git: 55e020a6f9fc - stable/13 - amd64: Reduce the amount of cpuset copying done for TLB shootdowns
- Go to: [ bottom of page ] [ top of archives ] [ this month ]
Date: Wed, 15 Dec 2021 13:38:05 UTC
The branch stable/13 has been updated by markj:
URL: https://cgit.FreeBSD.org/src/commit/?id=55e020a6f9fcbe80bfd8f532cae9c22cf4ebd74b
commit 55e020a6f9fcbe80bfd8f532cae9c22cf4ebd74b
Author: Mark Johnston <markj@FreeBSD.org>
AuthorDate: 2021-11-15 17:52:03 +0000
Commit: Mark Johnston <markj@FreeBSD.org>
CommitDate: 2021-12-15 13:31:48 +0000
amd64: Reduce the amount of cpuset copying done for TLB shootdowns
We use pmap_invalidate_cpu_mask() to get the set of active CPUs. This
(32-byte) set is copied by value through multiple frames until we get to
smp_targeted_tlb_shootdown(), where it is copied yet again.
Avoid this copying by having smp_targeted_tlb_shootdown() make a local
copy of the active CPUs for the pmap, and drop the cpuset parameter,
simplifying callers. Also leverage the use of the non-destructive
CPU_FOREACH_ISSET to avoid unneeded copying within
smp_targeted_tlb_shootdown().
Reviewed by: alc, kib
Tested by: pho
Sponsored by: The FreeBSD Foundation
(cherry picked from commit ab12e8db292c386a33445dcd95fa629413954192)
---
sys/amd64/amd64/mp_machdep.c | 39 ++++++++++++++++++---------------------
sys/amd64/amd64/pmap.c | 8 +++-----
sys/amd64/include/pmap.h | 9 +++++++--
sys/x86/include/x86_smp.h | 15 ++++++++++++---
4 files changed, 40 insertions(+), 31 deletions(-)
diff --git a/sys/amd64/amd64/mp_machdep.c b/sys/amd64/amd64/mp_machdep.c
index cbc4b33841ba..58f4a539c481 100644
--- a/sys/amd64/amd64/mp_machdep.c
+++ b/sys/amd64/amd64/mp_machdep.c
@@ -613,10 +613,10 @@ invl_scoreboard_slot(u_int cpu)
* completion.
*/
static void
-smp_targeted_tlb_shootdown(cpuset_t mask, pmap_t pmap, vm_offset_t addr1,
- vm_offset_t addr2, smp_invl_cb_t curcpu_cb, enum invl_op_codes op)
+smp_targeted_tlb_shootdown(pmap_t pmap, vm_offset_t addr1, vm_offset_t addr2,
+ smp_invl_cb_t curcpu_cb, enum invl_op_codes op)
{
- cpuset_t other_cpus;
+ cpuset_t mask;
uint32_t generation, *p_cpudone;
int cpu;
bool is_all;
@@ -631,10 +631,12 @@ smp_targeted_tlb_shootdown(cpuset_t mask, pmap_t pmap, vm_offset_t addr1,
KASSERT(curthread->td_pinned > 0, ("curthread not pinned"));
/*
- * Check for other cpus. Return if none.
+ * Make a stable copy of the set of CPUs on which the pmap is active.
+ * See if we have to interrupt other CPUs.
