/*
 * This file implements the perfmon-2 subsystem which is used
 * to program the IA-64 Performance Monitoring Unit (PMU).
 *
 * The initial version of perfmon.c was written by
 * Ganesh Venkitachalam, IBM Corp.
 *
 * Then it was modified for perfmon-1.x by Stephane Eranian and
 * David Mosberger, Hewlett Packard Co.
 *
 * Version Perfmon-2.x is a rewrite of perfmon-1.x
 * by Stephane Eranian, Hewlett Packard Co.
 *
 * Copyright (C) 1999-2003, 2005  Hewlett Packard Co
 *               Stephane Eranian <eranian@hpl.hp.com>
 *               David Mosberger-Tang <davidm@hpl.hp.com>
 *
 * More information about perfmon available at:
 *      http://www.hpl.hp.com/research/linux/perfmon
 */

#include <linux/config.h>
#include <linux/module.h>
#include <linux/kernel.h>
#include <linux/sched.h>
#include <linux/interrupt.h>
#include <linux/smp_lock.h>
#include <linux/proc_fs.h>
#include <linux/seq_file.h>
#include <linux/init.h>
#include <linux/vmalloc.h>
#include <linux/mm.h>
#include <linux/sysctl.h>
#include <linux/list.h>
#include <linux/file.h>
#include <linux/poll.h>
#include <linux/vfs.h>
#include <linux/pagemap.h>
#include <linux/mount.h>
#include <linux/version.h>
#include <linux/bitops.h>

#include <asm/errno.h>
#include <asm/intrinsics.h>
#include <asm/page.h>
#include <asm/perfmon.h>
#include <asm/processor.h>
#include <asm/signal.h>
#include <asm/system.h>
#include <asm/uaccess.h>
#include <asm/delay.h>

#ifdef CONFIG_PERFMON
/*
 * perfmon context state
 */
#define PFM_CTX_UNLOADED        1       /* context is not loaded onto any task */
#define PFM_CTX_LOADED          2       /* context is loaded onto a task */
#define PFM_CTX_MASKED          3       /* context is loaded but monitoring is masked due to overflow */
#define PFM_CTX_ZOMBIE          4       /* owner of the context is closing it */

#define PFM_INVALID_ACTIVATION  (~0UL)

/*
 * depth of message queue
 */
#define PFM_MAX_MSGS            32
#define PFM_CTXQ_EMPTY(g)       ((g)->ctx_msgq_head == (g)->ctx_msgq_tail)

/*
 * type of a PMU register (bitmask).
 * bitmask structure:
 *      bit0   : register implemented
 *      bit1   : end marker
 *      bit2-3 : reserved
 *      bit4   : pmc has pmc.pm
 *      bit5   : pmc controls a counter (has pmc.oi), pmd is used as counter
 *      bit6-7 : register type
 *      bit8-31: reserved
 */
#define PFM_REG_NOTIMPL         0x0 /* not implemented at all */
#define PFM_REG_IMPL            0x1 /* register implemented */
#define PFM_REG_END             0x2 /* end marker */
#define PFM_REG_MONITOR         (0x1<<4|PFM_REG_IMPL) /* a PMC with a pmc.pm field only */
#define PFM_REG_COUNTING        (0x2<<4|PFM_REG_MONITOR) /* a monitor + pmc.oi+ PMD used as a counter */
#define PFM_REG_CONTROL         (0x4<<4|PFM_REG_IMPL) /* PMU control register */
#define PFM_REG_CONFIG          (0x8<<4|PFM_REG_IMPL) /* configuration register */
#define PFM_REG_BUFFER          (0xc<<4|PFM_REG_IMPL) /* PMD used as buffer */

#define PMC_IS_LAST(i)  (pmu_conf->pmc_desc[i].type & PFM_REG_END)
#define PMD_IS_LAST(i)  (pmu_conf->pmd_desc[i].type & PFM_REG_END)

#define PMC_OVFL_NOTIFY(ctx, i) ((ctx)->ctx_pmds[i].flags &  PFM_REGFL_OVFL_NOTIFY)

/* i assumed unsigned */
#define PMC_IS_IMPL(i)    (i< PMU_MAX_PMCS && (pmu_conf->pmc_desc[i].type & PFM_REG_IMPL))
#define PMD_IS_IMPL(i)    (i< PMU_MAX_PMDS && (pmu_conf->pmd_desc[i].type & PFM_REG_IMPL))

/* XXX: these assume that register i is implemented */
#define PMD_IS_COUNTING(i) ((pmu_conf->pmd_desc[i].type & PFM_REG_COUNTING) == PFM_REG_COUNTING)
#define PMC_IS_COUNTING(i) ((pmu_conf->pmc_desc[i].type & PFM_REG_COUNTING) == PFM_REG_COUNTING)
#define PMC_IS_MONITOR(i)  ((pmu_conf->pmc_desc[i].type & PFM_REG_MONITOR)  == PFM_REG_MONITOR)
#define PMC_IS_CONTROL(i)  ((pmu_conf->pmc_desc[i].type & PFM_REG_CONTROL)  == PFM_REG_CONTROL)

#define PMC_DFL_VAL(i)     pmu_conf->pmc_desc[i].default_value
#define PMC_RSVD_MASK(i)   pmu_conf->pmc_desc[i].reserved_mask
#define PMD_PMD_DEP(i)     pmu_conf->pmd_desc[i].dep_pmd[0]
#define PMC_PMD_DEP(i)     pmu_conf->pmc_desc[i].dep_pmd[0]

#define PFM_NUM_IBRS      IA64_NUM_DBG_REGS
#define PFM_NUM_DBRS      IA64_NUM_DBG_REGS

#define CTX_OVFL_NOBLOCK(c)     ((c)->ctx_fl_block == 0)
#define CTX_HAS_SMPL(c)         ((c)->ctx_fl_is_sampling)
#define PFM_CTX_TASK(h)         (h)->ctx_task

#define PMU_PMC_OI              5 /* position of pmc.oi bit */

/* XXX: does not support more than 64 PMDs */
#define CTX_USED_PMD(ctx, mask) (ctx)->ctx_used_pmds[0] |= (mask)
#define CTX_IS_USED_PMD(ctx, c) (((ctx)->ctx_used_pmds[0] & (1UL << (c))) != 0UL)

#define CTX_USED_MONITOR(ctx, mask) (ctx)->ctx_used_monitors[0] |= (mask)

#define CTX_USED_IBR(ctx,n)     (ctx)->ctx_used_ibrs[(n)>>6] |= 1UL<< ((n) % 64)
#define CTX_USED_DBR(ctx,n)     (ctx)->ctx_used_dbrs[(n)>>6] |= 1UL<< ((n) % 64)
#define CTX_USES_DBREGS(ctx)    (((pfm_context_t *)(ctx))->ctx_fl_using_dbreg==1)
#define PFM_CODE_RR     0       /* requesting code range restriction */
#define PFM_DATA_RR     1       /* requestion data range restriction */

#define PFM_CPUINFO_CLEAR(v)    pfm_get_cpu_var(pfm_syst_info) &= ~(v)
#define PFM_CPUINFO_SET(v)      pfm_get_cpu_var(pfm_syst_info) |= (v)
#define PFM_CPUINFO_GET()       pfm_get_cpu_var(pfm_syst_info)

#define RDEP(x) (1UL<<(x))

/*
 * context protection macros
 * in SMP:
 *      - we need to protect against CPU concurrency (spin_lock)
 *      - we need to protect against PMU overflow interrupts (local_irq_disable)
 * in UP:
 *      - we need to protect against PMU overflow interrupts (local_irq_disable)
 *
 * spin_lock_irqsave()/spin_lock_irqrestore():
 *      in SMP: local_irq_disable + spin_lock
 *      in UP : local_irq_disable
 *
 * spin_lock()/spin_lock():
 *      in UP : removed automatically
 *      in SMP: protect against context accesses from other CPU. interrupts
 *              are not masked. This is useful for the PMU interrupt handler
 *              because we know we will not get PMU concurrency in that code.
 */
#define PROTECT_CTX(c, f) \
        do {  \
                DPRINT(("spinlock_irq_save ctx %p by [%d]\n", c, current->pid)); \
                spin_lock_irqsave(&(c)->ctx_lock, f); \
                DPRINT(("spinlocked ctx %p  by [%d]\n", c, current->pid)); \
        } while(0)

#define UNPROTECT_CTX(c, f) \
        do { \
                DPRINT(("spinlock_irq_restore ctx %p by [%d]\n", c, current->pid)); \
                spin_unlock_irqrestore(&(c)->ctx_lock, f); \
        } while(0)

#define PROTECT_CTX_NOPRINT(c, f) \
        do {  \
                spin_lock_irqsave(&(c)->ctx_lock, f); \
        } while(0)


#define UNPROTECT_CTX_NOPRINT(c, f) \
        do { \
                spin_unlock_irqrestore(&(c)->ctx_lock, f); \
        } while(0)


#define PROTECT_CTX_NOIRQ(c) \
        do {  \
                spin_lock(&(c)->ctx_lock); \
        } while(0)

#define UNPROTECT_CTX_NOIRQ(c) \
        do { \
                spin_unlock(&(c)->ctx_lock); \
        } while(0)


#ifdef CONFIG_SMP

#define GET_ACTIVATION()        pfm_get_cpu_var(pmu_activation_number)
#define INC_ACTIVATION()        pfm_get_cpu_var(pmu_activation_number)++
#define SET_ACTIVATION(c)       (c)->ctx_last_activation = GET_ACTIVATION()

#else /* !CONFIG_SMP */
#define SET_ACTIVATION(t)       do {} while(0)
#define GET_ACTIVATION(t)       do {} while(0)
#define INC_ACTIVATION(t)       do {} while(0)
#endif /* CONFIG_SMP */

#define SET_PMU_OWNER(t, c)     do { pfm_get_cpu_var(pmu_owner) = (t); pfm_get_cpu_var(pmu_ctx) = (c); } while(0)
#define GET_PMU_OWNER()         pfm_get_cpu_var(pmu_owner)
#define GET_PMU_CTX()           pfm_get_cpu_var(pmu_ctx)

#define LOCK_PFS(g)             spin_lock_irqsave(&pfm_sessions.pfs_lock, g)
#define UNLOCK_PFS(g)           spin_unlock_irqrestore(&pfm_sessions.pfs_lock, g)

#define PFM_REG_RETFLAG_SET(flags, val) do { flags &= ~PFM_REG_RETFL_MASK; flags |= (val); } while(0)

/*
 * cmp0 must be the value of pmc0
 */
#define PMC0_HAS_OVFL(cmp0)  (cmp0 & ~0x1UL)

#define PFMFS_MAGIC 0xa0b4d889

/*
 * debugging
 */
#define PFM_DEBUGGING 1
#ifdef PFM_DEBUGGING
#define DPRINT(a) \
        do { \
                if (unlikely(pfm_sysctl.debug >0)) { printk("%s.%d: CPU%d [%d] ", __FUNCTION__, __LINE__, smp_processor_id(), current->pid); printk a; } \
        } while (0)

#define DPRINT_ovfl(a) \
        do { \
                if (unlikely(pfm_sysctl.debug > 0 && pfm_sysctl.debug_ovfl >0)) { printk("%s.%d: CPU%d [%d] ", __FUNCTION__, __LINE__, smp_processor_id(), current->pid); printk a; } \
        } while (0)
#endif

/*
 * 64-bit software counter structure
 *
 * the next_reset_type is applied to the next call to pfm_reset_regs()
 */
typedef struct {
        unsigned long   val;            /* virtual 64bit counter value */
        unsigned long   lval;           /* last reset value */
        unsigned long   long_reset;     /* reset value on sampling overflow */
        unsigned long   short_reset;    /* reset value on overflow */
        unsigned long   reset_pmds[4];  /* which other pmds to reset when this counter overflows */
        unsigned long   smpl_pmds[4];   /* which pmds are accessed when counter overflow */
        unsigned long   seed;           /* seed for random-number generator */
        unsigned long   mask;           /* mask for random-number generator */
        unsigned int    flags;          /* notify/do not notify */
        unsigned long   eventid;        /* overflow event identifier */
} pfm_counter_t;

/*
 * context flags
 */
typedef struct {
        unsigned int block:1;           /* when 1, task will blocked on user notifications */
        unsigned int system:1;          /* do system wide monitoring */
        unsigned int using_dbreg:1;     /* using range restrictions (debug registers) */
        unsigned int is_sampling:1;     /* true if using a custom format */
        unsigned int excl_idle:1;       /* exclude idle task in system wide session */
        unsigned int going_zombie:1;    /* context is zombie (MASKED+blocking) */
        unsigned int trap_reason:2;     /* reason for going into pfm_handle_work() */
        unsigned int no_msg:1;          /* no message sent on overflow */
        unsigned int can_restart:1;     /* allowed to issue a PFM_RESTART */
        unsigned int reserved:22;
} pfm_context_flags_t;

#define PFM_TRAP_REASON_NONE            0x0     /* default value */
#define PFM_TRAP_REASON_BLOCK           0x1     /* we need to block on overflow */
#define PFM_TRAP_REASON_RESET           0x2     /* we need to reset PMDs */


/*
 * perfmon context: encapsulates all the state of a monitoring session
 */

typedef struct pfm_context {
        spinlock_t              ctx_lock;               /* context protection */

        pfm_context_flags_t     ctx_flags;              /* bitmask of flags  (block reason incl.) */
        unsigned int            ctx_state;              /* state: active/inactive (no bitfield) */

        struct task_struct      *ctx_task;              /* task to which context is attached */

        unsigned long           ctx_ovfl_regs[4];       /* which registers overflowed (notification) */

        struct semaphore        ctx_restart_sem;        /* use for blocking notification mode */

        unsigned long           ctx_used_pmds[4];       /* bitmask of PMD used            */
        unsigned long           ctx_all_pmds[4];        /* bitmask of all accessible PMDs */
        unsigned long           ctx_reload_pmds[4];     /* bitmask of force reload PMD on ctxsw in */

        unsigned long           ctx_all_pmcs[4];        /* bitmask of all accessible PMCs */
        unsigned long           ctx_reload_pmcs[4];     /* bitmask of force reload PMC on ctxsw in */
        unsigned long           ctx_used_monitors[4];   /* bitmask of monitor PMC being used */

        unsigned long           ctx_pmcs[IA64_NUM_PMC_REGS];    /*  saved copies of PMC values */

        unsigned int            ctx_used_ibrs[1];               /* bitmask of used IBR (speedup ctxsw in) */
        unsigned int            ctx_used_dbrs[1];               /* bitmask of used DBR (speedup ctxsw in) */
        unsigned long           ctx_dbrs[IA64_NUM_DBG_REGS];    /* DBR values (cache) when not loaded */
        unsigned long           ctx_ibrs[IA64_NUM_DBG_REGS];    /* IBR values (cache) when not loaded */

        pfm_counter_t           ctx_pmds[IA64_NUM_PMD_REGS]; /* software state for PMDS */

        u64                     ctx_saved_psr_up;       /* only contains psr.up value */

        unsigned long           ctx_last_activation;    /* context last activation number for last_cpu */
        unsigned int            ctx_last_cpu;           /* CPU id of current or last CPU used (SMP only) */
        unsigned int            ctx_cpu;                /* cpu to which perfmon is applied (system wide) */

        int                     ctx_fd;                 /* file descriptor used my this context */
        pfm_ovfl_arg_t          ctx_ovfl_arg;           /* argument to custom buffer format handler */

        pfm_buffer_fmt_t        *ctx_buf_fmt;           /* buffer format callbacks */
        void                    *ctx_smpl_hdr;          /* points to sampling buffer header kernel vaddr */
        unsigned long           ctx_smpl_size;          /* size of sampling buffer */
        void                    *ctx_smpl_vaddr;        /* user level virtual address of smpl buffer */

        wait_queue_head_t       ctx_msgq_wait;
        pfm_msg_t               ctx_msgq[PFM_MAX_MSGS];
        int                     ctx_msgq_head;
        int                     ctx_msgq_tail;
        struct fasync_struct    *ctx_async_queue;

        wait_queue_head_t       ctx_zombieq;            /* termination cleanup wait queue */
} pfm_context_t;

/*
 * magic number used to verify that structure is really
 * a perfmon context
 */
#define PFM_IS_FILE(f)          ((f)->f_op == &pfm_file_ops)

#define PFM_GET_CTX(t)          ((pfm_context_t *)(t)->thread.pfm_context)

#ifdef CONFIG_SMP
#define SET_LAST_CPU(ctx, v)    (ctx)->ctx_last_cpu = (v)
#define GET_LAST_CPU(ctx)       (ctx)->ctx_last_cpu
#else
#define SET_LAST_CPU(ctx, v)    do {} while(0)
#define GET_LAST_CPU(ctx)       do {} while(0)
#endif


