Add UFS_WAPBL_UPDATE() after block allocations to group the
allocation and the inode update into the same transaction.
PR kern/60621 (panic: ufs_inactive: dirty filesystem?)
The 68010, 68020, and 68030 are have strictly ordered memory operations.
The 68040 and 68060 do not, however -- stores can be observed after
subsequent loads.
This is to say: all loads are load-acquire, all stores are store-release,
but store-before-load must be explicitly serialized. The 68040 User's
Manual states that a NOP insn forces this serialization.
And so: membar_aquire -> do nothing, membar_release -> do nothing,
membar_sync -> do nothing *with feeling*
Chapter and verse of the manual quoted in comments.
The MMU-specific functions that load the user root pointer already take
care of any necessary ATC / cache invalidations, so we don't need to do
it again here.
mmu_load_urp20hp(): No need to explicitly invalidate the external VAC
or the MMU's TLB -- user VAC entries and user TLB entries are automatically
invalidated when the USTP register is written, and there's no need to
invalidate supervisor entries in either of those when switching the user
address space.
pckbc(4) at acpi: Fix build with -Os.
Fixes:
/home/riastradh/netbsd/current/src/sys/dev/acpi/pckbc_acpi.c: In function 'pckbc_acpi_intr_establish':
/home/riastradh/netbsd/current/src/sys/dev/acpi/pckbc_acpi.c:242:63: error: 'handle' may be used uninitialized [-Werror=maybe-uninitialized]
242 | rv = acpi_intr_establish(sc->sc_dv, (uint64_t)(uintptr_t)handle,
| ^~~~~~~~~~~~~~~~~
/home/riastradh/netbsd/current/src/sys/dev/acpi/pckbc_acpi.c:224:21: note: 'handle' was declared here
224 | ACPI_HANDLE handle;
| ^~~~~~
No functional change intended -- the code wasn't _wrong_ before; this
just makes it clearer, with less analysis effort, that handle is
initialized in the path in question.
While here, rename local variable `rv' to `ih' to make it clearer
what it is (not just some arbitrary return value of a subroutine but
specifically an interrupt handler cookie).
[3 lines not shown]
Overhaul ATC / cache manipulation.
ATC + cache invalidation information is now gathered in a per-CPU
information struture that gathers up to 8 VAs in a batch. If more
than 8 VAs need to be invalidated, we just hit the entire ATC, trading
off the cost of the per-entry vs all-entries operations.
The same change allows us to batch cache invalidations, which reduces
the number of redundant cache invalidations.
Another nice property of these changes is that it collapses the HP MMU
handling into the same code paths as garden-variety 68851/68030, simplifying
the code flow in all sorts of places.
use ret from raw test case in verbose to see return value
and then tests can reach normalization tests for in flight packet testing.
more in flight packets tests for routing and possibly logging soon
gas: vax: Cherry-pick upstream commit 2722595890a
PR port-vax/59326
---
Author: Jan-Benedict Glaw <jbglaw at getslash.de>
Date: Sun Nov 15 18:36:27 2026 +0100
VAX: In PIC code, use GOT for function pointers
Use GOT references for function pointers in PIC code. Actual fix by
Rin Okuyama, initially reported for NetBSD in https://gnats.netbsd.org/59326 .
PR gas/32941
gas/
* config/tc-vax.c (md_estimate_size_before_relax): Use GOT instead of PLT.
* testsuite/gas/vax/elf-got-func.s,
* testsuite/gas/vax/elf-got-func.d: New test.
* testsuite/gas/vax/vax.exp: Run it.