Commit 19250aed authored by David Mosberger's avatar David Mosberger

ia64: Fix potential perfmon deadlock. Patch by Stephane Eranian.

parent 2995a981
......@@ -267,8 +267,6 @@ typedef struct pfm_context {
unsigned long ctx_saved_cpus_allowed; /* copy of the task cpus_allowed (system wide) */
unsigned int ctx_cpu; /* CPU used by system wide session */
atomic_t ctx_saving_in_progress; /* flag indicating actual save in progress */
atomic_t ctx_is_busy; /* context accessed by overflow handler */
atomic_t ctx_last_cpu; /* CPU id of current or last CPU used */
} pfm_context_t;
......@@ -439,9 +437,6 @@ static struct {
* forward declarations
*/
static void pfm_reset_pmu(struct task_struct *);
#ifdef CONFIG_SMP
static void pfm_fetch_regs(int cpu, struct task_struct *task, pfm_context_t *ctx);
#endif
static void pfm_lazy_save_regs (struct task_struct *ta);
#if defined(CONFIG_ITANIUM)
......@@ -490,6 +485,19 @@ pfm_set_psr_l(unsigned long val)
__asm__ __volatile__ ("mov psr.l=%0;; srlz.i;;"::"r"(val): "memory");
}
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 unsigned long
pfm_read_soft_counter(pfm_context_t *ctx, int i)
......@@ -1230,10 +1238,6 @@ pfm_context_create(struct task_struct *task, pfm_context_t *ctx, void *req, int
atomic_set(&ctx->ctx_last_cpu,-1); /* SMP only, means no CPU */
/* may be redudant with memset() but at least it's easier to remember */
atomic_set(&ctx->ctx_saving_in_progress, 0);
atomic_set(&ctx->ctx_is_busy, 0);
sema_init(&ctx->ctx_restart_sem, 0); /* init this semaphore to locked */
if (__copy_to_user(req, &tmp, sizeof(tmp))) {
......@@ -1667,25 +1671,6 @@ pfm_read_pmds(struct task_struct *task, pfm_context_t *ctx, void *arg, int count
val = ia64_get_pmd(cnum);
DBprintk(("reading pmd[%u]=0x%lx from hw\n", cnum, val));
} else {
#ifdef CONFIG_SMP
int cpu;
/*
* for SMP system, the context may still be live on another
* CPU so we need to fetch it before proceeding with the read
* This call we only be made once for the whole loop because
* of ctx_last_cpu becoming == -1.
*
* We cannot reuse ctx_last_cpu as it may change before we get to the
* actual IPI call. In this case, we will do the call for nothing but
* there is no way around it. The receiving side will simply do nothing.
*/
cpu = atomic_read(&ctx->ctx_last_cpu);
if (cpu != -1) {
DBprintk(("must fetch on CPU%d for [%d]\n", cpu, task->pid));
pfm_fetch_regs(cpu, task, ctx);
}
#endif
/* context has been saved */
val = th->pmd[cnum];
}
if (PMD_IS_COUNTING(cnum)) {
......@@ -1862,8 +1847,7 @@ pfm_restart(struct task_struct *task, pfm_context_t *ctx, void *arg, int count,
}
/* simply unfreeze */
ia64_set_pmc(0, 0);
ia64_srlz_d();
pfm_unfreeze_pmu();
return 0;
}
......@@ -2416,8 +2400,7 @@ pfm_enable(struct task_struct *task, pfm_context_t *ctx, void *arg, int count,
atomic_set(&ctx->ctx_last_cpu, smp_processor_id());
/* simply unfreeze */
ia64_set_pmc(0, 0);
ia64_srlz_d();
pfm_unfreeze_pmu();
return 0;
}
......@@ -2665,8 +2648,7 @@ pfm_ovfl_block_reset(void)
ctx->ctx_psb->psb_index = 0;
}
ia64_set_pmc(0, 0);
ia64_srlz_d();
pfm_unfreeze_pmu();
/* state restored, can go back to work (user mode) */
}
......@@ -3073,19 +3055,6 @@ pfm_interrupt_handler(int irq, void *arg, struct pt_regs *regs)
"no PFM context\n", task->pid);
return;
}
#ifdef CONFIG_SMP
/*
* Because an IPI has higher priority than the PMU overflow interrupt, it is
* possible that the handler be interrupted by a request from another CPU to fetch
* the PMU state of the currently active context. The task may have just been
* migrated to another CPU which is trying to restore the context. If there was
* a pending overflow interrupt when the task left this CPU, it is possible for
* the handler to get interrupt by the IPI. In which case, we fetch request
* MUST be postponed until the interrupt handler is done. The ctx_is_busy
* flag indicates such a condition. The other CPU must busy wait until it's cleared.