#define ctx_fl_block            ctx_flags.block
#define ctx_fl_system           ctx_flags.system
#define ctx_fl_using_dbreg      ctx_flags.using_dbreg
#define ctx_fl_is_sampling      ctx_flags.is_sampling
#define ctx_fl_excl_idle        ctx_flags.excl_idle
#define ctx_fl_going_zombie     ctx_flags.going_zombie
#define ctx_fl_trap_reason      ctx_flags.trap_reason
#define ctx_fl_no_msg           ctx_flags.no_msg
#define ctx_fl_can_restart      ctx_flags.can_restart

#define PFM_SET_WORK_PENDING(t, v)      do { (t)->thread.pfm_needs_checking = v; } while(0);
#define PFM_GET_WORK_PENDING(t)         (t)->thread.pfm_needs_checking

/*
 * global information about all sessions
 * mostly used to synchronize between system wide and per-process
 */
typedef struct {
        spinlock_t              pfs_lock;                  /* lock the structure */

        unsigned int            pfs_task_sessions;         /* number of per task sessions */
        unsigned int            pfs_sys_sessions;          /* number of per system wide sessions */
        unsigned int            pfs_sys_use_dbregs;        /* incremented when a system wide session uses debug regs */
        unsigned int            pfs_ptrace_use_dbregs;     /* incremented when a process uses debug regs */
        struct task_struct      *pfs_sys_session[NR_CPUS]; /* point to task owning a system-wide session */
} pfm_session_t;

/*
 * information about a PMC or PMD.
 * dep_pmd[]: a bitmask of dependent PMD registers
 * dep_pmc[]: a bitmask of dependent PMC registers
 */
typedef int (*pfm_reg_check_t)(struct task_struct *task, pfm_context_t *ctx, unsigned int cnum, unsigned long *val, struct pt_regs *regs);
typedef struct {
        unsigned int            type;
        int                     pm_pos;
        unsigned long           default_value;  /* power-on default value */
        unsigned long           reserved_mask;  /* bitmask of reserved bits */
        pfm_reg_check_t         read_check;
        pfm_reg_check_t         write_check;
        unsigned long           dep_pmd[4];
        unsigned long           dep_pmc[4];
} pfm_reg_desc_t;

/* assume cnum is a valid monitor */
#define PMC_PM(cnum, val)       (((val) >> (pmu_conf->pmc_desc[cnum].pm_pos)) & 0x1)

/*
 * This structure is initialized at boot time and contains
 * a description of the PMU main characteristics.
 *
 * If the probe function is defined, detection is based
 * on its return value: 
 *      - 0 means recognized PMU
 *      - anything else means not supported
 * When the probe function is not defined, then the pmu_family field
 * is used and it must match the host CPU family such that:
 *      - cpu->family & config->pmu_family != 0
 */
typedef struct {
        unsigned long  ovfl_val;        /* overflow value for counters */

        pfm_reg_desc_t *pmc_desc;       /* detailed PMC register dependencies descriptions */
        pfm_reg_desc_t *pmd_desc;       /* detailed PMD register dependencies descriptions */

        unsigned int   num_pmcs;        /* number of PMCS: computed at init time */
        unsigned int   num_pmds;        /* number of PMDS: computed at init time */
        unsigned long  impl_pmcs[4];    /* bitmask of implemented PMCS */
        unsigned long  impl_pmds[4];    /* bitmask of implemented PMDS */

        char          *pmu_name;        /* PMU family name */
        unsigned int  pmu_family;       /* cpuid family pattern used to identify pmu */
        unsigned int  flags;            /* pmu specific flags */
        unsigned int  num_ibrs;         /* number of IBRS: computed at init time */
        unsigned int  num_dbrs;         /* number of DBRS: computed at init time */
        unsigned int  num_counters;     /* PMC/PMD counting pairs : computed at init time */
        int           (*probe)(void);   /* customized probe routine */
        unsigned int  use_rr_dbregs:1;  /* set if debug registers used for range restriction */
} pmu_config_t;
/*
 * PMU specific flags
 */
#define PFM_PMU_IRQ_RESEND      1       /* PMU needs explicit IRQ resend */

/*
 * debug register related type definitions
 */
typedef struct {
        unsigned long ibr_mask:56;
        unsigned long ibr_plm:4;
        unsigned long ibr_ig:3;
        unsigned long ibr_x:1;
} ibr_mask_reg_t;

typedef struct {
        unsigned long dbr_mask:56;
        unsigned long dbr_plm:4;
        unsigned long dbr_ig:2;
        unsigned long dbr_w:1;
        unsigned long dbr_r:1;
} dbr_mask_reg_t;

typedef union {
        unsigned long  val;
        ibr_mask_reg_t ibr;
        dbr_mask_reg_t dbr;
} dbreg_t;


/*
 * perfmon command descriptions
 */
typedef struct {
        int             (*cmd_func)(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs);
        char            *cmd_name;
        int             cmd_flags;
        unsigned int    cmd_narg;
        size_t          cmd_argsize;
        int             (*cmd_getsize)(void *arg, size_t *sz);
} pfm_cmd_desc_t;

#define PFM_CMD_FD              0x01    /* command requires a file descriptor */
#define PFM_CMD_ARG_READ        0x02    /* command must read argument(s) */
#define PFM_CMD_ARG_RW          0x04    /* command must read/write argument(s) */
#define PFM_CMD_STOP            0x08    /* command does not work on zombie context */


#define PFM_CMD_NAME(cmd)       pfm_cmd_tab[(cmd)].cmd_name
#define PFM_CMD_READ_ARG(cmd)   (pfm_cmd_tab[(cmd)].cmd_flags & PFM_CMD_ARG_READ)
#define PFM_CMD_RW_ARG(cmd)     (pfm_cmd_tab[(cmd)].cmd_flags & PFM_CMD_ARG_RW)
#define PFM_CMD_USE_FD(cmd)     (pfm_cmd_tab[(cmd)].cmd_flags & PFM_CMD_FD)
#define PFM_CMD_STOPPED(cmd)    (pfm_cmd_tab[(cmd)].cmd_flags & PFM_CMD_STOP)

#define PFM_CMD_ARG_MANY        -1 /* cannot be zero */

typedef struct {
        int     debug;          /* turn on/off debugging via syslog */
        int     debug_ovfl;     /* turn on/off debug printk in overflow handler */
        int     fastctxsw;      /* turn on/off fast (unsecure) ctxsw */
        int     expert_mode;    /* turn on/off value checking */
        int     debug_pfm_read;
} pfm_sysctl_t;

typedef struct {
        unsigned long pfm_spurious_ovfl_intr_count;     /* keep track of spurious ovfl interrupts */
        unsigned long pfm_replay_ovfl_intr_count;       /* keep track of replayed ovfl interrupts */
        unsigned long pfm_ovfl_intr_count;              /* keep track of ovfl interrupts */
        unsigned long pfm_ovfl_intr_cycles;             /* cycles spent processing ovfl interrupts */
        unsigned long pfm_ovfl_intr_cycles_min;         /* min cycles spent processing ovfl interrupts */
        unsigned long pfm_ovfl_intr_cycles_max;         /* max cycles spent processing ovfl interrupts */
        unsigned long pfm_smpl_handler_calls;
        unsigned long pfm_smpl_handler_cycles;
        char pad[SMP_CACHE_BYTES] ____cacheline_aligned;
} pfm_stats_t;

/*
 * perfmon internal variables
 */
static pfm_stats_t              pfm_stats[NR_CPUS];
static pfm_session_t            pfm_sessions;   /* global sessions information */

static struct proc_dir_entry    *perfmon_dir;
static pfm_uuid_t               pfm_null_uuid = {0,};

static spinlock_t               pfm_buffer_fmt_lock;
static LIST_HEAD(pfm_buffer_fmt_list);

static pmu_config_t             *pmu_conf;

/* sysctl() controls */
static pfm_sysctl_t pfm_sysctl;
int pfm_debug_var;

static ctl_table pfm_ctl_table[]={
        {1, "debug", &pfm_sysctl.debug, sizeof(int), 0666, NULL, &proc_dointvec, NULL,},
        {2, "debug_ovfl", &pfm_sysctl.debug_ovfl, sizeof(int), 0666, NULL, &proc_dointvec, NULL,},
        {3, "fastctxsw", &pfm_sysctl.fastctxsw, sizeof(int), 0600, NULL, &proc_dointvec, NULL,},
        {4, "expert_mode", &pfm_sysctl.expert_mode, sizeof(int), 0600, NULL, &proc_dointvec, NULL,},
        { 0, },
};
static ctl_table pfm_sysctl_dir[] = {
        {1, "perfmon", NULL, 0, 0755, pfm_ctl_table, },
        {0,},
};
static ctl_table pfm_sysctl_root[] = {
        {1, "kernel", NULL, 0, 0755, pfm_sysctl_dir, },
        {0,},
};
static struct ctl_table_header *pfm_sysctl_header;

static int pfm_context_unload(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs);
static int pfm_flush(struct file *filp);

#define pfm_get_cpu_var(v)              __ia64_per_cpu_var(v)
#define pfm_get_cpu_data(a,b)           per_cpu(a, b)

static inline void
pfm_put_task(struct task_struct *task)
{
        if (task != current) put_task_struct(task);
}

static inline void
pfm_set_task_notify(struct task_struct *task)
{
        struct thread_info *info;

        info = (struct thread_info *) ((char *) task + IA64_TASK_SIZE);
        set_bit(TIF_NOTIFY_RESUME, &info->flags);
}

static inline void
pfm_clear_task_notify(void)
{
        clear_thread_flag(TIF_NOTIFY_RESUME);
}

static inline void
pfm_reserve_page(unsigned long a)
{
        SetPageReserved(vmalloc_to_page((void *)a));
}
static inline void
pfm_unreserve_page(unsigned long a)
{
        ClearPageReserved(vmalloc_to_page((void*)a));
}

static inline unsigned long
pfm_protect_ctx_ctxsw(pfm_context_t *x)
{
        spin_lock(&(x)->ctx_lock);
        return 0UL;
}

static inline unsigned long
pfm_unprotect_ctx_ctxsw(pfm_context_t *x, unsigned long f)
{
        spin_unlock(&(x)->ctx_lock);
}

static inline unsigned int
pfm_do_munmap(struct mm_struct *mm, unsigned long addr, size_t len, int acct)
{
        return do_munmap(mm, addr, len);
}

static inline unsigned long 
pfm_get_unmapped_area(struct file *file, unsigned long addr, unsigned long len, unsigned long pgoff, unsigned long flags, unsigned long exec)
{
        return get_unmapped_area(file, addr, len, pgoff, flags);
}


static struct super_block *
pfmfs_get_sb(struct file_system_type *fs_type, int flags, const char *dev_name, void *data)
{
        return get_sb_pseudo(fs_type, "pfm:", NULL, PFMFS_MAGIC);
}

static struct file_system_type pfm_fs_type = {
        .name     = "pfmfs",
        .get_sb   = pfmfs_get_sb,
        .kill_sb  = kill_anon_super,
};

DEFINE_PER_CPU(unsigned long, pfm_syst_info);
DEFINE_PER_CPU(struct task_struct *, pmu_owner);
DEFINE_PER_CPU(pfm_context_t  *, pmu_ctx);
DEFINE_PER_CPU(unsigned long, pmu_activation_number);


/* forward declaration */
static struct file_operations pfm_file_ops;

/*
 * forward declarations
 */
#ifndef CONFIG_SMP
static void pfm_lazy_save_regs (struct task_struct *ta);
#endif

void dump_pmu_state(const char *);
static int pfm_write_ibr_dbr(int mode, pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs);

#include "perfmon_itanium.h"
#include "perfmon_mckinley.h"
#include "perfmon_generic.h"

static pmu_config_t *pmu_confs[]={
        &pmu_conf_mck,
        &pmu_conf_ita,
        &pmu_conf_gen, /* must be last */
        NULL
};


static int pfm_end_notify_user(pfm_context_t *ctx);

static inline void
pfm_clear_psr_pp(void)
{
        ia64_rsm(IA64_PSR_PP);
        ia64_srlz_i();
}

static inline void
pfm_set_psr_pp(void)
{
        ia64_ssm(IA64_PSR_PP);
        ia64_srlz_i();
}

static inline void
pfm_clear_psr_up(void)
{
        ia64_rsm(IA64_PSR_UP);
        ia64_srlz_i();
}

static inline void
pfm_set_psr_up(void)
{
        ia64_ssm(IA64_PSR_UP);
        ia64_srlz_i();
}

static inline unsigned long
pfm_get_psr(void)
{
        unsigned long tmp;
        tmp = ia64_getreg(_IA64_REG_PSR);
        ia64_srlz_i();
        return tmp;
}

static inline void
pfm_set_psr_l(unsigned long val)
{
        ia64_setreg(_IA64_REG_PSR_L, val);
        ia64_srlz_i();
}

static inline void
pfm_freeze_pmu(void)
{
        ia64_set_pmc(0,1UL);
        ia64_srlz_d();
}

static inline void
pfm_unfreeze_pmu(void)
{
        ia64_set_pmc(0,0UL);
        ia64_srlz_d();
}

static inline void
pfm_restore_ibrs(unsigned long *ibrs, unsigned int nibrs)
{
        int i;

        for (i=0; i < nibrs; i++) {
                ia64_set_ibr(i, ibrs[i]);
                ia64_dv_serialize_instruction();
        }
        ia64_srlz_i();
}

static inline void
pfm_restore_dbrs(unsigned long *dbrs, unsigned int ndbrs)
{
        int i;

        for (i=0; i < ndbrs; i++) {
                ia64_set_dbr(i, dbrs[i]);
                ia64_dv_serialize_data();
        }
        ia64_srlz_d();
}

/*
 * PMD[i] must be a counter. no check is made
 */
static inline unsigned long
pfm_read_soft_counter(pfm_context_t *ctx, int i)
{
        return ctx->ctx_pmds[i].val + (ia64_get_pmd(i) & pmu_conf->ovfl_val);
}

/*
 * PMD[i] must be a counter. no check is made
 */
static inline void
pfm_write_soft_counter(pfm_context_t *ctx, int i, unsigned long val)
{
        unsigned long ovfl_val = pmu_conf->ovfl_val;

        ctx->ctx_pmds[i].val = val  & ~ovfl_val;
        /*
         * writing to unimplemented part is ignore, so we do not need to
         * mask off top part
         */
        ia64_set_pmd(i, val & ovfl_val);
}

static pfm_msg_t *
pfm_get_new_msg(pfm_context_t *ctx)
{
        int idx, next;

        next = (ctx->ctx_msgq_tail+1) % PFM_MAX_MSGS;

        DPRINT(("ctx_fd=%p head=%d tail=%d\n", ctx, ctx->ctx_msgq_head, ctx->ctx_msgq_tail));
        if (next == ctx->ctx_msgq_head) return NULL;

        idx =   ctx->ctx_msgq_tail;
        ctx->ctx_msgq_tail = next;

        DPRINT(("ctx=%p head=%d tail=%d msg=%d\n", ctx, ctx->ctx_msgq_head, ctx->ctx_msgq_tail, idx));

        return ctx->ctx_msgq+idx;
}

static pfm_msg_t *
pfm_get_next_msg(pfm_context_t *ctx)
{
        pfm_msg_t *msg;

        DPRINT(("ctx=%p head=%d tail=%d\n", ctx, ctx->ctx_msgq_head, ctx->ctx_msgq_tail));

        if (PFM_CTXQ_EMPTY(ctx)) return NULL;

        /*
         * get oldest message
         */
        msg = ctx->ctx_msgq+ctx->ctx_msgq_head;

        /*
         * and move forward
         */
        ctx->ctx_msgq_head = (ctx->ctx_msgq_head+1) % PFM_MAX_MSGS;

        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));

        return msg;
}

static void
pfm_reset_msgq(pfm_context_t *ctx)
{
        ctx->ctx_msgq_head = ctx->ctx_msgq_tail = 0;
        DPRINT(("ctx=%p msgq reset\n", ctx));
}

static void *
pfm_rvmalloc(unsigned long size)
{
        void *mem;
        unsigned long addr;

        size = PAGE_ALIGN(size);
        mem  = vmalloc(size);
        if (mem) {
                //printk("perfmon: CPU%d pfm_rvmalloc(%ld)=%p\n", smp_processor_id(), size, mem);
                memset(mem, 0, size);
                addr = (unsigned long)mem;
                while (size > 0) {
                        pfm_reserve_page(addr);
                        addr+=PAGE_SIZE;
                        size-=PAGE_SIZE;
                }
        }
        return mem;
}

static void
pfm_rvfree(void *mem, unsigned long size)
{
        unsigned long addr;

        if (mem) {
                DPRINT(("freeing physical buffer @%p size=%lu\n", mem, size));
                addr = (unsigned long) mem;
                while ((long) size > 0) {
                        pfm_unreserve_page(addr);
                        addr+=PAGE_SIZE;
                        size-=PAGE_SIZE;
                }
                vfree(mem);
        }
        return;
}

static pfm_context_t *
pfm_context_alloc(void)
{
        pfm_context_t *ctx;

