Each malloc/free was testing dtrace_malloc_enabled and forcing
extra reads from the malloc type struct to see if perhaps a
dtmalloc probe was on.
Treat it like lockstat and sdt: have a global bolean.
malloc was showing at the top of profile during while running microbenchmarks.
#define DTMALLOC_PROBE_MAX 2
struct malloc_type_internal {
uint32_t mti_probes[DTMALLOC_PROBE_MAX];
u_char mti_zone;
struct malloc_type_stats mti_stats[MAXCPU];
};
Reading mti_zone it wastes a cacheline to hold mti_probes + mti_zone
(which we know is 0) + part of malloc stats of the first cpu which on top
induces false-sharing.
In particular will-it-scale lock1_processes -t 128 -s 10:
before: average:45879692
after: average:51655596
Note the counters can be padded but the right fix is to move them to
counter(9), leaving the struct read-only after creation (modulo dtrace
probes).
This patch adds a new socket option, SO_REUSEPORT_LB, which allow multiple
programs or threads to bind to the same port and incoming connections will be
load balanced using a hash function.
Most of the code was copied from a similar patch for DragonflyBSD.
However, in DragonflyBSD, load balancing is a global on/off setting and can not
be set per socket. This patch allows for simultaneous use of both the current
SO_REUSEPORT and the new SO_REUSEPORT_LB options on the same system.
Required changes to structures
Globally change so_options from 16 to 32 bit value to allow for more options.
Add hashtable in pcbinfo to hold all SO_REUSEPORT_LB sockets.
Limitations
As DragonflyBSD, a load balance group is limited to 256 pcbs
(256 programs or threads sharing the same socket).
Submitted by: Johannes Lundberg <johanlun0@gmail.com>
Sponsored by: Limelight Networks
Differential Revision: https://reviews.freebsd.org/D11003
This combined with previous changes significantly depessimizes the behaviour
under contentnion.
In particular the lock1_processes test (locking/unlocking separate files)
from the will-it-scale suite was executed with 128 concurrency on a
4-socket Broadwell with 128 hardware threads.
Operations/second (lock+unlock) go from ~750000 to ~45000000 (6000%)
For reference single-process is ~1680000 (i.e. on stock kernel the resulting
perf is less than *half* of the single-threaded run),
Note this still does not really scale all that well as the locks were just
bolted on top of the current implementation. Significant room for improvement
is still here. In particular the top performance fluctuates depending on the
extent of false sharing in given run (which extends beyond the file).
Added chain+lock pairs were not padded w.r.t. cacheline size.
One big ticket item is the hash used for spreading threads: it used to be the
process pid (which basically serialized all threaded ops). Temporarily the
vnode addr was slapped in instead.
Tested by: pho
kproc_suspend_check. In r329612 bufspacedaemon was turned into a thread
of the bufdaemon process causing both to call kproc_suspend_check with the
same proc argument and that function contains the following while loop:
while (SIGISMEMBER(p->p_siglist, SIGSTOP)) {
wakeup(&p->p_siglist);
msleep(&p->p_siglist, &p->p_mtx, PPAUSE, "kpsusp", 0);
}
So one thread wakes up the other and the other wakes up the first again,
locking up UP machines on shutdown.
Also register the shutdown handlers with SHUTDOWN_PRI_LAST + 100 so they
run after the syncer has shutdown, because the syncer can cause a
situation where bufdaemon help is needed to proceed.
PR: 227404
Reviewed by: kib
Tested by: cy, rmacklem
1. check if P_ADVLOCK is already set and if so, don't lock to set it
(stolen from DragonFly)
2. when trying for fast path unlock, check that we are doing unlock
first instead of taking the interlock for no reason (e.g. if we want
to *lock*). whilere make it more likely that falling fast path will
not take the interlock either by checking for state
Note the code is severely pessimized both single- and multithreaded.
these assumptions may not hold true once we've panic'd. Therefore, the
checks hold less value after a panic. Additionally, if one of the checks
fails while we are already panic'd, this creates a double-panic which can
interfere with debugging the original panic.