*/
- is_all = !CPU_CMP(&mask, &all_cpus);
- CPU_CLR(PCPU_GET(cpuid), &mask);
+ CPU_COPY(pmap_invalidate_cpu_mask(pmap), &mask);
+ is_all = CPU_CMP(&mask, &all_cpus) == 0;
+ CPU_CLR(curcpu, &mask);
if (CPU_EMPTY(&mask))
goto local_cb;
@@ -663,7 +665,7 @@ smp_targeted_tlb_shootdown(cpuset_t mask, pmap_t pmap, vm_offset_t addr1,
CPU_FOREACH_ISSET(cpu, &mask) {
KASSERT(*invl_scoreboard_slot(cpu) != 0,
("IPI scoreboard is zero, initiator %d target %d",
- PCPU_GET(cpuid), cpu));
+ curcpu, cpu));
*invl_scoreboard_slot(cpu) = 0;
}
@@ -674,14 +676,11 @@ smp_targeted_tlb_shootdown(cpuset_t mask, pmap_t pmap, vm_offset_t addr1,
*/
if (is_all) {
ipi_all_but_self(IPI_INVLOP);
- other_cpus = all_cpus;
- CPU_CLR(PCPU_GET(cpuid), &other_cpus);
} else {
- other_cpus = mask;
ipi_selected(mask, IPI_INVLOP);
}
curcpu_cb(pmap, addr1, addr2);
- CPU_FOREACH_ISSET(cpu, &other_cpus) {
+ CPU_FOREACH_ISSET(cpu, &mask) {
p_cpudone = invl_scoreboard_slot(cpu);
while (atomic_load_int(p_cpudone) != generation)
ia32_pause();
@@ -705,29 +704,28 @@ local_cb:
}
void
-smp_masked_invltlb(cpuset_t mask, pmap_t pmap, smp_invl_cb_t curcpu_cb)
+smp_masked_invltlb(pmap_t pmap, smp_invl_cb_t curcpu_cb)
{
- smp_targeted_tlb_shootdown(mask, pmap, 0, 0, curcpu_cb, invl_op_tlb);
+ smp_targeted_tlb_shootdown(pmap, 0, 0, curcpu_cb, invl_op_tlb);
#ifdef COUNT_XINVLTLB_HITS
ipi_global++;
#endif
}
void
-smp_masked_invlpg(cpuset_t mask, vm_offset_t addr, pmap_t pmap,
- smp_invl_cb_t curcpu_cb)
+smp_masked_invlpg(vm_offset_t addr, pmap_t pmap, smp_invl_cb_t curcpu_cb)
{
- smp_targeted_tlb_shootdown(mask, pmap, addr, 0, curcpu_cb, invl_op_pg);
+ smp_targeted_tlb_shootdown(pmap, addr, 0, curcpu_cb, invl_op_pg);
#ifdef COUNT_XINVLTLB_HITS
ipi_page++;
#endif
}
void
-smp_masked_invlpg_range(cpuset_t mask, vm_offset_t addr1, vm_offset_t addr2,
- pmap_t pmap, smp_invl_cb_t curcpu_cb)
+smp_masked_invlpg_range(vm_offset_t addr1, vm_offset_t addr2, pmap_t pmap,
+ smp_invl_cb_t curcpu_cb)
{
- smp_targeted_tlb_shootdown(mask, pmap, addr1, addr2, curcpu_cb,
+ smp_targeted_tlb_shootdown(pmap, addr1, addr2, curcpu_cb,
invl_op_pgrng);
#ifdef COUNT_XINVLTLB_HITS
ipi_range++;
@@ -738,8 +736,7 @@ smp_masked_invlpg_range(cpuset_t mask, vm_offset_t addr1, vm_offset_t addr2,
void
smp_cache_flush(smp_invl_cb_t curcpu_cb)
{
- smp_targeted_tlb_shootdown(all_cpus, NULL, 0, 0, curcpu_cb,
- INVL_OP_CACHE);
+ smp_targeted_tlb_shootdown(kernel_pmap, 0, 0, curcpu_cb, INVL_OP_CACHE);
}
/*
diff --git a/sys/amd64/amd64/pmap.c b/sys/amd64/amd64/pmap.c
index 986a1b670d60..4325acd1255a 100644
--- a/sys/amd64/amd64/pmap.c
+++ b/sys/amd64/amd64/pmap.c
@@ -3104,8 +3104,7 @@ pmap_invalidate_page(pmap_t pmap, vm_offset_t va)
("pmap_invalidate_page: invalid type %d", pmap->pm_type));
