I see the fllowing panic on AMD when exiting pmcstat:
panic: [pmc,1473] pp_pmcval outside of expected range cpu=2 ri=17
pp_pmcval=fffffffffa529f5b pm_reloadcount=10000
It seems that at least on AMD a performance counter keeps counting after
overflowing. When pmcstat exits it sets counters that it used to
PMC_STATE_DELETED and waits until their use count goes to zero.
amd_intr() wouldn't reload a counter in that state and, thus, a counter
would be allowed to overflow. That means that the counter's value would
be allowed to go outside the expected range.
MFC after: 2 weeks
When hwpmc stops sampling it will set the pm_state to something other
than PMC_STATE_RUNNING. This means the following sequence can happen:
CPU 0: Enter the interrupt handler
CPU 0: Set the thread TDP_CALLCHAIN pflag
CPU 1: Stop sampling
CPU 0: Call pmc_process_samples, sampling is stopped so clears ps_nsamples
CPU 0: Finishes interrupt processing with the TDP_CALLCHAIN flag set
CPU 0: Call pmc_capture_user_callchain to capture the user call chain
CPU 0: Find all the pmc sample are free so no call chains need to be captured
CPU 0: KASSERT because of this
This fixes the issue by checking if any of the samples have been stopped
and including this in te KASSERT.
PR: 204273
Reviewed by: bz, gnn
Obtained from: ABT Systems Ltd
Sponsored by: The FreeBSD Foundation
Differential Revision: https://reviews.freebsd.org/D6581
Currently, Application Processors (non-boot CPUs) are started by
MD code at SI_SUB_CPU, but they are kept waiting in a "pen" until
SI_SUB_SMP at which point they are released to run kernel threads.
SI_SUB_SMP is one of the last SYSINIT levels, so APs don't enter
the scheduler and start running threads until fairly late in the
boot.
This change moves SI_SUB_SMP up to just before software interrupt
threads are created allowing the APs to start executing kernel
threads much sooner (before any devices are probed). This allows
several initialization routines that need to perform initialization
on all CPUs to now perform that initialization in one step rather
than having to defer the AP initialization to a second SYSINIT run
at SI_SUB_SMP. It also permits all CPUs to be available for
handling interrupts before any devices are probed.
This last feature fixes a problem on with interrupt vector exhaustion.
Specifically, in the old model all device interrupts were routed
onto the boot CPU during boot. Later after the APs were released at
SI_SUB_SMP, interrupts were redistributed across all CPUs.
However, several drivers for multiqueue hardware allocate N interrupts
per CPU in the system. In a system with many CPUs, just a few drivers
doing this could exhaust the available pool of interrupt vectors on
the boot CPU as each driver was allocating N * mp_ncpu vectors on the
boot CPU. Now, drivers will allocate interrupts on their desired CPUs
during boot meaning that only N interrupts are allocated from the boot
CPU instead of N * mp_ncpu.
Some other bits of code can also be simplified as smp_started is
now true much earlier and will now always be true for these bits of
code. This removes the need to treat the single-CPU boot environment
as a special case.
As a transition aid, the new behavior is available under a new kernel
option (EARLY_AP_STARTUP). This will allow the option to be turned off
if need be during initial testing. I plan to enable this on x86 by
default in a followup commit in the next few days and to have all
platforms moved over before 11.0. Once the transition is complete,
the option will be removed along with the !EARLY_AP_STARTUP code.
These changes have only been tested on x86. Other platform maintainers
are encouraged to port their architectures over as well. The main
things to check for are any uses of smp_started in MD code that can be
simplified and SI_SUB_SMP SYSINITs in MD code that can be removed in
the EARLY_AP_STARTUP case (e.g. the interrupt shuffling).
PR: kern/199321
Reviewed by: markj, gnn, kib
Sponsored by: Netflix
when its result is immediately ignored, i.e. for kernel processes
forked from the user process. Do not test for non-null before freeing
string.