Therefore, this commit allows an administrator to suppress a response to
KASSERT checks after a panic by setting a tunable/sysctl. The
tunable/sysctl (debug.kassert.suppress_in_panic) defaults to being
enabled.
Reviewed by: kib
Sponsored by: Netflix, Inc.
Differential Revision: https://reviews.freebsd.org/D12920
-> PROC_PDEATHSIG_STATUS for consistency with other procctl(2)
operations names.
Requested by: emaste
Sponsored by: The FreeBSD Foundation
MFC after: 13 days
The lock is required to ensure that the switch to the new credentials
and the transfer of the process's accounting data from the old
credentials to the new ones is done atomically. Otherwise, some updates
may be applied to the new credentials and then additionally transferred
from the old credentials if the updates happen after proc_set_cred() and
before racct_proc_ucred_changed().
The problem is especially pronounced for RACCT_RSS because
- there is a strict accounting for this resource (it's reclaimable)
- it's updated asynchronously by the vm daemon
- it's updated by setting an absolute value instead of applying a delta
I had to remove a call to rctl_proc_ucred_changed() from
racct_proc_ucred_changed() and make all callers of latter call the
former as well. The reason is that rctl_proc_ucred_changed, as it is
implemented now, cannot be called while holding the proc lock, so the
lock is dropped after calling racct_proc_ucred_changed. Additionally,
I've added calls to crhold / crfree around the rctl call, because
without the proc lock there is no gurantee that the new credentials,
owned by the process, will stay stable. That does not eliminate a
possibility that the credentials passed to the rctl will get stale.
Ideally, rctl_proc_ucred_changed should be able to work under the proc
lock.
Many thanks to kib for pointing out the above problems.
PR: 222027
Discussed with: kib
No comment: trasz
MFC after: 2 weeks
Differential Revision: https://reviews.freebsd.org/D15048
Remove auxarg_size as it was only used once right after a confusing
assignment in each of the variants of exec_copyout_strings().
Reviewed by: emaste
MFC after: 1 month
Differential Revision: https://reviews.freebsd.org/D15123
Allow processes to request the delivery of a signal upon death of
their parent process. Supposed consumer of the feature is PostgreSQL.
Submitted by: Thomas Munro
Reviewed by: jilles, mjg
MFC after: 2 weeks
Differential revision: https://reviews.freebsd.org/D15106
fget_cap() tries to do a cheaper snapshot of a file descriptor without
holding the file descriptor lock. This snapshot does not do a deep
copy of the ioctls capability array, but instead uses a different
return value to inform the caller to retry the copy with the lock
held. However, filecaps_copy() was returning 1 to indicate that a
retry was required, and fget_cap() was checking for 0 (actually
'!filecaps_copy()'). As a result, fget_cap() did not do a deep copy
of the ioctls array and just reused the original pointer. This cause
multiple file descriptor entries to think they owned the same pointer
and eventually resulted in duplicate frees.
The only code path that I'm aware of that triggers this is to create a
listen socket that has a restricted list of ioctls and then call
accept() which calls fget_cap() with a valid filecaps structure from
getsock_cap().
To fix, change the return value of filecaps_copy() to return true if
it succeeds in copying the caps and false if it fails because the lock
is required. I find this more intuitive than fixing the caller in
this case. While here, change the return type from 'int' to 'bool'.
Finally, make filecaps_copy() more robust in the failure case by not
copying any of the source filecaps structure over. This avoids the
possibility of leaking a pointer into a structure if a similar future
caller doesn't properly handle the return value from filecaps_copy()
at the expense of one more branch.
I also added a test case that panics before this change and now passes.
Reviewed by: kib
Discussed with: mjg (not a fan of the extra branch)
MFC after: 1 week
Differential Revision: https://reviews.freebsd.org/D15047
Always take the AST path rather than calling MD functions which are
often implemented as always failing. The is the case on amd64, arm,
i386, and powerpc. This optimization (inherited from 4.4 Lite) is a
pessimization on those architectures and is the sole use of these
functions. They will be removed in a seperate commit.
