Skip to contentSkip to conversation

mm: introduce reference pages

PATCH In review Peter Collingbourne 3 participants 3 messages Jul 17, 2021 – Jul 30, 2021

Subsystems (23)
lore
Share & tools
lore
#0 lore raw

PATCH v5 1/1 mm: introduce reference pages

Introduce a new syscall, refpage_create, which returns a file descriptor which may be mapped using mmap. Such a mapping is similar to an anonymous mapping, but instead of clean pages being backed by the zero page, they are instead backed by a so-called reference page, whose contents are specified using an argument to refpage_create. Loads from the mapping will load directly from the reference page, and initial stores to the mapping will copy-on-write from the reference page.

Reference pages are useful in circumstances where anonymous mappings combined with manual stores to memory would impose undesirable costs, either in terms of performance or RSS. Use cases are focused on heap allocators and include:

- Pattern initialization for the heap. This is where malloc(3) gives you memory whose contents are filled with a non-zero pattern byte, in order to help detect and mitigate bugs involving use of uninitialized memory. Typically this is implemented by having the allocator memset the allocation with the pattern byte before returning it to the user, but for large allocations this can result in a significant increase in RSS, especially for allocations that are used sparsely. Even for dense allocations there is a needless impact to startup performance when it may be better to amortize it throughout the program. By creating allocations using a reference page filled with the pattern byte, we can avoid these costs.

- Pre-tagged heap memory. Memory tagging [1] is an upcoming ARMv8.5 feature which allows for memory to be tagged in order to detect certain kinds of memory errors with low overhead. In order to set up an allocation to allow memory errors to be detected, the entire allocation needs to have the same tag. The issue here is similar to pattern initialization in the sense that large tagged allocations will be expensive if the tagging is done up front. The idea is that the allocator would create reference pages with each of the possible memory tags, and use those reference pages for the large allocations.

This patch includes specific optimizations for these use cases in order to reduce memory traffic. If the reference page consists of a single repeating byte, the page is initialized using memset, and on arm64 if the reference page consists of a uniformly tagged zero page, the DC GZVA instruction is used to initialize the page.

In order to measure the performance and RSS impact of reference pages, I used the following microbenchmark program, which is intended to compare an implementation of heap pattern initialization that uses memset to initialize the pages against an implementation that uses reference pages:

  #include <stdio.h>
  #include <stdlib.h>
  #include <string.h>
  #include <sys/mman.h>
  #include <unistd.h>

  constexpr unsigned char pattern_byte = 0xaa;

  #define PAGE_SIZE 4096

  _Alignas(PAGE_SIZE) static unsigned char pattern[PAGE_SIZE];

  int main(int argc, char **argv) {
    if (argc < 3)
      return 1;
    bool use_refpage = argc > 3;
    size_t mmap_size = atoi(argv[1]);
    size_t touch_size = atoi(argv[2]);
    int refpage_fd;
    if (use_refpage) {
      memset(pattern, pattern_byte, PAGE_SIZE);
      refpage_fd = syscall(447, pattern, 0);
    }
    for (unsigned i = 0; i != 1000; ++i) {
      char *p;
      if (use_refpage) {
        p = (char *)mmap(0, mmap_size, PROT_READ | PROT_WRITE, MAP_PRIVATE,
                         refpage_fd, 0);
      } else {
        p = (char *)mmap(0, mmap_size, PROT_READ | PROT_WRITE,
                         MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
        memset(p, pattern_byte, mmap_size);
      }
      for (unsigned j = 0; j < touch_size; j += PAGE_SIZE)
        p[j] = 0;
      munmap(p, mmap_size);
    }

}

On a DragonBoard 845c with the powersave governor, and taking the median of 10 runs for each measurement, I measured the following results for real time (s):

touch_size/mmap_size memset refpages improvement (95% CI)

      4096/4096000    3.962194   0.026726   98.8015% +/- 1.14684%
   2048000/4096000    3.925309   1.48081    61.8271% +/- 1.11911%
   4096000/4096000    3.986275   3.385003   15.1205% +/- 0.227235%

And the following for max RSS (KiB):

touch_size/mmap_size memset refpages improvement (95% CI)

      4096/4096000      6656      3448      49.3815% +/- 1.30339%
   2048000/4096000      6696      4580      31.7053% +/- 1.16411%
   4096000/4096000      6716      6684              none

So we see a large improvement for sparsely used allocations, and even a modest perf improvement for fully utilized allocations as a result of touching the pages one fewer time (with memset: once in the kernel and once in userspace; with refpages: just once in the kernel).

---
v5:
- rebased to 5.15rc2
- changed the API to take a size argument, including the prep hook
- made tag_clear_highpage a helper function
- renamed optzn -> content_type, optzn_info -> content_info
- outlined part of the is_refpage_vma check
- removed is_zero_or_refpage_pfn
- added a comment at the top of refpage.c
- restructured syscall handler to use goto
- inlined put_refpage_private_data into remaining caller

v4:
- rebased to linux-next
- added arch hooks to support MTE tagged reference pages
- added optimizations for pages with pattern byte as well as uniformly MTE-tagged pages
- added helper functions to avoid open-coding the reference page detection
- wrote a microbenchmark program and got new perf results for the commit message

As an alternative to introducing this syscall, I considered using userfaultfd to implement reference pages. However, after having taken a detailed look at the interface, it does not seem suitable to be used in the context of a general purpose allocator. For example, UFFD_FEATURE_FORK support would be required in order to correctly support fork(2) in a process that uses the allocator (although POSIX does not guarantee support for allocating after fork, many allocators including Scudo support it, and nothing stops the forked process from page faulting pre-existing allocations after forking anyway), but UFFD_FEATURE_FORK has been restricted to root by commit 3c1c24d91ffd ("userfaultfd: require CAP_SYS_PTRACE for UFFD_FEATURE_EVENT_FORK"), making it unsuitable for use in an allocator. Furthermore, even if the interface issues are resolved, I suspect (but have not measured) that the cost of the multiple context switches between kernel and userspace would be too high to be used in an allocator anyway.