Sponsored by: The FreeBSD Foundation
MFC after: 2 weeks
The code tracks a counter which is the number of events until the next
sample. On context switch in, it loads the saved counter. On context
switch out, it tries to calculate a new saved counter.
Problems:
1. The saved counter was shared by all threads in a process. However, this
means that all threads would be initially loaded with the same saved
counter. However, that could result in sampling more often than once every
X number of events.
2. The calculation to determine a new saved counter was backwards. It
added when it should have subtracted, and subtracted when it should have
added. Assume a single-threaded process with a reload count of 1000 events.
Assuming the counter on context switch in was 100 and the counter on context
switch out was 50 (meaning the thread has "consumed" 50 more events), the
code would calculate a new saved counter of 150 (instead of the proper 50).
Fix:
1. As soon as the saved counter is used to initialize a monitor for a
thread on context switch in, set the saved counter to the reload count.
That way, subsequent threads to use the saved counter will get the full
reload count, assuring we sample at least once every X number of events
(across all threads).
2. Change the calculation of the saved counter. Due to the change to the
saved counter in #1, we simply need to add (modulo the reload count) the
remaining counter time we retrieve from the CPU when a thread is context
switched out.
Differential Revision: https://reviews.freebsd.org/D4122
Approved by: gnn (mentor)
MFC after: 1 month
Sponsored by: Juniper Networks
Changes to the code to gather user stacks:
* Delay setting pmc_cpumask until we actually have the stack.
* When recording user stack traces, only walk the portion of the ring
that should have samples for us.
Sponsored by: Juniper Networks
Approved by: gnn (mentor)
MFC after: 1 month
Currently, there is a single pm_stalled flag that tracks whether a
performance monitor was "stalled" due to insufficent ring buffer
space for samples. However, because the same performance monitor
can run on multiple processes or threads at the same time, a single
pm_stalled flag that impacts them all seems insufficient.
In particular, you can hit corner cases where the code fails to stop
performance monitors during a context switch out, because it thinks
the performance monitor is already stopped. However, in reality,
it may be that only the monitor running on a different CPU was stalled.
This patch attempts to fix that behavior by tracking on a per-CPU basis
whether a PM desires to run and whether it is "stalled". This lets the
code make better decisions about when to stop PMs and when to try to
restart them. Ideally, we should avoid the case where the code fails
to stop a PM during a context switch out.
Sponsored by: Juniper Networks
Reviewed by: jhb
Approved by: gnn (mentor)
Differential Revision: https://reviews.freebsd.org/D4124
in the parent, we will inherit the pmcids but cannot execute any operations
on them in the child. The reason for this is that pmc_find_pmc() only
tries to find the current process on the owners hash list, but given the
child does not own the attachment, we cannot find it.
Thus, in case the initial lookup fails, try to find the pmc_process state
affiliated with the child process, lookup the pmc from there using the
row index, and get the owner process from that pmc.
Then continue as normal and lookup the pmc context of the owner (process).
This allows us to call, e.g., pmc_start() in the child process before we
start the work there, but to collect the accumulated results later in
the parent.
Sponsored by: DARPA,AFRL
Obtained from: L41
Tested by: rwatson, L41
MFC after: 4 weeks
Reviewed by: gnn
Differential Revision: https://reviews.freebsd.org/D2052
When providing memory map information to userland, populate the vnode pointer
for tmpfs files. Set the memory mapping to appear as a vnode type, to match
FreeBSD 9 behavior.
This fixes the use of tmpfs files with the dtrace pid provider,
procstat -v, procfs, linprocfs, pmc (pmcstat), and ptrace (PT_VM_ENTRY).
Submitted by: Eric Badger <eric@badgerio.us> (initial revision)
Obtained from: Dell Inc.
PR: 198431
MFC after: 2 weeks
Reviewed by: jhb
Approved by: kib (mentor)
In both cases, the the effect of the bug was that a very small positive
number was written to the counter. This means that a large number of
events needed to occur before the next sampling interrupt would trigger.
