Skip to contentSkip to conversation

x86/mm/tlb: Revert retpoline avoidance approach

PATCH In review Dave Hansen 2 participants 3 messages Mar 18, 2022 – Mar 19, 2022

Subsystems (2)
lore
Share & tools
lore

Dave Hansen

author maintainer
#0 lore raw

PATCH 1/1 x86/mm/tlb: Revert retpoline avoidance approach

0day reported a regression on a microbenchmark which is intended to stress the TLB flushing path:

	https://lore.kernel.org/all/20220317090415.GE735@xsang-OptiPlex-9020/

It pointed at a commit from Nadav which intended to remove retpoline overhead in the TLB flushing path by taking the 'cond'-ition in on_each_cpu_cond_mask(), pre-calculating it, and incorporating it into 'cpumask'. That allowed the code to use a bunch of earlier direct calls instead of later indirect calls that need a retpoline.

But, in practice, threads can go idle (and into lazy TLB mode where they don't need to flush their TLB) between the early and late calls. It works in this direction and not in the other because TLB-flushing threads tend to hold mmap_lock for write. Contention on that lock causes threads to _go_ idle right in this early/late window.

There was not any performance data in the original commit specific to the retpoline overhead. I did a few tests on a system with retpolines:

	https://lore.kernel.org/all/[email protected]/

which showed a possible small win. But, that small win pales in comparison with the bigger loss induced on non-retpoline systems.

Revert the patch that removed the retpolines. This was not a clean revert, but it was self-contained enough not to be too painful.

Fixes 6035152d8eeb ("x86/mm/tlb: Open-code on_each_cpu_cond_mask() for tlb_is_not_lazy()")
Cc Nadav Amit <[email protected]>
Cc Ingo Molnar <[email protected]>
Cc Andy Lutomirski <[email protected]>
Cc Peter Zijlstra <[email protected]>
---
 arch/x86/mm/tlb.c | 37 +++++--------------------------------
 1 file changed, 5 insertions(+), 32 deletions(-)
arch/x86/mm/tlb.c +5 -32
index 1e6513f95133..161984b44331 100644
@@ -854,13 +854,11 @@ static void flush_tlb_func(void *info)
 			nr_invalidate);
 }
 
-static bool tlb_is_not_lazy(int cpu)
+static bool tlb_is_not_lazy(int cpu, void *data)
 {
 	return !per_cpu(cpu_tlbstate_shared.is_lazy, cpu);
 }
 
-static DEFINE_PER_CPU(cpumask_t, flush_tlb_mask);
-
 DEFINE_PER_CPU_SHARED_ALIGNED(struct tlb_state_shared, cpu_tlbstate_shared);
 EXPORT_PER_CPU_SYMBOL(cpu_tlbstate_shared);
 
@@ -889,36 +887,11 @@ STATIC_NOPV void native_flush_tlb_multi(const struct cpumask *cpumask,
 	 * up on the new contents of what used to be page tables, while
 	 * doing a speculative memory access.
 	 */
-	if (info->freed_tables) {
+	if (info->freed_tables)
 		on_each_cpu_mask(cpumask, flush_tlb_func, (void *)info, true);
-	} else {
-		/*
-		 * Although we could have used on_each_cpu_cond_mask(),
-		 * open-coding it has performance advantages, as it eliminates
-		 * the need for indirect calls or retpolines. In addition, it
-		 * allows to use a designated cpumask for evaluating the
-		 * condition, instead of allocating one.
-		 *
-		 * This code works under the assumption that there are no nested
-		 * TLB flushes, an assumption that is already made in
-		 * flush_tlb_mm_range().
-		 *
-		 * cond_cpumask is logically a stack-local variable, but it is
-		 * more efficient to have it off the stack and not to allocate
-		 * it on demand. Preemption is disabled and this code is
-		 * non-reentrant.
-		 */
-		struct cpumask *cond_cpumask = this_cpu_ptr(&flush_tlb_mask);
-		int cpu;
-
-		cpumask_clear(cond_cpumask);
-
-		for_each_cpu(cpu, cpumask) {
-			if (tlb_is_not_lazy(cpu))
-				__cpumask_set_cpu(cpu, cond_cpumask);
-		}
-		on_each_cpu_mask(cond_cpumask, flush_tlb_func, (void *)info, true);
-	}
+	else
+		on_each_cpu_cond_mask(tlb_is_not_lazy, flush_tlb_func,
+				(void *)info, 1, cpumask);
 }
 
 void flush_tlb_multi(const struct cpumask *cpumask,
Signature · 1 line
2.34.0
Acked-by on v1 1/1
#1 lore raw
Quoted 16 lines · Dave Hansen

⚠ External Email: This email originated from outside of the organization. Do not click links or open attachments unless you recognize the sender.

0day reported a regression on a microbenchmark which is intended to stress the TLB flushing path:

       https://nam04.safelinks.protection.outlook.com/?url=https%3A%2F%2Flore.kernel.org%2Fall%2F20220317090415.GE735%40xsang-OptiPlex-9020%2F&amp;data=04%7C01%7Cnamit%40vmware.com%7C4a2c382b5ef44105474308da08fd0c7f%7Cb39138ca3cee4b4aa4d6cd83d9dd62f0%7C0%7C0%7C637832180178751497%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C3000&amp;sdata=TA0iATQCnfDjIZ1lG3YdhjMZjelXrVatBjBE8Hz3AfE%3D&amp;reserved=0

It pointed at a commit from Nadav which intended to remove retpoline overhead in the TLB flushing path by taking the 'cond'-ition in on_each_cpu_cond_mask(), pre-calculating it, and incorporating it into 'cpumask'. That allowed the code to use a bunch of earlier direct calls instead of later indirect calls that need a retpoline.

Quoted 6 lines

But, in practice, threads can go idle (and into lazy TLB mode where they don't need to flush their TLB) between the early and late calls. It works in this direction and not in the other because TLB-flushing threads tend to hold mmap_lock for write. Contention on that lock causes threads to _go_ idle right in this early/late window.

#2 lore raw
Quoted 20 lines · Nadav Amit
Quoted 17 lines · Dave Hansen

⚠ External Email: This email originated from outside of the organization. Do not click links or open attachments unless you recognize the sender.

0day reported a regression on a microbenchmark which is intended to stress the TLB flushing path:

      https://nam04.safelinks.protection.outlook.com/?url=https%3A%2F%2Flore.kernel.org%2Fall%2F20220317090415.GE735%40xsang-OptiPlex-9020%2F&amp;data=04%7C01%7Cnamit%40vmware.com%7C4a2c382b5ef44105474308da08fd0c7f%7Cb39138ca3cee4b4aa4d6cd83d9dd62f0%7C0%7C0%7C637832180178751497%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C3000&amp;sdata=TA0iATQCnfDjIZ1lG3YdhjMZjelXrVatBjBE8Hz3AfE%3D&amp;reserved=0

It pointed at a commit from Nadav which intended to remove retpoline overhead in the TLB flushing path by taking the 'cond'-ition in on_each_cpu_cond_mask(), pre-calculating it, and incorporating it into 'cpumask'. That allowed the code to use a bunch of earlier direct calls instead of later indirect calls that need a retpoline.

Quoted 9 lines
Quoted 6 lines

But, in practice, threads can go idle (and into lazy TLB mode where they don't need to flush their TLB) between the early and late calls. It works in this direction and not in the other because TLB-flushing threads tend to hold mmap_lock for write. Contention on that lock causes threads to _go_ idle right in this early/late window.

Dave,

Note that your SOB has a wrong email address (intel.cm).

I noticed only because my mail server complained about wrong email address in the recipient list…