_state == PFM_CTX_MASKED) pfm_restore_monitoring(current);
4955                 } else {
4956                         DPRINT(("stopping monitoring\n"));
4957                         //pfm_stop_monitoring(current, regs);
4958                 }
4959                 ctx->ctx_state = PFM_CTX_LOADED;
4960         }
4961 }
4962 
4963 /*
4964  * context MUST BE LOCKED when calling
4965  * can only be called for current
4966  */
4967 static void
4968 pfm_context_force_terminate(pfm_context_t *ctx, struct pt_regs *regs)
4969 {
4970         int ret;
4971 
4972         DPRINT(("entering for [%d]\n", task_pid_nr(current)));
4973 
4974         ret = pfm_context_unload(ctx, NULL, 0, regs);
4975         if (ret) {
4976                 printk(KERN_ERR "pfm_context_force_terminate: [%d] unloaded failed with %d\n", task_pid_nr(current), ret);
4977         }
4978 
4979         /*
4980          * and wakeup controlling task, indicating we are now disconnected
4981          */
4982         wake_up_interruptible(&ctx->ctx_zombieq);
4983 
4984         /*
4985          * given that context is still locked, the controlling
4986          * task will only get access when we return from
4987          * pfm_handle_work().
4988          */
4989 }
4990 
4991 static int pfm_ovfl_notify_user(pfm_context_t *ctx, unsigned long ovfl_pmds);
4992 
4993  /*
4994   * pfm_handle_work() can be called with interrupts enabled
4995   * (TIF_NEED_RESCHED) or disabled. The down_interruptible
4996   * call may sleep, therefore we must re-enable interrupts
4997   * to avoid deadlocks. It is safe to do so because this function
4998   * is called ONLY when returning to user level (pUStk=1), in which case
4999   * there is no risk of kernel stack overflow due to deep
5000   * interrupt nesting.
5001   */
5002 void
5003 pfm_handle_work(void)
5004 {
5005         pfm_context_t *ctx;
5006         struct pt_regs *regs;
5007         unsigned long flags, dummy_flags;
5008         unsigned long ovfl_regs;
5009         unsigned int reason;
5010         int ret;
5011 
5012         ctx = PFM_GET_CTX(current);
5013         if (ctx == NULL) {
5014                 printk(KERN_ERR "perfmon: [%d] has no PFM context\n",
5015                         task_pid_nr(current));
5016                 return;
5017         }
5018 
5019         PROTECT_CTX(ctx, flags);
5020 
5021         PFM_SET_WORK_PENDING(current, 0);
5022 
5023         regs = task_pt_regs(current);
5024 
5025         /*
5026          * extract reason for being here and clear
5027          */
5028         reason = ctx->ctx_fl_trap_reason;
5029         ctx->ctx_fl_trap_reason = PFM_TRAP_REASON_NONE;
5030         ovfl_regs = ctx->ctx_ovfl_regs[0];
5031 
5032         DPRINT(("reason=%d state=%d\n", reason, ctx->ctx_state));
5033 
5034         /*
5035          * must be done before we check for simple-reset mode
5036          */
5037         if (ctx->ctx_fl_going_zombie || ctx->ctx_state == PFM_CTX_ZOMBIE)
5038                 goto do_zombie;
5039 
5040         //if (CTX_OVFL_NOBLOCK(ctx)) goto skip_blocking;
5041         if (reason == PFM_TRAP_REASON_RESET)
5042                 goto skip_blocking;
5043 
5044         /*
5045          * restore interrupt mask to what it was on entry.
5046          * Could be enabled/diasbled.
5047          */
5048         UNPROTECT_CTX(ctx, flags);
5049 
5050         /*
5051          * force interrupt enable because of down_interruptible()
5052          */
5053         local_irq_enable();
5054 
5055         DPRINT(("before block sleeping\n"));
5056 
5057         /*
5058          * may go through without blocking on SMP systems
5059          * if restart has been received already by the time we call down()
5060          */
5061         ret = wait_for_completion_interruptible(&ctx->ctx_restart_done);
5062 
5063         DPRINT(("after block sleeping ret=%d\n", ret));
5064 
5065         /*
5066          * lock context and mask interrupts again
5067          * We save flags into a dummy because we may have
5068          * altered interrupts mask compared to entry in this
5069          * function.
5070          */
5071         PROTECT_CTX(ctx, dummy_flags);
5072 
5073         /*
5074          * we need to read the ovfl_regs only after wake-up
5075          * because we may have had pfm_write_pmds() in between
5076          * and that can changed PMD values and therefore 
5077          * ovfl_regs is reset for these new PMD values.
