In commit f8f2835a9c we changed the behavior of XFS to use EFIs to
remove blocks from an overfilled AGFL because there were complaints
about transaction overruns that stemmed from trying to free multiple
blocks in a single transaction.
Unfortunately, that commit missed a subtlety in the debug-mode
transaction accounting when a realtime volume is attached. If a
realtime file undergoes a data fork mapping change such that realtime
extents are allocated (or freed) in the same transaction that a data
device block is also allocated (or freed), we can trip a debugging
assertion. This can happen (for example) if a realtime extent is
allocated and it is necessary to reshape the bmbt to hold the new
mapping.
When we go to allocate a bmbt block from an AG, the first thing the data
device block allocator does is ensure that the freelist is the proper
length. If the freelist is too long, it will trim the freelist to the
proper length.
In debug mode, trimming the freelist calls xfs_trans_agflist_delta() to
record the decrement in the AG free list count. Prior to f8f28 we would
put the free block back in the free space btrees in the same
transaction, which calls xfs_trans_agblocks_delta() to record the
increment in the AG free block count. Since AGFL blocks are included in
the global free block count (fdblocks), there is no corresponding
fdblocks update, so the AGFL free satisfies the following condition in
xfs_trans_apply_sb_deltas:
/*
* Check that superblock mods match the mods made to AGF counters.
*/
ASSERT((tp->t_fdblocks_delta + tp->t_res_fdblocks_delta) ==
(tp->t_ag_freeblks_delta + tp->t_ag_flist_delta +
tp->t_ag_btree_delta));
The comparison here used to be: (X + 0) == ((X+1) + -1 + 0), where X is
the number blocks that were allocated.
After commit f8f28 we defer the block freeing to the next chained
transaction, which means that the calls to xfs_trans_agflist_delta and
xfs_trans_agblocks_delta occur in separate transactions. The (first)
transaction that shortens the free list trips on the comparison, which
has now become:
(X + 0) == ((X) + -1 + 0)
because we haven't freed the AGFL block yet; we've only logged an
intention to free it. When the second transaction (the deferred free)
commits, it will evaluate the expression as:
(0 + 0) == (1 + 0 + 0)
and trip over that in turn.
At this point, the astute reader may note that the two commits tagged by
this patch have been in the kernel for a long time but haven't generated
any bug reports. How is it that the author became aware of this bug?
This originally surfaced as an intermittent failure when I was testing
realtime rmap, but a different bug report by Zorro Lang reveals the same
assertion occuring on !lazysbcount filesystems.
The common factor to both reports (and why this problem wasn't
previously reported) becomes apparent if we consider when
xfs_trans_apply_sb_deltas is called by __xfs_trans_commit():
if (tp->t_flags & XFS_TRANS_SB_DIRTY)
xfs_trans_apply_sb_deltas(tp);
With a modern lazysbcount filesystem, transactions update only the
percpu counters, so they don't need to set XFS_TRANS_SB_DIRTY, hence
xfs_trans_apply_sb_deltas is rarely called.
However, updates to the count of free realtime extents are not part of
lazysbcount, so XFS_TRANS_SB_DIRTY will be set on transactions adding or
removing data fork mappings to realtime files; similarly,
XFS_TRANS_SB_DIRTY is always set on !lazysbcount filesystems.
Dave mentioned in response to an earlier version of this patch:
"IIUC, what you are saying is that this debug code is simply not
exercised in normal testing and hasn't been for the past decade? And it
still won't be exercised on anything other than realtime device testing?
"...it was debugging code from 1994 that was largely turned into dead
code when lazysbcounters were introduced in 2007. Hence I'm not sure it
holds any value anymore."
This debugging code isn't especially helpful - you can modify the
flcount on one AG and the freeblks of another AG, and it won't trigger.
Add the fact that nobody noticed for a decade, and let's just get rid of
it (and start testing realtime :P).
This bug was found by running generic/051 on either a V4 filesystem
lacking lazysbcount; or a V5 filesystem with a realtime volume.
Cc: bfoster@redhat.com, zlang@redhat.com
Fixes: f8f2835a9c ("xfs: defer agfl block frees when dfops is available")
Signed-off-by: Darrick J. Wong <djwong@kernel.org>
Reviewed-by: Brian Foster <bfoster@redhat.com>
We were doing it in tools/build/Makefile.feature, after running the
feature checks, but then in tools/perf/Makefile.config we can call more
feature checks when we notice that some feature check failed, like when
libbfd wasn't detected and we add libraries to the LDFLAGS of its
feature check to try again, etc.