Reviewed by: kib
Sponsored by: DARPA, AFRL
Differential Revision: https://reviews.freebsd.org/D15101
This behavior is already documented by the man page, and suggested by POSIX.
Reviewed by: jhb
MFC after: 3 weeks
Differential Revision: https://reviews.freebsd.org/D15099
The change makes the user and kernel address spaces on i386
independent, giving each almost the full 4G of usable virtual addresses
except for one PDE at top used for trampoline and per-CPU trampoline
stacks, and system structures that must be always mapped, namely IDT,
GDT, common TSS and LDT, and process-private TSS and LDT if allocated.
By using 1:1 mapping for the kernel text and data, it appeared
possible to eliminate assembler part of the locore.S which bootstraps
initial page table and KPTmap. The code is rewritten in C and moved
into the pmap_cold(). The comment in vmparam.h explains the KVA
layout.
There is no PCID mechanism available in protected mode, so each
kernel/user switch forth and back completely flushes the TLB, except
for the trampoline PTD region. The TLB invalidations for userspace
becomes trivial, because IPI handlers switch page tables. On the other
hand, context switches no longer need to reload %cr3.
copyout(9) was rewritten to use vm_fault_quick_hold(). An issue for
new copyout(9) is compatibility with wiring user buffers around sysctl
handlers. This explains two kind of locks for copyout ptes and
accounting of the vslock() calls. The vm_fault_quick_hold() AKA slow
path, is only tried after the 'fast path' failed, which temporary
changes mapping to the userspace and copies the data to/from small
per-cpu buffer in the trampoline. If a page fault occurs during the
copy, it is short-circuit by exception.s to not even reach C code.
The change was motivated by the need to implement the Meltdown
mitigation, but instead of KPTI the full split is done. The i386
architecture already shows the sizing problems, in particular, it is
impossible to link clang and lld with debugging. I expect that the
issues due to the virtual address space limits would only exaggerate
and the split gives more liveness to the platform.
Tested by: pho
Discussed with: bde
Sponsored by: The FreeBSD Foundation
MFC after: 1 month
Differential revision: https://reviews.freebsd.org/D14633
Now that 10 years have passed since the original limit of 10000 was
committed, bump it a little bit.
Spinning waiting for writers is semi-informed in the sense that we always
know if the owner is running and base the decision to spin on that.
However, no such information is provided for read-locking. In particular
this means that it is possible for a write-spinner to completely waste cpu
time waiting for the lock to be released, while the reader holding it was
preempted and is now waiting for the spinner to go off cpu.
Nonetheless, in majority of cases it is an improvement to spin instead of
instantly giving up and going to sleep.
The current approach is pretty simple: snatch the number of current readers
and performs that many pauses before checking again. The total number of
pauses to execute is limited to 10k. If the lock is still not free by
that time, go to sleep.
Given the previously noted problem of not knowing whether spinning makes
any sense to begin with the new limit has to remain rather conservative.
But at the very least it should also be related to the machine. Waiting
for writers uses parameters selected based on the number of activated
hardware threads. The upper limit of pause instructions to be executed
in-between re-reads of the lock is typically 16384 or 32678. It was
selected as the limit of total spins. The lower bound is set to
already present 10000 as to not change it for smaller machines.
Bumping the limit reduces system time by few % during benchmarks like
buildworld, buildkernel and others. Tested on 2 and 4 socket machines
(Broadwell, Skylake).
Figuring out how to make a more informed decision while not pessimizing
the fast path is left as an exercise for the reader.
Sometimes the values contain geli passphrases being communicated from
loader(8) to the kernel, and some day the compiler may decide to start
eliding calls to memset() for a pointer which is not dereferenced again
before being passed to free().
This allows NIC drivers to sleep on polling config operations.
Submitted by: Matthew Macy <mmacy@mattmacy.io>
Reviewed by: shurd
Sponsored by: Limelight Networks
Differential Revision: https://reviews.freebsd.org/D14982
Previous limits were chosen when locking primitives had spurious lock
accesses.