5078          */
5079         ovfl_regs = ctx->ctx_ovfl_regs[0];
5080 
5081         if (ctx->ctx_fl_going_zombie) {
5082 do_zombie:
5083                 DPRINT(("context is zombie, bailing out\n"));
5084                 pfm_context_force_terminate(ctx, regs);
5085                 goto nothing_to_do;
5086         }
5087         /*
5088          * in case of interruption of down() we don't restart anything
5089          */
5090         if (ret < 0)
5091                 goto nothing_to_do;
5092 
5093 skip_blocking:
5094         pfm_resume_after_ovfl(ctx, ovfl_regs, regs);
5095         ctx->ctx_ovfl_regs[0] = 0UL;
5096 
5097 nothing_to_do:
5098         /*
5099          * restore flags as they were upon entry
5100          */
5101         UNPROTECT_CTX(ctx, flags);
5102 }
5103 
5104 static int
5105 pfm_notify_user(pfm_context_t *ctx, pfm_msg_t *msg)
5106 {
5107         if (ctx->ctx_state == PFM_CTX_ZOMBIE) {
5108                 DPRINT(("ignoring overflow notification, owner is zombie\n"));
5109                 return 0;
5110         }
5111 
5112         DPRINT(("waking up somebody\n"));
5113 
5114         if (msg) wake_up_interruptible(&ctx->ctx_msgq_wait);
5115 
5116         /*
5117          * safe, we are not in intr handler, nor in ctxsw when
5118          * we come here
5119          */
5120         kill_fasync (&ctx->ctx_async_queue, SIGIO, POLL_IN);
5121 
5122         return 0;
5123 }
5124 
5125 static int
5126 pfm_ovfl_notify_user(pfm_context_t *ctx, unsigned long ovfl_pmds)
5127 {
5128         pfm_msg_t *msg = NULL;
5129 
5130         if (ctx->ctx_fl_no_msg == 0) {
5131                 msg = pfm_get_new_msg(ctx);
5132                 if (msg == NULL) {
5133                         printk(KERN_ERR "perfmon: pfm_ovfl_notify_user no more notification msgs\n");
5134                         return -1;
5135                 }
5136 
5137                 msg->pfm_ovfl_msg.msg_type         = PFM_MSG_OVFL;
5138                 msg->pfm_ovfl_msg.msg_ctx_fd       = ctx->ctx_fd;
5139                 msg->pfm_ovfl_msg.msg_active_set   = 0;
5140                 msg->pfm_ovfl_msg.msg_ovfl_pmds[0] = ovfl_pmds;
5141                 msg->pfm_ovfl_msg.msg_ovfl_pmds[1] = 0UL;
5142                 msg->pfm_ovfl_msg.msg_ovfl_pmds[2] = 0UL;
5143                 msg->pfm_ovfl_msg.msg_ovfl_pmds[3] = 0UL;
5144                 msg->pfm_ovfl_msg.msg_tstamp       = 0UL;
5145         }
5146 
5147         DPRINT(("ovfl msg: msg=%p no_msg=%d fd=%d ovfl_pmds=0x%lx\n",
5148                 msg,
5149                 ctx->ctx_fl_no_msg,
5150                 ctx->ctx_fd,
5151                 ovfl_pmds));
5152 
5153         return pfm_notify_user(ctx, msg);
5154 }
5155 
5156 static int
5157 pfm_end_notify_user(pfm_context_t *ctx)
5158 {
5159         pfm_msg_t *msg;
5160 
5161         msg = pfm_get_new_msg(ctx);
5162         if (msg == NULL) {
5163                 printk(KERN_ERR "perfmon: pfm_end_notify_user no more notification msgs\n");
5164                 return -1;
5165         }
5166         /* no leak */
5167         memset(msg, 0, sizeof(*msg));
5168 
5169         msg->pfm_end_msg.msg_type    = PFM_MSG_END;
5170         msg->pfm_end_msg.msg_ctx_fd  = ctx->ctx_fd;
5171         msg->pfm_ovfl_msg.msg_tstamp = 0UL;
5172 
5173         DPRINT(("end msg: msg=%p no_msg=%d ctx_fd=%d\n",
5174                 msg,
5175                 ctx->ctx_fl_no_msg,
5176                 ctx->ctx_fd));
5177 
5178         return pfm_notify_user(ctx, msg);
5179 }
5180 
5181 /*
5182  * main overflow processing routine.
