Using perf with call graph method dwarf fails to provide backtrace
support for stripped binary even though .gnu_debuglink points to *.dbg
flavor with properly populated debug symbols.
Problem is reproduced on ARM (v7, v8), kernels 3.14.y, 4.4.y and
4.10.rc3. Perf is configured with libunwind, and unwind dwarf support
[1]. Test code (stress_bt.c) can be found on [2].
Running (explicitly disable other unwinding methods):
$ gcc -g -o stress_bt -fomit-frame-pointer -fno-unwind-tables \
-fno-asynchronous-unwind-tables stress_bt.c
$ perf record -N --call-graph dwarf ./stress_bt
$ perf report
results in properly generated call graph. Stripping the binary and running
it results with missing call graph. Expected result is to have call graph:
$ gcc -g -o stress_bt -fomit-frame-pointer -fno-unwind-tables \
-fno-asynchronous-unwind-tables stress_bt.c
$ objcopy --only-keep-debug stress_bt stress_bt.dbg
$ objcopy --strip-debug stress_bt
$ objcopy --add-gnu-debuglink=stress_bt.dbg stress_bt
$ perf record -N --call-graph dwarf ./stress_bt
$ perf report
Problem is that perf doesn't try to read symbols pointed by gnu
debuglink. Patch adds checking, and reading of the symbols from
debuglink and symsrc. Order of the check is to first check within dso,
then check whether symsrc is defined and try to read from it. Finally,
debuglink is checked. Default locations of debug files are discussed in
[3] and [4]. Comments on RFC are on [5].
[1] https://wiki.linaro.org/LEG/Engineering/TOOLS/perf-callstack-unwinding
[2] [1]#Backtrace_stress_application
[3] https://sourceware.org/gdb/onlinedocs/gdb/Separate-Debug-Files.html
[4] https://sourceware.org/binutils/docs/binutils/objcopy.html
[5] https://lkml.org/lkml/2016/8/22/473
Signed-off-by: Matija Glavinic Pecotic <matija.glavinic-pecotic.ext@nokia.com>
Acked-by: Jiri Olsa <jolsa@kernel.org>
Cc: Alexander Sverdlin <alexander.sverdlin@nokia.com>
Cc: Masami Hiramatsu <mhiramat@kernel.org>
Cc: Namhyung Kim <namhyung@kernel.org>
Link: http://lkml.kernel.org/r/d309d40a-463f-482b-68e1-1465326efdc1@nokia.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Pull perf/core improvements and fixes from Arnaldo Carvalho de Melo:
New features:
- Account thread wait time (off CPU time) separately: sleep, iowait and
preempt, based on the prev_state of the last event, show the breakdown
when using "perf sched timehist --state" (Namhyumg Kim)
Infrastructure changes:
- Factor out PMU scale conversion code (Andi Kleen)
- Remove unnecessary feature-dwarf warning (David Carrillo-Cisneros)
- Add missing member name in OPT_() macros (Soramichi AKIYAMA)
- Move variables referenced in libperf.a object files from perf's main()
file, so that other tools can use libperf.a with a different main()
(Soramichi AKIYAMA)
Documentation changes:
- Fix 'perf script' man page about --dump-raw-trace option (Michael Petlan)
- Also allow forcing reading of non-root owned files by root in 'perf
script' (Yannick Brosseau)
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
The use_browser and perf_version_string variables are both declared in
perf.c but they are also referenced by other functions of libperf.a.
Therefore a user linking an own main() with libperf.a must declare those
two variables in their files even if the files never use the browser or
the version information.
This patch fixes this issue by moving use_browser and
perf_version_string out of perf.c to some other files.
Signed-off-by: Soramichi Akiyama <akiyama@m.soramichi.jp>
Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Link: http://lkml.kernel.org/r/20170117002237.c1aec0ce3b4d675dca018deb@m.soramichi.jp
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
The --state option is to show task state when switched out. The state
is printed as a single character like in the /proc but I added 'I' for
idle state rather than 'R'.
