* 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
+ * Copyright (C) 1999-2005 Hewlett Packard Co
* Stephane Eranian <eranian@hpl.hp.com>
* David Mosberger-Tang <davidm@hpl.hp.com>
*
* 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/file.h>
#include <linux/poll.h>
#include <linux/vfs.h>
+#include <linux/smp.h>
#include <linux/pagemap.h>
#include <linux/mount.h>
-#include <linux/version.h>
#include <linux/bitops.h>
+#include <linux/capability.h>
+#include <linux/rcupdate.h>
+#include <linux/completion.h>
#include <linux/vs_memory.h>
-#include <linux/vs_cvirt.h>
#include <asm/errno.h>
#include <asm/intrinsics.h>
#define PFM_INVALID_ACTIVATION (~0UL)
+#define PFM_NUM_PMC_REGS 64 /* PMC save area for ctxsw */
+#define PFM_NUM_PMD_REGS 64 /* PMD save area for ctxsw */
+
/*
* depth of message queue
*/
unsigned long ctx_ovfl_regs[4]; /* which registers overflowed (notification) */
- struct semaphore ctx_restart_sem; /* use for blocking notification mode */
+ struct completion ctx_restart_done; /* 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_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 long ctx_pmcs[PFM_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 */
+ pfm_counter_t ctx_pmds[PFM_NUM_PMD_REGS]; /* software state for PMDS */
+
+ unsigned long th_pmcs[PFM_NUM_PMC_REGS]; /* PMC thread save state */
+ unsigned long th_pmds[PFM_NUM_PMD_REGS]; /* PMD thread save state */
u64 ctx_saved_psr_up; /* only contains psr.up value */
#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 */
static pfm_stats_t pfm_stats[NR_CPUS];
static pfm_session_t pfm_sessions; /* global sessions information */
+static DEFINE_SPINLOCK(pfm_alt_install_check);
+static pfm_intr_handler_desc_t *pfm_alt_intr_handler;
+
static struct proc_dir_entry *perfmon_dir;
static pfm_uuid_t pfm_null_uuid = {0,};
static pmu_config_t *pmu_conf;
/* sysctl() controls */
-static pfm_sysctl_t pfm_sysctl;
-int pfm_debug_var;
+pfm_sysctl_t pfm_sysctl;
+EXPORT_SYMBOL(pfm_sysctl);
static ctl_table pfm_ctl_table[]={
{1, "debug", &pfm_sysctl.debug, sizeof(int), 0666, NULL, &proc_dointvec, NULL,},
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)
return 0UL;
}
-static inline unsigned long
+static inline void
pfm_unprotect_ctx_ctxsw(pfm_context_t *x, unsigned long f)
{
spin_unlock(&(x)->ctx_lock);
}
-static struct super_block *
-pfmfs_get_sb(struct file_system_type *fs_type, int flags, const char *dev_name, void *data)
+static int
+pfmfs_get_sb(struct file_system_type *fs_type, int flags, const char *dev_name, void *data,
+ struct vfsmount *mnt)
{
- return get_sb_pseudo(fs_type, "pfm:", NULL, PFMFS_MAGIC);
+ return get_sb_pseudo(fs_type, "pfm:", NULL, PFMFS_MAGIC, mnt);
}
static struct file_system_type pfm_fs_type = {
DEFINE_PER_CPU(struct task_struct *, pmu_owner);
DEFINE_PER_CPU(pfm_context_t *, pmu_ctx);
DEFINE_PER_CPU(unsigned long, pmu_activation_number);
+EXPORT_PER_CPU_SYMBOL_GPL(pfm_syst_info);
/* forward declaration */
#include "perfmon_itanium.h"
#include "perfmon_mckinley.h"
+#include "perfmon_montecito.h"
#include "perfmon_generic.h"
static pmu_config_t *pmu_confs[]={
+ &pmu_conf_mont,
&pmu_conf_mck,
&pmu_conf_ita,
&pmu_conf_gen, /* must be last */
* allocate context descriptor
* must be able to free with interrupts disabled
*/
- ctx = kmalloc(sizeof(pfm_context_t), GFP_KERNEL);
+ ctx = kzalloc(sizeof(pfm_context_t), GFP_KERNEL);
if (ctx) {
- memset(ctx, 0, sizeof(pfm_context_t));
DPRINT(("alloc ctx @%p\n", ctx));
}
return ctx;
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;
* 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
+ * As a consequence to this call, the ctx->th_pmds[] array
* contains stale information which must be ignored
* when context is reloaded AND monitoring is active (see
* pfm_restart).