5183  * it can be called from the interrupt path or explicitly during the context switch code
5184  */
5185 static void pfm_overflow_handler(struct task_struct *task, pfm_context_t *ctx,
5186                                 unsigned long pmc0, struct pt_regs *regs)
5187 {
5188         pfm_ovfl_arg_t *ovfl_arg;
5189         unsigned long mask;
5190         unsigned long old_val, ovfl_val, new_val;
5191         unsigned long ovfl_notify = 0UL, ovfl_pmds = 0UL, smpl_pmds = 0UL, reset_pmds;
5192         unsigned long tstamp;
5193         pfm_ovfl_ctrl_t ovfl_ctrl;
5194         unsigned int i, has_smpl;
5195         int must_notify = 0;
5196 
5197         if (unlikely(ctx->ctx_state == PFM_CTX_ZOMBIE)) goto stop_monitoring;
5198 
5199         /*
5200          * sanity test. Should never happen
5201          */
5202         if (unlikely((pmc0 & 0x1) == 0)) goto sanity_check;
5203 
5204         tstamp   = ia64_get_itc();
5205         mask     = pmc0 >> PMU_FIRST_COUNTER;
5206         ovfl_val = pmu_conf->ovfl_val;
5207         has_smpl = CTX_HAS_SMPL(ctx);
5208 
5209         DPRINT_ovfl(("pmc0=0x%lx pid=%d iip=0x%lx, %s "
5210                      "used_pmds=0x%lx\n",
5211                         pmc0,
5212                         task ? task_pid_nr(task): -1,
5213                         (regs ? regs->cr_iip : 0),
5214                         CTX_OVFL_NOBLOCK(ctx) ? "nonblocking" : "blocking",
5215                         ctx->ctx_used_pmds[0]));
5216 
5217 
5218         /*
5219          * first we update the virtual counters
5220          * assume there was a prior ia64_srlz_d() issued
5221          */
5222         for (i = PMU_FIRST_COUNTER; mask ; i++, mask >>= 1) {
5223 
5224                 /* skip pmd which did not overflow */
5225                 if ((mask & 0x1) == 0) continue;
5226 
5227                 /*
5228                  * Note that the pmd is not necessarily 0 at this point as qualified events
5229                  * may have happened before the PMU was frozen. The residual count is not
5230                  * taken into consideration here but will be with any read of the pmd via
5231                  * pfm_read_pmds().
5232                  */
5233                 old_val              = new_val = ctx->ctx_pmds[i].val;
5234                 new_val             += 1 + ovfl_val;
5235                 ctx->ctx_pmds[i].val = new_val;
5236 
5237                 /*
5238                  * check for overflow condition
5239                  */
5240                 if (likely(old_val > new_val)) {
5241                         ovfl_pmds |= 1UL << i;
5242                         if (PMC_OVFL_NOTIFY(ctx, i)) ovfl_notify |= 1UL << i;
5243                 }
5244 
5245                 DPRINT_ovfl(("ctx_pmd[%d].val=0x%lx old_val=0x%lx pmd=0x%lx ovfl_pmds=0x%lx ovfl_notify=0x%lx\n",
5246                         i,
5247                         new_val,
5248                         old_val,
5249                         ia64_get_pmd(i) & ovfl_val,
5250                         ovfl_pmds,
5251                         ovfl_notify));
5252         }
5253 
5254         /*
5255          * there was no 64-bit overflow, nothing else to do
5256          */
5257         if (ovfl_pmds == 0UL) return;
5258 
5259         /* 
5260          * reset all control bits
5261          */
5262         ovfl_ctrl.val = 0;
5263         reset_pmds    = 0UL;
5264 
5265         /*
5266          * if a sampling format module exists, then we "cache" the overflow by 
5267          * calling the module's handler() routine.
5268          */
5269         if (has_smpl) {
5270                 unsigned long start_cycles, end_cycles;
5271                 unsigned long pmd_mask;
5272                 int j, k, ret = 0;
5273                 int this_cpu = smp_processor_id();
5274 
5275                 pmd_mask = ovfl_pmds >> PMU_FIRST_COUNTER;
5276                 ovfl_arg = &ctx->ctx_ovfl_arg;
5277 
5278                 prefetch(ctx->ctx_smpl_hdr);
5279 
5280                 for(i=PMU_FIRST_COUNTER; pmd_mask && ret == 0; i++, pmd_mask >>=1) {
5281 
5282                         mask = 1UL << i;
5283 
5284                         if ((pmd_mask & 0x1) == 0) continue;
5285 
5286                         ovfl_arg->ovfl_pmd      = (unsigned char )i;
5287                         ovfl_arg->ovfl_notify   = ovfl_notify & mask ? 1 : 0;
5288                         ovfl_arg->active_set    = 0;
5289                         ovfl_arg->ovfl_ctrl.val = 0; /* module must fill in all fields */
5290                         ovfl_arg->smpl_pmds[0]  = smpl_pmds = ctx->ctx_pmds[i].smpl_pmds[0];
5291 
5292                         ovfl_arg->pmd_value      = ctx->ctx_pmds[i].val;
5293                         ovfl_arg->pmd_last_reset = ctx->ctx_pmds[i].lval;
5294                         ovfl_arg->pmd_eventid    = ctx->ctx_pmds[i].eventid;
5295 
5296                         /*
5297                          * copy values of pmds of interest. Sampling format may copy them
5298                          * into sampling buffer.
