The hwpmc pcpu sample buffer is prone to head of line blocking
when waiting for user process to return to user space and
collect a pending callchain. If more than one tick has elapsed
between the time the sample entry was marked for collection and
the time that the hardclock pmc handler runs to copy the records
to a larger temporary buffer, mark the sample entry as not in
use.
This changes reduces the number of samples marked as not valid
when collecting under load from ~99.5% to 5-20%.
Reported by: mjg@
MFC after: 3 days
Remove malloc_domain(9) and most other _domain KPIs added in r327900.
The new functions allow the caller to specify a general NUMA domain
selection policy, rather than specifically requesting an allocation from
a specific domain. The latter policy tends to interact poorly with
M_WAITOK, resulting in situations where a caller is blocked indefinitely
because the specified domain is depleted. Most existing consumers of
the _domain KPIs are converted to instead use a DOMAINSET_PREF() policy,
in which we fall back to other domains to satisfy the allocation
request.
This change also defines a set of DOMAINSET_FIXED() policies, which
only permit allocations from the specified domain.
Discussed with: gallatin, jeff
Reported and tested by: pho (previous version)
MFC after: 2 weeks
Sponsored by: The FreeBSD Foundation
Differential Revision: https://reviews.freebsd.org/D17418
Refactor sample ring buffer ring handling to make it more robust to
long running callchain collection handling
r338112 introduced a (now fixed) regression that exposed a number of race
conditions within the management of the sample buffers. This
simplifies the handling and moves the decision to overwrite a
callchain sample that has taken too long out of the NMI in to the
hardlock handler. With this change the problem no longer shows up as a
ring corruption but as the code spending all of its time in callchain
collection.
- Makes the producer / consumer index incrementing monotonic, making it
easier (for me at least) to reason about.
- Moves the decision to overwrite a sample from NMI context to interrupt
context where we can enforce serialization.
- Puts a time limit on waiting to collect a user callchain - putting a
bound on head-of-line blocking causing samples to be dropped
- Removes the flush routine which was previously needed to purge
dangling references to the pmc from the sample buffers but now is only
a source of a race condition on unload.
Previously one could lock up or crash HEAD by running:
pmcstat -S inst_retired.any_p -T and then hitting ^C
After this change it is no longer possible.
PR: 231793
Reviewed by: markj@
Approved by: re (gjb@)
Differential Revision: https://reviews.freebsd.org/D17011
Not all event descriptions have a sample rate (such as inst_retired.any)
this will restore the legacy behavior of using 65536 in that case. It also
prevents accidental API misuse that could lead to panic.
PR: 230985
Reported by: markj
Reviewed by: markj
Approved by: re (gjb)
Sponsored by: Limelight Networks
Differential Revision: https://reviews.freebsd.org/D16958
r336773 removed all things xscale. However, some things xscale are
really armv5. Revert that entirely. A more modest removal will follow.
Noticed by: andrew@
The OLD XSCALE stuff hasn't been useful in a while. The original
committer (cognet@) was the only one that had boards for it. He's
blessed this removal. Newer XSCALE (GUMSTIX) is for hardware that's
quite old. After discussion on arm@, it was clear there was no support
for keeping it.
Differential Review: https://reviews.freebsd.org/D16313
Remove all the big-endian arm architectures (ixp425 and ixp435)
support in the kernel and associated drivers.
Differential Revision: https://reviews.freebsd.org/D16257
- Add tracker argument to preemptible epochs
- Inline epoch read path in kernel and tied modules
- Change in_epoch to take an epoch as argument
- Simplify tfb_tcp_do_segment to not take a ti_locked argument,
there's no longer any benefit to dropping the pcbinfo lock
and trying to do so just adds an error prone branchfest to
these functions
- Remove cases of same function recursion on the epoch as
recursing is no longer free.
- Remove the the TAILQ_ENTRY and epoch_section from struct
thread as the tracker field is now stack or heap allocated
as appropriate.
Tested by: pho and Limelight Networks
Reviewed by: kbowling at llnw dot com
Sponsored by: Limelight Networks
Differential Revision: https://reviews.freebsd.org/D16066
pmc_process_interrupt takes 5 arguments when only 3 are needed.
cpu is always available in curcpu and inuserspace can always be
derived from the passed trapframe.
