SuperBench's eight-GPU PyTorch LSTM training on Azure MI300X VMs regressed
between Linux 6.8 and 7.0. 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").

That change leaves a CPU's bit set after it switches to another mm.
Native flush callbacks remove these stale bits, but successful Hyper-V
hypercalls do not run those callbacks. A CPU actively running another
mm need not be in lazy TLB mode, so the existing lazy-TLB check does not
exclude it. If the mask includes all present CPUs, the backend selects
the all-processors shortcut before checking individual CPUs.

Use the native predicate to build a per-CPU target mask before selecting
the all-processors shortcut or either hypercall format. Keep the full
candidate mask when freeing page tables. Retain the empty basic-mask
check because filtering can leave no targets.

This patch does not change when trimming is requested.

Signed-off-by: Naman Jain <[email protected]>
---
 arch/x86/hyperv/mmu.c | 40 ++++++++++++++++++++++++++--------------
 1 file changed, 26 insertions(+), 14 deletions(-)

diff --git a/arch/x86/hyperv/mmu.c b/arch/x86/hyperv/mmu.c
index 375f23b157ab3..833eaea1d3fa8 100644
--- a/arch/x86/hyperv/mmu.c
+++ b/arch/x86/hyperv/mmu.c
@@ -1,6 +1,7 @@
 #define pr_fmt(fmt)  "Hyper-V: " fmt
 
 #include <linux/log2.h>
+#include <linux/percpu.h>
 #include <linux/slab.h>
 #include <linux/types.h>
 
@@ -16,6 +17,8 @@
 /* Each gva in gva_list encodes up to 4096 pages to flush */
 #define HV_TLB_FLUSH_UNIT (4096 * PAGE_SIZE)
 
+static DEFINE_PER_CPU(struct cpumask, hyperv_tlb_flush_mask);
+
 static u64 hyperv_flush_tlb_others_ex(const struct cpumask *cpus,
                                      const struct flush_tlb_info *info);
 
@@ -51,19 +54,14 @@ static inline int fill_gva_list(u64 gva_list[], int offset,
        return gva_n - offset;
 }
 
-static bool cpu_is_lazy(int cpu)
-{
-       return per_cpu(cpu_tlbstate_shared.is_lazy, cpu);
-}
-
 static void hyperv_flush_tlb_multi(const struct cpumask *cpus,
                                   const struct flush_tlb_info *info)
 {
        int cpu, vcpu, gva_n, max_gvas;
        struct hv_tlb_flush *flush;
+       struct cpumask *flush_cpus;
        u64 status;
        unsigned long flags;
-       bool do_lazy = !info->freed_tables;
 
        trace_hyperv_mmu_flush_tlb_multi(cpus, info);
 
@@ -87,6 +85,23 @@ static void hyperv_flush_tlb_multi(const struct cpumask 
*cpus,
                goto do_native;
        }
 
+       flush_cpus = this_cpu_ptr(&hyperv_tlb_flush_mask);
+       /*
+        * mm_cpumask() may retain CPUs that have switched to another mm.
+        * Use the native predicate for its context-switch ordering and lazy
+        * TLB exclusion. Page-table freeing must include lazy CPUs, so keep
+        * the full candidate mask in that case.
+        */
+       if (info->freed_tables) {
+               cpumask_copy(flush_cpus, cpus);
+       } else {
+               cpumask_clear(flush_cpus);
+               for_each_cpu(cpu, cpus) {
+                       if (should_flush_tlb(cpu, info))
+                               __cpumask_set_cpu(cpu, flush_cpus);
+               }
+       }
+
        if (info->mm) {
                /*
                 * AddressSpace argument must match the CR3 with PCID bits
@@ -101,7 +116,7 @@ static void hyperv_flush_tlb_multi(const struct cpumask 
*cpus,
        }
 
        flush->processor_mask = 0;
-       if (cpumask_equal(cpus, cpu_present_mask)) {
+       if (cpumask_equal(flush_cpus, cpu_present_mask)) {
                flush->flags |= HV_FLUSH_ALL_PROCESSORS;
        } else {
                /*
@@ -115,14 +130,12 @@ static void hyperv_flush_tlb_multi(const struct cpumask 
*cpus,
                 * must. We will also check all VP numbers when walking the
                 * supplied CPU set to remain correct in all cases.
                 */
-               cpu = cpumask_last(cpus);
+               cpu = cpumask_last(flush_cpus);
 
                if (cpu < nr_cpumask_bits && hv_cpu_number_to_vp_number(cpu) >= 
64)
                        goto do_ex_hypercall;
 
-               for_each_cpu(cpu, cpus) {
-                       if (do_lazy && cpu_is_lazy(cpu))
-                               continue;
+               for_each_cpu(cpu, flush_cpus) {
                        vcpu = hv_cpu_number_to_vp_number(cpu);
                        if (vcpu == VP_INVAL) {
                                local_irq_restore(flags);
@@ -166,7 +179,7 @@ static void hyperv_flush_tlb_multi(const struct cpumask 
*cpus,
        goto check_status;
 
 do_ex_hypercall:
-       status = hyperv_flush_tlb_others_ex(cpus, info);
+       status = hyperv_flush_tlb_others_ex(flush_cpus, info);
 
 check_status:
        local_irq_restore(flags);
@@ -205,8 +218,7 @@ static u64 hyperv_flush_tlb_others_ex(const struct cpumask 
*cpus,
        flush->hv_vp_set.valid_bank_mask = 0;
 
        flush->hv_vp_set.format = HV_GENERIC_SET_SPARSE_4K;
-       nr_bank = cpumask_to_vpset_skip(&flush->hv_vp_set, cpus,
-                       info->freed_tables ? NULL : cpu_is_lazy);
+       nr_bank = cpumask_to_vpset(&flush->hv_vp_set, cpus);
        if (nr_bank <= 0)
                return HV_STATUS_INVALID_PARAMETER;
 
-- 
2.43.0


Reply via email to