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]));
+ ia64_set_pmc(i, ctx->th_pmcs[i] & ~0xfUL);
+ ctx->th_pmcs[i] &= ~0xfUL;
+ DPRINT_ovfl(("pmc[%d]=0x%lx\n", i, ctx->th_pmcs[i]));
}
/*
* make all of this visible
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;
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]));
+ ctx->th_pmcs[i] = ctx->ctx_pmcs[i];
+ ia64_set_pmc(i, ctx->th_pmcs[i]);
+ DPRINT(("[%d] pmc[%d]=0x%lx\n", task->pid, i, ctx->th_pmcs[i]));
}
ia64_srlz_d();
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;
ctx->ctx_pmds[i].val = val & ~ovfl_val;
val &= ovfl_val;
}
- thread->pmds[i] = val;
+ ctx->th_pmds[i] = val;
DPRINT(("pmd[%d]=0x%lx soft_val=0x%lx\n",
i,
- thread->pmds[i],
+ ctx->th_pmds[i],
ctx->ctx_pmds[i].val));
}
}
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;
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]));
+ ctx->th_pmcs[i] = ctx->ctx_pmcs[i];
+ DPRINT(("pmc[%d]=0x%lx\n", i, ctx->th_pmcs[i]));
}
}
}
EXPORT_SYMBOL(pfm_unregister_buffer_fmt);
+extern void update_pal_halt_status(int);
+
static int
pfm_reserve_session(struct task_struct *task, int is_syswide, unsigned int cpu)
{
is_syswide,
cpu));
+ /*
+ * disable default_idle() to go to PAL_HALT
+ */
+ update_pal_halt_status(0);
+
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()));
+ cpu));
abort:
UNLOCK_PFS(flags);
is_syswide,
cpu));
+ /*
+ * if possible, enable default_idle() to go into PAL_HALT
+ */
+ if (pfm_sessions.pfs_task_sessions == 0 && pfm_sessions.pfs_sys_sessions == 0)
+ update_pal_halt_status(1);
+
UNLOCK_PFS(flags);
return 0;
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));
+ DPRINT(("fd=%d type=%d\n", 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);
pfm_syswide_force_stop(void *info)
{
pfm_context_t *ctx = (pfm_context_t *)info;
- struct pt_regs *regs = ia64_task_regs(current);
+ struct pt_regs *regs = task_pt_regs(current);
struct task_struct *owner;
unsigned long flags;
int ret;
* When caller is self-monitoring, the context is unloaded.
*/
static int
-pfm_flush(struct file *filp)
+pfm_flush(struct file *filp, fl_owner_t id)
{
pfm_context_t *ctx;
struct task_struct *task;
is_system = ctx->ctx_fl_system;
task = PFM_CTX_TASK(ctx);
- regs = ia64_task_regs(task);
+ regs = task_pt_regs(task);
DPRINT(("ctx_state=%d is_current=%d\n",
state,
is_system = ctx->ctx_fl_system;
task = PFM_CTX_TASK(ctx);
- regs = ia64_task_regs(task);
+ regs = task_pt_regs(task);
DPRINT(("ctx_state=%d is_current=%d\n",
state,
/*
* force task to wake up from MASKED state
*/
- up(&ctx->ctx_restart_sem);
+ complete(&ctx->ctx_restart_done);
DPRINT(("waking up ctx_state=%d\n", state));
/*
* 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_path.dentry = d_alloc(pfmfs_mnt->mnt_sb->s_root, &this);
+ if (!file->f_path.dentry) goto out;
- file->f_dentry->d_op = &pfmfs_dentry_operations;
+ file->f_path.dentry->d_op = &pfmfs_dentry_operations;
- d_add(file->f_dentry, inode);
- file->f_vfsmnt = mntget(pfmfs_mnt);
+ d_add(file->f_path.dentry, inode);
+ file->f_path.mnt = mntget(pfmfs_mnt);
file->f_mapping = inode->i_mapping;
file->f_op = &pfm_file_ops;
pfm_free_fd(int fd, struct file *file)
{
struct files_struct *files = current->files;
+ struct fdtable *fdt;
/*
* there ie no fd_uninstall(), so we do it here
*/
spin_lock(&files->file_lock);
- files->fd[fd] = NULL;
+ fdt = files_fdtable(files);
+ rcu_assign_pointer(fdt->fd[fd], NULL);
spin_unlock(&files->file_lock);
- if (file) put_filp(file);
+ if (file)
+ put_filp(file);
put_unused_fd(fd);
}
DPRINT(("smpl_buf @%p\n", smpl_buf));
/* allocate vma */
- vma = kmem_cache_alloc(vm_area_cachep, SLAB_KERNEL);
+ vma = kmem_cache_alloc(vm_area_cachep, GFP_KERNEL);
if (!vma) {
DPRINT(("Cannot allocate vma\n"));
goto error_kmem;
insert_vm_struct(mm, vma);
vx_vmpages_add(mm, size >> PAGE_SHIFT);
- vm_stat_account(vma);
+ vm_stat_account(vma->vm_mm, vma->vm_flags, vma->vm_file,
+ vma_pages(vma));
up_write(&task->mm->mmap_sem);
/*
/*
* init restart semaphore to locked
*/
- sema_init(&ctx->ctx_restart_sem, 0);
+ init_completion(&ctx->ctx_restart_done);
/*
* activation is used in SMP only
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;
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.
