[PATCH v2 0/5] x86/hyperv: Account for lazy mm_cpumask maintenance
From: Naman Jain
Date: Tue Nov 17 2026 - 04:15:26 EST
Lazy mm_cpumask maintenance leaves CPU bits set after switching to another
address space, with stale memberships removed by native flush callbacks.
Successful Hyper-V TLB flush hypercalls do not run those callbacks, so
historical memberships can persist and cause unnecessary remote flushes.
If the mask covers all present CPUs, the backend selects an all-processor
flush before reaching its per-CPU lazy-TLB checks.
This was found while investigating a regression between Linux 6.8 and 7.0
in SuperBench's eight-GPU PyTorch LSTM training on Azure MI300X VMs.
Recreating DataLoader workers each epoch caused host-to-GPU copy stalls
and timeouts. The reported bisection identified commit 209954cbc7d0
("x86/mm/tlb: Update mm_cpumask lazily"). This is distinct from lazy TLB
mode: a CPU actively running another address space need not be lazy.
Patches 1-3 prepare for target filtering in patch 4 while patch 5 adds
immediate cleanup.
Patch 1 rejects VP-set conversion failures that return zero, allowing
native fallback rather than issuing an empty-target extended hypercall.
Patch 2 makes the native target predicate reusable, retaining its
context-switch ordering and adding a clearly named callback wrapper.
Patch 3 uses native callbacks for requested mask trimming and global ASID
transitions. The latter need guest callbacks to switch CPUs still using
local ASIDs; a Hyper-V invalidation does not do that.
Patch 4 filters targets before selecting either hypercall format or the
all-processors shortcut. Page-table-freeing flushes remain unfiltered.
Patch 5 separately requests best-effort cleanup when no remote target
remains after filtering. Lazy memberships may survive this cleanup.
The cleanup policy remains open for discussion.
The suggested mask-weight > mm_users heuristic was tested as a replacement
for patch 5's approach on a 32-vCPU Hyper-V VM. With three remote readers,
mm_users was 4 while lazy memberships kept mask weights at 26-32. The
heuristic requested trimming for all 10,000 flushes, and each reader
received 10,000 remote callbacks. The scoped approach used Hyper-V
hypercalls instead. A doubled threshold or rate-limited policy has not
been tested.
Earlier testing of the monolithic patch reduced the 32-vCPU all-history
reproducer from 17.235 us to 1.622 us per iteration (two mprotect() calls
and a write). The reporter's application hit the 600-second timeout on the
unpatched 96-vCPU VM and completed in 424 and 436 seconds with the patch.
Changes since v1:
https://lore.kernel.org/all/20261029064234.362385-1-namjain@xxxxxxxxxxxxxxxxxxx/
* Split the change into five patches with separate explanations (Dave).
* Add the workload and regression context (Michael).
* Explain the ASID fallback separately from the reported regression (Dave).
* Keep related predicate/wrapper names and comment the definition rather
than the header declaration (Dave).
* Explain target filtering, page-table freeing, and the no-remote-target
cleanup condition. Document the const copy and lazy-membership limit
(Dave).
* Test the mm_users heuristic and retain scoped cleanup as a separate
policy discussion (Dave).
* Omit Fixes tags and stable requests (Dave).
* Rebase onto next-20261116.
Naman Jain (5):
x86/hyperv: Reject unrepresentable TLB flush VP sets
x86/tlb: Make the conditional flush predicate reusable
x86/hyperv: Use native callbacks for trimming and ASID transitions
x86/hyperv: Filter TLB targets before building hypercalls
x86/hyperv: Request mask cleanup when no remote target remains
arch/x86/hyperv/mmu.c | 71 +++++++++++++++++++++++----------
arch/x86/include/asm/tlbflush.h | 1 +
arch/x86/mm/tlb.c | 18 +++++++--
3 files changed, 65 insertions(+), 25 deletions(-)
base-commit: 749a4de370409464cf8501ec5b2451df45c84eb4
--
2.43.0