5299                          */
5300                         if (smpl_pmds) {
5301                                 for(j=0, k=0; smpl_pmds; j++, smpl_pmds >>=1) {
5302                                         if ((smpl_pmds & 0x1) == 0) continue;
5303                                         ovfl_arg->smpl_pmds_values[k++] = PMD_IS_COUNTING(j) ?  pfm_read_soft_counter(ctx, j) : ia64_get_pmd(j);
5304                                         DPRINT_ovfl(("smpl_pmd[%d]=pmd%u=0x%lx\n", k-1, j, ovfl_arg->smpl_pmds_values[k-1]));
5305                                 }
5306                         }
5307 
5308                         pfm_stats[this_cpu].pfm_smpl_handler_calls++;
5309 
5310                         start_cycles = ia64_get_itc();
5311 
5312                         /*
5313                          * call custom buffer format record (handler) routine
5314                          */
5315                         ret = (*ctx->ctx_buf_fmt->fmt_handler)(task, ctx->ctx_smpl_hdr, ovfl_arg, regs, tstamp);
5316 
5317                         end_cycles = ia64_get_itc();
5318 
5319                         /*
5320                          * For those controls, we take the union because they have
5321                          * an all or nothing behavior.
5322                          */
5323                         ovfl_ctrl.bits.notify_user     |= ovfl_arg->ovfl_ctrl.bits.notify_user;
5324                         ovfl_ctrl.bits.block_task      |= ovfl_arg->ovfl_ctrl.bits.block_task;
5325                         ovfl_ctrl.bits.mask_monitoring |= ovfl_arg->ovfl_ctrl.bits.mask_monitoring;
5326                         /*
5327                          * build the bitmask of pmds to reset now
5328                          */
5329                         if (ovfl_arg->ovfl_ctrl.bits.reset_ovfl_pmds) reset_pmds |= mask;
5330 
5331                         pfm_stats[this_cpu].pfm_smpl_handler_cycles += end_cycles - start_cycles;
5332                 }
5333                 /*
5334                  * when the module cannot handle the rest of the overflows, we abort right here
5335                  */
5336                 if (ret && pmd_mask) {
5337                         DPRINT(("handler aborts leftover ovfl_pmds=0x%lx\n",
5338                                 pmd_mask<<PMU_FIRST_COUNTER));
5339                 }
5340                 /*
5341                  * remove the pmds we reset now from the set of pmds to reset in pfm_restart()
5342                  */
5343                 ovfl_pmds &= ~reset_pmds;
5344         } else {
5345                 /*
5346                  * when no sampling module is used, then the default
5347                  * is to notify on overflow if requested by user
5348                  */
5349                 ovfl_ctrl.bits.notify_user     = ovfl_notify ? 1 : 0;
5350                 ovfl_ctrl.bits.block_task      = ovfl_notify ? 1 : 0;
5351                 ovfl_ctrl.bits.mask_monitoring = ovfl_notify ? 1 : 0; /* XXX: change for saturation */
5352                 ovfl_ctrl.bits.reset_ovfl_pmds = ovfl_notify ? 0 : 1;
5353                 /*
5354                  * if needed, we reset all overflowed pmds
5355                  */
5356                 if (ovfl_notify == 0) reset_pmds = ovfl_pmds;
5357         }
5358 
5359         DPRINT_ovfl(("ovfl_pmds=0x%lx reset_pmds=0x%lx\n", ovfl_pmds, reset_pmds));
5360 
5361         /*
5362          * reset the requested PMD registers using the short reset values
5363          */
5364         if (reset_pmds) {
5365                 unsigned long bm = reset_pmds;
5366                 pfm_reset_regs(ctx, &bm, PFM_PMD_SHORT_RESET);
5367         }
5368 
5369         if (ovfl_notify && ovfl_ctrl.bits.notify_user) {
5370                 /*
5371                  * keep track of what to reset when unblocking
5372                  */
5373                 ctx->ctx_ovfl_regs[0] = ovfl_pmds;
5374 
5375                 /*
5376                  * check for blocking context 
5377                  */
5378                 if (CTX_OVFL_NOBLOCK(ctx) == 0 && ovfl_ctrl.bits.block_task) {
5379 
5380                         ctx->ctx_fl_trap_reason = PFM_TRAP_REASON_BLOCK;
5381 
5382                         /*
5383                          * set the perfmon specific checking pending work for the task
5384                          */
5385                         PFM_SET_WORK_PENDING(task, 1);
5386 
5387                         /*
5388                          * when coming from ctxsw, current still points to the
5389                          * previous task, therefore we must work with task and not current.
5390                          */
5391                         set_notify_resume(task);
5392                 }
5393                 /*
5394                  * defer until state is changed (shorten spin window). the context is locked
5395                  * anyway, so the signal receiver would come spin for nothing.