*
* 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
+ * The value in th_pmcs[] may be modified on overflow, i.e., when
* monitoring needs to be stopped.
*/
if (is_monitor) CTX_USED_MONITOR(ctx, 1UL << cnum);
/*
* write thread state
*/
- if (is_system == 0) thread->pmcs[cnum] = value;
+ if (is_system == 0) ctx->th_pmcs[cnum] = value;
/*
* write hardware register if we can
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;
* 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.
/*
* write thread state
*/
- if (is_system == 0) thread->pmds[cnum] = hw_value;
+ if (is_system == 0) ctx->th_pmds[cnum] = hw_value;
/*
* write hardware register if we can
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;
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.
* 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;
+ val = is_loaded ? ctx->th_pmds[cnum] : 0UL;
}
rd_func = pmu_conf->pmd_desc[cnum].read_check;
*/
if (CTX_OVFL_NOBLOCK(ctx) == 0 && state == PFM_CTX_MASKED) {
DPRINT(("unblocking [%d] \n", task->pid));
- up(&ctx->ctx_restart_sem);
+ complete(&ctx->ctx_restart_done);
} else {
DPRINT(("[%d] armed exit trap\n", task->pid));
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) {
*/
ia64_psr(regs)->up = 0;
} else {
- tregs = ia64_task_regs(task);
+ tregs = task_pt_regs(task);
/*
* stop monitoring at the user level
ia64_psr(regs)->up = 1;
} else {
- tregs = ia64_task_regs(ctx->ctx_task);
+ tregs = task_pt_regs(ctx->ctx_task);
/*
* start monitoring at the kernel level the next
DPRINT(("cannot load to [%d], invalid ctx_state=%d\n",
req->load_pid,
ctx->ctx_state));
- return -EINVAL;
+ return -EBUSY;
}
DPRINT(("load_pid [%d] using_dbreg=%d\n", req->load_pid, ctx->ctx_fl_using_dbreg));
DPRINT(("before cmpxchg() old_ctx=%p new_ctx=%p\n",
thread->pfm_context, ctx));
+ ret = -EBUSY;
old = ia64_cmpxchg(acq, &thread->pfm_context, NULL, ctx, sizeof(pfm_context_t *));
if (old != NULL) {
DPRINT(("load_pid [%d] already has a context\n", req->load_pid));
pfm_copy_pmds(task, ctx);
pfm_copy_pmcs(task, ctx);
- pmcs_source = thread->pmcs;
- pmds_source = thread->pmds;
+ pmcs_source = ctx->th_pmcs;
+ pmds_source = ctx->th_pmds;
/*
* always the case for system-wide
/*
* when not current, task MUST be stopped, so this is safe
*/
- regs = ia64_task_regs(task);
+ regs = task_pt_regs(task);
/* force a full reload */
ctx->ctx_last_activation = PFM_INVALID_ACTIVATION;
/*
* per-task mode
*/
- tregs = task == current ? regs : ia64_task_regs(task);
+ tregs = task == current ? regs : task_pt_regs(task);
if (task == current) {
/*
{
pfm_context_t *ctx;
unsigned long flags;
- struct pt_regs *regs = ia64_task_regs(task);
+ struct pt_regs *regs = task_pt_regs(task);
int ret, state;
int free_ok = 0;
if (task == current || ctx->ctx_fl_system) return 0;
/*
- * if context is UNLOADED we are safe to go
+ * we are monitoring another thread
*/
- if (state == PFM_CTX_UNLOADED) return 0;
-
- /*
- * no command can operate on a zombie context
- */
- if (state == PFM_CTX_ZOMBIE) {
- DPRINT(("cmd %d state zombie cannot operate on context\n", cmd));
- return -EINVAL;
+ switch(state) {
+ case PFM_CTX_UNLOADED:
+ /*
+ * if context is UNLOADED we are safe to go
+ */
+ return 0;
+ case PFM_CTX_ZOMBIE:
+ /*
+ * no command can operate on a zombie context
+ */
+ DPRINT(("cmd %d state zombie cannot operate on context\n", cmd));
+ return -EINVAL;
+ case PFM_CTX_MASKED:
+ /*
+ * PMU state has been saved to software even though
+ * the thread may still be running.
