Allow (p)redefinitions of symbols in <machine/param.h>
sun3 (via machine/param3.h) defines UBC_NWINS UBC_WINSHIFT & MAXEXEC
before including <m68k/param.h> == and defines UBC_NWINS differently.
Allow for that (hopefully fix broken sun3 build).
[ Why sun3 has __mc680010__ defined is left as an unexplained mystery ]
Fix a longstanding bug on 64 bit ports with ABIs requiring sign-extension when returning 32-bit values in registers. Alpha, MIPS N64, RISCV.
With these ABIs it is necessary for the syscall return code to manually
sign-extend all 32-bit return values. If we don't, everything may look
just fine, until a syscall return value is spilled to the stack and then
reloaded. At that point, the zero-extended form produced by C assignment
into the 64-bit register for return to userspace will be sign-extended
upon reload, producing a different, corrupt value which will fail comparisons.
This was observed with a getuid() != geteuid() deep inside sshd on N64, on an
unusual system that had some very large UIDs with bit 31 set - but is quite
difficult to reproduce with small tests, since register pressure and compiler
behavior determine whether the conditions necessary to trigger the bug will
exist.
The easiest way to see whether a new target needs this treatment is to
look at its GCC configuratrion; if its PROMOTE_MODE has UNSIGNEDP = 0
for SImode, the sign-extension in syscall() is necessary.
Harmonize USPACE / UPAGES across m68k platforms.
There seems to be a complicated history rooted in the original Utah code
using UPAGES of 2 w/ NBPG of 4096 for an 8KB total u-area. The Amiga
port then appeared using NBPG of 8192, but UPAGES was not adjusted to
conpensate, resulting in a 16KB u-area for Amiga. From there, chaos
spread which resulted in us homehow tolerating a situation where the
Sun2 is also using a 16KB u-area despite only have 2048-byte pages.
End this madness by centrally defining USPACE as 8192 and deriving
UPAGES from USPACE. 8KB is totally enough for these platforms.
Address problems with MIPS Malta platform code found running under QEMU.
QEMU's "malta" system emulates a MIPS Malta with the Gallileo host bridge
and 32-bit or 64-bit CPUs of either endianness. It is one of the only
working QEMU system-level emulations that could run NetBSD with all
combinations of endianness and address size. After fixes to QEMU over the
past several years, NetBSD has been unable to use the emulated PCI bus in
big-endian and 64-bit configurations.
No actual Malta hardware with any Gallileo-based CPU card could be found
for testing. These changes have been checked against the databook and
some limited checking of the relevant QEMU changes (which seem to have
mostly come from former MIPS employees) was also performed.
Changes:
1. The GT-64120 host bridge _does_ byte-swap access to other PCI targets,
but _does not_ byte-swap access to itself (bus 0, device 0). QEMU
evidently used to get this wrong, but, I confirmed with the databook.
This means we need to manually byte-swap a bunch of access to the
[29 lines not shown]
Fix MKDEBUGKERNEL vs MKDEBUG for kernel debug file sets.
The problem manifests as checkflist failures when building ports that
have extensive ALL_KERNELS but not...building all the kernels; notably
the various "evb" ports with a bazillion kernels for a bazillion SoCs.
The mk.conf(5) man page documents MKDEBUGKERNEL as controlling
whether kernel debug files (netbsd-*.debug) appear in the
distribution sets. However, the prior implementation used MKDEBUG
(the general userland debug flag) instead.
This meant MKDEBUG=yes with MKDEBUGKERNEL=no incorrectly expected
kernel debug files for every kernel config listed in ALL_KERNELS.
When only a subset of kernels is built, checkflist fails with
missing files.
The fix is to make these variables fully conform to the longstanding
documentation. MKDEBUGKERNEL controls whether kernel debug symbols are
built; MKDEBUG controls everything else. If you want something like the
old behavior but minus the bugs, set both.
A few things done while chasing down mod/ref bugs:
- pmap_remove_mapping() can now take pointer to the vm_page, saving a
lookup and allowing some additional assertions when it's available
(which is "frequently" in this implementation).
- All of the PTE load/store/modify-in-PT helpers now are decorated
with "volatile".
- Don't bother with atomic_load / atomic_store.
- Simplify pmap_testbit() and pmap_changebit().
- Add more PMAP_DEBUG-only mod/ref tests (including a test that validates
MMU beavior that was used to find a Qemu m68k emulator bug).
Import postfix 3.11.2 (previous was 3.10.1)
Changes in 3.11.2
Bugfix (defect introduced: Postfix 3.11): the proxymap(8) daemon
dereferenced an uninitialized pointer after a request protocol
error. This daemon is not exposed to local or remote users. Found
by Claude Opus 4.6.
Bugfix (defect introduced: 20260309) a change, to set the service_name
default value to "amnesiac", violated a test that parameter names
in postconf output must match 1:1 with parameter names in the
postlink script.
Changes in 3.11.1
Bugfix (defect introduced: 20260219): alias_maps errors when
default_database_type was not set in main.cf. Fix by Michael Tokarev.
[97 lines not shown]
