auto-partitioning of /usr/obj has become a bit small on at least
amd64, where there's no room for the entire build and a few manual
debugging library builds. grow it for future system installs
ok otto
sndiod: Enable recording and monitoring at the same time
If both recording and monitoring are selected (ex. -m play,rec,mon),
then the clients get a mix of what programs play and what the audio
interface records.
sndiod: Mostly style change of dev_sub_bcopy()
Calculate cmap_do() arguments near the cmap_do() call. Makes the code
slightly more readable and easier to modify.
No behaviour change.
sndiod: Factor the code to join/expand channels
Move the join/expand logic to the cmap structure which is used by
both playback and recording code-paths.
Make some fixes to the amdgpu(4) activate function to better reflect
what Linux does for suspend/hibernate/resume. In particular, set the
in_s4 member earlier, remove some duplicated code that now gets called
as part of amdgpu_pmops_suspend() and call amdgpu_pmops_complete() at
the end of a resume/restore.
ok jsg@
sndiod: Fix calls to enc_sil_do() with uninitialized argument
If the encoding is native the conv structure is uninitialized and
memset() must be used instead.
libsndio: Simplify cookie handling
Simple asnprintf(3) calls can replace the "clever" malloc(3) and
memcpy(3) calls that are much harder to review.
ok armani, deraadt
sndiod: Simplify unix socket binding code
Use sizeof(sockaddr_un->sun_path) instead of the complicated
calculation of the max socket path length.
Use the socket's address family to determine if this is an Unix
socket: no need to store the path, allocate memory for it, then
remember to free it.
ok armani@
Fix vmd(8) pledge violation when opening /dev/null.
The "vmm" process responsible for vm process creation uses /dev/null
for std{in,out,err} if running daemonized. Where it was opening
/dev/null is no longer supported by its current set of pledges.
Instead of expanding pledge scope, open /dev/null early prior to
unveil(2) and reuse the same file descriptor if needed.
This resolves the pledge violation reported by Nikita Kandinov and
removes the need to call open(2) for every created vm.
ok deraadt@, bluhm@
unstub ttm_tt_swapout()
kettenis now sees this called in the amdgpu hibernate path with GC 11.0.1.
I've seen it called when running piglit on vega 10.
feedback and ok kettenis@, who had almost the same diff
openssl.1: remove mention of OPENSSL_EC_BIN_PT_COMP
The need for this compile time option enabling point compression for
binary curves despite patent issues has been removed in openssl 1.0.0
(released in 2010).
[It's really difficult to count the number of bad ideas in the above.]
plug a memory leak when detaching codel ifq disciplines on an interface.
fqcodel_alloc() takes (moves) ownership of memory allocated as part
of the pf queue code that sets up the queue discipline, but nothing
releases it when the discipline is removed from the interface. every
time you load a pf rulese that uses codel, it temporarily resets
an interface to priq before creating and attaching new codel
instances. this means every ruleset load would leak memory, to the
point where M_DEVBUF allocations exhaust the kernel map.
this has fqcodel_free() call fqcodel_pf_free to free the now unused
codel discipline.
found and fixed by lexi wilson
Fix a longstanding off-by-one error in hibernate unpack.
Last year, I made a commit to early-allocate the hibernate data structure
needed during actual hibernate instead of triyng to allocate it at the
time of 'ZZZ'. This change exposed an old bug that caused machines to
be unstable after unhibernate (typically manifesting in pool corruption
or other difficult to track down bugs).
The error was due to skipping a page of memory immediately adjacent to
the hibernate piglet during unpack, effectively leaving that page with
whatever content it had before image unpack started. If you got lucky, this
page wasn't in use and you didn't notice anything. But with the fix to do
early allocation, the chance that the page was in use increased, and
thus the bug was exposed.
Big thank you to Walter Alejandro Iglesias for tracking this down and
providing the diff/root cause.
[5 lines not shown]
Fix Tx performance on iwx(4) BZ devices.
BZ devices need to be told how long the MAC header is in the offload_assist
field of the Tx command, even if no offloading features are being used.
Found by chance while browsing iwlwifi driver git logs.
See linux.git commit af8bfc7e38a7ad4f8e1663de7ab1463e644050b2
When pledged, if a process receives a bad descriptor the receiver should
not be killed. The EPERM approach used for other conditions is good enough.
from deraadt@; Report from Henry Ford
this is errata/7.7/028_recvfd.patch.sig
