Skip to contentSkip to conversation

Optimize this_cpu_*() ops for non-x86 (ARM64 for this series)

RFC PATCH In review Yang Shi 14 participants 66 messages Jul 15 – Sep 2

Subsystems (9)
lore
Share & tools
lore

Yang Shi

author
#0 lore raw

v2 cover Optimize this_cpu_*() ops for non-x86 (ARM64 for this series)

Hi,

This is v2 RFC. In v2 a lot problems found out by Sashiko were fixed and more feature gaps were closed (please see the below changelog for the details). Although there are still some open issues, for example, it just can support 48 bits VA (for 4K and 64K) and 47 bits VA (for 16K), KPTI support has not been solved yet, etc, but I think the delta should be big enough and worth a new RFC to gather comments in order to make sure I'm on the right track.

Some more benchmarks were done, for example, some latency related benchmarks that I mentioned at LSFMM because I thought responsiveness should be improved due to the removal of preempt_disable. Collected more PMU counters as well. Please refer to the benchmark section for more details.

Look forward to comments.

Changelog
v2: * Added 3-level and 2-level page table support.
* Tested with 16K and 64K page size. But we just support 48 bits VA (4K and 64K) and 47 bits VA with 16K for now. Please refer to the below "known issue" section for the detail reason.
* Added support for memory hotplug.
* Added support for KASAN (generic).
* Treated percpu and local percpu area address as vmalloc address.
* Fixed build failure for x86.
* Fixed build failure for !CONFIG_NUMA.
* Added KASAN support for local percpu area.
* Some other misc bug fixes found out by Sashiko.
* More code refactor and cleanup.
* Regorganized the patches.
* More benchmarks, refer to benchmark section for more details.
* Rebased to v7.2-rc1.

Introduction
============
This patch series implemented the LSFMM 2026 proposal for optimizing this_cpu_*() ops on ARM64. For the details of the proposal, Please refer to:
https://lore.kernel.org/linux-mm/CAHbLzkpcN-T8MH6=W3jCxcFj1gVZp8fRqe231yzZT-rV_E_org@mail.gmail.com/
I didn't repeat it in the cover letter because there is no change to the proposal.

The series is based on 7.1-rc1. It is basically minimum viable patches. There are still a few hacks in this series and it may break something, for example, KPTI, SMT machines which shared TLB, etc. But it shoule be good enough for now to demonstrate the core idea. The main purpose of the RFC is to gather feedback, figure out missing parts and risks, and make sure we are on the right track, as well as hopefully it can help the discussion for the upcoming LSFMM.

I broke the patches down to arch-dependent and arch-independent parts so that hopefully the interested persons can do experiments on other architectures, for example, S390, easier.

A new kernel config is introduced, HAVE_LOCAL_PER_CPU_MAP. The architectures which can support this feature will select it. Allocating and freeing percpu local mapping is protected by this config so that others won't pay the cost.

Known Issues
============
1. KPTI
-------
We need determine what CPU we are on, then switch to the right page table. Currently arm64 kernel fetches tramp_pg_dir via swapper_pg_dir - fixed_offset, and fetches swapper_pg_dir from ttbr1. But ttbr1 may not hold swapper_pg_dir anymore except CPU #0. So we need to figure out the other way to handle it. Switching to tramp_pg_dir should be easy, but the reverse seems harder because tramp_pg_dir just maps the trampoline vectors.
Maybe we can do two steps switch. Switch to swapper_pg_dir at the first step, then switch to per cpu page table (for entry) or tramp page table (for exit). Nobody should call this_cpu_*() at either userspace -> kernel entry stage or kernel -> userspace exit stage.

2. SW PAN
---------
Has the similar issue as KPTI. It installs reserved_pg_dir to TTBR0 when running in kernel space, but fetching reserved_pg_dir via swapper_pg_dir - fixed_offset. Maybe we can save the physical address of swapper_pg_dir in a variable, then load it from that variable instead of ttbr1.

3. Shared TLB machines
----------------------
Some machines may share TLB between CPUs, for example, SMT machines may share TLB between the two hardware threads in one core.
The per cpu page table just can't work with it. Maybe we need a new cpufeature to indicate whether per cpu page table is allowed or not. Then just enable it for not-shared-TLB machines.

4. Don't support all VA bits
----------------------------
We just support 48 bits VA (4K and 64K) and 47 bits VA (16K) for now. For 4K and 64K, supporting other VA bits is not hard, we just need to determine the size for percpu and local percpu area.
But it is harder for supporting 48 bits VA + 16K page size. We just have two top level kernel page table entries with this configuration, but we assume we just need to sync up kernel page table at the top level for now. We need to sync up kernel page table at the second level in order to support it. I'm not sure whether it is worth it or not.

Benchmark
=========
The benchmarks are done on 160 core AmpereOne machine. The baseline is v7.2-rc1 kernel.

1. Reduction of kernel text size
--------------------------------
The patchset can reduce at least 11 instructions for this_cpu_*() ops. Both preempt_disable() and preempt_enable() need 4 instructions to manipulate the preempt count, and preempt_enable() needs more instructions (compare + READ + compare) to determine whether reschedule is needed or not. Because this_cpu_*() ops are inlined and called in a lot of places so we can save a lot of instructions.

The size of kernel text is reduced by ~184KB with default Fedora kernel config. This also helps reduce kernel icache miss rate and stalled frontend cycles as kernel build benchmark result showed.

2. Kernel Build
---------------
Run kernel build (make -j160) with the default Fedora kernel config in a memcg.
13% - 18% sys time improvment
3% - 7% wall time improvement

5% fewer kernel icache miss, 5% fewer executed kernel instructions and 15% fewer stalled frontend cycles for kernel.

3. stress-ng vm ops
-------------------
stress-ng --vm 160 --vm-bytes 128M --vm-ops 100000000 8.5% improvement

4. stress-ng vm ops + fork
--------------------------
stress-ng --mmapfork 160 --mmapfork-bytes 128M --mmapfork-ops 500 15% improvement

5. Specjbb
----------
The specjbb test latency curves showed the patched kernel has consistently lower p99 latency (the lower the better) than the baseline.

2.5% improvement on max-jOPS and 4% - 5% improvement on critical-jOPS. The specjbb benchmark is quite sensitive to latency and responsiveness, particularly critical-jOPS result. The patches are supposed to improve the responsiveness due to the reduction of preempt-disabled critical sections.

6. MySQL
--------
1% - 2% gains on read-only test, 2% - 4% gains on write-only test. Also see 15% decrease on frontend cache stall.

Regression test
===============
1. memcg creation
-----------------
Create 10K memcgs. Each memcg creation needs to allocate multiple percpu variables, for example, percpu refcnt, rstat and objcg percpu refcnt.

Consumed 2112K more virtual memory for percpu “local mapping” and a few more mega bytes consumed by per cpu page tables.
No noticeable regression was found for elapsed time.

2. fork test
------------
stress-ng --fork 160 --fork-ops 10000000
fork() needs to allocate multiple percpu variables, for example, rss counters and mm_cid_cpu.

Roughly 1% regression was found. However stress-ng fork test has quites small address space, the real life workloads typically have much larger address space and do more complicated works. The stress-ng mmapfork benchmark saw 15% improvement.

The organization of patches
===========================
The refactor and prepatory patches (patch 1 - patch 4) Percpu page table support patches (patch 5 - patch 8) Local percpu area support patches (patch 7 - patch 15) Use local percpu area for this_cpu ops (patch 16)

Yang Shi (16):
drivers: arch_numa: move percpu set up code to arch
arm64: kconfig: make percpu related configs not depend on NUMA
mm: pgalloc: introduce {pud|pmd}_populate_sync()
vmalloc: pass in pgd pointer for vmap{__vunmap}_range_noflush()
arm64: mm: enable percpu kernel page table
arm64: mm: defined {pud|pmd}_populate_sync()
arm64: mm: sync percpu page table for memory hotplug/unplug
arm64: kasan: sync up kasan shadow area page table
arm64: mm: define percpu virtual space area
mm: percpu: prepare to use dedicated percpu area
arm64: mm: map local percpu first chunk
mm: percpu: set up first chunk and reserve chunk
arm64: mm: introduce __per_cpu_local_off
mm: percpu: allocate and free local percpu vm area
arm64: kconfig: select HAVE_LOCAL_PER_CPU_MAP
arm64: percpu: use local percpu for this_cpu_*() APIs

 arch/arm64/Kconfig                   |  12 +++++++---
 arch/arm64/include/asm/mmu.h         |   5 ++++
 arch/arm64/include/asm/mmu_context.h |   9 +++++++-
 arch/arm64/include/asm/percpu.h      |  37 ++++++++++++++++++++++++++++-
 arch/arm64/include/asm/pgalloc.h     |  24 +++++++++++++++++++
 arch/arm64/include/asm/pgtable.h     |  37 ++++++++++++++++++++++++++---
 arch/arm64/kernel/setup.c            |   3 +++
 arch/arm64/kernel/smp.c              |  44 +++++++++++++++++++++++++++++++++++
 arch/arm64/mm/kasan_init.c           |  47 +++++++++++++++++++++++--------------
 arch/arm64/mm/mmu.c                  | 165 ++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++----------------------
 arch/arm64/mm/ptdump.c               |   4 ++++
 arch/riscv/kernel/smp.c              |  51 ++++++++++++++++++++++++++++++++++++++++
 drivers/base/arch_numa.c             |  51 +---------------------------------------
 include/linux/mm.h                   |  11 +++++++++
 include/linux/percpu.h               |   4 +++-
 include/linux/pgalloc.h              |  13 +++++++++++
 include/linux/vmalloc.h              |   3 +++
 mm/Kconfig                           |   9 ++++++++
 mm/internal.h                        |   5 +++-
 mm/kmsan/hooks.c                     |  14 +++++------
 mm/percpu-internal.h                 |  14 +++++++++++
 mm/percpu-vm.c                       |  94 ++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
 mm/percpu.c                          |  58 +++++++++++++++++++++++++++++++++++++---------
 mm/sparse-vmemmap.c                  |   4 ++--
 mm/vmalloc.c                         | 138 +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++---------------
 25 files changed, 712 insertions(+), 144 deletions(-)

Thanks,
Yang

author
#1 lore raw

PATCH v2 1/16 drivers: arch_numa: move percpu set up code to arch

The following patches will need to modify it to set up local percpu area on arm64 so the generic code won't work anymore. ARM64 and RISCV are the only users, so move the code to their own kernel/smp.c file.

We can't leave the generic code in arch_numa.c because both ARM64 and RISCV select CONFIG_GENERIC_ARCH_NUMA, so it will result in multiple definition problem.

There is no functional change.

---
 arch/arm64/kernel/smp.c  | 51 ++++++++++++++++++++++++++++++++++++++++
 arch/riscv/kernel/smp.c  | 51 ++++++++++++++++++++++++++++++++++++++++
 drivers/base/arch_numa.c | 51 +---------------------------------------
 3 files changed, 103 insertions(+), 50 deletions(-)
arch/arm64/kernel/smp.c +51 -0
index d46022f72075..0dab4e9c97e4 100644
@@ -819,6 +819,57 @@ void __init smp_prepare_cpus(unsigned int max_cpus)
 	}
 }
 
+#ifdef CONFIG_HAVE_SETUP_PER_CPU_AREA
+extern int cpu_to_node_map[NR_CPUS];
+
+unsigned long __per_cpu_offset[NR_CPUS] __read_mostly;
+EXPORT_SYMBOL(__per_cpu_offset);
+
+int early_cpu_to_node(int cpu)
+{
+	return cpu_to_node_map[cpu];
+}
+
+static int __init pcpu_cpu_distance(unsigned int from, unsigned int to)
+{
+	return node_distance(early_cpu_to_node(from), early_cpu_to_node(to));
+}
+
+void __init setup_per_cpu_areas(void)
+{
+	unsigned long delta;
+	unsigned int cpu;
+	int rc = -EINVAL;
+
+	if (pcpu_chosen_fc != PCPU_FC_PAGE) {
+		/*
+		 * Always reserve area for module percpu variables.  That's
+		 * what the legacy allocator did.
+		 */
+		rc = pcpu_embed_first_chunk(PERCPU_MODULE_RESERVE,
+					    PERCPU_DYNAMIC_RESERVE, PAGE_SIZE,
+					    pcpu_cpu_distance,
+					    early_cpu_to_node);
+#ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK
+		if (rc < 0)
+			pr_warn("PERCPU: %s allocator failed (%d), falling back to page size\n",
+				   pcpu_fc_names[pcpu_chosen_fc], rc);
+#endif
+	}
+
+#ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK
+	if (rc < 0)
+		rc = pcpu_page_first_chunk(PERCPU_MODULE_RESERVE, early_cpu_to_node);
+#endif
+	if (rc < 0)
+		panic("Failed to initialize percpu areas (err=%d).", rc);
+
+	delta = (unsigned long)pcpu_base_addr - (unsigned long)__per_cpu_start;
+	for_each_possible_cpu(cpu)
+		__per_cpu_offset[cpu] = delta + pcpu_unit_offsets[cpu];
+}
+#endif
+
 static const char *ipi_types[MAX_IPI] __tracepoint_string = {
 	[IPI_RESCHEDULE]	= "Rescheduling interrupts",
 	[IPI_CALL_FUNC]		= "Function call interrupts",
arch/riscv/kernel/smp.c +51 -0

54 lines not on this page · raw message

drivers/base/arch_numa.c +1 -50

63 lines not on this page · raw message

Signature · 1 line
2.47.0
author
#2 lore raw

PATCH v2 2/16 arm64: kconfig: make percpu related configs not depend on NUMA

In reply to Yang Shi

The percpu set up code has been moved out of arch_numa.c, so it is unnecessary to have percpu related kernel configs depend on NUMA.

---
 arch/arm64/Kconfig      |  6 +++---
 arch/arm64/kernel/smp.c | 10 ++++++----
 2 files changed, 9 insertions(+), 7 deletions(-)
arch/arm64/Kconfig +3 -3
index b3afe0688919..2e1bd79a579c 100644
@@ -224,6 +224,7 @@ config ARM64
 	select MMU_GATHER_RCU_TABLE_FREE
 	select HAVE_RSEQ
 	select HAVE_RUST if RUSTC_SUPPORTS_ARM64
+	select HAVE_SETUP_PER_CPU_AREA
 	select HAVE_STACKPROTECTOR
 	select HAVE_STATIC_CALL if CFI
 	select HAVE_SYSCALL_TRACEPOINTS
@@ -239,6 +240,8 @@ config ARM64
 	select LOCK_MM_AND_FIND_VMA
 	select MODULES_USE_ELF_RELA
 	select NEED_DMA_MAP_STATE
+	select NEED_PER_CPU_EMBED_FIRST_CHUNK
+	select NEED_PER_CPU_PAGE_FIRST_CHUNK
 	select NEED_SG_DMA_LENGTH
 	select OF
 	select OF_EARLY_FLATTREE
@@ -1581,9 +1584,6 @@ config NUMA
 	bool "NUMA Memory Allocation and Scheduler Support"
 	select GENERIC_ARCH_NUMA
 	select OF_NUMA
-	select HAVE_SETUP_PER_CPU_AREA
-	select NEED_PER_CPU_EMBED_FIRST_CHUNK
-	select NEED_PER_CPU_PAGE_FIRST_CHUNK
 	select USE_PERCPU_NUMA_NODE_ID
 	help
 	  Enable NUMA (Non-Uniform Memory Access) support.
arch/arm64/kernel/smp.c +6 -4
index 0dab4e9c97e4..28ff889658d3 100644
@@ -819,16 +819,18 @@ void __init smp_prepare_cpus(unsigned int max_cpus)
 	}
 }
 
-#ifdef CONFIG_HAVE_SETUP_PER_CPU_AREA
+#ifdef CONFIG_NUMA
 extern int cpu_to_node_map[NR_CPUS];
 
-unsigned long __per_cpu_offset[NR_CPUS] __read_mostly;
-EXPORT_SYMBOL(__per_cpu_offset);
-
 int early_cpu_to_node(int cpu)
 {
 	return cpu_to_node_map[cpu];
 }
+#endif
+
+#ifdef CONFIG_HAVE_SETUP_PER_CPU_AREA
+unsigned long __per_cpu_offset[NR_CPUS] __read_mostly;
+EXPORT_SYMBOL(__per_cpu_offset);
 
 static int __init pcpu_cpu_distance(unsigned int from, unsigned int to)
 {
Signature · 1 line
2.47.0
author
#3 lore raw

PATCH v2 3/16 mm: pgalloc: introduce {pud|pmd}_populate_sync()

In reply to Yang Shi

pud_populate() and pmd_populate_kernel() are called by percpu and vmemmap to populate PUD and PMD. They need to sync up kernel page table when percpu page table support for ARM64 is added in the later patch. So introduce {pud|pmd}_populate_sync() and protect them with ARCH_HAS_{PUD|PMD}_POPULATE_SYNC kernel configs. ARM64 specific implementation will be defined in the later patch.

We can change pud_populate() and pmd_populate_kernel() to take address as a new parameter, but this change is needed for all architectures and not all callsites of them need sync page table. So introducing new APIs for this usecase seems much more simple.

This is a preparation patch, no functional change.

---
 include/linux/pgalloc.h | 13 +++++++++++++
 mm/Kconfig              |  6 ++++++
 mm/percpu.c             |  4 ++--
 mm/sparse-vmemmap.c     |  4 ++--
 4 files changed, 23 insertions(+), 4 deletions(-)
include/linux/pgalloc.h +13 -0
index 9174fa59bbc5..062c64a5d4bd 100644
@@ -26,4 +26,17 @@
 			arch_sync_kernel_mappings(addr, addr);		\
 	} while (0)
 
+#ifndef CONFIG_ARCH_HAS_PUD_POPULATE_SYNC
+#define pud_populate_sync(addr, pud, pmd)				\
+	do {								\
+		pud_populate(&init_mm, pud, pmd);			\
+	} while (0)
+#endif
+
+#ifndef CONFIG_ARCH_HAS_PMD_POPULATE_SYNC
+#define pmd_populate_sync(addr, pmd, pte)				\
+	do {								\
+		pmd_populate_kernel(&init_mm, pmd, pte);		\
+	} while (0)
+#endif
 #endif /* _LINUX_PGALLOC_H */
mm/Kconfig +6 -0
index 9e0ca4824905..13ea4b29ee7b 100644
@@ -1509,6 +1509,12 @@ config LAZY_MMU_MODE_KUNIT_TEST
 
 	  If unsure, say N.
 
+config ARCH_HAS_PMD_POPULATE_SYNC
+	bool
+
+config ARCH_HAS_PUD_POPULATE_SYNC
+	bool
+
 source "mm/damon/Kconfig"
 
 endmenu
mm/percpu.c +2 -2
index b0676b8054ed..45c30c6531c5 100644
@@ -3163,7 +3163,7 @@ void __init __weak pcpu_populate_pte(unsigned long addr)
 	pud = pud_offset(p4d, addr);
 	if (pud_none(*pud)) {
 		pmd = memblock_alloc_or_panic(PMD_TABLE_SIZE, PMD_TABLE_SIZE);
-		pud_populate(&init_mm, pud, pmd);
+		pud_populate_sync(addr, pud, pmd);
 	}
 
 	pmd = pmd_offset(pud, addr);
@@ -3171,7 +3171,7 @@ void __init __weak pcpu_populate_pte(unsigned long addr)
 		pte_t *new;
 
 		new = memblock_alloc_or_panic(PTE_TABLE_SIZE, PTE_TABLE_SIZE);
-		pmd_populate_kernel(&init_mm, pmd, new);
+		pmd_populate_sync(addr, pmd, new);
 	}
 
 	return;
mm/sparse-vmemmap.c +2 -2
index 99e2be39671b..463c2f8b611c 100644
@@ -198,7 +198,7 @@ static pmd_t * __meminit vmemmap_pmd_populate(pud_t *pud, unsigned long addr, in
 		if (!p)
 			return NULL;
 		kernel_pte_init(p);
-		pmd_populate_kernel(&init_mm, pmd, p);
+		pmd_populate_sync(addr, pmd, p);
 	}
 	return pmd;
 }
@@ -211,7 +211,7 @@ static pud_t * __meminit vmemmap_pud_populate(p4d_t *p4d, unsigned long addr, in
 		if (!p)
 			return NULL;
 		pmd_init(p);
-		pud_populate(&init_mm, pud, p);
+		pud_populate_sync(addr, pud, p);
 	}
 	return pud;
 }
Signature · 1 line
2.47.0
author
#4 lore raw

PATCH v2 4/16 vmalloc: pass in pgd pointer for vmap{__vunmap}_range_noflush()

In reply to Yang Shi

vmap{__vunmap}_range_noflush() assume manipulate init_mm pgd. The following patch will map percpu local mapping into percpu page table by calling them, so the assumption will no longer stand. Make them take pgd pointer as an parameter.

Also make vmap_range_noflush() non static, it will be called outside vmalloc in the following patch.

There is no functional change.

---
 mm/internal.h    |  5 ++++-
 mm/kmsan/hooks.c | 14 +++++++-------
 mm/vmalloc.c     | 25 +++++++++++++------------
 3 files changed, 24 insertions(+), 20 deletions(-)
mm/internal.h +4 -1
index 181e79f1d6a2..9d9f08cb26a1 100644
@@ -1552,10 +1552,13 @@ void clear_vm_uninitialized_flag(struct vm_struct *vm);
 int __must_check __vmap_pages_range_noflush(unsigned long addr,
 			       unsigned long end, pgprot_t prot,
 			       struct page **pages, unsigned int page_shift);
+int __must_check vmap_range_noflush(pgd_t *pgdir, unsigned long addr,
+			unsigned long end, phys_addr_t phys_addr,
+			pgprot_t prot, unsigned int max_page_shift);
 
 void vunmap_range_noflush(unsigned long start, unsigned long end);
 
-void __vunmap_range_noflush(unsigned long start, unsigned long end);
+void __vunmap_range_noflush(pgd_t *pgdir, unsigned long start, unsigned long end);
 
 static inline bool vma_is_single_threaded_private(struct vm_area_struct *vma)
 {
mm/kmsan/hooks.c +7 -7
index 8f22d1f22981..e2a0faf344b9 100644
@@ -135,8 +135,8 @@ static unsigned long vmalloc_origin(unsigned long addr)
 
 void kmsan_vunmap_range_noflush(unsigned long start, unsigned long end)
 {
-	__vunmap_range_noflush(vmalloc_shadow(start), vmalloc_shadow(end));
-	__vunmap_range_noflush(vmalloc_origin(start), vmalloc_origin(end));
+	__vunmap_range_noflush(init_mm.pgd, vmalloc_shadow(start), vmalloc_shadow(end));
+	__vunmap_range_noflush(init_mm.pgd, vmalloc_origin(start), vmalloc_origin(end));
 	flush_cache_vmap(vmalloc_shadow(start), vmalloc_shadow(end));
 	flush_cache_vmap(vmalloc_origin(start), vmalloc_origin(end));
 }
@@ -181,7 +181,7 @@ int kmsan_ioremap_page_range(unsigned long start, unsigned long end,
 			vmalloc_origin(start + off + PAGE_SIZE), prot, &origin,
 			PAGE_SHIFT);
 		if (mapped) {
-			__vunmap_range_noflush(
+			__vunmap_range_noflush(init_mm.pgd,
 				vmalloc_shadow(start + off),
 				vmalloc_shadow(start + off + PAGE_SIZE));
 			err = mapped;
@@ -203,10 +203,10 @@ int kmsan_ioremap_page_range(unsigned long start, unsigned long end,
 			__free_pages(shadow, 1);
 		if (origin)
 			__free_pages(origin, 1);
-		__vunmap_range_noflush(
+		__vunmap_range_noflush(init_mm.pgd,
 			vmalloc_shadow(start),
 			vmalloc_shadow(start + clean * PAGE_SIZE));
-		__vunmap_range_noflush(
+		__vunmap_range_noflush(init_mm.pgd,
 			vmalloc_origin(start),
 			vmalloc_origin(start + clean * PAGE_SIZE));
 	}
@@ -233,8 +233,8 @@ void kmsan_iounmap_page_range(unsigned long start, unsigned long end)
 	     i++, v_shadow += PAGE_SIZE, v_origin += PAGE_SIZE) {
 		shadow = kmsan_vmalloc_to_page_or_null((void *)v_shadow);
 		origin = kmsan_vmalloc_to_page_or_null((void *)v_origin);
-		__vunmap_range_noflush(v_shadow, vmalloc_shadow(end));
-		__vunmap_range_noflush(v_origin, vmalloc_origin(end));
+		__vunmap_range_noflush(init_mm.pgd, v_shadow, vmalloc_shadow(end));
+		__vunmap_range_noflush(init_mm.pgd, v_origin, vmalloc_origin(end));
 		if (shadow)
 			__free_pages(shadow, 1);
 		if (origin)
mm/vmalloc.c +13 -12
index 1afca3568b9b..17c26e9796b2 100644
@@ -295,9 +295,9 @@ static int vmap_p4d_range(pgd_t *pgd, unsigned long addr, unsigned long end,
 	return err;
 }
 
-static int vmap_range_noflush(unsigned long addr, unsigned long end,
-			phys_addr_t phys_addr, pgprot_t prot,
-			unsigned int max_page_shift)
+int vmap_range_noflush(pgd_t *pgdir, unsigned long addr, unsigned long end,
+		       phys_addr_t phys_addr, pgprot_t prot,
+		       unsigned int max_page_shift)
 {
 	pgd_t *pgd;
 	unsigned long start;
@@ -314,7 +314,7 @@ static int vmap_range_noflush(unsigned long addr, unsigned long end,
 	BUG_ON(addr >= end);
 
 	start = addr;
-	pgd = pgd_offset_k(addr);
+	pgd = pgd_offset_pgd(pgdir, addr);
 	do {
 		next = pgd_addr_end(addr, end);
 		err = vmap_p4d_range(pgd, addr, next, phys_addr, prot,
@@ -334,8 +334,8 @@ int vmap_page_range(unsigned long addr, unsigned long end,
 {
 	int err;
 
-	err = vmap_range_noflush(addr, end, phys_addr, pgprot_nx(prot),
-				 ioremap_max_page_shift);
+	err = vmap_range_noflush(init_mm.pgd, addr, end, phys_addr,
+				 pgprot_nx(prot), ioremap_max_page_shift);
 	flush_cache_vmap(addr, end);
 	if (!err)
 		err = kmsan_ioremap_page_range(addr, end, phys_addr, prot,
@@ -478,7 +478,7 @@ static void vunmap_p4d_range(pgd_t *pgd, unsigned long addr, unsigned long end,
  *
  * This is an internal function only. Do not use outside mm/.
  */
-void __vunmap_range_noflush(unsigned long start, unsigned long end)
+void __vunmap_range_noflush(pgd_t *pgdir, unsigned long start, unsigned long end)
 {
 	unsigned long next;
 	pgd_t *pgd;
@@ -486,7 +486,7 @@ void __vunmap_range_noflush(unsigned long start, unsigned long end)
 	pgtbl_mod_mask mask = 0;
 
 	BUG_ON(addr >= end);
-	pgd = pgd_offset_k(addr);
+	pgd = pgd_offset_pgd(pgdir, addr);
 	do {
 		next = pgd_addr_end(addr, end);
 		if (pgd_bad(*pgd))
@@ -503,7 +503,7 @@ void __vunmap_range_noflush(unsigned long start, unsigned long end)
 void vunmap_range_noflush(unsigned long start, unsigned long end)
 {
 	kmsan_vunmap_range_noflush(start, end);
-	__vunmap_range_noflush(start, end);
+	__vunmap_range_noflush(init_mm.pgd, start, end);
 }
 
 /**
@@ -670,9 +670,10 @@ int __vmap_pages_range_noflush(unsigned long addr, unsigned long end,
 	for (i = 0; i < nr; i += 1U << (page_shift - PAGE_SHIFT)) {
 		int err;
 
-		err = vmap_range_noflush(addr, addr + (1UL << page_shift),
-					page_to_phys(pages[i]), prot,
-					page_shift);
+		err = vmap_range_noflush(init_mm.pgd, addr,
+					 addr + (1UL << page_shift),
+					 page_to_phys(pages[i]), prot,
+					 page_shift);
 		if (err)
 			return err;
 
Signature · 1 line
2.47.0
#5 lore raw

Re: [PATCH 04/16] vmalloc: pass in pgd pointer for vmap{__vunmap}_range_noflush()

Quoted 1 line · Yang Shi

vmap{__vunmap}_range_noflush() assume manipulate init_mm pgd. The

It's not an assumption, it's a requirement.

There is now (and really should have always been) strictly a locking requirement that the mm being manipulated is init_mm.

Quoted 3 lines

following patch will map percpu local mapping into percpu page table by calling them, so the assumption will no longer stand. Make them take pgd pointer as an parameter.

No, this isn't OK.

vmalloc is expressly for manipulating kernel mappings in init_mm.

Quoted 3 lines

Also make vmap_range_noflush() non static, it will be called outside vmalloc in the following patch.

Hmm not loving the exposure of an internal function ehre.

Quoted 4 lines

There is no functional change.

You're fundamentally changing something here, conceptually. I'm not sure why you specifically need vmalloc to do it but it doesn't strike me as worthwhile for what you're doing.

