perf/x86: Fix Intel shared extra MSR allocation
authorPeter Zijlstra <peterz@infradead.org>
Tue, 5 Jun 2012 13:30:31 +0000 (15:30 +0200)
committerIngo Molnar <mingo@kernel.org>
Wed, 6 Jun 2012 14:59:44 +0000 (16:59 +0200)
Zheng Yan reported that event group validation can wreck event state
when Intel extra_reg allocation changes event state.

Validation shouldn't change any persistent state. Cloning events in
validate_{event,group}() isn't really pretty either, so add a few
special cases to avoid modifying the event state.

The code is restructured to minimize the special case impact.

Reported-by: Zheng Yan <zheng.z.yan@linux.intel.com>
Acked-by: Stephane Eranian <eranian@google.com>
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Link: http://lkml.kernel.org/r/1338903031.28282.175.camel@twins
Signed-off-by: Ingo Molnar <mingo@kernel.org>
arch/x86/kernel/cpu/perf_event.c
arch/x86/kernel/cpu/perf_event.h
arch/x86/kernel/cpu/perf_event_intel.c

index e049d6da01832cfc91b5e2a45b922f592c7bb7ae..cb608383e4f608a4c5dd0ac987fe7777e0543fe1 100644 (file)
@@ -1496,6 +1496,7 @@ static struct cpu_hw_events *allocate_fake_cpuc(void)
                if (!cpuc->shared_regs)
                        goto error;
        }
+       cpuc->is_fake = 1;
        return cpuc;
 error:
        free_fake_cpuc(cpuc);
index 6638aaf5449302c2ea2d5d03073f11bbcaf5b6ba..83794d8e6af0fb20e98f7b4f4d90285f73d12d01 100644 (file)
@@ -117,6 +117,7 @@ struct cpu_hw_events {
        struct perf_event       *event_list[X86_PMC_IDX_MAX]; /* in enabled order */
 
        unsigned int            group_flag;
+       int                     is_fake;
 
        /*
         * Intel DebugStore bits
index 166546ec6aefe523a20fc5b4206d0ef9a87e4679..965baa2fa790ed4b6c6a9bddc12d575821cf86ec 100644 (file)
@@ -1119,27 +1119,33 @@ intel_bts_constraints(struct perf_event *event)
        return NULL;
 }
 
-static bool intel_try_alt_er(struct perf_event *event, int orig_idx)
+static int intel_alt_er(int idx)
 {
        if (!(x86_pmu.er_flags & ERF_HAS_RSP_1))
-               return false;
+               return idx;
 
-       if (event->hw.extra_reg.idx == EXTRA_REG_RSP_0) {
-               event->hw.config &= ~INTEL_ARCH_EVENT_MASK;
-               event->hw.config |= 0x01bb;
-               event->hw.extra_reg.idx = EXTRA_REG_RSP_1;
-               event->hw.extra_reg.reg = MSR_OFFCORE_RSP_1;
-       } else if (event->hw.extra_reg.idx == EXTRA_REG_RSP_1) {
+       if (idx == EXTRA_REG_RSP_0)
+               return EXTRA_REG_RSP_1;
+
+       if (idx == EXTRA_REG_RSP_1)
+               return EXTRA_REG_RSP_0;
+
+       return idx;
+}
+
+static void intel_fixup_er(struct perf_event *event, int idx)
+{
+       event->hw.extra_reg.idx = idx;
+
+       if (idx == EXTRA_REG_RSP_0) {
                event->hw.config &= ~INTEL_ARCH_EVENT_MASK;
                event->hw.config |= 0x01b7;
-               event->hw.extra_reg.idx = EXTRA_REG_RSP_0;
                event->hw.extra_reg.reg = MSR_OFFCORE_RSP_0;
+       } else if (idx == EXTRA_REG_RSP_1) {
+               event->hw.config &= ~INTEL_ARCH_EVENT_MASK;
+               event->hw.config |= 0x01bb;
+               event->hw.extra_reg.reg = MSR_OFFCORE_RSP_1;
        }
-
-       if (event->hw.extra_reg.idx == orig_idx)
-               return false;
-
-       return true;
 }
 
 /*
@@ -1157,14 +1163,18 @@ __intel_shared_reg_get_constraints(struct cpu_hw_events *cpuc,
        struct event_constraint *c = &emptyconstraint;
        struct er_account *era;
        unsigned long flags;
-       int orig_idx = reg->idx;
+       int idx = reg->idx;
 
-       /* already allocated shared msr */
-       if (reg->alloc)
+       /*
+        * reg->alloc can be set due to existing state, so for fake cpuc we
+        * need to ignore this, otherwise we might fail to allocate proper fake
+        * state for this extra reg constraint. Also see the comment below.
+        */
+       if (reg->alloc && !cpuc->is_fake)
                return NULL; /* call x86_get_event_constraint() */
 
 again:
-       era = &cpuc->shared_regs->regs[reg->idx];
+       era = &cpuc->shared_regs->regs[idx];
        /*
         * we use spin_lock_irqsave() to avoid lockdep issues when
         * passing a fake cpuc
@@ -1173,6 +1183,29 @@ again:
 
        if (!atomic_read(&era->ref) || era->config == reg->config) {
 
+               /*
+                * If its a fake cpuc -- as per validate_{group,event}() we
+                * shouldn't touch event state and we can avoid doing so
+                * since both will only call get_event_constraints() once
+                * on each event, this avoids the need for reg->alloc.
+                *
+                * Not doing the ER fixup will only result in era->reg being
+                * wrong, but since we won't actually try and program hardware
+                * this isn't a problem either.
+                */
+               if (!cpuc->is_fake) {
+                       if (idx != reg->idx)
+                               intel_fixup_er(event, idx);
+
+                       /*
+                        * x86_schedule_events() can call get_event_constraints()
+                        * multiple times on events in the case of incremental
+                        * scheduling(). reg->alloc ensures we only do the ER
+                        * allocation once.
+                        */
+                       reg->alloc = 1;
+               }
+
                /* lock in msr value */
                era->config = reg->config;
                era->reg = reg->reg;
@@ -1180,17 +1213,17 @@ again:
                /* one more user */
                atomic_inc(&era->ref);
 
-               /* no need to reallocate during incremental event scheduling */
-               reg->alloc = 1;
-
                /*
                 * need to call x86_get_event_constraint()
                 * to check if associated event has constraints
                 */
                c = NULL;
-       } else if (intel_try_alt_er(event, orig_idx)) {
-               raw_spin_unlock_irqrestore(&era->lock, flags);
-               goto again;
+       } else {
+               idx = intel_alt_er(idx);
+               if (idx != reg->idx) {
+                       raw_spin_unlock_irqrestore(&era->lock, flags);
+                       goto again;
+               }
        }
        raw_spin_unlock_irqrestore(&era->lock, flags);
 
@@ -1204,11 +1237,14 @@ __intel_shared_reg_put_constraints(struct cpu_hw_events *cpuc,
        struct er_account *era;
 
        /*
-        * only put constraint if extra reg was actually
-        * allocated. Also takes care of event which do
-        * not use an extra shared reg
+        * Only put constraint if extra reg was actually allocated. Also takes
+        * care of event which do not use an extra shared reg.
+        *
+        * Also, if this is a fake cpuc we shouldn't touch any event state
+        * (reg->alloc) and we don't care about leaving inconsistent cpuc state
+        * either since it'll be thrown out.
         */
-       if (!reg->alloc)
+       if (!reg->alloc || cpuc->is_fake)
                return;
 
        era = &cpuc->shared_regs->regs[reg->idx];