$ perf sched timehist --state | head
Samples do not have callchains.
time cpu task name wait time sch delay run time state
[tid/pid] (msec) (msec) (msec)
-------- --- ----------------------- -------- ------------------ -----
1.753791 [3] <idle> 0.000 0.000 0.000 I
1.753834 [1] perf[27469] 0.000 0.000 0.000 S
1.753904 [3] perf[27470] 0.000 0.006 0.112 S
1.753914 [1] <idle> 0.000 0.000 0.079 I
1.753915 [3] migration/3[23] 0.000 0.002 0.011 S
1.754287 [2] <idle> 0.000 0.000 0.000 I
1.754335 [2] transmission[1773/1739] 0.000 0.004 0.047 S
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
Tested-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: David Ahern <dsahern@gmail.com>
Cc: Jiri Olsa <jolsa@kernel.org>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Link: http://lkml.kernel.org/r/20170113104523.31212-2-namhyung@kernel.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Improve __kernel_text_address()/kernel_text_address() to return
true if the given address is on a kprobe's instruction slot
trampoline.
This can help stacktraces to determine the address is on a
text area or not.
To implement this atomically in is_kprobe_*_slot(), also change
the insn_cache page list to an RCU list.
This changes timings a bit (it delays page freeing to the RCU garbage
collection phase), but none of that is in the hot path.
Note: this change can add small overhead to stack unwinders because
it adds 2 additional checks to __kernel_text_address(). However, the
impact should be very small, because kprobe_insn_pages list has 1 entry
per 256 probes(on x86, on arm/arm64 it will be 1024 probes),
and kprobe_optinsn_pages has 1 entry per 32 probes(on x86).
In most use cases, the number of kprobe events may be less
than 20, which means that is_kprobe_*_slot() will check just one entry.
Tested-by: Josh Poimboeuf <jpoimboe@redhat.com>
Signed-off-by: Masami Hiramatsu <mhiramat@kernel.org>
Acked-by: Peter Zijlstra <peterz@infradead.org>
Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: Ananth N Mavinakayanahalli <ananth@linux.vnet.ibm.com>
Cc: Andrew Morton <akpm@linux-foundation.org>
Cc: Andrey Konovalov <andreyknvl@google.com>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Jiri Olsa <jolsa@redhat.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Link: http://lkml.kernel.org/r/148388747896.6869.6354262871751682264.stgit@devbox
[ Improved the changelog and coding style. ]
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Pull perf/core improvements and fixes from Arnaldo Carvalho de Melo:
New features:
- Add more triggers to switch the output file (perf.data.TIMESTAMP).
Now, in addition to switching to a different output file when
receiving a SIGUSR2, one can also specify file size and time based
triggers:
perf record -a --switch-output=signal
is equivalent to what we had before:
perf record -a --switch-output
While we can also ask for the file to be "sliced" by size, taking
into account that that will happen only when we get woken up by
the kernel, i.e. one has to take into account the --mmap-pages (the
size of the perf mmap ring buffer):
perf record -a --switch-output=2G
will break the perf.data output into multiple files limited to 2GB
of samples, right when generating the output.
For time based samples, alert() will be used, so to have 1 minute
limited perf.data output files:
perf record -a --switch-output=1m
(Jiri Olsa)
- Remove the need to use -e only for syscalls and --event only for
tracepoints/HW/SW/etc events, i.e. now one can use:
perf trace -e nanosleep,futex,sched:sched_switch ./workload
or:
perf trace --event nanosleep,futex,sched:sched_switch ./workload
And have it tracing raw_syscalls:sys_{enter,exit} for the nanosleep
and futex syscalls, formatting those as strace does while also
tracing sched:sched_switch, ordering it all into one strace like
output.
Using '!' as the first character in the -e/--event argument remains
a way to negate the list of syscalls, i.e. all syscalls except for
the ones specified, doesn't affect the other kinds of events.
E.g:
[root@jouet ~] # perf trace -e sched:sched_switch,nanosleep usleep 1
0.000 ( 0.028 ms): usleep/28150 nanosleep(rqtp: 0x7ffe4201b9f0) ...
0.028 ( ): sched:sched_switch:usleep:28150 [120] S ==> swapper/0:0 [120])
0.000 ( 0.065 ms): usleep/28150 ... [continued]: nanosleep()) = 0
[root@jouet ~]#
(Arnaldo Carvalho de Melo)
- 'perf kallsyms' toy tool to look for extended symbol information on
the running kernel and demonstrate the machine/thread/symbol APIs for
use in other tools, such as 'perf probe' (Arnaldo Carvalho de Melo)
Infrastructure improvements:
- Add missing linux/kernel.h include to subcmd.h (Arnaldo Carvalho de Melo)
tools: Sync x86's vmx.h with the kernel
- Create libdir directory before installing libperf-jvmti.so (Laura Abbott)
- Fix typo in perf_evlist__start_workload() (Soramichi Akiyama)
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
To pick the changes from:
1b07304c58 ("KVM: nVMX: support descriptor table exits")
That adds entries to VMX_EXIT_REASONS, that is used by
tools/perf/arch/x86/util/kvm-stat.c.