Quoted 152 lines
---
 mm/internal.h    |  5 ++++-
 mm/kmsan/hooks.c | 14 +++++++-------
 mm/vmalloc.c     | 25 +++++++++++++------------
 3 files changed, 24 insertions(+), 20 deletions(-)
mm/internal.h +4 -1
index 181e79f1d6a2..9d9f08cb26a1 100644
@@ -1552,10 +1552,13 @@ void clear_vm_uninitialized_flag(struct vm_struct *vm);
 int __must_check __vmap_pages_range_noflush(unsigned long addr,
 			       unsigned long end, pgprot_t prot,
 			       struct page **pages, unsigned int page_shift);
+int __must_check vmap_range_noflush(pgd_t *pgdir, unsigned long addr,
+			unsigned long end, phys_addr_t phys_addr,
+			pgprot_t prot, unsigned int max_page_shift);
 
 void vunmap_range_noflush(unsigned long start, unsigned long end);
 
-void __vunmap_range_noflush(unsigned long start, unsigned long end);
+void __vunmap_range_noflush(pgd_t *pgdir, unsigned long start, unsigned long end);
 
 static inline bool vma_is_single_threaded_private(struct vm_area_struct *vma)
 {
mm/kmsan/hooks.c +7 -7
index 8f22d1f22981..e2a0faf344b9 100644
@@ -135,8 +135,8 @@ static unsigned long vmalloc_origin(unsigned long addr)
 
 void kmsan_vunmap_range_noflush(unsigned long start, unsigned long end)
 {
-	__vunmap_range_noflush(vmalloc_shadow(start), vmalloc_shadow(end));
-	__vunmap_range_noflush(vmalloc_origin(start), vmalloc_origin(end));
+	__vunmap_range_noflush(init_mm.pgd, vmalloc_shadow(start), vmalloc_shadow(end));
+	__vunmap_range_noflush(init_mm.pgd, vmalloc_origin(start), vmalloc_origin(end));
 	flush_cache_vmap(vmalloc_shadow(start), vmalloc_shadow(end));
 	flush_cache_vmap(vmalloc_origin(start), vmalloc_origin(end));
 }
@@ -181,7 +181,7 @@ int kmsan_ioremap_page_range(unsigned long start, unsigned long end,
 			vmalloc_origin(start + off + PAGE_SIZE), prot, &origin,
 			PAGE_SHIFT);
 		if (mapped) {
-			__vunmap_range_noflush(
+			__vunmap_range_noflush(init_mm.pgd,
 				vmalloc_shadow(start + off),
 				vmalloc_shadow(start + off + PAGE_SIZE));
 			err = mapped;
@@ -203,10 +203,10 @@ int kmsan_ioremap_page_range(unsigned long start, unsigned long end,
 			__free_pages(shadow, 1);
 		if (origin)
 			__free_pages(origin, 1);
-		__vunmap_range_noflush(
+		__vunmap_range_noflush(init_mm.pgd,
 			vmalloc_shadow(start),
 			vmalloc_shadow(start + clean * PAGE_SIZE));
-		__vunmap_range_noflush(
+		__vunmap_range_noflush(init_mm.pgd,
 			vmalloc_origin(start),
 			vmalloc_origin(start + clean * PAGE_SIZE));
 	}
@@ -233,8 +233,8 @@ void kmsan_iounmap_page_range(unsigned long start, unsigned long end)
 	     i++, v_shadow += PAGE_SIZE, v_origin += PAGE_SIZE) {
 		shadow = kmsan_vmalloc_to_page_or_null((void *)v_shadow);
 		origin = kmsan_vmalloc_to_page_or_null((void *)v_origin);
-		__vunmap_range_noflush(v_shadow, vmalloc_shadow(end));
-		__vunmap_range_noflush(v_origin, vmalloc_origin(end));
+		__vunmap_range_noflush(init_mm.pgd, v_shadow, vmalloc_shadow(end));
+		__vunmap_range_noflush(init_mm.pgd, v_origin, vmalloc_origin(end));
 		if (shadow)
 			__free_pages(shadow, 1);
 		if (origin)
mm/vmalloc.c +13 -13
index 1afca3568b9b..17c26e9796b2 100644
@@ -295,9 +295,9 @@ static int vmap_p4d_range(pgd_t *pgd, unsigned long addr, unsigned long end,
 	return err;
 }
 
-static int vmap_range_noflush(unsigned long addr, unsigned long end,
-			phys_addr_t phys_addr, pgprot_t prot,
-			unsigned int max_page_shift)
+int vmap_range_noflush(pgd_t *pgdir, unsigned long addr, unsigned long end,
+		       phys_addr_t phys_addr, pgprot_t prot,
+		       unsigned int max_page_shift)
 {
 	pgd_t *pgd;
 	unsigned long start;
@@ -314,7 +314,7 @@ static int vmap_range_noflush(unsigned long addr, unsigned long end,
 	BUG_ON(addr >= end);
 
 	start = addr;
-	pgd = pgd_offset_k(addr);
+	pgd = pgd_offset_pgd(pgdir, addr);
 	do {
 		next = pgd_addr_end(addr, end);
 		err = vmap_p4d_range(pgd, addr, next, phys_addr, prot,
@@ -334,8 +334,8 @@ int vmap_page_range(unsigned long addr, unsigned long end,
 {
 	int err;
 
-	err = vmap_range_noflush(addr, end, phys_addr, pgprot_nx(prot),
-				 ioremap_max_page_shift);
+	err = vmap_range_noflush(init_mm.pgd, addr, end, phys_addr,
+				 pgprot_nx(prot), ioremap_max_page_shift);
 	flush_cache_vmap(addr, end);
 	if (!err)
 		err = kmsan_ioremap_page_range(addr, end, phys_addr, prot,
@@ -478,7 +478,7 @@ static void vunmap_p4d_range(pgd_t *pgd, unsigned long addr, unsigned long end,
  *
  * This is an internal function only. Do not use outside mm/.
  */
-void __vunmap_range_noflush(unsigned long start, unsigned long end)
+void __vunmap_range_noflush(pgd_t *pgdir, unsigned long start, unsigned long end)
 {
 	unsigned long next;
 	pgd_t *pgd;
@@ -486,7 +486,7 @@ void __vunmap_range_noflush(unsigned long start, unsigned long end)
 	pgtbl_mod_mask mask = 0;
 
 	BUG_ON(addr >= end);
-	pgd = pgd_offset_k(addr);
+	pgd = pgd_offset_pgd(pgdir, addr);
 	do {
 		next = pgd_addr_end(addr, end);
 		if (pgd_bad(*pgd))
@@ -503,7 +503,7 @@ void __vunmap_range_noflush(unsigned long start, unsigned long end)
 void vunmap_range_noflush(unsigned long start, unsigned long end)
 {
 	kmsan_vunmap_range_noflush(start, end);
-	__vunmap_range_noflush(start, end);
+	__vunmap_range_noflush(init_mm.pgd, start, end);
 }
 
 /**
@@ -670,9 +670,10 @@ int __vmap_pages_range_noflush(unsigned long addr, unsigned long end,
 	for (i = 0; i < nr; i += 1U << (page_shift - PAGE_SHIFT)) {
 		int err;
 
-		err = vmap_range_noflush(addr, addr + (1UL << page_shift),
-					page_to_phys(pages[i]), prot,
-					page_shift);
+		err = vmap_range_noflush(init_mm.pgd, addr,
+					 addr + (1UL << page_shift),
+					 page_to_phys(pages[i]), prot,
+					 page_shift);
 		if (err)
 			return err;
 
--
Quoted 4 lines

2.47.0

Cheers, Lorenzo

author
#6 lore raw

Re: [PATCH 04/16] vmalloc: pass in pgd pointer for vmap{__vunmap}_range_noflush()

Quoted 7 lines · Lorenzo Stoakes (ARM)
Quoted 1 line · Yang Shi

vmap{__vunmap}_range_noflush() assume manipulate init_mm pgd. The

It's not an assumption, it's a requirement.

There is now (and really should have always been) strictly a locking requirement that the mm being manipulated is init_mm.

Quoted 6 lines
Quoted 3 lines

following patch will map percpu local mapping into percpu page table by calling them, so the assumption will no longer stand. Make them take pgd pointer as an parameter.

No, this isn't OK.

vmalloc is expressly for manipulating kernel mappings in init_mm.

This commit message may be a little bit misleading. vmalloc still manipulates init_mm all the time. The percpu allocator uses some vmalloc APIs to map memory. So the APIs used by percpu allocator needs to manipulate percpu page table when mapping memory for the new percpu area added by the following patches. So percpu pgd is just used when mapping the specific percpu area by percpu allocator.

Quoted 4 lines
Quoted 3 lines

Also make vmap_range_noflush() non static, it will be called outside vmalloc in the following patch.

Hmm not loving the exposure of an internal function ehre.

The RFC reused and extended the vmalloc APIs. We don't have to do so if it is not preferred. I think I can create some percpu allocator specific APIs to manipulate percpu area mapping with percpu pgd.

Quoted 7 lines
Quoted 4 lines

There is no functional change.

You're fundamentally changing something here, conceptually. I'm not sure why you specifically need vmalloc to do it but it doesn't strike me as worthwhile for what you're doing.

The percpu allocator uses vmalloc APIs to map percpu variables currently. So the patch extended it to take pgd pointer as an extra parameter. This helped to reduce duplicate code. I can add percpu specific APIs if extending the APIs doesn't make too much sense.

Thanks,
Yang

Quoted 158 lines
Quoted 153 lines
---
  mm/internal.h    |  5 ++++-
  mm/kmsan/hooks.c | 14 +++++++-------
  mm/vmalloc.c     | 25 +++++++++++++------------
  3 files changed, 24 insertions(+), 20 deletions(-)
mm/internal.h +4 -1
index 181e79f1d6a2..9d9f08cb26a1 100644
@@ -1552,10 +1552,13 @@ void clear_vm_uninitialized_flag(struct vm_struct *vm);
  int __must_check __vmap_pages_range_noflush(unsigned long addr,
  			       unsigned long end, pgprot_t prot,
  			       struct page **pages, unsigned int page_shift);
+int __must_check vmap_range_noflush(pgd_t *pgdir, unsigned long addr,
+			unsigned long end, phys_addr_t phys_addr,
+			pgprot_t prot, unsigned int max_page_shift);
 
  void vunmap_range_noflush(unsigned long start, unsigned long end);
 
-void __vunmap_range_noflush(unsigned long start, unsigned long end);
+void __vunmap_range_noflush(pgd_t *pgdir, unsigned long start, unsigned long end);
 
  static inline bool vma_is_single_threaded_private(struct vm_area_struct *vma)
  {
mm/kmsan/hooks.c +7 -7
index 8f22d1f22981..e2a0faf344b9 100644
@@ -135,8 +135,8 @@ static unsigned long vmalloc_origin(unsigned long addr)
 
  void kmsan_vunmap_range_noflush(unsigned long start, unsigned long end)
  {
-	__vunmap_range_noflush(vmalloc_shadow(start), vmalloc_shadow(end));
-	__vunmap_range_noflush(vmalloc_origin(start), vmalloc_origin(end));
+	__vunmap_range_noflush(init_mm.pgd, vmalloc_shadow(start), vmalloc_shadow(end));
+	__vunmap_range_noflush(init_mm.pgd, vmalloc_origin(start), vmalloc_origin(end));
  	flush_cache_vmap(vmalloc_shadow(start), vmalloc_shadow(end));
  	flush_cache_vmap(vmalloc_origin(start), vmalloc_origin(end));
  }
@@ -181,7 +181,7 @@ int kmsan_ioremap_page_range(unsigned long start, unsigned long end,
  			vmalloc_origin(start + off + PAGE_SIZE), prot, &origin,
  			PAGE_SHIFT);
  		if (mapped) {
-			__vunmap_range_noflush(
+			__vunmap_range_noflush(init_mm.pgd,
  				vmalloc_shadow(start + off),
  				vmalloc_shadow(start + off + PAGE_SIZE));
  			err = mapped;
@@ -203,10 +203,10 @@ int kmsan_ioremap_page_range(unsigned long start, unsigned long end,
  			__free_pages(shadow, 1);
  		if (origin)
  			__free_pages(origin, 1);
-		__vunmap_range_noflush(
+		__vunmap_range_noflush(init_mm.pgd,
  			vmalloc_shadow(start),
  			vmalloc_shadow(start + clean * PAGE_SIZE));
-		__vunmap_range_noflush(
+		__vunmap_range_noflush(init_mm.pgd,
  			vmalloc_origin(start),
  			vmalloc_origin(start + clean * PAGE_SIZE));
  	}
@@ -233,8 +233,8 @@ void kmsan_iounmap_page_range(unsigned long start, unsigned long end)
  	     i++, v_shadow += PAGE_SIZE, v_origin += PAGE_SIZE) {
  		shadow = kmsan_vmalloc_to_page_or_null((void *)v_shadow);
  		origin = kmsan_vmalloc_to_page_or_null((void *)v_origin);
-		__vunmap_range_noflush(v_shadow, vmalloc_shadow(end));
-		__vunmap_range_noflush(v_origin, vmalloc_origin(end));
+		__vunmap_range_noflush(init_mm.pgd, v_shadow, vmalloc_shadow(end));
+		__vunmap_range_noflush(init_mm.pgd, v_origin, vmalloc_origin(end));
  		if (shadow)
  			__free_pages(shadow, 1);
  		if (origin)
mm/vmalloc.c +13 -13
index 1afca3568b9b..17c26e9796b2 100644
@@ -295,9 +295,9 @@ static int vmap_p4d_range(pgd_t *pgd, unsigned long addr, unsigned long end,
  	return err;
  }
 
-static int vmap_range_noflush(unsigned long addr, unsigned long end,
-			phys_addr_t phys_addr, pgprot_t prot,
-			unsigned int max_page_shift)
+int vmap_range_noflush(pgd_t *pgdir, unsigned long addr, unsigned long end,
+		       phys_addr_t phys_addr, pgprot_t prot,
+		       unsigned int max_page_shift)
  {
  	pgd_t *pgd;
  	unsigned long start;
@@ -314,7 +314,7 @@ static int vmap_range_noflush(unsigned long addr, unsigned long end,
  	BUG_ON(addr >= end);
 
  	start = addr;
-	pgd = pgd_offset_k(addr);
+	pgd = pgd_offset_pgd(pgdir, addr);
  	do {
  		next = pgd_addr_end(addr, end);
  		err = vmap_p4d_range(pgd, addr, next, phys_addr, prot,
@@ -334,8 +334,8 @@ int vmap_page_range(unsigned long addr, unsigned long end,
  {
  	int err;
 
-	err = vmap_range_noflush(addr, end, phys_addr, pgprot_nx(prot),
-				 ioremap_max_page_shift);
+	err = vmap_range_noflush(init_mm.pgd, addr, end, phys_addr,
+				 pgprot_nx(prot), ioremap_max_page_shift);
  	flush_cache_vmap(addr, end);
  	if (!err)
  		err = kmsan_ioremap_page_range(addr, end, phys_addr, prot,
@@ -478,7 +478,7 @@ static void vunmap_p4d_range(pgd_t *pgd, unsigned long addr, unsigned long end,
   *
   * This is an internal function only. Do not use outside mm/.
   */
-void __vunmap_range_noflush(unsigned long start, unsigned long end)
+void __vunmap_range_noflush(pgd_t *pgdir, unsigned long start, unsigned long end)
  {
  	unsigned long next;
  	pgd_t *pgd;
@@ -486,7 +486,7 @@ void __vunmap_range_noflush(unsigned long start, unsigned long end)
  	pgtbl_mod_mask mask = 0;
 
  	BUG_ON(addr >= end);
-	pgd = pgd_offset_k(addr);
+	pgd = pgd_offset_pgd(pgdir, addr);
  	do {
  		next = pgd_addr_end(addr, end);
  		if (pgd_bad(*pgd))
@@ -503,7 +503,7 @@ void __vunmap_range_noflush(unsigned long start, unsigned long end)
  void vunmap_range_noflush(unsigned long start, unsigned long end)
  {
  	kmsan_vunmap_range_noflush(start, end);
-	__vunmap_range_noflush(start, end);
+	__vunmap_range_noflush(init_mm.pgd, start, end);
  }
 
  /**
@@ -670,9 +670,10 @@ int __vmap_pages_range_noflush(unsigned long addr, unsigned long end,
  	for (i = 0; i < nr; i += 1U << (page_shift - PAGE_SHIFT)) {
  		int err;
 
-		err = vmap_range_noflush(addr, addr + (1UL << page_shift),
-					page_to_phys(pages[i]), prot,
-					page_shift);
+		err = vmap_range_noflush(init_mm.pgd, addr,
+					 addr + (1UL << page_shift),
+					 page_to_phys(pages[i]), prot,
+					 page_shift);
  		if (err)
  			return err;
 
--
Quoted 4 lines

2.47.0

Cheers, Lorenzo

#7 lore raw

Re: [PATCH 04/16] vmalloc: pass in pgd pointer for vmap{__vunmap}_range_noflush()

Quoted 10 lines · Yang Shi
Quoted 7 lines · Lorenzo Stoakes (ARM)
Quoted 1 line · Yang Shi

vmap{__vunmap}_range_noflush() assume manipulate init_mm pgd. The

It's not an assumption, it's a requirement.

There is now (and really should have always been) strictly a locking requirement that the mm being manipulated is init_mm.

Quoted 8 lines
Quoted 6 lines
Quoted 3 lines

following patch will map percpu local mapping into percpu page table by calling them, so the assumption will no longer stand. Make them take pgd pointer as an parameter.

No, this isn't OK.

vmalloc is expressly for manipulating kernel mappings in init_mm.

This commit message may be a little bit misleading. vmalloc still

It is - it's worse than that implies :) as you're actually allowing arbitrary mm's to be manipulated.

Quoted 5 lines

manipulates init_mm all the time. The percpu allocator uses some vmalloc APIs to map memory. So the APIs used by percpu allocator needs to manipulate percpu page table when mapping memory for the new percpu area added by the following patches. So percpu pgd is just used when mapping the specific percpu area by percpu allocator.

Yup I understand all that, but my objection remains the same. You're essentially hacking something in to the vmalloc code which adds complexity and confusion.

(The correct version of this, assuming the per-CPU page table stuff was acceptable, would be to separate out common code and have vmalloc and the per-CPU page table stuff use the common code :)

Quoted 14 lines
Quoted 5 lines
Quoted 4 lines

Also make vmap_range_noflush() non static, it will be called outside vmalloc in the following patch.

Hmm not loving the exposure of an internal function ehre.

The RFC reused and extended the vmalloc APIs. We don't have to do so if it is not preferred. I think I can create some percpu allocator specific APIs to manipulate percpu area mapping with percpu pgd.

Quoted 4 lines
Quoted 4 lines

There is no functional change.

Quoted 8 lines
Quoted 3 lines

You're fundamentally changing something here, conceptually. I'm not sure why you specifically need vmalloc to do it but it doesn't strike me as worthwhile for what you're doing.

The percpu allocator uses vmalloc APIs to map percpu variables currently. So the patch extended it to take pgd pointer as an extra parameter. This helped to reduce duplicate code. I can add percpu specific APIs if extending the APIs doesn't make too much sense.

I mean again, the correct version of this would be to separate out common code and refactor etc. etc.

But yes, I'd rather the vmalloc code is left as-is. Doing what you're proposing in this patch adds complexity and confusion as to whether remote mm's can be updated.

(Overall my impression is that the per-CPU page table approach isn't something Linus wants and there's push back from the commnity, so maybe worth discussing the approach in general first rather than respinning?)

Quoted 163 lines

Thanks,
Yang

Quoted 158 lines
Quoted 153 lines
---
  mm/internal.h    |  5 ++++-
  mm/kmsan/hooks.c | 14 +++++++-------
  mm/vmalloc.c     | 25 +++++++++++++------------
  3 files changed, 24 insertions(+), 20 deletions(-)
mm/internal.h +4 -1
index 181e79f1d6a2..9d9f08cb26a1 100644
@@ -1552,10 +1552,13 @@ void clear_vm_uninitialized_flag(struct vm_struct *vm);
  int __must_check __vmap_pages_range_noflush(unsigned long addr,
  			       unsigned long end, pgprot_t prot,
  			       struct page **pages, unsigned int page_shift);
+int __must_check vmap_range_noflush(pgd_t *pgdir, unsigned long addr,
+			unsigned long end, phys_addr_t phys_addr,
+			pgprot_t prot, unsigned int max_page_shift);
 
  void vunmap_range_noflush(unsigned long start, unsigned long end);
 
-void __vunmap_range_noflush(unsigned long start, unsigned long end);
+void __vunmap_range_noflush(pgd_t *pgdir, unsigned long start, unsigned long end);
 
  static inline bool vma_is_single_threaded_private(struct vm_area_struct *vma)
  {
mm/kmsan/hooks.c +7 -7
index 8f22d1f22981..e2a0faf344b9 100644
@@ -135,8 +135,8 @@ static unsigned long vmalloc_origin(unsigned long addr)
 
  void kmsan_vunmap_range_noflush(unsigned long start, unsigned long end)
  {
-	__vunmap_range_noflush(vmalloc_shadow(start), vmalloc_shadow(end));
-	__vunmap_range_noflush(vmalloc_origin(start), vmalloc_origin(end));
+	__vunmap_range_noflush(init_mm.pgd, vmalloc_shadow(start), vmalloc_shadow(end));
+	__vunmap_range_noflush(init_mm.pgd, vmalloc_origin(start), vmalloc_origin(end));
  	flush_cache_vmap(vmalloc_shadow(start), vmalloc_shadow(end));
  	flush_cache_vmap(vmalloc_origin(start), vmalloc_origin(end));
  }
@@ -181,7 +181,7 @@ int kmsan_ioremap_page_range(unsigned long start, unsigned long end,
  			vmalloc_origin(start + off + PAGE_SIZE), prot, &origin,
  			PAGE_SHIFT);
  		if (mapped) {
-			__vunmap_range_noflush(
+			__vunmap_range_noflush(init_mm.pgd,
  				vmalloc_shadow(start + off),
  				vmalloc_shadow(start + off + PAGE_SIZE));
  			err = mapped;
@@ -203,10 +203,10 @@ int kmsan_ioremap_page_range(unsigned long start, unsigned long end,
  			__free_pages(shadow, 1);
  		if (origin)
  			__free_pages(origin, 1);
-		__vunmap_range_noflush(
+		__vunmap_range_noflush(init_mm.pgd,
  			vmalloc_shadow(start),
  			vmalloc_shadow(start + clean * PAGE_SIZE));
-		__vunmap_range_noflush(
+		__vunmap_range_noflush(init_mm.pgd,
  			vmalloc_origin(start),
  			vmalloc_origin(start + clean * PAGE_SIZE));
  	}
@@ -233,8 +233,8 @@ void kmsan_iounmap_page_range(unsigned long start, unsigned long end)
  	     i++, v_shadow += PAGE_SIZE, v_origin += PAGE_SIZE) {
  		shadow = kmsan_vmalloc_to_page_or_null((void *)v_shadow);
  		origin = kmsan_vmalloc_to_page_or_null((void *)v_origin);
-		__vunmap_range_noflush(v_shadow, vmalloc_shadow(end));
-		__vunmap_range_noflush(v_origin, vmalloc_origin(end));
+		__vunmap_range_noflush(init_mm.pgd, v_shadow, vmalloc_shadow(end));
+		__vunmap_range_noflush(init_mm.pgd, v_origin, vmalloc_origin(end));
  		if (shadow)
  			__free_pages(shadow, 1);
  		if (origin)
mm/vmalloc.c +13 -13
index 1afca3568b9b..17c26e9796b2 100644
@@ -295,9 +295,9 @@ static int vmap_p4d_range(pgd_t *pgd, unsigned long addr, unsigned long end,
  	return err;
  }
 
-static int vmap_range_noflush(unsigned long addr, unsigned long end,
-			phys_addr_t phys_addr, pgprot_t prot,
-			unsigned int max_page_shift)
+int vmap_range_noflush(pgd_t *pgdir, unsigned long addr, unsigned long end,
+		       phys_addr_t phys_addr, pgprot_t prot,
+		       unsigned int max_page_shift)
  {
  	pgd_t *pgd;
  	unsigned long start;
@@ -314,7 +314,7 @@ static int vmap_range_noflush(unsigned long addr, unsigned long end,
  	BUG_ON(addr >= end);
 
  	start = addr;
-	pgd = pgd_offset_k(addr);
+	pgd = pgd_offset_pgd(pgdir, addr);
  	do {
  		next = pgd_addr_end(addr, end);
  		err = vmap_p4d_range(pgd, addr, next, phys_addr, prot,
@@ -334,8 +334,8 @@ int vmap_page_range(unsigned long addr, unsigned long end,
  {
  	int err;
 
-	err = vmap_range_noflush(addr, end, phys_addr, pgprot_nx(prot),
-				 ioremap_max_page_shift);
+	err = vmap_range_noflush(init_mm.pgd, addr, end, phys_addr,
+				 pgprot_nx(prot), ioremap_max_page_shift);
  	flush_cache_vmap(addr, end);
  	if (!err)
  		err = kmsan_ioremap_page_range(addr, end, phys_addr, prot,
@@ -478,7 +478,7 @@ static void vunmap_p4d_range(pgd_t *pgd, unsigned long addr, unsigned long end,
   *
   * This is an internal function only. Do not use outside mm/.
   */
-void __vunmap_range_noflush(unsigned long start, unsigned long end)
+void __vunmap_range_noflush(pgd_t *pgdir, unsigned long start, unsigned long end)
  {
  	unsigned long next;
  	pgd_t *pgd;
@@ -486,7 +486,7 @@ void __vunmap_range_noflush(unsigned long start, unsigned long end)
  	pgtbl_mod_mask mask = 0;
 
  	BUG_ON(addr >= end);
-	pgd = pgd_offset_k(addr);
+	pgd = pgd_offset_pgd(pgdir, addr);
  	do {
  		next = pgd_addr_end(addr, end);
  		if (pgd_bad(*pgd))
@@ -503,7 +503,7 @@ void __vunmap_range_noflush(unsigned long start, unsigned long end)
  void vunmap_range_noflush(unsigned long start, unsigned long end)
  {
  	kmsan_vunmap_range_noflush(start, end);
-	__vunmap_range_noflush(start, end);
+	__vunmap_range_noflush(init_mm.pgd, start, end);
  }
 
  /**
@@ -670,9 +670,10 @@ int __vmap_pages_range_noflush(unsigned long addr, unsigned long end,
  	for (i = 0; i < nr; i += 1U << (page_shift - PAGE_SHIFT)) {
  		int err;
 
-		err = vmap_range_noflush(addr, addr + (1UL << page_shift),
-					page_to_phys(pages[i]), prot,
-					page_shift);
+		err = vmap_range_noflush(init_mm.pgd, addr,
+					 addr + (1UL << page_shift),
+					 page_to_phys(pages[i]), prot,
+					 page_shift);
  		if (err)
  			return err;
 
--
Quoted 4 lines

2.47.0

Cheers, Lorenzo

--
Cheers, Lorenzo

author
#8 lore raw

PATCH v2 5/16 arm64: mm: enable percpu kernel page table

In reply to Yang Shi

Currently all cpus share the same kernel page table (swapper_pg_dir), this patch creates kernel page table for each cpu and each cpu uses its own kernel page table. The cpu 0 keeps using swapper_pg_dir. All the kernel page tables share the same content. So we don't have to duplicate the whole page table for all cpus, we just need to have different pgd page for each cpu. All kernel page table modification (split, creation, deletion, etc) actually still happens on swapper_pg_dir, the modification needs to be synchronized to other cpu's page tables when the top level is modified.

The percpu page table can't be shared across cores, so clear cnp bit too even though CNP is supported.

Some features may not work with it for now, for example, memory hotplug, KASAN, etc. They will be fixed in the following patches.

---
 arch/arm64/include/asm/mmu.h         |  1 +
 arch/arm64/include/asm/mmu_context.h |  9 ++++-
 arch/arm64/include/asm/pgtable.h     | 10 ++++++
 arch/arm64/kernel/setup.c            |  3 ++
 arch/arm64/kernel/smp.c              |  8 +++++
 arch/arm64/mm/mmu.c                  | 53 ++++++++++++++++++++++++++++
 6 files changed, 83 insertions(+), 1 deletion(-)
arch/arm64/include/asm/mmu.h +1 -0
index 5e1211c540ab..8ed3b5f3cf84 100644
@@ -63,6 +63,7 @@ static inline bool arm64_kernel_unmapped_at_el0(void)
 extern void arm64_memblock_init(void);
 extern void paging_init(void);
 extern void bootmem_init(void);
+extern void setup_percpu_pgd(void);
 extern void create_mapping_noalloc(phys_addr_t phys, unsigned long virt,
 				   phys_addr_t size, pgprot_t prot);
 extern void create_pgd_mapping(struct mm_struct *mm, phys_addr_t phys,
arch/arm64/include/asm/mmu_context.h +8 -1
index 803b68758152..0e991b98ed9e 100644
@@ -27,6 +27,7 @@
 #include <asm/tlbflush.h>
 
 extern bool rodata_full;
+extern pgd_t *percpu_pgd[NR_CPUS];
 
 static inline void contextidr_thread_switch(struct task_struct *next)
 {
@@ -138,7 +139,13 @@ void __cpu_replace_ttbr1(pgd_t *pgdp, bool cnp);
 
 static inline void cpu_enable_swapper_cnp(void)
 {
-	__cpu_replace_ttbr1(lm_alias(swapper_pg_dir), true);
+	unsigned int cpu = smp_processor_id();
+	pgd_t *ttbr1 = percpu_pgd[cpu];
+
+	if (cpu == 0)
+		ttbr1 = lm_alias(swapper_pg_dir);
+
+	__cpu_replace_ttbr1(ttbr1, false);
 }
 
 static inline void cpu_replace_ttbr1(pgd_t *pgdp)
arch/arm64/include/asm/pgtable.h +10 -0
index 27689c62bd25..13e069b01600 100644
@@ -1216,6 +1216,16 @@ p4d_t *p4d_offset_lockless_folded(pgd_t *pgdp, pgd_t pgd, unsigned long addr)
 
 #endif  /* CONFIG_PGTABLE_LEVELS > 4 */
 
+#if CONFIG_PGTABLE_LEVELS == 2
+#define ARCH_PAGE_TABLE_SYNC_MASK      PGTBL_PMD_MODIFIED
+#elif CONFIG_PGTABLE_LEVELS == 3
+#define ARCH_PAGE_TABLE_SYNC_MASK      PGTBL_PUD_MODIFIED
+#elif CONFIG_PGTABLE_LEVELS == 4
+#define ARCH_PAGE_TABLE_SYNC_MASK      PGTBL_P4D_MODIFIED
+#elif CONFIG_PGTABLE_LEVELS == 5
+#define ARCH_PAGE_TABLE_SYNC_MASK      PGTBL_PGD_MODIFIED
+#endif
+
 #define pgd_ERROR(e)	\
 	pr_err("%s:%d: bad pgd %016llx.\n", __FILE__, __LINE__, pgd_val(e))
 
arch/arm64/kernel/setup.c +3 -0
index 23c05dc7a8f2..6d420ad59af4 100644
@@ -360,6 +360,9 @@ void __init __no_sanitize_address setup_arch(char **cmdline_p)
 	smp_init_cpus();
 	smp_build_mpidr_hash();
 
+	/* Must be called after smp_init_cpus */
+	setup_percpu_pgd();
+
 #ifdef CONFIG_ARM64_SW_TTBR0_PAN
 	/*
 	 * Make sure init_thread_info.ttbr0 always generates translation
arch/arm64/kernel/smp.c +8 -0
index 28ff889658d3..49524a7176b2 100644
@@ -55,6 +55,8 @@
 
 #include <trace/events/ipi.h>
 
+extern void idmap_cpu_replace_ttbr1(phys_addr_t pgdir);
+
 /*
  * as from 2.5, kernels no longer have an init_tasks structure
  * so we need some other way of telling a new secondary core
@@ -198,6 +200,12 @@ asmlinkage notrace void secondary_start_kernel(void)
 	struct mm_struct *mm = &init_mm;
 	const struct cpu_operations *ops;
 	unsigned int cpu = smp_processor_id();
+	typedef void (ttbr_replace_func)(phys_addr_t);
+	ttbr_replace_func *replace_ttbr;
+
+	phys_addr_t ttbr1 = phys_to_ttbr(virt_to_phys(percpu_pgd[cpu]));
+	replace_ttbr = (void *)__pa_symbol(idmap_cpu_replace_ttbr1);
+	replace_ttbr(ttbr1);
 
 	/*
 	 * All kernel threads share the same mm context; grab a
arch/arm64/mm/mmu.c +53 -0

59 lines not on this page · raw message

Signature · 1 line
2.47.0
author
#9 lore raw

PATCH v2 6/16 arm64: mm: defined {pud|pmd}_populate_sync()

In reply to Yang Shi

Implement ARM64 specific {pud|pmd}_populate_sync(), which sync up page table.

---
 arch/arm64/Kconfig               |  2 ++
 arch/arm64/include/asm/pgalloc.h | 24 ++++++++++++++++++++++++
 2 files changed, 26 insertions(+)
arch/arm64/Kconfig +2 -0
index 2e1bd79a579c..23cad9fc7046 100644
@@ -41,9 +41,11 @@ config ARM64
 	select ARCH_HAS_NMI_SAFE_THIS_CPU_OPS
 	select ARCH_HAS_NON_OVERLAPPING_ADDRESS_SPACE
 	select ARCH_HAS_NONLEAF_PMD_YOUNG if ARM64_HAFT
+	select ARCH_HAS_PMD_POPULATE_SYNC
 	select ARCH_HAS_PREEMPT_LAZY
 	select ARCH_HAS_PTDUMP
 	select ARCH_HAS_PTE_SPECIAL
+	select ARCH_HAS_PUD_POPULATE_SYNC
 	select ARCH_HAS_HW_PTE_YOUNG
 	select ARCH_HAS_SETUP_DMA_OPS
 	select ARCH_HAS_SET_DIRECT_MAP
arch/arm64/include/asm/pgalloc.h +24 -0
index 1b4509d3382c..dd1ac84a1160 100644
@@ -33,11 +33,27 @@ static inline void pud_populate(struct mm_struct *mm, pud_t *pudp, pmd_t *pmdp)
 	pudval |= (mm == &init_mm) ? PUD_TABLE_UXN : PUD_TABLE_PXN;
 	__pud_populate(pudp, __pa(pmdp), pudval);
 }
+
+#ifdef CONFIG_ARCH_HAS_PUD_POPULATE_SYNC
+static inline void pud_populate_sync(unsigned long addr, pud_t *pudp, pmd_t *pmdp)
+{
+	pud_populate(&init_mm, pudp, pmdp);
+	if (ARCH_PAGE_TABLE_SYNC_MASK & PGTBL_PUD_MODIFIED)
+		arch_sync_kernel_mappings(addr, addr);
+}
+#endif
 #else
 static inline void __pud_populate(pud_t *pudp, phys_addr_t pmdp, pudval_t prot)
 {
 	BUILD_BUG();
 }
+
+#ifdef CONFIG_ARCH_HAS_PUD_POPULATE_SYNC
+static inline void pud_populate_sync(unsigned long addr, pud_t *pudp, pmd_t *pmdp)
+{
+	BUILD_BUG();
+}
+#endif
 #endif	/* CONFIG_PGTABLE_LEVELS > 2 */
 
 #if CONFIG_PGTABLE_LEVELS > 3
@@ -121,4 +137,12 @@ pmd_populate(struct mm_struct *mm, pmd_t *pmdp, pgtable_t ptep)
 		       PMD_TYPE_TABLE | PMD_TABLE_AF | PMD_TABLE_PXN);
 }
 
+#ifdef CONFIG_ARCH_HAS_PMD_POPULATE_SYNC
+static inline void pmd_populate_sync(unsigned long addr, pmd_t *pmdp, pte_t *ptep)
+{
+	pmd_populate_kernel(&init_mm, pmdp, ptep);
+	if (ARCH_PAGE_TABLE_SYNC_MASK & PGTBL_PMD_MODIFIED)
+		arch_sync_kernel_mappings(addr, addr);
+}
+#endif
 #endif
Signature · 1 line
2.47.0
author
#10 lore raw

PATCH v2 7/16 arm64: mm: sync percpu page table for memory hotplug/unplug

In reply to Yang Shi

Need to sync kernel percpu page table for memory hotplug/unplug, including both linear mapping and vmemmap.