+ */
+ if (cmd != PFM_UNLOAD_CONTEXT) return 0;
}
/*
if (unlikely(ret)) goto abort_locked;
skip_fd:
- ret = (*func)(ctx, args_k, count, ia64_task_regs(current));
+ ret = (*func)(ctx, args_k, count, task_pt_regs(current));
call_made = 1;
if (likely(ctx)) {
DPRINT(("context unlocked\n"));
UNPROTECT_CTX(ctx, flags);
- fput(file);
}
/* copy argument back to user, if needed */
if (call_made && PFM_CMD_RW_ARG(cmd) && copy_to_user(arg, args_k, base_sz*count)) ret = -EFAULT;
error_args:
- if (args_k) kfree(args_k);
+ if (file)
+ fput(file);
+
+ kfree(args_k);
DPRINT(("cmd=%s ret=%ld\n", PFM_CMD_NAME(cmd), ret));
}
static int pfm_ovfl_notify_user(pfm_context_t *ctx, unsigned long ovfl_pmds);
-
+ /*
+ * pfm_handle_work() can be called with interrupts enabled
+ * (TIF_NEED_RESCHED) or disabled. The down_interruptible
+ * call may sleep, therefore we must re-enable interrupts
+ * to avoid deadlocks. It is safe to do so because this function
+ * is called ONLY when returning to user level (PUStk=1), in which case
+ * there is no risk of kernel stack overflow due to deep
+ * interrupt nesting.
+ */
void
pfm_handle_work(void)
{
pfm_context_t *ctx;
struct pt_regs *regs;
- unsigned long flags;
+ unsigned long flags, dummy_flags;
unsigned long ovfl_regs;
unsigned int reason;
int ret;
pfm_clear_task_notify();
- regs = ia64_task_regs(current);
+ regs = task_pt_regs(current);
/*
* extract reason for being here and clear
//if (CTX_OVFL_NOBLOCK(ctx)) goto skip_blocking;
if (reason == PFM_TRAP_REASON_RESET) goto skip_blocking;
+ /*
+ * restore interrupt mask to what it was on entry.
+ * Could be enabled/diasbled.
+ */
UNPROTECT_CTX(ctx, flags);
- /*
- * pfm_handle_work() is currently called with interrupts disabled.
- * The down_interruptible call may sleep, therefore we
- * must re-enable interrupts to avoid deadlocks. It is
- * safe to do so because this function is called ONLY
- * when returning to user level (PUStk=1), in which case
- * there is no risk of kernel stack overflow due to deep
- * interrupt nesting.
- */
- BUG_ON(flags & IA64_PSR_I);
+ /*
+ * force interrupt enable because of down_interruptible()
+ */
local_irq_enable();
DPRINT(("before block sleeping\n"));
* may go through without blocking on SMP systems
* if restart has been received already by the time we call down()
*/
- ret = down_interruptible(&ctx->ctx_restart_sem);
+ ret = wait_for_completion_interruptible(&ctx->ctx_restart_done);
DPRINT(("after block sleeping ret=%d\n", ret));
/*
- * disable interrupts to restore state we had upon entering
- * this function
+ * lock context and mask interrupts again
+ * We save flags into a dummy because we may have
+ * altered interrupts mask compared to entry in this
+ * function.