Acked-by: Jiri Olsa <jolsa@kernel.org>
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
By setting FEATURE_DISPLAY_DEFERRED=1 a tool may ask for the printout
of the detected features in tools/build/Makefile.feature to be done
later adter extra feature checks are done that are tool specific.
The perf tool will do it via its tools/perf/Makefile.config, as it
performs such extra feature checks.
Acked-by: Jiri Olsa <jolsa@kernel.org>
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Feature detection is done in tools/build/Makefile.feature, we may exit
there with some features not detected and then, in
tools/perf/Makefile.config try adding extra libraries to link and then
do extra feature checks to see if we now find the feature.
This is the case with the disassembler-four-args that checks if the
diassembler() function in libopcodes (binutils) has a signature with
one or with four arguments, as this is not ABI and they changed it at
some point.
This is not a problem when doing normal builds, for instance:
$ make -C tools/perf O=/tmp/build/perf
As we don't use what is in FEATURE-DUMP at that point, but is a problem
if we pass FEATURE_DUMP=/previously-detected-features as we do in
'make -C tools/perf build-test' to reuse the feature detection in the
many build combinations we test there.
When that is done feature-disassembler-four-args will be set to 0, but
opensuse 15.1 has the four arguments function signature in
disassembler(). The build thus fails.
Fix it by rewriting the FEATURE-DUMP file at the end of
tools/perf/Makefile.config to register features we retested in that make
file.
Signed-off-by: Jiri Olsa <jolsa@redhat.com>
Reported-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Tested-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Currently we support only static linking with kernel's libtraceevent
(tools/lib/traceevent). This patch adds libtraceevent package detection
and support to link perf with it dynamically.
The libtraceevent package status is displayed with:
$ make VF=1 LIBTRACEEVENT_DYNAMIC=1
...
... libtraceevent: [ on ]
Default behavior remains the same (static linking).
Committer testing:
$ make LIBTRACEEVENT_DYNAMIC=1 VF=1 O=/tmp/build/perf -C tools/perf install-bin |& grep traceevent
Makefile.config:1090: *** Error: No libtraceevent devel library found, please install libtraceevent-devel. Stop.
$
Signed-off-by: Michael Petlan <mpetlan@redhat.com>
Tested-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Jiri Olsa <jolsa@redhat.com>
LPU-Reference: 20210428092023.4009-1-mpetlan@redhat.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
If a group has events which are from different hybrid PMUs,
shows a warning:
"WARNING: events in group from different hybrid PMUs!"
This is to remind the user not to put the core event and atom
event into one group.
Next, just disable grouping.
# perf stat -e "{cpu_core/cycles/,cpu_atom/cycles/}" -a -- sleep 1
WARNING: events in group from different hybrid PMUs!
WARNING: grouped events cpus do not match, disabling group:
anon group { cpu_core/cycles/, cpu_atom/cycles/ }
Performance counter stats for 'system wide':
5,438,125 cpu_core/cycles/
3,914,586 cpu_atom/cycles/
1.004250966 seconds time elapsed
Signed-off-by: Jin Yao <yao.jin@linux.intel.com>
Reviewed-by: Jiri Olsa <jolsa@kernel.org>
Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Kan Liang <kan.liang@intel.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Link: https://lore.kernel.org/r/20210427070139.25256-17-yao.jin@linux.intel.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Previously if '-e' is not specified in perf stat, some software events
and hardware events are added to evlist by default.
Before:
# perf stat -a -- sleep 1
Performance counter stats for 'system wide':
24,044.40 msec cpu-clock # 23.946 CPUs utilized
99 context-switches # 4.117 /sec
24 cpu-migrations # 0.998 /sec
3 page-faults # 0.125 /sec
7,000,244 cycles # 0.000 GHz
2,955,024 instructions # 0.42 insn per cycle
608,941 branches # 25.326 K/sec
31,991 branch-misses # 5.25% of all branches
1.004106859 seconds time elapsed
Among the events, cycles, instructions, branches and branch-misses
are hardware events.