---
 arch/arm64/mm/mmu.c | 85 ++++++++++++++++++++++++++++++---------------
 1 file changed, 57 insertions(+), 28 deletions(-)
arch/arm64/mm/mmu.c +57 -28
index 5a235491f268..e1290ccaa5be 100644
@@ -214,7 +214,7 @@ static void init_clear_pgtable(void *table)
 }
 
 static void init_pte(pte_t *ptep, unsigned long addr, unsigned long end,
-		     phys_addr_t phys, pgprot_t prot)
+		     phys_addr_t phys, pgprot_t prot, pgtbl_mod_mask *mask)
 {
 	do {
 		pte_t old_pte = __ptep_get(ptep);
@@ -224,6 +224,7 @@ static void init_pte(pte_t *ptep, unsigned long addr, unsigned long end,
 		 * are deferred to the end of alloc_init_cont_pte().
 		 */
 		__set_pte_nosync(ptep, pfn_pte(__phys_to_pfn(phys), prot));
+		*mask |= PGTBL_PTE_MODIFIED;
 
 		/*
 		 * After the PTE entry has been populated once, we
@@ -251,7 +252,7 @@ static int alloc_init_cont_pte(pmd_t *pmdp, unsigned long addr,
 			       unsigned long end, phys_addr_t phys,
 			       pgprot_t prot,
 			       phys_addr_t (*pgtable_alloc)(enum pgtable_level),
-			       int flags)
+			       int flags, pgtbl_mod_mask *mask)
 {
 	unsigned long next;
 	pmd_t pmd = READ_ONCE(*pmdp);
@@ -272,6 +273,7 @@ static int alloc_init_cont_pte(pmd_t *pmdp, unsigned long addr,
 		init_clear_pgtable(ptep);
 		ptep += pte_index(addr);
 		__pmd_populate(pmdp, pte_phys, pmdval);
+		*mask |= PGTBL_PMD_MODIFIED;
 	} else {
 		BUG_ON(pmd_bad(pmd));
 		ptep = pte_set_fixmap_offset(pmdp, addr);
@@ -288,7 +290,7 @@ static int alloc_init_cont_pte(pmd_t *pmdp, unsigned long addr,
 		    !pte_range_has_valid_noncont(ptep))
 			__prot = __pgprot(pgprot_val(prot) | PTE_CONT);
 
-		init_pte(ptep, addr, next, phys, __prot);
+		init_pte(ptep, addr, next, phys, __prot, mask);
 
 		ptep += pte_index(next) - pte_index(addr);
 		phys += next - addr;
@@ -306,7 +308,8 @@ static int alloc_init_cont_pte(pmd_t *pmdp, unsigned long addr,
 
 static int init_pmd(pmd_t *pmdp, unsigned long addr, unsigned long end,
 		    phys_addr_t phys, pgprot_t prot,
-		    phys_addr_t (*pgtable_alloc)(enum pgtable_level), int flags)
+		    phys_addr_t (*pgtable_alloc)(enum pgtable_level), int flags,
+		    pgtbl_mod_mask *mask)
 {
 	unsigned long next;
 
@@ -319,7 +322,10 @@ static int init_pmd(pmd_t *pmdp, unsigned long addr, unsigned long end,
 		if (((addr | next | phys) & ~PMD_MASK) == 0 &&
 		    (flags & NO_BLOCK_MAPPINGS) == 0 &&
 		    !pmd_table(old_pmd)) {
-			WARN_ON(!pmd_set_huge(pmdp, phys, prot));
+			if (pmd_set_huge(pmdp, phys, prot))
+				*mask |= PGTBL_PMD_MODIFIED;
+			else
+				WARN_ON(1);
 
 			/*
 			 * After the PMD entry has been populated once, we
@@ -331,7 +337,7 @@ static int init_pmd(pmd_t *pmdp, unsigned long addr, unsigned long end,
 			int ret;
 
 			ret = alloc_init_cont_pte(pmdp, addr, next, phys, prot,
-						  pgtable_alloc, flags);
+						  pgtable_alloc, flags, mask);
 			if (ret)
 				return ret;
 
@@ -359,7 +365,7 @@ static int alloc_init_cont_pmd(pud_t *pudp, unsigned long addr,
 			       unsigned long end, phys_addr_t phys,
 			       pgprot_t prot,
 			       phys_addr_t (*pgtable_alloc)(enum pgtable_level),
-			       int flags)
+			       int flags, pgtbl_mod_mask *mask)
 {
 	int ret;
 	unsigned long next;
@@ -384,6 +390,7 @@ static int alloc_init_cont_pmd(pud_t *pudp, unsigned long addr,
 		init_clear_pgtable(pmdp);
 		pmdp += pmd_index(addr);
 		__pud_populate(pudp, pmd_phys, pudval);
+		*mask |= PGTBL_PUD_MODIFIED;
 	} else {
 		BUG_ON(pud_bad(pud));
 		pmdp = pmd_set_fixmap_offset(pudp, addr);
@@ -400,7 +407,8 @@ static int alloc_init_cont_pmd(pud_t *pudp, unsigned long addr,
 		    !pmd_range_has_valid_noncont(pmdp))
 			__prot = __pgprot(pgprot_val(prot) | PTE_CONT);
 
-		ret = init_pmd(pmdp, addr, next, phys, __prot, pgtable_alloc, flags);
+		ret = init_pmd(pmdp, addr, next, phys, __prot, pgtable_alloc, flags,
+			       mask);
 		if (ret)
 			goto out;
 
@@ -417,7 +425,7 @@ static int alloc_init_cont_pmd(pud_t *pudp, unsigned long addr,
 static int alloc_init_pud(p4d_t *p4dp, unsigned long addr, unsigned long end,
 			  phys_addr_t phys, pgprot_t prot,
 			  phys_addr_t (*pgtable_alloc)(enum pgtable_level),
-			  int flags)
+			  int flags, pgtbl_mod_mask *mask)
 {
 	int ret = 0;
 	unsigned long next;
@@ -438,6 +446,7 @@ static int alloc_init_pud(p4d_t *p4dp, unsigned long addr, unsigned long end,
 		init_clear_pgtable(pudp);
 		pudp += pud_index(addr);
 		__p4d_populate(p4dp, pud_phys, p4dval);
+		*mask |= PGTBL_P4D_MODIFIED;
 	} else {
 		BUG_ON(p4d_bad(p4d));
 		pudp = pud_set_fixmap_offset(p4dp, addr);
@@ -455,7 +464,10 @@ static int alloc_init_pud(p4d_t *p4dp, unsigned long addr, unsigned long end,
 		   ((addr | next | phys) & ~PUD_MASK) == 0 &&
 		    (flags & NO_BLOCK_MAPPINGS) == 0 &&
 		    !pud_table(old_pud)) {
-			WARN_ON(!pud_set_huge(pudp, phys, prot));
+			if (pud_set_huge(pudp, phys, prot))
+				*mask |= PGTBL_PUD_MODIFIED;
+			else
+				WARN_ON(1);
 
 			/*
 			 * After the PUD entry has been populated once, we
@@ -465,7 +477,7 @@ static int alloc_init_pud(p4d_t *p4dp, unsigned long addr, unsigned long end,
 						      READ_ONCE(pud_val(*pudp))));
 		} else {
 			ret = alloc_init_cont_pmd(pudp, addr, next, phys, prot,
-						  pgtable_alloc, flags);
+						  pgtable_alloc, flags, mask);
 			if (ret)
 				goto out;
 
@@ -484,7 +496,7 @@ static int alloc_init_pud(p4d_t *p4dp, unsigned long addr, unsigned long end,
 static int alloc_init_p4d(pgd_t *pgdp, unsigned long addr, unsigned long end,
 			  phys_addr_t phys, pgprot_t prot,
 			  phys_addr_t (*pgtable_alloc)(enum pgtable_level),
-			  int flags)
+			  int flags, pgtbl_mod_mask *mask)
 {
 	int ret;
 	unsigned long next;
@@ -505,6 +517,7 @@ static int alloc_init_p4d(pgd_t *pgdp, unsigned long addr, unsigned long end,
 		init_clear_pgtable(p4dp);
 		p4dp += p4d_index(addr);
 		__pgd_populate(pgdp, p4d_phys, pgdval);
+		*mask |= PGTBL_PGD_MODIFIED;
 	} else {
 		BUG_ON(pgd_bad(pgd));
 		p4dp = p4d_set_fixmap_offset(pgdp, addr);
@@ -516,7 +529,7 @@ static int alloc_init_p4d(pgd_t *pgdp, unsigned long addr, unsigned long end,
 		next = p4d_addr_end(addr, end);
 
 		ret = alloc_init_pud(p4dp, addr, next, phys, prot,
-				     pgtable_alloc, flags);
+				     pgtable_alloc, flags, mask);
 		if (ret)
 			goto out;
 
@@ -536,7 +549,7 @@ static int __create_pgd_mapping_locked(pgd_t *pgdir, phys_addr_t phys,
 				       unsigned long virt, phys_addr_t size,
 				       pgprot_t prot,
 				       phys_addr_t (*pgtable_alloc)(enum pgtable_level),
-				       int flags)
+				       int flags, pgtbl_mod_mask *mask)
 {
 	int ret;
 	unsigned long addr, end, next;
@@ -556,7 +569,7 @@ static int __create_pgd_mapping_locked(pgd_t *pgdir, phys_addr_t phys,
 	do {
 		next = pgd_addr_end(addr, end);
 		ret = alloc_init_p4d(pgdp, addr, next, phys, prot, pgtable_alloc,
-				     flags);
+				     flags, mask);
 		if (ret)
 			return ret;
 		phys += next - addr;
@@ -569,13 +582,13 @@ static int __create_pgd_mapping(pgd_t *pgdir, phys_addr_t phys,
 				unsigned long virt, phys_addr_t size,
 				pgprot_t prot,
 				phys_addr_t (*pgtable_alloc)(enum pgtable_level),
-				int flags)
+				int flags, pgtbl_mod_mask *mask)
 {
 	int ret;
 
 	mutex_lock(&fixmap_lock);
 	ret = __create_pgd_mapping_locked(pgdir, phys, virt, size, prot,
-					  pgtable_alloc, flags);
+					  pgtable_alloc, flags, mask);
 	mutex_unlock(&fixmap_lock);
 
 	return ret;
@@ -588,9 +601,10 @@ static void early_create_pgd_mapping(pgd_t *pgdir, phys_addr_t phys,
 				     int flags)
 {
 	int ret;
+	pgtbl_mod_mask mask = 0;
 
 	ret = __create_pgd_mapping(pgdir, phys, virt, size, prot, pgtable_alloc,
-				   flags);
+				   flags, &mask);
 	if (ret)
 		panic("Failed to create page tables\n");
 }
@@ -1332,6 +1346,7 @@ static int __init __kpti_install_ng_mappings(void *__unused)
 
 	if (!cpu) {
 		int ret;
+		pgtbl_mod_mask mask = 0;
 
 		alloc = __get_free_pages(GFP_ATOMIC | __GFP_ZERO, order);
 		kpti_ng_temp_pgd = (pgd_t *)(alloc + (levels - 1) * PAGE_SIZE);
@@ -1355,7 +1370,7 @@ static int __init __kpti_install_ng_mappings(void *__unused)
 		//
 		ret = __create_pgd_mapping_locked(kpti_ng_temp_pgd, __pa(alloc),
 						  KPTI_NG_TEMP_VA, PAGE_SIZE, PAGE_KERNEL,
-						  kpti_ng_pgd_alloc, 0);
+						  kpti_ng_pgd_alloc, 0, &mask);
 		if (ret)
 			panic("Failed to create page tables\n");
 	}
@@ -1661,7 +1676,7 @@ static void unmap_hotplug_range(unsigned long addr, unsigned long end,
 
 static void free_empty_pte_table(pmd_t *pmdp, unsigned long addr,
 				 unsigned long end, unsigned long floor,
-				 unsigned long ceiling)
+				 unsigned long ceiling, pgtbl_mod_mask *mask)
 {
 	pte_t *ptep, pte;
 	unsigned long i, start = addr;
@@ -1677,6 +1692,7 @@ static void free_empty_pte_table(pmd_t *pmdp, unsigned long addr,
 		WARN_ON(!pte_none(pte));
 	} while (addr += PAGE_SIZE, addr < end);
 
+	*mask |= PGTBL_PTE_MODIFIED;
 	if (!pgtable_range_aligned(start, end, floor, ceiling, PMD_MASK))
 		return;
 
@@ -1692,13 +1708,14 @@ static void free_empty_pte_table(pmd_t *pmdp, unsigned long addr,
 	}
 
 	pmd_clear(pmdp);
+	*mask |= PGTBL_PMD_MODIFIED;
 	__flush_tlb_kernel_pgtable(start);
 	free_hotplug_pgtable_page(virt_to_page(ptep));
 }
 
 static void free_empty_pmd_table(pud_t *pudp, unsigned long addr,
 				 unsigned long end, unsigned long floor,
-				 unsigned long ceiling)
+				 unsigned long ceiling, pgtbl_mod_mask *mask)
 {
 	pmd_t *pmdp, pmd;
 	unsigned long i, next, start = addr;
@@ -1711,7 +1728,7 @@ static void free_empty_pmd_table(pud_t *pudp, unsigned long addr,
 			continue;
 
 		WARN_ON(!pmd_present(pmd) || !pmd_table(pmd));
-		free_empty_pte_table(pmdp, addr, next, floor, ceiling);
+		free_empty_pte_table(pmdp, addr, next, floor, ceiling, mask);
 	} while (addr = next, addr < end);
 
 	if (CONFIG_PGTABLE_LEVELS <= 2)
@@ -1732,13 +1749,14 @@ static void free_empty_pmd_table(pud_t *pudp, unsigned long addr,
 	}
 
 	pud_clear(pudp);
+	*mask |= PGTBL_PUD_MODIFIED;
 	__flush_tlb_kernel_pgtable(start);
 	free_hotplug_pgtable_page(virt_to_page(pmdp));
 }
 
 static void free_empty_pud_table(p4d_t *p4dp, unsigned long addr,
 				 unsigned long end, unsigned long floor,
-				 unsigned long ceiling)
+				 unsigned long ceiling, pgtbl_mod_mask *mask)
 {
 	pud_t *pudp, pud;
 	unsigned long i, next, start = addr;
@@ -1751,7 +1769,7 @@ static void free_empty_pud_table(p4d_t *p4dp, unsigned long addr,
 			continue;
 
 		WARN_ON(!pud_present(pud) || !pud_table(pud));
-		free_empty_pmd_table(pudp, addr, next, floor, ceiling);
+		free_empty_pmd_table(pudp, addr, next, floor, ceiling, mask);
 	} while (addr = next, addr < end);
 
 	if (!pgtable_l4_enabled())
@@ -1772,13 +1790,14 @@ static void free_empty_pud_table(p4d_t *p4dp, unsigned long addr,
 	}
 
 	p4d_clear(p4dp);
+	*mask |= PGTBL_P4D_MODIFIED;
 	__flush_tlb_kernel_pgtable(start);
 	free_hotplug_pgtable_page(virt_to_page(pudp));
 }
 
 static void free_empty_p4d_table(pgd_t *pgdp, unsigned long addr,
 				 unsigned long end, unsigned long floor,
-				 unsigned long ceiling)
+				 unsigned long ceiling, pgtbl_mod_mask *mask)
 {
 	p4d_t *p4dp, p4d;
 	unsigned long i, next, start = addr;
@@ -1791,7 +1810,7 @@ static void free_empty_p4d_table(pgd_t *pgdp, unsigned long addr,
 			continue;
 
 		WARN_ON(!p4d_present(p4d));
-		free_empty_pud_table(p4dp, addr, next, floor, ceiling);
+		free_empty_pud_table(p4dp, addr, next, floor, ceiling, mask);
 	} while (addr = next, addr < end);
 
 	if (!pgtable_l5_enabled())
@@ -1812,6 +1831,7 @@ static void free_empty_p4d_table(pgd_t *pgdp, unsigned long addr,
 	}
 
 	pgd_clear(pgdp);
+	*mask |= PGTBL_PGD_MODIFIED;
 	__flush_tlb_kernel_pgtable(start);
 	free_hotplug_pgtable_page(virt_to_page(p4dp));
 }
@@ -1821,6 +1841,8 @@ static void free_empty_tables(unsigned long addr, unsigned long end,
 {
 	unsigned long next;
 	pgd_t *pgdp, pgd;
+	pgtbl_mod_mask mask = 0;
+	unsigned long start = addr;
 
 	do {
 		next = pgd_addr_end(addr, end);
@@ -1830,8 +1852,11 @@ static void free_empty_tables(unsigned long addr, unsigned long end,
 			continue;
 
 		WARN_ON(!pgd_present(pgd));
-		free_empty_p4d_table(pgdp, addr, next, floor, ceiling);
+		free_empty_p4d_table(pgdp, addr, next, floor, ceiling, &mask);
 	} while (addr = next, addr < end);
+
+	if (mask & ARCH_PAGE_TABLE_SYNC_MASK)
+		arch_sync_kernel_mappings(start, end);
 }
 #endif
 
@@ -2039,6 +2064,7 @@ int arch_add_memory(int nid, u64 start, u64 size,
 		    struct mhp_params *params)
 {
 	int ret, flags = NO_EXEC_MAPPINGS;
+	pgtbl_mod_mask mask = 0;
 
 	VM_BUG_ON(!mhp_range_allowed(start, size, true));
 
@@ -2047,10 +2073,13 @@ int arch_add_memory(int nid, u64 start, u64 size,
 
 	ret = __create_pgd_mapping(swapper_pg_dir, start, __phys_to_virt(start),
 				   size, params->pgprot, pgd_pgtable_alloc_init_mm,
-				   flags);
+				   flags, &mask);
 	if (ret)
 		goto err;
 
+	if (mask & ARCH_PAGE_TABLE_SYNC_MASK)
+		arch_sync_kernel_mappings(start, start + size);
+
 	memblock_clear_nomap(start, size);
 
 	ret = __add_pages(nid, start >> PAGE_SHIFT, size >> PAGE_SHIFT,
Signature · 1 line
2.47.0
author
#11 lore raw

PATCH v2 8/16 arm64: kasan: sync up kasan shadow area page table

In reply to Yang Shi

Allocating vmalloc and percpu needs to populate kasan shadow area page table, needs to sync up to percpu page table.

---
 arch/arm64/mm/kasan_init.c | 45 ++++++++++++++++++++++++--------------
 1 file changed, 29 insertions(+), 16 deletions(-)
arch/arm64/mm/kasan_init.c +29 -16
index 3fcad956fdf7..960020343760 100644
@@ -60,13 +60,14 @@ static phys_addr_t __init kasan_alloc_raw_page(int node)
 }
 
 static pte_t *__init kasan_pte_offset(pmd_t *pmdp, unsigned long addr, int node,
-				      bool early)
+				      bool early, pgtbl_mod_mask *mask)
 {
 	if (pmd_none(READ_ONCE(*pmdp))) {
 		phys_addr_t pte_phys = early ?
 				__pa_symbol(kasan_early_shadow_pte)
 					: kasan_alloc_zeroed_page(node);
 		__pmd_populate(pmdp, pte_phys, PMD_TYPE_TABLE);
+		*mask |= PGTBL_PMD_MODIFIED;
 	}
 
 	return early ? pte_offset_kimg(pmdp, addr)
@@ -74,49 +75,53 @@ static pte_t *__init kasan_pte_offset(pmd_t *pmdp, unsigned long addr, int node,
 }
 
 static pmd_t *__init kasan_pmd_offset(pud_t *pudp, unsigned long addr, int node,
-				      bool early)
+				      bool early, pgtbl_mod_mask *mask)
 {
 	if (pud_none(READ_ONCE(*pudp))) {
 		phys_addr_t pmd_phys = early ?
 				__pa_symbol(kasan_early_shadow_pmd)
 					: kasan_alloc_zeroed_page(node);
 		__pud_populate(pudp, pmd_phys, PUD_TYPE_TABLE);
+		*mask |= PGTBL_PUD_MODIFIED;
 	}
 
 	return early ? pmd_offset_kimg(pudp, addr) : pmd_offset(pudp, addr);
 }
 
 static pud_t *__init kasan_pud_offset(p4d_t *p4dp, unsigned long addr, int node,
-				      bool early)
+				      bool early, pgtbl_mod_mask *mask)
 {
 	if (p4d_none(READ_ONCE(*p4dp))) {
 		phys_addr_t pud_phys = early ?
 				__pa_symbol(kasan_early_shadow_pud)
 					: kasan_alloc_zeroed_page(node);
 		__p4d_populate(p4dp, pud_phys, P4D_TYPE_TABLE);
+		*mask |= PGTBL_P4D_MODIFIED;
 	}
 
 	return early ? pud_offset_kimg(p4dp, addr) : pud_offset(p4dp, addr);
 }
 
 static p4d_t *__init kasan_p4d_offset(pgd_t *pgdp, unsigned long addr, int node,
-				      bool early)
+				      bool early, pgtbl_mod_mask *mask)
 {
 	if (pgd_none(READ_ONCE(*pgdp))) {
 		phys_addr_t p4d_phys = early ?
 				__pa_symbol(kasan_early_shadow_p4d)
 					: kasan_alloc_zeroed_page(node);
 		__pgd_populate(pgdp, p4d_phys, PGD_TYPE_TABLE);
+		*mask |= PGTBL_PGD_MODIFIED;
 	}
 
 	return early ? p4d_offset_kimg(pgdp, addr) : p4d_offset(pgdp, addr);
 }
 
 static void __init kasan_pte_populate(pmd_t *pmdp, unsigned long addr,
-				      unsigned long end, int node, bool early)
+				      unsigned long end, int node, bool early,
+				      pgtbl_mod_mask *mask)
 {
 	unsigned long next;
-	pte_t *ptep = kasan_pte_offset(pmdp, addr, node, early);
+	pte_t *ptep = kasan_pte_offset(pmdp, addr, node, early, mask);
 
 	do {
 		phys_addr_t page_phys = early ?
@@ -130,38 +135,41 @@ static void __init kasan_pte_populate(pmd_t *pmdp, unsigned long addr,
 }
 
 static void __init kasan_pmd_populate(pud_t *pudp, unsigned long addr,
-				      unsigned long end, int node, bool early)
+				      unsigned long end, int node, bool early,
+				      pgtbl_mod_mask *mask)
 {
 	unsigned long next;
-	pmd_t *pmdp = kasan_pmd_offset(pudp, addr, node, early);
+	pmd_t *pmdp = kasan_pmd_offset(pudp, addr, node, early, mask);
 
 	do {
 		next = pmd_addr_end(addr, end);
-		kasan_pte_populate(pmdp, addr, next, node, early);
+		kasan_pte_populate(pmdp, addr, next, node, early, mask);
 	} while (pmdp++, addr = next, addr != end && pmd_none(READ_ONCE(*pmdp)));
 }
 
 static void __init kasan_pud_populate(p4d_t *p4dp, unsigned long addr,
-				      unsigned long end, int node, bool early)
+				      unsigned long end, int node, bool early,
+				      pgtbl_mod_mask *mask)
 {
 	unsigned long next;
-	pud_t *pudp = kasan_pud_offset(p4dp, addr, node, early);
+	pud_t *pudp = kasan_pud_offset(p4dp, addr, node, early, mask);
 
 	do {
 		next = pud_addr_end(addr, end);
-		kasan_pmd_populate(pudp, addr, next, node, early);
+		kasan_pmd_populate(pudp, addr, next, node, early, mask);
 	} while (pudp++, addr = next, addr != end && pud_none(READ_ONCE(*pudp)));
 }
 
 static void __init kasan_p4d_populate(pgd_t *pgdp, unsigned long addr,
-				      unsigned long end, int node, bool early)
+				      unsigned long end, int node, bool early,
+				      pgtbl_mod_mask *mask)
 {
 	unsigned long next;
-	p4d_t *p4dp = kasan_p4d_offset(pgdp, addr, node, early);
+	p4d_t *p4dp = kasan_p4d_offset(pgdp, addr, node, early, mask);
 
 	do {
 		next = p4d_addr_end(addr, end);
-		kasan_pud_populate(p4dp, addr, next, node, early);
+		kasan_pud_populate(p4dp, addr, next, node, early, mask);
 	} while (p4dp++, addr = next, addr != end && p4d_none(READ_ONCE(*p4dp)));
 }
 
@@ -170,12 +178,17 @@ static void __init kasan_pgd_populate(unsigned long addr, unsigned long end,
 {
 	unsigned long next;
 	pgd_t *pgdp;
+	pgtbl_mod_mask mask = 0;
+	unsigned long start = addr;
 
 	pgdp = pgd_offset_k(addr);
 	do {
 		next = pgd_addr_end(addr, end);
-		kasan_p4d_populate(pgdp, addr, next, node, early);
+		kasan_p4d_populate(pgdp, addr, next, node, early, &mask);
 	} while (pgdp++, addr = next, addr != end);
+
+	if (mask & ARCH_PAGE_TABLE_SYNC_MASK)
+		arch_sync_kernel_mappings(start, end);
 }
 
 #if defined(CONFIG_ARM64_64K_PAGES) || CONFIG_PGTABLE_LEVELS > 4
Signature · 1 line
2.47.0
author
#12 lore raw

PATCH v2 9/16 arm64: mm: define percpu virtual space area

In reply to Yang Shi

The percpu allocator returns offset from percpu base address. The percpu base address is determined by the first chunk which is typically in the low address of vmalloc space, however percpu varriables are typically allocated from the high address of vmalloc space. So the offset could be quite big. It may be the whole size of vmalloc space. To support local percpu mapping in order to optimize this_cpu_*() ops, the percpu allocator needs to allocate memory from local percpu area too in the following patch and the offset to local percpu base address must be same because the offset returned by percpu allocator must be used to access both global percpu and local percpu.

We can half vmalloc space to have either half dedicated to local percpu, but it wastes too much address space. So carve out dedicated global percpu and local percpu areas. Each area size is 2 * PGDIR_SIZE if page size is 4K. It is 1TB with 4K page size, should be big enough for percpu. It has different size for 16K and 64K page size respectively.

The percpu areas are aligned to top level page table entry size in order to just need to sync percpu page table at pgd level to minimize page table sync overhead.

The kernel virtual address space layout now looks like:

+-----------------+
|  Linear mapping |
+-----------------+
|  Modules        |
+-----------------+
|  Vmalloc        |
+-----------------+
|  Global Percpu  |
+-----------------+
|  Local Percpu   |
+-----------------+
|  Vmemap         |
+-----------------+
|  PCI I/O        |
+-----------------+
|  Fixed map      |
+-----------------+
---
 arch/arm64/include/asm/pgtable.h | 27 ++++++++++++++++++++++++---
 arch/arm64/mm/kasan_init.c       |  2 +-
 arch/arm64/mm/mmu.c              |  4 ++++
 arch/arm64/mm/ptdump.c           |  4 ++++
 4 files changed, 33 insertions(+), 4 deletions(-)
arch/arm64/include/asm/pgtable.h +24 -3
index 13e069b01600..83cd65610d7f 100644
@@ -18,14 +18,35 @@
  * VMALLOC range.
  *
  * VMALLOC_START: beginning of the kernel vmalloc space
- * VMALLOC_END: extends to the available space below vmemmap
+ * VMALLOC_END: extends to the space below global percpu area
  */
 #define VMALLOC_START		(MODULES_END)
+#define VMALLOC_END		(PERCPU_START - SZ_8M)
+
+/*
+ * PERCPU range
+ *
+ * PERCPU_START: beginning of global percpu area
+ * PERCPU_END: end of global percpu area
+ * LOCAL_PERCPU_START: beginning of local percpu area
+ * LOCAL_PERCPU_END: end of local percpu area, extend to the available
+ *                   space below vmemmap
+ */
+#if defined(CONFIG_ARM64_4K_PAGES)
+#define PERCPU_SIZE		(2 * PGDIR_SIZE)
+#elif defined(CONFIG_ARM64_16K_PAGES)
+#define PERCPU_SIZE            (16 * PUD_SIZE)
+#elif defined(CONFIG_ARM64_64K_PAGES)
+#define PERCPU_SIZE            (PGDIR_SIZE)
+#endif
+#define PERCPU_START		(PERCPU_END - PERCPU_SIZE)
+#define PERCPU_END		(LOCAL_PERCPU_START)
+#define LOCAL_PERCPU_START	(LOCAL_PERCPU_END - PERCPU_SIZE)
 #if VA_BITS == VA_BITS_MIN
-#define VMALLOC_END		(VMEMMAP_START - SZ_8M)
+#define LOCAL_PERCPU_END	(ALIGN_DOWN(VMEMMAP_START, PERCPU_SIZE))
 #else
 #define VMEMMAP_UNUSED_NPAGES	((_PAGE_OFFSET(vabits_actual) - PAGE_OFFSET) >> PAGE_SHIFT)
-#define VMALLOC_END		(VMEMMAP_START + VMEMMAP_UNUSED_NPAGES * sizeof(struct page) - SZ_8M)
+#define LOCAL_PERCPU_END	(ALIGN_DOWN((VMEMMAP_START + VMEMMAP_UNUSED_NPAGES * sizeof(struct page)), PERCPU_SIZE))
 #endif
 
 #define vmemmap			((struct page *)VMEMMAP_START - (memstart_addr >> PAGE_SHIFT))
arch/arm64/mm/kasan_init.c +1 -1
index 960020343760..a8e272bfefe1 100644
@@ -324,7 +324,7 @@ static void __init kasan_init_shadow(void)
 
 	mod_shadow_start = (u64)kasan_mem_to_shadow((void *)MODULES_VADDR);
 
-	vmalloc_shadow_end = (u64)kasan_mem_to_shadow((void *)VMALLOC_END);
+	vmalloc_shadow_end = (u64)kasan_mem_to_shadow((void *)LOCAL_PERCPU_END);
 
 	/*
 	 * We are going to perform proper setup of shadow memory.
arch/arm64/mm/mmu.c +4 -0
index e1290ccaa5be..f14ba512fd8f 100644
@@ -108,6 +108,10 @@ void arch_sync_kernel_mappings(unsigned long start, unsigned long end)
 	if (!percpu_pgd_setup_done)
 		return;
 
+	/* Don't sync local percpu area page table */
+	if (start >= LOCAL_PERCPU_START && end < LOCAL_PERCPU_END)
+		return;
+
 	addr = start;
 	do {
 		pgd = READ_ONCE(*pgdp);
arch/arm64/mm/ptdump.c +4 -0
index 1c20144700d7..dd293319823e 100644
@@ -389,6 +389,10 @@ static int __init ptdump_init(void)
 		{ MODULES_END,		"Modules end" },
 		{ VMALLOC_START,	"vmalloc() area" },
 		{ VMALLOC_END,		"vmalloc() end" },
+		{ PERCPU_START,		"Global percpu start" },
+		{ PERCPU_END,		"Global percpu end" },
+		{ LOCAL_PERCPU_START,	"Local percpu start" },
+		{ LOCAL_PERCPU_END,	"Local percpu end" },
 		{ vmemmap_start,	"vmemmap start" },
 		{ VMEMMAP_END,		"vmemmap end" },
 		{ PCI_IO_START,		"PCI I/O start" },
Signature · 1 line
2.47.0
author
#13 lore raw

PATCH v2 10/16 mm: percpu: prepare to use dedicated percpu area

In reply to Yang Shi

The percpu variables are allocated from vmalloc area by default. The architectures which support local percpu map will allocate percpu variables from dedicated percpu area, for example, ARM64.