*/
- local_irq_disable();
-
- PROTECT_CTX(ctx, flags);
+ PROTECT_CTX(ctx, dummy_flags);
/*
* we need to read the ovfl_regs only after wake-up
ctx->ctx_ovfl_regs[0] = 0UL;
nothing_to_do:
-
+ /*
+ * restore flags as they were upon entry
+ */
UNPROTECT_CTX(ctx, flags);
}
}
static irqreturn_t
-pfm_interrupt_handler(int irq, void *arg, struct pt_regs *regs)
+pfm_interrupt_handler(int irq, void *arg)
{
unsigned long start_cycles, total_cycles;
unsigned long min, max;
int this_cpu;
int ret;
+ struct pt_regs *regs = get_irq_regs();
this_cpu = get_cpu();
- min = pfm_stats[this_cpu].pfm_ovfl_intr_cycles_min;
- max = pfm_stats[this_cpu].pfm_ovfl_intr_cycles_max;
+ if (likely(!pfm_alt_intr_handler)) {
+ min = pfm_stats[this_cpu].pfm_ovfl_intr_cycles_min;
+ max = pfm_stats[this_cpu].pfm_ovfl_intr_cycles_max;
- start_cycles = ia64_get_itc();
+ start_cycles = ia64_get_itc();
- ret = pfm_do_interrupt_handler(irq, arg, regs);
+ ret = pfm_do_interrupt_handler(irq, arg, regs);
- total_cycles = ia64_get_itc();
+ total_cycles = ia64_get_itc();
- /*
- * don't measure spurious interrupts
- */
- if (likely(ret == 0)) {
- total_cycles -= start_cycles;
+ /*
+ * don't measure spurious interrupts
+ */
+ if (likely(ret == 0)) {
+ total_cycles -= start_cycles;
- if (total_cycles < min) pfm_stats[this_cpu].pfm_ovfl_intr_cycles_min = total_cycles;
- if (total_cycles > max) pfm_stats[this_cpu].pfm_ovfl_intr_cycles_max = total_cycles;
+ if (total_cycles < min) pfm_stats[this_cpu].pfm_ovfl_intr_cycles_min = total_cycles;
+ if (total_cycles > max) pfm_stats[this_cpu].pfm_ovfl_intr_cycles_max = total_cycles;
- pfm_stats[this_cpu].pfm_ovfl_intr_cycles += total_cycles;
+ pfm_stats[this_cpu].pfm_ovfl_intr_cycles += total_cycles;
+ }
+ }
+ else {
+ (*pfm_alt_intr_handler->handler)(irq, arg, regs);
}
+
put_cpu_no_resched();
return IRQ_HANDLED;
}
* on every CPU, so we can rely on the pid to identify the idle task.
*/
if ((info & PFM_CPUINFO_EXCL_IDLE) == 0 || task->pid) {
- regs = ia64_task_regs(task);
+ regs = task_pt_regs(task);
ia64_psr(regs)->pp = is_ctxswin ? dcr_pp : 0;
return;
}
pfm_save_regs(struct task_struct *task)
{
pfm_context_t *ctx;
- struct thread_struct *t;
unsigned long flags;
u64 psr;
ctx = PFM_GET_CTX(task);
if (ctx == NULL) return;
- t = &task->thread;
/*
* we always come here with interrupts ALREADY disabled by
flags = pfm_protect_ctx_ctxsw(ctx);
if (ctx->ctx_state == PFM_CTX_ZOMBIE) {
- struct pt_regs *regs = ia64_task_regs(task);
+ struct pt_regs *regs = task_pt_regs(task);
pfm_clear_psr_up();
* guarantee we will be schedule at that same
* CPU again.
*/
- pfm_save_pmds(t->pmds, ctx->ctx_used_pmds[0]);
+ pfm_save_pmds(ctx->th_pmds, ctx->ctx_used_pmds[0]);
/*
* save pmc0 ia64_srlz_d() done in pfm_save_pmds()
* we will need it on the restore path to check
* for pending overflow.
