BSD Commit Log Search

   hlim must be initialized.

   ok mvs

   Enable outline atomics on OpenBSD/arm64 by default.

   ok patrick@, deraadt@

   aarch64 supports Float16; also add two missing files for BFloat16 support.

   ok rsadowski@, miod@

   Add a 'mach fwsetup' command, which uses the EFI OsIndications feature
   to reboot the machine into the firmware setup interface, if supported.
   This is an optional feature introduced in UEFI 2.4 (released 2013).

   ok kn@

   SEV-ES: Move some ghcb_sync_out arguments into a struct

   Move the extra register values to ghcb_sync_out() into a struct. This will
   make the code simpler when more arguments are added and the same
   parameters are also passed to ghcb_sync_in().

   From Sebastian Sturm

   ok mlarkin@ hshoexer@

   strptime: support strftime's %v conversion
   From enh at google

   strptime: move gmt[] and utc[] constants to rodata
   From enh at google

   Make the external PCIe port work on the rpi5.

   ok mglocker@

   ec_asn1_test: fix error message

   ec_asn1_test: d'oh. actually ensure all builtin curves are of prime order

   ec_asn1_test: ensure all builtin curves are of prime order

   Finish rpi4 support.  Configure the SCB to respond to incoming transactions,
   ping the Raspberry Pi firmware to load the firmware for the VIA VL805 USB
   controller (needed on the Raspberry Pi 4B rev 1.4 and later), and enable
   SSC (Spread Spectrum Clock) mode.

   ok mglocker@

   ec_mult.c: place this code under my ISC license

   I have effectively rewritten the entirety of this file end of 2024.
   This isn't code I'm particularly proud of, but it's much better than
   it was before (it's not as if that involved any sort of challenge...)

   requested by/ok jsing

   typo fix from Christian Schulte, ok schwarze@

   sys/uvideo: synchronized definition both H264 structures

   OK: mglocker@

   Correctly calculate lines to clear for deletions, from Pavel Roskin.

   input_stop_utf8 can move the cursor, so it is now wrong to store the
   cursor position before calling it. Problem found by Jayakrishna
   Vadayath.

   Be more robust against misbehaving clients, prompted by deraadt based on
   a report from sai02 at student dot ubc dot ca.

   use an rwlock around the list of pfkey sockets instead of an SRPL.

   this keeps the locking (relatively) simple and avoids the need to
   refcnt a bunch of stuff.

   tweaks from mvs@
   ok mvs@ jmatthew@

   rsa_method_test: point at correct function in error string

   make amdgpu S3 suspend more reliable

   Laurence Tratt and matthieu@ reported that S3 suspend on amdgpu
   had become unreliable.

   Laurence with a Ryzen 9 9950X APU (GC 10.3.6)
   matthieu with a Ryzen 5 PRO 3500U APU (Picasso)

   this was tracked to:
   Revert "drm/amd: Stop evicting resources on APUs in suspend"
   5d9d62323e1893e6f429fad7da19a652088cac21 in linux-6.12.y/6.12.30
   d0ce1aaa8531a4a4707711cab5721374751c51b0 in mainline linux

   Instead of reverting the revert, change the activate function
   to be equivalent.  Makes the local change easier to maintain.

   Thanks to Laurence for patiently trying many patches, as I could
   not reproduce this on my Picasso machine.
   ok kettenis@

   The switch to FRAME_GROWS_DOWNWARD requires that the spill area for
   registers, when expanding the prologue of a variadic function, can no
   longer use room on top of the logical frame pointer, after the local
   variables, by returning a non-zero pretend_args_size of the room needed.
   Doing this confuses the rest of the compiler code and bad addresses get
   computed, especially on platforms where variadic arguments, depending
   on their type, can be passed in registers, then on the stack, then in
   more registers, then on the stack, then in registers again, and m88k is
   probably the most versatile example of this.

   Move the register save area immediately below the hard frame pointer,
   and do not set pretend_args_size. This fixes the worst cases of variadic
   function torture, while still letting __builtin_apply_args() compute the
   right address.

   After these changes, the (logical) frame pointer and the argument
   pointer are now identical, so it is possible to partly undo the addition
   of the logical frame pointer by sharing the same logical register
   between the fp and the ap.

   zap stray space

   Move opensslconf.h to a machine-independent place

   After drilling through many layers of fossilized turds from a
   long-forgotten millenium, jsing and I finally found oil^Wa
   machine-independent version of opensslconf.h.

   Remove the no longer needed versions in arch/*/ and move one copy
   to the top level.  Add an RCS tag and place the remaining garbage
   in the public domain.

   ok jsing

   sync

   Use the safe idiom from snprintf(3) manpage on how to check for string
   overflow.
   OK tb@ millert@

   Remove unused variable.
   OK tb@

   freenull: missing target deps, use -Wall -Werror like elsewhere

   freenull.c.head: include mlkem.h for MLKEM_{private,public}_key_free()

   Stop the xonly crash in libunwind on powerpc64

   The powerpc64 unwinder was trying to read the instruction at the
   return address.  This crashed, because you can't read from
   execute-only memory.

   This has probably been broken since January 2023 (before OpenBSD 7.3),
   when powerpc64 became xonly.  This broke anything that got a
   backtrace(3) from libexecinfo or threw a C++ exception with libc++abi.
   editors/emacs,no_x11 crashed in libexecinfo.  devel/gdb crashed in
   libc++abi.  If you tried to examine the core dupm in gdb, then gdb
   crashed again.

   The unwinder is asking whether the return address has an instruction
   to restore the TOC register.  Instead of crashing, call dladdr(3)
   twice to find the shared objects.  Assume that each object has only
   one TOC.  Restore TOC if returning to a different object.

   ok kettenis@