Introduce a new kernel config, CONFIG_HAVE_LOCAL_PER_CPU_MAP. The architectures which support local percpu map will need to select it. If it is enabled, allocate percpu variables from the dedicated percpu area.

But still treat percpu area address as vmalloc address.

---
 include/linux/mm.h | 11 +++++++++++
 mm/Kconfig         |  3 +++
 mm/percpu.c        |  8 ++++++++
 mm/vmalloc.c       | 23 ++++++++++++++++++++---
 4 files changed, 42 insertions(+), 3 deletions(-)
include/linux/mm.h +11 -0
index 485df9c2dbdd..c8c3800fbd9f 100644
@@ -1858,6 +1858,17 @@ unsigned long vmalloc_to_pfn(const void *addr);
 #ifdef CONFIG_MMU
 extern bool is_vmalloc_addr(const void *x);
 extern int is_vmalloc_or_module_addr(const void *x);
+#ifdef CONFIG_HAVE_LOCAL_PER_CPU_MAP
+static inline bool is_percpu_addr(unsigned long addr)
+{
+	return addr >= PERCPU_START && addr < LOCAL_PERCPU_END;
+}
+#else
+static inline bool is_percpu_addr(unsigned long addr)
+{
+	return false;
+}
+#endif
 #else
 static inline bool is_vmalloc_addr(const void *x)
 {
mm/Kconfig +3 -0
index 13ea4b29ee7b..e1910c332638 100644
@@ -1061,6 +1061,9 @@ config NEED_PER_CPU_PAGE_FIRST_CHUNK
 config USE_PERCPU_NUMA_NODE_ID
 	bool
 
+config HAVE_LOCAL_PER_CPU_MAP
+	bool
+
 config HAVE_SETUP_PER_CPU_AREA
 	bool
 
mm/percpu.c +8 -0
index 45c30c6531c5..b27828602c32 100644
@@ -3243,9 +3243,17 @@ int __init pcpu_page_first_chunk(size_t reserved_size, pcpu_fc_cpu_to_node_fn_t
 	}
 
 	/* allocate vm area, map the pages and copy static data */
+#ifdef CONFIG_HAVE_LOCAL_PER_CPU_MAP
+	vm.flags = VM_ALLOC;
+	vm.addr = (void *)ALIGN(PERCPU_START, PAGE_SIZE);
+	vm.size = num_possible_cpus() * ai->unit_size;
+	vm_area_add_early(&vm);
+	kasan_populate_early_vm_area_shadow(vm.addr, vm.size);
+#else
 	vm.flags = VM_ALLOC;
 	vm.size = num_possible_cpus() * ai->unit_size;
 	vm_area_register_early(&vm, PAGE_SIZE);
+#endif
 
 	for (unit = 0; unit < num_possible_cpus(); unit++) {
 		unsigned long unit_addr =
mm/vmalloc.c +20 -3
index 17c26e9796b2..9ec401918736 100644
@@ -80,6 +80,9 @@ bool is_vmalloc_addr(const void *x)
 {
 	unsigned long addr = (unsigned long)kasan_reset_tag(x);
 
+	if (is_percpu_addr(addr))
+		return true;
+
 	return addr >= VMALLOC_START && addr < VMALLOC_END;
 }
 EXPORT_SYMBOL(is_vmalloc_addr);
@@ -4926,9 +4929,15 @@ pvm_find_va_enclose_addr(unsigned long addr)
 static unsigned long
 pvm_determine_end_from_reverse(struct vmap_area **va, unsigned long align)
 {
-	unsigned long vmalloc_end = VMALLOC_END & ~(align - 1);
+	unsigned long vmalloc_end;
 	unsigned long addr;
 
+#ifdef CONFIG_HAVE_LOCAL_PER_CPU_MAP
+	vmalloc_end = PERCPU_END & ~(align - 1);
+#else
+	vmalloc_end = VMALLOC_END & ~(align - 1);
+#endif
+
 	if (likely(*va)) {
 		list_for_each_entry_from_reverse((*va),
 				&free_vmap_area_list, list) {
@@ -4969,14 +4978,22 @@ struct vm_struct **pcpu_get_vm_areas(const unsigned long *offsets,
 				     const size_t *sizes, int nr_vms,
 				     size_t align)
 {
-	const unsigned long vmalloc_start = ALIGN(VMALLOC_START, align);
-	const unsigned long vmalloc_end = VMALLOC_END & ~(align - 1);
+	unsigned long vmalloc_start;
+	unsigned long vmalloc_end;
 	struct vmap_area **vas, *va;
 	struct vm_struct **vms;
 	int area, area2, last_area, term_area;
 	unsigned long base, start, size, end, last_end, orig_start, orig_end;
 	bool purged = false;
 
+#ifdef CONFIG_HAVE_LOCAL_PER_CPU_MAP
+	vmalloc_start = ALIGN(PERCPU_START, align);
+	vmalloc_end = PERCPU_END & ~(align - 1);
+#else
+	vmalloc_start = ALIGN(VMALLOC_START, align);
+	vmalloc_end = VMALLOC_END & ~(align - 1);
+#endif
+
 	/* verify parameters and allocate data structures */
 	BUG_ON(offset_in_page(align) || !is_power_of_2(align));
 	for (last_area = 0, area = 0; area < nr_vms; area++) {
Signature · 1 line
2.47.0
author
#14 lore raw

PATCH v2 11/16 arm64: mm: map local percpu first chunk

In reply to Yang Shi

Allocate local percpu area and map to percpu page table for the first chunk.

It doesn't work for PCPU_FC_EMBED because the percpu base adddress may be in linear mapping space in this case, it will result in returning huge offset for percpu allocator. So percpu local map just can work with PCPU_FC_PAGE which allocates percpu variables from vmalloc area or the dedicated percpu area. So unselect NEED_PER_CPU_EMBED_FIRST_CHUNK if the architectures support percpu local map.

---
 arch/arm64/Kconfig           |  1 -
 arch/arm64/include/asm/mmu.h |  4 ++++
 arch/arm64/kernel/smp.c      | 25 ++-----------------------
 arch/arm64/mm/mmu.c          | 23 +++++++++++++++++++++++
 mm/percpu-internal.h         | 11 +++++++++++
 mm/percpu.c                  | 16 ++++++++++++++++
 6 files changed, 56 insertions(+), 24 deletions(-)
arch/arm64/Kconfig +0 -1
index 23cad9fc7046..df55e6bdf81f 100644
@@ -242,7 +242,6 @@ config ARM64
 	select LOCK_MM_AND_FIND_VMA
 	select MODULES_USE_ELF_RELA
 	select NEED_DMA_MAP_STATE
-	select NEED_PER_CPU_EMBED_FIRST_CHUNK
 	select NEED_PER_CPU_PAGE_FIRST_CHUNK
 	select NEED_SG_DMA_LENGTH
 	select OF
arch/arm64/include/asm/mmu.h +4 -0
index 8ed3b5f3cf84..aa0818f7ec58 100644
@@ -73,6 +73,10 @@ extern void *fixmap_remap_fdt(phys_addr_t dt_phys, int *size, pgprot_t prot);
 extern void mark_linear_text_alias_ro(void);
 extern int split_kernel_leaf_mapping(unsigned long start, unsigned long end);
 extern void linear_map_maybe_split_to_ptes(void);
+#ifdef CONFIG_HAVE_LOCAL_PER_CPU_MAP
+extern void map_local_percpu_first_chunk(unsigned int cpu, unsigned long virt,
+				struct page **pages, unsigned int nr);
+#endif
 
 /*
  * This check is triggered during the early boot before the cpufeature
arch/arm64/kernel/smp.c +2 -23
index 49524a7176b2..4c112bab8356 100644
@@ -840,36 +840,15 @@ int early_cpu_to_node(int cpu)
 unsigned long __per_cpu_offset[NR_CPUS] __read_mostly;
 EXPORT_SYMBOL(__per_cpu_offset);
 
-static int __init pcpu_cpu_distance(unsigned int from, unsigned int to)
-{
-	return node_distance(early_cpu_to_node(from), early_cpu_to_node(to));
-}
-
 void __init setup_per_cpu_areas(void)
 {
 	unsigned long delta;
 	unsigned int cpu;
 	int rc = -EINVAL;
 
-	if (pcpu_chosen_fc != PCPU_FC_PAGE) {
-		/*
-		 * Always reserve area for module percpu variables.  That's
-		 * what the legacy allocator did.
-		 */
-		rc = pcpu_embed_first_chunk(PERCPU_MODULE_RESERVE,
-					    PERCPU_DYNAMIC_RESERVE, PAGE_SIZE,
-					    pcpu_cpu_distance,
-					    early_cpu_to_node);
-#ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK
-		if (rc < 0)
-			pr_warn("PERCPU: %s allocator failed (%d), falling back to page size\n",
-				   pcpu_fc_names[pcpu_chosen_fc], rc);
-#endif
-	}
-
 #ifdef CONFIG_NEED_PER_CPU_PAGE_FIRST_CHUNK
-	if (rc < 0)
-		rc = pcpu_page_first_chunk(PERCPU_MODULE_RESERVE, early_cpu_to_node);
+	/* PCPU page table just can support PCPU_FC_PAGE */
+	rc = pcpu_page_first_chunk(PERCPU_MODULE_RESERVE, early_cpu_to_node);
 #endif
 	if (rc < 0)
 		panic("Failed to initialize percpu areas (err=%d).", rc);
arch/arm64/mm/mmu.c +23 -0
index f14ba512fd8f..ee4ab031ebf8 100644
@@ -1080,6 +1080,29 @@ void __init linear_map_maybe_split_to_ptes(void)
 	}
 }
 
+#ifdef CONFIG_HAVE_LOCAL_PER_CPU_MAP
+void __init map_local_percpu_first_chunk(unsigned int cpu, unsigned long virt,
+			    struct page **pages, unsigned int nr)
+{
+	int i, ret;
+	pgtbl_mod_mask mask = 0;
+
+	arch_enter_lazy_mmu_mode();
+
+	for (i = 0; i < nr; i++) {
+		phys_addr_t phys = page_to_phys(pages[i]);
+		ret = __create_pgd_mapping(percpu_pgd[cpu], phys, virt, PAGE_SIZE, PAGE_KERNEL,
+					   early_pgtable_alloc, NO_EXEC_MAPPINGS, &mask);
+		if (ret)
+			panic("Failed to create local percpu first chunk\n");
+
+		virt += PAGE_SIZE;
+	}
+
+	arch_leave_lazy_mmu_mode();
+}
+#endif
+
 /*
  * This function can only be used to modify existing table entries,
  * without allocating new levels of table. Note that this permits the
mm/percpu-internal.h +11 -0
index 8cbe039bf847..6d0542fb9e49 100644
@@ -162,6 +162,17 @@ static inline size_t pcpu_obj_full_size(size_t size)
 	return size * num_possible_cpus() + extra_size;
 }
 
+#ifdef CONFIG_HAVE_LOCAL_PER_CPU_MAP
+extern void __init map_local_percpu_first_chunk(unsigned int cpu, unsigned long virt,
+                            struct page **pages, unsigned int nr);
+#else
+static inline void __init map_local_percpu_first_chunk(unsigned int cpu, unsigned long virt,
+                            struct page **pages, unsigned int nr)
+{
+	return;
+}
+#endif
+
 #ifdef CONFIG_PERCPU_STATS
 
 #include <linux/spinlock.h>
mm/percpu.c +16 -0
index b27828602c32..df2daa57309b 100644
@@ -3194,6 +3194,7 @@ void __init __weak pcpu_populate_pte(unsigned long addr)
 int __init pcpu_page_first_chunk(size_t reserved_size, pcpu_fc_cpu_to_node_fn_t cpu_to_nd_fn)
 {
 	static struct vm_struct vm;
+	static struct vm_struct pcpu_vm;
 	struct pcpu_alloc_info *ai;
 	char psize_str[16];
 	int unit_pages;
@@ -3249,6 +3250,12 @@ int __init pcpu_page_first_chunk(size_t reserved_size, pcpu_fc_cpu_to_node_fn_t
 	vm.size = num_possible_cpus() * ai->unit_size;
 	vm_area_add_early(&vm);
 	kasan_populate_early_vm_area_shadow(vm.addr, vm.size);
+
+	pcpu_vm.flags = VM_ALLOC;
+	pcpu_vm.addr = (void *)ALIGN(LOCAL_PERCPU_START, PAGE_SIZE);
+	pcpu_vm.size = ai->unit_size;
+	vm_area_add_early(&pcpu_vm);
+	kasan_populate_early_vm_area_shadow(pcpu_vm.addr, pcpu_vm.size);
 #else
 	vm.flags = VM_ALLOC;
 	vm.size = num_possible_cpus() * ai->unit_size;
@@ -3256,6 +3263,7 @@ int __init pcpu_page_first_chunk(size_t reserved_size, pcpu_fc_cpu_to_node_fn_t
 #endif
 
 	for (unit = 0; unit < num_possible_cpus(); unit++) {
+		unsigned int cpu = ai->groups[0].cpu_map[unit];
 		unsigned long unit_addr =
 			(unsigned long)vm.addr + unit * ai->unit_size;
 
@@ -3272,6 +3280,14 @@ int __init pcpu_page_first_chunk(size_t reserved_size, pcpu_fc_cpu_to_node_fn_t
 
 		/* copy static data */
 		memcpy((void *)unit_addr, __per_cpu_start, ai->static_size);
+
+		/*
+		 * Map percpu data to PERCPU map.
+		 *
+		 * PCPU_FC_EMBED can't support it.
+		 */
+		map_local_percpu_first_chunk(cpu, (unsigned long)pcpu_vm.addr,
+				&pages[unit * unit_pages], unit_pages);
 	}
 
 	/* we're ready, commit */
Signature · 1 line
2.47.0
author
#15 lore raw

PATCH v2 12/16 mm: percpu: set up first chunk and reserve chunk

In reply to Yang Shi

Set up the first chunk and reserve chunk with local percpu map.

---
 include/linux/percpu.h |  2 +-
 mm/percpu-internal.h   |  2 ++
 mm/percpu.c            | 27 ++++++++++++++++++---------
 3 files changed, 21 insertions(+), 10 deletions(-)
include/linux/percpu.h +1 -1
index 4f36b8585995..2f0e190aca29 100644
@@ -114,7 +114,7 @@ extern struct pcpu_alloc_info * __init pcpu_alloc_alloc_info(int nr_groups,
 extern void __init pcpu_free_alloc_info(struct pcpu_alloc_info *ai);
 
 extern void __init pcpu_setup_first_chunk(const struct pcpu_alloc_info *ai,
-					 void *base_addr);
+					 void *base_addr, void *local_base);
 
 extern int __init pcpu_embed_first_chunk(size_t reserved_size, size_t dyn_size,
 				size_t atom_size,
mm/percpu-internal.h +2 -0
index 6d0542fb9e49..ec2e55a2f836 100644
@@ -63,6 +63,8 @@ struct pcpu_chunk {
 	 * chunk_md.
 	 */
 	void			*base_addr ____cacheline_aligned_in_smp;
+	/* percpu local base address of the chunk */
+	void                    *local_base;
 
 	unsigned long		*alloc_map;	/* allocation map */
 	struct pcpu_block_md	*md_blocks;	/* metadata blocks */
mm/percpu.c +18 -9
index df2daa57309b..15303611ecaf 100644
@@ -1338,6 +1338,7 @@ static void pcpu_init_md_blocks(struct pcpu_chunk *chunk)
 /**
  * pcpu_alloc_first_chunk - creates chunks that serve the first chunk
  * @tmp_addr: the start of the region served
+ * @local_tmp: the start of the local percpu region served
  * @map_size: size of the region served
  *
  * This is responsible for creating the chunks that serve the first chunk.  The
@@ -1349,15 +1350,16 @@ static void pcpu_init_md_blocks(struct pcpu_chunk *chunk)
  * Chunk serving the region at @tmp_addr of @map_size.
  */
 static struct pcpu_chunk * __init pcpu_alloc_first_chunk(unsigned long tmp_addr,
-							 int map_size)
+							 unsigned long local_tmp, int map_size)
 {
 	struct pcpu_chunk *chunk;
-	unsigned long aligned_addr;
+	unsigned long aligned_addr, aligned_local;
 	int start_offset, offset_bits, region_size, region_bits;
 	size_t alloc_size;
 
 	/* region calculations */
 	aligned_addr = tmp_addr & PAGE_MASK;
+	aligned_local = local_tmp & PAGE_MASK;
 
 	start_offset = tmp_addr - aligned_addr;
 	region_size = ALIGN(start_offset + map_size, PAGE_SIZE);
@@ -1370,6 +1372,7 @@ static struct pcpu_chunk * __init pcpu_alloc_first_chunk(unsigned long tmp_addr,
 	INIT_LIST_HEAD(&chunk->list);
 
 	chunk->base_addr = (void *)aligned_addr;
+	chunk->local_base = (void *)aligned_local;
 	chunk->start_offset = start_offset;
 	chunk->end_offset = region_size - chunk->start_offset - map_size;
 
@@ -2562,7 +2565,7 @@ static void pcpu_dump_alloc_info(const char *lvl,
  * and available for dynamic allocation like any other chunk.
  */
 void __init pcpu_setup_first_chunk(const struct pcpu_alloc_info *ai,
-				   void *base_addr)
+				   void *base_addr, void *local_base)
 {
 	size_t size_sum = ai->static_size + ai->reserved_size + ai->dyn_size;
 	size_t static_size, dyn_size;
@@ -2572,7 +2575,7 @@ void __init pcpu_setup_first_chunk(const struct pcpu_alloc_info *ai,
 	unsigned int cpu;
 	int *unit_map;
 	int group, unit, i;
-	unsigned long tmp_addr;
+	unsigned long tmp_addr, local_tmp = 0;
 	size_t alloc_size;
 
 #define PCPU_SETUP_BUG_ON(cond)	do {					\
@@ -2713,11 +2716,15 @@ void __init pcpu_setup_first_chunk(const struct pcpu_alloc_info *ai,
 	 *   chunk.
 	 */
 	tmp_addr = (unsigned long)base_addr + static_size;
+	if (local_base)
+		local_tmp = (unsigned long)local_base + static_size;
 	if (ai->reserved_size)
-		pcpu_reserved_chunk = pcpu_alloc_first_chunk(tmp_addr,
+		pcpu_reserved_chunk = pcpu_alloc_first_chunk(tmp_addr, local_tmp,
 						ai->reserved_size);
 	tmp_addr = (unsigned long)base_addr + static_size + ai->reserved_size;
-	pcpu_first_chunk = pcpu_alloc_first_chunk(tmp_addr, dyn_size);
+	if (local_base)
+		local_tmp = (unsigned long)local_base + static_size + ai->reserved_size;
+	pcpu_first_chunk = pcpu_alloc_first_chunk(tmp_addr, local_tmp, dyn_size);
 
 	pcpu_nr_empty_pop_pages = pcpu_first_chunk->nr_empty_pop_pages;
 	pcpu_chunk_relocate(pcpu_first_chunk, -1);
@@ -3108,7 +3115,7 @@ int __init pcpu_embed_first_chunk(size_t reserved_size, size_t dyn_size,
 		PFN_DOWN(size_sum), ai->static_size, ai->reserved_size,
 		ai->dyn_size, ai->unit_size);
 
-	pcpu_setup_first_chunk(ai, base);
+	pcpu_setup_first_chunk(ai, base, NULL);
 	goto out_free;
 
 out_free_areas:
@@ -3260,6 +3267,8 @@ int __init pcpu_page_first_chunk(size_t reserved_size, pcpu_fc_cpu_to_node_fn_t
 	vm.flags = VM_ALLOC;
 	vm.size = num_possible_cpus() * ai->unit_size;
 	vm_area_register_early(&vm, PAGE_SIZE);
+
+	pcpu_vm.addr = NULL;
 #endif
 
 	for (unit = 0; unit < num_possible_cpus(); unit++) {
@@ -3295,7 +3304,7 @@ int __init pcpu_page_first_chunk(size_t reserved_size, pcpu_fc_cpu_to_node_fn_t
 		unit_pages, psize_str, ai->static_size,
 		ai->reserved_size, ai->dyn_size);
 
-	pcpu_setup_first_chunk(ai, vm.addr);
+	pcpu_setup_first_chunk(ai, vm.addr, pcpu_vm.addr);
 	goto out_free_ar;
 
 enomem:
@@ -3377,7 +3386,7 @@ void __init setup_per_cpu_areas(void)
 	ai->groups[0].nr_units = 1;
 	ai->groups[0].cpu_map[0] = 0;
 
-	pcpu_setup_first_chunk(ai, fc);
+	pcpu_setup_first_chunk(ai, fc, NULL);
 	pcpu_free_alloc_info(ai);
 }
 
Signature · 1 line
2.47.0
author
#16 lore raw

PATCH v2 13/16 arm64: mm: introduce __per_cpu_local_off

In reply to Yang Shi

this_cpu_*() ops will use it to get the local percpu address. It has the same value for all CPUs.

Also introduce pcpu_local_base, which is the base address of local percpu map.

---
 arch/arm64/kernel/smp.c | 4 ++++
 include/linux/percpu.h  | 2 ++
 mm/percpu.c             | 3 +++
 3 files changed, 9 insertions(+)
arch/arm64/kernel/smp.c +4 -0
index 4c112bab8356..4a811379c191 100644
@@ -839,6 +839,9 @@ int early_cpu_to_node(int cpu)
 #ifdef CONFIG_HAVE_SETUP_PER_CPU_AREA
 unsigned long __per_cpu_offset[NR_CPUS] __read_mostly;
 EXPORT_SYMBOL(__per_cpu_offset);
+/* Used to calculate pcpu local address, the offset is same for all CPUs */
+unsigned long __per_cpu_local_off __ro_after_init;
+EXPORT_SYMBOL(__per_cpu_local_off);
 
 void __init setup_per_cpu_areas(void)
 {
@@ -856,6 +859,7 @@ void __init setup_per_cpu_areas(void)
 	delta = (unsigned long)pcpu_base_addr - (unsigned long)__per_cpu_start;
 	for_each_possible_cpu(cpu)
 		__per_cpu_offset[cpu] = delta + pcpu_unit_offsets[cpu];
+	__per_cpu_local_off = (unsigned long)pcpu_local_base - (unsigned long)__per_cpu_start;
 }
 #endif
 
include/linux/percpu.h +2 -0
index 2f0e190aca29..5a27031dff1a 100644
@@ -75,6 +75,8 @@
 
 extern void *pcpu_base_addr;
 extern const unsigned long *pcpu_unit_offsets;
+/* percpu local mapping base */
+extern void *pcpu_local_base;
 
 struct pcpu_group_info {
 	int			nr_units;	/* aligned # of units */
mm/percpu.c +3 -0
index 15303611ecaf..8b273ed8e5a8 100644
@@ -145,6 +145,8 @@ static unsigned int pcpu_high_unit_cpu __ro_after_init;
 
 /* the address of the first chunk which starts with the kernel static area */
 void *pcpu_base_addr __ro_after_init;
+/* The address of the first chunk local mapping */
+void *pcpu_local_base __ro_after_init;
 
 static const int *pcpu_unit_map __ro_after_init;		/* cpu -> unit */
 const unsigned long *pcpu_unit_offsets __ro_after_init;	/* cpu -> unit offset */
@@ -3305,6 +3307,7 @@ int __init pcpu_page_first_chunk(size_t reserved_size, pcpu_fc_cpu_to_node_fn_t
 		ai->reserved_size, ai->dyn_size);
 
 	pcpu_setup_first_chunk(ai, vm.addr, pcpu_vm.addr);
+	pcpu_local_base = pcpu_vm.addr;
 	goto out_free_ar;
 
 enomem:
Signature · 1 line
2.47.0
author
#17 lore raw

PATCH v2 14/16 mm: percpu: allocate and free local percpu vm area

In reply to Yang Shi

Allocate local percpu vm area. The delta between the allocated addr (chunk local base) and pcpu_local_base must be same with the delta between chunk base and pcpu_base_addr. Each CPU's local percpu area will be mapped to its own page table. This section of page table is not shared between CPUs.

And free local percpu vm area. Also unmap from percpu page table.

---
 include/linux/vmalloc.h |  3 ++
 mm/percpu-internal.h    |  1 +
 mm/percpu-vm.c          | 94 +++++++++++++++++++++++++++++++++++++++++
 mm/vmalloc.c            | 92 +++++++++++++++++++++++++++++++++++++---
 4 files changed, 185 insertions(+), 5 deletions(-)
include/linux/vmalloc.h +3 -0
index d87dc7f77f4e..daf9965a111d 100644
@@ -313,6 +313,9 @@ struct vm_struct **pcpu_get_vm_areas(const unsigned long *offsets,
 				     size_t align);
 
 void pcpu_free_vm_areas(struct vm_struct **vms, int nr_vms);
+struct vm_struct *pcpu_get_local_vm_area(unsigned long hint,
+				     int unit_size, size_t align);
+
 # else
 static inline struct vm_struct **
 pcpu_get_vm_areas(const unsigned long *offsets,
mm/percpu-internal.h +1 -0
index ec2e55a2f836..4d461305243b 100644
@@ -70,6 +70,7 @@ struct pcpu_chunk {
 	struct pcpu_block_md	*md_blocks;	/* metadata blocks */
 
 	void			*data;		/* chunk data */
+	void			*local_data;	/* chunk local vm */
 	bool			immutable;	/* no [de]population allowed */
 	bool			isolated;	/* isolated from active chunk
 						   slots */
mm/percpu-vm.c +94 -0

160 lines not on this page · raw message

mm/vmalloc.c +87 -5

131 lines not on this page · raw message

Signature · 1 line
2.47.0
author
#18 lore raw

PATCH v2 15/16 arm64: kconfig: select HAVE_LOCAL_PER_CPU_MAP

In reply to Yang Shi

ARM64 supports local percpu map, so select HAVE_LOCAL_PER_CPU_MAP. But it depends on VA bits and page size for now.

---
 arch/arm64/Kconfig | 5 +++++
 1 file changed, 5 insertions(+)
arch/arm64/Kconfig +5 -0
index df55e6bdf81f..95ffd151187f 100644
@@ -415,6 +415,11 @@ config KASAN_SHADOW_OFFSET
 config UNWIND_TABLES
 	bool
 
+config HAVE_LOCAL_PER_CPU_MAP
+	def_bool y
+	depends on (ARM64_VA_BITS_48 && (ARM64_4K_PAGES || ARM64_64K_PAGES)) || \
+		   (ARM64_VA_BITS_47 && ARM64_16K_PAGES)
+
 source "arch/arm64/Kconfig.platforms"
 
 menu "Kernel Features"
Signature · 1 line
2.47.0
author
#19 lore raw

PATCH v2 16/16 arm64: percpu: use local percpu for this_cpu_*() APIs

In reply to Yang Shi

Use local percpu address for this_cpu_*() APIs. Because the percpu variable is mapped to the same virtual address, their address can be calculated by using __per_cpu_local_off which has same value for all CPUs. So preempt_disable/preempt_enable is not needed anymore. This optimization can improve the performance for this_cpu_*() operations.

Kernel build test on AmpereOne (160 cores) with default Fedora kernel config in a memcg roughly showed 13% - 15% sys time improvement.

---
 arch/arm64/include/asm/percpu.h | 37 ++++++++++++++++++++++++++++++++-
 1 file changed, 36 insertions(+), 1 deletion(-)
arch/arm64/include/asm/percpu.h +36 -1
index b57b2bb00967..03d149f9177c 100644
@@ -142,6 +142,27 @@ PERCPU_RET_OP(add, add, ldadd)
 #undef PERCPU_OP
 #undef PERCPU_RET_OP
 
+#ifdef CONFIG_HAVE_LOCAL_PER_CPU_MAP
+extern unsigned long __per_cpu_local_off;
+
+#define local_cpu_ptr(ptr)						\
+({									\
+	__verify_pcpu_ptr(ptr);						\
+	SHIFT_PERCPU_PTR(ptr, __per_cpu_local_off);			\
+})
+
+#define _pcp_protect(op, pcp, ...)					\
+({									\
+	op(local_cpu_ptr(&(pcp)), __VA_ARGS__);				\
+})
+
+#define _pcp_protect_return(op, pcp, args...)				\
+({									\
+	typeof(pcp) __retval;						\
+	__retval = (typeof(pcp))op(local_cpu_ptr(&(pcp)), ##args);	\
+	__retval;							\
+})
+#else
 /*
  * It would be nice to avoid the conditional call into the scheduler when
  * re-enabling preemption for preemptible kernels, but doing that in a way
@@ -152,7 +173,6 @@ PERCPU_RET_OP(add, add, ldadd)
  * preemption point when TIF_NEED_RESCHED gets set while preemption is
  * disabled.
  */
-
 #define _pcp_protect(op, pcp, ...)					\
 ({									\
 	preempt_disable_notrace();					\
@@ -168,6 +188,7 @@ PERCPU_RET_OP(add, add, ldadd)
 	preempt_enable_notrace();					\
 	__retval;							\
 })
+#endif
 
 #define this_cpu_read_1(pcp)		\
 	_pcp_protect_return(__percpu_read_8, pcp)
@@ -243,6 +264,19 @@ PERCPU_RET_OP(add, add, ldadd)
 
 #define this_cpu_cmpxchg64(pcp, o, n)	this_cpu_cmpxchg_8(pcp, o, n)
 
+#ifdef CONFIG_HAVE_LOCAL_PER_CPU_MAP
+#define this_cpu_cmpxchg128(pcp, o, n)					\
+({									\
+	typedef typeof(pcp) pcp_op_T__;					\
+	u128 old__, new__, ret__;					\
+	pcp_op_T__ *ptr__;						\
+	old__ = o;							\
+	new__ = n;							\
+	ptr__ = local_cpu_ptr(&(pcp));					\
+	ret__ = cmpxchg128_local((void *)ptr__, old__, new__);		\
+	ret__;								\
+})
+#else
 #define this_cpu_cmpxchg128(pcp, o, n)					\
 ({									\
 	typedef typeof(pcp) pcp_op_T__;					\
@@ -256,6 +290,7 @@ PERCPU_RET_OP(add, add, ldadd)
 	preempt_enable_notrace();					\
 	ret__;								\
 })
+#endif
 
 #ifdef __KVM_NVHE_HYPERVISOR__
 extern unsigned long __hyp_per_cpu_offset(unsigned int cpu);
Signature · 1 line
2.47.0
#20 lore raw
In reply to Yang Shi
Quoted 10 lines · Yang Shi

Hi,

This is v2 RFC. In v2 a lot problems found out by Sashiko were fixed and more feature gaps were closed (please see the below changelog for the details). Although there are still some open issues, for example, it just can support 48 bits VA (for 4K and 64K) and 47 bits VA (for 16K), KPTI support has not been solved yet, etc, but I think the delta should be big enough and worth a new RFC to gather comments in order to make sure I'm on the right track.

Quoted 4 lines

Some more benchmarks were done, for example, some latency related benchmarks that I mentioned at LSFMM because I thought responsiveness should be improved due to the removal of preempt_disable. Collected more PMU counters as well. Please refer to the benchmark section for more details.

Hi Yang,

When we spoke about this at LSFMM, I said I would take the series for a spin on other hardware to see if the performance gains generalise. I did make a start on that but the kernel wouldn't even boot and I never found time to debug. Sorry that I didn't let you know sooner.

But I think that there are other issues with this concept that likely block it from being accepted upstream; see below...

Quoted 38 lines

Look forward to comments.