While facially a reasonable cleanup this change was motivated
by the need to workaround a compiler bug.
core2_intr(cpu, tf) ->
pmc_process_interrupt(cpu, ring, pmc, tf, inuserspace) ->
pmc_add_sample(cpu, ring, pm, tf, inuserspace)
In the process of optimizing the tail call the tf pointer was getting
clobbered:
(kgdb) up
at /storage/mmacy/devel/freebsd/sys/dev/hwpmc/hwpmc_mod.c:4709
4709 pmc_save_kernel_callchain(ps->ps_pc,
(kgdb) up
1205 error = pmc_process_interrupt(cpu, PMC_HR, pm, tf,
resulting in a crash in pmc_save_kernel_callchain.
- add '-j' options to filter to enable converting native pmc
log format to json lines format to enable the use of scripts
and external tooling
% pmc filter -j pmc.log pmc.jsonl
- Record the tsc value in sampling interrupts as opposed to
recording nanotime when the sample is copied to a global log
in hardclock - potentially many milliseconds later.
- At initialize record the tsc_freq and the time of day to give
us an offset for translating the tsc values in callchain records
By logging all threads and processes 'pmc filter'
can now filter on process or thread name, relieving
the user of the burden of determining which tid or
pid was which when the sample was taken.
% pmc filter -T if_io_tqg -P nginx pmc.log pmc-iflib.log
% pmc filter -x -T idle pmc.log pmc-noidle.log
- increase pmc cpuid field from 8 to 12 bits
- add cpuid version string to initialize entry in the log
so that filter can identify which counter index an
event name maps to
- GC unused config flags
- make fixed counter assignment more robust as well as the
changes needed to be properly identified for filter
This adds the -U options to pmcstat which will attribute in-kernel samples
back to the user stack that invoked the system call. It is not the default,
because when looking at kernel profiles it is generally more desirable to
merge all instances of a given system call together.
Although heavily revised, this change is directly derived from D7350 by
Jonathan T. Looney.
Obtained from: jtl
Sponsored by: Juniper Networks, Limelight Networks
Intel now provides comprehensive tables for all performance counters
and the various valid configuration permutations as text .json files.
Libpmc has been converted to use these and hwpmc_core has been greatly
simplified by moving to passthrough of the table values.
The one gotcha is that said tables don't support pentium pro and and pentium
IV. There's very few users of hwpmc on _amd64_ kernels on new hardware. It is
unlikely that anyone is doing low level optimization on 15 year old Intel
hardware. Nonetheless, if someone feels strongly enough to populate the
corresponding tables for p4 and ppro I will reinstate the files in to the
build.
Code for the K8 counters and !x86 architectures remains unchanged.
vendor provided pmu-events tables and sundry cleanups.
The vendor pmu-events tables provide counter descriptions, default
sample rates, event, umask, and flag values for all the counter
configuration permutations. Using this gives us:
- much simpler kernel code for the MD component
- helpful long and short event descriptions
- simpler user code
- sample rates that won't overload the system
Update man page with newer sample types and remove unused sample type.
vendor provided pmu-events tables and sundry cleanups.
The vendor pmu-events tables provide counter descriptions, default
sample rates, event, umask, and flag values for all the counter
configuration permutations. Using this gives us:
- much simpler kernel code for the MD component
- helpful long and short event descriptions
- simpler user code
- sample rates that won't overload the system
Update man page with newer sample types and remove unused sample type.