 arch/alpha/kernel/syscalls/syscall.tbl      |   1 +
 arch/arm/tools/syscall.tbl                  |   1 +
 arch/arm64/include/asm/mman.h               |  16 +++
 arch/arm64/include/asm/mte.h                |  10 +-
 arch/arm64/include/asm/page.h               |   4 +-
 arch/arm64/include/asm/unistd.h             |   2 +-
 arch/arm64/include/asm/unistd32.h           |   2 +
 arch/arm64/kernel/mte.c                     |  24 ++++
 arch/arm64/lib/mte.S                        |  26 +++-
 arch/arm64/mm/fault.c                       |  48 ++++++-
 arch/ia64/kernel/syscalls/syscall.tbl       |   1 +
 arch/m68k/kernel/syscalls/syscall.tbl       |   1 +
 arch/microblaze/kernel/syscalls/syscall.tbl |   1 +
 arch/mips/kernel/syscalls/syscall_n32.tbl   |   1 +
 arch/mips/kernel/syscalls/syscall_n64.tbl   |   1 +
 arch/mips/kernel/syscalls/syscall_o32.tbl   |   1 +
 arch/parisc/kernel/syscalls/syscall.tbl     |   1 +
 arch/powerpc/kernel/syscalls/syscall.tbl    |   1 +
 arch/s390/kernel/syscalls/syscall.tbl       |   1 +
 arch/sh/kernel/syscalls/syscall.tbl         |   1 +
 arch/sparc/kernel/syscalls/syscall.tbl      |   1 +
 arch/x86/entry/syscalls/syscall_32.tbl      |   1 +
 arch/x86/entry/syscalls/syscall_64.tbl      |   1 +
 arch/xtensa/kernel/syscalls/syscall.tbl     |   1 +
 include/linux/gfp.h                         |  11 +-
 include/linux/highmem.h                     |   9 +-
 include/linux/huge_mm.h                     |   7 +
 include/linux/mm.h                          |  32 +++++
 include/linux/mman.h                        |  19 +++
 include/linux/syscalls.h                    |   3 +
 include/uapi/asm-generic/unistd.h           |   5 +-
 kernel/sys_ni.c                             |   1 +
 mm/Makefile                                 |   4 +-
 mm/gup.c                                    |   3 +-
 mm/memory.c                                 |  47 ++++--
 mm/migrate.c                                |   4 +-
 mm/refpage.c                                | 151 ++++++++++++++++++++
 37 files changed, 413 insertions(+), 31 deletions(-)
 create mode 100644 mm/refpage.c
arch/alpha/kernel/syscalls/syscall.tbl +1 -0
index a17687ed4b51..494edc5ca61c 100644
@@ -486,3 +486,4 @@
 554	common	landlock_create_ruleset		sys_landlock_create_ruleset
 555	common	landlock_add_rule		sys_landlock_add_rule
 556	common	landlock_restrict_self		sys_landlock_restrict_self
+558	common	refpage_create			sys_refpage_create
arch/arm/tools/syscall.tbl +1 -0
index c5df1179fc5d..8fd7045f46b9 100644
@@ -460,3 +460,4 @@
 444	common	landlock_create_ruleset		sys_landlock_create_ruleset
 445	common	landlock_add_rule		sys_landlock_add_rule
 446	common	landlock_restrict_self		sys_landlock_restrict_self
+448	common	refpage_create			sys_refpage_create
arch/arm64/include/asm/mman.h +16 -0
index e3e28f7daf62..018e3c8913ea 100644
@@ -84,4 +84,20 @@ static inline bool arch_validate_flags(unsigned long vm_flags)
 }
 #define arch_validate_flags(vm_flags) arch_validate_flags(vm_flags)
 
+struct refpage_private_data;
+
+void arch_prep_refpage_private_data(struct refpage_private_data *priv,
+				    char *content_kaddr, unsigned long size);
+#define arch_prep_refpage_private_data arch_prep_refpage_private_data
+
+static inline void arch_prep_refpage_vma(struct vm_area_struct *vma)
+{
+	vma->vm_flags |= VM_MTE_ALLOWED;
+}
+#define arch_prep_refpage_vma arch_prep_refpage_vma
+
+void arch_copy_refpage(struct page *page, unsigned long addr,
+				     struct vm_area_struct *vma);
+#define arch_copy_refpage arch_copy_refpage
+
 #endif /* ! __ASM_MMAN_H__ */
arch/arm64/include/asm/mte.h +8 -2
index 58c7f80f5596..e37bcb0e285e 100644
@@ -31,13 +31,14 @@ void mte_invalidate_tags(int type, pgoff_t offset);
 void mte_invalidate_tags_area(int type);
 void *mte_allocate_tag_storage(void);
 void mte_free_tag_storage(char *storage);
+void mte_memcpy(char *to, char *from, unsigned long size);
 
 #ifdef CONFIG_ARM64_MTE
 
 /* track which pages have valid allocation tags */
 #define PG_mte_tagged	PG_arch_2
 
-void mte_zero_clear_page_tags(void *addr);
+void mte_zero_set_page_tags(void *addr);
 void mte_sync_tags(pte_t old_pte, pte_t pte);
 void mte_copy_page_tags(void *kto, const void *kfrom);
 void mte_thread_init_user(void);
@@ -48,13 +49,14 @@ long set_mte_ctrl(struct task_struct *task, unsigned long arg);
 long get_mte_ctrl(struct task_struct *task);
 int mte_ptrace_copy_tags(struct task_struct *child, long request,
 			 unsigned long addr, unsigned long data);
+u8 mte_check_tag_zero_page(struct page *userpage);
 