Even with very frequently occurring events like clock cycles wrapping all
the way around could take a long time. Both bugs occurred when updating
the saved reload count for an outgoing thread on a context switch.
First, the counter-independent code compares the current reload count
against the count set when the thread switched in and generates a delta
to apply to the saved count. If this delta causes the reload counter
to go negative, it would add a full reload interval to wrap it around to
a positive value. The fix is to add the full reload interval if the
resulting counter is zero.
Second, occasionally the raw counter value read during a context switch
has actually wrapped, but an interrupt has not yet triggered. In this
case the existing logic would return a very large reload count (e.g.
2^48 - 2 if the counter had overflowed by a count of 2). This was seen
both for fixed-function and programmable counters on an E5-2643.
Workaround this case by returning a reload count of zero.
PR: 198149
Differential Revision: https://reviews.freebsd.org/D2557
Reviewed by: emaste
MFC after: 1 week
Sponsored by: Norse Corp, Inc.
This removes one of the frequent causes of ABI breakage when new CPU
types are added to hwpmc(4).
Differential Revision: https://reviews.freebsd.org/D2586
Reviewed by: davide, emaste, gnn (earlier version)
MFC after: 2 weeks
go back through HASWELL, IVY_BRIDGE, IVY_BRIDGE_XEON and SANDY_BRIDGE
to straighten out all the missing PMCs. We also add a new pmc tool
pmcstudy, this allows one to run the various formulas from
the documents "Using Intel Vtune Amplifier XE on XXX Generation platforms" for
IB/SB and Haswell. The tool also allows one to postulate your own
formulas with any of the various PMC's. At some point I will enahance
this to work with Brendan Gregg's flame-graphs so we can flamegraph
various PMC interactions. Note the manual page also needs some
work (lots of work) but gnn has committed to help me with that ;-)
Reviewed by: gnn
MFC after:1 month
Sponsored by: Netflix Inc.
consistent with pmc_destroy_owner_descriptor(). Also be sure to destroy
PMCs if a process exits or execs without explicitly releasing them.
Reviewed by: bz, gnn
MFC after: 2 weeks
Sponsored by: EMC / Isilon Storage Division
Differential Revision: https://reviews.freebsd.org/D958
These changes prevent sysctl(8) from returning proper output,
such as:
1) no output from sysctl(8)
2) erroneously returning ENOMEM with tools like truss(1)
or uname(1)
truss: can not get etype: Cannot allocate memory
there is an environment variable which shall initialize the SYSCTL
during early boot. This works for all SYSCTL types both statically and
dynamically created ones, except for the SYSCTL NODE type and SYSCTLs
which belong to VNETs. A new flag, CTLFLAG_NOFETCH, has been added to
be used in the case a tunable sysctl has a custom initialisation
function allowing the sysctl to still be marked as a tunable. The
kernel SYSCTL API is mostly the same, with a few exceptions for some
special operations like iterating childrens of a static/extern SYSCTL
node. This operation should probably be made into a factored out
common macro, hence some device drivers use this. The reason for
changing the SYSCTL API was the need for a SYSCTL parent OID pointer
and not only the SYSCTL parent OID list pointer in order to quickly
generate the sysctl path. The motivation behind this patch is to avoid
parameter loading cludges inside the OFED driver subsystem. Instead of
adding special code to the OFED driver subsystem to post-load tunables
into dynamically created sysctls, we generalize this in the kernel.
Other changes:
- Corrected a possibly incorrect sysctl name from "hw.cbb.intr_mask"
to "hw.pcic.intr_mask".
- Removed redundant TUNABLE statements throughout the kernel.
- Some minor code rewrites in connection to removing not needed
TUNABLE statements.
- Added a missing SYSCTL_DECL().
- Wrapped two very long lines.
