[PATCH v4 0/2] alpha: complete the direct-mm shootdown fixes

[Magnus Lindholm]

This is the second of two series. Apply all eight patches of "alpha: fix
stale TLB translations breaking copy-on-write and writeback" v4 first,
then these two patches. The base-commit and prerequisite-patch-id lines
below describe that ordering from for-next at 33465f6ab697.

Patch 1 removes the remote context-clearing loop after the migration
shootdown rendezvous. The callbacks have already updated their own slots;
clearing them remotely can hide a CPU still running the mm.

Patch 2 fixes the mm_users <= 1 shootdown shortcut. A kthread_use_mm()
borrower takes mmgrab(), not mmget(), so mm_users does not count it.
Inspect remote context slots instead of clearing them, and skip IPIs
only when no other CPU has a context. Together the two patches remove
runtime remote writes to those slots.

The full-mm, icache and migration callbacks now test current->mm, like
the page-flush IPI handler. A CPU retaining the mm only lazily clears its
slot instead of publishing a new ASN. The migration callback still does
the per-VA tbi() afterwards. Returning to the mm allocates a new ASN
through ev5_switch_mm(), or through the direct-load path for an explicit
borrow by current.

The caller-side tests in flush_tlb_mm() and flush_icache_user_page()
remain current->active_mm. A lazy caller can publish its own slot, but
it also performs the shortcut check, which excludes the calling CPU.
The claim about retiring lazy slots applies to the callbacks; it does
not mean that every path now tests current->mm.

Context publication from ev5_switch_mm() is ordered before use of the
mm by the mb() in Alpha's arch_spin_unlock(), when finish_lock_switch()
drops the rq lock. This is an Alpha property, not a generic guarantee
that unlocking the scheduler lock is a full barrier. A comment at the
store now records that dependency. The smp_mb() added to
__load_new_mm_context() in v3 remains necessary for publications from
check_mmu_context(), which runs after the scheduler has dropped the lock.

Changes since v3:

  - Drop the stale v2 Message-ID trailer from patch 1. Its code is
    unchanged and Matt's Reviewed-by is retained.
  - Test current->mm in ipi_flush_mm_and_page(), so migration callbacks
    also retire lazy slots rather than repopulating them on each page.
  - Explain why the caller-side active_mm tests remain unchanged.
  - Name Alpha's arch_spin_unlock() mb() in the ordering argument and
    document the dependency next to the ev5_switch_mm() publication.
  - Add Matt's Reviewed-by to patch 2 with the migration-handler change
    he requested as the condition for that tag.

Testing

Fresh fork-throughput measurements and an SMP run under compaction are
still pending for this revision. Targeted Alpha SMP, UP and SMP with
compaction builds passed, as did a full ES40 CONFIG_COMPACTION=y build. The earlier fork numbers do not measure
these handler changes and are not presented as results for v4.

For comparison only, on 2026-10-09 the preceding MMU v3 plus stale-TLB v4
passed nine lost-store rounds, nine stale-read rounds and ten writeback
rounds on an ES40, EV68AL Tsunami, with four CPUs online and
CONFIG_COMPACTION=n. No new kernel errors were observed. That run did not
exercise the migration callback changed here and is not validation of
this revision under compaction.

The direct context-load fix moved to stale-TLB v4 patch 1. Its existing
review/test tags remain there; it is not duplicated in this series.
Both patches here retain Cc: stable, including the migration cleanup as
the shortcut's prerequisite.

Thanks to Matt Turner for the review and EV7 testing.

Earlier revision:
  v2: https://lore.kernel.org/linux-alpha/20261008195608.965266-1-linmag7@xxxxxxxxx/
  v3: https://lore.kernel.org/linux-alpha/20261009212501.971158-1-linmag7@xxxxxxxxx/

Magnus Lindholm (2):
  alpha: do not clear remote MMU contexts in migrate_flush_tlb_page()
  alpha: do not skip TLB shootdown IPIs for kthread-borrowed mms

 arch/alpha/include/asm/mmu_context.h |  7 +++-
 arch/alpha/kernel/smp.c              | 52 ++++++++++++++--------------
 arch/alpha/mm/fault.c                |  4 ++-
 arch/alpha/mm/tlbflush.c             | 18 +---------
 4 files changed, 36 insertions(+), 45 deletions(-)


base-commit: 33465f6ab697abceafd9a340067a544d551c4412
prerequisite-patch-id: d194ee1170a45793fa8533400dbc5599fc0e59ae
prerequisite-patch-id: 17df30a2da268e1bd07c7b44de4ef061cb744c50
prerequisite-patch-id: 94856bd5734d96521ef19e59cdb77af1d577fefa
prerequisite-patch-id: 7326de756f1a933cd3c44e4f7669c5a7c4519678
prerequisite-patch-id: 9e39c93f5b7267d1e8ffd176988803681ba03eeb
prerequisite-patch-id: c3f48cfa74e60b966610f88f73f7a393b5adf0b2
prerequisite-patch-id: 73a4faf81a8a0929ba8ba03a6d9a3286e0d9bd2b
prerequisite-patch-id: 55fb49cef84ea661b5340f14c56d73c04556f565
-- 
2.43.0