 #else /* CONFIG_ARM64_MTE */
 
 /* unused if !CONFIG_ARM64_MTE, silence the compiler */
 #define PG_mte_tagged	0
 
-static inline void mte_zero_clear_page_tags(void *addr)
+static inline void mte_zero_set_page_tags(void *addr)
 {
 }
 static inline void mte_sync_tags(pte_t old_pte, pte_t pte)
@@ -89,6 +91,10 @@ static inline int mte_ptrace_copy_tags(struct task_struct *child,
 {
 	return -EIO;
 }
+static inline u8 mte_check_tag_zero_page(struct page *userpage)
+{
+	return 0;
+}
 
 #endif /* CONFIG_ARM64_MTE */
 
arch/arm64/include/asm/page.h +2 -2
index 993a27ea6f54..80e74d9ecaf0 100644
@@ -33,8 +33,8 @@ struct page *alloc_zeroed_user_highpage_movable(struct vm_area_struct *vma,
 						unsigned long vaddr);
 #define __HAVE_ARCH_ALLOC_ZEROED_USER_HIGHPAGE_MOVABLE
 
-void tag_clear_highpage(struct page *to);
-#define __HAVE_ARCH_TAG_CLEAR_HIGHPAGE
+void tag_set_highpage(struct page *to, unsigned long tag);
+#define __HAVE_ARCH_TAG_SET_HIGHPAGE
 
 #define clear_user_page(page, vaddr, pg)	clear_page(page)
 #define copy_user_page(to, from, vaddr, pg)	copy_page(to, from)
arch/arm64/include/asm/unistd.h +1 -1
index 727bfc3be99b..3cb206aea3db 100644
@@ -38,7 +38,7 @@
 #define __ARM_NR_compat_set_tls		(__ARM_NR_COMPAT_BASE + 5)
 #define __ARM_NR_COMPAT_END		(__ARM_NR_COMPAT_BASE + 0x800)
 
-#define __NR_compat_syscalls		447
+#define __NR_compat_syscalls		449
 #endif
 
 #define __ARCH_WANT_SYS_CLONE
arch/arm64/include/asm/unistd32.h +2 -0
index 99ffcafc736c..2a116aa17fe7 100644
@@ -901,6 +901,8 @@ __SYSCALL(__NR_landlock_create_ruleset, sys_landlock_create_ruleset)
 __SYSCALL(__NR_landlock_add_rule, sys_landlock_add_rule)
 #define __NR_landlock_restrict_self 446
 __SYSCALL(__NR_landlock_restrict_self, sys_landlock_restrict_self)
+#define __NR_refpage_create 448
+__SYSCALL(__NR_refpage_create, sys_refpage_create)
 
 /*
  * Please add new compat syscalls above this comment and update
arch/arm64/kernel/mte.c +24 -0
index 36f51b0e438a..876997703d7b 100644
@@ -450,3 +450,27 @@ int mte_ptrace_copy_tags(struct task_struct *child, long request,
 
 	return ret;
 }
+
+u8 mte_check_tag_zero_page(struct page *userpage)
+{
+	char *userpage_addr = page_address(userpage);
+	u8 tag;
+	int i;
+
+	if (!test_bit(PG_mte_tagged, &userpage->flags))
+		return 0;
+
+	tag = mte_get_mem_tag(userpage_addr) & 0xF;
+	if (tag == 0)
+		return 0;
+
+	for (i = 0; i != PAGE_SIZE; ++i)
+		if (userpage_addr[i] != 0)
+			return 0;
+
+	for (i = MTE_GRANULE_SIZE; i != PAGE_SIZE; i += MTE_GRANULE_SIZE)
+		if ((mte_get_mem_tag(userpage_addr + i) & 0xF) != tag)
+			return 0;
+
+	return tag;
+}
arch/arm64/lib/mte.S +22 -4
index e83643b3995f..47de7f3c0baf 100644
@@ -37,24 +37,23 @@ SYM_FUNC_START(mte_clear_page_tags)
 SYM_FUNC_END(mte_clear_page_tags)
 
 /*
- * Zero the page and tags at the same time
+ * Zero the page and set tags at the same time
  *
  * Parameters:
  *	x0 - address to the beginning of the page
  */
-SYM_FUNC_START(mte_zero_clear_page_tags)
+SYM_FUNC_START(mte_zero_set_page_tags)
 	mrs	x1, dczid_el0
 	and	w1, w1, #0xf
 	mov	x2, #4
 	lsl	x1, x2, x1
-	and	x0, x0, #(1 << MTE_TAG_SHIFT) - 1	// clear the tag
 
 1:	dc	gzva, x0
 	add	x0, x0, x1
 	tst	x0, #(PAGE_SIZE - 1)
 	b.ne	1b
 	ret
-SYM_FUNC_END(mte_zero_clear_page_tags)
+SYM_FUNC_END(mte_zero_set_page_tags)
 
 /*
  * Copy the tags from the source page to the destination one
@@ -169,3 +168,22 @@ SYM_FUNC_START(mte_restore_page_tags)
 
 	ret
 SYM_FUNC_END(mte_restore_page_tags)
+
+/*
+ * Copies memory and tags
+ *   x0 - destination address
+ *   x1 - source address
+ *   x2 - size
+ * Preconditions: size != 0, all arguments must be aligned to 16.
+ */
+SYM_FUNC_START(mte_memcpy)
+1:
+	ldg	x3, [x1]
+	stg	x3, [x0]
+	ldp	x3, x4, [x1], #16
+	stp	x3, x4, [x0], #16
+	sub	x2, x2, #16
+	cbnz	x2, 1b
+
+	ret
+SYM_FUNC_END(mte_memcpy)
arch/arm64/mm/fault.c +46 -2