        /* 
         * allocate context descriptor 
         * must be able to free with interrupts disabled
         */
        ctx = kmalloc(sizeof(pfm_context_t), GFP_KERNEL);
        if (ctx) {
                memset(ctx, 0, sizeof(pfm_context_t));
                DPRINT(("alloc ctx @%p\n", ctx));
        }
        return ctx;
}

static void
pfm_context_free(pfm_context_t *ctx)
{
        if (ctx) {
                DPRINT(("free ctx @%p\n", ctx));
                kfree(ctx);
        }
}

static void
pfm_mask_monitoring(struct task_struct *task)
{
        pfm_context_t *ctx = PFM_GET_CTX(task);
        struct thread_struct *th = &task->thread;
        unsigned long mask, val, ovfl_mask;
        int i;

        DPRINT_ovfl(("masking monitoring for [%d]\n", task->pid));

        ovfl_mask = pmu_conf->ovfl_val;
        /*
         * monitoring can only be masked as a result of a valid
         * counter overflow. In UP, it means that the PMU still
         * has an owner. Note that the owner can be different
         * from the current task. However the PMU state belongs
         * to the owner.
         * In SMP, a valid overflow only happens when task is
         * current. Therefore if we come here, we know that
         * the PMU state belongs to the current task, therefore
         * we can access the live registers.
         *
         * So in both cases, the live register contains the owner's
         * state. We can ONLY touch the PMU registers and NOT the PSR.
         *
         * As a consequence to this call, the thread->pmds[] array
         * contains stale information which must be ignored
         * when context is reloaded AND monitoring is active (see
         * pfm_restart).
         */
        mask = ctx->ctx_used_pmds[0];
        for (i = 0; mask; i++, mask>>=1) {
                /* skip non used pmds */
                if ((mask & 0x1) == 0) continue;
                val = ia64_get_pmd(i);

                if (PMD_IS_COUNTING(i)) {
                        /*
                         * we rebuild the full 64 bit value of the counter
                         */
                        ctx->ctx_pmds[i].val += (val & ovfl_mask);
                } else {
                        ctx->ctx_pmds[i].val = val;
                }
                DPRINT_ovfl(("pmd[%d]=0x%lx hw_pmd=0x%lx\n",
                        i,
                        ctx->ctx_pmds[i].val,
                        val & ovfl_mask));
        }
        /*
         * mask monitoring by setting the privilege level to 0
         * we cannot use psr.pp/psr.up for this, it is controlled by
         * the user
         *
         * if task is current, modify actual registers, otherwise modify
         * thread save state, i.e., what will be restored in pfm_load_regs()
         */
        mask = ctx->ctx_used_monitors[0] >> PMU_FIRST_COUNTER;
        for(i= PMU_FIRST_COUNTER; mask; i++, mask>>=1) {
                if ((mask & 0x1) == 0UL) continue;
                ia64_set_pmc(i, th->pmcs[i] & ~0xfUL);
                th->pmcs[i] &= ~0xfUL;
                DPRINT_ovfl(("pmc[%d]=0x%lx\n", i, th->pmcs[i]));
        }
        /*
         * make all of this visible
         */
        ia64_srlz_d();
}

/*
 * must always be done with task == current
 *
 * context must be in MASKED state when calling
 */
static void
pfm_restore_monitoring(struct task_struct *task)
{
        pfm_context_t *ctx = PFM_GET_CTX(task);
        struct thread_struct *th = &task->thread;
        unsigned long mask, ovfl_mask;
        unsigned long psr, val;
        int i, is_system;

        is_system = ctx->ctx_fl_system;
        ovfl_mask = pmu_conf->ovfl_val;

        if (task != current) {
                printk(KERN_ERR "perfmon.%d: invalid task[%d] current[%d]\n", __LINE__, task->pid, current->pid);
                return;
        }
        if (ctx->ctx_state != PFM_CTX_MASKED) {
                printk(KERN_ERR "perfmon.%d: task[%d] current[%d] invalid state=%d\n", __LINE__,
                        task->pid, current->pid, ctx->ctx_state);
                return;
        }
        psr = pfm_get_psr();
        /*
         * monitoring is masked via the PMC.
         * As we restore their value, we do not want each counter to
         * restart right away. We stop monitoring using the PSR,
         * restore the PMC (and PMD) and then re-establish the psr
         * as it was. Note that there can be no pending overflow at
         * this point, because monitoring was MASKED.
         *
         * system-wide session are pinned and self-monitoring
         */
        if (is_system && (PFM_CPUINFO_GET() & PFM_CPUINFO_DCR_PP)) {
                /* disable dcr pp */
                ia64_setreg(_IA64_REG_CR_DCR, ia64_getreg(_IA64_REG_CR_DCR) & ~IA64_DCR_PP);
                pfm_clear_psr_pp();
        } else {
                pfm_clear_psr_up();
        }
        /*
         * first, we restore the PMD
         */
        mask = ctx->ctx_used_pmds[0];
        for (i = 0; mask; i++, mask>>=1) {
                /* skip non used pmds */
                if ((mask & 0x1) == 0) continue;

                if (PMD_IS_COUNTING(i)) {
                        /*
                         * we split the 64bit value according to
                         * counter width
                         */
                        val = ctx->ctx_pmds[i].val & ovfl_mask;
                        ctx->ctx_pmds[i].val &= ~ovfl_mask;
                } else {
                        val = ctx->ctx_pmds[i].val;
                }
                ia64_set_pmd(i, val);

                DPRINT(("pmd[%d]=0x%lx hw_pmd=0x%lx\n",
                        i,
                        ctx->ctx_pmds[i].val,
                        val));
        }
        /*
         * restore the PMCs
         */
        mask = ctx->ctx_used_monitors[0] >> PMU_FIRST_COUNTER;
        for(i= PMU_FIRST_COUNTER; mask; i++, mask>>=1) {
                if ((mask & 0x1) == 0UL) continue;
                th->pmcs[i] = ctx->ctx_pmcs[i];
                ia64_set_pmc(i, th->pmcs[i]);
                DPRINT(("[%d] pmc[%d]=0x%lx\n", task->pid, i, th->pmcs[i]));
        }
        ia64_srlz_d();

        /*
         * must restore DBR/IBR because could be modified while masked
         * XXX: need to optimize 
         */
        if (ctx->ctx_fl_using_dbreg) {
                pfm_restore_ibrs(ctx->ctx_ibrs, pmu_conf->num_ibrs);
                pfm_restore_dbrs(ctx->ctx_dbrs, pmu_conf->num_dbrs);
        }

        /*
         * now restore PSR
         */
        if (is_system && (PFM_CPUINFO_GET() & PFM_CPUINFO_DCR_PP)) {
                /* enable dcr pp */
                ia64_setreg(_IA64_REG_CR_DCR, ia64_getreg(_IA64_REG_CR_DCR) | IA64_DCR_PP);
                ia64_srlz_i();
        }
        pfm_set_psr_l(psr);
}

static inline void
pfm_save_pmds(unsigned long *pmds, unsigned long mask)
{
        int i;

        ia64_srlz_d();

        for (i=0; mask; i++, mask>>=1) {
                if (mask & 0x1) pmds[i] = ia64_get_pmd(i);
        }
}

/*
 * reload from thread state (used for ctxw only)
 */
static inline void
pfm_restore_pmds(unsigned long *pmds, unsigned long mask)
{
        int i;
        unsigned long val, ovfl_val = pmu_conf->ovfl_val;

        for (i=0; mask; i++, mask>>=1) {
                if ((mask & 0x1) == 0) continue;
                val = PMD_IS_COUNTING(i) ? pmds[i] & ovfl_val : pmds[i];
                ia64_set_pmd(i, val);
        }
        ia64_srlz_d();
}

/*
 * propagate PMD from context to thread-state
 */
static inline void
pfm_copy_pmds(struct task_struct *task, pfm_context_t *ctx)
{
        struct thread_struct *thread = &task->thread;
        unsigned long ovfl_val = pmu_conf->ovfl_val;
        unsigned long mask = ctx->ctx_all_pmds[0];
        unsigned long val;
        int i;

        DPRINT(("mask=0x%lx\n", mask));

        for (i=0; mask; i++, mask>>=1) {

                val = ctx->ctx_pmds[i].val;

                /*
                 * We break up the 64 bit value into 2 pieces
                 * the lower bits go to the machine state in the
                 * thread (will be reloaded on ctxsw in).
                 * The upper part stays in the soft-counter.
                 */
                if (PMD_IS_COUNTING(i)) {
                        ctx->ctx_pmds[i].val = val & ~ovfl_val;
                         val &= ovfl_val;
                }
                thread->pmds[i] = val;

                DPRINT(("pmd[%d]=0x%lx soft_val=0x%lx\n",
                        i,
                        thread->pmds[i],
                        ctx->ctx_pmds[i].val));
        }
}

/*
 * propagate PMC from context to thread-state
 */
static inline void
pfm_copy_pmcs(struct task_struct *task, pfm_context_t *ctx)
{
        struct thread_struct *thread = &task->thread;
        unsigned long mask = ctx->ctx_all_pmcs[0];
        int i;

        DPRINT(("mask=0x%lx\n", mask));

        for (i=0; mask; i++, mask>>=1) {
                /* masking 0 with ovfl_val yields 0 */
                thread->pmcs[i] = ctx->ctx_pmcs[i];
                DPRINT(("pmc[%d]=0x%lx\n", i, thread->pmcs[i]));
        }
}



static inline void
pfm_restore_pmcs(unsigned long *pmcs, unsigned long mask)
{
        int i;

        for (i=0; mask; i++, mask>>=1) {
                if ((mask & 0x1) == 0) continue;
                ia64_set_pmc(i, pmcs[i]);
        }
        ia64_srlz_d();
}

static inline int
pfm_uuid_cmp(pfm_uuid_t a, pfm_uuid_t b)
{
        return memcmp(a, b, sizeof(pfm_uuid_t));
}

static inline int
pfm_buf_fmt_exit(pfm_buffer_fmt_t *fmt, struct task_struct *task, void *buf, struct pt_regs *regs)
{
        int ret = 0;
        if (fmt->fmt_exit) ret = (*fmt->fmt_exit)(task, buf, regs);
        return ret;
}

static inline int
pfm_buf_fmt_getsize(pfm_buffer_fmt_t *fmt, struct task_struct *task, unsigned int flags, int cpu, void *arg, unsigned long *size)
{
        int ret = 0;
        if (fmt->fmt_getsize) ret = (*fmt->fmt_getsize)(task, flags, cpu, arg, size);
        return ret;
}


static inline int
pfm_buf_fmt_validate(pfm_buffer_fmt_t *fmt, struct task_struct *task, unsigned int flags,
                     int cpu, void *arg)
{
        int ret = 0;
        if (fmt->fmt_validate) ret = (*fmt->fmt_validate)(task, flags, cpu, arg);
        return ret;
}

static inline int
pfm_buf_fmt_init(pfm_buffer_fmt_t *fmt, struct task_struct *task, void *buf, unsigned int flags,
                     int cpu, void *arg)
{
        int ret = 0;
        if (fmt->fmt_init) ret = (*fmt->fmt_init)(task, buf, flags, cpu, arg);
        return ret;
}

static inline int
pfm_buf_fmt_restart(pfm_buffer_fmt_t *fmt, struct task_struct *task, pfm_ovfl_ctrl_t *ctrl, void *buf, struct pt_regs *regs)
{
        int ret = 0;
        if (fmt->fmt_restart) ret = (*fmt->fmt_restart)(task, ctrl, buf, regs);
        return ret;
}

static inline int
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)
{
        int ret = 0;
        if (fmt->fmt_restart_active) ret = (*fmt->fmt_restart_active)(task, ctrl, buf, regs);
        return ret;
}

static pfm_buffer_fmt_t *
__pfm_find_buffer_fmt(pfm_uuid_t uuid)
{
        struct list_head * pos;
        pfm_buffer_fmt_t * entry;

        list_for_each(pos, &pfm_buffer_fmt_list) {
                entry = list_entry(pos, pfm_buffer_fmt_t, fmt_list);
                if (pfm_uuid_cmp(uuid, entry->fmt_uuid) == 0)
                        return entry;
        }
        return NULL;
}
 
/*
 * find a buffer format based on its uuid
 */
static pfm_buffer_fmt_t *
pfm_find_buffer_fmt(pfm_uuid_t uuid)
{
        pfm_buffer_fmt_t * fmt;
        spin_lock(&pfm_buffer_fmt_lock);
        fmt = __pfm_find_buffer_fmt(uuid);
        spin_unlock(&pfm_buffer_fmt_lock);
        return fmt;
}
 
int
pfm_register_buffer_fmt(pfm_buffer_fmt_t *fmt)
{
        int ret = 0;

        /* some sanity checks */
        if (fmt == NULL || fmt->fmt_name == NULL) return -EINVAL;

        /* we need at least a handler */
        if (fmt->fmt_handler == NULL) return -EINVAL;

        /*
         * XXX: need check validity of fmt_arg_size
         */

        spin_lock(&pfm_buffer_fmt_lock);

        if (__pfm_find_buffer_fmt(fmt->fmt_uuid)) {
                printk(KERN_ERR "perfmon: duplicate sampling format: %s\n", fmt->fmt_name);
                ret = -EBUSY;
                goto out;
        } 
        list_add(&fmt->fmt_list, &pfm_buffer_fmt_list);
        printk(KERN_INFO "perfmon: added sampling format %s\n", fmt->fmt_name);

out:
        spin_unlock(&pfm_buffer_fmt_lock);
        return ret;
}
EXPORT_SYMBOL(pfm_register_buffer_fmt);

int
pfm_unregister_buffer_fmt(pfm_uuid_t uuid)
{
        pfm_buffer_fmt_t *fmt;
        int ret = 0;

        spin_lock(&pfm_buffer_fmt_lock);

        fmt = __pfm_find_buffer_fmt(uuid);
        if (!fmt) {
                printk(KERN_ERR "perfmon: cannot unregister format, not found\n");
                ret = -EINVAL;
                goto out;
        }
        list_del_init(&fmt->fmt_list);
        printk(KERN_INFO "perfmon: removed sampling format: %s\n", fmt->fmt_name);

out:
        spin_unlock(&pfm_buffer_fmt_lock);
        return ret;

}
EXPORT_SYMBOL(pfm_unregister_buffer_fmt);

static int
pfm_reserve_session(struct task_struct *task, int is_syswide, unsigned int cpu)
{
        unsigned long flags;
        /*
         * validy checks on cpu_mask have been done upstream
         */
        LOCK_PFS(flags);

        DPRINT(("in sys_sessions=%u task_sessions=%u dbregs=%u syswide=%d cpu=%u\n",
                pfm_sessions.pfs_sys_sessions,
                pfm_sessions.pfs_task_sessions,
                pfm_sessions.pfs_sys_use_dbregs,
                is_syswide,
                cpu));

        if (is_syswide) {
                /*
                 * cannot mix system wide and per-task sessions
                 */
                if (pfm_sessions.pfs_task_sessions > 0UL) {
                        DPRINT(("system wide not possible, %u conflicting task_sessions\n",
                                pfm_sessions.pfs_task_sessions));
                        goto abort;
                }

                if (pfm_sessions.pfs_sys_session[cpu]) goto error_conflict;

                DPRINT(("reserving system wide session on CPU%u currently on CPU%u\n", cpu, smp_processor_id()));

                pfm_sessions.pfs_sys_session[cpu] = task;

                pfm_sessions.pfs_sys_sessions++ ;

        } else {
                if (pfm_sessions.pfs_sys_sessions) goto abort;
                pfm_sessions.pfs_task_sessions++;
        }

        DPRINT(("out sys_sessions=%u task_sessions=%u dbregs=%u syswide=%d cpu=%u\n",
                pfm_sessions.pfs_sys_sessions,
                pfm_sessions.pfs_task_sessions,
                pfm_sessions.pfs_sys_use_dbregs,
                is_syswide,
                cpu));