Squashed commit of the following:
commit 4459d43eff815bec08ccc5533dbe5de846f03128
Author: Matt Macy <mmacy@mattmacy.io>
Date: Sat May 26 00:06:31 2018 -0700
libpmc: fix pmu function signatures for non amd64
commit a2cb8bbc586c65d41f9b291430a2261ec67b59fe
Author: Matt Macy <mmacy@mattmacy.io>
Date: Fri May 25 22:38:11 2018 -0700
pmcstat: fix indentation of usage
commit f686954b15ff56a833ac80404898977cb80a265b
Author: Matt Macy <mmacy@mattmacy.io>
Date: Fri May 25 22:19:49 2018 -0700
pmclog(3): add callchain and pmcallocatedyn, remove pcsample
commit 73e13a0d2e9498c81c150d14d022050cee7511bb
Author: Matt Macy <mmacy@mattmacy.io>
Date: Fri May 25 22:19:00 2018 -0700
pmclog.h: GC pcsample field
commit 3e93ffd65da641fa657539dad3c48e281f8b5798
Author: Matt Macy <mmacy@mattmacy.io>
Date: Fri May 25 22:05:57 2018 -0700
hwpmc: make Intel core CPUs use external event tables
commit 634f5fae1e1644ac324003136c66cd9c619d1c93
Author: Matt Macy <mmacy@mattmacy.io>
Date: Fri May 25 22:00:06 2018 -0700
pmclog: update log record types, bump PMC_MAJOR
- explicitly make log record types a multiple of 8 bytes
- hook in pmu event types for pmc_allocate records
- remove references to no longer PCSAMPLE record
commit 83d84fcd2d65bdf6ddcb2e155a22f0cfa2a9c225
Author: Matt Macy <mmacy@mattmacy.io>
Date: Fri May 25 21:52:10 2018 -0700
libpmc: add support for having vendor table driven pmc_allocate
commit 9e6ad63c40c2fce8404847ace5078ca6cb33a736
Author: Matt Macy <mmacy@mattmacy.io>
Date: Fri May 25 19:11:33 2018 -0700
hwpmc_core: add accessors for EVSEL & UMASK, make IAP_UMASK useful to user
commit 859dceb93daa6419a48c794db99b6758e5b041c9
Author: Matt Macy <mmacy@mattmacy.io>
Date: Fri May 25 19:09:45 2018 -0700
pmcstat: update usage and man page as well as make -L consistent with pmccontrol
commit 79c7d8597e28c2eb13f5f9113e65ec2792ca57b1
Author: Matt Macy <mmacy@mattmacy.io>
Date: Fri May 25 18:07:03 2018 -0700
pmu_util: add support for all current intel event keywords
commit d8089c7f6a6c8527f38324252b1ffb47004694c6
Author: Matt Macy <mmacy@mattmacy.io>
Date: Fri May 25 17:45:00 2018 -0700
add description for new arguments
commit 058336740bab53c62ec88a3a026ea848cf3878c6
Author: Matt Macy <mmacy@mattmacy.io>
Date: Fri May 25 17:38:15 2018 -0700
libpmc: move pmu_events table and pmu_utils out of libpmcstat so that they can be used by pmc_allocate
commit 049b66b382e2f833c3f47bc8df9e750cb265709f
Author: Matt Macy <mmacy@mattmacy.io>
Date: Fri May 25 16:12:41 2018 -0700
pmcstat: hook pmu_events counter description utility routines in
commit f5e01e7b37a691dc045e1aa16b3ebdd162515de8
Author: Matt Macy <mmacy@mattmacy.io>
Date: Fri May 25 16:11:59 2018 -0700
pmu_events: add utility routines for listing counters and their descriptions
commit cba4d4f8907f772279f86f18f915e0d74d33ac56
Author: Matt Macy <mmacy@mattmacy.io>
Date: Fri May 25 16:09:50 2018 -0700
pmu-events: expand out skylake regex to simplify string matches
There are risks associated with waiting on a preemptible epoch section.
Change the name to make them not be the default and document the issue
under CAVEATS.
Reported by: markj
This implements per-thread counters for PMC sampling. The thread
descriptors are stored in a list attached to the process descriptor.
These thread descriptors can store any per-thread information necessary
for current or future features. For the moment, they just store the counters
for sampling.
The thread descriptors are created when the process descriptor is created.
Additionally, thread descriptors are created or freed when threads
are started or stopped. Because the thread exit function is called in a
critical section, we can't directly free the thread descriptors. Hence,
they are freed to a cache, which is also used as a source of allocations
when needed for new threads.
Approved by: sbruno
Obtained from: jtl
Sponsored by: Juniper Networks, Limelight Networks
Differential Revision: https://reviews.freebsd.org/D15335
Once a pmc owner is added to the pmc_ss_owners list it is
visible for all to see. We don't want this to happen until
setup is complete.
Reported by: mjg
Approved by: sbruno
- fix load/unload race by allocating the per-domain list structure at boot
- fix long extant vm map LOR by replacing pmc_sx sx_slock with global_epoch
to protect the liveness of elements of the pmc_ss_owners list
Reported by: pho
Approved by: sbruno
It appears that domain information is set correctly independent
of whether or not NUMA is defined. However, there is no memory
backing secondary domains leading to allocation failure.
Reported by: pho@, np@
Approved by: sbruno@