*/
- t->pmcs[0] = ia64_get_pmc(0);
+ ctx->th_pmcs[0] = ia64_get_pmc(0);
/*
* unfreeze PMU if had pending overflows
*/
- if (t->pmcs[0] & ~0x1UL) pfm_unfreeze_pmu();
+ if (ctx->th_pmcs[0] & ~0x1UL) pfm_unfreeze_pmu();
/*
* finally, allow context access.
pfm_lazy_save_regs (struct task_struct *task)
{
pfm_context_t *ctx;
- struct thread_struct *t;
unsigned long flags;
{ u64 psr = pfm_get_psr();
}
ctx = PFM_GET_CTX(task);
- t = &task->thread;
/*
* we need to mask PMU overflow here to
/*
* save all the pmds we use
*/
- pfm_save_pmds(t->pmds, ctx->ctx_used_pmds[0]);
+ pfm_save_pmds(ctx->th_pmds, ctx->ctx_used_pmds[0]);
/*
* save pmc0 ia64_srlz_d() done in pfm_save_pmds()
* it is needed to check for pended overflow
* on the restore path
*/
- t->pmcs[0] = ia64_get_pmc(0);
+ ctx->th_pmcs[0] = ia64_get_pmc(0);
/*
* unfreeze PMU if had pending overflows
*/
- if (t->pmcs[0] & ~0x1UL) pfm_unfreeze_pmu();
+ if (ctx->th_pmcs[0] & ~0x1UL) pfm_unfreeze_pmu();
/*
* now get can unmask PMU interrupts, they will
pfm_load_regs (struct task_struct *task)
{
pfm_context_t *ctx;
- struct thread_struct *t;
unsigned long pmc_mask = 0UL, pmd_mask = 0UL;
unsigned long flags;
u64 psr, psr_up;
BUG_ON(GET_PMU_OWNER());
- t = &task->thread;
/*
* possible on unload
*/
- if (unlikely((t->flags & IA64_THREAD_PM_VALID) == 0)) return;
+ if (unlikely((task->thread.flags & IA64_THREAD_PM_VALID) == 0)) return;
/*
* we always come here with interrupts ALREADY disabled by
BUG_ON(psr & IA64_PSR_I);
if (unlikely(ctx->ctx_state == PFM_CTX_ZOMBIE)) {
- struct pt_regs *regs = ia64_task_regs(task);
+ struct pt_regs *regs = task_pt_regs(task);
BUG_ON(ctx->ctx_smpl_hdr);
*
* XXX: optimize here
*/
- if (pmd_mask) pfm_restore_pmds(t->pmds, pmd_mask);
- if (pmc_mask) pfm_restore_pmcs(t->pmcs, pmc_mask);
+ if (pmd_mask) pfm_restore_pmds(ctx->th_pmds, pmd_mask);
+ if (pmc_mask) pfm_restore_pmcs(ctx->th_pmcs, pmc_mask);
/*
* check for pending overflow at the time the state
* was saved.
*/
- if (unlikely(PMC0_HAS_OVFL(t->pmcs[0]))) {
+ if (unlikely(PMC0_HAS_OVFL(ctx->th_pmcs[0]))) {
/*
* reload pmc0 with the overflow information
* On McKinley PMU, this will trigger a PMU interrupt
*/
- ia64_set_pmc(0, t->pmcs[0]);
+ ia64_set_pmc(0, ctx->th_pmcs[0]);
ia64_srlz_d();
- t->pmcs[0] = 0UL;
+ ctx->th_pmcs[0] = 0UL;
/*
* will replay the PMU interrupt
*/
- if (need_irq_resend) hw_resend_irq(NULL, IA64_PERFMON_VECTOR);
+ if (need_irq_resend) ia64_resend_irq(IA64_PERFMON_VECTOR);
pfm_stats[smp_processor_id()].pfm_replay_ovfl_intr_count++;
}
void
pfm_load_regs (struct task_struct *task)
{
- struct thread_struct *t;
pfm_context_t *ctx;
struct task_struct *owner;
unsigned long pmd_mask, pmc_mask;
owner = GET_PMU_OWNER();
ctx = PFM_GET_CTX(task);
- t = &task->thread;
psr = pfm_get_psr();
BUG_ON(psr & (IA64_PSR_UP|IA64_PSR_PP));
*/
pmc_mask = ctx->ctx_all_pmcs[0];
- pfm_restore_pmds(t->pmds, pmd_mask);
- pfm_restore_pmcs(t->pmcs, pmc_mask);
+ pfm_restore_pmds(ctx->th_pmds, pmd_mask);
+ pfm_restore_pmcs(ctx->th_pmcs, pmc_mask);
/*
* check for pending overflow at the time the state
* was saved.