        UNLOCK_PFS(flags);

        return 0;

error_conflict:
        DPRINT(("system wide not possible, conflicting session [%d] on CPU%d\n",
                pfm_sessions.pfs_sys_session[cpu]->pid,
                smp_processor_id()));
abort:
        UNLOCK_PFS(flags);

        return -EBUSY;

}

static int
pfm_unreserve_session(pfm_context_t *ctx, int is_syswide, unsigned int cpu)
{
        unsigned long flags;
        /*
         * validy checks on cpu_mask have been done upstream
         */
        LOCK_PFS(flags);

        DPRINT(("in sys_sessions=%u task_sessions=%u dbregs=%u syswide=%d cpu=%u\n",
                pfm_sessions.pfs_sys_sessions,
                pfm_sessions.pfs_task_sessions,
                pfm_sessions.pfs_sys_use_dbregs,
                is_syswide,
                cpu));


        if (is_syswide) {
                pfm_sessions.pfs_sys_session[cpu] = NULL;
                /*
                 * would not work with perfmon+more than one bit in cpu_mask
                 */
                if (ctx && ctx->ctx_fl_using_dbreg) {
                        if (pfm_sessions.pfs_sys_use_dbregs == 0) {
                                printk(KERN_ERR "perfmon: invalid release for ctx %p sys_use_dbregs=0\n", ctx);
                        } else {
                                pfm_sessions.pfs_sys_use_dbregs--;
                        }
                }
                pfm_sessions.pfs_sys_sessions--;
        } else {
                pfm_sessions.pfs_task_sessions--;
        }
        DPRINT(("out sys_sessions=%u task_sessions=%u dbregs=%u syswide=%d cpu=%u\n",
                pfm_sessions.pfs_sys_sessions,
                pfm_sessions.pfs_task_sessions,
                pfm_sessions.pfs_sys_use_dbregs,
                is_syswide,
                cpu));

        UNLOCK_PFS(flags);

        return 0;
}

/*
 * removes virtual mapping of the sampling buffer.
 * IMPORTANT: cannot be called with interrupts disable, e.g. inside
 * a PROTECT_CTX() section.
 */
static int
pfm_remove_smpl_mapping(struct task_struct *task, void *vaddr, unsigned long size)
{
        int r;

        /* sanity checks */
        if (task->mm == NULL || size == 0UL || vaddr == NULL) {
                printk(KERN_ERR "perfmon: pfm_remove_smpl_mapping [%d] invalid context mm=%p\n", task->pid, task->mm);
                return -EINVAL;
        }

        DPRINT(("smpl_vaddr=%p size=%lu\n", vaddr, size));

        /*
         * does the actual unmapping
         */
        down_write(&task->mm->mmap_sem);

        DPRINT(("down_write done smpl_vaddr=%p size=%lu\n", vaddr, size));

        r = pfm_do_munmap(task->mm, (unsigned long)vaddr, size, 0);

        up_write(&task->mm->mmap_sem);
        if (r !=0) {
                printk(KERN_ERR "perfmon: [%d] unable to unmap sampling buffer @%p size=%lu\n", task->pid, vaddr, size);
        }

        DPRINT(("do_unmap(%p, %lu)=%d\n", vaddr, size, r));

        return 0;
}

/*
 * free actual physical storage used by sampling buffer
 */
#if 0
static int
pfm_free_smpl_buffer(pfm_context_t *ctx)
{
        pfm_buffer_fmt_t *fmt;

        if (ctx->ctx_smpl_hdr == NULL) goto invalid_free;

        /*
         * we won't use the buffer format anymore
         */
        fmt = ctx->ctx_buf_fmt;

        DPRINT(("sampling buffer @%p size %lu vaddr=%p\n",
                ctx->ctx_smpl_hdr,
                ctx->ctx_smpl_size,
                ctx->ctx_smpl_vaddr));

        pfm_buf_fmt_exit(fmt, current, NULL, NULL);

        /*
         * free the buffer
         */
        pfm_rvfree(ctx->ctx_smpl_hdr, ctx->ctx_smpl_size);

        ctx->ctx_smpl_hdr  = NULL;
        ctx->ctx_smpl_size = 0UL;

        return 0;

invalid_free:
        printk(KERN_ERR "perfmon: pfm_free_smpl_buffer [%d] no buffer\n", current->pid);
        return -EINVAL;
}
#endif

static inline void
pfm_exit_smpl_buffer(pfm_buffer_fmt_t *fmt)
{
        if (fmt == NULL) return;

        pfm_buf_fmt_exit(fmt, current, NULL, NULL);

}

/*
 * pfmfs should _never_ be mounted by userland - too much of security hassle,
 * no real gain from having the whole whorehouse mounted. So we don't need
 * any operations on the root directory. However, we need a non-trivial
 * d_name - pfm: will go nicely and kill the special-casing in procfs.
 */
static struct vfsmount *pfmfs_mnt;

static int __init
init_pfm_fs(void)
{
        int err = register_filesystem(&pfm_fs_type);
        if (!err) {
                pfmfs_mnt = kern_mount(&pfm_fs_type);
                err = PTR_ERR(pfmfs_mnt);
                if (IS_ERR(pfmfs_mnt))
                        unregister_filesystem(&pfm_fs_type);
                else
                        err = 0;
        }
        return err;
}

static void __exit
exit_pfm_fs(void)
{
        unregister_filesystem(&pfm_fs_type);
        mntput(pfmfs_mnt);
}

static ssize_t
pfm_read(struct file *filp, char __user *buf, size_t size, loff_t *ppos)
{
        pfm_context_t *ctx;
        pfm_msg_t *msg;
        ssize_t ret;
        unsigned long flags;
        DECLARE_WAITQUEUE(wait, current);
        if (PFM_IS_FILE(filp) == 0) {
                printk(KERN_ERR "perfmon: pfm_poll: bad magic [%d]\n", current->pid);
                return -EINVAL;
        }

        ctx = (pfm_context_t *)filp->private_data;
        if (ctx == NULL) {
                printk(KERN_ERR "perfmon: pfm_read: NULL ctx [%d]\n", current->pid);
                return -EINVAL;
        }

        /*
         * check even when there is no message
         */
        if (size < sizeof(pfm_msg_t)) {
                DPRINT(("message is too small ctx=%p (>=%ld)\n", ctx, sizeof(pfm_msg_t)));
                return -EINVAL;
        }

        PROTECT_CTX(ctx, flags);

        /*
         * put ourselves on the wait queue
         */
        add_wait_queue(&ctx->ctx_msgq_wait, &wait);


        for(;;) {
                /*
                 * check wait queue
                 */

                set_current_state(TASK_INTERRUPTIBLE);

                DPRINT(("head=%d tail=%d\n", ctx->ctx_msgq_head, ctx->ctx_msgq_tail));

                ret = 0;
                if(PFM_CTXQ_EMPTY(ctx) == 0) break;

                UNPROTECT_CTX(ctx, flags);

                /*
                 * check non-blocking read
                 */
                ret = -EAGAIN;
                if(filp->f_flags & O_NONBLOCK) break;

                /*
                 * check pending signals
                 */
                if(signal_pending(current)) {
                        ret = -EINTR;
                        break;
                }
                /*
                 * no message, so wait
                 */
                schedule();

                PROTECT_CTX(ctx, flags);
        }
        DPRINT(("[%d] back to running ret=%ld\n", current->pid, ret));
        set_current_state(TASK_RUNNING);
        remove_wait_queue(&ctx->ctx_msgq_wait, &wait);

        if (ret < 0) goto abort;

        ret = -EINVAL;
        msg = pfm_get_next_msg(ctx);
        if (msg == NULL) {
                printk(KERN_ERR "perfmon: pfm_read no msg for ctx=%p [%d]\n", ctx, current->pid);
                goto abort_locked;
        }

        DPRINT(("[%d] fd=%d type=%d\n", current->pid, msg->pfm_gen_msg.msg_ctx_fd, msg->pfm_gen_msg.msg_type));

        ret = -EFAULT;
        if(copy_to_user(buf, msg, sizeof(pfm_msg_t)) == 0) ret = sizeof(pfm_msg_t);

abort_locked:
        UNPROTECT_CTX(ctx, flags);
abort:
        return ret;
}

static ssize_t
pfm_write(struct file *file, const char __user *ubuf,
                          size_t size, loff_t *ppos)
{
        DPRINT(("pfm_write called\n"));
        return -EINVAL;
}

static unsigned int
pfm_poll(struct file *filp, poll_table * wait)
{
        pfm_context_t *ctx;
        unsigned long flags;
        unsigned int mask = 0;

        if (PFM_IS_FILE(filp) == 0) {
                printk(KERN_ERR "perfmon: pfm_poll: bad magic [%d]\n", current->pid);
                return 0;
        }

        ctx = (pfm_context_t *)filp->private_data;
        if (ctx == NULL) {
                printk(KERN_ERR "perfmon: pfm_poll: NULL ctx [%d]\n", current->pid);
                return 0;
        }


        DPRINT(("pfm_poll ctx_fd=%d before poll_wait\n", ctx->ctx_fd));

        poll_wait(filp, &ctx->ctx_msgq_wait, wait);

        PROTECT_CTX(ctx, flags);

        if (PFM_CTXQ_EMPTY(ctx) == 0)
                mask =  POLLIN | POLLRDNORM;

        UNPROTECT_CTX(ctx, flags);

        DPRINT(("pfm_poll ctx_fd=%d mask=0x%x\n", ctx->ctx_fd, mask));

        return mask;
}

static int
pfm_ioctl(struct inode *inode, struct file *file, unsigned int cmd, unsigned long arg)
{
        DPRINT(("pfm_ioctl called\n"));
        return -EINVAL;
}

/*
 * interrupt cannot be masked when coming here
 */
static inline int
pfm_do_fasync(int fd, struct file *filp, pfm_context_t *ctx, int on)
{
        int ret;

        ret = fasync_helper (fd, filp, on, &ctx->ctx_async_queue);

        DPRINT(("pfm_fasync called by [%d] on ctx_fd=%d on=%d async_queue=%p ret=%d\n",
                current->pid,
                fd,
                on,
                ctx->ctx_async_queue, ret));

        return ret;
}

static int
pfm_fasync(int fd, struct file *filp, int on)
{
        pfm_context_t *ctx;
        int ret;

        if (PFM_IS_FILE(filp) == 0) {
                printk(KERN_ERR "perfmon: pfm_fasync bad magic [%d]\n", current->pid);
                return -EBADF;
        }

        ctx = (pfm_context_t *)filp->private_data;
        if (ctx == NULL) {
                printk(KERN_ERR "perfmon: pfm_fasync NULL ctx [%d]\n", current->pid);
                return -EBADF;
        }
        /*
         * we cannot mask interrupts during this call because this may
         * may go to sleep if memory is not readily avalaible.
         *
         * We are protected from the conetxt disappearing by the get_fd()/put_fd()
         * done in caller. Serialization of this function is ensured by caller.
         */
        ret = pfm_do_fasync(fd, filp, ctx, on);


        DPRINT(("pfm_fasync called on ctx_fd=%d on=%d async_queue=%p ret=%d\n",
                fd,
                on,
                ctx->ctx_async_queue, ret));

        return ret;
}

#ifdef CONFIG_SMP
/*
 * this function is exclusively called from pfm_close().
 * The context is not protected at that time, nor are interrupts
 * on the remote CPU. That's necessary to avoid deadlocks.
 */
static void
pfm_syswide_force_stop(void *info)
{
        pfm_context_t   *ctx = (pfm_context_t *)info;
        struct pt_regs *regs = ia64_task_regs(current);
        struct task_struct *owner;
        unsigned long flags;
        int ret;

        if (ctx->ctx_cpu != smp_processor_id()) {
                printk(KERN_ERR "perfmon: pfm_syswide_force_stop for CPU%d  but on CPU%d\n",
                        ctx->ctx_cpu,
                        smp_processor_id());
                return;
        }
        owner = GET_PMU_OWNER();
        if (owner != ctx->ctx_task) {
                printk(KERN_ERR "perfmon: pfm_syswide_force_stop CPU%d unexpected owner [%d] instead of [%d]\n",
                        smp_processor_id(),
                        owner->pid, ctx->ctx_task->pid);
                return;
        }
        if (GET_PMU_CTX() != ctx) {
                printk(KERN_ERR "perfmon: pfm_syswide_force_stop CPU%d unexpected ctx %p instead of %p\n",
                        smp_processor_id(),
                        GET_PMU_CTX(), ctx);
                return;
        }

        DPRINT(("on CPU%d forcing system wide stop for [%d]\n", smp_processor_id(), ctx->ctx_task->pid));       
        /*
         * the context is already protected in pfm_close(), we simply
         * need to mask interrupts to avoid a PMU interrupt race on
         * this CPU
         */
        local_irq_save(flags);

        ret = pfm_context_unload(ctx, NULL, 0, regs);
        if (ret) {
                DPRINT(("context_unload returned %d\n", ret));
        }

        /*
         * unmask interrupts, PMU interrupts are now spurious here
         */
        local_irq_restore(flags);
}

static void
pfm_syswide_cleanup_other_cpu(pfm_context_t *ctx)
{
        int ret;

        DPRINT(("calling CPU%d for cleanup\n", ctx->ctx_cpu));
        ret = smp_call_function_single(ctx->ctx_cpu, pfm_syswide_force_stop, ctx, 0, 1);
        DPRINT(("called CPU%d for cleanup ret=%d\n", ctx->ctx_cpu, ret));
}
#endif /* CONFIG_SMP */

/*
 * called for each close(). Partially free resources.
 * When caller is self-monitoring, the context is unloaded.
 */
static int
pfm_flush(struct file *filp)
{
        pfm_context_t *ctx;
        struct task_struct *task;
        struct pt_regs *regs;
        unsigned long flags;
        unsigned long smpl_buf_size = 0UL;
        void *smpl_buf_vaddr = NULL;
        int state, is_system;

        if (PFM_IS_FILE(filp) == 0) {
                DPRINT(("bad magic for\n"));
                return -EBADF;
        }

        ctx = (pfm_context_t *)filp->private_data;
        if (ctx == NULL) {
                printk(KERN_ERR "perfmon: pfm_flush: NULL ctx [%d]\n", current->pid);
                return -EBADF;
        }

        /*
         * remove our file from the async queue, if we use this mode.
         * This can be done without the context being protected. We come
         * here when the context has become unreacheable by other tasks.
         *
         * We may still have active monitoring at this point and we may
         * end up in pfm_overflow_handler(). However, fasync_helper()
         * operates with interrupts disabled and it cleans up the
         * queue. If the PMU handler is called prior to entering
         * fasync_helper() then it will send a signal. If it is
         * invoked after, it will find an empty queue and no
         * signal will be sent. In both case, we are safe
         */
        if (filp->f_flags & FASYNC) {
                DPRINT(("cleaning up async_queue=%p\n", ctx->ctx_async_queue));
                pfm_do_fasync (-1, filp, ctx, 0);
        }

        PROTECT_CTX(ctx, flags);

        state     = ctx->ctx_state;
        is_system = ctx->ctx_fl_system;

        task = PFM_CTX_TASK(ctx);
        regs = ia64_task_regs(task);

        DPRINT(("ctx_state=%d is_current=%d\n",
                state,
                task == current ? 1 : 0));

        /*
         * if state == UNLOADED, then task is NULL
         */

        /*
         * we must stop and unload because we are losing access to the context.
         */
        if (task == current) {
#ifdef CONFIG_SMP
                /*
                 * the task IS the owner but it migrated to another CPU: that's bad
                 * but we must handle this cleanly. Unfortunately, the kernel does
                 * not provide a mechanism to block migration (while the context is loaded).
                 *
                 * We need to release the resource on the ORIGINAL cpu.
                 */
                if (is_system && ctx->ctx_cpu != smp_processor_id()) {

                        DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
                        /*
                         * keep context protected but unmask interrupt for IPI
                         */
                        local_irq_restore(flags);

                        pfm_syswide_cleanup_other_cpu(ctx);

                        /*
                         * restore interrupt masking
                         */
                        local_irq_save(flags);

                        /*
                         * context is unloaded at this point
                         */
                } else
#endif /* CONFIG_SMP */
                {

                        DPRINT(("forcing unload\n"));
                        /*
                        * stop and unload, returning with state UNLOADED
                        * and session unreserved.
                        */
                        pfm_context_unload(ctx, NULL, 0, regs);

                        DPRINT(("ctx_state=%d\n", ctx->ctx_state));
                }
        }

        /*
         * remove virtual mapping, if any, for the calling task.
         * cannot reset ctx field until last user is calling close().
         *
         * ctx_smpl_vaddr must never be cleared because it is needed
         * by every task with access to the context
         *
         * When called from do_exit(), the mm context is gone already, therefore
         * mm is NULL, i.e., the VMA is already gone  and we do not have to
         * do anything here
         */
        if (ctx->ctx_smpl_vaddr && current->mm) {
                smpl_buf_vaddr = ctx->ctx_smpl_vaddr;
                smpl_buf_size  = ctx->ctx_smpl_size;
        }