pmap_invalidate_preipi(pmap);
- smp_masked_invlpg(pmap_invalidate_cpu_mask(pmap), va, pmap,
- pmap_invalidate_page_curcpu_cb);
+ smp_masked_invlpg(va, pmap, pmap_invalidate_page_curcpu_cb);
}
/* 4k PTEs -- Chosen to exceed the total size of Broadwell L2 TLB */
@@ -3203,7 +3202,7 @@ pmap_invalidate_range(pmap_t pmap, vm_offset_t sva, vm_offset_t eva)
("pmap_invalidate_range: invalid type %d", pmap->pm_type));
pmap_invalidate_preipi(pmap);
- smp_masked_invlpg_range(pmap_invalidate_cpu_mask(pmap), sva, eva, pmap,
+ smp_masked_invlpg_range(sva, eva, pmap,
pmap_invalidate_range_curcpu_cb);
}
@@ -3289,8 +3288,7 @@ pmap_invalidate_all(pmap_t pmap)
("pmap_invalidate_all: invalid type %d", pmap->pm_type));
pmap_invalidate_preipi(pmap);
- smp_masked_invltlb(pmap_invalidate_cpu_mask(pmap), pmap,
- pmap_invalidate_all_curcpu_cb);
+ smp_masked_invltlb(pmap, pmap_invalidate_all_curcpu_cb);
}
static void
diff --git a/sys/amd64/include/pmap.h b/sys/amd64/include/pmap.h
index 14ff4cf3cde9..8f1e77806a25 100644
--- a/sys/amd64/include/pmap.h
+++ b/sys/amd64/include/pmap.h
@@ -535,10 +535,15 @@ void pmap_kasan_enter(vm_offset_t);
void pmap_kmsan_enter(vm_offset_t);
#endif
-static __inline cpuset_t
+/*
+ * Returns a pointer to a set of CPUs on which the pmap is currently active.
+ * Note that the set can be modified without any mutual exclusion, so a copy
+ * must be made if a stable value is required.
+ */
+static __inline volatile cpuset_t *
pmap_invalidate_cpu_mask(pmap_t pmap)
{
- return (pmap->pm_active);
+ return (&pmap->pm_active);
}
#endif /* _KERNEL */
diff --git a/sys/x86/include/x86_smp.h b/sys/x86/include/x86_smp.h
index b9a1febb70f2..2cf0ff97eae0 100644
--- a/sys/x86/include/x86_smp.h
+++ b/sys/x86/include/x86_smp.h
@@ -107,14 +107,23 @@ void ipi_swi_handler(struct trapframe frame);
void ipi_selected(cpuset_t cpus, u_int ipi);
void ipi_self_from_nmi(u_int vector);
void set_interrupt_apic_ids(void);
+void mem_range_AP_init(void);
+void topo_probe(void);
+
+/* functions in mp_machdep.c */
void smp_cache_flush(smp_invl_cb_t curcpu_cb);
+#ifdef __i386__
void smp_masked_invlpg(cpuset_t mask, vm_offset_t addr, struct pmap *pmap,
smp_invl_cb_t curcpu_cb);
void smp_masked_invlpg_range(cpuset_t mask, vm_offset_t startva,
vm_offset_t endva, struct pmap *pmap, smp_invl_cb_t curcpu_cb);
void smp_masked_invltlb(cpuset_t mask, struct pmap *pmap,
smp_invl_cb_t curcpu_cb);
-void mem_range_AP_init(void);
-void topo_probe(void);
-
+#else
+void smp_masked_invlpg(vm_offset_t addr, struct pmap *pmap,
+ smp_invl_cb_t curcpu_cb);
+void smp_masked_invlpg_range(vm_offset_t startva, vm_offset_t endva,
+ struct pmap *pmap, smp_invl_cb_t curcpu_cb);
+void smp_masked_invltlb(struct pmap *pmap, smp_invl_cb_t curcpu_cb);
+#endif
#endif