Changelog
v2: * Added 3-level and 2-level page table support.
* Tested with 16K and 64K page size. But we just support 48 bits VA (4K and 64K) and 47 bits VA with 16K for now. Please refer to the below "known issue" section for the detail reason.
* Added support for memory hotplug.
* Added support for KASAN (generic).
* Treated percpu and local percpu area address as vmalloc address.
* Fixed build failure for x86.
* Fixed build failure for !CONFIG_NUMA.
* Added KASAN support for local percpu area.
* Some other misc bug fixes found out by Sashiko.
* More code refactor and cleanup.
* Regorganized the patches.
* More benchmarks, refer to benchmark section for more details.
* Rebased to v7.2-rc1.

Introduction
============
This patch series implemented the LSFMM 2026 proposal for optimizing this_cpu_*() ops on ARM64. For the details of the proposal, Please refer to:
https://lore.kernel.org/linux-mm/CAHbLzkpcN-T8MH6=W3jCxcFj1gVZp8fRqe231yzZT-rV_E_org@mail.gmail.com/
I didn't repeat it in the cover letter because there is no change to the proposal.

The series is based on 7.1-rc1. It is basically minimum viable patches. There are still a few hacks in this series and it may break something, for example, KPTI, SMT machines which shared TLB, etc. But it shoule be good enough for now to demonstrate the core idea. The main purpose of the RFC is to gather feedback, figure out missing parts and risks, and make sure we are on the right track, as well as hopefully it can help the discussion for the upcoming LSFMM.

Quoted 7 lines

I broke the patches down to arch-dependent and arch-independent parts so that hopefully the interested persons can do experiments on other architectures, for example, S390, easier.

A new kernel config is introduced, HAVE_LOCAL_PER_CPU_MAP. The architectures which can support this feature will select it. Allocating and freeing percpu local mapping is protected by this config so that others won't pay the cost.

I think this approach is a bit of a sledge hammer to crack a nut. If the problem is the cost this_cpu_* ops on arm64 due to the need to disable/enable preemption, I think we can solve that much more simply with [1], which introduces in-kernel restartable sequences. Using this, the operation no longer needs to disable preemption but can instead just detect when it is preempted and retry.

[1]
https://lore.kernel.org/all/[email protected]/

We have observed a few performance regressions recently, for which the root cause is increased use of this_cpu_*. We have somebody at Arm about to start an investigation into whether in-kernel rseq can solve the problem.

Quoted 24 lines

Known Issues
============
1. KPTI
-------
We need determine what CPU we are on, then switch to the right page table. Currently arm64 kernel fetches tramp_pg_dir via swapper_pg_dir - fixed_offset, and fetches swapper_pg_dir from ttbr1. But ttbr1 may not hold swapper_pg_dir anymore except CPU #0. So we need to figure out the other way to handle it. Switching to tramp_pg_dir should be easy, but the reverse seems harder because tramp_pg_dir just maps the trampoline vectors.
Maybe we can do two steps switch. Switch to swapper_pg_dir at the first step, then switch to per cpu page table (for entry) or tramp page table (for exit). Nobody should call this_cpu_*() at either userspace -> kernel entry stage or kernel -> userspace exit stage.

2. SW PAN
---------
Has the similar issue as KPTI. It installs reserved_pg_dir to TTBR0 when running in kernel space, but fetching reserved_pg_dir via swapper_pg_dir - fixed_offset. Maybe we can save the physical address of swapper_pg_dir in a variable, then load it from that variable instead of ttbr1.

Quoted 7 lines

3. Shared TLB machines
----------------------
Some machines may share TLB between CPUs, for example, SMT machines may share TLB between the two hardware threads in one core.
The per cpu page table just can't work with it. Maybe we need a new cpufeature to indicate whether per cpu page table is allowed or not. Then just enable it for not-shared-TLB machines.

The architectural feature you're referring to here is FEAT_TTCNP (common not private). There are many CPUs out there that use this feature (not just those that support SMT).

FEAT_TTCNP is mandatory from Armv8.2 and the presence of the feature only tells you whether you can legally set the CnP bit - it doesn't actually tell you if the HW does any sharing - there is no way to detect this. So making per-cpu pgtables mutually exclusive with CNP is a non-starter as all CPUs from v8.2 onwards advertise CNP.

Even ignoring the CNP problem, you'll end up installing the per-cpu TLB entries tagged with whatever user ASID happens to be installed at the time. Which means you could end up with lots of duplicated TLB entries all differing only by ASID. It's probably not the end of the world, but it seems... inefficient.

The only way I can see to make this work fully is to rely on FEAT_ASID2, which allows specifying separate ASIDs for TTBR0 and TTBR1. Then each CPU can have it's own (permanently installed) ASID for local mappings in TTBR1. That would solve the CNP issue. FEAT_ASID2 is optional from v9.4 and mandatory from v9.5.

Quoted 11 lines

4. Don't support all VA bits
----------------------------
We just support 48 bits VA (4K and 64K) and 47 bits VA (16K) for now. For 4K and 64K, supporting other VA bits is not hard, we just need to determine the size for percpu and local percpu area.
But it is harder for supporting 48 bits VA + 16K page size. We just have two top level kernel page table entries with this configuration, but we assume we just need to sync up kernel page table at the top level for now. We need to sync up kernel page table at the second level in order to support it. I'm not sure whether it is worth it or not.

If your proposal is to only support per-cpu pgtables for certain VA combinations, I don't think that will fly. We'd have too much of a test burden for these different configs.

Quoted 50 lines

Benchmark
=========
The benchmarks are done on 160 core AmpereOne machine. The baseline is v7.2-rc1 kernel.

1. Reduction of kernel text size
--------------------------------
The patchset can reduce at least 11 instructions for this_cpu_*() ops. Both preempt_disable() and preempt_enable() need 4 instructions to manipulate the preempt count, and preempt_enable() needs more instructions (compare + READ + compare) to determine whether reschedule is needed or not. Because this_cpu_*() ops are inlined and called in a lot of places so we can save a lot of instructions.

The size of kernel text is reduced by ~184KB with default Fedora kernel config. This also helps reduce kernel icache miss rate and stalled frontend cycles as kernel build benchmark result showed.

2. Kernel Build
---------------
Run kernel build (make -j160) with the default Fedora kernel config in a memcg.
13% - 18% sys time improvment
3% - 7% wall time improvement

5% fewer kernel icache miss, 5% fewer executed kernel instructions and 15% fewer stalled frontend cycles for kernel.

3. stress-ng vm ops
-------------------
stress-ng --vm 160 --vm-bytes 128M --vm-ops 100000000 8.5% improvement

4. stress-ng vm ops + fork
--------------------------
stress-ng --mmapfork 160 --mmapfork-bytes 128M --mmapfork-ops 500 15% improvement

5. Specjbb
----------
The specjbb test latency curves showed the patched kernel has consistently lower p99 latency (the lower the better) than the baseline.

2.5% improvement on max-jOPS and 4% - 5% improvement on critical-jOPS. The specjbb benchmark is quite sensitive to latency and responsiveness, particularly critical-jOPS result. The patches are supposed to improve the responsiveness due to the reduction of preempt-disabled critical sections.

Quoted 4 lines

6. MySQL
--------
1% - 2% gains on read-only test, 2% - 4% gains on write-only test. Also see 15% decrease on frontend cache stall.

These gains all look great but I expect we will see similar gains using the rseq approach, which is simpler and can be applied universally.

Thanks,
Ryan

Quoted 82 lines

Regression test
===============
1. memcg creation
-----------------
Create 10K memcgs. Each memcg creation needs to allocate multiple percpu variables, for example, percpu refcnt, rstat and objcg percpu refcnt.

Consumed 2112K more virtual memory for percpu “local mapping” and a few more mega bytes consumed by per cpu page tables.
No noticeable regression was found for elapsed time.

2. fork test
------------
stress-ng --fork 160 --fork-ops 10000000
fork() needs to allocate multiple percpu variables, for example, rss counters and mm_cid_cpu.

Roughly 1% regression was found. However stress-ng fork test has quites small address space, the real life workloads typically have much larger address space and do more complicated works. The stress-ng mmapfork benchmark saw 15% improvement.

The organization of patches
===========================
The refactor and prepatory patches (patch 1 - patch 4) Percpu page table support patches (patch 5 - patch 8) Local percpu area support patches (patch 7 - patch 15) Use local percpu area for this_cpu ops (patch 16)

Yang Shi (16):
drivers: arch_numa: move percpu set up code to arch
arm64: kconfig: make percpu related configs not depend on NUMA
mm: pgalloc: introduce {pud|pmd}_populate_sync()
vmalloc: pass in pgd pointer for vmap{__vunmap}_range_noflush()
arm64: mm: enable percpu kernel page table
arm64: mm: defined {pud|pmd}_populate_sync()
arm64: mm: sync percpu page table for memory hotplug/unplug
arm64: kasan: sync up kasan shadow area page table
arm64: mm: define percpu virtual space area
mm: percpu: prepare to use dedicated percpu area
arm64: mm: map local percpu first chunk
mm: percpu: set up first chunk and reserve chunk
arm64: mm: introduce __per_cpu_local_off
mm: percpu: allocate and free local percpu vm area
arm64: kconfig: select HAVE_LOCAL_PER_CPU_MAP
arm64: percpu: use local percpu for this_cpu_*() APIs

 arch/arm64/Kconfig                   |  12 +++++++---
 arch/arm64/include/asm/mmu.h         |   5 ++++
 arch/arm64/include/asm/mmu_context.h |   9 +++++++-
 arch/arm64/include/asm/percpu.h      |  37 ++++++++++++++++++++++++++++-
 arch/arm64/include/asm/pgalloc.h     |  24 +++++++++++++++++++
 arch/arm64/include/asm/pgtable.h     |  37 ++++++++++++++++++++++++++---
 arch/arm64/kernel/setup.c            |   3 +++
 arch/arm64/kernel/smp.c              |  44 +++++++++++++++++++++++++++++++++++
 arch/arm64/mm/kasan_init.c           |  47 +++++++++++++++++++++++--------------
 arch/arm64/mm/mmu.c                  | 165 ++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++----------------------
 arch/arm64/mm/ptdump.c               |   4 ++++
 arch/riscv/kernel/smp.c              |  51 ++++++++++++++++++++++++++++++++++++++++
 drivers/base/arch_numa.c             |  51 +---------------------------------------
 include/linux/mm.h                   |  11 +++++++++
 include/linux/percpu.h               |   4 +++-
 include/linux/pgalloc.h              |  13 +++++++++++
 include/linux/vmalloc.h              |   3 +++
 mm/Kconfig                           |   9 ++++++++
 mm/internal.h                        |   5 +++-
 mm/kmsan/hooks.c                     |  14 +++++------
 mm/percpu-internal.h                 |  14 +++++++++++
 mm/percpu-vm.c                       |  94 ++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
 mm/percpu.c                          |  58 +++++++++++++++++++++++++++++++++++++---------
 mm/sparse-vmemmap.c                  |   4 ++--
 mm/vmalloc.c                         | 138 +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++---------------
 25 files changed, 712 insertions(+), 144 deletions(-)

Thanks,
Yang

reviewer
#21 lore raw
Quoted 4 lines · Ryan Roberts

We have observed a few performance regressions recently, for which the root cause is increased use of this_cpu_*. We have somebody at Arm about to start an investigation into whether in-kernel rseq can solve the problem.

FWIW, I had a look into something rseq like (resetting the PC upon preemption or any exception), and I ended up with a GPR fixup scheme. That avoids the need to disable/enable preemption, and only requires a read of current and a couple of stores, which *should* be cheap.

The approach is similar to Peter Zijlstra's in-kernel rseq scheme [1], and Heiko Carsten's approach for s390 [2]:

[1] https://lore.kernel.org/lkml/[email protected]/
[2] https://lore.kernel.org/lkml/[email protected]/

A key difference is that I ignore the PC entirely, and *only* track whether GPRs are in use by a critical section. The GPR fixup is idempotent, and can safely be applied anywhere in the critical section. Ignoring the PC means that (in theory at least) this should work with kprobes, etc.

I've rebased and cleaned that up a bit, and pushed a WIP version to my arm64/percpu-fixup branch:

git://git.kernel.org/pub/scm/linux/kernel/git/mark/linux.git

I'll aim to have a more complete/polished version out in the near future.

The key idea is that every percpu op has a critical section during which it maintains three distinct GPRs:

<pcp> : The original __percpu pointer.
<off> : The percpu offset.
<addr>: The final pointer (<pcp> + <off>).

Whenever the kernel performs an exception return back into a critical section, it can fix up <off> and <addr>, as described below. The critical section is roughly as follows:

	/*
	 * Prologue.
	 *
	 * At the start, <pcp> is already set, but <off> and <addr> are
	 * uninitialised.
	 *
	 * After this STR, upon an exception return into this critical
	 * section, the kernel will:
	 * 1. Update <off> to the current CPU's offset.
	 * 2. Update <addr> to be <pcp> + <off>.
	 */
	mrs	<tsk>, sp_el0
	mov	<tmp>, ENCODE_REGISTER_NUMBERS(<pcp>, <off>, <addr>)
	strh	<tmp>, [<tsk>, #TSK_TI_PCPU_GPRS]

	/*
	 * Generate the address. Any preemption within this critical
	 * section will result in <off> and <addr> being updated to
	 * match the CPU that this is resumed upon.
	 */
	mrs	<off>, TPIDR_ELx
	add	<addr>, <pcp>, <off>

	/* The actual operation, e.g. read/write/cmpxchg */
	DO_SOMETHING_WITH(<addr>)

	/*
	 * Epilogue.
	 *
	 * After this STR, the kernel will no longer apply the fixups.
	 */
	strh	wzr, [<tsk>, #TSK_TI_PCPU_GPRS]

To handle nesting, the exception code is updated to save/restore the register numbers in pt_regs (and clearing the active value at entry) so that this safely nests.

With that, a simple test case such as:

  u64 outline_this_cpu_read_u64(u64 __percpu *p)
  {
          return this_cpu_read(*p);
  }

... is reduced from:

<outline_this_cpu_read_u64>:

         paciasp
         stp     x29, x30, [sp, #-32]!
         mrs     x1, sp_el0
         mov     x29, sp
         ldr     w2, [x1, #8]
         add     w2, w2, #0x1
         str     w2, [x1, #8]
         mrs     x2, tpidr_el1
         ldr     x0, [x0, x2]
         ldr     x2, [x1, #8]
         sub     x2, x2, #0x1
         str     w2, [x1, #8]
         cbz     x2, 1f
         ldr     x1, [x1, #8]
         cbnz    x1, 2f
  1:
         str     x0, [sp, #24]
         bl      0 <preempt_schedule_notrace>
         ldr     x0, [sp, #24]
  2:
         ldp     x29, x30, [sp], #32
         autiasp
         ret

... down to:

<outline_this_cpu_read_u64>:

         mrs     x2, sp_el0        // Prologue
         mov     x4, #0xc80        // Prologue
         strh    w4, [x2, #20]     // Prologue
         mrs     x4, tpidr_el1
         add     x3, x0, x4
         ldr     x1, [x3]
         strh    wzr, [x2, #20]    // Epilogue
         mov     x0, x1

ret

... which is clearly much better.

As noted above, the branch is a WIP. There are a bunch of things to do (in particular, factoring out the xchg and cmpxchg assembly), but I don't currently see a major technical blocker for the fixup approach.

Mark.

author
#22 lore raw
In reply to Ryan Roberts
Quoted 14 lines · Ryan Roberts
Quoted 9 lines · Yang Shi

Hi,

This is v2 RFC. In v2 a lot problems found out by Sashiko were fixed and more feature gaps were closed (please see the below changelog for the details). Although there are still some open issues, for example, it just can support 48 bits VA (for 4K and 64K) and 47 bits VA (for 16K), KPTI support has not been solved yet, etc, but I think the delta should be big enough and worth a new RFC to gather comments in order to make sure I'm on the right track.

Quoted 4 lines

Some more benchmarks were done, for example, some latency related benchmarks that I mentioned at LSFMM because I thought responsiveness should be improved due to the removal of preempt_disable. Collected more PMU counters as well. Please refer to the benchmark section for more details.

Quoted 6 lines

Hi Yang,

When we spoke about this at LSFMM, I said I would take the series for a spin on other hardware to see if the performance gains generalise. I did make a start on that but the kernel wouldn't even boot and I never found time to debug. Sorry that I didn't let you know sooner.

No worries. There are some known issues, for example, KPTI doesn't work yet so I need to disable CONFIG_UNMAP_KERNEL_AT_EL0 in order to make it boot on Altra/AltraMax or any others which need KPTI, and 16K/64K page sizes didn't work in v1, but they are fixed in this spin.

Quoted 54 lines

But I think that there are other issues with this concept that likely block it from being accepted upstream; see below...

Quoted 37 lines

Look forward to comments.

Changelog
v2: * Added 3-level and 2-level page table support.
* Tested with 16K and 64K page size. But we just support 48 bits VA (4K and 64K) and 47 bits VA with 16K for now. Please refer to the below "known issue" section for the detail reason.
* Added support for memory hotplug.
* Added support for KASAN (generic).
* Treated percpu and local percpu area address as vmalloc address.
* Fixed build failure for x86.
* Fixed build failure for !CONFIG_NUMA.
* Added KASAN support for local percpu area.
* Some other misc bug fixes found out by Sashiko.
* More code refactor and cleanup.
* Regorganized the patches.
* More benchmarks, refer to benchmark section for more details.
* Rebased to v7.2-rc1.

Introduction
============
This patch series implemented the LSFMM 2026 proposal for optimizing this_cpu_*() ops on ARM64. For the details of the proposal, Please refer to:
https://lore.kernel.org/linux-mm/CAHbLzkpcN-T8MH6=W3jCxcFj1gVZp8fRqe231yzZT-rV_E_org@mail.gmail.com/
I didn't repeat it in the cover letter because there is no change to the proposal.

The series is based on 7.1-rc1. It is basically minimum viable patches. There are still a few hacks in this series and it may break something, for example, KPTI, SMT machines which shared TLB, etc. But it shoule be good enough for now to demonstrate the core idea. The main purpose of the RFC is to gather feedback, figure out missing parts and risks, and make sure we are on the right track, as well as hopefully it can help the discussion for the upcoming LSFMM.

Quoted 7 lines

I broke the patches down to arch-dependent and arch-independent parts so that hopefully the interested persons can do experiments on other architectures, for example, S390, easier.

A new kernel config is introduced, HAVE_LOCAL_PER_CPU_MAP. The architectures which can support this feature will select it. Allocating and freeing percpu local mapping is protected by this config so that others won't pay the cost.

I think this approach is a bit of a sledge hammer to crack a nut. If the problem is the cost this_cpu_* ops on arm64 due to the need to disable/enable preemption, I think we can solve that much more simply with [1], which introduces in-kernel restartable sequences. Using this, the operation no longer needs to disable preemption but can instead just detect when it is preempted and retry.

Quoted 6 lines

[1]
https://lore.kernel.org/all/[email protected]/

We have observed a few performance regressions recently, for which the root cause is increased use of this_cpu_*. We have somebody at Arm about to start an investigation into whether in-kernel rseq can solve the problem.

It is good to know someone is exploring the alternative path. But I doubt restart sequence can really boost the performance. It needs check whether preemption (or cpu migration) happens in the middle of percpu operations then manipulate the ip to restart the operation. It sounds like it just moved the cost from one place to the other place and it also seems hacky TBH. It may reduce the preempt_disable/preempt_enable cost somehow for some less contended cases, but it may be worse for high contended cases. The percpu page table can work well for both.

S390 implemented restart sequence. Heiko didn't share too much benchmark data, just confirmed there is no regression and I had a lengthy discussion with him (https://lore.kernel.org/lkml/[email protected]/).

Improving this_cpu performance is just one of the usecase of percpu page table. The other potential usecase is kernel text replication I mentioned at LSFMM. The idea is not new, someone else has explored it, see https://lore.kernel.org/linux-arm-kernel/[email protected]/. Percpu page table can make kernel support it more naturally.

Quoted 40 lines
Quoted 23 lines

Known Issues
============
1. KPTI
-------
We need determine what CPU we are on, then switch to the right page table. Currently arm64 kernel fetches tramp_pg_dir via swapper_pg_dir - fixed_offset, and fetches swapper_pg_dir from ttbr1. But ttbr1 may not hold swapper_pg_dir anymore except CPU #0. So we need to figure out the other way to handle it. Switching to tramp_pg_dir should be easy, but the reverse seems harder because tramp_pg_dir just maps the trampoline vectors.
Maybe we can do two steps switch. Switch to swapper_pg_dir at the first step, then switch to per cpu page table (for entry) or tramp page table (for exit). Nobody should call this_cpu_*() at either userspace -> kernel entry stage or kernel -> userspace exit stage.

2. SW PAN
---------
Has the similar issue as KPTI. It installs reserved_pg_dir to TTBR0 when running in kernel space, but fetching reserved_pg_dir via swapper_pg_dir - fixed_offset. Maybe we can save the physical address of swapper_pg_dir in a variable, then load it from that variable instead of ttbr1.

Quoted 7 lines

3. Shared TLB machines
----------------------
Some machines may share TLB between CPUs, for example, SMT machines may share TLB between the two hardware threads in one core.
The per cpu page table just can't work with it. Maybe we need a new cpufeature to indicate whether per cpu page table is allowed or not. Then just enable it for not-shared-TLB machines.

The architectural feature you're referring to here is FEAT_TTCNP (common not private). There are many CPUs out there that use this feature (not just those that support SMT).

FEAT_TTCNP is mandatory from Armv8.2 and the presence of the feature only tells you whether you can legally set the CnP bit - it doesn't actually tell you if the HW does any sharing - there is no way to detect this. So making per-cpu pgtables mutually exclusive with CNP is a non-starter as all CPUs from v8.2 onwards advertise CNP.

Quoted 4 lines

Even ignoring the CNP problem, you'll end up installing the per-cpu TLB entries tagged with whatever user ASID happens to be installed at the time. Which means you could end up with lots of duplicated TLB entries all differing only by ASID. It's probably not the end of the world, but it seems... inefficient.

Really? I thought ASID will be ignored if nG == 0.

Quoted 4 lines

The only way I can see to make this work fully is to rely on FEAT_ASID2, which allows specifying separate ASIDs for TTBR0 and TTBR1. Then each CPU can have it's own (permanently installed) ASID for local mappings in TTBR1. That would solve the CNP issue. FEAT_ASID2 is optional from v9.4 and mandatory from v9.5.

I agree it is safer to handle CNP problem by depending on FEAT_ASID2.

Quoted 10 lines
Quoted 10 lines

4. Don't support all VA bits
----------------------------
We just support 48 bits VA (4K and 64K) and 47 bits VA (16K) for now. For 4K and 64K, supporting other VA bits is not hard, we just need to determine the size for percpu and local percpu area.
But it is harder for supporting 48 bits VA + 16K page size. We just have two top level kernel page table entries with this configuration, but we assume we just need to sync up kernel page table at the top level for now. We need to sync up kernel page table at the second level in order to support it. I'm not sure whether it is worth it or not.

Quoted 3 lines

If your proposal is to only support per-cpu pgtables for certain VA combinations, I don't think that will fly. We'd have too much of a test burden for these different configs.

OK, I can try to add support for other VA combinations. That should require sync'ing up percpu page tables at both the top level and the second level.

Thanks,
Yang

Quoted 138 lines
Quoted 48 lines

Benchmark
=========
The benchmarks are done on 160 core AmpereOne machine. The baseline is v7.2-rc1 kernel.

1. Reduction of kernel text size
--------------------------------
The patchset can reduce at least 11 instructions for this_cpu_*() ops. Both preempt_disable() and preempt_enable() need 4 instructions to manipulate the preempt count, and preempt_enable() needs more instructions (compare + READ + compare) to determine whether reschedule is needed or not. Because this_cpu_*() ops are inlined and called in a lot of places so we can save a lot of instructions.

The size of kernel text is reduced by ~184KB with default Fedora kernel config. This also helps reduce kernel icache miss rate and stalled frontend cycles as kernel build benchmark result showed.

2. Kernel Build
---------------
Run kernel build (make -j160) with the default Fedora kernel config in a memcg.
13% - 18% sys time improvment
3% - 7% wall time improvement

5% fewer kernel icache miss, 5% fewer executed kernel instructions and 15% fewer stalled frontend cycles for kernel.

3. stress-ng vm ops
-------------------
stress-ng --vm 160 --vm-bytes 128M --vm-ops 100000000 8.5% improvement

4. stress-ng vm ops + fork
--------------------------
stress-ng --mmapfork 160 --mmapfork-bytes 128M --mmapfork-ops 500 15% improvement

5. Specjbb
----------
The specjbb test latency curves showed the patched kernel has consistently lower p99 latency (the lower the better) than the baseline.

2.5% improvement on max-jOPS and 4% - 5% improvement on critical-jOPS. The specjbb benchmark is quite sensitive to latency and responsiveness, particularly critical-jOPS result. The patches are supposed to improve the responsiveness due to the reduction of preempt-disabled critical sections.

Quoted 4 lines

6. MySQL
--------
1% - 2% gains on read-only test, 2% - 4% gains on write-only test. Also see 15% decrease on frontend cache stall.

These gains all look great but I expect we will see similar gains using the rseq approach, which is simpler and can be applied universally.

Thanks,
Ryan

Quoted 80 lines

Regression test
===============
1. memcg creation
-----------------
Create 10K memcgs. Each memcg creation needs to allocate multiple percpu variables, for example, percpu refcnt, rstat and objcg percpu refcnt.

Consumed 2112K more virtual memory for percpu “local mapping” and a few more mega bytes consumed by per cpu page tables.
No noticeable regression was found for elapsed time.

2. fork test
------------
stress-ng --fork 160 --fork-ops 10000000
fork() needs to allocate multiple percpu variables, for example, rss counters and mm_cid_cpu.

Roughly 1% regression was found. However stress-ng fork test has quites small address space, the real life workloads typically have much larger address space and do more complicated works. The stress-ng mmapfork benchmark saw 15% improvement.

The organization of patches
===========================
The refactor and prepatory patches (patch 1 - patch 4) Percpu page table support patches (patch 5 - patch 8) Local percpu area support patches (patch 7 - patch 15) Use local percpu area for this_cpu ops (patch 16)

Yang Shi (16):
drivers: arch_numa: move percpu set up code to arch
arm64: kconfig: make percpu related configs not depend on NUMA
mm: pgalloc: introduce {pud|pmd}_populate_sync()
vmalloc: pass in pgd pointer for vmap{__vunmap}_range_noflush()
arm64: mm: enable percpu kernel page table
arm64: mm: defined {pud|pmd}_populate_sync()
arm64: mm: sync percpu page table for memory hotplug/unplug
arm64: kasan: sync up kasan shadow area page table
arm64: mm: define percpu virtual space area
mm: percpu: prepare to use dedicated percpu area
arm64: mm: map local percpu first chunk
mm: percpu: set up first chunk and reserve chunk
arm64: mm: introduce __per_cpu_local_off
mm: percpu: allocate and free local percpu vm area
arm64: kconfig: select HAVE_LOCAL_PER_CPU_MAP
arm64: percpu: use local percpu for this_cpu_*() APIs

  arch/arm64/Kconfig                   |  12 +++++++---
  arch/arm64/include/asm/mmu.h         |   5 ++++
  arch/arm64/include/asm/mmu_context.h |   9 +++++++-
  arch/arm64/include/asm/percpu.h      |  37 ++++++++++++++++++++++++++++-
  arch/arm64/include/asm/pgalloc.h     |  24 +++++++++++++++++++
  arch/arm64/include/asm/pgtable.h     |  37 ++++++++++++++++++++++++++---
  arch/arm64/kernel/setup.c            |   3 +++
  arch/arm64/kernel/smp.c              |  44 +++++++++++++++++++++++++++++++++++
  arch/arm64/mm/kasan_init.c           |  47 +++++++++++++++++++++++--------------
  arch/arm64/mm/mmu.c                  | 165 ++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++----------------------
  arch/arm64/mm/ptdump.c               |   4 ++++
  arch/riscv/kernel/smp.c              |  51 ++++++++++++++++++++++++++++++++++++++++
  drivers/base/arch_numa.c             |  51 +---------------------------------------
  include/linux/mm.h                   |  11 +++++++++
  include/linux/percpu.h               |   4 +++-
  include/linux/pgalloc.h              |  13 +++++++++++
  include/linux/vmalloc.h              |   3 +++
  mm/Kconfig                           |   9 ++++++++
  mm/internal.h                        |   5 +++-
  mm/kmsan/hooks.c                     |  14 +++++------
  mm/percpu-internal.h                 |  14 +++++++++++
  mm/percpu-vm.c                       |  94 ++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
  mm/percpu.c                          |  58 +++++++++++++++++++++++++++++++++++++---------
  mm/sparse-vmemmap.c                  |   4 ++--
  mm/vmalloc.c                         | 138 +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++---------------
  25 files changed, 712 insertions(+), 144 deletions(-)

Thanks,
Yang

reviewer
#23 lore raw
Quoted 5 lines · Yang Shi
Quoted 3 lines · Ryan Roberts

On 15/07/2026 19:04, Yang Shi wrote:
But I think that there are other issues with this concept that likely block it from being accepted upstream; see below...

Quoted 8 lines
Quoted 3 lines

We have observed a few performance regressions recently, for which the root cause is increased use of this_cpu_*. We have somebody at Arm about to start an investigation into whether in-kernel rseq can solve the problem.

It is good to know someone is exploring the alternative path. But I doubt restart sequence can really boost the performance. It needs check whether preemption (or cpu migration) happens in the middle of percpu operations then manipulate the ip to restart the operation.

With the rseq approach, that check happens out-of-line in exception handling code, *only* when preemption occurs, so in the fast path there is *no code whatsoever* to check for preemption/migration, and no branches to cause a restart.

Please note that with the approach I've prototyped, the exception handling code doesn't even adjust the PC; we just adjust a GPR containing the address. The forward progress guarantees (and behaviour under contention) should be the same as a single instruction. The fixup logic itself is trivial.

Please see my message where I outlined that approach in detail:

https://lore.kernel.org/linux-arm-kernel/al_DpFJFcmVhxpvW@J2N7QTR9R3/

I expect that should come with a reasonable benefit, but I don't have benchmark figures yet as I haven't finished converting the xchg and cmpxchg implementations.

Quoted 2 lines

It sounds like it just moved the cost from one place to the other place and it also seems hacky TBH.

The existing (high) cost is removed from the fast path. A new (low) cost is added to the rare path where we perform an exception return into the middle of a critical section.

I don't think this is any more hacky than percpu page tables. It's far more contained to the relevant pieces of code (only the this_cpu_*() ops and exception entry/return).

Quoted 3 lines

It may reduce the preempt_disable/preempt_enable cost somehow for some less contended cases, but it may be worse for high contended cases. The percpu page table can work well for both.

It entirely removes the preempt_disable/preempt_enable cost.

I don't follow your concern with contention. Any contention on the memory location is going to be independent of preemption, and if a thread is preempted mid-operation, the impact is going to be the same with either the adress fixup or percpu page tables -- I don't think you avoid that.

Quoted 4 lines

S390 implemented restart sequence. Heiko didn't share too much benchmark data, just confirmed there is no regression and I had a lengthy discussion with him
(https://lore.kernel.org/lkml/[email protected]/).

As above, for s390 Heiko didn't implement a restart sequence; he implemented an address fixup sequence. There is no restart.

The discussion you had there seemed to focus on the s390 implementation details (e.g. why he used an address fixup rather than a restart), and the benefits of removing the preempt_count manipulation (which all the approaches, including mine, do).

Am I missing some key detail?

Quoted 3 lines

Improving this_cpu performance is just one of the usecase of percpu page
table. The other potential usecase is kernel text replication I mentioned at
LSFMM. The idea is not new, someone else has explored it, see https://lore.kernel.org/linux-arm-kernel/[email protected]/.

That was looked at in the past, but people weren't keen given the complexity it would lead to, which is why it never saw much traction.

Quoted 1 line

Percpu page table can make kernel support it more naturally.

Perhaps, but the only difference is that you've moving the support code into the core kernel mm code. I think the general concerns people have with percpu page tables still apply.

[...]

Quoted 20 lines
Quoted 20 lines
Quoted 10 lines

Known Issues
============

3. Shared TLB machines
----------------------
Some machines may share TLB between CPUs, for example, SMT machines may share TLB between the two hardware threads in one core.
The per cpu page table just can't work with it. Maybe we need a new cpufeature to indicate whether per cpu page table is allowed or not. Then just enable it for not-shared-TLB machines.