5396                  */
5397                 must_notify = 1;
5398         }
5399 
5400         DPRINT_ovfl(("owner [%d] pending=%ld reason=%u ovfl_pmds=0x%lx ovfl_notify=0x%lx masked=%d\n",
5401                         GET_PMU_OWNER() ? task_pid_nr(GET_PMU_OWNER()) : -1,
5402                         PFM_GET_WORK_PENDING(task),
5403                         ctx->ctx_fl_trap_reason,
5404                         ovfl_pmds,
5405                         ovfl_notify,
5406                         ovfl_ctrl.bits.mask_monitoring ? 1 : 0));
5407         /*
5408          * in case monitoring must be stopped, we toggle the psr bits
5409          */
5410         if (ovfl_ctrl.bits.mask_monitoring) {
5411                 pfm_mask_monitoring(task);
5412                 ctx->ctx_state = PFM_CTX_MASKED;
5413                 ctx->ctx_fl_can_restart = 1;
5414         }
5415 
5416         /*
5417          * send notification now
5418          */
5419         if (must_notify) pfm_ovfl_notify_user(ctx, ovfl_notify);
5420 
5421         return;
5422 
5423 sanity_check:
5424         printk(KERN_ERR "perfmon: CPU%d overflow handler [%d] pmc0=0x%lx\n",
5425                         smp_processor_id(),
5426                         task ? task_pid_nr(task) : -1,
5427                         pmc0);
5428         return;
5429 
5430 stop_monitoring:
5431         /*
5432          * in SMP, zombie context is never restored but reclaimed in pfm_load_regs().
5433          * Moreover, zombies are also reclaimed in pfm_save_regs(). Therefore we can
5434          * come here as zombie only if the task is the current task. In which case, we
5435          * can access the PMU  hardware directly.
5436          *
5437          * Note that zombies do have PM_VALID set. So here we do the minimal.
5438          *
5439          * In case the context was zombified it could not be reclaimed at the time
5440          * the monitoring program exited. At this point, the PMU reservation has been
5441          * returned, the sampiing buffer has been freed. We must convert this call
5442          * into a spurious interrupt. However, we must also avoid infinite overflows
5443          * by stopping monitoring for this task. We can only come here for a per-task
5444          * context. All we need to do is to stop monitoring using the psr bits which
5445          * are always task private. By re-enabling secure montioring, we ensure that
5446          * the monitored task will not be able to re-activate monitoring.
5447          * The task will eventually be context switched out, at which point the context
5448          * will be reclaimed (that includes releasing ownership of the PMU).
5449          *
5450          * So there might be a window of time where the number of per-task session is zero
5451          * yet one PMU might have a owner and get at most one overflow interrupt for a zombie
5452          * context. This is safe because if a per-task session comes in, it will push this one
5453          * out and by the virtue on pfm_save_regs(), this one will disappear. If a system wide
5454          * session is force on that CPU, given that we use task pinning, pfm_save_regs() will
5455          * also push our zombie context out.
5456          *
5457          * Overall pretty hairy stuff....
5458          */
5459         DPRINT(("ctx is zombie for [%d], converted to spurious\n", task ? task_pid_nr(task): -1));
5460         pfm_clear_psr_up();
5461         ia64_psr(regs)->up = 0;
5462         ia64_psr(regs)->sp = 1;
5463         return;
5464 }
5465 
5466 static int
5467 pfm_do_interrupt_handler(void *arg, struct pt_regs *regs)
5468 {
5469         struct task_struct *task;
5470         pfm_context_t *ctx;
5471         unsigned long flags;
5472         u64 pmc0;
5473         int this_cpu = smp_processor_id();
5474         int retval = 0;
5475 
5476         pfm_stats[this_cpu].pfm_ovfl_intr_count++;
5477 
5478         /*
5479          * srlz.d done before arriving here
5480          */
5481         pmc0 = ia64_get_pmc(0);
5482 
5483         task = GET_PMU_OWNER();
5484         ctx  = GET_PMU_CTX();
5485 
5486         /*
5487          * if we have some pending bits set
5488          * assumes : if any PMC0.bit[63-1] is set, then PMC0.fr = 1
5489          */
5490         if (PMC0_HAS_OVFL(pmc0) && task) {
5491                 /*
5492                  * we assume that pmc0.fr is always set here
5493                  */
5494 
5495                 /* sanity check */
5496                 if (!ctx) goto report_spurious1;
5497 
5498                 if (ctx->ctx_fl_system == 0 && (task->thread.flags & IA64_THREAD_PM_VALID) == 0) 
5499                         goto report_spurious2;
5500 
5501                 PROTECT_CTX_NOPRINT(ctx, flags);