*/
atomic_set(&ctx->ctx_is_busy, 1);
#endif
/*
* assume PMC[0].fr = 1 at this point
......@@ -3099,12 +3068,6 @@ pfm_interrupt_handler(int irq, void *arg, struct pt_regs *regs)
ia64_set_pmc(0, pmc0);
ia64_srlz_d();
#ifdef CONFIG_SMP
/*
* announce that we are doing with the context
*/
atomic_set(&ctx->ctx_is_busy, 0);
#endif
} else {
pfm_stats[smp_processor_id()].pfm_spurious_ovfl_intr_count++;
}
......@@ -3222,10 +3185,13 @@ void
pfm_save_regs (struct task_struct *task)
{
pfm_context_t *ctx;
unsigned long mask;
u64 psr;
int i;
ctx = task->thread.pfm_context;
/*
* save current PSR: needed because we modify it
*/
......@@ -3238,129 +3204,61 @@ pfm_save_regs (struct task_struct *task)
* We do not need to set psr.sp because, it is irrelevant in kernel.
* It will be restored from ipsr when going back to user level
*/
__asm__ __volatile__ ("rum psr.up;;"::: "memory");
pfm_clear_psr_up();
ia64_srlz_i();
ctx->ctx_saved_psr = psr;
//ctx->ctx_last_cpu = smp_processor_id();
}
static void
pfm_lazy_save_regs (struct task_struct *task)
{
pfm_context_t *ctx;
struct thread_struct *t;
unsigned long mask;
int i;
DBprintk(("on [%d] by [%d]\n", task->pid, current->pid));
t = &task->thread;
ctx = task->thread.pfm_context;
#ifdef CONFIG_SMP
/*
* announce we are saving this PMU state
* This will cause other CPU, to wait until we're done
* before using the context.h
*
* must be an atomic operation
*/
atomic_set(&ctx->ctx_saving_in_progress, 1);
/*
* if owner is NULL, it means that the other CPU won the race
* and the IPI has caused the context to be saved in pfm_handle_fectch_regs()
* instead of here. We have nothing to do
* We do not use a lazy scheme in SMP because
* of the new scheduler which masks interrupts
* during low-level context switch. So we save
* all the PMD register we use and restore on
* ctxsw in.
*
* note that this is safe, because the other CPU NEVER modifies saving_in_progress.
* release ownership of this PMU.
* must be done before we save the registers.
*/
if (PMU_OWNER() == NULL) goto do_nothing;
#endif
SET_PMU_OWNER(NULL);
/*
* do not own the PMU
* save PMDs
*/
SET_PMU_OWNER(NULL);
ia64_srlz_d();
/*
* XXX needs further optimization.
