The global array has prohibitive performance impact on multicore systems.
The same data (and more) can be obtained with dtrace.
Reviewed by: markj
Differential Revision: https://reviews.freebsd.org/D27199
Restart syscalls and some sync operations when filesystem indicated
ERELOOKUP condition, mostly for VOPs operating on metdata. In
particular, lookup results cached in the inode/v_data is no longer
valid and needs recalculating. Right now this should be nop.
Assert that ERELOOKUP is catched everywhere and not returned to
userspace, by asserting that td_errno != ERELOOKUP on syscall return
path.
In collaboration with: pho
Reviewed by: mckusick (previous version), markj
Tested by: markj (syzkaller), pho
Sponsored by: The FreeBSD Foundation
Differential revision: https://reviews.freebsd.org/D26136
The routine does not serve any practical purpose.
Memory can be allocated in many other ways and most consumers pass the
M_WAITOK flag, making malloc not fail in the first place.
Reviewed by: markj
Differential Revision: https://reviews.freebsd.org/D27143
This works for amd64, but none others -- drop it, because we already have a
proper definition in sys/compat/freebsd32/freebsd32.h that correctly uses
time32_t.
MFC after: 1 week
This gets rid of the most contended spinlock seen when creating/destroying
threads in a loop. (modulo kstack)
Tested by: alfredo (ppc64), bdragon (ppc64)
First, funsetownlst() list looks at the first element of the list to see
whether it's processing a process or a process group list. Then it
acquires the global sigio lock and processes the list. However, nothing
prevents the first sigio tracker from being freed by a concurrent
funsetown() before the sigio lock is acquired.
Fix this by acquiring the global sigio lock immediately after checking
whether the list is empty. Callers of funsetownlst() ensure that new
sigio trackers cannot be added concurrently.
Second, fsetown() uses funsetown() to remove an existing sigio structure
from a file object. However, funsetown() uses a racy check to avoid the
sigio lock, so two threads may call fsetown() on the same file object,
both observe that no sigio tracker is present, and enqueue two sigio
trackers for the same file object. However, if the file object is
destroyed, funsetown() will only remove one sigio tracker, and
funsetownlst() may later trigger a use-after-free when it clears the
file object reference for each entry in the list.
Fix this by introducing funsetown_locked(), which avoids the racy check.
Reviewed by: kib
Reported by: pho
Tested by: pho
MFC after: 1 week
Sponsored by: The FreeBSD Foundation
Differential Revision: https://reviews.freebsd.org/D27157
Note this still does not scale but is enough to move it out of the way
for the foreseable future.
In particular a trivial benchmark spawning/killing threads stops contesting
on tidhash.
This in particular unbreaks rtkit.
The limitation was a leftover of previous state, to quote a
comment:
/*
* Though lwpid is unique, only current process is supported
* since there is no efficient way to look up a LWP yet.
*/
Long since then a global tid hash was introduced to remedy
the problem.
Permission checks still apply.
Submitted by: greg_unrelenting.technology (Greg V)
Differential Revision: https://reviews.freebsd.org/D27158
There are workloads with very bursty tid allocation and since unr tries very
hard to have small-sized bitmaps it keeps reallocating memory. Just doing
buildkernel gives almost 150k calls to free coming from unr.
This also gets rid of the hack which tried to postpone TID reuse.
Reviewed by: kib, markj
Tested by: pho
Differential Revision: https://reviews.freebsd.org/D27101
The intent is to replace the current id allocation method and a known upper
bound will be useful.
Reviewed by: kib (previous version), markj (previous version)
Tested by: pho
Differential Revision: https://reviews.freebsd.org/D27100
imgact_binmisc matches magic/mask from imgp->image_header, which is only a
single page in size mapped from the first page of an image. One can specify
an interpreter that matches on, e.g., --offset 4096 --size 256 to read up to
256 bytes past the mapped first page.
The limitation is that we cannot specify a magic string that exceeds a
single page, and we can't allow offset + size to exceed a single page
either. A static assert has been added in case someone finds it useful to
try and expand the size, but it does seem a little unlikely.
While this looks kind of exploitable at a sideways squinty-glance, there are
a couple of mitigating factors:
1.) imgact_binmisc is not enabled by default,
2.) entries may only be added by the superuser,
3.) trying to exploit this information to read what's mapped past the end
would be worse than a root canal or some other relatably painful
experience, and
4.) there's no way one could pull this off without it being completely
obvious.
The first page is mapped out of an sf_buf, the implementation of which (or
lack thereof) depends on your platform.
MFC after: 1 week
access the socket send or receive buffer. This is not possible for
listening sockets since r319722.
Because send()/recv() calls fail on listening sockets, fail also ioctl()
indicating EINVAL.
PR: 250366
Reported by: Yong-Hao Zou
Reviewed by: glebius, rscheff
MFC after: 1 week
Sponsored by: Netflix, Inc.
Differential Revision: https://reviews.freebsd.org/D26897
The offset we need to account for in the interpreter string comes in two
variants:
1. Fixed - macros other than #a that will not vary from invocation to
invocation
2. Variable - #a, which is substitued with the argv0 that we're replacing
Note that we don't have a mechanism to modify an existing entry. By
recording both of these offset requirements when the interpreter is added,
we can avoid some unnecessary calculations in the exec path.
Most importantly, we can know up-front whether we need to grab
calculate/grab the the filename for this interpreter. We also get to avoid
walking the string a first time looking for macros. For most invocations,
it's a swift exit as they won't have any, but there's no point entering a
loop and searching for the macro indicator if we already know there will not
be one.
While we're here, go ahead and only calculate the argv0 name length once per
invocation. While it's unlikely that we'll have more than one #a, there's no
reason to recalculate it every time we encounter an #a when it will not
change.