        UNPROTECT_CTX(ctx, flags);

        /*
         * if there was a mapping, then we systematically remove it
         * at this point. Cannot be done inside critical section
         * because some VM function reenables interrupts.
         *
         */
        if (smpl_buf_vaddr) pfm_remove_smpl_mapping(current, smpl_buf_vaddr, smpl_buf_size);

        return 0;
}
/*
 * called either on explicit close() or from exit_files(). 
 * Only the LAST user of the file gets to this point, i.e., it is
 * called only ONCE.
 *
 * IMPORTANT: we get called ONLY when the refcnt on the file gets to zero 
 * (fput()),i.e, last task to access the file. Nobody else can access the 
 * file at this point.
 *
 * When called from exit_files(), the VMA has been freed because exit_mm()
 * is executed before exit_files().
 *
 * When called from exit_files(), the current task is not yet ZOMBIE but we
 * flush the PMU state to the context. 
 */
static int
pfm_close(struct inode *inode, struct file *filp)
{
        pfm_context_t *ctx;
        struct task_struct *task;
        struct pt_regs *regs;
        DECLARE_WAITQUEUE(wait, current);
        unsigned long flags;
        unsigned long smpl_buf_size = 0UL;
        void *smpl_buf_addr = NULL;
        int free_possible = 1;
        int state, is_system;

        DPRINT(("pfm_close called private=%p\n", filp->private_data));

        if (PFM_IS_FILE(filp) == 0) {
                DPRINT(("bad magic\n"));
                return -EBADF;
        }
        
        ctx = (pfm_context_t *)filp->private_data;
        if (ctx == NULL) {
                printk(KERN_ERR "perfmon: pfm_close: NULL ctx [%d]\n", current->pid);
                return -EBADF;
        }

        PROTECT_CTX(ctx, flags);

        state     = ctx->ctx_state;
        is_system = ctx->ctx_fl_system;

        task = PFM_CTX_TASK(ctx);
        regs = ia64_task_regs(task);

        DPRINT(("ctx_state=%d is_current=%d\n", 
                state,
                task == current ? 1 : 0));

        /*
         * if task == current, then pfm_flush() unloaded the context
         */
        if (state == PFM_CTX_UNLOADED) goto doit;

        /*
         * context is loaded/masked and task != current, we need to
         * either force an unload or go zombie
         */

        /*
         * The task is currently blocked or will block after an overflow.
         * we must force it to wakeup to get out of the
         * MASKED state and transition to the unloaded state by itself.
         *
         * This situation is only possible for per-task mode
         */
        if (state == PFM_CTX_MASKED && CTX_OVFL_NOBLOCK(ctx) == 0) {

                /*
                 * set a "partial" zombie state to be checked
                 * upon return from down() in pfm_handle_work().
                 *
                 * We cannot use the ZOMBIE state, because it is checked
                 * by pfm_load_regs() which is called upon wakeup from down().
                 * In such case, it would free the context and then we would
                 * return to pfm_handle_work() which would access the
                 * stale context. Instead, we set a flag invisible to pfm_load_regs()
                 * but visible to pfm_handle_work().
                 *
                 * For some window of time, we have a zombie context with
                 * ctx_state = MASKED  and not ZOMBIE
                 */
                ctx->ctx_fl_going_zombie = 1;

                /*
                 * force task to wake up from MASKED state
                 */
                up(&ctx->ctx_restart_sem);

                DPRINT(("waking up ctx_state=%d\n", state));

                /*
                 * put ourself to sleep waiting for the other
                 * task to report completion
                 *
                 * the context is protected by mutex, therefore there
                 * is no risk of being notified of completion before
                 * begin actually on the waitq.
                 */
                set_current_state(TASK_INTERRUPTIBLE);
                add_wait_queue(&ctx->ctx_zombieq, &wait);

                UNPROTECT_CTX(ctx, flags);

                /*
                 * XXX: check for signals :

                 *      - ok for explicit close
                 *      - not ok when coming from exit_files()
                 */
                schedule();


                PROTECT_CTX(ctx, flags);


                remove_wait_queue(&ctx->ctx_zombieq, &wait);
                set_current_state(TASK_RUNNING);

                /*
                 * context is unloaded at this point
                 */
                DPRINT(("after zombie wakeup ctx_state=%d for\n", state));
        }
        else if (task != current) {
#ifdef CONFIG_SMP
                /*
                 * switch context to zombie state
                 */
                ctx->ctx_state = PFM_CTX_ZOMBIE;

                DPRINT(("zombie ctx for [%d]\n", task->pid));
                /*
                 * cannot free the context on the spot. deferred until
                 * the task notices the ZOMBIE state
                 */
                free_possible = 0;
#else
                pfm_context_unload(ctx, NULL, 0, regs);
#endif
        }

doit:
        /* reload state, may have changed during  opening of critical section */
        state = ctx->ctx_state;

        /*
         * the context is still attached to a task (possibly current)
         * we cannot destroy it right now
         */

        /*
         * we must free the sampling buffer right here because
         * we cannot rely on it being cleaned up later by the
         * monitored task. It is not possible to free vmalloc'ed
         * memory in pfm_load_regs(). Instead, we remove the buffer
         * now. should there be subsequent PMU overflow originally
         * meant for sampling, the will be converted to spurious
         * and that's fine because the monitoring tools is gone anyway.
         */
        if (ctx->ctx_smpl_hdr) {
                smpl_buf_addr = ctx->ctx_smpl_hdr;
                smpl_buf_size = ctx->ctx_smpl_size;
                /* no more sampling */
                ctx->ctx_smpl_hdr = NULL;
                ctx->ctx_fl_is_sampling = 0;
        }

        DPRINT(("ctx_state=%d free_possible=%d addr=%p size=%lu\n",
                state,
                free_possible,
                smpl_buf_addr,
                smpl_buf_size));

        if (smpl_buf_addr) pfm_exit_smpl_buffer(ctx->ctx_buf_fmt);

        /*
         * UNLOADED that the session has already been unreserved.
         */
        if (state == PFM_CTX_ZOMBIE) {
                pfm_unreserve_session(ctx, ctx->ctx_fl_system , ctx->ctx_cpu);
        }

        /*
         * disconnect file descriptor from context must be done
         * before we unlock.
         */
        filp->private_data = NULL;

        /*
         * if we free on the spot, the context is now completely unreacheable
         * from the callers side. The monitored task side is also cut, so we
         * can freely cut.
         *
         * If we have a deferred free, only the caller side is disconnected.
         */
        UNPROTECT_CTX(ctx, flags);

        /*
         * All memory free operations (especially for vmalloc'ed memory)
         * MUST be done with interrupts ENABLED.
         */
        if (smpl_buf_addr)  pfm_rvfree(smpl_buf_addr, smpl_buf_size);

        /*
         * return the memory used by the context
         */
        if (free_possible) pfm_context_free(ctx);

        return 0;
}

static int
pfm_no_open(struct inode *irrelevant, struct file *dontcare)
{
        DPRINT(("pfm_no_open called\n"));
        return -ENXIO;
}



static struct file_operations pfm_file_ops = {
        .llseek   = no_llseek,
        .read     = pfm_read,
        .write    = pfm_write,
        .poll     = pfm_poll,
        .ioctl    = pfm_ioctl,
        .open     = pfm_no_open,        /* special open code to disallow open via /proc */
        .fasync   = pfm_fasync,
        .release  = pfm_close,
        .flush    = pfm_flush
};

static int
pfmfs_delete_dentry(struct dentry *dentry)
{
        return 1;
}

static struct dentry_operations pfmfs_dentry_operations = {
        .d_delete = pfmfs_delete_dentry,
};


static int
pfm_alloc_fd(struct file **cfile)
{
        int fd, ret = 0;
        struct file *file = NULL;
        struct inode * inode;
        char name[32];
        struct qstr this;

        fd = get_unused_fd();
        if (fd < 0) return -ENFILE;

        ret = -ENFILE;

        file = get_empty_filp();
        if (!file) goto out;

        /*
         * allocate a new inode
         */
        inode = new_inode(pfmfs_mnt->mnt_sb);
        if (!inode) goto out;

        DPRINT(("new inode ino=%ld @%p\n", inode->i_ino, inode));

        inode->i_mode = S_IFCHR|S_IRUGO;
        inode->i_uid  = current->fsuid;
        inode->i_gid  = current->fsgid;

        sprintf(name, "[%lu]", inode->i_ino);
        this.name = name;
        this.len  = strlen(name);
        this.hash = inode->i_ino;

        ret = -ENOMEM;

        /*
         * allocate a new dcache entry
         */
        file->f_dentry = d_alloc(pfmfs_mnt->mnt_sb->s_root, &this);
        if (!file->f_dentry) goto out;

        file->f_dentry->d_op = &pfmfs_dentry_operations;

        d_add(file->f_dentry, inode);
        file->f_vfsmnt = mntget(pfmfs_mnt);
        file->f_mapping = inode->i_mapping;

        file->f_op    = &pfm_file_ops;
        file->f_mode  = FMODE_READ;
        file->f_flags = O_RDONLY;
        file->f_pos   = 0;

        /*
         * may have to delay until context is attached?
         */
        fd_install(fd, file);

        /*
         * the file structure we will use
         */
        *cfile = file;

        return fd;
out:
        if (file) put_filp(file);
        put_unused_fd(fd);
        return ret;
}

static void
pfm_free_fd(int fd, struct file *file)
{
        struct files_struct *files = current->files;

        /* 
         * there ie no fd_uninstall(), so we do it here
         */
        spin_lock(&files->file_lock);
        files->fd[fd] = NULL;
        spin_unlock(&files->file_lock);

        if (file) put_filp(file);
        put_unused_fd(fd);
}

static int
pfm_remap_buffer(struct vm_area_struct *vma, unsigned long buf, unsigned long addr, unsigned long size)
{
        DPRINT(("CPU%d buf=0x%lx addr=0x%lx size=%ld\n", smp_processor_id(), buf, addr, size));

        while (size > 0) {
                unsigned long pfn = ia64_tpa(buf) >> PAGE_SHIFT;


                if (remap_pfn_range(vma, addr, pfn, PAGE_SIZE, PAGE_READONLY))
                        return -ENOMEM;

                addr  += PAGE_SIZE;
                buf   += PAGE_SIZE;
                size  -= PAGE_SIZE;
        }
        return 0;
}

/*
 * allocate a sampling buffer and remaps it into the user address space of the task
 */
static int
pfm_smpl_buffer_alloc(struct task_struct *task, pfm_context_t *ctx, unsigned long rsize, void **user_vaddr)
{
        struct mm_struct *mm = task->mm;
        struct vm_area_struct *vma = NULL;
        unsigned long size;
        void *smpl_buf;


        /*
         * the fixed header + requested size and align to page boundary
         */
        size = PAGE_ALIGN(rsize);

        DPRINT(("sampling buffer rsize=%lu size=%lu bytes\n", rsize, size));

        /*
         * check requested size to avoid Denial-of-service attacks
         * XXX: may have to refine this test
         * Check against address space limit.
         *
         * if ((mm->total_vm << PAGE_SHIFT) + len> task->rlim[RLIMIT_AS].rlim_cur)
         *      return -ENOMEM;
         */
        if (size > task->signal->rlim[RLIMIT_MEMLOCK].rlim_cur)
                return -ENOMEM;

        /*
         * We do the easy to undo allocations first.
         *
         * pfm_rvmalloc(), clears the buffer, so there is no leak
         */
        smpl_buf = pfm_rvmalloc(size);
        if (smpl_buf == NULL) {
                DPRINT(("Can't allocate sampling buffer\n"));
                return -ENOMEM;
        }

        DPRINT(("smpl_buf @%p\n", smpl_buf));

        /* allocate vma */
        vma = kmem_cache_alloc(vm_area_cachep, SLAB_KERNEL);
        if (!vma) {
                DPRINT(("Cannot allocate vma\n"));
                goto error_kmem;
        }
        memset(vma, 0, sizeof(*vma));

        /*
         * partially initialize the vma for the sampling buffer
         */
        vma->vm_mm           = mm;
        vma->vm_flags        = VM_READ| VM_MAYREAD |VM_RESERVED;
        vma->vm_page_prot    = PAGE_READONLY; /* XXX may need to change */

        /*
         * Now we have everything we need and we can initialize
         * and connect all the data structures
         */

        ctx->ctx_smpl_hdr   = smpl_buf;
        ctx->ctx_smpl_size  = size; /* aligned size */

        /*
         * Let's do the difficult operations next.
         *
         * now we atomically find some area in the address space and
         * remap the buffer in it.
         */
        down_write(&task->mm->mmap_sem);

        /* find some free area in address space, must have mmap sem held */
        vma->vm_start = pfm_get_unmapped_area(NULL, 0, size, 0, MAP_PRIVATE|MAP_ANONYMOUS, 0);
        if (vma->vm_start == 0UL) {
                DPRINT(("Cannot find unmapped area for size %ld\n", size));
                up_write(&task->mm->mmap_sem);
                goto error;
        }
        vma->vm_end = vma->vm_start + size;
        vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT;

        DPRINT(("aligned size=%ld, hdr=%p mapped @0x%lx\n", size, ctx->ctx_smpl_hdr, vma->vm_start));

        /* can only be applied to current task, need to have the mm semaphore held when called */
        if (pfm_remap_buffer(vma, (unsigned long)smpl_buf, vma->vm_start, size)) {
                DPRINT(("Can't remap buffer\n"));
                up_write(&task->mm->mmap_sem);
                goto error;
        }

        /*
         * now insert the vma in the vm list for the process, must be
         * done with mmap lock held
         */
        insert_vm_struct(mm, vma);

        mm->total_vm  += size >> PAGE_SHIFT;
        vm_stat_account(vma);
        up_write(&task->mm->mmap_sem);

        /*
         * keep track of user level virtual address
         */
        ctx->ctx_smpl_vaddr = (void *)vma->vm_start;
        *(unsigned long *)user_vaddr = vma->vm_start;

        return 0;

error:
        kmem_cache_free(vm_area_cachep, vma);
error_kmem:
        pfm_rvfree(smpl_buf, size);

        return -ENOMEM;
}

/*
 * XXX: do something better here
 */
static int
pfm_bad_permissions(struct task_struct *task)
{
        /* inspired by ptrace_attach() */
        DPRINT(("cur: uid=%d gid=%d task: euid=%d suid=%d uid=%d egid=%d sgid=%d\n",
                current->uid,
                current->gid,
                task->euid,
                task->suid,
                task->uid,
                task->egid,
                task->sgid));

        return ((current->uid != task->euid)
            || (current->uid != task->suid)
            || (current->uid != task->uid)
            || (current->gid != task->egid)
            || (current->gid != task->sgid)
            || (current->gid != task->gid)) && !capable(CAP_SYS_PTRACE);
}

static int
pfarg_is_sane(struct task_struct *task, pfarg_context_t *pfx)
{
        int ctx_flags;

        /* valid signal */

        ctx_flags = pfx->ctx_flags;

        if (ctx_flags & PFM_FL_SYSTEM_WIDE) {

                /*
                 * cannot block in this mode
                 */
                if (ctx_flags & PFM_FL_NOTIFY_BLOCK) {
                        DPRINT(("cannot use blocking mode when in system wide monitoring\n"));
                        return -EINVAL;
                }
        } else {
        }
        /* probably more to add here */

        return 0;
}

static int
pfm_setup_buffer_fmt(struct task_struct *task, pfm_context_t *ctx, unsigned int ctx_flags,
                     unsigned int cpu, pfarg_context_t *arg)
{
        pfm_buffer_fmt_t *fmt = NULL;
        unsigned long size = 0UL;
        void *uaddr = NULL;
        void *fmt_arg = NULL;
        int ret = 0;
#define PFM_CTXARG_BUF_ARG(a)   (pfm_buffer_fmt_t *)(a+1)

        /* invoke and lock buffer format, if found */
        fmt = pfm_find_buffer_fmt(arg->ctx_smpl_buf_id);
        if (fmt == NULL) {
                DPRINT(("[%d] cannot find buffer format\n", task->pid));
                return -EINVAL;
        }

        /*
         * buffer argument MUST be contiguous to pfarg_context_t
         */
        if (fmt->fmt_arg_size) fmt_arg = PFM_CTXARG_BUF_ARG(arg);

        ret = pfm_buf_fmt_validate(fmt, task, ctx_flags, cpu, fmt_arg);