* Also must take holes into account
*/
mask = ctx->ctx_used_pmds[0];
for (i=0; mask; i++, mask>>=1) {
if (mask & 0x1) t->pmd[i] =ia64_get_pmd(i);
if (mask & 0x1) task->thread.pmd[i] =ia64_get_pmd(i);
}
/* save pmc0 */
t->pmc[0] = ia64_get_pmc(0);
/* not owned by this CPU */
atomic_set(&ctx->ctx_last_cpu, -1);
#ifdef CONFIG_SMP
do_nothing:
#endif
/*
* declare we are done saving this context
*
* must be an atomic operation
* save pmc0
*/
atomic_set(&ctx->ctx_saving_in_progress,0);
task->thread.pmc[0] = ia64_get_pmc(0);
/*
* force a full reload
*/
atomic_set(&ctx->ctx_last_cpu, -1);
#endif
}
#ifdef CONFIG_SMP
/*
* Handles request coming from other CPUs
*/
static void
pfm_handle_fetch_regs(void *info)
pfm_lazy_save_regs (struct task_struct *task)
{
pfm_smp_ipi_arg_t *arg = info;
struct thread_struct *t;
pfm_context_t *ctx;
struct thread_struct *t;
unsigned long mask;
int i;
ctx = arg->task->thread.pfm_context;
t = &arg->task->thread;
DBprintk(("task=%d owner=%d saving=%d\n",
arg->task->pid,
PMU_OWNER() ? PMU_OWNER()->pid: -1,
atomic_read(&ctx->ctx_saving_in_progress)));
/* must wait until not busy before retrying whole request */
if (atomic_read(&ctx->ctx_is_busy)) {
arg->retval = 2;
return;
}
/* must wait if saving was interrupted */
if (atomic_read(&ctx->ctx_saving_in_progress)) {
arg->retval = 1;
return;
}
/* can proceed, done with context */
if (PMU_OWNER() != arg->task) {
arg->retval = 0;
return;
}
DBprintk(("on [%d] by [%d]\n", task->pid, current->pid));
DBprintk(("saving state for [%d] used_pmcs=0x%lx reload_pmcs=0x%lx used_pmds=0x%lx\n",
arg->task->pid,
ctx->ctx_used_pmcs[0],
ctx->ctx_reload_pmcs[0],
ctx->ctx_used_pmds[0]));
t = &task->thread;
ctx = task->thread.pfm_context;
/*
* XXX: will be replaced with pure assembly call
* do not own the PMU
*/
SET_PMU_OWNER(NULL);
......@@ -3368,10 +3266,11 @@ pfm_handle_fetch_regs(void *info)
/*
* XXX needs further optimization.
* Also must take holes into account
*/
mask = ctx->ctx_used_pmds[0];
for (i=0; mask; i++, mask>>=1) {
if (mask & 0x1) t->pmd[i] = ia64_get_pmd(i);
if (mask & 0x1) t->pmd[i] =ia64_get_pmd(i);
}
/* save pmc0 */
......@@ -3379,67 +3278,7 @@ pfm_handle_fetch_regs(void *info)
/* not owned by this CPU */
atomic_set(&ctx->ctx_last_cpu, -1);
/* can proceed */
arg->retval = 0;
}
/*
* Function call to fetch PMU state from another CPU identified by 'cpu'.
* If the context is being saved on the remote CPU, then we busy wait until
* the saving is done and then we return. In this case, non IPI is sent.
* Otherwise, we send an IPI to the remote CPU, potentially interrupting
* pfm_lazy_save_regs() over there.
*
* If the retval==1, then it means that we interrupted remote save and that we must
* wait until the saving is over before proceeding.
* Otherwise, we did the saving on the remote CPU, and it was done by the time we got there.
* in either case, we can proceed.
*/
static void
pfm_fetch_regs(int cpu, struct task_struct *task, pfm_context_t *ctx)
{
pfm_smp_ipi_arg_t arg;
int ret;
arg.task = task;
arg.retval = -1;
if (atomic_read(&ctx->ctx_is_busy)) {
must_wait_busy:
while (atomic_read(&ctx->ctx_is_busy));
}
if (atomic_read(&ctx->ctx_saving_in_progress)) {
DBprintk(("no IPI, must wait for [%d] to be saved on [%d]\n", task->pid, cpu));
must_wait_saving:
/* busy wait */
while (atomic_read(&ctx->ctx_saving_in_progress));
DBprintk(("done saving for [%d] on [%d]\n", task->pid, cpu));
return;
}
DBprintk(("calling CPU %d from CPU %d\n", cpu, smp_processor_id()));
if (cpu == -1) {
printk("refusing to use -1 for [%d]\n", task->pid);
return;
}
/* will send IPI to other CPU and wait for completion of remote call */
if ((ret=smp_call_function_single(cpu, pfm_handle_fetch_regs, &arg, 0, 1))) {
printk(KERN_ERR "perfmon: remote CPU call from %d to %d error %d\n",
smp_processor_id(), cpu, ret);
return;
}
/*
* we must wait until saving is over on the other CPU
* This is the case, where we interrupted the saving which started just at the time we sent the
* IPI.