I have not bothered trying to benchmark this at all, because it's arguably a
minor and straightforward/obvious improvement.
MFC after: 1 week
This adds a dedicated counter updated with atomics when INVARIANTS
is used. As a side effect one can reliably determine the lock is held
for reading by at least one thread, but it's still not possible to
find out whether curthread has the lock in said mode.
This should be good enough in practice.
Problem spotted by avg.
This doesn't change anything at the moment since the out-of-order elements
were a pair of uint32_t, but future additions may have caused unnecessary
padding by following the existing precedent.
MFC after: 1 week
If we hadn't been traced in the first place when syscallenter()
started executing, we can ignore TDB_USERWR. TDB_USERWR can get set,
sure, but if it does, it's because the debugger raced with the syscall,
and it cannot depend on winning that race.
Reviewed by: kib
MFC after: 2 weeks
Sponsored by: EPSRC
Differential Revision: https://reviews.freebsd.org/D26585
This module handles relatively few execs (initial qemu-user-static, then
qemu-user-static handles exec'ing itself for binaries it's already running),
but all execs pay the price of at least taking the relatively expensive
sx/slock to check for a match when this module is loaded. Future work will
almost certainly swap this out for another lock, perhaps an rmslock.
The RLOCK/WLOCK phrasing was chosen based on what the callers are really
wanting, rather than using the verbiage typically appropriate for an sx.
MFC after: 1 week
We may want to reserve bits in the future for kernel-only use, so start
rejecting any that aren't the two that we're currently expecting from
userland.
MFC after: 1 week
Previously, non-preemptible epochs could not check; in_epoch() would always
fail, usually because non-preemptible epochs don't imply THREAD_NO_SLEEPING.
For default epochs, it's easy enough to verify that we're in the given
epoch: if we're in a critical section and our record for the given epoch
is active, then we're in it.
This patch also adds some additional INVARIANTS bookkeeping. Notably, we set
and check the recorded thread in epoch_enter/epoch_exit to try and catch
some edge-cases for the caller. It also checks upon freeing that none of the
records had a thread in the epoch, which may make it a little easier to
diagnose some improper use if epoch_free() took place while some other
thread was inside.
This version differs slightly from what was just previously reviewed by the
below-listed, in that in_epoch() will assert that no CPU has this thread
recorded even if it *is* currently in a critical section. This is intended
to catch cases where the caller might have somehow messed up critical
section nesting, we can catch both if they exited the critical section or if
they exited, migrated, then re-entered (on the wrong CPU).
Reviewed by: kib, markj (both previous version)
MFC after: 1 week
Differential Revision: https://reviews.freebsd.org/D27098
Notably, streamline error paths through the existing 'done' label, making it
easier to quickly verify correct cleanup.
Future work might add a kernel-only flag to indicate that a interpreter uses
#a. Currently, all executions via imgact_binmisc pay the penalty of
constructing sname/fname, even if they will not use it. qemu-user-static
doesn't need it, the stock rc script for qemu-user-static certainly doesn't
use it, and I suspect these are the vast majority of (if not the only)
current users.
MFC after: 1 week
According to code comments the original motivation was to allow for
malloc_type_internal changes without ABI breakage. This can be trivially
accomplished by providing spare fields and versioning the struct, as
implemented in the patch below.
The upshots are one less memory indirection on each alloc and disappearance
of mt_zone.
Reviewed by: markj
Differential Revision: https://reviews.freebsd.org/D27104
This ensures that no writes are pending in memory, either metadata or
user data, but not including dirty pages not yet converted to fs writes.
Only filesystems declared local are suspended.
Note that this does not guarantee absence of the metadata errors or
leaks if resume is not done: for instance, on UFS unlinked but opened
inodes are leaked and require fsck to gc.
Reviewed by: markj
Discussed with: imp
Tested by: imp (previous version), pho
Sponsored by: The FreeBSD Foundation
MFC after: 2 weeks
Differential revision: https://reviews.freebsd.org/D27054
Sample usage: kernel modules can decide whether to stick to malloc or
create their own zone.
Reviewed by: markj
Differential Revision: https://reviews.freebsd.org/D27097
The 2 provided zones had inconsistent naming between each other
("int" and "64") and other allocator zones (which use bytes).
Follow malloc by naming them "pcpu-" + size in bytes.
This is a step towards replacing ad-hoc per-cpu zones with
general slabs.
On a sample box vmstat -z shows:
ITEM SIZE LIMIT USED FREE REQ
64: 64, 0, 1043784, 4367538,3698187229
selfd: 64, 0, 1520, 13726,182729008
But at the same time:
vm.uma.selfd.keg.domain.1.pages: 121
vm.uma.selfd.keg.domain.0.pages: 121
Thus 242 pages got pulled even though the malloc zone would likely accomodate
the load without using extra memory.
- Removed a bunch of redundant headers
- Don't explicitly initialize to 0
- The !error check prior to setting imgp->interpreter_name is redundant, all
error paths should and do return or go to 'done'. We have larger problems
otherwise.
Linux allows polling without any events specified and it happens to be the case
in FreeBSD as well. POLLHUP has to be delivered regardless of the event mask
and this works fine if the condition is already present. However, if it is
missing, selrecord is only called if the eventmask has relevant bits set. This
in particular leads to a conditon where pipe_poll can return 0 events and
neglect to selrecord, while kern_poll takes it as an indication it has to go to
sleep, but then there is nobody to wake it up.
While the problem seems systemic to *_poll handlers the least we can do is fix
it up for pipes.
Reported by: Jeremie Galarneau <jeremie.galarneau at efficios.com>
Reviewed by: kib
Differential Revision: https://reviews.freebsd.org/D27094