5502 
5503                 pfm_overflow_handler(task, ctx, pmc0, regs);
5504 
5505                 UNPROTECT_CTX_NOPRINT(ctx, flags);
5506 
5507         } else {
5508                 pfm_stats[this_cpu].pfm_spurious_ovfl_intr_count++;
5509                 retval = -1;
5510         }
5511         /*
5512          * keep it unfrozen at all times
5513          */
5514         pfm_unfreeze_pmu();
5515 
5516         return retval;
5517 
5518 report_spurious1:
5519         printk(KERN_INFO "perfmon: spurious overflow interrupt on CPU%d: process %d has no PFM context\n",
5520                 this_cpu, task_pid_nr(task));
5521         pfm_unfreeze_pmu();
5522         return -1;
5523 report_spurious2:
5524         printk(KERN_INFO "perfmon: spurious overflow interrupt on CPU%d: process %d, invalid flag\n", 
5525                 this_cpu, 
5526                 task_pid_nr(task));
5527         pfm_unfreeze_pmu();
5528         return -1;
5529 }
5530 
5531 static irqreturn_t
5532 pfm_interrupt_handler(int irq, void *arg)
5533 {
5534         unsigned long start_cycles, total_cycles;
5535         unsigned long min, max;
5536         int this_cpu;
5537         int ret;
5538         struct pt_regs *regs = get_irq_regs();
5539 
5540         this_cpu = get_cpu();
5541         if (likely(!pfm_alt_intr_handler)) {
5542                 min = pfm_stats[this_cpu].pfm_ovfl_intr_cycles_min;
5543                 max = pfm_stats[this_cpu].pfm_ovfl_intr_cycles_max;
5544 
5545                 start_cycles = ia64_get_itc();
5546 
5547                 ret = pfm_do_interrupt_handler(arg, regs);
5548 
5549                 total_cycles = ia64_get_itc();
5550 
5551                 /*
5552                  * don't measure spurious interrupts
5553                  */
5554                 if (likely(ret == 0)) {
5555                         total_cycles -= start_cycles;
5556 
5557                         if (total_cycles < min) pfm_stats[this_cpu].pfm_ovfl_intr_cycles_min = total_cycles;
5558                         if (total_cycles > max) pfm_stats[this_cpu].pfm_ovfl_intr_cycles_max = total_cycles;
5559 
5560                         pfm_stats[this_cpu].pfm_ovfl_intr_cycles += total_cycles;
5561                 }
5562         }
5563         else {
5564                 (*pfm_alt_intr_handler->handler)(irq, arg, regs);
5565         }
5566 
5567         put_cpu();
5568         return IRQ_HANDLED;
5569 }
5570 
5571 /*
5572  * /proc/perfmon interface, for debug only
5573  */
5574 
5575 #define PFM_PROC_SHOW_HEADER    ((void *)(long)nr_cpu_ids+1)
5576 
5577 static void *
5578 pfm_proc_start(struct seq_file *m, loff_t *pos)
5579 {
5580         if (*pos == 0) {
5581                 return PFM_PROC_SHOW_HEADER;
5582         }
5583 
5584         while (*pos <= nr_cpu_ids) {
5585                 if (cpu_online(*pos - 1)) {
5586                         return (void *)*pos;
5587                 }
5588                 ++*pos;
5589         }
5590         return NULL;
5591 }
5592 
5593 static void *
5594 pfm_proc_next(struct seq_file *m, void *v, loff_t *pos)
5595 {
5596         ++*pos;
5597         return pfm_proc_start(m, pos);
5598 }
5599 
5600 static void
5601 pfm_proc_stop(struct seq_file *m, void *v)
5602 {
5603 }
5604 
5605 static void
5606 pfm_proc_show_header(struct seq_file *m)
5607 {
5608         struct list_head * pos;
5609         pfm_buffer_fmt_t * entry;
5610         unsigned long flags;
5611 
5612         seq_printf(m,
5613                 "perfmon version           : %u.%u\n"
5614                 "model                     : %s\n"
5615                 "fastctxsw                 : %s\n"
5616                 "expert mode               : %s\n"
5617                 "ovfl_mask                 : 0x%lx\n"
5618                 "PMU flags                 : 0x%x\n",
5619                 PFM_VERSION_MAJ, PFM_VERSION_MIN,
5620                 pmu_conf->pmu_name,
5621                 pfm_sysctl.fastctxsw > 0 ? "Yes": "No",
5622                 pfm_sysctl.expert_mode > 0 ? "Yes": "No",
5623                 pmu_conf->ovfl_val,
5624                 pmu_conf->flags);
5625 
5626         LOCK_PFS(flags);
5627 
5628         seq_printf(m,
5629                 "proc_sessions             : %u\n"
5630                 "sys_sessions              : %u\n"
5631                 "sys_use_dbregs            : %u\n"
5632                 "ptrace_use_dbregs         : %u\n",
5633                 pfm_sessions.pfs_task_sessions,
5634                 pfm_sessions.pfs_sys_sessions,
5635                 pfm_sessions.pfs_sys_use_dbregs,
5636                 pfm_sessions.pfs_ptrace_use_dbregs);
5637 
5638         UNLOCK_PFS(flags);
5639 
5640         spin_lock(&pfm_buffer_fmt_lock);
5641 