- Avoid malloc()/free() inside sysctl string handling, in case it is
called to initialize a sysctl from a tunable, hence malloc()/free() is
not ready when sysctls from the sysctl dataset are registered.
- Bumped FreeBSD version to indicate SYSCTL API change.
MFC after: 2 weeks
Sponsored by: Mellanox Technologies
current lock classes KPI it was really difficult because there was no
way to pass an rmtracker object to the lock/unlock routines. In order
to accomplish the task, modify the aforementioned functions so that
they can return (or pass as argument) an uinptr_t, which is in the rm
case used to hold a pointer to struct rm_priotracker for current
thread. As an added bonus, this fixes rm_sleep() in the rm shared
case, which right now can communicate priotracker structure between
lc_unlock()/lc_lock().
Suggested by: jhb
Reviewed by: jhb
Approved by: re (delphij)
kld_unload event handler which gets invoked after a linker file has been
successfully unloaded. The kld_unload and kld_load event handlers are now
invoked with the shared linker lock held, while kld_unload_try is invoked
with the lock exclusively held.
Convert hwpmc(4) to use these event handlers instead of having
kern_kldload() and kern_kldunload() invoke hwpmc(4) hooks whenever files are
loaded or unloaded. This has no functional effect, but simplifes the linker
code somewhat.
Reviewed by: jhb
future further optimizations where the vm_object lock will be held
in read mode most of the time the page cache resident pool of pages
are accessed for reading purposes.
The change is mostly mechanical but few notes are reported:
* The KPI changes as follow:
- VM_OBJECT_LOCK() -> VM_OBJECT_WLOCK()
- VM_OBJECT_TRYLOCK() -> VM_OBJECT_TRYWLOCK()
- VM_OBJECT_UNLOCK() -> VM_OBJECT_WUNLOCK()
- VM_OBJECT_LOCK_ASSERT(MA_OWNED) -> VM_OBJECT_ASSERT_WLOCKED()
(in order to avoid visibility of implementation details)
- The read-mode operations are added:
VM_OBJECT_RLOCK(), VM_OBJECT_TRYRLOCK(), VM_OBJECT_RUNLOCK(),
VM_OBJECT_ASSERT_RLOCKED(), VM_OBJECT_ASSERT_LOCKED()
* The vm/vm_pager.h namespace pollution avoidance (forcing requiring
sys/mutex.h in consumers directly to cater its inlining functions
using VM_OBJECT_LOCK()) imposes that all the vm/vm_pager.h
consumers now must include also sys/rwlock.h.
* zfs requires a quite convoluted fix to include FreeBSD rwlocks into
the compat layer because the name clash between FreeBSD and solaris
versions must be avoided.
At this purpose zfs redefines the vm_object locking functions
directly, isolating the FreeBSD components in specific compat stubs.
The KPI results heavilly broken by this commit. Thirdy part ports must
be updated accordingly (I can think off-hand of VirtualBox, for example).
Sponsored by: EMC / Isilon storage division
Reviewed by: jeff
Reviewed by: pjd (ZFS specific review)
Discussed with: alc
Tested by: pho
executed. This means past the point where userret() is generally
executed.
Skip the td_pinned check if a callchain tracing is currently happening
and add a more robust check to pmc_capture_user_callchain() in order to
catch td_pinned leak past ast() in hwpmc case.
Reported and tested by: fabient
MFC after: 1 week
X-MFC: r240246
In particular, do not lock Giant conditionally when calling into the
filesystem module, remove the VFS_LOCK_GIANT() and related
macros. Stop handling buffers belonging to non-mpsafe filesystems.
The VFS_VERSION is bumped to indicate the interface change which does
not result in the interface signatures changes.
Conducted and reviewed by: attilio
Tested by: pho
New kernel events can be added at various location for sampling or counting.
This will for example allow easy system profiling whatever the processor is
with known tools like pmcstat(8).
Simultaneous usage of software PMC and hardware PMC is possible, for example
looking at the lock acquire failure, page fault while sampling on
instructions.