61 lines not on this page · raw message

arch/ia64/kernel/syscalls/syscall.tbl +1 -0
index 6d07742c57b8..c2209d83f3c3 100644
@@ -367,3 +367,4 @@
 444	common	landlock_create_ruleset		sys_landlock_create_ruleset
 445	common	landlock_add_rule		sys_landlock_add_rule
 446	common	landlock_restrict_self		sys_landlock_restrict_self
+448	common	refpage_create			sys_refpage_create
arch/m68k/kernel/syscalls/syscall.tbl +1 -0
index 541bc1b3a8f9..0360cf474a49 100644
@@ -446,3 +446,4 @@
 444	common	landlock_create_ruleset		sys_landlock_create_ruleset
 445	common	landlock_add_rule		sys_landlock_add_rule
 446	common	landlock_restrict_self		sys_landlock_restrict_self
+448	common	refpage_create			sys_refpage_create
arch/microblaze/kernel/syscalls/syscall.tbl +1 -0
index a176faca2927..de85d758e564 100644
@@ -452,3 +452,4 @@
 444	common	landlock_create_ruleset		sys_landlock_create_ruleset
 445	common	landlock_add_rule		sys_landlock_add_rule
 446	common	landlock_restrict_self		sys_landlock_restrict_self
+448	common	refpage_create			sys_refpage_create
arch/mips/kernel/syscalls/syscall_n32.tbl +1 -0
index c2d2e19abea8..b07c7293d2a3 100644
@@ -385,3 +385,4 @@
 444	n32	landlock_create_ruleset		sys_landlock_create_ruleset
 445	n32	landlock_add_rule		sys_landlock_add_rule
 446	n32	landlock_restrict_self		sys_landlock_restrict_self
+448	n32	refpage_create			sys_refpage_create
arch/mips/kernel/syscalls/syscall_n64.tbl +1 -0
index ac653d08b1ea..7ebabb99dd06 100644
@@ -361,3 +361,4 @@
 444	n64	landlock_create_ruleset		sys_landlock_create_ruleset
 445	n64	landlock_add_rule		sys_landlock_add_rule
 446	n64	landlock_restrict_self		sys_landlock_restrict_self
+448	n64	refpage_create			sys_refpage_create
arch/mips/kernel/syscalls/syscall_o32.tbl +1 -0
index 253f2cd70b6b..a51149ac101c 100644
@@ -434,3 +434,4 @@
 444	o32	landlock_create_ruleset		sys_landlock_create_ruleset
 445	o32	landlock_add_rule		sys_landlock_add_rule
 446	o32	landlock_restrict_self		sys_landlock_restrict_self
+448	o32	refpage_create			sys_refpage_create
arch/parisc/kernel/syscalls/syscall.tbl +1 -0
index e26187b9ab87..385565864861 100644
@@ -444,3 +444,4 @@
 444	common	landlock_create_ruleset		sys_landlock_create_ruleset
 445	common	landlock_add_rule		sys_landlock_add_rule
 446	common	landlock_restrict_self		sys_landlock_restrict_self
+448	common	refpage_create			sys_refpage_create
arch/powerpc/kernel/syscalls/syscall.tbl +1 -0
index aef2a290e71a..95cdd9f7dc06 100644
@@ -526,3 +526,4 @@
 444	common	landlock_create_ruleset		sys_landlock_create_ruleset
 445	common	landlock_add_rule		sys_landlock_add_rule
 446	common	landlock_restrict_self		sys_landlock_restrict_self
+448	common	refpage_create			sys_refpage_create
arch/s390/kernel/syscalls/syscall.tbl +1 -0
index 64d51ab5a8b4..92ed1260ffd9 100644
@@ -449,3 +449,4 @@
 444  common	landlock_create_ruleset	sys_landlock_create_ruleset	sys_landlock_create_ruleset
 445  common	landlock_add_rule	sys_landlock_add_rule		sys_landlock_add_rule
 446  common	landlock_restrict_self	sys_landlock_restrict_self	sys_landlock_restrict_self
+448  common	refpage_create		sys_refpage_create		sys_refpage_create
arch/sh/kernel/syscalls/syscall.tbl +1 -0
index e0a70be77d84..f9d198cc2541 100644
@@ -449,3 +449,4 @@
 444	common	landlock_create_ruleset		sys_landlock_create_ruleset
 445	common	landlock_add_rule		sys_landlock_add_rule
 446	common	landlock_restrict_self		sys_landlock_restrict_self
+448	common	refpage_create			sys_refpage_create
arch/sparc/kernel/syscalls/syscall.tbl +1 -0
index 603f5a821502..83533aa49340 100644
@@ -492,3 +492,4 @@
 444	common	landlock_create_ruleset		sys_landlock_create_ruleset
 445	common	landlock_add_rule		sys_landlock_add_rule
 446	common	landlock_restrict_self		sys_landlock_restrict_self
+448	common	refpage_create			sys_refpage_create
arch/x86/entry/syscalls/syscall_32.tbl +1 -0
index ce763a12311c..054c69e395b5 100644
@@ -452,3 +452,4 @@
 445	i386	landlock_add_rule	sys_landlock_add_rule
 446	i386	landlock_restrict_self	sys_landlock_restrict_self
 447	i386	memfd_secret		sys_memfd_secret
+448	i386	refpage_create		sys_refpage_create
arch/x86/entry/syscalls/syscall_64.tbl +1 -0
index f6b57799c1ea..1f24f0b66cbd 100644
@@ -369,6 +369,7 @@
 445	common	landlock_add_rule	sys_landlock_add_rule
 446	common	landlock_restrict_self	sys_landlock_restrict_self
 447	common	memfd_secret		sys_memfd_secret
+448	common	refpage_create		sys_refpage_create
 