5642         list_for_each(pos, &pfm_buffer_fmt_list) {
5643                 entry = list_entry(pos, pfm_buffer_fmt_t, fmt_list);
5644                 seq_printf(m, "format                    : %16phD %s\n",
5645                            entry->fmt_uuid, entry->fmt_name);
5646         }
5647         spin_unlock(&pfm_buffer_fmt_lock);
5648 
5649 }
5650 
5651 static int
5652 pfm_proc_show(struct seq_file *m, void *v)
5653 {
5654         unsigned long psr;
5655         unsigned int i;
5656         int cpu;
5657 
5658         if (v == PFM_PROC_SHOW_HEADER) {
5659                 pfm_proc_show_header(m);
5660                 return 0;
5661         }
5662 
5663         /* show info for CPU (v - 1) */
5664 
5665         cpu = (long)v - 1;
5666         seq_printf(m,
5667                 "CPU%-2d overflow intrs      : %lu\n"
5668                 "CPU%-2d overflow cycles     : %lu\n"
5669                 "CPU%-2d overflow min        : %lu\n"
5670                 "CPU%-2d overflow max        : %lu\n"
5671                 "CPU%-2d smpl handler calls  : %lu\n"
5672                 "CPU%-2d smpl handler cycles : %lu\n"
5673                 "CPU%-2d spurious intrs      : %lu\n"
5674                 "CPU%-2d replay   intrs      : %lu\n"
5675                 "CPU%-2d syst_wide           : %d\n"
5676                 "CPU%-2d dcr_pp              : %d\n"
5677                 "CPU%-2d exclude idle        : %d\n"
5678                 "CPU%-2d owner               : %d\n"
5679                 "CPU%-2d context             : %p\n"
5680                 "CPU%-2d activations         : %lu\n",
5681                 cpu, pfm_stats[cpu].pfm_ovfl_intr_count,
5682                 cpu, pfm_stats[cpu].pfm_ovfl_intr_cycles,
5683                 cpu, pfm_stats[cpu].pfm_ovfl_intr_cycles_min,
5684                 cpu, pfm_stats[cpu].pfm_ovfl_intr_cycles_max,
5685                 cpu, pfm_stats[cpu].pfm_smpl_handler_calls,
5686                 cpu, pfm_stats[cpu].pfm_smpl_handler_cycles,
5687                 cpu, pfm_stats[cpu].pfm_spurious_ovfl_intr_count,
5688                 cpu, pfm_stats[cpu].pfm_replay_ovfl_intr_count,
5689                 cpu, pfm_get_cpu_data(pfm_syst_info, cpu) & PFM_CPUINFO_SYST_WIDE ? 1 : 0,
5690                 cpu, pfm_get_cpu_data(pfm_syst_info, cpu) & PFM_CPUINFO_DCR_PP ? 1 : 0,
5691                 cpu, pfm_get_cpu_data(pfm_syst_info, cpu) & PFM_CPUINFO_EXCL_IDLE ? 1 : 0,
5692                 cpu, pfm_get_cpu_data(pmu_owner, cpu) ? pfm_get_cpu_data(pmu_owner, cpu)->pid: -1,
5693                 cpu, pfm_get_cpu_data(pmu_ctx, cpu),
5694                 cpu, pfm_get_cpu_data(pmu_activation_number, cpu));
5695 
5696         if (num_online_cpus() == 1 && pfm_sysctl.debug > 0) {
5697 
5698                 psr = pfm_get_psr();
5699 
5700                 ia64_srlz_d();
5701 
5702                 seq_printf(m, 
5703                         "CPU%-2d psr                 : 0x%lx\n"
5704                         "CPU%-2d pmc0                : 0x%lx\n", 
5705                         cpu, psr,
5706                         cpu, ia64_get_pmc(0));
5707 
5708                 for (i=0; PMC_IS_LAST(i) == 0;  i++) {
5709                         if (PMC_IS_COUNTING(i) == 0) continue;
5710                         seq_printf(m, 
5711                                 "CPU%-2d pmc%u                : 0x%lx\n"
5712                                 "CPU%-2d pmd%u                : 0x%lx\n", 
5713                                 cpu, i, ia64_get_pmc(i),
5714                                 cpu, i, ia64_get_pmd(i));
5715                 }
5716         }
5717         return 0;
5718 }
5719 
5720 const struct seq_operations pfm_seq_ops = {
5721         .start =        pfm_proc_start,
5722         .next =         pfm_proc_next,
5723         .stop =         pfm_proc_stop,
5724         .show =         pfm_proc_show
5725 };
5726 
5727 static int
5728 pfm_proc_open(struct inode *inode, struct file *file)
5729 {
5730         return seq_open(file, &pfm_seq_ops);
5731 }
5732 
5733 
5734 /*
5735  * we come here as soon as local_cpu_data->pfm_syst_wide is set. this happens
5736  * during pfm_enable() hence before pfm_start(). We cannot assume monitoring
5737  * is active or inactive based on mode. We must rely on the value in
5738  * local_cpu_data->pfm_syst_info
5739  */
5740 void
5741 pfm_syst_wide_update_task(struct task_struct *task, unsigned long info, int is_ctxswin)
5742 {
5743         struct pt_regs *regs;
5744         unsigned long dcr;
5745         unsigned long dcr_pp;
5746 
5747         dcr_pp = info & PFM_CPUINFO_DCR_PP ? 1 : 0;
5748 
5749         /*
5750          * pid 0 is guaranteed to be the idle task. There is one such task with pid 0
5751          * on every CPU, so we can rely on the pid to identify the idle task.