        DPRINT(("[%d] after validate(0x%x,%d,%p)=%d\n", task->pid, ctx_flags, cpu, fmt_arg, ret));

        if (ret) goto error;

        /* link buffer format and context */
        ctx->ctx_buf_fmt = fmt;

        /*
         * check if buffer format wants to use perfmon buffer allocation/mapping service
         */
        ret = pfm_buf_fmt_getsize(fmt, task, ctx_flags, cpu, fmt_arg, &size);
        if (ret) goto error;

        if (size) {
                /*
                 * buffer is always remapped into the caller's address space
                 */
                ret = pfm_smpl_buffer_alloc(current, ctx, size, &uaddr);
                if (ret) goto error;

                /* keep track of user address of buffer */
                arg->ctx_smpl_vaddr = uaddr;
        }
        ret = pfm_buf_fmt_init(fmt, task, ctx->ctx_smpl_hdr, ctx_flags, cpu, fmt_arg);

error:
        return ret;
}

static void
pfm_reset_pmu_state(pfm_context_t *ctx)
{
        int i;

        /*
         * install reset values for PMC.
         */
        for (i=1; PMC_IS_LAST(i) == 0; i++) {
                if (PMC_IS_IMPL(i) == 0) continue;
                ctx->ctx_pmcs[i] = PMC_DFL_VAL(i);
                DPRINT(("pmc[%d]=0x%lx\n", i, ctx->ctx_pmcs[i]));
        }
        /*
         * PMD registers are set to 0UL when the context in memset()
         */

        /*
         * On context switched restore, we must restore ALL pmc and ALL pmd even
         * when they are not actively used by the task. In UP, the incoming process
         * may otherwise pick up left over PMC, PMD state from the previous process.
         * As opposed to PMD, stale PMC can cause harm to the incoming
         * process because they may change what is being measured.
         * Therefore, we must systematically reinstall the entire
         * PMC state. In SMP, the same thing is possible on the
         * same CPU but also on between 2 CPUs.
         *
         * The problem with PMD is information leaking especially
         * to user level when psr.sp=0
         *
         * There is unfortunately no easy way to avoid this problem
         * on either UP or SMP. This definitively slows down the
         * pfm_load_regs() function.
         */

         /*
          * bitmask of all PMCs accessible to this context
          *
          * PMC0 is treated differently.
          */
        ctx->ctx_all_pmcs[0] = pmu_conf->impl_pmcs[0] & ~0x1;

        /*
         * bitmask of all PMDs that are accesible to this context
         */
        ctx->ctx_all_pmds[0] = pmu_conf->impl_pmds[0];

        DPRINT(("<%d> all_pmcs=0x%lx all_pmds=0x%lx\n", ctx->ctx_fd, ctx->ctx_all_pmcs[0],ctx->ctx_all_pmds[0]));

        /*
         * useful in case of re-enable after disable
         */
        ctx->ctx_used_ibrs[0] = 0UL;
        ctx->ctx_used_dbrs[0] = 0UL;
}

static int
pfm_ctx_getsize(void *arg, size_t *sz)
{
        pfarg_context_t *req = (pfarg_context_t *)arg;
        pfm_buffer_fmt_t *fmt;

        *sz = 0;

        if (!pfm_uuid_cmp(req->ctx_smpl_buf_id, pfm_null_uuid)) return 0;

        fmt = pfm_find_buffer_fmt(req->ctx_smpl_buf_id);
        if (fmt == NULL) {
                DPRINT(("cannot find buffer format\n"));
                return -EINVAL;
        }
        /* get just enough to copy in user parameters */
        *sz = fmt->fmt_arg_size;
        DPRINT(("arg_size=%lu\n", *sz));

        return 0;
}



/*
 * cannot attach if :
 *      - kernel task
 *      - task not owned by caller
 *      - task incompatible with context mode
 */
static int
pfm_task_incompatible(pfm_context_t *ctx, struct task_struct *task)
{
        /*
         * no kernel task or task not owner by caller
         */
        if (task->mm == NULL) {
                DPRINT(("task [%d] has not memory context (kernel thread)\n", task->pid));
                return -EPERM;
        }
        if (pfm_bad_permissions(task)) {
                DPRINT(("no permission to attach to  [%d]\n", task->pid));
                return -EPERM;
        }
        /*
         * cannot block in self-monitoring mode
         */
        if (CTX_OVFL_NOBLOCK(ctx) == 0 && task == current) {
                DPRINT(("cannot load a blocking context on self for [%d]\n", task->pid));
                return -EINVAL;
        }

        if (task->exit_state == EXIT_ZOMBIE) {
                DPRINT(("cannot attach to  zombie task [%d]\n", task->pid));
                return -EBUSY;
        }

        /*
         * always ok for self
         */
        if (task == current) return 0;

        if ((task->state != TASK_STOPPED) && (task->state != TASK_TRACED)) {
                DPRINT(("cannot attach to non-stopped task [%d] state=%ld\n", task->pid, task->state));
                return -EBUSY;
        }
        /*
         * make sure the task is off any CPU
         */
        wait_task_inactive(task);

        /* more to come... */

        return 0;
}

static int
pfm_get_task(pfm_context_t *ctx, pid_t pid, struct task_struct **task)
{
        struct task_struct *p = current;
        int ret;

        /* XXX: need to add more checks here */
        if (pid < 2) return -EPERM;

        if (pid != current->pid) {

                read_lock(&tasklist_lock);

                p = find_task_by_pid(pid);

                /* make sure task cannot go away while we operate on it */
                if (p) get_task_struct(p);

                read_unlock(&tasklist_lock);

                if (p == NULL) return -ESRCH;
        }

        ret = pfm_task_incompatible(ctx, p);
        if (ret == 0) {
                *task = p;
        } else if (p != current) {
                pfm_put_task(p);
        }
        return ret;
}



static int
pfm_context_create(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
{
        pfarg_context_t *req = (pfarg_context_t *)arg;
        struct file *filp;
        int ctx_flags;
        int ret;

        /* let's check the arguments first */
        ret = pfarg_is_sane(current, req);
        if (ret < 0) return ret;

        ctx_flags = req->ctx_flags;

        ret = -ENOMEM;

        ctx = pfm_context_alloc();
        if (!ctx) goto error;

        ret = pfm_alloc_fd(&filp);
        if (ret < 0) goto error_file;

        req->ctx_fd = ctx->ctx_fd = ret;

        /*
         * attach context to file
         */
        filp->private_data = ctx;

        /*
         * does the user want to sample?
         */
        if (pfm_uuid_cmp(req->ctx_smpl_buf_id, pfm_null_uuid)) {
                ret = pfm_setup_buffer_fmt(current, ctx, ctx_flags, 0, req);
                if (ret) goto buffer_error;
        }

        /*
         * init context protection lock
         */
        spin_lock_init(&ctx->ctx_lock);

        /*
         * context is unloaded
         */
        ctx->ctx_state = PFM_CTX_UNLOADED;

        /*
         * initialization of context's flags
         */
        ctx->ctx_fl_block       = (ctx_flags & PFM_FL_NOTIFY_BLOCK) ? 1 : 0;
        ctx->ctx_fl_system      = (ctx_flags & PFM_FL_SYSTEM_WIDE) ? 1: 0;
        ctx->ctx_fl_is_sampling = ctx->ctx_buf_fmt ? 1 : 0; /* assume record() is defined */
        ctx->ctx_fl_no_msg      = (ctx_flags & PFM_FL_OVFL_NO_MSG) ? 1: 0;
        /*
         * will move to set properties
         * ctx->ctx_fl_excl_idle   = (ctx_flags & PFM_FL_EXCL_IDLE) ? 1: 0;
         */

        /*
         * init restart semaphore to locked
         */
        sema_init(&ctx->ctx_restart_sem, 0);

        /*
         * activation is used in SMP only
         */
        ctx->ctx_last_activation = PFM_INVALID_ACTIVATION;
        SET_LAST_CPU(ctx, -1);

        /*
         * initialize notification message queue
         */
        ctx->ctx_msgq_head = ctx->ctx_msgq_tail = 0;
        init_waitqueue_head(&ctx->ctx_msgq_wait);
        init_waitqueue_head(&ctx->ctx_zombieq);

        DPRINT(("ctx=%p flags=0x%x system=%d notify_block=%d excl_idle=%d no_msg=%d ctx_fd=%d \n",
                ctx,
                ctx_flags,
                ctx->ctx_fl_system,
                ctx->ctx_fl_block,
                ctx->ctx_fl_excl_idle,
                ctx->ctx_fl_no_msg,
                ctx->ctx_fd));

        /*
         * initialize soft PMU state
         */
        pfm_reset_pmu_state(ctx);

        return 0;

buffer_error:
        pfm_free_fd(ctx->ctx_fd, filp);

        if (ctx->ctx_buf_fmt) {
                pfm_buf_fmt_exit(ctx->ctx_buf_fmt, current, NULL, regs);
        }
error_file:
        pfm_context_free(ctx);

error:
        return ret;
}

static inline unsigned long
pfm_new_counter_value (pfm_counter_t *reg, int is_long_reset)
{
        unsigned long val = is_long_reset ? reg->long_reset : reg->short_reset;
        unsigned long new_seed, old_seed = reg->seed, mask = reg->mask;
        extern unsigned long carta_random32 (unsigned long seed);

        if (reg->flags & PFM_REGFL_RANDOM) {
                new_seed = carta_random32(old_seed);
                val -= (old_seed & mask);       /* counter values are negative numbers! */
                if ((mask >> 32) != 0)
                        /* construct a full 64-bit random value: */
                        new_seed |= carta_random32(old_seed >> 32) << 32;
                reg->seed = new_seed;
        }
        reg->lval = val;
        return val;
}

static void
pfm_reset_regs_masked(pfm_context_t *ctx, unsigned long *ovfl_regs, int is_long_reset)
{
        unsigned long mask = ovfl_regs[0];
        unsigned long reset_others = 0UL;
        unsigned long val;
        int i;

        /*
         * now restore reset value on sampling overflowed counters
         */
        mask >>= PMU_FIRST_COUNTER;
        for(i = PMU_FIRST_COUNTER; mask; i++, mask >>= 1) {

                if ((mask & 0x1UL) == 0UL) continue;

                ctx->ctx_pmds[i].val = val = pfm_new_counter_value(ctx->ctx_pmds+ i, is_long_reset);
                reset_others        |= ctx->ctx_pmds[i].reset_pmds[0];

                DPRINT_ovfl((" %s reset ctx_pmds[%d]=%lx\n", is_long_reset ? "long" : "short", i, val));
        }

        /*
         * Now take care of resetting the other registers
         */
        for(i = 0; reset_others; i++, reset_others >>= 1) {

                if ((reset_others & 0x1) == 0) continue;

                ctx->ctx_pmds[i].val = val = pfm_new_counter_value(ctx->ctx_pmds + i, is_long_reset);

                DPRINT_ovfl(("%s reset_others pmd[%d]=%lx\n",
                          is_long_reset ? "long" : "short", i, val));
        }
}

static void
pfm_reset_regs(pfm_context_t *ctx, unsigned long *ovfl_regs, int is_long_reset)
{
        unsigned long mask = ovfl_regs[0];
        unsigned long reset_others = 0UL;
        unsigned long val;
        int i;

        DPRINT_ovfl(("ovfl_regs=0x%lx is_long_reset=%d\n", ovfl_regs[0], is_long_reset));

        if (ctx->ctx_state == PFM_CTX_MASKED) {
                pfm_reset_regs_masked(ctx, ovfl_regs, is_long_reset);
                return;
        }

        /*
         * now restore reset value on sampling overflowed counters
         */
        mask >>= PMU_FIRST_COUNTER;
        for(i = PMU_FIRST_COUNTER; mask; i++, mask >>= 1) {

                if ((mask & 0x1UL) == 0UL) continue;

                val           = pfm_new_counter_value(ctx->ctx_pmds+ i, is_long_reset);
                reset_others |= ctx->ctx_pmds[i].reset_pmds[0];

                DPRINT_ovfl((" %s reset ctx_pmds[%d]=%lx\n", is_long_reset ? "long" : "short", i, val));

                pfm_write_soft_counter(ctx, i, val);
        }

        /*
         * Now take care of resetting the other registers
         */
        for(i = 0; reset_others; i++, reset_others >>= 1) {

                if ((reset_others & 0x1) == 0) continue;

                val = pfm_new_counter_value(ctx->ctx_pmds + i, is_long_reset);

                if (PMD_IS_COUNTING(i)) {
                        pfm_write_soft_counter(ctx, i, val);
                } else {
                        ia64_set_pmd(i, val);
                }
                DPRINT_ovfl(("%s reset_others pmd[%d]=%lx\n",
                          is_long_reset ? "long" : "short", i, val));
        }
        ia64_srlz_d();
}

static int
pfm_write_pmcs(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
{
        struct thread_struct *thread = NULL;
        struct task_struct *task;
        pfarg_reg_t *req = (pfarg_reg_t *)arg;
        unsigned long value, pmc_pm;
        unsigned long smpl_pmds, reset_pmds, impl_pmds;
        unsigned int cnum, reg_flags, flags, pmc_type;
        int i, can_access_pmu = 0, is_loaded, is_system, expert_mode;
        int is_monitor, is_counting, state;
        int ret = -EINVAL;
        pfm_reg_check_t wr_func;
#define PFM_CHECK_PMC_PM(x, y, z) ((x)->ctx_fl_system ^ PMC_PM(y, z))

        state     = ctx->ctx_state;
        is_loaded = state == PFM_CTX_LOADED ? 1 : 0;
        is_system = ctx->ctx_fl_system;
        task      = ctx->ctx_task;
        impl_pmds = pmu_conf->impl_pmds[0];

        if (state == PFM_CTX_ZOMBIE) return -EINVAL;

        if (is_loaded) {
                thread = &task->thread;
                /*
                 * In system wide and when the context is loaded, access can only happen
                 * when the caller is running on the CPU being monitored by the session.
                 * It does not have to be the owner (ctx_task) of the context per se.
                 */
                if (is_system && ctx->ctx_cpu != smp_processor_id()) {
                        DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
                        return -EBUSY;
                }
                can_access_pmu = GET_PMU_OWNER() == task || is_system ? 1 : 0;
        }
        expert_mode = pfm_sysctl.expert_mode; 

        for (i = 0; i < count; i++, req++) {

                cnum       = req->reg_num;
                reg_flags  = req->reg_flags;
                value      = req->reg_value;
                smpl_pmds  = req->reg_smpl_pmds[0];
                reset_pmds = req->reg_reset_pmds[0];
                flags      = 0;


                if (cnum >= PMU_MAX_PMCS) {
                        DPRINT(("pmc%u is invalid\n", cnum));
                        goto error;
                }

                pmc_type   = pmu_conf->pmc_desc[cnum].type;
                pmc_pm     = (value >> pmu_conf->pmc_desc[cnum].pm_pos) & 0x1;
                is_counting = (pmc_type & PFM_REG_COUNTING) == PFM_REG_COUNTING ? 1 : 0;
                is_monitor  = (pmc_type & PFM_REG_MONITOR) == PFM_REG_MONITOR ? 1 : 0;

                /*
                 * we reject all non implemented PMC as well
                 * as attempts to modify PMC[0-3] which are used
                 * as status registers by the PMU
                 */
                if ((pmc_type & PFM_REG_IMPL) == 0 || (pmc_type & PFM_REG_CONTROL) == PFM_REG_CONTROL) {
                        DPRINT(("pmc%u is unimplemented or no-access pmc_type=%x\n", cnum, pmc_type));
                        goto error;
                }
                wr_func = pmu_conf->pmc_desc[cnum].write_check;
                /*
                 * If the PMC is a monitor, then if the value is not the default:
                 *      - system-wide session: PMCx.pm=1 (privileged monitor)
                 *      - per-task           : PMCx.pm=0 (user monitor)
                 */
                if (is_monitor && value != PMC_DFL_VAL(cnum) && is_system ^ pmc_pm) {
                        DPRINT(("pmc%u pmc_pm=%lu is_system=%d\n",
                                cnum,
                                pmc_pm,
                                is_system));
                        goto error;
                }

                if (is_counting) {
                        /*
                         * enforce generation of overflow interrupt. Necessary on all
                         * CPUs.
                         */
                        value |= 1 << PMU_PMC_OI;

                        if (reg_flags & PFM_REGFL_OVFL_NOTIFY) {
                                flags |= PFM_REGFL_OVFL_NOTIFY;
                        }

                        if (reg_flags & PFM_REGFL_RANDOM) flags |= PFM_REGFL_RANDOM;

