change some busy wait delays to tsleep
Previously tsleep was avoided in usleep_range(), drm_msleep() and
fsleep() on precision grounds. That shouldn't be a problem after
kern_synch.c rev 1.225.
This change prevents inteldrm GuC initialisation errors seen on
Alder Lake by daniel@ and rsadowski@
Problem found by and initial patch from daniel@
profil(2): release the dirfd reference, and publish pr_cdir once
Fixes both leaks reported by Acts1631, plus the race that fixing the
first one exposes.
- Take the vnode reference while the file reference is still held, then
FRELE() it.
- Acquire pr_ucred together with pr_cdir, after validation, so an
EBADF/ENOTDIR return cannot leave a cred held with pr_cdir still NULL
- Re-test pr_cdir before publishing to and release the loser's resource
deraadt@ Looks good
Do a prefix re-evaluation if the validation state of a route changes.
The validation state of a route does not change the route evaluation
but if the state changes the result of the outbound filters may change.
Because of this do the same dance we do for the filtered flag and
re-evaluate the prefix.
Move the validation state also into the middle of the re-evaluation dance
even though it is not strictly needed. At least that way it is more
consistent with the rest of the code.
Reported by N0zoM1z0
OK tb@
sys/arm: avoid cache faults on untranslated aliases
An executable PV entry does not establish a usable hardware translation;
for an unreferenced prefaulted page, pmap_enter() can install an L2 PTE
with its access flag clear while leaving the L1 descriptor invalid.
pmap_clean_page() nevertheless uses such aliases for cache maintenance;
during address space teardown, a translation abort can reenter
uvm_fault() while the current thread owns the map write lock, provoking
recursive acquisition of vmmaplk.
Inspect each executable alias's L1 descriptor before issuing cache
maintenance by virtual address; if it is invalid, use
cpu_idcache_wbinv_all() and return, since the operation covers the
remaining aliases. This preserves data cache cleaning and instruction
cache invalidation without translating the unusable address; aliases
with valid L1 descriptors retain the existing per page operation.
This idea follow discovered discussion in linux-arm-kernel:
https://lists.infradead.org/pipermail/linux-arm-kernel/2018-January/556458.html
[4 lines not shown]
iked: Remove invalid SPI size 8 in delete path
DELETE does not provide the SPI of the IKE SA to be deleted. It
is the IKE SA that is used to send the DELETE notify. Therefore,
SPI size 8 is not valid. With the previous change, such DELETE
notifies will be dropped. Therefore remove the now dead code.
See https://www.rfc-editor.org/rfc/rfc7296.html#section-3.11
ok markus@
vmm(4): Do not access raw PSL_I
With SEV-ES enabled we can not directly access PSL_I in RFLAGS.
Instead we have to use the GUEST_INTR_MASK bit in vmcb->v_intr_shadow
which indicates the actual PSL_I state.
ok dv@ mlarkin@
rpki-client: rename valid_x509() to valid_cert()
x509 usually indicates an X509 * argument and cert a struct cert *.
This got changed in the surgery today and valid_cert() is now free,
so we can fix this.
ok claudio
rpki-client: check all !TA certs for resource coverage
For historic reasons, EE certs did not have their resources checked for
coverage in the parent. Since we have the exception of EE certs that need
special casing for inheritance (MFT/TAK/GBR), this can be done at the end
of valid_x509(). The check is not expensive since this only goes up as
many issuers as needed to hit explicit resources.
Rename the misnamed valid_cert() to valid_resources() for lack of a better
idea and call it from the end of valid_x509(). Remove the remaining calls
since they're now redundant, and make valid_resources() static.
ok job
uaudio: Make sure the reported format and channels are usable
A bogus device may trigger a division by zero if the number of
channels or sample size are 0. Similarly, memory corruption may occur
if the reported sample size is larger than four bytes.
Found by Acts1631 <acts1631kjv at proton.me>, who proposed a diff that
inspired this fix. Thanks!
The MTP firmware needs to stay around after "loading" it; the comprocessor
that runs the firmware seems to reference the firmware image that we pass
it directly instead of making its own copy. This fixes a regression on
laptops with a dockchannel touchpad after the recent changes to apldart(4).
ok tobhe@
As well as tracking modified (dirty) lines during a synchronized update,
also track scrolled lines. This means they can be scrolled as one
instead of forcing each line to be redrawn - much less expensive for a
very common operation. From Ben Maurer in GitHub issue 5611.
rpki-client: match CRLDP and referring Manifest for all !TA certs
Now that crl->mftcrldp is always set, we can move the comparison of
cert->crl with crl->mftdp into valid_x509() so that all certs which
come through here must satisfy this requirement. While TA certs are
not validated here, add a check for cert->purpose to avoid potential
NULL accesses.
ok job
rpki-client: teach filemode about crl->mftcrldp
The following commit relies on mftcrldp being set on every crl.
In filemode use the repo from which the CRL was loaded.
ok job
key= in the signify gzip comment used to include the path, but this was
dropped in signify.c:1.137 so that only bare filenames are used. teach
pkg_add about this so that the signer is still recorded in +CONTENTS
for installed packages. from espie
uaudio: Handle bogus endpoint intervals
Fixes possible division by zero that bogus uaudio hardware may cause.
If the interval is 0 (not allowed for isoc endpoints), assume that the
descriptor is not filled and use 1ms, which works on all devices.
Skip endpoints with intervals of more than 10ms, which make no sense
for audio.
Found by Acts1631 <acts1631kjv at proton.me>
sndiod: Don't abort if a MIDI endpoint connects with non-empty buffers
When the server.thru control is set, the midi endpoints may have
non-empty buffers, which would cause the server to abort.
Found and fixed by Acts1631 <acts1631kjv at proton.me>, thanks!