This also picks the changes in:
1dc35dacc1 ("KVM: nVMX: check host CR3 on vmentry and vmexit")
But these are not used in 'perf kvm stat', do it just to silence the
kernel/tools file cache coherency detector:
$ make -C tools/perf
make: Entering directory '/home/acme/git/linux/tools/perf'
BUILD: Doing 'make -j4' parallel build
Warning: arch/x86/include/uapi/asm/vmx.h differs from kernel
Cc: Adrian Hunter <adrian.hunter@intel.com>
Cc: David Ahern <dsahern@gmail.com>
Cc: Jiri Olsa <jolsa@kernel.org>
Cc: Ladi Prosek <lprosek@redhat.com>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Paolo Bonzini <pbonzini@redhat.com>
Cc: Wang Nan <wangnan0@huawei.com>
Link: http://lkml.kernel.org/n/tip-56uowkk8t5zje49a42asffcy@git.kernel.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
It's now possible to specify the threshold time for perf.data like:
$ perf record --switch-output=30s ...
Once it's reached, the current data are dumped in to the
perf.data.<timestamp> file and session does on.
$ perf record --switch-output=30s ...
[ perf record: dump data: Woken up 44 times ]
[ perf record: Dump perf.data.2017010213043746 ]
...
The time is expected to be a number with appended unit
character - s/m/h/d.
Signed-off-by: Jiri Olsa <jolsa@kernel.org>
Tested-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Acked-by: Wang Nan <wangnan0@huawei.com>
Cc: David Ahern <dsahern@gmail.com>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Link: http://lkml.kernel.org/r/1483955520-29063-7-git-send-email-jolsa@kernel.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
It's now possible to specify the threshold size for perf.data like:
$ perf record --switch-output=2G ...
Once it's reached, the current data are dumped in to the
perf.data.<timestamp> file and session does on.
$ perf record --switch-output=2G ...
[ perf record: dump data: Woken up 7244 times ]
[ perf record: Dump perf.data.2017010214093746 ]
...
The size is expected to be a number with appended unit character -
B/K/M/G.
Signed-off-by: Jiri Olsa <jolsa@kernel.org>
Acked-by: Wang Nan <wangnan0@huawei.com>
Tested-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: David Ahern <dsahern@gmail.com>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Link: http://lkml.kernel.org/r/1483955520-29063-5-git-send-email-jolsa@kernel.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Makes it easier to specify both events and syscalls (to be formatter
strace-like), i.e. previously one would have to do:
# perf trace -e nanosleep --event sched:sched_switch usleep 1
Now it is possible to do:
# perf trace -e nanosleep,sched:sched_switch usleep 1
0.000 ( 0.021 ms): usleep/17962 nanosleep(rqtp: 0x7ffdedd61ec0) ...
0.021 ( ): sched:sched_switch:usleep:17962 [120] S ==> swapper/1:0 [120])
0.000 ( 0.066 ms): usleep/17962 ... [continued]: nanosleep()) = 0
#
The old style --expr and using both -e and --event continues to work.
Cc: Adrian Hunter <adrian.hunter@intel.com>
Cc: David Ahern <dsahern@gmail.com>
Cc: Jiri Olsa <jolsa@kernel.org>
Cc: Milian Wolff <milian.wolff@kdab.com>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Wang Nan <wangnan0@huawei.com>
Link: http://lkml.kernel.org/n/tip-ieg6bakub4657l9e6afn85r4@git.kernel.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Its similar to doing grep on a /proc/kallsyms, but it also shows extra
information like the path to the kernel module and the unrelocated
addresses in it, to help in diagnosing problems.
It is also helps demonstrate the use of the symbols routines so that
tool writers can use them more effectively.