                        /* verify validity of smpl_pmds */
                        if ((smpl_pmds & impl_pmds) != smpl_pmds) {
                                DPRINT(("invalid smpl_pmds 0x%lx for pmc%u\n", smpl_pmds, cnum));
                                goto error;
                        }

                        /* verify validity of reset_pmds */
                        if ((reset_pmds & impl_pmds) != reset_pmds) {
                                DPRINT(("invalid reset_pmds 0x%lx for pmc%u\n", reset_pmds, cnum));
                                goto error;
                        }
                } else {
                        if (reg_flags & (PFM_REGFL_OVFL_NOTIFY|PFM_REGFL_RANDOM)) {
                                DPRINT(("cannot set ovfl_notify or random on pmc%u\n", cnum));
                                goto error;
                        }
                        /* eventid on non-counting monitors are ignored */
                }

                /*
                 * execute write checker, if any
                 */
                if (likely(expert_mode == 0 && wr_func)) {
                        ret = (*wr_func)(task, ctx, cnum, &value, regs);
                        if (ret) goto error;
                        ret = -EINVAL;
                }

                /*
                 * no error on this register
                 */
                PFM_REG_RETFLAG_SET(req->reg_flags, 0);

                /*
                 * Now we commit the changes to the software state
                 */

                /*
                 * update overflow information
                 */
                if (is_counting) {
                        /*
                         * full flag update each time a register is programmed
                         */
                        ctx->ctx_pmds[cnum].flags = flags;

                        ctx->ctx_pmds[cnum].reset_pmds[0] = reset_pmds;
                        ctx->ctx_pmds[cnum].smpl_pmds[0]  = smpl_pmds;
                        ctx->ctx_pmds[cnum].eventid       = req->reg_smpl_eventid;

                        /*
                         * Mark all PMDS to be accessed as used.
                         *
                         * We do not keep track of PMC because we have to
                         * systematically restore ALL of them.
                         *
                         * We do not update the used_monitors mask, because
                         * if we have not programmed them, then will be in
                         * a quiescent state, therefore we will not need to
                         * mask/restore then when context is MASKED.
                         */
                        CTX_USED_PMD(ctx, reset_pmds);
                        CTX_USED_PMD(ctx, smpl_pmds);
                        /*

                         * make sure we do not try to reset on
                         * restart because we have established new values
                         */
                        if (state == PFM_CTX_MASKED) ctx->ctx_ovfl_regs[0] &= ~1UL << cnum;
                }
                /*
                 * Needed in case the user does not initialize the equivalent
                 * PMD. Clearing is done indirectly via pfm_reset_pmu_state() so there is no
                 * possible leak here.
                 */
                CTX_USED_PMD(ctx, pmu_conf->pmc_desc[cnum].dep_pmd[0]);

                /*
                 * keep track of the monitor PMC that we are using.
                 * we save the value of the pmc in ctx_pmcs[] and if
                 * the monitoring is not stopped for the context we also
                 * place it in the saved state area so that it will be
                 * picked up later by the context switch code.
                 *
                 * The value in ctx_pmcs[] can only be changed in pfm_write_pmcs().
                 *
                 * The value in thread->pmcs[] may be modified on overflow, i.e.,  when
                 * monitoring needs to be stopped.
                 */
                if (is_monitor) CTX_USED_MONITOR(ctx, 1UL << cnum);

                /*
                 * update context state
                 */
                ctx->ctx_pmcs[cnum] = value;

                if (is_loaded) {
                        /*
                         * write thread state
                         */
                        if (is_system == 0) thread->pmcs[cnum] = value;

                        /*
                         * write hardware register if we can
                         */
                        if (can_access_pmu) {
                                ia64_set_pmc(cnum, value);
                        }
#ifdef CONFIG_SMP
                        else {
                                /*
                                 * per-task SMP only here
                                 *
                                 * we are guaranteed that the task is not running on the other CPU,
                                 * we indicate that this PMD will need to be reloaded if the task
                                 * is rescheduled on the CPU it ran last on.
                                 */
                                ctx->ctx_reload_pmcs[0] |= 1UL << cnum;
                        }
#endif
                }

                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",
                          cnum,
                          value,
                          is_loaded,
                          can_access_pmu,
                          flags,
                          ctx->ctx_all_pmcs[0],
                          ctx->ctx_used_pmds[0],
                          ctx->ctx_pmds[cnum].eventid,
                          smpl_pmds,
                          reset_pmds,
                          ctx->ctx_reload_pmcs[0],
                          ctx->ctx_used_monitors[0],
                          ctx->ctx_ovfl_regs[0]));
        }

        /*
         * make sure the changes are visible
         */
        if (can_access_pmu) ia64_srlz_d();

        return 0;
error:
        PFM_REG_RETFLAG_SET(req->reg_flags, PFM_REG_RETFL_EINVAL);
        return ret;
}

static int
pfm_write_pmds(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
{
        struct thread_struct *thread = NULL;
        struct task_struct *task;
        pfarg_reg_t *req = (pfarg_reg_t *)arg;
        unsigned long value, hw_value, ovfl_mask;
        unsigned int cnum;
        int i, can_access_pmu = 0, state;
        int is_counting, is_loaded, is_system, expert_mode;
        int ret = -EINVAL;
        pfm_reg_check_t wr_func;


        state     = ctx->ctx_state;
        is_loaded = state == PFM_CTX_LOADED ? 1 : 0;
        is_system = ctx->ctx_fl_system;
        ovfl_mask = pmu_conf->ovfl_val;
        task      = ctx->ctx_task;

        if (unlikely(state == PFM_CTX_ZOMBIE)) return -EINVAL;

        /*
         * on both UP and SMP, we can only write to the PMC when the task is
         * the owner of the local PMU.
         */
        if (likely(is_loaded)) {
                thread = &task->thread;
                /*
                 * In system wide and when the context is loaded, access can only happen
                 * when the caller is running on the CPU being monitored by the session.
                 * It does not have to be the owner (ctx_task) of the context per se.
                 */
                if (unlikely(is_system && ctx->ctx_cpu != smp_processor_id())) {
                        DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
                        return -EBUSY;
                }
                can_access_pmu = GET_PMU_OWNER() == task || is_system ? 1 : 0;
        }
        expert_mode = pfm_sysctl.expert_mode; 

        for (i = 0; i < count; i++, req++) {

                cnum  = req->reg_num;
                value = req->reg_value;

                if (!PMD_IS_IMPL(cnum)) {
                        DPRINT(("pmd[%u] is unimplemented or invalid\n", cnum));
                        goto abort_mission;
                }
                is_counting = PMD_IS_COUNTING(cnum);
                wr_func     = pmu_conf->pmd_desc[cnum].write_check;

                /*
                 * execute write checker, if any
                 */
                if (unlikely(expert_mode == 0 && wr_func)) {
                        unsigned long v = value;

                        ret = (*wr_func)(task, ctx, cnum, &v, regs);
                        if (ret) goto abort_mission;

                        value = v;
                        ret   = -EINVAL;
                }

                /*
                 * no error on this register
                 */
                PFM_REG_RETFLAG_SET(req->reg_flags, 0);

                /*
                 * now commit changes to software state
                 */
                hw_value = value;

                /*
                 * update virtualized (64bits) counter
                 */
                if (is_counting) {
                        /*
                         * write context state
                         */
                        ctx->ctx_pmds[cnum].lval = value;

                        /*
                         * when context is load we use the split value
                         */
                        if (is_loaded) {
                                hw_value = value &  ovfl_mask;
                                value    = value & ~ovfl_mask;
                        }
                }
                /*
                 * update reset values (not just for counters)
                 */
                ctx->ctx_pmds[cnum].long_reset  = req->reg_long_reset;
                ctx->ctx_pmds[cnum].short_reset = req->reg_short_reset;

                /*
                 * update randomization parameters (not just for counters)
                 */
                ctx->ctx_pmds[cnum].seed = req->reg_random_seed;
                ctx->ctx_pmds[cnum].mask = req->reg_random_mask;

                /*
                 * update context value
                 */
                ctx->ctx_pmds[cnum].val  = value;

                /*
                 * Keep track of what we use
                 *
                 * We do not keep track of PMC because we have to
                 * systematically restore ALL of them.
                 */
                CTX_USED_PMD(ctx, PMD_PMD_DEP(cnum));

                /*
                 * mark this PMD register used as well
                 */
                CTX_USED_PMD(ctx, RDEP(cnum));

                /*
                 * make sure we do not try to reset on
                 * restart because we have established new values
                 */
                if (is_counting && state == PFM_CTX_MASKED) {
                        ctx->ctx_ovfl_regs[0] &= ~1UL << cnum;
                }

                if (is_loaded) {
                        /*
                         * write thread state
                         */
                        if (is_system == 0) thread->pmds[cnum] = hw_value;

                        /*
                         * write hardware register if we can
                         */
                        if (can_access_pmu) {
                                ia64_set_pmd(cnum, hw_value);
                        } else {
#ifdef CONFIG_SMP
                                /*
                                 * we are guaranteed that the task is not running on the other CPU,
                                 * we indicate that this PMD will need to be reloaded if the task
                                 * is rescheduled on the CPU it ran last on.
                                 */
                                ctx->ctx_reload_pmds[0] |= 1UL << cnum;
#endif
                        }
                }

                DPRINT(("pmd[%u]=0x%lx ld=%d apmu=%d, hw_value=0x%lx ctx_pmd=0x%lx  short_reset=0x%lx "
                          "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",
                        cnum,
                        value,
                        is_loaded,
                        can_access_pmu,
                        hw_value,
                        ctx->ctx_pmds[cnum].val,
                        ctx->ctx_pmds[cnum].short_reset,
                        ctx->ctx_pmds[cnum].long_reset,
                        PMC_OVFL_NOTIFY(ctx, cnum) ? 'Y':'N',
                        ctx->ctx_pmds[cnum].seed,
                        ctx->ctx_pmds[cnum].mask,
                        ctx->ctx_used_pmds[0],
                        ctx->ctx_pmds[cnum].reset_pmds[0],
                        ctx->ctx_reload_pmds[0],
                        ctx->ctx_all_pmds[0],
                        ctx->ctx_ovfl_regs[0]));
        }

        /*
         * make changes visible
         */
        if (can_access_pmu) ia64_srlz_d();

        return 0;

abort_mission:
        /*
         * for now, we have only one possibility for error
         */
        PFM_REG_RETFLAG_SET(req->reg_flags, PFM_REG_RETFL_EINVAL);
        return ret;
}

/*
 * By the way of PROTECT_CONTEXT(), interrupts are masked while we are in this function.
 * Therefore we know, we do not have to worry about the PMU overflow interrupt. If an
 * interrupt is delivered during the call, it will be kept pending until we leave, making
 * it appears as if it had been generated at the UNPROTECT_CONTEXT(). At least we are
 * guaranteed to return consistent data to the user, it may simply be old. It is not
 * trivial to treat the overflow while inside the call because you may end up in
 * some module sampling buffer code causing deadlocks.
 */
static int
pfm_read_pmds(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
{
        struct thread_struct *thread = NULL;
        struct task_struct *task;
        unsigned long val = 0UL, lval, ovfl_mask, sval;
        pfarg_reg_t *req = (pfarg_reg_t *)arg;
        unsigned int cnum, reg_flags = 0;
        int i, can_access_pmu = 0, state;
        int is_loaded, is_system, is_counting, expert_mode;
        int ret = -EINVAL;
        pfm_reg_check_t rd_func;

        /*
         * access is possible when loaded only for
         * self-monitoring tasks or in UP mode
         */

        state     = ctx->ctx_state;
        is_loaded = state == PFM_CTX_LOADED ? 1 : 0;
        is_system = ctx->ctx_fl_system;
        ovfl_mask = pmu_conf->ovfl_val;
        task      = ctx->ctx_task;

        if (state == PFM_CTX_ZOMBIE) return -EINVAL;

        if (likely(is_loaded)) {
                thread = &task->thread;
                /*
                 * In system wide and when the context is loaded, access can only happen
                 * when the caller is running on the CPU being monitored by the session.
                 * It does not have to be the owner (ctx_task) of the context per se.
                 */
                if (unlikely(is_system && ctx->ctx_cpu != smp_processor_id())) {
                        DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
                        return -EBUSY;
                }
                /*
                 * this can be true when not self-monitoring only in UP
                 */
                can_access_pmu = GET_PMU_OWNER() == task || is_system ? 1 : 0;

                if (can_access_pmu) ia64_srlz_d();
        }
        expert_mode = pfm_sysctl.expert_mode; 

        DPRINT(("ld=%d apmu=%d ctx_state=%d\n",
                is_loaded,
                can_access_pmu,
                state));

        /*
         * on both UP and SMP, we can only read the PMD from the hardware register when
         * the task is the owner of the local PMU.
         */

        for (i = 0; i < count; i++, req++) {

                cnum        = req->reg_num;
                reg_flags   = req->reg_flags;

                if (unlikely(!PMD_IS_IMPL(cnum))) goto error;
                /*
                 * we can only read the register that we use. That includes
                 * the one we explicitely initialize AND the one we want included
                 * in the sampling buffer (smpl_regs).
                 *
                 * Having this restriction allows optimization in the ctxsw routine
                 * without compromising security (leaks)
                 */
                if (unlikely(!CTX_IS_USED_PMD(ctx, cnum))) goto error;

                sval        = ctx->ctx_pmds[cnum].val;
                lval        = ctx->ctx_pmds[cnum].lval;
                is_counting = PMD_IS_COUNTING(cnum);

                /*
                 * If the task is not the current one, then we check if the
                 * PMU state is still in the local live register due to lazy ctxsw.
                 * If true, then we read directly from the registers.
                 */
                if (can_access_pmu){
                        val = ia64_get_pmd(cnum);
                } else {
                        /*
                         * context has been saved
                         * if context is zombie, then task does not exist anymore.
                         * In this case, we use the full value saved in the context (pfm_flush_regs()).
                         */
                        val = is_loaded ? thread->pmds[cnum] : 0UL;
                }
                rd_func = pmu_conf->pmd_desc[cnum].read_check;

                if (is_counting) {
                        /*
                         * XXX: need to check for overflow when loaded
                         */
                        val &= ovfl_mask;
                        val += sval;
                }

                /*
                 * execute read checker, if any
                 */
                if (unlikely(expert_mode == 0 && rd_func)) {
                        unsigned long v = val;
                        ret = (*rd_func)(ctx->ctx_task, ctx, cnum, &v, regs);
                        if (ret) goto error;
                        val = v;
                        ret = -EINVAL;
                }

                PFM_REG_RETFLAG_SET(reg_flags, 0);

                DPRINT(("pmd[%u]=0x%lx\n", cnum, val));

                /*
                 * update register return value, abort all if problem during copy.
                 * we only modify the reg_flags field. no check mode is fine because
                 * access has been verified upfront in sys_perfmonctl().
                 */
                req->reg_value            = val;
                req->reg_flags            = reg_flags;
                req->reg_last_reset_val   = lval;
        }

        return 0;

error:
        PFM_REG_RETFLAG_SET(req->reg_flags, PFM_REG_RETFL_EINVAL);
        return ret;
}

int
pfm_mod_write_pmcs(struct task_struct *task, void *req, unsigned int nreq, struct pt_regs *regs)
{
        pfm_context_t *ctx;

        if (req == NULL) return -EINVAL;

        ctx = GET_PMU_CTX();

        if (ctx == NULL) return -EINVAL;

        /*
         * for now limit to current task, which is enough when calling
         * from overflow handler
         */
        if (task != current && ctx->ctx_fl_system == 0) return -EBUSY;

        return pfm_write_pmcs(ctx, req, nreq, regs);
}
EXPORT_SYMBOL(pfm_mod_write_pmcs);

int
pfm_mod_read_pmds(struct task_struct *task, void *req, unsigned int nreq, struct pt_regs *regs)
{
        pfm_context_t *ctx;

        if (req == NULL) return -EINVAL;

        ctx = GET_PMU_CTX();

        if (ctx == NULL) return -EINVAL;

        /*
         * for now limit to current task, which is enough when calling
         * from overflow handler
         */
        if (task != current && ctx->ctx_fl_system == 0) return -EBUSY;

        return pfm_read_pmds(ctx, req, nreq, regs);
}
EXPORT_SYMBOL(pfm_mod_read_pmds);

/*
 * Only call this function when a process it trying to
 * write the debug registers (reading is always allowed)
 */
int
pfm_use_debug_registers(struct task_struct *task)
{
        pfm_context_t *ctx = task->thread.pfm_context;
        unsigned long flags;
        int ret = 0;

        if (pmu_conf->use_rr_dbregs == 0) return 0;

        DPRINT(("called for [%d]\n", task->pid));

        /*
         * do it only once
         */
        if (task->thread.flags & IA64_THREAD_DBG_VALID) return 0;