Sponsored by: NETASQ
MFC after: 1 month
As the underlying block is 4KB if the PMC throughput is low the measurement
will be reported on the next tick. pmcstat(8) use the modified flush API to
reclaim current buffer before displaying next top.
MFC after: 1 month
patch modifies makesyscalls.sh to prefix all of the non-compatibility
calls (e.g. not linux_, freebsd32_) with sys_ and updates the kernel
entry points and all places in the code that use them. It also
fixes an additional name space collision between the kernel function
psignal and the libc function of the same name by renaming the kernel
psignal kern_psignal(). By introducing this change now we will ease future
MFCs that change syscalls.
Reviewed by: rwatson
Approved by: re (bz)
cpuset_t objects.
That is going to offer the underlying support for a simple bump of
MAXCPU and then support for number of cpus > 32 (as it is today).
Right now, cpumask_t is an int, 32 bits on all our supported architecture.
cpumask_t on the other side is implemented as an array of longs, and
easilly extendible by definition.
The architectures touched by this commit are the following:
- amd64
- i386
- pc98
- arm
- ia64
- XEN
while the others are still missing.
Userland is believed to be fully converted with the changes contained
here.
Some technical notes:
- This commit may be considered an ABI nop for all the architectures
different from amd64 and ia64 (and sparc64 in the future)
- per-cpu members, which are now converted to cpuset_t, needs to be
accessed avoiding migration, because the size of cpuset_t should be
considered unknown
- size of cpuset_t objects is different from kernel and userland (this is
primirally done in order to leave some more space in userland to cope
with KBI extensions). If you need to access kernel cpuset_t from the
userland please refer to example in this patch on how to do that
correctly (kgdb may be a good source, for example).
- Support for other architectures is going to be added soon
- Only MAXCPU for amd64 is bumped now
The patch has been tested by sbruno and Nicholas Esborn on opteron
4 x 12 pack CPUs. More testing on big SMP is expected to came soon.
pluknet tested the patch with his 8-ways on both amd64 and i386.
Tested by: pluknet, sbruno, gianni, Nicholas Esborn
Reviewed by: jeff, jhb, sbruno
This bug would cause allocating sample-mode PMCs to fail with ENOMEM after allocating several counting-mode PMCs.
Approved by: jkoshy (mentor)
MFC after: 2 weeks
for the pmclog.
Reported by: Ryan Stone <rstone at sandvine dot com>
Tested by: Ryan Stone <rstone at sandvine dot com>
Sponsored by: Sandvine Incorporated
time it is marked for user space callchain capture in the NMI
handler and the time the callchain capture callback runs.
- Improve code and control flow clarity by invoking hwpmc(4)'s user
space callchain capture callback directly from low-level code.
Reviewed by: jhb (kern/subr_trap.c)
Testing (various patch revisions): gnn,
Fabien Thomas <fabien dot thomas at netasq dot com>,
Artem Belevich <artemb at gmail dot com>
and Core Duo), models 0xF (Core2), model 0x17 (Core2Extreme) and
model 0x1C (Atom).
In these CPUs, the actual numbers, kinds and widths of PMCs present
need to queried at run time. Support for specific "architectural"
events also needs to be queried at run time.
Model 0xE CPUs support programmable PMCs, subsequent CPUs
additionally support "fixed-function" counters.
- Use event names that are close to vendor documentation, taking in
account that:
- events with identical semantics on two or more CPUs in this family
can have differing names in vendor documentation,
- identical vendor event names may map to differing events across
CPUs,
- each type of CPU supports a different subset of measurable
events.
Fixed-function and programmable counters both use the same vendor
names for events. The use of a class name prefix ("iaf-" or
"iap-" respectively) permits these to be distinguished.
- In libpmc, refactor pmc_name_of_event() into a public interface
and an internal helper function, for use by log handling code.
- Minor code tweaks: staticize a global, freshen a few comments.
Tested by: gnn