Flipping the starting point to 1 (or rather 2 as the first call shifts it)
provides a modest win when mild contention is seen while not hurting worse
cases. Tested on a bunch of one, two and four socket old and new systems
(Westmere, Skylake, Threadreaper and others) by doing concurrent page faults,
buildkernel/buildworld and other stuff (although not all systems got all the
tests).
Another thing is the upper limit. It is semi-arbitrarily chosen as it was
getting out of hand for slightly less small systems (e.g. a 128-thread one).
Note that backoff is fundamentally a speculative bandaid and this change just
makes it fit a little bit better. It remains completely oblivious to the
hardware topology or the contention pattern. This is being experimented with.
opt_compat.h is mentioned in nearly 180 files. In-progress network
driver compabibility improvements may add over 100 more so this is
closer to "just about everywhere" than "only some files" per the
guidance in sys/conf/options.
Keep COMPAT_LINUX32 in opt_compat.h as it is confined to a subset of
sys/compat/linux/*.c. A fake _COMPAT_LINUX option ensure opt_compat.h
is created on all architectures.
Move COMPAT_LINUXKPI to opt_dontuse.h as it is only used to control the
set of compiled files.
Reviewed by: kib, cem, jhb, jtl
Sponsored by: DARPA, AFRL
Differential Revision: https://reviews.freebsd.org/D14941
Modify makesyscalls.sh to strip out SAL annotations.
No functional change.
This is based on work I started in CheriBSD and use to validate fat
pointers at the syscall boundary. Tal Garfinkel reviewed the changes,
added annotations to COMPAT* syscalls and is using them in a record and
playback framework. One can envision other uses such as a WITNESS-like
validator for copyin/out as speculated on in the review.
As this time we are only annotating sys/kern/syscalls.master as that is
sufficient for userspace work. If kernel use cases materialize, we can
annotate other syscalls.master as needed.
Submitted by: Tal Garfinkel <talg@cs.stanford.edu>
Sponsored by: DARPA, AFRL (in part)
Differential Revision: https://reviews.freebsd.org/D14285
Add a new "interleave" allocation policy which stripes pages across
domains with a stride or width keeping contiguity within a multi-page
region.
Move the kernel to the dedicated numbered cpuset #2 making it possible
to assign kernel threads and memory policy separately from user. This
also eliminates the need for the complicated interrupt binding code.
Add a sysctl API for viewing and manipulating domainsets. Refactor some
of the cpuset_t manipulation code using the generic bitset type so that
it can be used for both. This probably belongs in a dedicated subr file.
Attempt to improve the include situation.
Reviewed by: kib
Discussed with: jhb (cpuset parts)
Tested by: pho (before review feedback)
Sponsored by: Netflix, Dell/EMC Isilon
Differential Revision: https://reviews.freebsd.org/D14839
platforms. Original commit message as follows:
Only use CPUs in the domain the device is attached to for default
assignment. Device drivers are able to override the default assignment
if they bind directly. There are severe performance penalties for
handling interrupts on remote CPUs and this should only be done in
very controlled circumstances.
Reviewed by: jhb, kib
Tested by: pho
Sponsored by: Netflix, Dell/EMC Isilon
Differential Revision: https://reviews.freebsd.org/D14838
It's not sufficient nor required to use the vnode interlock when
checking if we are going to drop the last use count as the code in
vputx() uses refcount (atomic) operations for both checking and
decrementing the use code. Apply the same method to vn_rele_async().
While here, remove vn_rele_inactive(), a wrapper around vrele() that
didn't add any value.
Also, the change required making vfs_refcount_release_if_not_last()
public. I've made vfs_refcount_acquire_if_not_zero() public as well.
They are in sys/refcount.h now. While making the move I've dropped the
vfs_ prefix.
Reviewed by: mjg
MFC after: 2 weeks
Sponsored by: Panzura
Differential Revision: https://reviews.freebsd.org/D14869
It is random collection of fixes for issues not yet corrected,
reported at https://tsyrklevi.ch/clang_analyzer/freebsd_013017/. Many
issues from that list were already corrected. Most of them are for
compat32, old compat32 or affect both primary host ABI and compat32.