 #
 # Due to a historical design error, certain syscalls are numbered differently
arch/xtensa/kernel/syscalls/syscall.tbl +1 -0
index 235d67d6ceb4..96c27fb404ca 100644
@@ -417,3 +417,4 @@
 444	common	landlock_create_ruleset		sys_landlock_create_ruleset
 445	common	landlock_add_rule		sys_landlock_add_rule
 446	common	landlock_restrict_self		sys_landlock_restrict_self
+448	common	refpage_create			sys_refpage_create
include/linux/gfp.h +9 -2
index 55b2ec1f965a..ae3c763eb9e9 100644
@@ -55,8 +55,9 @@ struct vm_area_struct;
 #define ___GFP_ACCOUNT		0x400000u
 #define ___GFP_ZEROTAGS		0x800000u
 #define ___GFP_SKIP_KASAN_POISON	0x1000000u
+#define ___GFP_NOZERO		0x2000000u
 #ifdef CONFIG_LOCKDEP
-#define ___GFP_NOLOCKDEP	0x2000000u
+#define ___GFP_NOLOCKDEP	0x4000000u
 #else
 #define ___GFP_NOLOCKDEP	0
 #endif
@@ -238,18 +239,24 @@ struct vm_area_struct;
  * %__GFP_SKIP_KASAN_POISON returns a page which does not need to be poisoned
  * on deallocation. Typically used for userspace pages. Currently only has an
  * effect in HW tags mode.
+ *
+ * %__GFP_NOZERO disables any implicit zeroing of the page (e.g. as a result
+ * of init_on_alloc=on). This flag should only be used to address specific
+ * performance bottlenecks and only if the page is clearly being fully
+ * initialized following the allocation.
  */
 #define __GFP_NOWARN	((__force gfp_t)___GFP_NOWARN)
 #define __GFP_COMP	((__force gfp_t)___GFP_COMP)
 #define __GFP_ZERO	((__force gfp_t)___GFP_ZERO)
 #define __GFP_ZEROTAGS	((__force gfp_t)___GFP_ZEROTAGS)
 #define __GFP_SKIP_KASAN_POISON	((__force gfp_t)___GFP_SKIP_KASAN_POISON)
+#define __GFP_NOZERO	((__force gfp_t)___GFP_NOZERO)
 
 /* Disable lockdep for GFP context tracking */
 #define __GFP_NOLOCKDEP ((__force gfp_t)___GFP_NOLOCKDEP)
 
 /* Room for N __GFP_FOO bits */
-#define __GFP_BITS_SHIFT (25 + IS_ENABLED(CONFIG_LOCKDEP))
+#define __GFP_BITS_SHIFT (26 + IS_ENABLED(CONFIG_LOCKDEP))
 #define __GFP_BITS_MASK ((__force gfp_t)((1 << __GFP_BITS_SHIFT) - 1))
 
 /**
include/linux/highmem.h +7 -2
index 8c6e8e996c87..038c235fd146 100644
@@ -185,14 +185,19 @@ static inline void clear_highpage(struct page *page)
 	kunmap_atomic(kaddr);
 }
 
-#ifndef __HAVE_ARCH_TAG_CLEAR_HIGHPAGE
+#ifndef __HAVE_ARCH_TAG_SET_HIGHPAGE
 
-static inline void tag_clear_highpage(struct page *page)
+static inline void tag_set_highpage(struct page *page, unsigned long tag)
 {
 }
 
 #endif
 
+static inline void tag_clear_highpage(struct page *page)
+{
+	tag_set_highpage(page, 0);
+}
+
 /*
  * If we pass in a base or tail page, we can zero up to PAGE_SIZE.
  * If we pass in a head page, we can zero up to the size of the compound page.
include/linux/huge_mm.h +7 -0
index f123e15d966e..36ecfc391b46 100644
@@ -127,6 +127,13 @@ static inline bool transhuge_vma_suitable(struct vm_area_struct *vma,
 
 	if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end)
 		return false;
+
+	/*
+	 * Transparent hugepages not currently supported for anonymous VMAs with
+	 * reference pages
+	 */
+	if (unlikely(is_refpage_vma(vma)))
+		return false;
 	return true;
 }
 
include/linux/mm.h +32 -0
index 7ca22e6e694a..4a49ed7099cd 100644
@@ -722,6 +722,36 @@ int vma_is_stack_for_current(struct vm_area_struct *vma);
 /* flush_tlb_range() takes a vma, not a mm, and can care about flags */
 #define TLB_FLUSH_VMA(mm,flags) { .vm_mm = (mm), .vm_flags = (flags) }
 
+struct refpage_private_data {
+	struct page *refpage;
+	u8 content_type;
+	u8 content_info;
+};
+
+#define REFPAGE_CONTENT_TYPE_USER_SET	0
+#define REFPAGE_CONTENT_TYPE_PATTERN	1
+#define REFPAGE_CONTENT_TYPE_MTE_TAGGED	2
+
+bool is_refpage_file(struct file *file);
+
+static inline bool is_refpage_vma(struct vm_area_struct *vma)
+{
+	return vma->vm_file && is_refpage_file(vma->vm_file);
+}
+
+static inline struct page *get_vma_refpage(struct vm_area_struct *vma)
+{
+	struct refpage_private_data *priv = vma->vm_private_data;
+
+	BUG_ON(!is_refpage_vma(vma));
+	return priv->refpage;
+}
+
+static inline int is_refpage_pfn(struct vm_area_struct *vma, unsigned long pfn)
+{
+	return is_refpage_vma(vma) && pfn == page_to_pfn(get_vma_refpage(vma));
+}
+
 struct mmu_gather;
 struct inode;
 
@@ -2977,6 +3007,8 @@ static inline void kernel_unpoison_pages(struct page *page, int numpages) { }
 DECLARE_STATIC_KEY_MAYBE(CONFIG_INIT_ON_ALLOC_DEFAULT_ON, init_on_alloc);
 static inline bool want_init_on_alloc(gfp_t flags)
 {
+	if (flags & __GFP_NOZERO)
+		return false;
 	if (static_branch_maybe(CONFIG_INIT_ON_ALLOC_DEFAULT_ON,
 				&init_on_alloc))
 		return true;
include/linux/mman.h +19 -0
index ebb09a964272..03a4678d93c6 100644
@@ -123,6 +123,25 @@ static inline bool arch_validate_flags(unsigned long flags)
 #define arch_validate_flags arch_validate_flags
 #endif
 