Using it:
$ perf kallsyms e1000_xmit_frame netif_rx usb_stor_set_xfer_buf
e1000_xmit_frame: [e1000e] /lib/modules/4.9.0+/kernel/drivers/net/ethernet/intel/e1000e/e1000e.ko 0xffffffffc046fc10-0xffffffffc0470bb0 (0x19c80-0x1ac20)
netif_rx: [kernel] [kernel.kallsyms] 0xffffffff916f03a0-0xffffffff916f0410 (0xffffffff916f03a0-0xffffffff916f0410)
usb_stor_set_xfer_buf: [usb_storage] /lib/modules/4.9.0+/kernel/drivers/usb/storage/usb-storage.ko 0xffffffffc057aea0-0xffffffffc057af19 (0xf10-0xf89)
$
Cc: Adrian Hunter <adrian.hunter@intel.com>
Cc: David Ahern <dsahern@gmail.com>
Cc: Jiri Olsa <jolsa@kernel.org>
Cc: Masami Hiramatsu <mhiramat@kernel.org>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Wang Nan <wangnan0@huawei.com>
Link: http://lkml.kernel.org/n/tip-79bk9pakujn4l4vq0f90klv3@git.kernel.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Pull perf/urgent fixes and one improvement from Arnaldo Carvalho de Melo:
Fixes:
- Fix prev/next_prio formatting for deadline tasks in libtraceevent (Daniel Bristot de Oliveira)
- Robustify reading of build-ids from /sys/kernel/note (Arnaldo Carvalho de Melo)
- Fix building some sample/bpf in Alpine Linux 3.4 (Arnaldo Carvalho de Melo)
- Fix 'make install-bin' to install libtraceevent plugins (Arnaldo Carvalho de Melo)
- Fix 'perf record --switch-output' documentation and comment (Jiri Olsa)
- Fix 'perf probe' for cross arch probing (Masami Hiramatsu)
Improvement:
- Show total scheduling time in 'perf sched timehist' (Namhyumg Kim)
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Fix perf-probe to show probe definition on gcc generated symbols for
offline kernel (including cross-arch kernel image).
gcc sometimes optimizes functions and generate new symbols with suffixes
such as ".constprop.N" or ".isra.N" etc. Since those symbol names are
not recorded in DWARF, we have to find correct generated symbols from
offline ELF binary to probe on it (kallsyms doesn't correct it). For
online kernel or uprobes we don't need it because those are rebased on
_text, or a section relative address.
E.g. Without this:
$ perf probe -k build-arm/vmlinux -F __slab_alloc*
__slab_alloc.constprop.9
$ perf probe -k build-arm/vmlinux -D __slab_alloc
p:probe/__slab_alloc __slab_alloc+0
If you put above definition on target machine, it should fail
because there is no __slab_alloc in kallsyms.
With this fix, perf probe shows correct probe definition on
__slab_alloc.constprop.9:
$ perf probe -k build-arm/vmlinux -D __slab_alloc
p:probe/__slab_alloc __slab_alloc.constprop.9+0
Signed-off-by: Masami Hiramatsu <mhiramat@kernel.org>
Cc: Jiri Olsa <jolsa@redhat.com>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Link: http://lkml.kernel.org/r/148350060434.19001.11864836288580083501.stgit@devbox
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Fix --funcs (-F) option to show correct symbols for offline module.
Since previous perf-probe uses machine__findnew_module_map() for offline
module, even if user passes a module file (with full path) which is for
other architecture, perf-probe always tries to load symbol map for
current kernel module.
This fix uses dso__new_map() to load the map from given binary as same
as a map for user applications.
Signed-off-by: Masami Hiramatsu <mhiramat@kernel.org>
Cc: Jiri Olsa <jolsa@redhat.com>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Link: http://lkml.kernel.org/r/148350053478.19001.15435255244512631545.stgit@devbox
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Currently, the sched:sched_switch tracepoint reports deadline tasks with
priority -1. But when reading the trace via perf script I've got the
following output:
# ./d & # (d is a deadline task, see [1])
# perf record -e sched:sched_switch -a sleep 1
# perf script
...
swapper 0 [000] 2146.962441: sched:sched_switch: swapper/0:0 [120] R ==> d:2593 [4294967295]
d 2593 [000] 2146.972472: sched:sched_switch: d:2593 [4294967295] R ==> g:2590 [4294967295]
The task d reports the wrong priority [4294967295]. This happens because
the "int prio" is stored in an unsigned long long val. Although it is
set as a %lld, as int is shorter than unsigned long long,
trace_seq_printf prints it as a positive number.
The fix is just to cast the val as an int, and print it as a %d,
as in the sched:sched_switch tracepoint's "format".