One hybrid platform, two hardware events are created for one
hardware event.
cpu_core/cycles/,
cpu_atom/cycles/,
cpu_core/instructions/,
cpu_atom/instructions/,
cpu_core/branches/,
cpu_atom/branches/,
cpu_core/branch-misses/,
cpu_atom/branch-misses/
These events would be added to evlist on hybrid platform.
Since parse_events() has been supported to create two hardware events
for one event on hybrid platform, so we just use parse_events(evlist,
"cycles,instructions,branches,branch-misses") to create the default
events and add them to evlist.
After:
# perf stat -a -- sleep 1
Performance counter stats for 'system wide':
24,043.99 msec cpu-clock # 23.991 CPUs utilized
139 context-switches # 5.781 /sec
25 cpu-migrations # 1.040 /sec
6 page-faults # 0.250 /sec
10,381,751 cpu_core/cycles/ # 431.782 K/sec
1,264,216 cpu_atom/cycles/ # 52.579 K/sec
3,406,958 cpu_core/instructions/ # 141.697 K/sec
414,588 cpu_atom/instructions/ # 17.243 K/sec
705,149 cpu_core/branches/ # 29.327 K/sec
82,358 cpu_atom/branches/ # 3.425 K/sec
40,821 cpu_core/branch-misses/ # 1.698 K/sec
9,086 cpu_atom/branch-misses/ # 377.891 /sec
1.002228863 seconds time elapsed
We can see two events are created for one hardware event.
One TODO is, the shadow stats looks a bit different, now it's just
'M/sec'.
The perf_stat__update_shadow_stats and perf_stat__print_shadow_stats
need to be improved in future if we want to get the original shadow
stats.
Signed-off-by: Jin Yao <yao.jin@linux.intel.com>
Reviewed-by: Jiri Olsa <jolsa@kernel.org>
Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Kan Liang <kan.liang@intel.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Link: https://lore.kernel.org/r/20210427070139.25256-15-yao.jin@linux.intel.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
On hybrid platform, user may want to enable events on one pmu.
Following syntax are supported:
cpu_core/<event>/
cpu_atom/<event>/
But the syntax doesn't work for cache event.
Before:
# perf stat -e cpu_core/LLC-loads/ -a -- sleep 1
event syntax error: 'cpu_core/LLC-loads/'
\___ unknown term 'LLC-loads' for pmu 'cpu_core'
Cache events are a bit complex. We can't create aliases for them.
We use another solution. For example, if we use "cpu_core/LLC-loads/",
in parse_events_add_pmu(), term->config is "LLC-loads".
Then we create a new parser to scan "LLC-loads". The
parse_events_add_cache() would be called during parsing.
The parse_state->hybrid_pmu_name is used to identify the pmu
where the event should be enabled on.
After:
# perf stat -e cpu_core/LLC-loads/ -a -- sleep 1
Performance counter stats for 'system wide':
24,593 cpu_core/LLC-loads/
1.003911601 seconds time elapsed
Signed-off-by: Jin Yao <yao.jin@linux.intel.com>
Reviewed-by: Jiri Olsa <jolsa@kernel.org>
Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Kan Liang <kan.liang@intel.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Link: https://lore.kernel.org/r/20210427070139.25256-13-yao.jin@linux.intel.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
The functions perf_pmu__is_hybrid and perf_pmu__find_hybrid_pmu
can be used to identify the hybrid platform and return the found
hybrid cpu pmu. All the detected hybrid pmus have been saved in
'perf_pmu__hybrid_pmus' list. So we just need to search this list.
perf_pmu__hybrid_type_to_pmu converts the user specified string
to hybrid pmu name. This is used to support the '--cputype' option
in next patches.
perf_pmu__has_hybrid checks the existing of hybrid pmu. Note that,
we have to define it in pmu.c (make pmu-hybrid.c no more symbol
dependency), otherwise perf test python would be failed.
Signed-off-by: Jin Yao <yao.jin@linux.intel.com>
Reviewed-by: Jiri Olsa <jolsa@kernel.org>
Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Kan Liang <kan.liang@intel.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Link: https://lore.kernel.org/r/20210427070139.25256-7-yao.jin@linux.intel.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
We identify the cpu_core pmu and cpu_atom pmu by explicitly
checking following files:
For cpu_core, checks:
"/sys/bus/event_source/devices/cpu_core/cpus"
For cpu_atom, checks:
"/sys/bus/event_source/devices/cpu_atom/cpus"
If the 'cpus' file exists and it has data, the pmu exists.
