user/sven/linux.git/include/linux/perf_event.h, branch v5.10.124

perf: Fix perf_event_read_local() time

2022-02-01T16:25:45Z

[ Upstream commit 09f5e7dc7ad705289e1b1ec065439aa3c42951c4 ] Time readers that cannot take locks (due to NMI etc..) currently make use of perf_event::shadow_ctx_time, which, for that event gives: time' = now + (time - timestamp) or, alternatively arranged: time' = time + (now - timestamp) IOW, the progression of time since the last time the shadow_ctx_time was updated. There's problems with this: A) the shadow_ctx_time is per-event, even though the ctx_time it reflects is obviously per context. The direct concequence of this is that the context needs to iterate all events all the time to keep the shadow_ctx_time in sync. B) even with the prior point, the context itself might not be active meaning its time should not advance to begin with. C) shadow_ctx_time isn't consistently updated when ctx_time is There are 3 users of this stuff, that suffer differently from this: - calc_timer_values() - perf_output_read() - perf_event_update_userpage() /* A */ - perf_event_read_local() /* A,B */ In particular, perf_output_read() doesn't suffer at all, because it's sample driven and hence only relevant when the event is actually running. This same was supposed to be true for perf_event_update_userpage(), after all self-monitoring implies the context is active *HOWEVER*, as per commit f79256532682 ("perf/core: fix userpage->time_enabled of inactive events") this goes wrong when combined with counter overcommit, in that case those events that do not get scheduled when the context becomes active (task events typically) miss out on the EVENT_TIME update and ENABLED time is inflated (for a little while) with the time the context was inactive. Once the event gets rotated in, this gets corrected, leading to a non-monotonic timeflow. perf_event_read_local() made things even worse, it can request time at any point, suffering all the problems perf_event_update_userpage() does and more. Because while perf_event_update_userpage() is limited by the context being active, perf_event_read_local() users have no such constraint. Therefore, completely overhaul things and do away with perf_event::shadow_ctx_time. Instead have regular context time updates keep track of this offset directly and provide perf_event_time_now() to complement perf_event_time(). perf_event_time_now() will, in adition to being context wide, also take into account if the context is active. For inactive context, it will not advance time. This latter property means the cgroup perf_cgroup_info context needs to grow addition state to track this. Additionally, since all this is strictly per-cpu, we can use barrier() to order context activity vs context time. Fixes: 7d9285e82db5 ("perf/bpf: Extend the perf_event_read_local() interface, a.k.a. "bpf: perf event change needed for subsequent bpf helpers"") Signed-off-by: Peter Zijlstra (Intel) Tested-by: Song Liu Tested-by: Namhyung Kim Link: https://lkml.kernel.org/r/YcB06DasOBtU0b00@hirez.programming.kicks-ass.net Signed-off-by: Sasha Levin

perf: Protect perf_guest_cbs with RCU

2022-01-20T08:17:50Z

commit ff083a2d972f56bebfd82409ca62e5dfce950961 upstream. Protect perf_guest_cbs with RCU to fix multiple possible errors. Luckily, all paths that read perf_guest_cbs already require RCU protection, e.g. to protect the callback chains, so only the direct perf_guest_cbs touchpoints need to be modified. Bug #1 is a simple lack of WRITE_ONCE/READ_ONCE behavior to ensure perf_guest_cbs isn't reloaded between a !NULL check and a dereference. Fixed via the READ_ONCE() in rcu_dereference(). Bug #2 is that on weakly-ordered architectures, updates to the callbacks themselves are not guaranteed to be visible before the pointer is made visible to readers. Fixed by the smp_store_release() in rcu_assign_pointer() when the new pointer is non-NULL. Bug #3 is that, because the callbacks are global, it's possible for readers to run in parallel with an unregisters, and thus a module implementing the callbacks can be unloaded while readers are in flight, resulting in a use-after-free. Fixed by a synchronize_rcu() call when unregistering callbacks. Bug #1 escaped notice because it's extremely unlikely a compiler will reload perf_guest_cbs in this sequence. perf_guest_cbs does get reloaded for future derefs, e.g. for ->is_user_mode(), but the ->is_in_guest() guard all but guarantees the consumer will win the race, e.g. to nullify perf_guest_cbs, KVM has to completely exit the guest and teardown down all VMs before KVM start its module unload / unregister sequence. This also makes it all but impossible to encounter bug #3. Bug #2 has not been a problem because all architectures that register callbacks are strongly ordered and/or have a static set of callbacks. But with help, unloading kvm_intel can trigger bug #1 e.g. wrapping perf_guest_cbs with READ_ONCE in perf_misc_flags() while spamming kvm_intel module load/unload leads to: BUG: kernel NULL pointer dereference, address: 0000000000000000 #PF: supervisor read access in kernel mode #PF: error_code(0x0000) - not-present page PGD 0 P4D 0 Oops: 0000 [#1] PREEMPT SMP CPU: 6 PID: 1825 Comm: stress Not tainted 5.14.0-rc2+ #459 Hardware name: QEMU Standard PC (Q35 + ICH9, 2009), BIOS 0.0.0 02/06/2015 RIP: 0010:perf_misc_flags+0x1c/0x70 Call Trace: perf_prepare_sample+0x53/0x6b0 perf_event_output_forward+0x67/0x160 __perf_event_overflow+0x52/0xf0 handle_pmi_common+0x207/0x300 intel_pmu_handle_irq+0xcf/0x410 perf_event_nmi_handler+0x28/0x50 nmi_handle+0xc7/0x260 default_do_nmi+0x6b/0x170 exc_nmi+0x103/0x130 asm_exc_nmi+0x76/0xbf Fixes: 39447b386c84 ("perf: Enhance perf to allow for guest statistic collection from host") Signed-off-by: Sean Christopherson Signed-off-by: Peter Zijlstra (Intel) Reviewed-by: Paolo Bonzini Cc: stable@vger.kernel.org Link: https://lore.kernel.org/r/20211111020738.2512932-2-seanjc@google.com Signed-off-by: Greg Kroah-Hartman

Revert "perf: Rework perf_event_exit_event()"

2021-11-26T09:39:22Z

This reverts commit 94902ee2996a7f71471138093495df452dab87b6 which is upstream commit ef54c1a476aef7eef26fe13ea10dc090952c00f8. Reverting for now due to issues that need to get fixed upstream. Signed-off-by: Sasha Levin

perf/core: fix userpage->time_enabled of inactive events

2021-10-17T08:43:33Z

[ Upstream commit f792565326825ed806626da50c6f9a928f1079c1 ] Users of rdpmc rely on the mmapped user page to calculate accurate time_enabled. Currently, userpage->time_enabled is only updated when the event is added to the pmu. As a result, inactive event (due to counter multiplexing) does not have accurate userpage->time_enabled. This can be reproduced with something like: /* open 20 task perf_event "cycles", to create multiplexing */ fd = perf_event_open(); /* open task perf_event "cycles" */ userpage = mmap(fd); /* use mmap and rdmpc */ while (true) { time_enabled_mmap = xxx; /* use logic in perf_event_mmap_page */ time_enabled_read = read(fd).time_enabled; if (time_enabled_mmap > time_enabled_read) BUG(); } Fix this by updating userpage for inactive events in merge_sched_in. Suggested-by: Peter Zijlstra (Intel) Reported-and-tested-by: Lucian Grijincu Signed-off-by: Song Liu Signed-off-by: Peter Zijlstra (Intel) Link: https://lkml.kernel.org/r/20210929194313.2398474-1-songliubraving@fb.com Signed-off-by: Sasha Levin

perf: Rework perf_event_exit_event()

2021-05-11T12:47:31Z

[ Upstream commit ef54c1a476aef7eef26fe13ea10dc090952c00f8 ] Make perf_event_exit_event() more robust, such that we can use it from other contexts. Specifically the up and coming remove_on_exec. For this to work we need to address a few issues. Remove_on_exec will not destroy the entire context, so we cannot rely on TASK_TOMBSTONE to disable event_function_call() and we thus have to use perf_remove_from_context(). When using perf_remove_from_context(), there's two races to consider. The first is against close(), where we can have concurrent tear-down of the event. The second is against child_list iteration, which should not find a half baked event. To address this, teach perf_remove_from_context() to special case !ctx->is_active and about DETACH_CHILD. [ elver@google.com: fix racing parent/child exit in sync_child_event(). ] Signed-off-by: Marco Elver Signed-off-by: Peter Zijlstra (Intel) Link: https://lkml.kernel.org/r/20210408103605.1676875-2-elver@google.com Signed-off-by: Sasha Levin

perf/core: Flush PMU internal buffers for per-CPU events

2021-03-17T16:06:34Z

[ Upstream commit a5398bffc01fe044848c5024e5e867e407f239b8 ] Sometimes the PMU internal buffers have to be flushed for per-CPU events during a context switch, e.g., large PEBS. Otherwise, the perf tool may report samples in locations that do not belong to the process where the samples are processed in, because PEBS does not tag samples with PID/TID. The current code only flush the buffers for a per-task event. It doesn't check a per-CPU event. Add a new event state flag, PERF_ATTACH_SCHED_CB, to indicate that the PMU internal buffers have to be flushed for this event during a context switch. Add sched_cb_entry and perf_sched_cb_usages back to track the PMU/cpuctx which is required to be flushed. Only need to invoke the sched_task() for per-CPU events in this patch. The per-task events have been handled in perf_event_context_sched_in/out already. Fixes: 9c964efa4330 ("perf/x86/intel: Drain the PEBS buffer during context switches") Reported-by: Gabriel Marin Originally-by: Namhyung Kim Signed-off-by: Kan Liang Signed-off-by: Peter Zijlstra (Intel) Signed-off-by: Ingo Molnar Link: https://lkml.kernel.org/r/20201130193842.10569-1-kan.liang@linux.intel.com Signed-off-by: Sasha Levin

perf/arch: Remove perf_sample_data::regs_user_copy

2020-11-09T17:12:34Z

struct perf_sample_data lives on-stack, we should be careful about it's size. Furthermore, the pt_regs copy in there is only because x86_64 is a trainwreck, solve it differently. Reported-by: Thomas Gleixner Signed-off-by: Peter Zijlstra (Intel) Tested-by: Steven Rostedt Link: https://lkml.kernel.org/r/20201030151955.258178461@infradead.org

perf: Reduce stack usage of perf_output_begin()

2020-11-09T17:12:33Z

__perf_output_begin() has an on-stack struct perf_sample_data in the unlikely case it needs to generate a LOST record. However, every call to perf_output_begin() must already have a perf_sample_data on-stack. Reported-by: Thomas Gleixner Signed-off-by: Peter Zijlstra (Intel) Link: https://lkml.kernel.org/r/20201030151954.985416146@infradead.org

perf/core: Pull pmu::sched_task() into perf_event_context_sched_out()

2020-09-10T09:19:34Z

The pmu::sched_task() is a context switch callback. It passes the cpuctx->task_ctx as a parameter to the lower code. To find the cpuctx->task_ctx, the current code iterates a cpuctx list. The same context will iterated in perf_event_context_sched_out() soon. Share the cpuctx->task_ctx can avoid the unnecessary iteration of the cpuctx list. The pmu::sched_task() is also required for the optimization case for equivalent contexts. The task_ctx_sched_out() will eventually disable and reenable the PMU when schedule out events. Add perf_pmu_disable() and perf_pmu_enable() around task_ctx_sched_out() don't break anything. Drop the cpuctx->ctx.lock for the pmu::sched_task(). The lock is for per-CPU context, which is not necessary for the per-task context schedule. No one uses sched_cb_entry, perf_sched_cb_usages, sched_cb_list, and perf_pmu_sched_task() any more. Suggested-by: Peter Zijlstra (Intel) Signed-off-by: Kan Liang Signed-off-by: Peter Zijlstra (Intel) Link: https://lkml.kernel.org/r/20200821195754.20159-2-kan.liang@linux.intel.com

perf/x86/intel: Support per-thread RDPMC TopDown metrics

2020-08-18T14:34:37Z

Starts from Ice Lake, the TopDown metrics are directly available as fixed counters and do not require generic counters. Also, the TopDown metrics can be collected per thread. Extend the RDPMC usage to support per-thread TopDown metrics. The RDPMC index of the PERF_METRICS will be output if RDPMC users ask for the RDPMC index of the metrics events. To support per thread RDPMC TopDown, the metrics and slots counters have to be saved/restored during the context switching. The last_period and period_left are not used in the counting mode. Use the fields for saved_metric and saved_slots. Signed-off-by: Kan Liang Signed-off-by: Peter Zijlstra (Intel) Link: https://lkml.kernel.org/r/20200723171117.9918-12-kan.liang@linux.intel.com