The output with the fix is:
# ./d &
# perf record -e sched:sched_switch -a sleep 1
# perf script
...
swapper 0 [000] 4306.374037: sched:sched_switch: swapper/0:0 [120] R ==> d:10941 [-1]
d 10941 [000] 4306.383823: sched:sched_switch: d:10941 [-1] R ==> swapper/0:0 [120]
[1] d.c
---
#include <stdio.h>
#include <unistd.h>
#include <sys/syscall.h>
#include <linux/types.h>
#include <linux/sched.h>
struct sched_attr {
__u32 size, sched_policy;
__u64 sched_flags;
__s32 sched_nice;
__u32 sched_priority;
__u64 sched_runtime, sched_deadline, sched_period;
};
int sched_setattr(pid_t pid, const struct sched_attr *attr, unsigned int flags)
{
return syscall(__NR_sched_setattr, pid, attr, flags);
}
int main(void)
{
struct sched_attr attr = {
.size = sizeof(attr),
.sched_policy = SCHED_DEADLINE, /* This creates a 10ms/30ms reservation */
.sched_runtime = 10 * 1000 * 1000,
.sched_period = attr.sched_deadline = 30 * 1000 * 1000,
};
if (sched_setattr(0, &attr, 0) < 0) {
perror("sched_setattr");
return -1;
}
for(;;);
}
---
Committer notes:
Got the program from the provided URL, http://bristot.me/lkml/d.c,
trimmed it and included in the cset log above, so that we have
everything needed to test it in one place.
Signed-off-by: Daniel Bristot de Oliveira <bristot@redhat.com>
Acked-by: Steven Rostedt <rostedt@goodmis.org>
Tested-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: Daniel Bristot de Oliveira <bristot@redhat.com>
Cc: Jiri Olsa <jolsa@redhat.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Link: http://lkml.kernel.org/r/866ef75bcebf670ae91c6a96daa63597ba981f0d.1483443552.git.bristot@redhat.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Since 'perf probe' supports cross-arch probes, it is possible to analyze
different arch kernel image which has different bits-per-long.
In that case, it fails to get the module name because it uses the
MOD_NAME_OFFSET macro based on the host machine bits-per-long, instead
of the target arch bits-per-long.
This fixes above issue by changing modname-offset based on the target
archs bit width. This is ok because linux kernel uses LP64 model on
64bit arch.
E.g. without this (on x86_64, and target module is arm32):
$ perf probe -m build-arm/fs/configfs/configfs.ko -D configfs_lookup
p:probe/configfs_lookup :configfs_lookup+0
^-Here is an empty module name.
With this fix, you can see correct module name:
$ perf probe -m build-arm/fs/configfs/configfs.ko -D configfs_lookup
p:probe/configfs_lookup configfs:configfs_lookup+0
Signed-off-by: Masami Hiramatsu <mhiramat@kernel.org>
Cc: Jiri Olsa <jolsa@redhat.com>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Link: http://lkml.kernel.org/r/148337043836.6752.383495516397005695.stgit@devbox
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Pull DAX updates from Dan Williams:
"The completion of Jan's DAX work for 4.10.
As I mentioned in the libnvdimm-for-4.10 pull request, these are some
final fixes for the DAX dirty-cacheline-tracking invalidation work
that was merged through the -mm, ext4, and xfs trees in -rc1. These
patches were prepared prior to the merge window, but we waited for
4.10-rc1 to have a stable merge base after all the prerequisites were
merged.
Quoting Jan on the overall changes in these patches:
"So I'd like all these 6 patches to go for rc2. The first three
patches fix invalidation of exceptional DAX entries (a bug which
is there for a long time) - without these patches data loss can
occur on power failure even though user called fsync(2). The other
three patches change locking of DAX faults so that ->iomap_begin()
is called in a more relaxed locking context and we are safe to
start a transaction there for ext4"
These have received a build success notification from the kbuild
robot, and pass the latest libnvdimm unit tests. There have not been
any -next releases since -rc1, so they have not appeared there"
* 'libnvdimm-fixes' of git://git.kernel.org/pub/scm/linux/kernel/git/nvdimm/nvdimm:
ext4: Simplify DAX fault path
dax: Call ->iomap_begin without entry lock during dax fault
dax: Finish fault completely when loading holes
dax: Avoid page invalidation races and unnecessary radix tree traversals
mm: Invalidate DAX radix tree entries only if appropriate
ext2: Return BH_New buffers for zeroed blocks
Pull documentation fixes from Jonathan Corbet:
"Two small fixes:
- A merge error on my part broke the DocBook build. I've
requisitioned one of tglx's frozen sharks for appropriate
disciplinary action and resolved to be more careful about testing
the DocBook stuff as long as it's still around.