5752          */
5753         if ((info & PFM_CPUINFO_EXCL_IDLE) == 0 || task->pid) {
5754                 regs = task_pt_regs(task);
5755                 ia64_psr(regs)->pp = is_ctxswin ? dcr_pp : 0;
5756                 return;
5757         }
5758         /*
5759          * if monitoring has started
5760          */
5761         if (dcr_pp) {
5762                 dcr = ia64_getreg(_IA64_REG_CR_DCR);
5763                 /*
5764                  * context switching in?
5765                  */
5766                 if (is_ctxswin) {
5767                         /* mask monitoring for the idle task */
5768                         ia64_setreg(_IA64_REG_CR_DCR, dcr & ~IA64_DCR_PP);
5769                         pfm_clear_psr_pp();
5770                         ia64_srlz_i();
5771                         return;
5772                 }
5773                 /*
5774                  * context switching out
5775                  * restore monitoring for next task
5776                  *
5777                  * Due to inlining this odd if-then-else construction generates
5778                  * better code.
5779                  */
5780                 ia64_setreg(_IA64_REG_CR_DCR, dcr |IA64_DCR_PP);
5781                 pfm_set_psr_pp();
5782                 ia64_srlz_i();
5783         }
5784 }
5785 
5786 #ifdef CONFIG_SMP
5787 
5788 static void
5789 pfm_force_cleanup(pfm_context_t *ctx, struct pt_regs *regs)
5790 {
5791         struct task_struct *task = ctx->ctx_task;
5792 
5793         ia64_psr(regs)->up = 0;
5794         ia64_psr(regs)->sp = 1;
5795 
5796         if (GET_PMU_OWNER() == task) {
5797                 DPRINT(("cleared ownership for [%d]\n",
5798                                         task_pid_nr(ctx->ctx_task)));
5799                 SET_PMU_OWNER(NULL, NULL);
5800         }
5801 
5802         /*
5803          * disconnect the task from the context and vice-versa
5804          */
5805         PFM_SET_WORK_PENDING(task, 0);
5806 
5807         task->thread.pfm_context  = NULL;
5808         task->thread.flags       &= ~IA64_THREAD_PM_VALID;
5809 
5810         DPRINT(("force cleanup for [%d]\n",  task_pid_nr(task)));
5811 }
5812 
5813 
5814 /*
5815  * in 2.6, interrupts are masked when we come here and the runqueue lock is held
5816  */
5817 void
5818 pfm_save_regs(struct task_struct *task)
5819 {
5820         pfm_context_t *ctx;
5821         unsigned long flags;
5822         u64 psr;
5823 
5824 
5825         ctx = PFM_GET_CTX(task);
5826         if (ctx == NULL) return;
5827 
5828         /*
5829          * we always come here with interrupts ALREADY disabled by
5830          * the scheduler. So we simply need to protect against concurrent
5831          * access, not CPU concurrency.
5832          */
5833         flags = pfm_protect_ctx_ctxsw(ctx);
5834 
5835         if (ctx->ctx_state == PFM_CTX_ZOMBIE) {
5836                 struct pt_regs *regs = task_pt_regs(task);
5837 
5838                 pfm_clear_psr_up();
5839 
5840                 pfm_force_cleanup(ctx, regs);
5841 
5842                 BUG_ON(ctx->ctx_smpl_hdr);
5843 
5844                 pfm_unprotect_ctx_ctxsw(ctx, flags);
5845 
5846                 pfm_context_free(ctx);
5847                 return;
5848         }
5849 
5850         /*
5851          * save current PSR: needed because we modify it
5852          */
5853         ia64_srlz_d();
5854         psr = pfm_get_psr();