[PATCH v1] riscv: bring up batched unmap tlb flush
Jinyu Tang
tjytimi at 163.com
Wed Oct 5 07:43:24 PDT 2022
For riscv, this feature can decrease the times of
flush_tlb_pages(), and decrease the times of boardcasting for SMP.
Add @start and @end in struct arch_tlbflush_unmap_barch in riscv
compared to x86, which can record the minimum and maximum addresses
in this flush batch to reduce the flush range.If there are a lot of
pages to reclaim, or pages are shared with many tasks like server
application,this feature is more beneficial.
Signed-off-by: Jinyu Tang <tjytimi at 163.com>
---
This PATCH maybe conflict with patch v4 of arm64 batched unmap tlb flush
and patch v9 of IPI support for riscv.If this patch is ok,I will rebase
it on them.
I have test this patch in QEMU base on:
https://lore.kernel.org/all/20220921084302.43631-3-yangyicong@huawei.com/
#include <sys/types.h>
#include <unistd.h>
#include <sys/mman.h>
#include <string.h>
#include <time.h>
#include <stdio.h>
#include <stdlib.h>
int main()
{
clock_t start,finish;
double dur;
#define SIZE (1 * 1024 * 1024 * 500)
volatile unsigned char *p = mmap(NULL, SIZE, PROT_READ | PROT_WRITE,
MAP_SHARED | MAP_ANONYMOUS, -1, 0);
for(int tt = 0; tt < 5; tt ++){
memset(p, 0x88, SIZE);
start = clock();
for (int k = 0; k < 10; k++) {
/* swap in */
for (int i = 0; i < SIZE; i += 4096) {
(void)p[i];
}
/* swap out */
madvise(p, SIZE, MADV_PAGEOUT);
}
finish = clock();
dur = (double)(finish - start) / CLOCKS_PER_SEC;
printf("%f seconds this time\n",dur);
}
return 0;
}
And in riscv QEMU, this patch decrease the average 27.1s to 26.7s of time cost
about this test app.Maybe the decrease will be more obvious for real CPU.
arch/riscv/Kconfig | 8 ++++++++
arch/riscv/include/asm/tlbbatch.h | 18 ++++++++++++++++++
arch/riscv/include/asm/tlbflush.h | 20 ++++++++++++++++++++
arch/riscv/mm/tlbflush.c | 18 ++++++++++++++++++
arch/x86/include/asm/tlbflush.h | 2 +-
mm/rmap.c | 9 ++++-----
6 files changed, 69 insertions(+), 6 deletions(-)
create mode 100644 arch/riscv/include/asm/tlbbatch.h
diff --git a/arch/riscv/Kconfig b/arch/riscv/Kconfig
index 59d18881f35b..03304d0e0c0a 100644
--- a/arch/riscv/Kconfig
+++ b/arch/riscv/Kconfig
@@ -498,6 +498,14 @@ config ARCH_HAS_KEXEC_PURGATORY
depends on CRYPTO=y
depends on CRYPTO_SHA256=y
+config ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH
+ bool "Support batched unmap tlb flush"
+ help
+ This will make tlb flush batched when unmap pages. If there are
+ a lot of pages to reclaim, or pages are shared with many tasks
+ like server device, this will decrease the times of flushing tlbs
+ and broadcasting.
+
config CRASH_DUMP
bool "Build kdump crash kernel"
help
diff --git a/arch/riscv/include/asm/tlbbatch.h b/arch/riscv/include/asm/tlbbatch.h
new file mode 100644
index 000000000000..357bb0f46cff
--- /dev/null
+++ b/arch/riscv/include/asm/tlbbatch.h
@@ -0,0 +1,18 @@
+/* SPDX-License-Identifier: GPL-2.0 */
+#ifndef _ARCH_RISCV_TLBBATCH_H
+#define _ARCH_RISCV_TLBBATCH_H
+
+#include <linux/cpumask.h>
+
+struct arch_tlbflush_unmap_batch {
+ /*
+ * Each bit set is a CPU that potentially has a TLB entry for one of
+ * the PFNs being flushed. And the start and end will record the range
+ * to be flush this batch.
+ */
+ struct cpumask cpumask;
+ unsigned long start;
+ unsigned long end;
+};
+
+#endif /* _ARCH_RISCV_TLBBATCH_H */
diff --git a/arch/riscv/include/asm/tlbflush.h b/arch/riscv/include/asm/tlbflush.h
index 801019381dea..e2d9dad08bd0 100644
--- a/arch/riscv/include/asm/tlbflush.h
+++ b/arch/riscv/include/asm/tlbflush.h
@@ -10,6 +10,8 @@
#include <linux/mm_types.h>
#include <asm/smp.h>
#include <asm/errata_list.h>
+#include <asm/tlbbatch.h>
+#include <asm/page.h>
#ifdef CONFIG_MMU
static inline void local_flush_tlb_all(void)
@@ -59,4 +61,22 @@ static inline void flush_tlb_kernel_range(unsigned long start,
flush_tlb_all();
}
+static inline void arch_tlbbatch_add_mm(struct arch_tlbflush_unmap_batch *batch,
+ unsigned long address, struct mm_struct *mm)
+{
+ /*
+ * Each bit of cpumask is a CPU that potentially has a TLB entry for one of
+ * the PFNs being flushed. And the start and end will record the range
+ * to be flush this batch.
