/*
 * 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 ne…