The architectural feature you're referring to here is FEAT_TTCNP (common not private). There are many CPUs out there that use this feature (not just those that support SMT).

FEAT_TTCNP is mandatory from Armv8.2 and the presence of the feature only tells you whether you can legally set the CnP bit - it doesn't actually tell you if the HW does any sharing - there is no way to detect this. So making per-cpu pgtables mutually exclusive with CNP is a non-starter as all CPUs from v8.2 onwards advertise CNP.

Quoted 6 lines
Quoted 4 lines

Even ignoring the CNP problem, you'll end up installing the per-cpu TLB entries tagged with whatever user ASID happens to be installed at the time. Which means you could end up with lots of duplicated TLB entries all differing only by ASID. It's probably not the end of the world, but it seems... inefficient.

Really? I thought ASID will be ignored if nG == 0.

The nG bit only applies to leaf entries (Page or Block descriptors). All intermediate entries (Table descriptors) are ASID tagged.

Quoted 6 lines
Quoted 4 lines

The only way I can see to make this work fully is to rely on FEAT_ASID2, which allows specifying separate ASIDs for TTBR0 and TTBR1. Then each CPU can have it's own (permanently installed) ASID for local mappings in TTBR1. That would solve the CNP issue. FEAT_ASID2 is optional from v9.4 and mandatory from v9.5.

I agree it is safer to handle CNP problem by depending on FEAT_ASID2.

I think that depending on FEAT_ASID2 is a non-starter.

There is no extant hardware with FEAT_ASID2, and I don't think we want to maintain distinct versions of hte percpu code accross architecture versions.

Mark.

author
#24 lore raw
Quoted 23 lines · Mark Rutland
Quoted 11 lines · Yang Shi
Quoted 6 lines · Ryan Roberts

On 15/07/2026 19:04, Yang Shi wrote:
But I think that there are other issues with this concept that likely block it from being accepted upstream; see below...
We have observed a few performance regressions recently, for which the root cause is increased use of this_cpu_*. We have somebody at Arm about to start an investigation into whether in-kernel rseq can solve the problem.

It is good to know someone is exploring the alternative path. But I doubt restart sequence can really boost the performance. It needs check whether preemption (or cpu migration) happens in the middle of percpu operations then manipulate the ip to restart the operation.

With the rseq approach, that check happens out-of-line in exception handling code, *only* when preemption occurs, so in the fast path there is *no code whatsoever* to check for preemption/migration, and no branches to cause a restart.

Please note that with the approach I've prototyped, the exception handling code doesn't even adjust the PC; we just adjust a GPR containing the address. The forward progress guarantees (and behaviour under contention) should be the same as a single instruction. The fixup logic itself is trivial.

Quoted 7 lines

Please see my message where I outlined that approach in detail:

https://lore.kernel.org/linux-arm-kernel/al_DpFJFcmVhxpvW@J2N7QTR9R3/

I expect that should come with a reasonable benefit, but I don't have benchmark figures yet as I haven't finished converting the xchg and cmpxchg implementations.

Yes, but it still needs prologue and epilogue which consumes 4 instructions in every inlined this_cpu ops if I read it correctly. I agree it is still a very good improvement. I just tried to compare it with my approach. My approach doesn't need prologue and epilogue, and the fixup in exception handling code as well.

Quoted 14 lines
Quoted 3 lines

It sounds like it just moved the cost from one place to the other place and it also seems hacky TBH.

The existing (high) cost is removed from the fast path. A new (low) cost is added to the rare path where we perform an exception return into the middle of a critical section.

I don't think this is any more hacky than percpu page tables. It's far more contained to the relevant pieces of code (only the this_cpu_*() ops and exception entry/return).

Quoted 3 lines

It may reduce the preempt_disable/preempt_enable cost somehow for some less contended cases, but it may be worse for high contended cases. The percpu page table can work well for both.

Quoted 7 lines

It entirely removes the preempt_disable/preempt_enable cost.

I don't follow your concern with contention. Any contention on the memory location is going to be independent of preemption, and if a thread is preempted mid-operation, the impact is going to be the same with either the adress fixup or percpu page tables -- I don't think you avoid that.

Contention means this_cpu ops is preempted and migrated to other CPUs quite often. If so the fix up cost may be high because you need to reload percpu offset and recalculate the pointer every time, at least extra two instructions. This case may be rare, but the percpu page table approach doesn't need to worry about the worst case scenario at all. Manipulating the local percpu counter is guaranteed by page table no matter how often preemption and migration happen.

Quoted 8 lines
Quoted 5 lines

S390 implemented restart sequence. Heiko didn't share too much benchmark data, just confirmed there is no regression and I had a lengthy discussion with him
(https://lore.kernel.org/lkml/[email protected]/).

As above, for s390 Heiko didn't implement a restart sequence; he implemented an address fixup sequence. There is no restart.

Quoted 6 lines

The discussion you had there seemed to focus on the s390 implementation details (e.g. why he used an address fixup rather than a restart), and the benefits of removing the preempt_count manipulation (which all the approaches, including mine, do).

Am I missing some key detail?

I was trying to understand S390's implementation so that I can compare to the percpu page table approach. As we discussed above, both S390 and your proposal does reduce the cost, but not remove the cost. This is the key detail.

I also admit fixup sequence approach is more simple and self-contained than percpu page table. So we were also seeking other usecases for percpu page table so that it can be worth the complexity.

Quoted 6 lines
Quoted 4 lines

Improving this_cpu performance is just one of the usecase of percpu page
table. The other potential usecase is kernel text replication I mentioned at
LSFMM. The idea is not new, someone else has explored it, see https://lore.kernel.org/linux-arm-kernel/[email protected]/.

That was looked at in the past, but people weren't keen given the complexity it would lead to, which is why it never saw much traction.

Some our customers are interested in this. Russell's 2023 RFC said "the performance results from kernel text replication are workload specific, but appear to show a gain of between 6% and 17% for database-centric like workloads. When combined with userspace awareness of NUMA, this can result in a gain of over 50%.". The performance gain sounds impressive.

Brendan Jackman also mentioned percpu page table could make his "address space isolation" work better at LSFMM, but I have not explored it too much yet.

Multiple usercases may make percpu page table more convincing?

Quoted 27 lines
Quoted 2 lines

Percpu page table can make kernel support it more naturally.

Perhaps, but the only difference is that you've moving the support code into the core kernel mm code. I think the general concerns people have with percpu page tables still apply.

[...]

Quoted 19 lines
Quoted 19 lines
Quoted 9 lines

Known Issues
============
3. Shared TLB machines
----------------------
Some machines may share TLB between CPUs, for example, SMT machines may share TLB between the two hardware threads in one core.
The per cpu page table just can't work with it. Maybe we need a new cpufeature to indicate whether per cpu page table is allowed or not. Then just enable it for not-shared-TLB machines.

The architectural feature you're referring to here is FEAT_TTCNP (common not private). There are many CPUs out there that use this feature (not just those that support SMT).

FEAT_TTCNP is mandatory from Armv8.2 and the presence of the feature only tells you whether you can legally set the CnP bit - it doesn't actually tell you if the HW does any sharing - there is no way to detect this. So making per-cpu pgtables mutually exclusive with CNP is a non-starter as all CPUs from v8.2 onwards advertise CNP.

Quoted 7 lines
Quoted 5 lines
Quoted 4 lines

Even ignoring the CNP problem, you'll end up installing the per-cpu TLB entries tagged with whatever user ASID happens to be installed at the time. Which means you could end up with lots of duplicated TLB entries all differing only by ASID. It's probably not the end of the world, but it seems... inefficient.

Really? I thought ASID will be ignored if nG == 0.

The nG bit only applies to leaf entries (Page or Block descriptors). All intermediate entries (Table descriptors) are ASID tagged.

OK, good to know it. But I still don't quite understand how come percpu page table can cause duplicated TLB entries with different ASIDs. Doesn't the current kernel (one single kernel page table) have ASID tagged intermediate entries too?

Quoted 11 lines
Quoted 6 lines
Quoted 5 lines

The only way I can see to make this work fully is to rely on FEAT_ASID2, which allows specifying separate ASIDs for TTBR0 and TTBR1. Then each CPU can have it's own (permanently installed) ASID for local mappings in TTBR1. That would solve the CNP issue. FEAT_ASID2 is optional from v9.4 and mandatory from v9.5.

I agree it is safer to handle CNP problem by depending on FEAT_ASID2.

I think that depending on FEAT_ASID2 is a non-starter.

There is no extant hardware with FEAT_ASID2, and I don't think we want to maintain distinct versions of hte percpu code accross architecture versions.

I agree it is not perfect to have different version of percpu code by depending on architecture versions. But AFAICT percpu is not the first one. We already have KPTI, SW PAN, bbml2_noabort, etc.

Thanks,
Yang

Quoted 2 lines

Mark.

#25 lore raw
In reply to Mark Rutland
Quoted 6 lines · Mark Rutland

I expect that should come with a reasonable benefit, but I don't have benchmark figures yet as I haven't finished converting the xchg and cmpxchg implementations.

Quoted 2 lines

It sounds like it just moved the cost from one place to the other place and it also seems hacky TBH.

Yang Shi's patch has *no* critical section. There is no additional code for the RMV instruction. The RMV instruction is executed on the correct per cpu area. One of the reasons for the performance win is the eliminattion of these critical sections. Your approach still has some form of prologue and posthandling like the current preempt approach and therefore will not be able to have the same performance gains.

The code is more efficient, there is no restart necessary and the technique is already widely used on x86 for a long time.

Having the ability in general to map mmemory differently depending on the cpu opens up a number of other optimization like

1. Per Node areas. Calculations of addresses for per node data and RMV operations on per node data becomes as trivial as the per cpu data handling. This will further reduce and eliminate critical sections currently necessary to handle per node data modifications for NUMA configurations which are increasingly becoming important for large core configurations on ARM64.

2. Kernel text replication. Kernel text can be replicated per node or per whatever memory is closer to the executing code. This avoids transfers via the on chip memory busses and increases performance. We have seen 20-50% on that one.

3. Readonly data replication. A similar approach is possible for kernel read only data.

3. Read-mostly replication. This is a bit more complex and the writes become more expensive since updates have to be made to all copies but a read-mostly variable is rarely written and the replication will reduce the latencies to reach these variables.

4. Custom sets of cores that operate on shared data that is replicated per node or some custom set of cores. This is for example useful for network devices that have the ability to do I/O via a split PCI bus or other construct where I/O lanes are duplicated to sets of cores.

What we are proposing here is a basic new feature that simplifies code and allows addititonal performance and functional features that are so far not possible on ARM64.

#26 lore raw
Quoted 51 lines · Christoph Lameter (Ampere)
Quoted 6 lines · Mark Rutland

I expect that should come with a reasonable benefit, but I don't have benchmark figures yet as I haven't finished converting the xchg and cmpxchg implementations.

Quoted 2 lines

It sounds like it just moved the cost from one place to the other place and it also seems hacky TBH.

Yang Shi's patch has *no* critical section. There is no additional code for the RMV instruction. The RMV instruction is executed on the correct per cpu area. One of the reasons for the performance win is the eliminattion of these critical sections. Your approach still has some form of prologue and posthandling like the current preempt approach and therefore will not be able to have the same performance gains.

The code is more efficient, there is no restart necessary and the technique is already widely used on x86 for a long time.

Having the ability in general to map mmemory differently depending on the cpu opens up a number of other optimization like

1. Per Node areas. Calculations of addresses for per node data and RMV operations on per node data becomes as trivial as the per cpu data handling. This will further reduce and eliminate critical sections currently necessary to handle per node data modifications for NUMA configurations which are increasingly becoming important for large core configurations on ARM64.

2. Kernel text replication. Kernel text can be replicated per node or per whatever memory is closer to the executing code. This avoids transfers via the on chip memory busses and increases performance. We have seen 20-50% on that one.

3. Readonly data replication. A similar approach is possible for kernel read only data.

3. Read-mostly replication. This is a bit more complex and the writes become more expensive since updates have to be made to all copies but a read-mostly variable is rarely written and the replication will reduce the latencies to reach these variables.

4. Custom sets of cores that operate on shared data that is replicated per node or some custom set of cores. This is for example useful for network devices that have the ability to do I/O via a split PCI bus or other construct where I/O lanes are duplicated to sets of cores.

Quoted 3 lines

What we are proposing here is a basic new feature that simplifies code and allows addititonal performance and functional features that are so far not possible on ARM64.

Most of the features you mentioned above are not supposed to be architecture-specific things. IIUC, Linus strongly objected per-cpu page tables in the past, which suggests to me that such an arm64-only thing is not the way to go?

Signature · 3 lines
Cheers,

David
#27 lore raw

+cc Linus, Jason

Quoted 52 lines · David Hildenbrand (Arm)
Quoted 51 lines · Christoph Lameter (Ampere)
Quoted 6 lines · Mark Rutland

I expect that should come with a reasonable benefit, but I don't have benchmark figures yet as I haven't finished converting the xchg and cmpxchg implementations.

Quoted 2 lines

It sounds like it just moved the cost from one place to the other place and it also seems hacky TBH.

Yang Shi's patch has *no* critical section. There is no additional code for the RMV instruction. The RMV instruction is executed on the correct per cpu area. One of the reasons for the performance win is the eliminattion of these critical sections. Your approach still has some form of prologue and posthandling like the current preempt approach and therefore will not be able to have the same performance gains.

The code is more efficient, there is no restart necessary and the technique is already widely used on x86 for a long time.

Having the ability in general to map mmemory differently depending on the cpu opens up a number of other optimization like

1. Per Node areas. Calculations of addresses for per node data and RMV operations on per node data becomes as trivial as the per cpu data handling. This will further reduce and eliminate critical sections currently necessary to handle per node data modifications for NUMA configurations which are increasingly becoming important for large core configurations on ARM64.

2. Kernel text replication. Kernel text can be replicated per node or per whatever memory is closer to the executing code. This avoids transfers via the on chip memory busses and increases performance. We have seen 20-50% on that one.

3. Readonly data replication. A similar approach is possible for kernel read only data.

3. Read-mostly replication. This is a bit more complex and the writes become more expensive since updates have to be made to all copies but a read-mostly variable is rarely written and the replication will reduce the latencies to reach these variables.

4. Custom sets of cores that operate on shared data that is replicated per node or some custom set of cores. This is for example useful for network devices that have the ability to do I/O via a split PCI bus or other construct where I/O lanes are duplicated to sets of cores.

Quoted 8 lines
Quoted 3 lines

What we are proposing here is a basic new feature that simplifies code and allows addititonal performance and functional features that are so far not possible on ARM64.

Most of the features you mentioned above are not supposed to be architecture-specific things. IIUC, Linus strongly objected per-cpu page tables in the past, which suggests to me that such an arm64-only thing is not the way to go?

Linus - I assume you still find per-cpu page tables totally unacceptable? (had a look around to see your take on this recently and found [0] which seems to confirm it :)

Yang/Christoph - You ran a LSF session on this but it seems there was significant push-back there also - https://lwn.net/Articles/1073395/ - it seemed pretty damning on the approach as a whole, so given Linus's and the community's rejection of it is it really worth pursuing this any further?

(Adding Jason as he was involved in that discussion.)

From my side, having recently done a bunch of work with kernel page tables and the absolute mess they're in, with intent to move to RCU kernel page table freeing, I'm very concerned about this adding additional complexity and making the already very fraught page table situation we have in the kernel even worse.

Since this work seems to be very much arm64-focused, perhaps it's therefore worth looking at an alterative solution that's specific to the arch, like the one suggested by Mark ([1])?

[0]:https://lore.kernel.org/all/CAHk-=wire3dzhHx=KiL_f5Rj0=1u9ustsa33QoR-F9-v-NU9Ng@mail.gmail.com/ [1]:https://lore.kernel.org/linux-arm-kernel/al_DpFJFcmVhxpvW@J2N7QTR9R3/

Quoted 6 lines

--
Cheers,

David

--
Cheers, Lorenzo

#28 lore raw
Quoted 3 lines · Lorenzo Stoakes (ARM)

Since this work seems to be very much arm64-focused, perhaps it's therefore worth looking at an alterative solution that's specific to the arch, like the one suggested by Mark ([1])?

Oh I hadn't realised Mark had already posted the series here:

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

Quoted 1 line

[1]:https://lore.kernel.org/linux-arm-kernel/al_DpFJFcmVhxpvW@J2N7QTR9R3/

--
Cheers, Lorenzo

#29 lore raw
Quoted 8 lines · Lorenzo Stoakes (ARM)
Quoted 3 lines · Lorenzo Stoakes (ARM)

Since this work seems to be very much arm64-focused, perhaps it's therefore worth looking at an alterative solution that's specific to the arch, like the one suggested by Mark ([1])?

Oh I hadn't realised Mark had already posted the series here:

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

It is a really neat approach for CPUs that lack the x86 like load from computed ea trick and should solve the original request.

That said there is still a merit for things like replicating .text, and I think we will also have some multi-socket systems that would benefit from this too..

So, I think the question you ask: if we should ever consider divergent per-cpu kernel page tables - is still interesting..

Jason

#30 lore raw
Quoted 3 lines · Jason Gunthorpe

That said there is still a merit for things like replicating .text, and I think we will also have some multi-socket systems that would benefit from this too..

Are you referring to kernel text replication or userspace text replication? If the latter, I don't think we need per-CPU page tables to do it.

#31 lore raw
Quoted 8 lines · Matthew Wilcox
Quoted 3 lines · Jason Gunthorpe

That said there is still a merit for things like replicating .text, and I think we will also have some multi-socket systems that would benefit from this too..

Are you referring to kernel text replication or userspace text replication? If the latter, I don't think we need per-CPU page tables to do it.

In a sense both, if you build a system with poor cross-socket performance then there can be big wins to replicate .text.

I've thought kernel needs to do it with text replication, there are too many absolute relocations to make two .text VAs work?

Userspace can meaningfully work on a non-replicated basis where single processes are localized to sockets. I think people do things like this already, but AFAIK it is not ergonomic to setup.

Jason

#32 lore raw
In reply to Lorenzo Stoakes (ARM)
Quoted 4 lines · Lorenzo Stoakes (ARM)

Linus - I assume you still find per-cpu page tables totally unacceptable? (had a look around to see your take on this recently and found [0] which seems to confirm it :)

Yes, I still think that per-cpu page tables as some kind of architectural design is a horrible idea, because it paints us into a corner where you have intentional duplication and then end up having to have coherency issues where you need to lock for updates just to keep things in sync - and may need to do extra work on hardware that may otherwise already have hardware coherency support.

I honestly think that people who think that cache coherency in software is a great idea in general are misguided and should not work on kernels, but should go back to eating crayons and decide if purple tastes better than yellow.

That said - per-cpu page tables as some kind of internal architecture choice that limits the damage to some simpler case is not necessarily wrong. We've obviously done things like that for page table isolation etc. It can be horribly expensive, but sometimes that is the price you are willing to pay - and sometimes you can avoid the expense entirely by knowing how the particular hardware works.

Very specific example: on 32-bit x86 with PAE, the top-level page table directory is fundamentally per-cpu in nature. It's just four entries that get loaded into the CPU on page table load, but it means that you can do things like having one quarter of the page tables be per-cpu, and the rest be shared. Several other architectures have similar "regions". So it can be an architectural choice to have some kind of per-cpu TLB. But I absolutely do *not* want to see a model where the general MM layer dictates that design.

And there are historical reasons to really hate it. Plan-9 had a completely broken threading model that depended on per-thread page tables. It was literal garbage. It was broken beyond words exactly due to coherency issues, and I still do not understand how otherwise very smart people ever came up with such a stupid idea.

We are not making _that_ mistake. We have made enough other ones to not neet to shoot ourselves in the head that badly.

Linus

#33 lore raw
Quoted 6 lines · Linus Torvalds
Quoted 4 lines · Lorenzo Stoakes (ARM)

Linus - I assume you still find per-cpu page tables totally unacceptable? (had a look around to see your take on this recently and found [0] which seems to confirm it :)

Quoted 6 lines

Yes, I still think that per-cpu page tables as some kind of architectural design is a horrible idea, because it paints us into a corner where you have intentional duplication and then end up having to have coherency issues where you need to lock for updates just to keep things in sync - and may need to do extra work on hardware that may otherwise already have hardware coherency support.

We already have had the intentional duplication, meaning multiple virtual addresses for a single physical address with vmalloc and with vmemmap for a very long time. Issues with hardware coherency that may have existed for these cases have been fixed a long time ago.

Quoted 4 lines

I honestly think that people who think that cache coherency in software is a great idea in general are misguided and should not work on kernels, but should go back to eating crayons and decide if purple tastes better than yellow.

Certain a bad idea. The per cpu pagew tables do not rely on cache coherency in software.

Quoted 6 lines

That said - per-cpu page tables as some kind of internal architecture choice that limits the damage to some simpler case is not necessarily wrong. We've obviously done things like that for page table isolation etc. It can be horribly expensive, but sometimes that is the price you are willing to pay - and sometimes you can avoid the expense entirely by knowing how the particular hardware works.

Have the same virtual address on all processors to reach the local per cpu instance of a variable simplifies kernel code and avoids preempt_enable/disable sections. Especially operations on counters in the VM become much more efficient.

X86 can accomplish that with a segment override. That is not available on other platforms. Page tables can be used to do the proper mapping so that kernel code does not have to theses address calculations. These are frequent operations in the kernel memory management.

Quoted 8 lines

Very specific example: on 32-bit x86 with PAE, the top-level page table directory is fundamentally per-cpu in nature. It's just four entries that get loaded into the CPU on page table load, but it means that you can do things like having one quarter of the page tables be per-cpu, and the rest be shared. Several other architectures have similar "regions". So it can be an architectural choice to have some kind of per-cpu TLB. But I absolutely do *not* want to see a model where the general MM layer dictates that design.

X86 creates highly optimized code for per cpu operations without preempt enable/disable sections and I hope its ok to use per cpu page table to get to a similar degress of compactness and performance on ARM64.

Quoted 8 lines

And there are historical reasons to really hate it. Plan-9 had a completely broken threading model that depended on per-thread page tables. It was literal garbage. It was broken beyond words exactly due to coherency issues, and I still do not understand how otherwise very smart people ever came up with such a stupid idea.

We are not making _that_ mistake. We have made enough other ones to not neet to shoot ourselves in the head that badly.

We certainly do not want to replicate that approach. We are using the per cpu page tables to avoid address calculations in the VM that other platforms can do with a segment override.

#34 lore raw
Quoted 4 lines · Christoph Lameter (Ampere)

We certainly do not want to replicate that approach. We are using the per cpu page tables to avoid address calculations in the VM that other platforms can do with a segment override.

Right, and I said that that's ok as long as it's a internal architecture thing. Not a "this is how the VM works".

Because people really have wanted to make it a "this is how the VM works" thing. And that is very much what I object to.

Linus

#35 lore raw
Quoted 8 lines · Linus Torvalds
Quoted 4 lines · Christoph Lameter (Ampere)

We certainly do not want to replicate that approach. We are using the per cpu page tables to avoid address calculations in the VM that other platforms can do with a segment override.

Right, and I said that that's ok as long as it's a internal architecture thing. Not a "this is how the VM works".

Just for completeness: I claim that things like "kernel text replication" are a completely different beast that we should not even put into the same bag as what is being discussed here ("using the per cpu page tables to avoid address calculations in the VM that other platforms can do with a segment override"), in some limited internal architecture-specific thing.

(I assume some people here don't realize how tricky kernel text replication actually is)

Signature · 3 lines
Cheers,

David
#36 lore raw
In reply to Linus Torvalds
Quoted 3 lines · Linus Torvalds

Because people really have wanted to make it a "this is how the VM works" thing. And that is very much what I object to.

Side note: it's not that I object to small helper functions for some common case that a couple of architectures take (ie arm64 and s390).

But I do object to what looks like some kind of "this is how architectures should do it". Because I don't think that's the proper path at all. This is an area where there are lots of commonalities, but there are also cases where an architecture simply doesn't need it at all (x86 segments are an example, but so is any architected base register) or might take a completely different approach simply because of how the architecture does TLB fills (ie different areas may have simply different TLB walking: the pattern of "top few bits of the virtual address picks between different "segments" - different from x86 segments - is not that unusual afaik).

And I think David is right that there are different forms of this all, where that kernel text is another special case where some particular architectural choice might mean that it's trivially cheap to do *that* entirely differently (typically because you know there are no writes, and no aliasing issues).

So I'm just trying to clarify my stance - I don't hate percpu TLB contents per se. I hate the notion of codifying them.

For example, if you have a SW-filled TLB - which is admittedly a horribly broken idea even though it feels so clean and nice and has been done multiple times - that changes how everything like this looks for that architecture.

Linus

#37 lore raw
Quoted 2 lines · Linus Torvalds

So I'm just trying to clarify my stance - I don't hate percpu TLB contents per se. I hate the notion of codifying them.

I do worry that changing per-CPU data structures to support this is exactly codifying this (and will encourage other arches to follow suit).

And I really don't think there's any way to implement this without fundamentally changing core mm or assumptions that core mm can make.

--
Cheers, Lorenzo

#38 lore raw
Quoted 9 lines · Lorenzo Stoakes (ARM)
Quoted 2 lines · Linus Torvalds

So I'm just trying to clarify my stance - I don't hate percpu TLB contents per se. I hate the notion of codifying them.

I do worry that changing per-CPU data structures to support this is exactly codifying this (and will encourage other arches to follow suit).

And I really don't think there's any way to implement this without fundamentally changing core mm or assumptions that core mm can make.

Would it be possible to discuss the technical issues instead of throwing up this and that foggy objection? Lets not get into the complexities of restarting a critical section from an interrupt handler that you are proposing but lets just look at the code generated for this_cpu_* operations:

The fundamental problem is that your patchset does

<preeamble establishing restart point>
RMV per cpu op
<post handling cleaning up setup of restart point>

What x86 does and what we can get ARM to do is:

RMW per cpu op

Why do you think that your solution that requires a complex restart mechanism is better?

These are operations that are key to the performance in critical parts of memory management and this code is replicated by being inlined thousands of times in the kernel.

#39 lore raw
Quoted 3 lines

These are operations that are key to the performance in critical parts of memory management and this code is replicated by being inlined thousands of times in the kernel.

Cool, then it is easy to show the difference by running micro + macro benchmarks.

Signature · 3 lines
Cheers,

David
#40 lore raw
Quoted 5 lines · David Hildenbrand (Arm)
Quoted 3 lines

These are operations that are key to the performance in critical parts of memory management and this code is replicated by being inlined thousands of times in the kernel.

Cool, then it is easy to show the difference by running micro + macro benchmarks.

And we have done so repeatedly. Look at Yang's posts.

#41 lore raw
Quoted 10 lines · Christoph Lameter (Ampere)
Quoted 5 lines · David Hildenbrand (Arm)
Quoted 3 lines

These are operations that are key to the performance in critical parts of memory management and this code is replicated by being inlined thousands of times in the kernel.

Cool, then it is easy to show the difference by running micro + macro benchmarks.

And we have done so repeatedly. Look at Yang's posts.

I'm getting tired of this.

So to be clear: claiming that the additional instructions are unacceptable I'd expect something minimal like: "removing these 3 instructions speeds up micro-benchmark (a, b, c) by X% and macro-benchmarks (d, e, f) by Y% compared to Mark's approach". [1]

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

Signature · 3 lines
Cheers,

David
#42 lore raw
In reply to Christoph Lameter (Ampere)
Quoted 6 lines · Christoph Lameter (Ampere)
Quoted 4 lines · Lorenzo Stoakes (ARM)
Quoted 2 lines · Linus Torvalds

So I'm just trying to clarify my stance - I don't hate percpu TLB contents per se. I hate the notion of codifying them.

Quoted 8 lines
Quoted 5 lines

I do worry that changing per-CPU data structures to support this is exactly codifying this (and will encourage other arches to follow suit).

And I really don't think there's any way to implement this without fundamentally changing core mm or assumptions that core mm can make.

Would it be possible to discuss the technical issues instead of throwing up this and that foggy objection? Lets not get into the complexities of

Christoph this is very unhelpful. I and other maintainers have raised _specific_ and _technical_ objections. If you are having comprehension issues it'd be more constructive to ask questions.

It's really simple - convince us or the series doesn't go upstream.

Quoted 11 lines

The fundamental problem is that your patchset does

<preeamble establishing restart point>
RMV per cpu op
<post handling cleaning up setup of restart point>

What x86 does and what we can get ARM to do is:

Quoted 8 lines

RMW per cpu op
Why do you think that your solution that requires a complex restart mechanism is better?

These are operations that are key to the performance in critical parts of memory management and this code is replicated by being inlined thousands of times in the kernel.

I mean this is just silly :)

Provide data to back any of this, and no your benchmarks of your approach vs. nothing isn't it, nor do they demonstrate anything about what you claim here.

You've been asked for this multiple times now.

(And please no hand waving about 'critical parts of mm'. Be specific please.)

Quoted 2 lines

--
Cheers, Lorenzo

#43 lore raw
In reply to Linus Torvalds
Quoted 11 lines · Linus Torvalds
Quoted 4 lines · Christoph Lameter (Ampere)

We certainly do not want to replicate that approach. We are using the per cpu page tables to avoid address calculations in the VM that other platforms can do with a segment override.

Right, and I said that that's ok as long as it's a internal architecture thing. Not a "this is how the VM works".

Because people really have wanted to make it a "this is how the VM works" thing. And that is very much what I object to.

Well :) it fundamentally changes how mm works - now we have a whole new set of kernel page tables and PGD's that we have to think about.

And the series either tries to hack changes that fundamentally alter how vmalloc works or will add its own duplicative kernel page table code to do the same kind of thing.

So now core mm has to worry about TLB coherency and synchronisation, how this might interact with things like page table isolation, and a lot of other headaches (and I'm not really convinced they've been thought through here).

IOW - breaking fundamental assumptions about kernel page tables is inherently a whole-VM change (See [0] for instance for an example of what can go wrong).

If Christoph + Yang can provide compelling data that advocates for this change AND the core mm and arm64 communities agree to move forward with this, then things are different.

However, I think the best approach here is to find the least invasive way of limiting the change to the architecture itself.

Mark has put forward an approach that eliminates the exact overhead that this series aims to address but does so without having to fundamentally alter core mm assumptions (v2 posted at [1]).

So that seems very clearly to be a better alternative to me.

Ultimately we need mm and arm64 maintainer agreement on the way forwards (or in the case of a truly isolated arm64 solution, arm64 maintainer agreement).

The LSF session was not encouraging ([2]) and as an mm maintainer I'm not at all happy with this myself, nor does the arm64 community seem all that enthused so I would against gently suggest to Christoph to be a little more open-minded about the the proposed alternative which he seems to have dismissed out of hand.

Quoted 2 lines

Linus

--
Cheers, Lorenzo

[0]:https://lore.kernel.org/linux-mm/[email protected]/ [1]:https://lore.kernel.org/linux-arm-kernel/[email protected]/ [2]:https://lwn.net/Articles/1073395/

#44 lore raw
Quoted 3 lines · Lorenzo Stoakes (ARM)

Well :) it fundamentally changes how mm works - now we have a whole new set of kernel page tables and PGD's that we have to think about.

Yes, I'm certainly not enamored with that patch series under discussion.

That said, I only scanned it quickly, and I'm also admittedly less involved in the MM world than I historically was, so I don't think you should read *too* much into my opinions.

Linus

author
#45 lore raw
In reply to Lorenzo Stoakes (ARM)
Quoted 8 lines · Lorenzo Stoakes (ARM)
Quoted 7 lines · Linus Torvalds
Quoted 3 lines · Christoph Lameter (Ampere)

We certainly do not want to replicate that approach. We are using the per cpu page tables to avoid address calculations in the VM that other platforms can do with a segment override.

Right, and I said that that's ok as long as it's a internal architecture thing. Not a "this is how the VM works".

Quoted 8 lines
Quoted 2 lines

Because people really have wanted to make it a "this is how the VM works" thing. And that is very much what I object to.

Well :) it fundamentally changes how mm works - now we have a whole new set of kernel page tables and PGD's that we have to think about.