- Fix an error in unaligned-memory-access.txt"
* tag 'docs-4.10-rc1-fix' of git://git.lwn.net/linux:
Documentation/unaligned-memory-access.txt: fix incorrect comparison operator
docs: Fix build failure
Pull crypto fix from Herbert Xu:
"This fixes a boot failure on some platforms when crypto self test is
enabled along with the new acomp interface"
* 'linus' of git://git.kernel.org/pub/scm/linux/kernel/git/herbert/crypto-2.6:
crypto: testmgr - Use heap buffer for acomp test input
mm/filemap.c: In function 'clear_bit_unlock_is_negative_byte':
mm/filemap.c:933:9: error: too few arguments to function 'test_bit'
return test_bit(PG_waiters);
^~~~~~~~
Fixes: b91e1302ad ('mm: optimize PageWaiters bit use for unlock_page()')
Signed-off-by: Olof Johansson <olof@lixom.net>
Brown-paper-bag-by: Linus Torvalds <dummy@duh.com>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
In commit 6290602709 ("mm: add PageWaiters indicating tasks are
waiting for a page bit") Nick Piggin made our page locking no longer
unconditionally touch the hashed page waitqueue, which not only helps
performance in general, but is particularly helpful on NUMA machines
where the hashed wait queues can bounce around a lot.
However, the "clear lock bit atomically and then test the waiters bit"
sequence turns out to be much more expensive than it needs to be,
because you get a nasty stall when trying to access the same word that
just got updated atomically.
On architectures where locking is done with LL/SC, this would be trivial
to fix with a new primitive that clears one bit and tests another
atomically, but that ends up not working on x86, where the only atomic
operations that return the result end up being cmpxchg and xadd. The
atomic bit operations return the old value of the same bit we changed,
not the value of an unrelated bit.
On x86, we could put the lock bit in the high bit of the byte, and use
"xadd" with that bit (where the overflow ends up not touching other
bits), and look at the other bits of the result. However, an even
simpler model is to just use a regular atomic "and" to clear the lock
bit, and then the sign bit in eflags will indicate the resulting state
of the unrelated bit #7.
So by moving the PageWaiters bit up to bit #7, we can atomically clear
the lock bit and test the waiters bit on x86 too. And architectures
with LL/SC (which is all the usual RISC suspects), the particular bit
doesn't matter, so they are fine with this approach too.
This avoids the extra access to the same atomic word, and thus avoids
the costly stall at page unlock time.
The only downside is that the interface ends up being a bit odd and
specialized: clear a bit in a byte, and test the sign bit. Nick doesn't
love the resulting name of the new primitive, but I'd rather make the
name be descriptive and very clear about the limitation imposed by
trying to work across all relevant architectures than make it be some
generic thing that doesn't make the odd semantics explicit.
So this introduces the new architecture primitive
clear_bit_unlock_is_negative_byte();
and adds the trivial implementation for x86. We have a generic
non-optimized fallback (that just does a "clear_bit()"+"test_bit(7)"
combination) which can be overridden by any architecture that can do
better. According to Nick, Power has the same hickup x86 has, for
example, but some other architectures may not even care.
All these optimizations mean that my page locking stress-test (which is
just executing a lot of small short-lived shell scripts: "make test" in
the git source tree) no longer makes our page locking look horribly bad.
Before all these optimizations, just the unlock_page() costs were just
over 3% of all CPU overhead on "make test". After this, it's down to
0.66%, so just a quarter of the cost it used to be.
(The difference on NUMA is bigger, but there this micro-optimization is
likely less noticeable, since the big issue on NUMA was not the accesses
to 'struct page', but the waitqueue accesses that were already removed
by Nick's earlier commit).
Acked-by: Nick Piggin <npiggin@gmail.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Bob Peterson <rpeterso@redhat.com>
Cc: Steven Whitehouse <swhiteho@redhat.com>
Cc: Andrew Lutomirski <luto@kernel.org>
Cc: Andreas Gruenbacher <agruenba@redhat.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Mel Gorman <mgorman@techsingularity.net>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Pull crypto fix from Herbert Xu:
"This fixes a hash corruption bug in the marvell driver"
* 'linus' of git://git.kernel.org/pub/scm/linux/kernel/git/herbert/crypto-2.6:
crypto: marvell - Copy IVDIG before launching partial DMA ahash requests