The freebsd32_kldstat(), for instance, was already fixed by using
malloc(M_ZERO). Patch includes correction to report the supplied
version back, which is just pedantic.
Reviewed by: brooks, emaste (previous version)
Sponsored by: The FreeBSD Foundation
MFC after: 1 week
Differential revision: https://reviews.freebsd.org/D14868
Include _uio.h instead of uio.h in several headers to reduce header
polution.
Fix a few places that relied on header polution to get the uio.h header.
I have not moved struct uio as many more things that use it rely on
header polution to get other definitions from uio.h.
Reviewed by: cem, kib, markj
Sponsored by: DARPA, AFRL
Differential Revision: https://reviews.freebsd.org/D14811
If the operation is not an update, if neither r/w nor r/o mode is
explicitly requested, if the error code hints at the possibility of the
media being read-only, and if the fallback is allowed, then we can try
to automatically downgrade to the readonly mode.
This is especially useful for auto-mounting of removable media that
sometimes can happen to be write-protected.
The fallback to r/o is not enabled by default. It can be requested on a
per-mount basis with a new mount option, 'autoro'. Or it can be
globally allowed by setting vfs.default_autoro.
Reviewed by: cem, kib
MFC after: 3 weeks
Relnotes: yes
Differential Revision: https://reviews.freebsd.org/D13361
We may remove a sleepqueue from the hash table in
sleepq_resume_thread().
Reviewed by: kib
MFC after: 3 days
Sponsored by: Dell EMC Isilon
Differential Revision: https://reviews.freebsd.org/D14847
Assert that all such memory is unwired on return to usermode.
The count of the wired memory will be used to detect the copyout mode.
Tested by: pho (as part of the larger patch)
Sponsored by: The FreeBSD Foundation
MFC after: 1 week
Most important for the future use, do not call
vm_fault_quick_hold_pages() with disabled pagefaults.
Reported and tested by: pho (as part of the larger patch)
Sponsored by: The FreeBSD Foundation
MFC after: 1 week
the task context.
shutdown_nice() is used from the fast interrupt handlers, mostly for
console drivers, where we cannot lock blockable locks. Schedule the
task in the fast queue to send the signal from the proper context.
Reviewed by: imp
Discussed with: bde
Sponsored by: The FreeBSD Foundation
MFC after: 1 week
summits at BSDCan and BSDCam in 2017.
The TCP Blackbox Recorder allows you to capture events on a TCP connection
in a ring buffer. It stores metadata with the event. It optionally stores
the TCP header associated with an event (if the event is associated with a
packet) and also optionally stores information on the sockets.
It supports setting a log ID on a TCP connection and using this to correlate
multiple connections that share a common log ID.
You can log connections in different modes. If you are doing a coordinated
test with a particular connection, you may tell the system to put it in
mode 4 (continuous dump). Or, if you just want to monitor for errors, you
can put it in mode 1 (ring buffer) and dump all the ring buffers associated
with the connection ID when we receive an error signal for that connection
ID. You can set a default mode that will be applied to a particular ratio
of incoming connections. You can also manually set a mode using a socket
option.
This commit includes only basic probes. rrs@ has added quite an abundance
of probes in his TCP development work. He plans to commit those soon.
There are user-space programs which we plan to commit as ports. These read
the data from the log device and output pcapng files, and then let you
analyze the data (and metadata) in the pcapng files.
Reviewed by: gnn (previous version)
Obtained from: Netflix, Inc.
Relnotes: yes
Differential Revision: https://reviews.freebsd.org/D11085
The object lock was only needed when attempting to free B_DIRECT
buffer pages, and for testing for invalid pages (and freeing them
if so). Handle the latter by instead moving invalid pages near the head
of the inactive queue, where they will be reclaimed quickly.
Reviewed by: alc, kib, jeff
MFC after: 3 weeks
Differential Revision: https://reviews.freebsd.org/D14778