+void prep_refpage_private_data(struct refpage_private_data *priv,
+			       char *content_kaddr, unsigned long size);
+#ifndef arch_prep_refpage_private_data
+#define arch_prep_refpage_private_data prep_refpage_private_data
+#endif
+
+#ifndef arch_prep_refpage_vma
+static inline void arch_prep_refpage_vma(struct vm_area_struct *vma)
+{
+}
+#define arch_prep_refpage_vma arch_prep_refpage_vma
+#endif
+
+void copy_refpage(struct page *page, unsigned long addr,
+		  struct vm_area_struct *vma);
+#ifndef arch_copy_refpage
+#define arch_copy_refpage copy_refpage
+#endif
+
 /*
  * Optimisation macro.  It is equivalent to:
  *      (x & bit1) ? bit2 : 0
include/linux/syscalls.h +3 -0
index 69c9a7010081..303a28a86500 100644
@@ -864,6 +864,9 @@ asmlinkage long sys_mremap(unsigned long addr,
 			   unsigned long old_len, unsigned long new_len,
 			   unsigned long flags, unsigned long new_addr);
 
+/* mm/refpage.c */
+asmlinkage long sys_refpage_create(const void __user *content, unsigned long flags);
+
 /* security/keys/keyctl.c */
 asmlinkage long sys_add_key(const char __user *_type,
 			    const char __user *_description,
include/uapi/asm-generic/unistd.h +4 -1
index a9d6fcd95f42..54cede7db5f0 100644
@@ -878,8 +878,11 @@ __SYSCALL(__NR_landlock_restrict_self, sys_landlock_restrict_self)
 __SYSCALL(__NR_memfd_secret, sys_memfd_secret)
 #endif
 
+#define __NR_refpage_create 448
+__SYSCALL(__NR_refpage_create, sys_refpage_create)
+
 #undef __NR_syscalls
-#define __NR_syscalls 448
+#define __NR_syscalls 449
 
 /*
  * 32 bit systems traditionally used different
kernel/sys_ni.c +1 -0
index 30971b1dd4a9..bc65a54eb2a4 100644
@@ -300,6 +300,7 @@ COND_SYSCALL(migrate_pages);
 COND_SYSCALL_COMPAT(migrate_pages);
 COND_SYSCALL(move_pages);
 COND_SYSCALL_COMPAT(move_pages);
+COND_SYSCALL(refpage_create);
 
 COND_SYSCALL(perf_event_open);
 COND_SYSCALL(accept4);
mm/Makefile +2 -2
index e3436741d539..137adc22bf50 100644
@@ -35,10 +35,10 @@ CFLAGS_init-mm.o += $(call cc-disable-warning, override-init)
 CFLAGS_init-mm.o += $(call cc-disable-warning, initializer-overrides)
 
 mmu-y			:= nommu.o
-mmu-$(CONFIG_MMU)	:= highmem.o memory.o mincore.o \
+mmu-$(CONFIG_MMU)	:= highmem.o ioremap.o memory.o mincore.o \
 			   mlock.o mmap.o mmu_gather.o mprotect.o mremap.o \
 			   msync.o page_vma_mapped.o pagewalk.o \
-			   pgtable-generic.o rmap.o vmalloc.o ioremap.o
+			   pgtable-generic.o refpage.o rmap.o vmalloc.o
 
 
 ifdef CONFIG_CROSS_MEMORY_ATTACH
mm/gup.c +2 -1
index 42b8b1fa6521..3beffc84c449 100644
@@ -548,7 +548,8 @@ static struct page *follow_page_pte(struct vm_area_struct *vma,
 			goto out;
 		}
 
-		if (is_zero_pfn(pte_pfn(pte))) {
+		if (is_zero_pfn(pte_pfn(pte)) ||
+		    is_refpage_pfn(vma, pte_pfn(pte))) {
 			page = pte_page(pte);
 		} else {
 			ret = follow_pfn_pte(vma, address, ptep, flags);
mm/memory.c +37 -10

97 lines not on this page · raw message

mm/migrate.c +2 -2
index 34a9ad3e0a4f..2ad41a69dd0e 100644
@@ -2726,8 +2726,8 @@ static void migrate_vma_insert_page(struct migrate_vma *migrate,
 	pmd_t *pmdp;
 	pte_t *ptep;
 
-	/* Only allow populating anonymous memory */
-	if (!vma_is_anonymous(vma))
+	/* Only allow populating anonymous memory without a reference page */
+	if (!vma_is_anonymous(vma) || is_refpage_vma(vma))
 		goto abort;
 
 	pgdp = pgd_offset(mm, addr);
mm/refpage.c new file +151 -0

151 lines not on this page · raw message

Signature · 1 line
2.32.0.402.g57bb445576-goog
#1 lore raw
Quoted 7 lines · Peter Collingbourne

Introduce a new syscall, refpage_create, which returns a file descriptor which may be mapped using mmap. Such a mapping is similar to an anonymous mapping, but instead of clean pages being backed by the zero page, they are instead backed by a so-called reference page, whose contents are specified using an argument to refpage_create. Loads from the mapping will load directly from the reference page, and initial stores to the mapping will copy-on-write from the reference page.

I'm wondering, does the target use case really require the COW optimization like we have for the shared zeropage?

If we'd avoid having a reference page at all and only store the pattern, we could significantly reduce the memory consumption when using a lot of reference pages, especially per process multiple ones. I'm asking because ...