And the series either tries to hack changes that fundamentally alter how vmalloc works or will add its own duplicative kernel page table code to do the same kind of thing.

I don't think this series fundamentally alter how vmalloc works IMHO. vmalloc still works in the same way and the page tables share the same contents (except local percpu area), it just needs to sync up kernel page tables when the top level is changed. This is not new, x86 has been doing so for years, see arch_sync_kernel_mappings(). The patch just added the implementation for arm64.

Quoted 4 lines

So now core mm has to worry about TLB coherency and synchronisation, how this might interact with things like page table isolation, and a lot of other headaches (and I'm not really convinced they've been thought through here).

I don't think this series changed how TLB is sync'ed either. When we invalidate TLB entries for kernel (i.e. vmalloc area), the TLBI is sent to all CPUs, it is unchanged.

The series does change how percpu area is setup for ARM64, but this is already sort of architecture dependent.

Quoted 7 lines

IOW - breaking fundamental assumptions about kernel page tables is inherently a whole-VM change (See [0] for instance for an example of what can go wrong).

If Christoph + Yang can provide compelling data that advocates for this change AND the core mm and arm64 communities agree to move forward with this, then things are different.

We shared some benchmark data in the cover letter. Kernel build is a quite common benchmark used by kernel developers. Specjbb and MySQL benchmark results are considered important benchmarks as well. I wish I could run benchmark on more ARM64 cpus from other vendors, but I do have limited access to other hardware.

Quoted 10 lines

However, I think the best approach here is to find the least invasive way of limiting the change to the architecture itself.

Mark has put forward an approach that eliminates the exact overhead that this series aims to address but does so without having to fundamentally alter core mm assumptions (v2 posted at [1]).

So that seems very clearly to be a better alternative to me.

Quoted 7 lines

Ultimately we need mm and arm64 maintainer agreement on the way forwards (or in the case of a truly isolated arm64 solution, arm64 maintainer agreement).

The LSF session was not encouraging ([2]) and as an mm maintainer I'm not at all happy with this myself, nor does the arm64 community seem all that enthused so I would against gently suggest to Christoph to be a little more open-minded about the the proposed alternative which he seems to have dismissed out of hand.

Nobody disagreed Mark's approach also achieved good result, we just thought the series can do better. I agree there is always trade off between gain and cost, this is why we are seeking more usecases although there are different opinions.

Thanks,
Yang

Quoted 8 lines
Quoted 2 lines

Linus

--
Cheers, Lorenzo

[0]:https://lore.kernel.org/linux-mm/[email protected]/ [1]:https://lore.kernel.org/linux-arm-kernel/[email protected]/ [2]:https://lwn.net/Articles/1073395/

#46 lore raw
Quoted 7 lines
Quoted 4 lines

The LSF session was not encouraging ([2]) and as an mm maintainer I'm not at all happy with this myself, nor does the arm64 community seem all that enthused so I would against gently suggest to Christoph to be a little more open-minded about the the proposed alternative which he seems to have dismissed out of hand.

Nobody disagreed Mark's approach also achieved good result, we just thought the series can do better.

Okay, you thought it can do better. Did you only think that or did you measure it?

The cover letter obviously only compares to upstream Linux, not to Mark's series. Are there performance results I am missing that directly show that the additional instructions have an actual measurable performance impact?

That's really all we are asking for. If there is a measurable performance difference, great. Share the results or point us at them if we missed them.

Signature · 3 lines
Cheers,

David
#47 lore raw
Quoted 3 lines

That's really all we are asking for. If there is a measurable performance difference, great. Share the results or point us at them if we missed them.

Scanning over the patchset: It seems that you are using a special instruction that is unfamiliar to me. Which processors support that?

#48 lore raw
Quoted 6 lines · Christoph Lameter (Ampere)
Quoted 3 lines

That's really all we are asking for. If there is a measurable performance difference, great. Share the results or point us at them if we missed them.

Scanning over the patchset: It seems that you are using a special instruction that is unfamiliar to me. Which processors support that?

That's best asked in reply to Mark's patch set I think.

Signature · 3 lines
Cheers,

David
reviewer
#49 lore raw
In reply to Christoph Lameter (Ampere)
Quoted 6 lines · Christoph Lameter (Ampere)
Quoted 3 lines

That's really all we are asking for. If there is a measurable performance difference, great. Share the results or point us at them if we missed them.

Scanning over the patchset: It seems that you are using a special instruction that is unfamiliar to me. Which processors support that?

Which patch and which instruction are you referring to? There is no information to go on here.

Asking on that specific patch would be helpful.

Mark.

#50 lore raw
In reply to Yang Shi
Quoted 20 lines · Yang Shi
Quoted 8 lines · Lorenzo Stoakes (ARM)
Quoted 7 lines · Linus Torvalds
Quoted 3 lines · Christoph Lameter (Ampere)

We certainly do not want to replicate that approach. We are using the per cpu page tables to avoid address calculations in the VM that other platforms can do with a segment override.

Right, and I said that that's ok as long as it's a internal architecture thing. Not a "this is how the VM works".

Quoted 8 lines
Quoted 2 lines

Because people really have wanted to make it a "this is how the VM works" thing. And that is very much what I object to.

Well :) it fundamentally changes how mm works - now we have a whole new set of kernel page tables and PGD's that we have to think about.

And the series either tries to hack changes that fundamentally alter how vmalloc works or will add its own duplicative kernel page table code to do the same kind of thing.

Quoted 6 lines

I don't think this series fundamentally alter how vmalloc works IMHO. vmalloc still works in the same way and the page tables share the same contents (except local percpu area), it just needs to sync up kernel page tables when the top level is changed. This is not new, x86 has been doing so for years, see arch_sync_kernel_mappings(). The patch just added the implementation for arm64.

Your patch fundamentally changes how vmalloc works (!init_mm) as I pointed out in the patch.

Your response was essentially 'OK I'll duplicate the code' which was not really encouraging.

So either you have a variant of vmalloc that works on a brand new type of kernel page table mapping, or you have _yet more_ duplicative kernel page table code, explicitly the kind of code that has caused latent races and issues that I've very recently had to address.

People have repeatedly treated the ordinary or garden variety kernel page tables as 'never go away, nobody else touches' and rolled their own walking/manipulation code which I'm now having to fix up, so I'm really not in favour of going down that road again.

For instance, there's a contract that kernel mapping manipulation and race resolution _has_ to happen on init_mm.mmap_lock. I'm not sure exactly how that will interact with the per CPU page table concept.

arch_sync_kernel_mappings() and friends are different - that's about synchronising kernel mappings across process PGDs, you're proposing an _entirely distinct_ set of kernel page tables.

Core mm really does have to worry about this stuff even if only one arch implements them.

Quoted 9 lines
Quoted 5 lines

So now core mm has to worry about TLB coherency and synchronisation, how this might interact with things like page table isolation, and a lot of other headaches (and I'm not really convinced they've been thought through here).

I don't think this series changed how TLB is sync'ed either. When we invalidate TLB entries for kernel (i.e. vmalloc area), the TLBI is sent to all CPUs, it is unchanged.

The v2 change log which was a laundry list of 'I didn't think about X' (which is understandable - this stuff is hard) really doesn't fill me with confidence that you've thought everything through.

And inevitably you have to worry about TLB synchronisation of these page tables, as you're essentially inventing a new class of page tables that have different semantics from everything else.

Quoted 3 lines

The series does change how percpu area is setup for ARM64, but this is already sort of architecture dependent.

The diffstat disagrees with you :) you are changing core percpu and mm for this.

Quoted 9 lines
Quoted 8 lines

IOW - breaking fundamental assumptions about kernel page tables is inherently a whole-VM change (See [0] for instance for an example of what can go wrong).

If Christoph + Yang can provide compelling data that advocates for this change AND the core mm and arm64 communities agree to move forward with this, then things are different.

Quoted 5 lines

We shared some benchmark data in the cover letter. Kernel build is a quite common benchmark used by kernel developers. Specjbb and MySQL benchmark results are considered important benchmarks as well. I wish I could run benchmark on more ARM64 cpus from other vendors, but I do have limited access to other hardware.

I understand that and saw them (thanks for that). But can you also run them against Mark's series?

Quoted 19 lines
Quoted 11 lines

However, I think the best approach here is to find the least invasive way of limiting the change to the architecture itself.

Mark has put forward an approach that eliminates the exact overhead that this series aims to address but does so without having to fundamentally alter core mm assumptions (v2 posted at [1]).

So that seems very clearly to be a better alternative to me.

Quoted 7 lines

Ultimately we need mm and arm64 maintainer agreement on the way forwards (or in the case of a truly isolated arm64 solution, arm64 maintainer agreement).

The LSF session was not encouraging ([2]) and as an mm maintainer I'm not at all happy with this myself, nor does the arm64 community seem all that enthused so I would against gently suggest to Christoph to be a little more open-minded about the the proposed alternative which he seems to have dismissed out of hand.

Quoted 4 lines

Nobody disagreed Mark's approach also achieved good result, we just thought the series can do better. I agree there is always trade off between gain and cost, this is why we are seeking more usecases although there are different opinions.

I'm glad you raised tradeoffs, absolutely agree :)

Your proposal incurs high costs in terms of maintenance headaches, so it's not 'which is theoretically quicker based on instruction count' (a poor metric anyway), but rather an assessment of overall RoI.

And every series must be considered on its own merits, nothing is guaranteed to be taken.

Let's see how the alternative stacks up based on data please.

Quoted 13 lines

Thanks,
Yang

Quoted 8 lines
Quoted 2 lines

Linus

--
Cheers, Lorenzo

[0]:https://lore.kernel.org/linux-mm/[email protected]/ [1]:https://lore.kernel.org/linux-arm-kernel/[email protected]/ [2]:https://lwn.net/Articles/1073395/

--
Cheers, Lorenzo

maintainer
#51 lore raw
In reply to Lorenzo Stoakes (ARM)
Quoted 34 lines · Lorenzo Stoakes (ARM)
Quoted 11 lines · Linus Torvalds
Quoted 4 lines · Christoph Lameter (Ampere)

We certainly do not want to replicate that approach. We are using the per cpu page tables to avoid address calculations in the VM that other platforms can do with a segment override.

Right, and I said that that's ok as long as it's a internal architecture thing. Not a "this is how the VM works".

Because people really have wanted to make it a "this is how the VM works" thing. And that is very much what I object to.

Well :) it fundamentally changes how mm works - now we have a whole new set of kernel page tables and PGD's that we have to think about.

And the series either tries to hack changes that fundamentally alter how vmalloc works or will add its own duplicative kernel page table code to do the same kind of thing.

So now core mm has to worry about TLB coherency and synchronisation, how this might interact with things like page table isolation, and a lot of other headaches (and I'm not really convinced they've been thought through here).

IOW - breaking fundamental assumptions about kernel page tables is inherently a whole-VM change (See [0] for instance for an example of what can go wrong).

If Christoph + Yang can provide compelling data that advocates for this change AND the core mm and arm64 communities agree to move forward with this, then things are different.

However, I think the best approach here is to find the least invasive way of limiting the change to the architecture itself.

Quoted 8 lines

Mark has put forward an approach that eliminates the exact overhead that this series aims to address but does so without having to fundamentally alter core mm assumptions (v2 posted at [1]).

So that seems very clearly to be a better alternative to me.

Ultimately we need mm and arm64 maintainer agreement on the way forwards (or in the case of a truly isolated arm64 solution, arm64 maintainer agreement).

So the nice thing about having two implementations (i.e. the per-cpu page-tables *and* the preemption stuff from Mark) is that we can pitch them against each other to help us make a decision.

However, I don't see how anybody could argue that Mark's series isn't cleaner and easier to maintain. I've seen it described as "hacky" but I can't tell whether or not that's supposed to be a criticism. If we seriously want to consider the per-cpu page-table approach on arm64, the numbers need to be _really_ good and across a variety of hardware, not just the stuff with a memory system made of baling twine. That extends to the kernel text replication efforts too.

Will

#52 lore raw
Quoted 3 lines · Will Deacon

So the nice thing about having two implementations (i.e. the per-cpu page-tables *and* the preemption stuff from Mark) is that we can pitch them against each other to help us make a decision.

Yes exactly :)

This point has been put to the submitters a number of times so I hope that an additional repetition from an arm64 maintainer helps underline it.

Quoted 3 lines

However, I don't see how anybody could argue that Mark's series isn't cleaner and easier to maintain.

I completely agree.

Quoted 2 lines

I've seen it described as "hacky" but I
can't tell whether or not that's supposed to be a criticism. If we

Allow me to translate: 'I want to ignore this' :)

Similarly the comment about mm maintainers wanting maintainable not-broken code being a 'foggy' position. Translation: 'I want to ignore this'.

Since maintainers decide what gets merged the translation gets flipped if the submitters fail to address the feedback.

Let's hope sanity prevails.

Quoted 4 lines

seriously want to consider the per-cpu page-table approach on arm64, the numbers need to be _really_ good and across a variety of hardware, not just the stuff with a memory system made of baling twine. That extends to the kernel text replication efforts too.

Thanks, and equally so for the core mm changes that are required by the page table approach (that requirement being attested to by the diffstat).

Quoted 2 lines

Will

--
Cheers, Lorenzo

#53 lore raw
Quoted 7 lines · Lorenzo Stoakes (ARM)
Quoted 3 lines · Will Deacon

So the nice thing about having two implementations (i.e. the per-cpu page-tables *and* the preemption stuff from Mark) is that we can pitch them against each other to help us make a decision.

Yes exactly :)

Quoted 8 lines

This point has been put to the submitters a number of times so I hope that an additional repetition from an arm64 maintainer helps underline it.

Quoted 3 lines

However, I don't see how anybody could argue that Mark's series isn't cleaner and easier to maintain.

I completely agree.

Changing a register used by code sometimes in a timer interrupt is cleaner than accessing a virtually mapped address?

A simple RMV instruction is simpler and cleaner than prefixing a code sequence with flags on how to modify registers if something happens like a timer interrupt.

Quoted 1 line

Let's hope sanity prevails.

I sure hope so. But there is the ingroup behavior that rationalizes the wildest approaches as sane.

Quoted 7 lines
Quoted 4 lines

seriously want to consider the per-cpu page-table approach on arm64, the numbers need to be _really_ good and across a variety of hardware, not just the stuff with a memory system made of baling twine. That extends to the kernel text replication efforts too.

Thanks, and equally so for the core mm changes that are required by the page table approach (that requirement being attested to by the diffstat).

The numbers are already approaching 20% and will go much higher because there is more potential here with the ability to localize various memory accesses in the future.

Scaling to large core counts requires the reduction of traffic on the fabric connecting the cores. And the per cpu page tables allow that to happen and scale to high core counts while minimizing traffc.

If you keep on with your hacky approach of changing registers in the tick then we wont be getting there. We may have entirely new classes of bugs because now the timer interrupts changes some registers. Think of the hacking potential. Wow.

Please stop the madness.

#54 lore raw
In reply to Lorenzo Stoakes (ARM)
Quoted 6 lines · Lorenzo Stoakes (ARM)

Since this work seems to be very much arm64-focused, perhaps it's therefore worth looking at an alterative solution that's specific to the arch, like the one suggested by Mark ([1])?

[0]:https://lore.kernel.org/all/CAHk-=wire3dzhHx=KiL_f5Rj0=1u9ustsa33QoR-F9-v-NU9Ng@mail.gmail.com/ [1]:https://lore.kernel.org/linux-arm-kernel/al_DpFJFcmVhxpvW@J2N7QTR9R3/

Mark's solution does replace the preempt_enable/disable sections with a rather hacky restart logic. It relies on a long preemable and postscript to each per cpu operations.

It does not remove the code entirely like the proposed patchset here and
cannot produce the same compact code as x86.

#55 lore raw
Quoted 6 lines · Christoph Lameter (Ampere)
Quoted 4 lines · Lorenzo Stoakes (ARM)

Since this work seems to be very much arm64-focused, perhaps it's therefore worth looking at an alterative solution that's specific to the arch, like the one suggested by Mark ([1])?

Quoted 6 lines
Quoted 2 lines

[0]:https://lore.kernel.org/all/CAHk-=wire3dzhHx=KiL_f5Rj0=1u9ustsa33QoR-F9-v-NU9Ng@mail.gmail.com/ [1]:https://lore.kernel.org/linux-arm-kernel/al_DpFJFcmVhxpvW@J2N7QTR9R3/

Mark's solution does replace the preempt_enable/disable sections with a rather hacky restart logic. It relies on a long preemable and postscript to each per cpu operations.

Okay, so 3 simple instructions of preemable is "long preemable"? In which universe?

But I am sure you did you homework and have data to back up your claims. Please share that data, because I am very curious.

Quoted 3 lines

It does not remove the code entirely like the proposed patchset here and
cannot produce the same compact code as x86.

Mark's solution is the obvious improvement to the problem, doing it just like s390 already does.

If the proposed solution has a significant benefit over Mark's solution it could be discussed as a next step; and it must be backed with sufficient data to warrant all this complexity here.

Signature · 3 lines
Cheers,

David
#56 lore raw
Quoted 5 lines · David Hildenbrand (Arm)
Quoted 2 lines

rather hacky restart logic. It relies on a long preemable and postscript to each per cpu operations.

But I am sure you did you homework and have data to back up your claims. Please share that data, because I am very curious.

The comparison here is a single instruction on one side and abnd the per cpu operation prefixed with a preeamble and another instruction / set of instructions on the other side.

Even if its only one before and one after then we are still looking at 3 instructions instead of one. This is a frequently used operation in the VM.

You need data to show that one instruction is faster than 3???

Quoted 2 lines

Mark's solution is the obvious improvement to the problem, doing it just like s390 already does.

Well there is interest by the S390 folks to move to what we proposed from what I can tell.

Quoted 3 lines

If the proposed solution has a significant benefit over Mark's solution it could be discussed as a next step; and it must be backed with sufficient data to warrant all this complexity here.

You cannot accept that one instruction is faster than the same instruction prefixed with another and something else at its tail?

#57 lore raw
Quoted 12 lines · Christoph Lameter (Ampere)
Quoted 5 lines · David Hildenbrand (Arm)
Quoted 2 lines

rather hacky restart logic. It relies on a long preemable and postscript to each per cpu operations.

But I am sure you did you homework and have data to back up your claims. Please share that data, because I am very curious.

The comparison here is a single instruction on one side and abnd the per cpu operation prefixed with a preeamble and another instruction / set of instructions on the other side.

Quoted 5 lines

Even if its only one before and one after then we are still looking at 3 instructions instead of one. This is a frequently used operation in the VM.

You need data to show that one instruction is faster than 3???

I want you to show that the difference *matters* in *practice* enough that we should *care* about accepting the additional *complexity*.

No hand-waving about "3 > 1". IIRC people were asking that previously but received no results back.

So to be clear: claiming that the additional instructions are unacceptable I'd expect something minimal like: "removing these 3 instructions speeds up micro-benchmark (a, b, c) by X% and macro-benchmarks (d, e, f) by Y% compared to Mark's approach".

Maybe there really is a big difference.

Quoted 6 lines
Quoted 3 lines

Mark's solution is the obvious improvement to the problem, doing it just like s390 already does.

Well there is interest by the S390 folks to move to what we proposed from what I can tell.

Heiko said that as reply to v1, bit he's been replying a lot on Mark's patches. Only Heiko can tell. (on CC)

Quoted 8 lines
Quoted 5 lines

If the proposed solution has a significant benefit over Mark's solution it could be discussed as a next step; and it must be backed with sufficient data to warrant all this complexity here.

You cannot accept that one instruction is faster than the same instruction prefixed with another and something else at its tail?

I am pretty sure that you got exactly what I meant, you likely just don't have the data to back it up.

Proof me wrong!

Signature · 3 lines
Cheers,

David
#58 lore raw
Quoted 11 lines · David Hildenbrand (Arm)
Quoted 6 lines
Quoted 3 lines

If the proposed solution has a significant benefit over Mark's solution it could be discussed as a next step; and it must be backed with sufficient data to warrant all this complexity here.

You cannot accept that one instruction is faster than the same instruction prefixed with another and something else at its tail?

I am pretty sure that you got exactly what I meant, you likely just don't have the data to back it up.

Proof me wrong!

Well you seem to have alrady seen the benchmarks that we ran and are running. So I had doubts on your sincerety with this comment. Seemes to be be just there for effect.

#59 lore raw
Quoted 6 lines · Christoph Lameter (Ampere)
Quoted 4 lines · David Hildenbrand (Arm)
Quoted 3 lines

You cannot accept that one instruction is faster than the same instruction prefixed with another and something else at its tail?

Quoted 8 lines
Quoted 4 lines

I am pretty sure that you got exactly what I meant, you likely just don't have the data to back it up.

Proof me wrong!

Well you seem to have alrady seen the benchmarks that we ran and are running. So I had doubts on your sincerety with this comment. Seemes to be be just there for effect.

Let me repeat a third time:

So to be clear: claiming that the additional instructions are unacceptable I'd expect something minimal like: "removing these 3 instructions speeds up micro-benchmark (a, b, c) by X% and macro-benchmarks (d, e, f) by Y% compared to Mark's approach". [1]

[1] https://lore.kernel.org/r/[email protected]

Signature · 3 lines
Cheers,

David
#60 lore raw
In reply to David Hildenbrand (Arm)
Quoted 10 lines · David Hildenbrand (Arm)
Quoted 6 lines · Christoph Lameter (Ampere)
Quoted 2 lines · David Hildenbrand (Arm)

Mark's solution is the obvious improvement to the problem, doing it just like s390 already does.

Well there is interest by the S390 folks to move to what we proposed from what I can tell.

Heiko said that as reply to v1, bit he's been replying a lot on Mark's patches. Only Heiko can tell. (on CC)

I said that before Peter Zijlstra proposed the restartable sequence approach for the kernel. I took that as input, and provided the current solution for s390, which does not restart anything, but only fix up register contents if required. Now Mark provided an improved implementation for arm64.

That said, there are some performance numbers [1] for s390 available as posted by Mete. A performance improvement is indeed there, but looking at the current s390 implementation I'm not sure there will be much of a benefit if we now go to percpu page tables.

Currently we have this code sequence for e.g. this_cpu_add(...) (%r2 contains the to be added value).

larl    %r4,1b33300     <-- load address of percpu var
mviy    960,4           <-- mark start of percpu op section
ag      %r4,952         <-- add percpu offset
laag    %r5,%r2,0(%r4)  <-- atomic add
mviy    960,0           <-- mark end of percpu op section

With the proposed percpu page tables I would guess / hope we would end up with something like this:

larl    %r4,1b33300     <-- load address of percpu var (same on all cpus)
laag    %r5,%r2,0(%r4)  <-- atomic add

That's certainly better to what we have compared to now (no performance numbers available), but I'm not sure if only this would justify all the added complexity to common code currently discussed.

Read: I'm more or less fine with the current s390 solution, and not pushing
for percpu page tables.

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

#61 lore raw
In reply to David Hildenbrand (Arm)
Quoted 10 lines · On Tue, 4 Aug 2026 18:47:26 +0200 "David Hildenbrand (Arm)" <[email protected]>…
Quoted 9 lines · Christoph Lameter (Ampere)
Quoted 6 lines · Lorenzo Stoakes (ARM)

Since this work seems to be very much arm64-focused, perhaps it's therefore worth looking at an alterative solution that's specific to the arch, like the one suggested by Mark ([1])?

[0]:https://lore.kernel.org/all/CAHk-=wire3dzhHx=KiL_f5Rj0=1u9ustsa33QoR-F9-v-NU9Ng@mail.gmail.com/ [1]:https://lore.kernel.org/linux-arm-kernel/al_DpFJFcmVhxpvW@J2N7QTR9R3/

Quoted 8 lines
Quoted 3 lines

Mark's solution does replace the preempt_enable/disable sections with a rather hacky restart logic. It relies on a long preemable and postscript to each per cpu operations.

Okay, so 3 simple instructions of preemable is "long preemable"? In which universe?

But I am sure you did you homework and have data to back up your claims. Please share that data, because I am very curious.

The proposed sequence is:

Quoted 6 lines
          // Prologue. Enable fixups for <off> and <addr>.
 1        mrs	<tsk>, sp_el0
 2        mov	<tmp>, #__VAL_PCPU_GPRS(<pcp>, <off>, <addr>)
 3        strh	<tmp>, [<tsk>, #TSK_TI_PCPU_GPRS]
          // Generate cpu-specific address
Quoted 8 lines

4 mrs <off>, TPIDR_ELx
5 add <addr>, <pcp>, <off>

          // Perform access sequence

6 ldr <val>, [<addr>]

          // Epilogue. Disable fixups

7 strh wzr, [<tsk>, #TSK_TI_PCPU_GPRS]

Think about how that actually gets execute by a real cpu.
Instructions will be read from the I-cache in 'chunks' (maybe half a cache line). They are then fed to multiple decoders that generate u-ops for the execution units. The decoder is unlikely to be a bottleneck.
I've numbered the instructions:
First clock can run instructions 1, 2 and 4.
Assuming the mrs have no extra latency the second runs 3 and 5.
The third will then run 6 and 7.
The cpu then probably has to wait for the result of the ldr.

If the access is a write then there may be a stall waiting for the value to be written to be available.

The only real effect of the extra instructions is likely to be code size.

David

reviewer
#62 lore raw
In reply to Christoph Lameter (Ampere)
Quoted 11 lines · Christoph Lameter (Ampere)
Quoted 6 lines · Mark Rutland

I expect that should come with a reasonable benefit, but I don't have benchmark figures yet as I haven't finished converting the xchg and cmpxchg implementations.

Quoted 2 lines

It sounds like it just moved the cost from one place to the other place and it also seems hacky TBH.

Yang Shi's patch has *no* critical section. There is no additional code for the RMV instruction. The RMV instruction is executed on the correct per cpu area.

None of that was in question.

Quoted 4 lines

One of the reasons for the performance win is the
eliminattion of these critical sections. Your approach still has some form of prologue and posthandling like the current preempt approach and therefore will not be able to have the same performance gains.

In absolute terms, yes.

However, I'm fairly confident that the vast majority of the overhead we have today can be eliminated with simpler alternatives.

There is a trade-off, and there are surprisingly complex interactions between page tables and other things (e.g. entry code). There is risk and maintenance burden associated with that. Hence people want to understand how much of the benefit is attributable to what. So far, the statements haven't convinced me people actually know what portion of the overhead come from which factor, e.g.

* How much of that attributable to conditional work when re-enabling preemption?

* How much of that is attributable to RMW sequences to modify the preempt count itself?

* How much of that is attributable to system register accesses (SP_EL0 and TPIDR_ELx)?

Any of those could easily dominate the other factors and might easily be avoidable. Most of that should be measurable today. For example you could restore the preempt_{enable,disable} calls atop Yang Shi's patches.

Quoted 2 lines

The code is more efficient, there is no restart necessary and the technique is already widely used on x86 for a long time.

It's true that the per-cpu page table approach will have fewer instructions in the fast path.

However, the other statements here are potentially misleading:

(1) There is no restart in the scheme I have proposed, so restarting is irrelevant to the comparison.

(2) On x86, this_cpu*() operations use segment relative addressing, NOT per-cpu page tables. If arm64 had a similar addressing scheme, I expect we would use it.

(3) There are a number of novel problems associated with per-cpu page tables (e.g the various unsolved issues Yang has described), which do not apply to x86's implementation of this_cpu_*() operations.

Quoted 6 lines

Having the ability in general to map mmemory differently depending on the cpu opens up a number of other optimization like

What we are proposing here is a basic new feature that simplifies code and allows addititonal performance and functional features that are so far not possible on ARM64.

While this simplifies the this_cpu_*() operations, I don't believe this is a simplification overall, and IMO, describing it as such is misleading.

Mark.

#63 lore raw

Re: [RFC v2 PATCH 0/16] Optimize this_cpu_*() ops for non-x86 (ARM64 for this series)~

In reply to Yang Shi
Quoted 2 lines · Yang Shi

S390 implemented restart sequence. Heiko didn't share too much benchmark data, just confirmed there is no regression and I had a lengthy discussion with him (https://lore.kernel.org/ lkml/[email protected]/).

Hi,

regarding performance numbers from s390 and Heiko's implementation. It is safe to say we see noticeable improvements on micro benchmarks but nothing really pops out on regular workloads.

on 24 Cores - 48 CPUs

1. Kernel Build
---------------------------------------------------------------------
time make -j $(nproc)
~0-2% improvement - within noise ratio

2. stress-ng vm ops
---------------------------------------------------------------------
stress-ng --vm $(nproc) --vm-bytes 128M --vm-ops 100000000 ~16% improvement on bogo/ops

3. stress-ng vm ops + fork
---------------------------------------------------------------------
stress-ng -M --mmapfork $(nproc) --mmapfork-bytes 128M --mmapfork-ops 10000 ~6% improvement on bogo/ops

4. fork test
---------------------------------------------------------------------
stress-ng -M --fork $(nproc) --fork-ops 10000000
~13% improvement on bogo/ops

Basecommit: a2b94afeb516 ("s390/zcrypt: Replace get_zeroed_page() with kzalloc()")

Thanks,
-Mete Durlu

#64 lore raw
In reply to Yang Shi
Quoted 151 lines · Yang Shi

Hi,

This is v2 RFC. In v2 a lot problems found out by Sashiko were fixed and more feature gaps were closed (please see the below changelog for the details). Although there are still some open issues, for example, it just can support 48 bits VA (for 4K and 64K) and 47 bits VA (for 16K), KPTI support has not been solved yet, etc, but I think the delta should be big enough and worth a new RFC to gather comments in order to make sure I'm on the right track.

Some more benchmarks were done, for example, some latency related benchmarks that I mentioned at LSFMM because I thought responsiveness should be improved due to the removal of preempt_disable. Collected more PMU counters as well. Please refer to the benchmark section for more details.

Look forward to comments.

Changelog
v2: * Added 3-level and 2-level page table support.
* Tested with 16K and 64K page size. But we just support 48 bits VA (4K and 64K) and 47 bits VA with 16K for now. Please refer to the below "known issue" section for the detail reason.
* Added support for memory hotplug.
* Added support for KASAN (generic).
* Treated percpu and local percpu area address as vmalloc address.
* Fixed build failure for x86.
* Fixed build failure for !CONFIG_NUMA.
* Added KASAN support for local percpu area.
* Some other misc bug fixes found out by Sashiko.
* More code refactor and cleanup.
* Regorganized the patches.
* More benchmarks, refer to benchmark section for more details.
* Rebased to v7.2-rc1.

Introduction
============
This patch series implemented the LSFMM 2026 proposal for optimizing this_cpu_*() ops on ARM64. For the details of the proposal, Please refer to:
https://lore.kernel.org/linux-mm/CAHbLzkpcN-T8MH6=W3jCxcFj1gVZp8fRqe231yzZT-rV_E_org@mail.gmail.com/
I didn't repeat it in the cover letter because there is no change to the proposal.

The series is based on 7.1-rc1. It is basically minimum viable patches. There are still a few hacks in this series and it may break something, for example, KPTI, SMT machines which shared TLB, etc. But it shoule be good enough for now to demonstrate the core idea. The main purpose of the RFC is to gather feedback, figure out missing parts and risks, and make sure we are on the right track, as well as hopefully it can help the discussion for the upcoming LSFMM.

I broke the patches down to arch-dependent and arch-independent parts so that hopefully the interested persons can do experiments on other architectures, for example, S390, easier.

A new kernel config is introduced, HAVE_LOCAL_PER_CPU_MAP. The architectures which can support this feature will select it. Allocating and freeing percpu local mapping is protected by this config so that others won't pay the cost.

Known Issues
============
1. KPTI
-------
We need determine what CPU we are on, then switch to the right page table. Currently arm64 kernel fetches tramp_pg_dir via swapper_pg_dir - fixed_offset, and fetches swapper_pg_dir from ttbr1. But ttbr1 may not hold swapper_pg_dir anymore except CPU #0. So we need to figure out the other way to handle it. Switching to tramp_pg_dir should be easy, but the reverse seems harder because tramp_pg_dir just maps the trampoline vectors.
Maybe we can do two steps switch. Switch to swapper_pg_dir at the first step, then switch to per cpu page table (for entry) or tramp page table (for exit). Nobody should call this_cpu_*() at either userspace -> kernel entry stage or kernel -> userspace exit stage.