+ */
+ cpumask_or(&batch->cpumask, &batch->cpumask, mm_cpumask(mm));
+
+ if (address < batch->start)
+ batch->start = address;
+ else if (address + PAGE_SIZE > batch->end)
+ batch->end = address + PAGE_SIZE;
+}
+
+extern void arch_tlbbatch_flush(struct arch_tlbflush_unmap_batch *batch);
+
#endif /* _ASM_RISCV_TLBFLUSH_H */
diff --git a/arch/riscv/mm/tlbflush.c b/arch/riscv/mm/tlbflush.c
index 37ed760d007c..6b117732200b 100644
--- a/arch/riscv/mm/tlbflush.c
+++ b/arch/riscv/mm/tlbflush.c
@@ -86,3 +86,21 @@ void flush_pmd_tlb_range(struct vm_area_struct *vma, unsigned long start,
__sbi_tlb_flush_range(vma->vm_mm, start, end - start, PMD_SIZE);
}
#endif
+void arch_tlbbatch_flush(struct arch_tlbflush_unmap_batch *batch)
+{
+ unsigned int cpuid;
+ bool broadcast;
+
+ cpuid = get_cpu();
+
+ broadcast = cpumask_any_but(&batch->cpumask, cpuid) < nr_cpu_ids;
+ if (broadcast) {
+ sbi_remote_sfence_vma(&batch->cpumask, batch->start, batch->end - batch->start);
+ } else {
+ local_flush_tlb_all();
+ }
+
+ cpumask_clear(&batch->cpumask);
+
+ put_cpu();
+}
diff --git a/arch/x86/include/asm/tlbflush.h b/arch/x86/include/asm/tlbflush.h
index cda3118f3b27..9e7027ae256e 100644
--- a/arch/x86/include/asm/tlbflush.h
+++ b/arch/x86/include/asm/tlbflush.h
@@ -252,7 +252,7 @@ static inline u64 inc_mm_tlb_gen(struct mm_struct *mm)
}
static inline void arch_tlbbatch_add_mm(struct arch_tlbflush_unmap_batch *batch,
- struct mm_struct *mm)
+ unsigned long address, struct mm_struct *mm)
{
inc_mm_tlb_gen(mm);
cpumask_or(&batch->cpumask, &batch->cpumask, mm_cpumask(mm));
diff --git a/mm/rmap.c b/mm/rmap.c
index 93d5a6f793d2..edae27aa20a9 100644
--- a/mm/rmap.c
+++ b/mm/rmap.c
@@ -645,12 +645,12 @@ void try_to_unmap_flush_dirty(void)
#define TLB_FLUSH_BATCH_PENDING_LARGE \
(TLB_FLUSH_BATCH_PENDING_MASK / 2)
-static void set_tlb_ubc_flush_pending(struct mm_struct *mm, bool writable)
+static void set_tlb_ubc_flush_pending(struct mm_struct *mm, unsigned long address, bool writable)
{
struct tlbflush_unmap_batch *tlb_ubc = ¤t->tlb_ubc;
int batch, nbatch;
- arch_tlbbatch_add_mm(&tlb_ubc->arch, mm);
+ arch_tlbbatch_add_mm(&tlb_ubc->arch, address, mm);
tlb_ubc->flush_required = true;
/*
@@ -735,7 +735,7 @@ void flush_tlb_batched_pending(struct mm_struct *mm)
}
}
#else
-static void set_tlb_ubc_flush_pending(struct mm_struct *mm, bool writable)
+static void set_tlb_ubc_flush_pending(struct mm_struct *mm, unsigned long address, bool writable)
{
}
@@ -1596,8 +1596,7 @@ static bool try_to_unmap_one(struct folio *folio, struct vm_area_struct *vma,
* and traps if the PTE is unmapped.
*/
pteval = ptep_get_and_clear(mm, address, pvmw.pte);
-
- set_tlb_ubc_flush_pending(mm, pte_dirty(pteval));
+ set_tlb_ubc_flush_pending(mm, address, pte_dirty(pteval));
} else {
pteval = ptep_clear_flush(vma, address, pvmw.pte);
}
--
2.34.1
More information about the linux-riscv
mailing list