*/
- if (unlikely(PMC0_HAS_OVFL(t->pmcs[0]))) {
+ if (unlikely(PMC0_HAS_OVFL(ctx->th_pmcs[0]))) {
/*
* reload pmc0 with the overflow information
* On McKinley PMU, this will trigger a PMU interrupt
*/
- ia64_set_pmc(0, t->pmcs[0]);
+ ia64_set_pmc(0, ctx->th_pmcs[0]);
ia64_srlz_d();
- t->pmcs[0] = 0UL;
+ ctx->th_pmcs[0] = 0UL;
/*
* will replay the PMU interrupt
*/
- if (need_irq_resend) hw_resend_irq(NULL, IA64_PERFMON_VECTOR);
+ if (need_irq_resend) ia64_resend_irq(IA64_PERFMON_VECTOR);
pfm_stats[smp_processor_id()].pfm_replay_ovfl_intr_count++;
}
*/
pfm_unfreeze_pmu();
} else {
- pmc0 = task->thread.pmcs[0];
+ pmc0 = ctx->th_pmcs[0];
/*
* clear whatever overflow status bits there were
*/
- task->thread.pmcs[0] = 0;
+ ctx->th_pmcs[0] = 0;
}
ovfl_val = pmu_conf->ovfl_val;
/*
/*
* can access PMU always true in system wide mode
*/
- val = pmd_val = can_access_pmu ? ia64_get_pmd(i) : task->thread.pmds[i];
+ val = pmd_val = can_access_pmu ? ia64_get_pmd(i) : ctx->th_pmds[i];
if (PMD_IS_COUNTING(i)) {
DPRINT(("[%d] pmd[%d] ctx_pmd=0x%lx hw_pmd=0x%lx\n",
DPRINT(("[%d] ctx_pmd[%d]=0x%lx pmd_val=0x%lx\n", task->pid, i, val, pmd_val));
- if (is_self) task->thread.pmds[i] = pmd_val;
+ if (is_self) ctx->th_pmds[i] = pmd_val;
ctx->ctx_pmds[i].val = val;
}
static struct irqaction perfmon_irqaction = {
.handler = pfm_interrupt_handler,
- .flags = SA_INTERRUPT,
+ .flags = IRQF_DISABLED,
.name = "perfmon"
};
+static void
+pfm_alt_save_pmu_state(void *data)
+{
+ struct pt_regs *regs;
+
+ regs = task_pt_regs(current);
+
+ DPRINT(("called\n"));
+
+ /*
+ * should not be necessary but
+ * let's take not risk
+ */
+ pfm_clear_psr_up();
+ pfm_clear_psr_pp();
+ ia64_psr(regs)->pp = 0;
+
+ /*
+ * This call is required
+ * May cause a spurious interrupt on some processors
+ */
+ pfm_freeze_pmu();
+
+ ia64_srlz_d();
+}
+
+void
+pfm_alt_restore_pmu_state(void *data)
+{
+ struct pt_regs *regs;
+
+ regs = task_pt_regs(current);
+
+ DPRINT(("called\n"));
+
+ /*
+ * put PMU back in state expected
+ * by perfmon
+ */
+ pfm_clear_psr_up();
+ pfm_clear_psr_pp();
+ ia64_psr(regs)->pp = 0;
+
+ /*
+ * perfmon runs with PMU unfrozen at all times
+ */
+ pfm_unfreeze_pmu();
+
+ ia64_srlz_d();
+}
+
+int
+pfm_install_alt_pmu_interrupt(pfm_intr_handler_desc_t *hdl)
+{
+ int ret, i;
+ int reserve_cpu;
+
+ /* some sanity checks */
+ if (hdl == NULL || hdl->handler == NULL) return -EINVAL;
+
+ /* do the easy test first */
+ if (pfm_alt_intr_handler) return -EBUSY;
+
+ /* one at a time in the install or remove, just fail the others */
+ if (!spin_trylock(&pfm_alt_install_check)) {
+ return -EBUSY;
+ }
+
+ /* reserve our session */
+ for_each_online_cpu(reserve_cpu) {
+ ret = pfm_reserve_session(NULL, 1, reserve_cpu);
+ if (ret) goto cleanup_reserve;
+ }
+
+ /* save the current system wide pmu states */
+ ret = on_each_cpu(pfm_alt_save_pmu_state, NULL, 0, 1);
+ if (ret) {
+ DPRINT(("on_each_cpu() failed: %d\n", ret));
+ goto cleanup_reserve;