2. SW PAN
---------
Has the similar issue as KPTI. It installs reserved_pg_dir to TTBR0 when running in kernel space, but fetching reserved_pg_dir via swapper_pg_dir - fixed_offset. Maybe we can save the physical address of swapper_pg_dir in a variable, then load it from that variable instead of ttbr1.

3. Shared TLB machines
----------------------
Some machines may share TLB between CPUs, for example, SMT machines may share TLB between the two hardware threads in one core.
The per cpu page table just can't work with it. Maybe we need a new cpufeature to indicate whether per cpu page table is allowed or not. Then just enable it for not-shared-TLB machines.

4. Don't support all VA bits
----------------------------
We just support 48 bits VA (4K and 64K) and 47 bits VA (16K) for now. For 4K and 64K, supporting other VA bits is not hard, we just need to determine the size for percpu and local percpu area.
But it is harder for supporting 48 bits VA + 16K page size. We just have two top level kernel page table entries with this configuration, but we assume we just need to sync up kernel page table at the top level for now. We need to sync up kernel page table at the second level in order to support it. I'm not sure whether it is worth it or not.

Benchmark
=========
The benchmarks are done on 160 core AmpereOne machine. The baseline is v7.2-rc1 kernel.

1. Reduction of kernel text size
--------------------------------
The patchset can reduce at least 11 instructions for this_cpu_*() ops. Both preempt_disable() and preempt_enable() need 4 instructions to manipulate the preempt count, and preempt_enable() needs more instructions (compare + READ + compare) to determine whether reschedule is needed or not. Because this_cpu_*() ops are inlined and called in a lot of places so we can save a lot of instructions.

The size of kernel text is reduced by ~184KB with default Fedora kernel config. This also helps reduce kernel icache miss rate and stalled frontend cycles as kernel build benchmark result showed.

2. Kernel Build
---------------
Run kernel build (make -j160) with the default Fedora kernel config in a memcg.
13% - 18% sys time improvment
3% - 7% wall time improvement

5% fewer kernel icache miss, 5% fewer executed kernel instructions and 15% fewer stalled frontend cycles for kernel.

3. stress-ng vm ops
-------------------
stress-ng --vm 160 --vm-bytes 128M --vm-ops 100000000 8.5% improvement

4. stress-ng vm ops + fork
--------------------------
stress-ng --mmapfork 160 --mmapfork-bytes 128M --mmapfork-ops 500 15% improvement

5. Specjbb
----------
The specjbb test latency curves showed the patched kernel has consistently lower p99 latency (the lower the better) than the baseline.

2.5% improvement on max-jOPS and 4% - 5% improvement on critical-jOPS. The specjbb benchmark is quite sensitive to latency and responsiveness, particularly critical-jOPS result. The patches are supposed to improve the responsiveness due to the reduction of preempt-disabled critical sections.

Quoted 4 lines

6. MySQL
--------
1% - 2% gains on read-only test, 2% - 4% gains on write-only test. Also see 15% decrease on frontend cache stall.

tl;dr
Comparing GPR approach [1] with this series gives 8 improvements and 3 regressions.

Fastpath is a Linux kernel performance benchmarking service. The table compares the both patched kernels: positive values are faster, negative values are slower, and (I)/(R) indicate statistically significant improvements/regressions. Unmarked differences are not significant after accounting for confidence and noise thresholds.

+---------------------------------+--------------------+-----------------+-----------------------+
| Benchmark                       | percpu-pgtable [2] |   gpr-fixup [3] |         gpr-fixup [3] |
|                                 |    vs baseline [1] | vs baseline [1] | vs percpu-pgtable [2] |
+=================================+====================+=================+=======================+
| lmbench/lat-mem-rd              |             -0.26% |       (I) 1.24% |             (I) 1.50% |
| micromm/fork                    |          (I) 1.10% |       (I) 2.46% |             (I) 1.35% |
| micromm/munmap                  |         (I) 18.10% |      (I) 19.78% |             (I) 1.43% |
| micromm/vmalloc                 |         (I) 11.92% |      (I) 14.68% |             (I) 2.47% |
| mmtests/hackbench               |              0.27% |       (I) 1.09% |                 0.81% |
| mmtests/kernbench               |          (I) 1.19% |       (I) 1.18% |                -0.01% |
| mmtests/sysbench-cpu            |              0.01% |           0.06% |                 0.05% |
| mmtests/sysbench-mutex          |             -0.85% |          -0.14% |                 0.72% |
| mmtests/sysbench-thread         |         (R) -4.49% |       (I) 3.02% |             (I) 7.86% |
| perf/futex                      |              0.55% |      (R) -2.80% |            (R) -3.34% |
| perf/sched                      |          (I) 1.61% |          -0.39% |            (R) -1.97% |
| perf/syscall                    |          (I) 1.54% |           0.98% |                -0.55% |
| pts/memtier-benchmark           |          (I) 2.37% |       (I) 2.51% |                 0.14% |
| pts/nginx                       |              0.42% |           0.96% |                 0.54% |
| pts/perl-benchmark              |          (I) 1.83% |       (I) 1.63% |                -0.20% |
| pts/pgbench                     |             -0.09% |           0.55% |                 0.65% |
| pts/pybench                     |             -0.05% |          -0.07% |                -0.02% |
| pts/redis                       |             -0.00% |          -0.04% |                -0.04% |
| pts/sqlite-speedtest            |              0.15% |           0.44% |                 0.28% |
| repro-collection/mysql-workload |              0.73% |           0.25% |                -0.48% |
| schbench/thread-contention      |              0.42% |       (I) 1.04% |                 0.61% |
| sockperf/echo-lat-tcp           |          (I) 2.88% |       (I) 1.69% |            (R) -1.16% |
| sockperf/echo-lat-udp           |          (I) 3.93% |       (I) 5.11% |             (I) 1.14% |
| sockperf/packet-tp-tcp          |             -0.41% |       (I) 1.37% |             (I) 1.79% |
| sockperf/packet-tp-udp          |          (I) 1.45% |       (I) 1.49% |                 0.04% |
| specjbb/composite               |          (I) 1.39% |       (I) 1.65% |                 0.26% |
| speedometer/v2.0                |         (R) -1.04% |      (R) -1.04% |                 0.00% |
| speedometer/v2.1                |             -0.10% |           0.10% |                 0.20% |
| syscall/getpid                  |             -0.98% |           0.52% |             (I) 1.51% |
| syscall/getppid                 |             -0.43% |           0.14% |                 0.57% |
| syscall/invalid                 |          (I) 3.06% |       (I) 3.56% |                 0.48% |
+---------------------------------+--------------------+-----------------+-----------------------+

[1] v7.2
[2] v7.2-percpu-pgtable (this series)
[3] v7.2-percpu-gpr [1] (Based on email review, we fixed SDEI to restore x26 instead of corrupting x22, and adjusted this_cpu_write() helpers to avoid GCC overflow warnings.)

The constraints/workarounds for the per-CPU page-table series were supplied as this Kconfig fragment for all the different runs.

CONFIG_EXPERT=y
CONFIG_ARM64_4K_PAGES=y
CONFIG_ARM64_VA_BITS_48=y
CONFIG_ARM64_VA_BITS_39=n
CONFIG_ARM64_VA_BITS_52=n
CONFIG_ARM64_PA_BITS_48=y
CONFIG_ARM64_PA_BITS_52=n
CONFIG_UNMAP_KERNEL_AT_EL0=n
CONFIG_ARM64_SW_TTBR0_PAN=n
CONFIG_KASAN=n

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

Quoted 83 lines

Regression test
===============
1. memcg creation
-----------------
Create 10K memcgs. Each memcg creation needs to allocate multiple percpu variables, for example, percpu refcnt, rstat and objcg percpu refcnt.

Consumed 2112K more virtual memory for percpu “local mapping” and a few more mega bytes consumed by per cpu page tables.
No noticeable regression was found for elapsed time.

2. fork test
------------
stress-ng --fork 160 --fork-ops 10000000
fork() needs to allocate multiple percpu variables, for example, rss counters and mm_cid_cpu.

Roughly 1% regression was found. However stress-ng fork test has quites small address space, the real life workloads typically have much larger address space and do more complicated works. The stress-ng mmapfork benchmark saw 15% improvement.

The organization of patches
===========================
The refactor and prepatory patches (patch 1 - patch 4) Percpu page table support patches (patch 5 - patch 8) Local percpu area support patches (patch 7 - patch 15) Use local percpu area for this_cpu ops (patch 16)

Yang Shi (16):
drivers: arch_numa: move percpu set up code to arch
arm64: kconfig: make percpu related configs not depend on NUMA
mm: pgalloc: introduce {pud|pmd}_populate_sync()
vmalloc: pass in pgd pointer for vmap{__vunmap}_range_noflush()
arm64: mm: enable percpu kernel page table
arm64: mm: defined {pud|pmd}_populate_sync()
arm64: mm: sync percpu page table for memory hotplug/unplug
arm64: kasan: sync up kasan shadow area page table
arm64: mm: define percpu virtual space area
mm: percpu: prepare to use dedicated percpu area
arm64: mm: map local percpu first chunk
mm: percpu: set up first chunk and reserve chunk
arm64: mm: introduce __per_cpu_local_off
mm: percpu: allocate and free local percpu vm area
arm64: kconfig: select HAVE_LOCAL_PER_CPU_MAP
arm64: percpu: use local percpu for this_cpu_*() APIs

 arch/arm64/Kconfig                   |  12 +++++++---
 arch/arm64/include/asm/mmu.h         |   5 ++++
 arch/arm64/include/asm/mmu_context.h |   9 +++++++-
 arch/arm64/include/asm/percpu.h      |  37 ++++++++++++++++++++++++++++-
 arch/arm64/include/asm/pgalloc.h     |  24 +++++++++++++++++++
 arch/arm64/include/asm/pgtable.h     |  37 ++++++++++++++++++++++++++---
 arch/arm64/kernel/setup.c            |   3 +++
 arch/arm64/kernel/smp.c              |  44 +++++++++++++++++++++++++++++++++++
 arch/arm64/mm/kasan_init.c           |  47 +++++++++++++++++++++++--------------
 arch/arm64/mm/mmu.c                  | 165 ++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++----------------------
 arch/arm64/mm/ptdump.c               |   4 ++++
 arch/riscv/kernel/smp.c              |  51 ++++++++++++++++++++++++++++++++++++++++
 drivers/base/arch_numa.c             |  51 +---------------------------------------
 include/linux/mm.h                   |  11 +++++++++
 include/linux/percpu.h               |   4 +++-
 include/linux/pgalloc.h              |  13 +++++++++++
 include/linux/vmalloc.h              |   3 +++
 mm/Kconfig                           |   9 ++++++++
 mm/internal.h                        |   5 +++-
 mm/kmsan/hooks.c                     |  14 +++++------
 mm/percpu-internal.h                 |  14 +++++++++++
 mm/percpu-vm.c                       |  94 ++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
 mm/percpu.c                          |  58 +++++++++++++++++++++++++++++++++++++---------
 mm/sparse-vmemmap.c                  |   4 ++--
 mm/vmalloc.c                         | 138 +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++---------------
 25 files changed, 712 insertions(+), 144 deletions(-)

Thanks,
Yang

author
#65 lore raw
Quoted 221 lines · Usama Anjum
Quoted 150 lines · Yang Shi

Hi,

This is v2 RFC. In v2 a lot problems found out by Sashiko were fixed and more feature gaps were closed (please see the below changelog for the details). Although there are still some open issues, for example, it just can support 48 bits VA (for 4K and 64K) and 47 bits VA (for 16K), KPTI support has not been solved yet, etc, but I think the delta should be big enough and worth a new RFC to gather comments in order to make sure I'm on the right track.

Some more benchmarks were done, for example, some latency related benchmarks that I mentioned at LSFMM because I thought responsiveness should be improved due to the removal of preempt_disable. Collected more PMU counters as well. Please refer to the benchmark section for more details.

Look forward to comments.

Changelog
v2: * Added 3-level and 2-level page table support.
* Tested with 16K and 64K page size. But we just support 48 bits VA (4K and 64K) and 47 bits VA with 16K for now. Please refer to the below "known issue" section for the detail reason.
* Added support for memory hotplug.
* Added support for KASAN (generic).
* Treated percpu and local percpu area address as vmalloc address.
* Fixed build failure for x86.
* Fixed build failure for !CONFIG_NUMA.
* Added KASAN support for local percpu area.
* Some other misc bug fixes found out by Sashiko.
* More code refactor and cleanup.
* Regorganized the patches.
* More benchmarks, refer to benchmark section for more details.
* Rebased to v7.2-rc1.

Introduction
============
This patch series implemented the LSFMM 2026 proposal for optimizing this_cpu_*() ops on ARM64. For the details of the proposal, Please refer to:
https://lore.kernel.org/linux-mm/CAHbLzkpcN-T8MH6=W3jCxcFj1gVZp8fRqe231yzZT-rV_E_org@mail.gmail.com/
I didn't repeat it in the cover letter because there is no change to the proposal.

The series is based on 7.1-rc1. It is basically minimum viable patches. There are still a few hacks in this series and it may break something, for example, KPTI, SMT machines which shared TLB, etc. But it shoule be good enough for now to demonstrate the core idea. The main purpose of the RFC is to gather feedback, figure out missing parts and risks, and make sure we are on the right track, as well as hopefully it can help the discussion for the upcoming LSFMM.

I broke the patches down to arch-dependent and arch-independent parts so that hopefully the interested persons can do experiments on other architectures, for example, S390, easier.

A new kernel config is introduced, HAVE_LOCAL_PER_CPU_MAP. The architectures which can support this feature will select it. Allocating and freeing percpu local mapping is protected by this config so that others won't pay the cost.

Known Issues
============
1. KPTI
-------
We need determine what CPU we are on, then switch to the right page table. Currently arm64 kernel fetches tramp_pg_dir via swapper_pg_dir - fixed_offset, and fetches swapper_pg_dir from ttbr1. But ttbr1 may not hold swapper_pg_dir anymore except CPU #0. So we need to figure out the other way to handle it. Switching to tramp_pg_dir should be easy, but the reverse seems harder because tramp_pg_dir just maps the trampoline vectors.
Maybe we can do two steps switch. Switch to swapper_pg_dir at the first step, then switch to per cpu page table (for entry) or tramp page table (for exit). Nobody should call this_cpu_*() at either userspace -> kernel entry stage or kernel -> userspace exit stage.

2. SW PAN
---------
Has the similar issue as KPTI. It installs reserved_pg_dir to TTBR0 when running in kernel space, but fetching reserved_pg_dir via swapper_pg_dir - fixed_offset. Maybe we can save the physical address of swapper_pg_dir in a variable, then load it from that variable instead of ttbr1.

3. Shared TLB machines
----------------------
Some machines may share TLB between CPUs, for example, SMT machines may share TLB between the two hardware threads in one core.
The per cpu page table just can't work with it. Maybe we need a new cpufeature to indicate whether per cpu page table is allowed or not. Then just enable it for not-shared-TLB machines.

4. Don't support all VA bits
----------------------------
We just support 48 bits VA (4K and 64K) and 47 bits VA (16K) for now. For 4K and 64K, supporting other VA bits is not hard, we just need to determine the size for percpu and local percpu area.
But it is harder for supporting 48 bits VA + 16K page size. We just have two top level kernel page table entries with this configuration, but we assume we just need to sync up kernel page table at the top level for now. We need to sync up kernel page table at the second level in order to support it. I'm not sure whether it is worth it or not.

Benchmark
=========
The benchmarks are done on 160 core AmpereOne machine. The baseline is v7.2-rc1 kernel.

1. Reduction of kernel text size
--------------------------------
The patchset can reduce at least 11 instructions for this_cpu_*() ops. Both preempt_disable() and preempt_enable() need 4 instructions to manipulate the preempt count, and preempt_enable() needs more instructions (compare + READ + compare) to determine whether reschedule is needed or not. Because this_cpu_*() ops are inlined and called in a lot of places so we can save a lot of instructions.

The size of kernel text is reduced by ~184KB with default Fedora kernel config. This also helps reduce kernel icache miss rate and stalled frontend cycles as kernel build benchmark result showed.

2. Kernel Build
---------------
Run kernel build (make -j160) with the default Fedora kernel config in a memcg.
13% - 18% sys time improvment
3% - 7% wall time improvement

5% fewer kernel icache miss, 5% fewer executed kernel instructions and 15% fewer stalled frontend cycles for kernel.

3. stress-ng vm ops
-------------------
stress-ng --vm 160 --vm-bytes 128M --vm-ops 100000000 8.5% improvement

4. stress-ng vm ops + fork
--------------------------
stress-ng --mmapfork 160 --mmapfork-bytes 128M --mmapfork-ops 500 15% improvement

5. Specjbb
----------
The specjbb test latency curves showed the patched kernel has consistently lower p99 latency (the lower the better) than the baseline.

2.5% improvement on max-jOPS and 4% - 5% improvement on critical-jOPS. The specjbb benchmark is quite sensitive to latency and responsiveness, particularly critical-jOPS result. The patches are supposed to improve the responsiveness due to the reduction of preempt-disabled critical sections.

Quoted 4 lines

6. MySQL
--------
1% - 2% gains on read-only test, 2% - 4% gains on write-only test. Also see 15% decrease on frontend cache stall.

tl;dr
Comparing GPR approach [1] with this series gives 8 improvements and 3 regressions.

Fastpath is a Linux kernel performance benchmarking service. The table compares the both patched kernels: positive values are faster, negative values are slower, and (I)/(R) indicate statistically significant improvements/regressions. Unmarked differences are not significant after accounting for confidence and noise thresholds.

+---------------------------------+--------------------+-----------------+-----------------------+
| Benchmark                       | percpu-pgtable [2] |   gpr-fixup [3] |         gpr-fixup [3] |
|                                 |    vs baseline [1] | vs baseline [1] | vs percpu-pgtable [2] |
+=================================+====================+=================+=======================+
| lmbench/lat-mem-rd              |             -0.26% |       (I) 1.24% |             (I) 1.50% |
| micromm/fork                    |          (I) 1.10% |       (I) 2.46% |             (I) 1.35% |
| micromm/munmap                  |         (I) 18.10% |      (I) 19.78% |             (I) 1.43% |
| micromm/vmalloc                 |         (I) 11.92% |      (I) 14.68% |             (I) 2.47% |
| mmtests/hackbench               |              0.27% |       (I) 1.09% |                 0.81% |
| mmtests/kernbench               |          (I) 1.19% |       (I) 1.18% |                -0.01% |
| mmtests/sysbench-cpu            |              0.01% |           0.06% |                 0.05% |
| mmtests/sysbench-mutex          |             -0.85% |          -0.14% |                 0.72% |
| mmtests/sysbench-thread         |         (R) -4.49% |       (I) 3.02% |             (I) 7.86% |
| perf/futex                      |              0.55% |      (R) -2.80% |            (R) -3.34% |
| perf/sched                      |          (I) 1.61% |          -0.39% |            (R) -1.97% |
| perf/syscall                    |          (I) 1.54% |           0.98% |                -0.55% |
| pts/memtier-benchmark           |          (I) 2.37% |       (I) 2.51% |                 0.14% |
| pts/nginx                       |              0.42% |           0.96% |                 0.54% |
| pts/perl-benchmark              |          (I) 1.83% |       (I) 1.63% |                -0.20% |
| pts/pgbench                     |             -0.09% |           0.55% |                 0.65% |
| pts/pybench                     |             -0.05% |          -0.07% |                -0.02% |
| pts/redis                       |             -0.00% |          -0.04% |                -0.04% |
| pts/sqlite-speedtest            |              0.15% |           0.44% |                 0.28% |
| repro-collection/mysql-workload |              0.73% |           0.25% |                -0.48% |
| schbench/thread-contention      |              0.42% |       (I) 1.04% |                 0.61% |
| sockperf/echo-lat-tcp           |          (I) 2.88% |       (I) 1.69% |            (R) -1.16% |
| sockperf/echo-lat-udp           |          (I) 3.93% |       (I) 5.11% |             (I) 1.14% |
| sockperf/packet-tp-tcp          |             -0.41% |       (I) 1.37% |             (I) 1.79% |
| sockperf/packet-tp-udp          |          (I) 1.45% |       (I) 1.49% |                 0.04% |
| specjbb/composite               |          (I) 1.39% |       (I) 1.65% |                 0.26% |
| speedometer/v2.0                |         (R) -1.04% |      (R) -1.04% |                 0.00% |
| speedometer/v2.1                |             -0.10% |           0.10% |                 0.20% |
| syscall/getpid                  |             -0.98% |           0.52% |             (I) 1.51% |
| syscall/getppid                 |             -0.43% |           0.14% |                 0.57% |
| syscall/invalid                 |          (I) 3.06% |       (I) 3.56% |                 0.48% |
+---------------------------------+--------------------+-----------------+-----------------------+

[1] v7.2
[2] v7.2-percpu-pgtable (this series)
[3] v7.2-percpu-gpr [1] (Based on email review, we fixed SDEI to restore x26 instead of corrupting x22, and adjusted this_cpu_write() helpers to avoid GCC overflow warnings.)

The constraints/workarounds for the per-CPU page-table series were supplied as this Kconfig fragment for all the different runs.

CONFIG_EXPERT=y
CONFIG_ARM64_4K_PAGES=y
CONFIG_ARM64_VA_BITS_48=y
CONFIG_ARM64_VA_BITS_39=n
CONFIG_ARM64_VA_BITS_52=n
CONFIG_ARM64_PA_BITS_48=y
CONFIG_ARM64_PA_BITS_52=n
CONFIG_UNMAP_KERNEL_AT_EL0=n
CONFIG_ARM64_SW_TTBR0_PAN=n
CONFIG_KASAN=n

Quoted 1 line

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

Hi Usama,

Thank you for running the benchmarks. I'm done some benchmarks and wrapped up the results too. I did run some same benchmarks as yours, like perf bench and kernel build. I also did some profiling for some benchmarks and did some investigation about where the performance difference came from. This is why the work took longer than I expected in the first place.

Test Environment
===========
Hardware
--------
1P AmpereOne machine with 192 cores, 512GB memory.

Kernel
------
v7.2 based.
I applied the first 6 patches from Mark's v2 series, which are some common micro optimization IIRC, on top of my percpu page table series. I'm not sure whether they can make measurable difference or not, but I think I'd better do it because we were benchmarking some subtle difference (4 instructions difference in some hot path) so we can have some apple to apple comparison.

4K page size + 48 VA bits.

Result Analysis
-------------
I use ministat (https://github.com/leahneukirchen/ministat), which is a small tool to do the statistics legwork on benchmarks, to process the benchmark results. It should help us rule out the misleading noises.

The ministat output looks like:
    N           Min           Max        Median           Avg Stddev x   10         46998         48366         48192     47986.333  499.14714 +   10         48438         49333         49030     48952.333  304.42711 Difference at 95.0% confidence
    966 +/- 531.791
    2.01307% +/- 1.10821%
    (Student's t, pooled s = 413.415)

I use the percpu page table series as baseline, which is shown as "x" row in ministat output. The "+" row is Mark's series. The column "N" means the number of samples (basically equals to the number of run iterations). Other columns are quite self-explained. But I will just show the percentage reported by ministat in this email.

Benchmarks
========
I have done some different benchmarks which cover from macro to micro benchmarks and different type of applications (kernel build, Java, database, networking, etc). They are:
  - Kernel text size (just compare the kernel text size for the two kernels)
  - Kernel build
  - Specjbb
  - MySQL (sysbench)
  - Memcached (via Mellanox NIC)
  - iperf (localhost)
  - perf bench
  - will-it-scale (page fault and mmap)

Because we are benchmarking some subtle code difference, so all the benchmarks were run multiple times (the number of iterations depend on benchmarks duration and variance) and a couple of reboots in order to rule out reboot by reboot spread (I did see some significant reboot by reboot spread for some test cases). I also bind the tests to some specific CPUs if they don't use all CPUs in order to minimize the core-to-core latency variance and scheduler interference.

Basically I didn't see measurable difference for the most benchmarks listed above (ministat showed no difference), I think this aligned with your results too. But some benchmarks showed regressions for Mark's series. Please see the below for the details (just the regressed benchmarks).

Kernel text size
-------------
With the same kernel configuration (the default Fedora config), the kernel text size for the percpu page table series is 92K smaller than Mark's series. This is kind of expected because all this_cpu ops are basically inlined and are called quite frequently, so even small difference could make big impact.

Kernel build
----------
Ran kernel build with "make -j192" in a memcg with cold page cache. "Drop cache" is called for each run before building kernel.

I did 5 runs. The percpu page table series has roughly 2% better sys time (the smaller the better). Difference at 95.0% confidence
    966 +/- 531.791
    2.01307% +/- 1.10821%
    (Student's t, pooled s = 413.415)

And slightly better wall time.
Difference at 95.0% confidence
    2.2 +/- 1.92934
    0.553877% +/- 0.485735%
    (Student's t, pooled s = 1.32288)

perf sched pipe
-------------
Bind two threads to CPU 78 and 79 (randomly picked) in order to avoid core-2-core latency variance and some scheduler overhead. I did 200 runs. The percpu page table series is better. I didn't see measurable difference with process mode.

Throughput (The bigger the better)
Difference at 95.0% confidence
    -24549.2 +/- 4900.32
    -9.73769% +/- 1.78314%
    (Student's t, pooled s = 25001.6)

Latency (The smaller the better)
Difference at 95.0% confidence
    0.377888 +/- 0.0753446
    9.37686% +/- 2.01494%
    (Student's t, pooled s = 0.384411)

perf futex wake-parallel (wake up one thread from 192 waiters)
------------------------------------------------------
The result shows the wake up latency. I did 1000 runs. It is the smaller the better. Difference at 95.0% confidence
    0.0009767 +/- 0.000134466
    17.0749% +/- 2.53593%
    (Student's t, pooled s = 0.00153406)

The histograms showed obvious long tail latency from Mark's series for both perf bench pipe and perf futex wake-parallel.

I also did some profiling for the above two benchmarks. The profiling showed the percpu page table series has lower overhead in schedule(). I got the profiling with pseudo NMI which has extra overhead, it may distort the profiling a little bit. The profiling showed finish_task_switch() as the hot spot. But it actually doesn't do too many things other than enabling IRQ. This is a typical misleading profiling due to not supported NMI on ARM64. But anyway both profiling results (w/ pseudo NMI and w/o it) show the overhead is in schedule().

perf bench fork
-------------
The result is the smaller the better. I did 20 runs. And the profiling showed the percpu page table series spent less time in __percpu_add_case_64(). Difference at 95.0% confidence
    48.5027 +/- 4.16055
    1.93944% +/- 0.167849%
    (Student's t, pooled s = 6.50039)

will-it-scale/page fault
--------------------
All the page fault micro benchmarks from will-it-scale were run with one single thread and bound to the same CPU in order to minimize the overhead from page allocator and all the VM locks (for example, lru lock, zone lock, etc). I did 100 runs for each test case. The result is the bigger the better.

page_fault1 (Anon page fault)
-------------------------
The percpu page table series is slightly better.
Difference at 95.0% confidence
    -13391.9 +/- 944.563
    -1.67893% +/- 0.117849%
    (Student's t, pooled s = 3407.69)

page fault2 (CoW page fault)
------------------------
The percpu page table series is slightly better.
Difference at 95.0% confidence
    -4736.52 +/- 438.185
    -1.16333% +/- 0.10686%
    (Student's t, pooled s = 1580.83)

page fault3 (tmpfs page fault)
-------------------------
The percpu page table series is slightly better.
Difference at 95.0% confidence
    -21667.9 +/- 1352.51
    -1.91305% +/- 0.1186%
    (Student's t, pooled s = 4879.44)

The profiling for all the page fault benchmarks showed the percpu page table series spent less time in __percpu_add_case_64().

will-it-scale/mmap
----------------
Like page fault tests, the mmap benchmarks were run with one single thread and bound to the same CPU in order to minimize the overhead from VM locks. I did 50 runs for each. The results are the bigger the better.

mmap1
------
The percpu page table series is better.
Difference at 95.0% confidence
    -15916.4 +/- 6094.58
    -3.53615% +/- 1.32214%
    (Student's t, pooled s = 15509.2)

mmap2
------
The percpu page table series is better.
Difference at 95.0% confidence
    -14662.4 +/- 9359.95
    -3.91005% +/- 2.44462%
    (Student's t, pooled s = 23588.6)

Thanks,
Yang

Quoted 81 lines
Quoted 81 lines

Regression test
===============
1. memcg creation
-----------------
Create 10K memcgs. Each memcg creation needs to allocate multiple percpu variables, for example, percpu refcnt, rstat and objcg percpu refcnt.

Consumed 2112K more virtual memory for percpu “local mapping” and a few more mega bytes consumed by per cpu page tables.
No noticeable regression was found for elapsed time.

2. fork test
------------
stress-ng --fork 160 --fork-ops 10000000
fork() needs to allocate multiple percpu variables, for example, rss counters and mm_cid_cpu.

Roughly 1% regression was found. However stress-ng fork test has quites small address space, the real life workloads typically have much larger address space and do more complicated works. The stress-ng mmapfork benchmark saw 15% improvement.

The organization of patches
===========================
The refactor and prepatory patches (patch 1 - patch 4) Percpu page table support patches (patch 5 - patch 8) Local percpu area support patches (patch 7 - patch 15) Use local percpu area for this_cpu ops (patch 16)

Yang Shi (16):
drivers: arch_numa: move percpu set up code to arch
arm64: kconfig: make percpu related configs not depend on NUMA
mm: pgalloc: introduce {pud|pmd}_populate_sync()
vmalloc: pass in pgd pointer for vmap{__vunmap}_range_noflush()
arm64: mm: enable percpu kernel page table
arm64: mm: defined {pud|pmd}_populate_sync()
arm64: mm: sync percpu page table for memory hotplug/unplug
arm64: kasan: sync up kasan shadow area page table
arm64: mm: define percpu virtual space area
mm: percpu: prepare to use dedicated percpu area
arm64: mm: map local percpu first chunk
mm: percpu: set up first chunk and reserve chunk
arm64: mm: introduce __per_cpu_local_off
mm: percpu: allocate and free local percpu vm area
arm64: kconfig: select HAVE_LOCAL_PER_CPU_MAP
arm64: percpu: use local percpu for this_cpu_*() APIs

  arch/arm64/Kconfig                   |  12 +++++++---
  arch/arm64/include/asm/mmu.h         |   5 ++++
  arch/arm64/include/asm/mmu_context.h |   9 +++++++-
  arch/arm64/include/asm/percpu.h      |  37 ++++++++++++++++++++++++++++-
  arch/arm64/include/asm/pgalloc.h     |  24 +++++++++++++++++++
  arch/arm64/include/asm/pgtable.h     |  37 ++++++++++++++++++++++++++---
  arch/arm64/kernel/setup.c            |   3 +++
  arch/arm64/kernel/smp.c              |  44 +++++++++++++++++++++++++++++++++++
  arch/arm64/mm/kasan_init.c           |  47 +++++++++++++++++++++++--------------
  arch/arm64/mm/mmu.c                  | 165 ++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++----------------------
  arch/arm64/mm/ptdump.c               |   4 ++++
  arch/riscv/kernel/smp.c              |  51 ++++++++++++++++++++++++++++++++++++++++
  drivers/base/arch_numa.c             |  51 +---------------------------------------
  include/linux/mm.h                   |  11 +++++++++
  include/linux/percpu.h               |   4 +++-
  include/linux/pgalloc.h              |  13 +++++++++++
  include/linux/vmalloc.h              |   3 +++
  mm/Kconfig                           |   9 ++++++++
  mm/internal.h                        |   5 +++-
  mm/kmsan/hooks.c                     |  14 +++++------
  mm/percpu-internal.h                 |  14 +++++++++++
  mm/percpu-vm.c                       |  94 ++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
  mm/percpu.c                          |  58 +++++++++++++++++++++++++++++++++++++---------
  mm/sparse-vmemmap.c                  |   4 ++--
  mm/vmalloc.c                         | 138 +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++---------------
  25 files changed, 712 insertions(+), 144 deletions(-)

Thanks,
Yang