From fa7315871046b9a4c48627905691dbde57e51033 Mon Sep 17 00:00:00 2001 From: Peter Zijlstra Date: Thu, 19 Sep 2013 10:16:42 +0200 Subject: perf: Fix capabilities bitfield compatibility in 'struct perf_event_mmap_page' Solve the problems around the broken definition of perf_event_mmap_page:: cap_usr_time and cap_usr_rdpmc fields which used to overlap, partially fixed by: 860f085b74e9 ("perf: Fix broken union in 'struct perf_event_mmap_page'") The problem with the fix (merged in v3.12-rc1 and not yet released officially), noticed by Vince Weaver is that the new behavior is not detectable by new user-space, and that due to the reuse of the field names it's easy to mis-compile a binary if old headers are used on a new kernel or new headers are used on an old kernel. To solve all that make this change explicit, detectable and self-contained, by iterating the ABI the following way: - Always clear bit 0, and rename it to usrpage->cap_bit0, to at least not confuse old user-space binaries. RDPMC will be marked as unavailable to old binaries but that's within the ABI, this is a capability bit. - Rename bit 1 to ->cap_bit0_is_deprecated and always set it to 1, so new libraries can reliably detect that bit 0 is deprecated and perma-zero without having to check the kernel version. - Use bits 2, 3, 4 for the newly defined, correct functionality: cap_user_rdpmc : 1, /* The RDPMC instruction can be used to read counts */ cap_user_time : 1, /* The time_* fields are used */ cap_user_time_zero : 1, /* The time_zero field is used */ - Rename all the bitfield names in perf_event.h to be different from the old names, to make sure it's not possible to mis-compile it accidentally with old assumptions. The 'size' field can then be used in the future to add new fields and it will act as a natural ABI version indicator as well. Also adjust tools/perf/ userspace for the new definitions, noticed by Adrian Hunter. Reported-by: Vince Weaver Signed-off-by: Peter Zijlstra Also-Fixed-by: Adrian Hunter Link: http://lkml.kernel.org/n/tip-zr03yxjrpXesOzzupszqglbv@git.kernel.org Signed-off-by: Ingo Molnar --- arch/x86/kernel/cpu/perf_event.c | 10 +++++----- 1 file changed, 5 insertions(+), 5 deletions(-) (limited to 'arch/x86/kernel/cpu/perf_event.c') diff --git a/arch/x86/kernel/cpu/perf_event.c b/arch/x86/kernel/cpu/perf_event.c index 8355c84b9729..a9c606bb4945 100644 --- a/arch/x86/kernel/cpu/perf_event.c +++ b/arch/x86/kernel/cpu/perf_event.c @@ -1883,9 +1883,9 @@ static struct pmu pmu = { void arch_perf_update_userpage(struct perf_event_mmap_page *userpg, u64 now) { - userpg->cap_usr_time = 0; - userpg->cap_usr_time_zero = 0; - userpg->cap_usr_rdpmc = x86_pmu.attr_rdpmc; + userpg->cap_user_time = 0; + userpg->cap_user_time_zero = 0; + userpg->cap_user_rdpmc = x86_pmu.attr_rdpmc; userpg->pmc_width = x86_pmu.cntval_bits; if (!boot_cpu_has(X86_FEATURE_CONSTANT_TSC)) @@ -1894,13 +1894,13 @@ void arch_perf_update_userpage(struct perf_event_mmap_page *userpg, u64 now) if (!boot_cpu_has(X86_FEATURE_NONSTOP_TSC)) return; - userpg->cap_usr_time = 1; + userpg->cap_user_time = 1; userpg->time_mult = this_cpu_read(cyc2ns); userpg->time_shift = CYC2NS_SCALE_FACTOR; userpg->time_offset = this_cpu_read(cyc2ns_offset) - now; if (sched_clock_stable && !check_tsc_disabled()) { - userpg->cap_usr_time_zero = 1; + userpg->cap_user_time_zero = 1; userpg->time_zero = this_cpu_read(cyc2ns_offset); } } -- cgit v1.2.3 From 8a3da6c7d0031fcb6a0d17f9c7a68b0e01f52855 Mon Sep 17 00:00:00 2001 From: Ingo Molnar Date: Sat, 28 Sep 2013 15:48:48 +0200 Subject: perf/x86: Fix PMU detection printout when no PMU is detected Ran into this cryptic PMU bootup log recently: [ 0.124047] Performance Events: [ 0.125000] smpboot: ... Turns out we print this if no PMU is detected. Fall back to the right condition so that the following is printed: [ 0.122381] Performance Events: no PMU driver, software events only. Cc: Peter Zijlstra Cc: Arnaldo Carvalho de Melo Link: http://lkml.kernel.org/n/tip-u2fwaUffakjp0qkpRfqljgsn@git.kernel.org Signed-off-by: Ingo Molnar --- arch/x86/kernel/cpu/perf_event.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) (limited to 'arch/x86/kernel/cpu/perf_event.c') diff --git a/arch/x86/kernel/cpu/perf_event.c b/arch/x86/kernel/cpu/perf_event.c index a9c606bb4945..897783b3302a 100644 --- a/arch/x86/kernel/cpu/perf_event.c +++ b/arch/x86/kernel/cpu/perf_event.c @@ -1506,7 +1506,7 @@ static int __init init_hw_perf_events(void) err = amd_pmu_init(); break; default: - return 0; + err = -ENOTSUPP; } if (err != 0) { pr_cont("no PMU driver, software events only.\n"); -- cgit v1.2.3 From d8b11a0cbd1c66ce283eb9dabe0498dfa6483f32 Mon Sep 17 00:00:00 2001 From: Peter Zijlstra Date: Thu, 3 Oct 2013 16:00:14 +0200 Subject: perf/x86: Clean up cap_user_time* setting Currently the cap_user_time_zero capability has different tests than cap_user_time; even though they expose the exact same data. Switch from CONSTANT && NONSTOP to sched_clock_stable to also deal with multi cabinet machines and drop the tsc_disabled() check.. non of this will work sanely without tsc anyway. Signed-off-by: Peter Zijlstra Link: http://lkml.kernel.org/n/tip-nmgn0j0muo1r4c94vlfh23xy@git.kernel.org Signed-off-by: Ingo Molnar --- arch/x86/kernel/cpu/perf_event.c | 11 +++-------- 1 file changed, 3 insertions(+), 8 deletions(-) (limited to 'arch/x86/kernel/cpu/perf_event.c') diff --git a/arch/x86/kernel/cpu/perf_event.c b/arch/x86/kernel/cpu/perf_event.c index 897783b3302a..9d8449158cf9 100644 --- a/arch/x86/kernel/cpu/perf_event.c +++ b/arch/x86/kernel/cpu/perf_event.c @@ -1888,10 +1888,7 @@ void arch_perf_update_userpage(struct perf_event_mmap_page *userpg, u64 now) userpg->cap_user_rdpmc = x86_pmu.attr_rdpmc; userpg->pmc_width = x86_pmu.cntval_bits; - if (!boot_cpu_has(X86_FEATURE_CONSTANT_TSC)) - return; - - if (!boot_cpu_has(X86_FEATURE_NONSTOP_TSC)) + if (!sched_clock_stable) return; userpg->cap_user_time = 1; @@ -1899,10 +1896,8 @@ void arch_perf_update_userpage(struct perf_event_mmap_page *userpg, u64 now) userpg->time_shift = CYC2NS_SCALE_FACTOR; userpg->time_offset = this_cpu_read(cyc2ns_offset) - now; - if (sched_clock_stable && !check_tsc_disabled()) { - userpg->cap_user_time_zero = 1; - userpg->time_zero = this_cpu_read(cyc2ns_offset); - } + userpg->cap_user_time_zero = 1; + userpg->time_zero = this_cpu_read(cyc2ns_offset); } /* -- cgit v1.2.3 From e8a923cc1fff6e627f906655ad52ee694ef2f6d7 Mon Sep 17 00:00:00 2001 From: Peter Zijlstra Date: Thu, 17 Oct 2013 15:32:10 +0200 Subject: 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(-) (limited to 'arch/x86/kernel/cpu/perf_event.c') 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) -- cgit v1.2.3