Quoted 6 lines

Reference pages are useful in circumstances where anonymous mappings combined with manual stores to memory would impose undesirable costs, either in terms of performance or RSS. Use cases are focused on heap allocators and include:

Quoted 11 lines

- Pattern initialization for the heap. This is where malloc(3) gives you memory whose contents are filled with a non-zero pattern byte, in order to help detect and mitigate bugs involving use of uninitialized memory. Typically this is implemented by having the allocator memset the allocation with the pattern byte before returning it to the user, but for large allocations this can result in a significant increase in RSS, especially for allocations that are used sparsely. Even for dense allocations there is a needless impact to startup performance when it may be better to amortize it throughout the program. By creating allocations using a reference page filled with the pattern byte, we can avoid these costs.

... I assume the first *sane* access to such a page is a write, and not a read.

Quoted 10 lines

- Pre-tagged heap memory. Memory tagging [1] is an upcoming ARMv8.5 feature which allows for memory to be tagged in order to detect certain kinds of memory errors with low overhead. In order to set up an allocation to allow memory errors to be detected, the entire allocation needs to have the same tag. The issue here is similar to pattern initialization in the sense that large tagged allocations will be expensive if the tagging is done up front. The idea is that the allocator would create reference pages with each of the possible memory tags, and use those reference pages for the large allocations.

... and here as well.

Having a first access being a read sound more like an actual BUG (e.g., detect and mitigate bugs), which doesn't scream for needing a performance improvement or sacrificing a whole (unmovable/unswappable) reference page.

So, what would you lose when not populating a real reference pages at all and instead only populating the pattern when populating a fresh page? (and populating a fresh page even on read faults)

Signature · 3 lines
Thanks,

David / dhildenb
maintainer
#2 lore raw
In reply to Peter Collingbourne
Quoted 7 lines · Peter Collingbourne

Introduce a new syscall, refpage_create, which returns a file descriptor which may be mapped using mmap. Such a mapping is similar to an anonymous mapping, but instead of clean pages being backed by the zero page, they are instead backed by a so-called reference page, whose contents are specified using an argument to refpage_create. Loads from the mapping will load directly from the reference page, and initial stores to the mapping will copy-on-write from the reference page.

I think you mentioned in an earlier version, we can't use memfd_create() as we want a single reference page to back arbitrarily sized mmap() regions. Random API thought: could we instead add an MFD_MMAP_WRAP flag or something to memfd_create() which pretends on mmap(MAP_PRIVATE) that it's an infinite file and just replicates the data it already has in that fd on read? On write, it would do a CoW.

Not sure it's any better than your proposal, just wondering whether we can contain it to shmem_fault() mostly and maybe a new shmem_map_pages(). Anyway, more API design comments below.

Quoted 17 lines

Reference pages are useful in circumstances where anonymous mappings combined with manual stores to memory would impose undesirable costs, either in terms of performance or RSS. Use cases are focused on heap allocators and include:

- Pattern initialization for the heap. This is where malloc(3) gives you memory whose contents are filled with a non-zero pattern byte, in order to help detect and mitigate bugs involving use of uninitialized memory. Typically this is implemented by having the allocator memset the allocation with the pattern byte before returning it to the user, but for large allocations this can result in a significant increase in RSS, especially for allocations that are used sparsely. Even for dense allocations there is a needless impact to startup performance when it may be better to amortize it throughout the program. By creating allocations using a reference page filled with the pattern byte, we can avoid these costs.

Quoted 9 lines

- Pre-tagged heap memory. Memory tagging [1] is an upcoming ARMv8.5 feature which allows for memory to be tagged in order to detect certain kinds of memory errors with low overhead. In order to set up an allocation to allow memory errors to be detected, the entire allocation needs to have the same tag. The issue here is similar to pattern initialization in the sense that large tagged allocations will be expensive if the tagging is done up front. The idea is that the allocator would create reference pages with each of the possible memory tags, and use those reference pages for the large allocations.

So for MTE you'd need the same tag other than tag 0 (which is what mmap() gives you). An alternative here is to allow passing a tag hint to mmap() with the kernel allocating and maintaining a reference page (we had some non-optimal internal attempts at this). But if the first case (data pattern initialisation) is equally important, a reference pattern may be a better option.

Quoted 5 lines

This patch includes specific optimizations for these use cases in order to reduce memory traffic. If the reference page consists of a single repeating byte, the page is initialized using memset, and on arm64 if the reference page consists of a uniformly tagged zero page, the DC GZVA instruction is used to initialize the page.

I don't particularly like such heuristics. If the memset/single-tag is your main use-case, just bake it into the API from the start. As David said, how important is the read prior to a write of such page? I suspect it's not relevant at all, so we could have the page allocated and populated on fault (even a read fault). The pattern and its size could be stored in the kernel as I suspect even if you need something other than a single byte, it would still be a repetitive pattern.

Also please split this into multiple patches as it's hard to follow. Something like (1) refpage_create, (2) basic arm64 tagging support, (3) optimisation for specific cases etc.

Quoted 22 lines
arch/arm64/kernel/mte.c +14 -0
index 36f51b0e438a..876997703d7b 100644
@@ -450,3 +450,27 @@ int mte_ptrace_copy_tags(struct task_struct *child, long request,
 
 	return ret;
 }
+
+u8 mte_check_tag_zero_page(struct page *userpage)
+{
+	char *userpage_addr = page_address(userpage);
+	u8 tag;
+	int i;
+
+	if (!test_bit(PG_mte_tagged, &userpage->flags))
+		return 0;
+
+	tag = mte_get_mem_tag(userpage_addr) & 0xF;
+	if (tag == 0)
+		return 0;
+

arch/arm64/kernel/mte.c int mte_ptrace_copy_tags(struct task_struct *child, long request, · line 476

Quoted 10 lines
arch/arm64/kernel/mte.c +10 -0
+	for (i = 0; i != PAGE_SIZE; ++i)
+		if (userpage_addr[i] != 0)
+			return 0;
+
+	for (i = MTE_GRANULE_SIZE; i != PAGE_SIZE; i += MTE_GRANULE_SIZE)
+		if ((mte_get_mem_tag(userpage_addr + i) & 0xF) != tag)
+			return 0;
+
+	return tag;
+}

What's this function for? I couldn't see a user in this patch (maybe I didn't look harder).

