From bf378d341e4873ed928dc3c636252e6895a21f50 Mon Sep 17 00:00:00 2001 From: Peter Zijlstra Date: Mon, 28 Oct 2013 13:55:29 +0100 Subject: [PATCH 1/2] perf: Fix perf ring buffer memory ordering The PPC64 people noticed a missing memory barrier and crufty old comments in the perf ring buffer code. So update all the comments and add the missing barrier. When the architecture implements local_t using atomic_long_t there will be double barriers issued; but short of introducing more conditional barrier primitives this is the best we can do. Reported-by: Victor Kaplansky Tested-by: Victor Kaplansky Signed-off-by: Peter Zijlstra Cc: Mathieu Desnoyers Cc: michael@ellerman.id.au Cc: Paul McKenney Cc: Michael Neuling Cc: Frederic Weisbecker Cc: anton@samba.org Cc: benh@kernel.crashing.org Link: http://lkml.kernel.org/r/20131025173749.GG19466@laptop.lan Signed-off-by: Ingo Molnar --- include/uapi/linux/perf_event.h | 12 +++++++----- kernel/events/ring_buffer.c | 31 +++++++++++++++++++++++++++---- 2 files changed, 34 insertions(+), 9 deletions(-) diff --git a/include/uapi/linux/perf_event.h b/include/uapi/linux/perf_event.h index 009a655a5d35..2fc1602e23bb 100644 --- a/include/uapi/linux/perf_event.h +++ b/include/uapi/linux/perf_event.h @@ -456,13 +456,15 @@ struct perf_event_mmap_page { /* * Control data for the mmap() data buffer. * - * User-space reading the @data_head value should issue an rmb(), on - * SMP capable platforms, after reading this value -- see - * perf_event_wakeup(). + * User-space reading the @data_head value should issue an smp_rmb(), + * after reading this value. * * When the mapping is PROT_WRITE the @data_tail value should be - * written by userspace to reflect the last read data. In this case - * the kernel will not over-write unread data. + * written by userspace to reflect the last read data, after issueing + * an smp_mb() to separate the data read from the ->data_tail store. + * In this case the kernel will not over-write unread data. + * + * See perf_output_put_handle() for the data ordering. */ __u64 data_head; /* head in the data section */ __u64 data_tail; /* user-space written tail */ diff --git a/kernel/events/ring_buffer.c b/kernel/events/ring_buffer.c index cd55144270b5..9c2ddfbf4525 100644 --- a/kernel/events/ring_buffer.c +++ b/kernel/events/ring_buffer.c @@ -87,10 +87,31 @@ static void perf_output_put_handle(struct perf_output_handle *handle) goto out; /* - * Publish the known good head. Rely on the full barrier implied - * by atomic_dec_and_test() order the rb->head read and this - * write. + * Since the mmap() consumer (userspace) can run on a different CPU: + * + * kernel user + * + * READ ->data_tail READ ->data_head + * smp_mb() (A) smp_rmb() (C) + * WRITE $data READ $data + * smp_wmb() (B) smp_mb() (D) + * STORE ->data_head WRITE ->data_tail + * + * Where A pairs with D, and B pairs with C. + * + * I don't think A needs to be a full barrier because we won't in fact + * write data until we see the store from userspace. So we simply don't + * issue the data WRITE until we observe it. Be conservative for now. + * + * OTOH, D needs to be a full barrier since it separates the data READ + * from the tail WRITE. + * + * For B a WMB is sufficient since it separates two WRITEs, and for C + * an RMB is sufficient since it separates two READs. + * + * See perf_output_begin(). */ + smp_wmb(); rb->user_page->data_head = head; /* @@ -154,9 +175,11 @@ int perf_output_begin(struct perf_output_handle *handle, * Userspace could choose to issue a mb() before updating the * tail pointer. So that all reads will be completed before the * write is issued. + * + * See perf_output_put_handle(). */ tail = ACCESS_ONCE(rb->user_page->data_tail); - smp_rmb(); + smp_mb(); offset = head = local_read(&rb->head); head += size; if (unlikely(!perf_output_space(rb, tail, offset, head))) From e8a923cc1fff6e627f906655ad52ee694ef2f6d7 Mon Sep 17 00:00:00 2001 From: Peter Zijlstra Date: Thu, 17 Oct 2013 15:32:10 +0200 Subject: [PATCH 2/2] perf/x86: Fix NMI measurements OK, so what I'm actually seeing on my WSM is that sched/clock.c is 'broken' for the purpose we're using it for. What triggered it is that my WSM-EP is broken :-( [ 0.001000] tsc: Fast TSC calibration using PIT [ 0.002000] tsc: Detected 2533.715 MHz processor [ 0.500180] TSC synchronization [CPU#0 -> CPU#6]: [ 0.505197] Measured 3 cycles TSC warp between CPUs, turning off TSC clock. [ 0.004000] tsc: Marking TSC unstable due to check_tsc_sync_source failed For some reason it consistently detects TSC skew, even though NHM+ should have a single clock domain for 'reasonable' systems. This marks sched_clock_stable=0, which means that we do fancy stuff to try and get a 'sane' clock. Part of this fancy stuff relies on the tick, clearly that's gone when NOHZ=y. So for idle cpus time gets stuck, until it either wakes up or gets kicked by another cpu. While this is perfectly fine for the scheduler -- it only cares about actually running stuff, and when we're running stuff we're obviously not idle. This does somewhat break down for perf which can trigger events just fine on an otherwise idle cpu. So I've got NMIs get get 'measured' as taking ~1ms, which actually don't last nearly that long: -0 [013] d.h. 886.311970: rcu_nmi_enter <-do_nmi ... -0 [013] d.h. 886.311997: perf_sample_event_took: HERE!!! : 1040990 So ftrace (which uses sched_clock(), not the fancy bits) only sees ~27us, but we measure ~1ms !! Now since all this measurement stuff lives in x86 code, we can actually fix it. Signed-off-by: Peter Zijlstra Cc: mingo@kernel.org Cc: dave.hansen@linux.intel.com Cc: eranian@google.com Cc: Don Zickus Cc: jmario@redhat.com Cc: acme@infradead.org Link: http://lkml.kernel.org/r/20131017133350.GG3364@laptop.programming.kicks-ass.net Signed-off-by: Ingo Molnar --- arch/x86/kernel/cpu/perf_event.c | 6 +++--- arch/x86/kernel/nmi.c | 4 ++-- 2 files changed, 5 insertions(+), 5 deletions(-) diff --git a/arch/x86/kernel/cpu/perf_event.c b/arch/x86/kernel/cpu/perf_event.c index 9d8449158cf9..8a87a3224121 100644 --- a/arch/x86/kernel/cpu/perf_event.c +++ b/arch/x86/kernel/cpu/perf_event.c @@ -1276,16 +1276,16 @@ void perf_events_lapic_init(void) static int __kprobes perf_event_nmi_handler(unsigned int cmd, struct pt_regs *regs) { - int ret; u64 start_clock; u64 finish_clock; + int ret; if (!atomic_read(&active_events)) return NMI_DONE; - start_clock = local_clock(); + start_clock = sched_clock(); ret = x86_pmu.handle_irq(regs); - finish_clock = local_clock(); + finish_clock = sched_clock(); perf_sample_event_took(finish_clock - start_clock); diff --git a/arch/x86/kernel/nmi.c b/arch/x86/kernel/nmi.c index ba77ebc2c353..6fcb49ce50a1 100644 --- a/arch/x86/kernel/nmi.c +++ b/arch/x86/kernel/nmi.c @@ -113,10 +113,10 @@ static int __kprobes nmi_handle(unsigned int type, struct pt_regs *regs, bool b2 u64 before, delta, whole_msecs; int remainder_ns, decimal_msecs, thishandled; - before = local_clock(); + before = sched_clock(); thishandled = a->handler(type, regs); handled += thishandled; - delta = local_clock() - before; + delta = sched_clock() - before; trace_nmi_handler(a->handler, (int)delta, thishandled); if (delta < nmi_longest_ns)