*/
if (arg.retval == 1) goto must_wait_saving;
if (arg.retval == 2) goto must_wait_busy;
}
#endif /* CONFIG_SMP */
void
pfm_load_regs (struct task_struct *task)
......@@ -3450,14 +3289,16 @@ pfm_load_regs (struct task_struct *task)
unsigned long mask;
u64 psr;
int i;
#ifdef CONFIG_SMP
int cpu;
#endif
owner = PMU_OWNER();
ctx = task->thread.pfm_context;
t = &task->thread;
if (ctx == NULL) {
printk("perfmon: pfm_load_regs: null ctx for [%d]\n", task->pid);
return;
}
/*
* we restore ALL the debug registers to avoid picking up
* stale state.
......@@ -3483,6 +3324,7 @@ pfm_load_regs (struct task_struct *task)
/*
* if we were the last user, then nothing to do except restore psr
* this path cannot be used in SMP
*/
if (owner == task) {
if (atomic_read(&ctx->ctx_last_cpu) != smp_processor_id())
......@@ -3490,32 +3332,19 @@ pfm_load_regs (struct task_struct *task)
atomic_read(&ctx->ctx_last_cpu), task->pid));
psr = ctx->ctx_saved_psr;
__asm__ __volatile__ ("mov psr.l=%0;; srlz.i;;"::"r"(psr): "memory");
pfm_set_psr_l(psr);
return;
}
DBprintk(("load_regs: must reload for [%d] owner=%d\n",
task->pid, owner ? owner->pid : -1 ));
/*
* someone else is still using the PMU, first push it out and
* then we'll be able to install our stuff !
*
* not possible in SMP
*/
if (owner) pfm_lazy_save_regs(owner);
#ifdef CONFIG_SMP
/*
* check if context on another CPU (-1 means saved)
* We MUST use the variable, as last_cpu may change behind our
* back. If it changes to -1 (not on a CPU anymore), then in cpu
* we have the last CPU the context was on. We may be sending the
* IPI for nothing, but we have no way of verifying this.
*/
cpu = atomic_read(&ctx->ctx_last_cpu);
if (cpu != -1) {
pfm_fetch_regs(cpu, task, ctx);
}
#endif
/*
* To avoid leaking information to the user level when psr.sp=0,
* we must reload ALL implemented pmds (even the ones we don't use).
......@@ -3552,8 +3381,7 @@ pfm_load_regs (struct task_struct *task)
* fl_frozen==1 when we are in blocking mode waiting for restart
*/
if (ctx->ctx_fl_frozen == 0) {
ia64_set_pmc(0, 0);
ia64_srlz_d();
pfm_unfreeze_pmu();
}
atomic_set(&ctx->ctx_last_cpu, smp_processor_id());
......@@ -3563,8 +3391,7 @@ pfm_load_regs (struct task_struct *task)
* restore the psr we changed in pfm_save_regs()
*/
psr = ctx->ctx_saved_psr;
__asm__ __volatile__ ("mov psr.l=%0;; srlz.i;;"::"r"(psr): "memory");
pfm_set_psr_l(psr);
}
/*
......@@ -3583,7 +3410,7 @@ pfm_reset_pmu(struct task_struct *task)
}
/* Let's make sure the PMU is frozen */
ia64_set_pmc(0,1);
pfm_freeze_pmu();
/*
* install reset values for PMC. We skip PMC0 (done above)
......@@ -3750,8 +3577,7 @@ pfm_flush_regs (struct task_struct *task)
* This destroys the overflow information. This is required to make sure
* next process does not start with monitoring on if not requested
*/
ia64_set_pmc(0, 1);
ia64_srlz_d();
pfm_freeze_pmu();
/*
* We don't need to restore psr, because we are on our way out
......@@ -4433,8 +4259,7 @@ pfm_init_percpu(void)
if (PMD_IS_IMPL(i) == 0) continue;
ia64_set_pmd(i, 0UL);
}
ia64_set_pmc(0,1UL);
ia64_srlz_d();
pfm_freeze_pmu();
}
#else /* !CONFIG_PERFMON */
......
Markdown is supported
0%
or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment