There's a perf stat bug easy to observer on a machine with only one cgroup:
$ perf stat -e cycles -I 1000 -C 0 -G /
# time counts unit events
1.000161699 <not counted> cycles /
2.000355591 <not counted> cycles /
3.000565154 <not counted> cycles /
4.000951350 <not counted> cycles /
We'd expect some output there.
The underlying problem is that there is an optimization in
perf_cgroup_sched_{in,out}() that skips the switch of cgroup events
if the old and new cgroups in a task switch are the same.
This optimization interacts with the current code in two ways
that cause a CPU context's cgroup (cpuctx->cgrp) to be NULL even if a
cgroup event matches the current task. These are:
1. On creation of the first cgroup event in a CPU: In current code,
cpuctx->cpu is only set in perf_cgroup_sched_in, but due to the
aforesaid optimization, perf_cgroup_sched_in will run until the next
cgroup switches in that CPU. This may happen late or never happen,
depending on system's number of cgroups, CPU load, etc.
2. On deletion of the last cgroup event in a cpuctx: In list_del_event,
cpuctx->cgrp is set NULL. Any new cgroup event will not be sched in
because cpuctx->cgrp == NULL until a cgroup switch occurs and
perf_cgroup_sched_in is executed (updating cpuctx->cgrp).
This patch fixes both problems by setting cpuctx->cgrp in list_add_event,
mirroring what list_del_event does when removing a cgroup event from CPU
context, as introduced in:
commit 68cacd2916 ("perf_events: Fix stale ->cgrp pointer in update_cgrp_time_from_cpuctx()")
With this patch, cpuctx->cgrp is always set/clear when installing/removing
the first/last cgroup event in/from the CPU context. With cpuctx->cgrp
correctly set, event_filter_match works as intended when events are
sched in/out.
After the fix, the output is as expected:
$ perf stat -e cycles -I 1000 -a -G /
# time counts unit events
1.004699159 627342882 cycles /
2.007397156 615272690 cycles /
3.010019057 616726074 cycles /
Signed-off-by: David Carrillo-Cisneros <davidcc@google.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Jiri Olsa <jolsa@redhat.com>
Cc: Kan Liang <kan.liang@intel.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Paul Turner <pjt@google.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Stephane Eranian <eranian@google.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Vegard Nossum <vegard.nossum@gmail.com>
Cc: Vince Weaver <vincent.weaver@maine.edu>
Link: http://lkml.kernel.org/r/1470124092-113192-1-git-send-email-davidcc@google.com
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Pull smp hotplug updates from Thomas Gleixner:
"This is the next part of the hotplug rework.
- Convert all notifiers with a priority assigned
- Convert all CPU_STARTING/DYING notifiers
The final removal of the STARTING/DYING infrastructure will happen
when the merge window closes.
Another 700 hundred line of unpenetrable maze gone :)"
* 'smp-hotplug-for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip: (70 commits)
timers/core: Correct callback order during CPU hot plug
leds/trigger/cpu: Move from CPU_STARTING to ONLINE level
powerpc/numa: Convert to hotplug state machine
arm/perf: Fix hotplug state machine conversion
irqchip/armada: Avoid unused function warnings
ARC/time: Convert to hotplug state machine
clocksource/atlas7: Convert to hotplug state machine
clocksource/armada-370-xp: Convert to hotplug state machine
clocksource/exynos_mct: Convert to hotplug state machine
clocksource/arm_global_timer: Convert to hotplug state machine
rcu: Convert rcutree to hotplug state machine
KVM/arm/arm64/vgic-new: Convert to hotplug state machine
smp/cfd: Convert core to hotplug state machine
x86/x2apic: Convert to CPU hotplug state machine
profile: Convert to hotplug state machine
timers/core: Convert to hotplug state machine
hrtimer: Convert to hotplug state machine
x86/tboot: Convert to hotplug state machine
arm64/armv8 deprecated: Convert to hotplug state machine
hwtracing/coresight-etm4x: Convert to hotplug state machine
...
Pull networking updates from David Miller:
1) Unified UDP encapsulation offload methods for drivers, from
Alexander Duyck.
2) Make DSA binding more sane, from Andrew Lunn.
3) Support QCA9888 chips in ath10k, from Anilkumar Kolli.
4) Several workqueue usage cleanups, from Bhaktipriya Shridhar.
5) Add XDP (eXpress Data Path), essentially running BPF programs on RX
packets as soon as the device sees them, with the option to mirror
the packet on TX via the same interface. From Brenden Blanco and
others.
6) Allow qdisc/class stats dumps to run lockless, from Eric Dumazet.
7) Add VLAN support to b53 and bcm_sf2, from Florian Fainelli.
8) Simplify netlink conntrack entry layout, from Florian Westphal.
9) Add ipv4 forwarding support to mlxsw spectrum driver, from Ido
Schimmel, Yotam Gigi, and Jiri Pirko.
10) Add SKB array infrastructure and convert tun and macvtap over to it.
From Michael S Tsirkin and Jason Wang.
11) Support qdisc packet injection in pktgen, from John Fastabend.
12) Add neighbour monitoring framework to TIPC, from Jon Paul Maloy.
13) Add NV congestion control support to TCP, from Lawrence Brakmo.
14) Add GSO support to SCTP, from Marcelo Ricardo Leitner.
15) Allow GRO and RPS to function on macsec devices, from Paolo Abeni.
16) Support MPLS over IPV4, from Simon Horman.
* git://git.kernel.org/pub/scm/linux/kernel/git/davem/net-next: (1622 commits)
xgene: Fix build warning with ACPI disabled.
be2net: perform temperature query in adapter regardless of its interface state
l2tp: Correctly return -EBADF from pppol2tp_getname.
net/mlx5_core/health: Remove deprecated create_singlethread_workqueue
net: ipmr/ip6mr: update lastuse on entry change
macsec: ensure rx_sa is set when validation is disabled
tipc: dump monitor attributes
tipc: add a function to get the bearer name
tipc: get monitor threshold for the cluster
tipc: make cluster size threshold for monitoring configurable
tipc: introduce constants for tipc address validation
net: neigh: disallow transition to NUD_STALE if lladdr is unchanged in neigh_update()
MAINTAINERS: xgene: Add driver and documentation path
Documentation: dtb: xgene: Add MDIO node
dtb: xgene: Add MDIO node
drivers: net: xgene: ethtool: Use phy_ethtool_gset and sset
drivers: net: xgene: Use exported functions
drivers: net: xgene: Enable MDIO driver
drivers: net: xgene: Add backward compatibility
drivers: net: phy: xgene: Add MDIO driver
...
This patch fixes the __output_custom() routine we currently use with
bpf_skb_copy(). I missed that when len is larger than the size of the
current handle, we can issue multiple invocations of copy_func, and
__output_custom() advances destination but also source buffer by the
written amount of bytes. When we have __output_custom(), this is actually
wrong since in that case the source buffer points to a non-linear object,
in our case an skb, which the copy_func helper is supposed to walk.
Therefore, since this is non-linear we thus need to pass the offset into
the helper, so that copy_func can use it for extracting the data from
the source object.
Therefore, adjust the callback signatures properly and pass offset
into the skb_header_pointer() invoked from bpf_skb_copy() callback. The
__DEFINE_OUTPUT_COPY_BODY() is adjusted to accommodate for two things:
i) to pass in whether we should advance source buffer or not; this is
a compile-time constant condition, ii) to pass in the offset for
__output_custom(), which we do with help of __VA_ARGS__, so everything
can stay inlined as is currently. Both changes allow for adapting the
__output_* fast-path helpers w/o extra overhead.
Fixes: 555c8a8623 ("bpf: avoid stack copy and use skb ctx for event output")
Fixes: 7e3f977edd ("perf, events: add non-linear data support for raw records")
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch adds support for non-linear data on raw records. It
extends raw records to have one or multiple fragments that will
be written linearly into the ring slot, where each fragment can
optionally have a custom callback handler to walk and extract
complex, possibly non-linear data.
If a callback handler is provided for a fragment, then the new
__output_custom() will be used instead of __output_copy() for
the perf_output_sample() part. perf_prepare_sample() does all
the size calculation only once, so perf_output_sample() doesn't
need to redo the same work anymore, meaning real_size and padding
will be cached in the raw record. The raw record becomes 32 bytes
in size without holes; to not increase it further and to avoid
doing unnecessary recalculations in fast-path, we can reuse
next pointer of the last fragment, idea here is borrowed from
ZERO_OR_NULL_PTR(), which should keep the perf_output_sample()
path for PERF_SAMPLE_RAW minimal.
This facility is needed for BPF's event output helper as a first
user that will, in a follow-up, add an additional perf_raw_frag
to its perf_raw_record in order to be able to more efficiently
dump skb context after a linear head meta data related to it.
skbs can be non-linear and thus need a custom output function to
dump buffers. Currently, the skb data needs to be copied twice;
with the help of __output_custom() this work only needs to be
done once. Future users could be things like XDP/BPF programs
that work on different context though and would thus also have
a different callback function.
The few users of raw records are adapted to initialize their frag
data from the raw record itself, no change in behavior for them.
The code is based upon a PoC diff provided by Peter Zijlstra [1].
[1] http://thread.gmane.org/gmane.linux.network/421294
Suggested-by: Peter Zijlstra <peterz@infradead.org>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
Actually a nice symmetric startup/teardown pair which fits properly into
the state machine concept. In the long run we should be able to invoke
the startup callback for the boot CPU via the state machine and get
rid of the init function which invokes it on the boot CPU.
Note: This comes actually before the perf hardware callbacks. In the notifier
model the hardware callbacks have a higher priority than the core
callback. But that's solely for CPU offline so that hardware migration of
events happens before the core is notified about the outgoing CPU.
With the symetric state array model we have the following ordering:
UP: core -> hardware
DOWN: hardware -> core
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Anna-Maria Gleixner <anna-maria@linutronix.de>
Reviewed-by: Sebastian Siewior <bigeasy@linutronix.de>
Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: Arnaldo Carvalho de Melo <acme@kernel.org>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Jiri Olsa <jolsa@redhat.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Oleg Nesterov <oleg@redhat.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Rasmus Villemoes <linux@rasmusvillemoes.dk>
Cc: rt@linutronix.de
Link: http://lkml.kernel.org/r/20160713153333.587514098@linutronix.de
Signed-off-by: Ingo Molnar <mingo@kernel.org>
The perf_event_aux() function iterates all PMUs and all events in
their respective per-CPU contexts to find the events to deliver
side-band records to.
For example, the brk test case in lkp triggers many mmap() operations,
which, if we're also running perf, results in many perf_event_aux()
invocations.
If we enable uncore PMU support (even when uncore events are not used),
dozens of uncore PMUs will be iterated, which can significantly
decrease brk_test's throughput.
For example, the brk throughput:
without uncore PMUs: 2647573 ops_per_sec
with uncore PMUs: 1768444 ops_per_sec
... a 33% reduction.
To get at the per-CPU events that need side-band records, this patch
puts these events on a per-CPU list, this avoids iterating the PMUs
and any events that do not need side-band records.
Per task events are unchanged to avoid extra overhead on the context
switch paths.
Suggested-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Reported-by: Huang, Ying <ying.huang@linux.intel.com>
Signed-off-by: Kan Liang <kan.liang@intel.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Jiri Olsa <jolsa@redhat.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Stephane Eranian <eranian@google.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Vince Weaver <vincent.weaver@maine.edu>
Link: http://lkml.kernel.org/r/1458757477-3781-1-git-send-email-kan.liang@intel.com
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Pull perf updates from Ingo Molnar:
"Mostly tooling and PMU driver fixes, but also a number of late updates
such as the reworking of the call-chain size limiting logic to make
call-graph recording more robust, plus tooling side changes for the
new 'backwards ring-buffer' extension to the perf ring-buffer"
* 'perf-urgent-for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/tip: (34 commits)
perf record: Read from backward ring buffer
perf record: Rename variable to make code clear
perf record: Prevent reading invalid data in record__mmap_read
perf evlist: Add API to pause/resume
perf trace: Use the ptr->name beautifier as default for "filename" args
perf trace: Use the fd->name beautifier as default for "fd" args
perf report: Add srcline_from/to branch sort keys
perf evsel: Record fd into perf_mmap
perf evsel: Add overwrite attribute and check write_backward
perf tools: Set buildid dir under symfs when --symfs is provided
perf trace: Only auto set call-graph to "dwarf" when syscalls are being traced
perf annotate: Sort list of recognised instructions
perf annotate: Fix identification of ARM blt and bls instructions
perf tools: Fix usage of max_stack sysctl
perf callchain: Stop validating callchains by the max_stack sysctl
perf trace: Fix exit_group() formatting
perf top: Use machine->kptr_restrict_warned
perf trace: Warn when trying to resolve kernel addresses with kptr_restrict=1
perf machine: Do not bail out if not managing to read ref reloc symbol
perf/x86/intel/p4: Trival indentation fix, remove space
...
Pull networking updates from David Miller:
"Highlights:
1) Support SPI based w5100 devices, from Akinobu Mita.
2) Partial Segmentation Offload, from Alexander Duyck.
3) Add GMAC4 support to stmmac driver, from Alexandre TORGUE.
4) Allow cls_flower stats offload, from Amir Vadai.
5) Implement bpf blinding, from Daniel Borkmann.
6) Optimize _ASYNC_ bit twiddling on sockets, unless the socket is
actually using FASYNC these atomics are superfluous. From Eric
Dumazet.
7) Run TCP more preemptibly, also from Eric Dumazet.
8) Support LED blinking, EEPROM dumps, and rxvlan offloading in mlx5e
driver, from Gal Pressman.
9) Allow creating ppp devices via rtnetlink, from Guillaume Nault.
10) Improve BPF usage documentation, from Jesper Dangaard Brouer.
11) Support tunneling offloads in qed, from Manish Chopra.
12) aRFS offloading in mlx5e, from Maor Gottlieb.
13) Add RFS and RPS support to SCTP protocol, from Marcelo Ricardo
Leitner.
14) Add MSG_EOR support to TCP, this allows controlling packet
coalescing on application record boundaries for more accurate
socket timestamp sampling. From Martin KaFai Lau.
15) Fix alignment of 64-bit netlink attributes across the board, from
Nicolas Dichtel.
16) Per-vlan stats in bridging, from Nikolay Aleksandrov.
17) Several conversions of drivers to ethtool ksettings, from Philippe
Reynes.
18) Checksum neutral ILA in ipv6, from Tom Herbert.
19) Factorize all of the various marvell dsa drivers into one, from
Vivien Didelot
20) Add VF support to qed driver, from Yuval Mintz"
* git://git.kernel.org/pub/scm/linux/kernel/git/davem/net-next: (1649 commits)
Revert "phy dp83867: Fix compilation with CONFIG_OF_MDIO=m"
Revert "phy dp83867: Make rgmii parameters optional"
r8169: default to 64-bit DMA on recent PCIe chips
phy dp83867: Make rgmii parameters optional
phy dp83867: Fix compilation with CONFIG_OF_MDIO=m
bpf: arm64: remove callee-save registers use for tmp registers
asix: Fix offset calculation in asix_rx_fixup() causing slow transmissions
switchdev: pass pointer to fib_info instead of copy
net_sched: close another race condition in tcf_mirred_release()
tipc: fix nametable publication field in nl compat
drivers: net: Don't print unpopulated net_device name
qed: add support for dcbx.
ravb: Add missing free_irq() calls to ravb_close()
qed: Remove a stray tab
net: ethernet: fec-mpc52xx: use phy_ethtool_{get|set}_link_ksettings
net: ethernet: fec-mpc52xx: use phydev from struct net_device
bpf, doc: fix typo on bpf_asm descriptions
stmmac: hardware TX COE doesn't work when force_thresh_dma_mode is set
net: ethernet: fs-enet: use phy_ethtool_{get|set}_link_ksettings
net: ethernet: fs-enet: use phydev from struct net_device
...
We will use it to count how many addresses are in the entry->ip[] array,
excluding PERF_CONTEXT_{KERNEL,USER,etc} entries, so that we can really
return the number of entries specified by the user via the relevant
sysctl, kernel.perf_event_max_contexts, or via the per event
perf_event_attr.sample_max_stack knob.
This way we keep the perf_sample->ip_callchain->nr meaning, that is the
number of entries, be it real addresses or PERF_CONTEXT_ entries, while
honouring the max_stack knobs, i.e. the end result will be max_stack
entries if we have at least that many entries in a given stack trace.
Cc: David Ahern <dsahern@gmail.com>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: Jiri Olsa <jolsa@redhat.com>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Link: http://lkml.kernel.org/n/tip-s8teto51tdqvlfhefndtat9r@git.kernel.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Commit:
2665784850 ("perf/core: Verify we have a single perf_hw_context PMU")
forcefully prevents multiple PMUs from sharing perf_hw_context, as this
generally doesn't make sense. It is a common bug for uncore PMUs to
use perf_hw_context rather than perf_invalid_context, which this detects.
However, systems exist with heterogeneous CPUs (and hence heterogeneous
HW PMUs), for which sharing perf_hw_context is necessary, and possible
in some limited cases.
To make this work we have to perform some gymnastics, as we did in these
commits:
66eb579e66 ("perf: allow for PMU-specific event filtering")
c904e32a69 ("arm: perf: filter unschedulable events")
To allow those systems to work, we must allow PMUs for heterogeneous
CPUs to share perf_hw_context, though we must still disallow sharing
otherwise to detect the common misuse of perf_hw_context.
This patch adds a new PERF_PMU_CAP_HETEROGENEOUS_CPUS for this, updates
the core logic to account for this, and makes use of it in the arm_pmu
code that is used for systems with heterogeneous CPUs. Comments are
added to make the rationale clear and hopefully avoid accidental abuse.
Signed-off-by: Mark Rutland <mark.rutland@arm.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Jiri Olsa <jolsa@redhat.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Stephane Eranian <eranian@google.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Vince Weaver <vincent.weaver@maine.edu>
Cc: Will Deacon <will.deacon@arm.com>
Cc: linux-arm-kernel@lists.infradead.org
Link: http://lkml.kernel.org/r/20160426103346.GA20836@leverpostej
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Many instruction tracing PMUs out there support address range-based
filtering, which would, for example, generate trace data only for a
given range of instruction addresses, which is useful for tracing
individual functions, modules or libraries. Other PMUs may also
utilize this functionality to allow filtering to or filtering out
code at certain address ranges.
This patch introduces the interface for userspace to specify these
filters and for the PMU drivers to apply these filters to hardware
configuration.
The user interface is an ASCII string that is passed via an ioctl()
and specifies (in the form of an ASCII string) address ranges within
certain object files or within kernel. There is no special treatment
for kernel modules yet, but it might be a worthy pursuit.
The PMU driver interface basically adds two extra callbacks to the
PMU driver structure, one of which validates the filter configuration
proposed by the user against what the hardware is actually capable of
doing and the other one translates hardware-independent filter
configuration into something that can be programmed into the
hardware.
Signed-off-by: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Reviewed-by: Mathieu Poirier <mathieu.poirier@linaro.org>
Cc: Arnaldo Carvalho de Melo <acme@infradead.org>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Borislav Petkov <bp@alien8.de>
Cc: Jiri Olsa <jolsa@redhat.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Stephane Eranian <eranian@google.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Vince Weaver <vincent.weaver@maine.edu>
Cc: vince@deater.net
Link: http://lkml.kernel.org/r/1461771888-10409-6-git-send-email-alexander.shishkin@linux.intel.com
Signed-off-by: Ingo Molnar <mingo@kernel.org>
The default remains 127, which is good for most cases, and not even hit
most of the time, but then for some cases, as reported by Brendan, 1024+
deep frames are appearing on the radar for things like groovy, ruby.
And in some workloads putting a _lower_ cap on this may make sense. One
that is per event still needs to be put in place tho.
The new file is:
# cat /proc/sys/kernel/perf_event_max_stack
127
Chaging it:
# echo 256 > /proc/sys/kernel/perf_event_max_stack
# cat /proc/sys/kernel/perf_event_max_stack
256
But as soon as there is some event using callchains we get:
# echo 512 > /proc/sys/kernel/perf_event_max_stack
-bash: echo: write error: Device or resource busy
#
Because we only allocate the callchain percpu data structures when there
is a user, which allows for changing the max easily, its just a matter
of having no callchain users at that point.
Reported-and-Tested-by: Brendan Gregg <brendan.d.gregg@gmail.com>
Reviewed-by: Frederic Weisbecker <fweisbec@gmail.com>
Acked-by: Alexei Starovoitov <ast@kernel.org>
Acked-by: David Ahern <dsahern@gmail.com>
Cc: Adrian Hunter <adrian.hunter@intel.com>
Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: He Kuang <hekuang@huawei.com>
Cc: Jiri Olsa <jolsa@redhat.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Masami Hiramatsu <mhiramat@kernel.org>
Cc: Milian Wolff <milian.wolff@kdab.com>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Stephane Eranian <eranian@google.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Vince Weaver <vincent.weaver@maine.edu>
Cc: Wang Nan <wangnan0@huawei.com>
Cc: Zefan Li <lizefan@huawei.com>
Link: http://lkml.kernel.org/r/20160426002928.GB16708@kernel.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
This patch introduces 'write_backward' bit to perf_event_attr, which
controls the direction of a ring buffer. After set, the corresponding
ring buffer is written from end to beginning. This feature is design to
support reading from overwritable ring buffer.
Ring buffer can be created by mapping a perf event fd. Kernel puts event
records into ring buffer, user tooling like perf fetch them from
address returned by mmap(). To prevent racing between kernel and tooling,
they communicate to each other through 'head' and 'tail' pointers.
Kernel maintains 'head' pointer, points it to the next free area (tail
of the last record). Tooling maintains 'tail' pointer, points it to the
tail of last consumed record (record has already been fetched). Kernel
determines the available space in a ring buffer using these two
pointers to avoid overwrite unfetched records.
By mapping without 'PROT_WRITE', an overwritable ring buffer is created.
Different from normal ring buffer, tooling is unable to maintain 'tail'
pointer because writing is forbidden. Therefore, for this type of ring
buffers, kernel overwrite old records unconditionally, works like flight
recorder. This feature would be useful if reading from overwritable ring
buffer were as easy as reading from normal ring buffer. However,
there's an obscure problem.
The following figure demonstrates a full overwritable ring buffer. In
this figure, the 'head' pointer points to the end of last record, and a
long record 'E' is pending. For a normal ring buffer, a 'tail' pointer
would have pointed to position (X), so kernel knows there's no more
space in the ring buffer. However, for an overwritable ring buffer,
kernel ignore the 'tail' pointer.
(X) head
. |
. V
+------+-------+----------+------+---+
|A....A|B.....B|C........C|D....D| |
+------+-------+----------+------+---+
Record 'A' is overwritten by event 'E':
head
|
V
+--+---+-------+----------+------+---+
|.E|..A|B.....B|C........C|D....D|E..|
+--+---+-------+----------+------+---+
Now tooling decides to read from this ring buffer. However, none of these
two natural positions, 'head' and the start of this ring buffer, are
pointing to the head of a record. Even the full ring buffer can be
accessed by tooling, it is unable to find a position to start decoding.
The first attempt tries to solve this problem AFAIK can be found from
[1]. It makes kernel to maintain 'tail' pointer: updates it when ring
buffer is half full. However, this approach introduces overhead to
fast path. Test result shows a 1% overhead [2]. In addition, this method
utilizes no more tham 50% records.
Another attempt can be found from [3], which allows putting the size of
an event at the end of each record. This approach allows tooling to find
records in a backward manner from 'head' pointer by reading size of a
record from its tail. However, because of alignment requirement, it
needs 8 bytes to record the size of a record, which is a huge waste. Its
performance is also not good, because more data need to be written.
This approach also introduces some extra branch instructions to fast
path.
'write_backward' is a better solution to this problem.
Following figure demonstrates the state of the overwritable ring buffer
when 'write_backward' is set before overwriting:
head
|
V
+---+------+----------+-------+------+
| |D....D|C........C|B.....B|A....A|
+---+------+----------+-------+------+
and after overwriting:
head
|
V
+---+------+----------+-------+---+--+
|..E|D....D|C........C|B.....B|A..|E.|
+---+------+----------+-------+---+--+
In each situation, 'head' points to the beginning of the newest record.
From this record, tooling can iterate over the full ring buffer and fetch
records one by one.
The only limitation that needs to be considered is back-to-back reading.
Due to the non-deterministic of user programs, it is impossible to ensure
the ring buffer keeps stable during reading. Consider an extreme situation:
tooling is scheduled out after reading record 'D', then a burst of events
come, eat up the whole ring buffer (one or multiple rounds). When the
tooling process comes back, reading after 'D' is incorrect now.
To prevent this problem, we need to find a way to ensure the ring buffer
is stable during reading. ioctl(PERF_EVENT_IOC_PAUSE_OUTPUT) is
suggested because its overhead is lower than
ioctl(PERF_EVENT_IOC_ENABLE).
By carefully verifying 'header' pointer, reader can avoid pausing the
ring-buffer. For example:
/* A union of all possible events */
union perf_event event;
p = head = perf_mmap__read_head();
while (true) {
/* copy header of next event */
fetch(&event.header, p, sizeof(event.header));
/* read 'head' pointer */
head = perf_mmap__read_head();
/* check overwritten: is the header good? */
if (!verify(sizeof(event.header), p, head))
break;
/* copy the whole event */
fetch(&event, p, event.header.size);
/* read 'head' pointer again */
head = perf_mmap__read_head();
/* is the whole event good? */
if (!verify(event.header.size, p, head))
break;
p += event.header.size;
}
However, the overhead is high because:
a) In-place decoding is not safe.
Copying-verifying-decoding is required.
b) Fetching 'head' pointer requires additional synchronization.
(From Alexei Starovoitov:
Even when this trick works, pause is needed for more than stability of
reading. When we collect the events into overwrite buffer we're waiting
for some other trigger (like all cpu utilization spike or just one cpu
running and all others are idle) and when it happens the buffer has
valuable info from the past. At this point new events are no longer
interesting and buffer should be paused, events read and unpaused until
next trigger comes.)
This patch utilizes event's default overflow_handler introduced
previously. perf_event_output_backward() is created as the default
overflow handler for backward ring buffers. To avoid extra overhead to
fast path, original perf_event_output() becomes __perf_event_output()
and marked '__always_inline'. In theory, there's no extra overhead
introduced to fast path.
Performance testing:
Calling 3000000 times of 'close(-1)', use gettimeofday() to check
duration. Use 'perf record -o /dev/null -e raw_syscalls:*' to capture
system calls. In ns.
Testing environment:
CPU : Intel(R) Core(TM) i7-4790 CPU @ 3.60GHz
Kernel : v4.5.0
MEAN STDVAR
BASE 800214.950 2853.083
PRE1 2253846.700 9997.014
PRE2 2257495.540 8516.293
POST 2250896.100 8933.921
Where 'BASE' is pure performance without capturing. 'PRE1' is test
result of pure 'v4.5.0' kernel. 'PRE2' is test result before this
patch. 'POST' is test result after this patch. See [4] for the detailed
experimental setup.
Considering the stdvar, this patch doesn't introduce performance
overhead to the fast path.
[1] http://lkml.iu.edu/hypermail/linux/kernel/1304.1/04584.html
[2] http://lkml.iu.edu/hypermail/linux/kernel/1307.1/00535.html
[3] http://lkml.iu.edu/hypermail/linux/kernel/1512.0/01265.html
[4] http://lkml.kernel.org/g/56F89DCD.1040202@huawei.com
Signed-off-by: Wang Nan <wangnan0@huawei.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Acked-by: Alexei Starovoitov <ast@kernel.org>
Cc: <acme@kernel.org>
Cc: <pi3orama@163.com>
Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Brendan Gregg <brendan.d.gregg@gmail.com>
Cc: He Kuang <hekuang@huawei.com>
Cc: Jiri Olsa <jolsa@kernel.org>
Cc: Jiri Olsa <jolsa@redhat.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Masami Hiramatsu <masami.hiramatsu.pt@hitachi.com>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Stephane Eranian <eranian@google.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Vince Weaver <vincent.weaver@maine.edu>
Cc: Zefan Li <lizefan@huawei.com>
Link: http://lkml.kernel.org/r/1459865478-53413-1-git-send-email-wangnan0@huawei.com
[ Fixed the changelog some more. ]
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
split allows to move expensive update of 'struct trace_entry' to later phase.
Repurpose unused 1st argument of perf_tp_event() to indicate event type.
While splitting use temp variable 'rctx' instead of '*rctx' to avoid
unnecessary loads done by the compiler due to -fno-strict-aliasing
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
avoid memset in perf_fetch_caller_regs, since it's the critical path of all tracepoints.
It's called from perf_sw_event_sched, perf_event_task_sched_in and all of perf_trace_##call
with this_cpu_ptr(&__perf_regs[..]) which are zero initialized by perpcu init logic and
subsequent call to perf_arch_fetch_caller_regs initializes the same fields on all archs,
so we can safely drop memset from all of the above cases and move it into
perf_ftrace_function_call that calls it with stack allocated pt_regs.
Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