+ }
+
+ /* officially change to the alternate interrupt handler */
+ pfm_alt_intr_handler = hdl;
+
+ spin_unlock(&pfm_alt_install_check);
+
+ return 0;
+
+cleanup_reserve:
+ for_each_online_cpu(i) {
+ /* don't unreserve more than we reserved */
+ if (i >= reserve_cpu) break;
+
+ pfm_unreserve_session(NULL, 1, i);
+ }
+
+ spin_unlock(&pfm_alt_install_check);
+
+ return ret;
+}
+EXPORT_SYMBOL_GPL(pfm_install_alt_pmu_interrupt);
+
+int
+pfm_remove_alt_pmu_interrupt(pfm_intr_handler_desc_t *hdl)
+{
+ int i;
+ int ret;
+
+ if (hdl == NULL) return -EINVAL;
+
+ /* cannot remove someone else's handler! */
+ if (pfm_alt_intr_handler != hdl) return -EINVAL;
+
+ /* one at a time in the install or remove, just fail the others */
+ if (!spin_trylock(&pfm_alt_install_check)) {
+ return -EBUSY;
+ }
+
+ pfm_alt_intr_handler = NULL;
+
+ ret = on_each_cpu(pfm_alt_restore_pmu_state, NULL, 0, 1);
+ if (ret) {
+ DPRINT(("on_each_cpu() failed: %d\n", ret));
+ }
+
+ for_each_online_cpu(i) {
+ pfm_unreserve_session(NULL, 1, i);
+ }
+
+ spin_unlock(&pfm_alt_install_check);
+
+ return 0;
+}
+EXPORT_SYMBOL_GPL(pfm_remove_alt_pmu_interrupt);
+
/*
* perfmon initialization routine, called from the initcall() table
*/
ffz(pmu_conf->ovfl_val));
/* sanity check */
- if (pmu_conf->num_pmds >= IA64_NUM_PMD_REGS || pmu_conf->num_pmcs >= IA64_NUM_PMC_REGS) {
+ if (pmu_conf->num_pmds >= PFM_NUM_PMD_REGS || pmu_conf->num_pmcs >= PFM_NUM_PMC_REGS) {
printk(KERN_ERR "perfmon: not enough pmc/pmd, perfmon disabled\n");
pmu_conf = NULL;
return -1;
void
pfm_init_percpu (void)
{
+ static int first_time=1;
/*
* make sure no measurement is active
* (may inherit programmed PMCs from EFI).
*/
pfm_unfreeze_pmu();
- if (smp_processor_id() == 0)
+ if (first_time) {
register_percpu_irq(IA64_PERFMON_VECTOR, &perfmon_irqaction);
+ first_time=0;
+ }
ia64_setreg(_IA64_REG_CR_PMV, IA64_PERFMON_VECTOR);
ia64_srlz_d();
dump_pmu_state(const char *from)
{
struct task_struct *task;
- struct thread_struct *t;
struct pt_regs *regs;
pfm_context_t *ctx;
unsigned long psr, dcr, info, flags;
local_irq_save(flags);
this_cpu = smp_processor_id();
- regs = ia64_task_regs(current);
+ regs = task_pt_regs(current);
info = PFM_CPUINFO_GET();
dcr = ia64_getreg(_IA64_REG_CR_DCR);
ia64_psr(regs)->up = 0;
ia64_psr(regs)->pp = 0;
- t = ¤t->thread;
-
for (i=1; PMC_IS_LAST(i) == 0; i++) {
if (PMC_IS_IMPL(i) == 0) continue;
- printk("->CPU%d pmc[%d]=0x%lx thread_pmc[%d]=0x%lx\n", this_cpu, i, ia64_get_pmc(i), i, t->pmcs[i]);
+ printk("->CPU%d pmc[%d]=0x%lx thread_pmc[%d]=0x%lx\n", this_cpu, i, ia64_get_pmc(i), i, ctx->th_pmcs[i]);
}
for (i=1; PMD_IS_LAST(i) == 0; i++) {
if (PMD_IS_IMPL(i) == 0) continue;
- printk("->CPU%d pmd[%d]=0x%lx thread_pmd[%d]=0x%lx\n", this_cpu, i, ia64_get_pmd(i), i, t->pmds[i]);
+ printk("->CPU%d pmd[%d]=0x%lx thread_pmd[%d]=0x%lx\n", this_cpu, i, ia64_get_pmd(i), i, ctx->th_pmds[i]);
}
if (ctx) {
git://git.onelab.eu / linux-2.6.git / blobdiff