Quoted 32 lines
arch/arm64/mm/fault.c +11 -2
index 349c488765ca..73fe728787f0 100644
@@ -25,6 +25,7 @@
 #include <linux/perf_event.h>
 #include <linux/preempt.h>
 #include <linux/hugetlb.h>
+#include <linux/mman.h>
 
 #include <asm/acpi.h>
 #include <asm/bug.h>
@@ -939,9 +940,52 @@ struct page *alloc_zeroed_user_highpage_movable(struct vm_area_struct *vma,
 	return alloc_page_vma(flags, vma, vaddr);
 }
 
-void tag_clear_highpage(struct page *page)
+void tag_set_highpage(struct page *page, unsigned long tag)
 {
-	mte_zero_clear_page_tags(page_address(page));
+	unsigned long addr = (unsigned long)page_address(page);
+
+	addr &= ~MTE_TAG_MASK;
+	addr |= tag << MTE_TAG_SHIFT;
+	mte_zero_set_page_tags((void *)addr);
 	page_kasan_tag_reset(page);
 	set_bit(PG_mte_tagged, &page->flags);
 }
+
+void arch_prep_refpage_private_data(struct refpage_private_data *priv,
+				    char *content_kaddr, unsigned long size)
+{

arch/arm64/mm/fault.c struct page *alloc_zeroed_user_highpage_movable(struct vm_area_struct *vma, · line 966

Quoted 10 lines
arch/arm64/mm/fault.c +10 -0
+	if (system_supports_mte()) {
+		u64 *content_kaddr_u64 = (u64 *)content_kaddr;
+
+		if (size == MTE_GRANULE_SIZE && content_kaddr_u64[0] == 0 &&
+		    content_kaddr_u64[1] == 0) {
+			priv->content_type = REFPAGE_CONTENT_TYPE_MTE_TAGGED;
+			priv->content_info = mte_get_mem_tag(content_kaddr);
+		}
+
+		set_bit(PG_mte_tagged, &priv->refpage->flags);

As per your example, the reference page may be in the data section. We explicitly avoid tagged pages in there (could be a file mmap, not necessarily bss) but you just force it here. I'd rather skip the tagging altogether if the page was not mapped with PROT_MTE.

Quoted 4 lines
arch/arm64/mm/fault.c +4 -0
+		if (size >= MTE_GRANULE_SIZE) {
+			char *refpage_kaddr = page_address(priv->refpage);
+			unsigned long i;
+

arch/arm64/mm/fault.c struct page *alloc_zeroed_user_highpage_movable(struct vm_area_struct *vma, · line 980

Quoted 10 lines
arch/arm64/mm/fault.c +10 -0
+			for (i = 0; i != PAGE_SIZE; i += size)
+				mte_memcpy(refpage_kaddr + i, content_kaddr,
+					   size);
+			return;
+		}
+		mte_clear_page_tags(page_address(priv->refpage));
+	}
+
+	prep_refpage_private_data(priv, content_kaddr, size);
+}

IIUC this function writes to the reference page provided by the user but get_user_pages() did not use a FOLL_WRITE. Even worse, the refpage_create() syscall does not seem to reject size != PAGE_SIZE values, so the user may think it's fine to have a shorter buffer with the kernel overriding whatever is around it. If the API states that the buffer must be PAGE_SIZE, why not get the user to populate it and avoid this arch_prep_refpage_private_data() altogether. I think it overcomplicates the kernel for something the user could do.

arch/arm64/mm/fault.c struct page *alloc_zeroed_user_highpage_movable(struct vm_area_struct *vma, · line 991

Quoted 11 lines
arch/arm64/mm/fault.c +11 -0
+
+void arch_copy_refpage(struct page *page, unsigned long addr,
+		       struct vm_area_struct *vma)
+{
+	struct refpage_private_data *priv = vma->vm_private_data;
+
+	if (priv->content_type == REFPAGE_CONTENT_TYPE_MTE_TAGGED)
+		tag_set_highpage(page, priv->content_info);
+	else
+		copy_refpage(page, addr, vma);
+}

Another heuristics targeted at your specific allocator. Just leave it as a later optimisation or at least a separate patch.

Quoted 6 lines
mm/refpage.c new file +0 -0
new file mode 100644
index 000000000000..13c0930dc926
@@ -0,0 +1,151 @@

[...]

Quoted 5 lines
mm/refpage.c new file +5 -0
+SYSCALL_DEFINE3(refpage_create, const void *__user, content, unsigned int,
+		size, unsigned long, flags)
+{
+	unsigned long content_addr = (unsigned long)content;
+	char *content_kaddr;

mm/refpage.c line 115

Quoted 10 lines
mm/refpage.c new file +10 -0
+	struct page *userpage;
+	struct refpage_private_data *private_data;
+	long retval;
+
+	if (flags != 0)
+		return -EINVAL;
+
+	if (!is_power_of_2(size) || size > PAGE_SIZE ||
+	    (content_addr & (size - 1)) != 0)
+		return -EINVAL;

Alternative: copy the pattern from the user content to a kernel buffer and in the file ops .mmap, instead of vma_set_anonymous(), just implement your own vma ops with a .fault callback that takes care of copying the pattern to a newly allocated page on any fault.

mm/refpage.c line 119

Quoted 4 lines
mm/refpage.c new file +4 -0
+
+	if (get_user_pages(content_addr & ~(PAGE_SIZE - 1), 1, 0, &userpage,
+			   0) != 1)
+		return -EFAULT;

FOLL_WRITE to the gup flags?

Signature · 1 line
Catalin