        /*
         * Even on SMP, we do not need to use an atomic here because
         * the only way in is via ptrace() and this is possible only when the
         * process is stopped. Even in the case where the ctxsw out is not totally
         * completed by the time we come here, there is no way the 'stopped' process
         * could be in the middle of fiddling with the pfm_write_ibr_dbr() routine.
         * So this is always safe.
         */
        if (ctx && ctx->ctx_fl_using_dbreg == 1) return -1;

        LOCK_PFS(flags);

        /*
         * We cannot allow setting breakpoints when system wide monitoring
         * sessions are using the debug registers.
         */
        if (pfm_sessions.pfs_sys_use_dbregs> 0)
                ret = -1;
        else
                pfm_sessions.pfs_ptrace_use_dbregs++;

        DPRINT(("ptrace_use_dbregs=%u  sys_use_dbregs=%u by [%d] ret = %d\n",
                  pfm_sessions.pfs_ptrace_use_dbregs,
                  pfm_sessions.pfs_sys_use_dbregs,
                  task->pid, ret));

        UNLOCK_PFS(flags);

        return ret;
}

/*
 * This function is called for every task that exits with the
 * IA64_THREAD_DBG_VALID set. This indicates a task which was
 * able to use the debug registers for debugging purposes via
 * ptrace(). Therefore we know it was not using them for
 * perfmormance monitoring, so we only decrement the number
 * of "ptraced" debug register users to keep the count up to date
 */
int
pfm_release_debug_registers(struct task_struct *task)
{
        unsigned long flags;
        int ret;

        if (pmu_conf->use_rr_dbregs == 0) return 0;

        LOCK_PFS(flags);
        if (pfm_sessions.pfs_ptrace_use_dbregs == 0) {
                printk(KERN_ERR "perfmon: invalid release for [%d] ptrace_use_dbregs=0\n", task->pid);
                ret = -1;
        }  else {
                pfm_sessions.pfs_ptrace_use_dbregs--;
                ret = 0;
        }
        UNLOCK_PFS(flags);

        return ret;
}

static int
pfm_restart(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
{
        struct task_struct *task;
        pfm_buffer_fmt_t *fmt;
        pfm_ovfl_ctrl_t rst_ctrl;
        int state, is_system;
        int ret = 0;

        state     = ctx->ctx_state;
        fmt       = ctx->ctx_buf_fmt;
        is_system = ctx->ctx_fl_system;
        task      = PFM_CTX_TASK(ctx);

        switch(state) {
                case PFM_CTX_MASKED:
                        break;
                case PFM_CTX_LOADED: 
                        if (CTX_HAS_SMPL(ctx) && fmt->fmt_restart_active) break;
                        /* fall through */
                case PFM_CTX_UNLOADED:
                case PFM_CTX_ZOMBIE:
                        DPRINT(("invalid state=%d\n", state));
                        return -EBUSY;
                default:
                        DPRINT(("state=%d, cannot operate (no active_restart handler)\n", state));
                        return -EINVAL;
        }

        /*
         * In system wide and when the context is loaded, access can only happen
         * when the caller is running on the CPU being monitored by the session.
         * It does not have to be the owner (ctx_task) of the context per se.
         */
        if (is_system && ctx->ctx_cpu != smp_processor_id()) {
                DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
                return -EBUSY;
        }

        /* sanity check */
        if (unlikely(task == NULL)) {
                printk(KERN_ERR "perfmon: [%d] pfm_restart no task\n", current->pid);
                return -EINVAL;
        }

        if (task == current || is_system) {

                fmt = ctx->ctx_buf_fmt;

                DPRINT(("restarting self %d ovfl=0x%lx\n",
                        task->pid,
                        ctx->ctx_ovfl_regs[0]));

                if (CTX_HAS_SMPL(ctx)) {

                        prefetch(ctx->ctx_smpl_hdr);

                        rst_ctrl.bits.mask_monitoring = 0;
                        rst_ctrl.bits.reset_ovfl_pmds = 0;

                        if (state == PFM_CTX_LOADED)
                                ret = pfm_buf_fmt_restart_active(fmt, task, &rst_ctrl, ctx->ctx_smpl_hdr, regs);
                        else
                                ret = pfm_buf_fmt_restart(fmt, task, &rst_ctrl, ctx->ctx_smpl_hdr, regs);
                } else {
                        rst_ctrl.bits.mask_monitoring = 0;
                        rst_ctrl.bits.reset_ovfl_pmds = 1;
                }

                if (ret == 0) {
                        if (rst_ctrl.bits.reset_ovfl_pmds)
                                pfm_reset_regs(ctx, ctx->ctx_ovfl_regs, PFM_PMD_LONG_RESET);

                        if (rst_ctrl.bits.mask_monitoring == 0) {
                                DPRINT(("resuming monitoring for [%d]\n", task->pid));

                                if (state == PFM_CTX_MASKED) pfm_restore_monitoring(task);
                        } else {
                                DPRINT(("keeping monitoring stopped for [%d]\n", task->pid));

                                // cannot use pfm_stop_monitoring(task, regs);
                        }
                }
                /*
                 * clear overflowed PMD mask to remove any stale information
                 */
                ctx->ctx_ovfl_regs[0] = 0UL;

                /*
                 * back to LOADED state
                 */
                ctx->ctx_state = PFM_CTX_LOADED;

                /*
                 * XXX: not really useful for self monitoring
                 */
                ctx->ctx_fl_can_restart = 0;

                return 0;
        }

        /* 
         * restart another task
         */

        /*
         * When PFM_CTX_MASKED, we cannot issue a restart before the previous 
         * one is seen by the task.
         */
        if (state == PFM_CTX_MASKED) {
                if (ctx->ctx_fl_can_restart == 0) return -EINVAL;
                /*
                 * will prevent subsequent restart before this one is
                 * seen by other task
                 */
                ctx->ctx_fl_can_restart = 0;
        }

        /*
         * if blocking, then post the semaphore is PFM_CTX_MASKED, i.e.
         * the task is blocked or on its way to block. That's the normal
         * restart path. If the monitoring is not masked, then the task
         * can be actively monitoring and we cannot directly intervene.
         * Therefore we use the trap mechanism to catch the task and
         * force it to reset the buffer/reset PMDs.
         *
         * if non-blocking, then we ensure that the task will go into
         * pfm_handle_work() before returning to user mode.
         *
         * We cannot explicitely reset another task, it MUST always
         * be done by the task itself. This works for system wide because
         * the tool that is controlling the session is logically doing 
         * "self-monitoring".
         */
        if (CTX_OVFL_NOBLOCK(ctx) == 0 && state == PFM_CTX_MASKED) {
                DPRINT(("unblocking [%d] \n", task->pid));
                up(&ctx->ctx_restart_sem);
        } else {
                DPRINT(("[%d] armed exit trap\n", task->pid));

                ctx->ctx_fl_trap_reason = PFM_TRAP_REASON_RESET;

                PFM_SET_WORK_PENDING(task, 1);

                pfm_set_task_notify(task);

                /*
                 * XXX: send reschedule if task runs on another CPU
                 */
        }
        return 0;
}

static int
pfm_debug(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
{
        unsigned int m = *(unsigned int *)arg;

        pfm_sysctl.debug = m == 0 ? 0 : 1;

        pfm_debug_var = pfm_sysctl.debug;

        printk(KERN_INFO "perfmon debugging %s (timing reset)\n", pfm_sysctl.debug ? "on" : "off");

        if (m == 0) {
                memset(pfm_stats, 0, sizeof(pfm_stats));
                for(m=0; m < NR_CPUS; m++) pfm_stats[m].pfm_ovfl_intr_cycles_min = ~0UL;
        }
        return 0;
}

/*
 * arg can be NULL and count can be zero for this function
 */
static int
pfm_write_ibr_dbr(int mode, pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
{
        struct thread_struct *thread = NULL;
        struct task_struct *task;
        pfarg_dbreg_t *req = (pfarg_dbreg_t *)arg;
        unsigned long flags;
        dbreg_t dbreg;
        unsigned int rnum;
        int first_time;
        int ret = 0, state;
        int i, can_access_pmu = 0;
        int is_system, is_loaded;

        if (pmu_conf->use_rr_dbregs == 0) return -EINVAL;

        state     = ctx->ctx_state;
        is_loaded = state == PFM_CTX_LOADED ? 1 : 0;
        is_system = ctx->ctx_fl_system;
        task      = ctx->ctx_task;

        if (state == PFM_CTX_ZOMBIE) return -EINVAL;

        /*
         * on both UP and SMP, we can only write to the PMC when the task is
         * the owner of the local PMU.
         */
        if (is_loaded) {
                thread = &task->thread;
                /*
                 * In system wide and when the context is loaded, access can only happen
                 * when the caller is running on the CPU being monitored by the session.
                 * It does not have to be the owner (ctx_task) of the context per se.
                 */
                if (unlikely(is_system && ctx->ctx_cpu != smp_processor_id())) {
                        DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
                        return -EBUSY;
                }
                can_access_pmu = GET_PMU_OWNER() == task || is_system ? 1 : 0;
        }

        /*
         * we do not need to check for ipsr.db because we do clear ibr.x, dbr.r, and dbr.w
         * ensuring that no real breakpoint can be installed via this call.
         *
         * IMPORTANT: regs can be NULL in this function
         */

        first_time = ctx->ctx_fl_using_dbreg == 0;

        /*
         * don't bother if we are loaded and task is being debugged
         */
        if (is_loaded && (thread->flags & IA64_THREAD_DBG_VALID) != 0) {
                DPRINT(("debug registers already in use for [%d]\n", task->pid));
                return -EBUSY;
        }

        /*
         * check for debug registers in system wide mode
         *
         * If though a check is done in pfm_context_load(),
         * we must repeat it here, in case the registers are
         * written after the context is loaded
         */
        if (is_loaded) {
                LOCK_PFS(flags);

                if (first_time && is_system) {
                        if (pfm_sessions.pfs_ptrace_use_dbregs)
                                ret = -EBUSY;
                        else
                                pfm_sessions.pfs_sys_use_dbregs++;
                }
                UNLOCK_PFS(flags);
        }

        if (ret != 0) return ret;

        /*
         * mark ourself as user of the debug registers for
         * perfmon purposes.
         */
        ctx->ctx_fl_using_dbreg = 1;

        /*
         * clear hardware registers to make sure we don't
         * pick up stale state.
         *
         * for a system wide session, we do not use
         * thread.dbr, thread.ibr because this process
         * never leaves the current CPU and the state
         * is shared by all processes running on it
         */
        if (first_time && can_access_pmu) {
                DPRINT(("[%d] clearing ibrs, dbrs\n", task->pid));
                for (i=0; i < pmu_conf->num_ibrs; i++) {
                        ia64_set_ibr(i, 0UL);
                        ia64_dv_serialize_instruction();
                }
                ia64_srlz_i();
                for (i=0; i < pmu_conf->num_dbrs; i++) {
                        ia64_set_dbr(i, 0UL);
                        ia64_dv_serialize_data();
                }
                ia64_srlz_d();
        }

        /*
         * Now install the values into the registers
         */
        for (i = 0; i < count; i++, req++) {

                rnum      = req->dbreg_num;
                dbreg.val = req->dbreg_value;

                ret = -EINVAL;

                if ((mode == PFM_CODE_RR && rnum >= PFM_NUM_IBRS) || ((mode == PFM_DATA_RR) && rnum >= PFM_NUM_DBRS)) {
                        DPRINT(("invalid register %u val=0x%lx mode=%d i=%d count=%d\n",
                                  rnum, dbreg.val, mode, i, count));

                        goto abort_mission;
                }

                /*
                 * make sure we do not install enabled breakpoint
                 */
                if (rnum & 0x1) {
                        if (mode == PFM_CODE_RR)
                                dbreg.ibr.ibr_x = 0;
                        else
                                dbreg.dbr.dbr_r = dbreg.dbr.dbr_w = 0;
                }

                PFM_REG_RETFLAG_SET(req->dbreg_flags, 0);

                /*
                 * Debug registers, just like PMC, can only be modified
                 * by a kernel call. Moreover, perfmon() access to those
                 * registers are centralized in this routine. The hardware
                 * does not modify the value of these registers, therefore,
                 * if we save them as they are written, we can avoid having
                 * to save them on context switch out. This is made possible
                 * by the fact that when perfmon uses debug registers, ptrace()
                 * won't be able to modify them concurrently.
                 */
                if (mode == PFM_CODE_RR) {
                        CTX_USED_IBR(ctx, rnum);

                        if (can_access_pmu) {
                                ia64_set_ibr(rnum, dbreg.val);
                                ia64_dv_serialize_instruction();
                        }

                        ctx->ctx_ibrs[rnum] = dbreg.val;

                        DPRINT(("write ibr%u=0x%lx used_ibrs=0x%x ld=%d apmu=%d\n",
                                rnum, dbreg.val, ctx->ctx_used_ibrs[0], is_loaded, can_access_pmu));
                } else {
                        CTX_USED_DBR(ctx, rnum);

                        if (can_access_pmu) {
                                ia64_set_dbr(rnum, dbreg.val);
                                ia64_dv_serialize_data();
                        }
                        ctx->ctx_dbrs[rnum] = dbreg.val;

                        DPRINT(("write dbr%u=0x%lx used_dbrs=0x%x ld=%d apmu=%d\n",
                                rnum, dbreg.val, ctx->ctx_used_dbrs[0], is_loaded, can_access_pmu));
                }
        }

        return 0;

abort_mission:
        /*
         * in case it was our first attempt, we undo the global modifications
         */
        if (first_time) {
                LOCK_PFS(flags);
                if (ctx->ctx_fl_system) {
                        pfm_sessions.pfs_sys_use_dbregs--;
                }
                UNLOCK_PFS(flags);
                ctx->ctx_fl_using_dbreg = 0;
        }
        /*
         * install error return flag
         */
        PFM_REG_RETFLAG_SET(req->dbreg_flags, PFM_REG_RETFL_EINVAL);

        return ret;
}

static int
pfm_write_ibrs(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
{
        return pfm_write_ibr_dbr(PFM_CODE_RR, ctx, arg, count, regs);
}

static int
pfm_write_dbrs(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
{
        return pfm_write_ibr_dbr(PFM_DATA_RR, ctx, arg, count, regs);
}

int
pfm_mod_write_ibrs(struct task_struct *task, void *req, unsigned int nreq, struct pt_regs *regs)
{
        pfm_context_t *ctx;

        if (req == NULL) return -EINVAL;

        ctx = GET_PMU_CTX();

        if (ctx == NULL) return -EINVAL;

        /*
         * for now limit to current task, which is enough when calling
         * from overflow handler
         */
        if (task != current && ctx->ctx_fl_system == 0) return -EBUSY;

        return pfm_write_ibrs(ctx, req, nreq, regs);
}
EXPORT_SYMBOL(pfm_mod_write_ibrs);

int
pfm_mod_write_dbrs(struct task_struct *task, void *req, unsigned int nreq, struct pt_regs *regs)
{
        pfm_context_t *ctx;

        if (req == NULL) return -EINVAL;

        ctx = GET_PMU_CTX();

        if (ctx == NULL) return -EINVAL;

        /*
         * for now limit to current task, which is enough when calling
         * from overflow handler
         */
        if (task != current && ctx->ctx_fl_system == 0) return -EBUSY;

        return pfm_write_dbrs(ctx, req, nreq, regs);
}
EXPORT_SYMBOL(pfm_mod_write_dbrs);


static int
pfm_get_features(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
{
        pfarg_features_t *req = (pfarg_features_t *)arg;

        req->ft_version = PFM_VERSION;
        return 0;
}

static int
pfm_stop(pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs)
{
        struct pt_regs *tregs;
        struct task_struct *task = PFM_CTX_TASK(ctx);
        int state, is_system;

        state     = ctx->ctx_state;
        is_system = ctx->ctx_fl_system;

        /*
         * context must be attached to issue the stop command (includes LOADED,MASKED,ZOMBIE)
         */
        if (state == PFM_CTX_UNLOADED) return -EINVAL;

        /*
         * In system wide and when the context is loaded, access can only happen
         * when the caller is running on the CPU being monitored by the session.
         * It does not have to be the owner (ctx_task) of the context per se.
         */
        if (is_system && ctx->ctx_cpu != smp_processor_id()) {
                DPRINT(("should be running on CPU%d\n", ctx->ctx_cpu));
                return -EBUSY;
        }
        DPRINT(("task [%d] ctx_state=%d is_system=%d\n",
                PFM_CTX_TASK(ctx)->pid,
                state,
                is_system));
        /*
         * in system mode, we need to update the PMU directly