But in order not to hardcode the "cpu_core" and "cpu_atom",
and make the code in a generic way.
So if the path "/sys/bus/event_source/devices/cpu_xxx/cpus" exists, the
hybrid pmu exists. All the detected hybrid pmus are linked to a global
list 'perf_pmu__hybrid_pmus' and then next we just need to iterate the
list to get all hybrid pmu by using perf_pmu__for_each_hybrid_pmu.
Signed-off-by: Jin Yao <yao.jin@linux.intel.com>
Reviewed-by: Jiri Olsa <jolsa@kernel.org>
Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Kan Liang <kan.liang@intel.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Link: https://lore.kernel.org/r/20210427070139.25256-6-yao.jin@linux.intel.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
For some Intel platforms, such as Alderlake, which is a hybrid platform
and it consists of atom cpu and core cpu. Each cpu has dedicated event
list. Part of events are available on core cpu, part of events are
available on atom cpu.
The kernel exports new cpu pmus: cpu_core and cpu_atom. The event in
json is added with a new field "Unit" to indicate which pmu the event
is available on.
For example, one event in cache.json,
{
"BriefDescription": "Counts the number of load ops retired that",
"CollectPEBSRecord": "2",
"Counter": "0,1,2,3",
"EventCode": "0xd2",
"EventName": "MEM_LOAD_UOPS_RETIRED_MISC.MMIO",
"PEBScounters": "0,1,2,3",
"SampleAfterValue": "1000003",
"UMask": "0x80",
"Unit": "cpu_atom"
},
The unit "cpu_atom" indicates this event is only available on "cpu_atom".
In generated pmu-events.c, we can see:
{
.name = "mem_load_uops_retired_misc.mmio",
.event = "period=1000003,umask=0x80,event=0xd2",
.desc = "Counts the number of load ops retired that. Unit: cpu_atom ",
.topic = "cache",
.pmu = "cpu_atom",
},
But if without this patch, the "uncore_" prefix is added before "cpu_atom",
such as:
.pmu = "uncore_cpu_atom"
That would be a wrong pmu.
Signed-off-by: Jin Yao <yao.jin@linux.intel.com>
Reviewed-by: Jiri Olsa <jolsa@kernel.org>
Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Kan Liang <kan.liang@intel.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Link: https://lore.kernel.org/r/20210427070139.25256-3-yao.jin@linux.intel.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
To make the output more readable, I think it's better to remove 0's in
the output. Also the dummy event has no event stats so it just wasts
the space. Let's use the --skip-empty option to suppress it.
$ perf report --stat --skip-empty
Aggregated stats:
TOTAL events: 16530
MMAP events: 226
COMM events: 1596
EXIT events: 2
THROTTLE events: 121
UNTHROTTLE events: 117
FORK events: 1595
SAMPLE events: 719
MMAP2 events: 12147
CGROUP events: 2
FINISHED_ROUND events: 2
THREAD_MAP events: 1
CPU_MAP events: 1
TIME_CONV events: 1
cycles stats:
SAMPLE events: 719
Reviewed-by: Andi Kleen <ak@linux.intel.com>
Signed-off-by: Namhyung Kim <namhyung@kernel.org>
Tested-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Acked-by: Jiri Olsa <jolsa@redhat.com>
Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: Ian Rogers <irogers@google.com>
Cc: Ingo Molnar <mingo@kernel.org>
Cc: Mark Rutland <mark.rutland@arm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Link: https://lore.kernel.org/r/20210427013717.1651674-5-namhyung@kernel.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
Currently, to use BPF to aggregate perf event counters, the user uses
--bpf-counters option. Enable "use bpf by default" events with a config
option, stat.bpf-counter-events. Events with name in the option will use
BPF.
This also enables mixed BPF event and regular event in the same sesssion.
For example:
perf config stat.bpf-counter-events=instructions
perf stat -e instructions,cs
The second command will use BPF for "instructions" but not "cs".
Signed-off-by: Song Liu <song@kernel.org>
Cc: Jiri Olsa <jolsa@kernel.org>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Song Liu <songliubraving@fb.com>
Link: https://lore.kernel.org/r/20210425214333.1090950-4-song@kernel.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
