[PATCH v3 0/5] Speed up mremap on large regions

[PATCH v3 0/5] Speed up mremap on large regions

[Kalesh Singh]

This version 3 of the mremap speed up patches previously posted at:
v1 - https://lore.kernel.org/r/20200930222130.4175584-1-kaleshsingh@google.com
v2 - https://lore.kernel.org/r/20201002162101.665549-1-kaleshsingh@google.com

mremap time can be optimized by moving entries at the PMD/PUD level if
the source and destination addresses are PMD/PUD-aligned and
PMD/PUD-sized. Enable moving at the PMD and PUD levels on arm64 and
x86. Other architectures where this type of move is supported and known to
be safe can also opt-in to these optimizations by enabling HAVE_MOVE_PMD
and HAVE_MOVE_PUD.

Observed Performance Improvements for remapping a PUD-aligned 1GB-sized
region on x86 and arm64:

    - HAVE_MOVE_PMD is already enabled on x86 : N/A
    - Enabling HAVE_MOVE_PUD on x86   : ~13x speed up

    - Enabling HAVE_MOVE_PMD on arm64 : ~ 8x speed up
    - Enabling HAVE_MOVE_PUD on arm64 : ~19x speed up

          Altogether, HAVE_MOVE_PMD and HAVE_MOVE_PUD
          give a total of ~150x speed up on arm64.

Changes in v2:
  - Reduce mremap_test time by only validating a configurable
    threshold of the remapped region, as per John.
  - Use a random pattern for mremap validation. Provide pattern
    seed in test output, as per John.
  - Moved set_pud_at() to separate patch, per Kirill.
  - Use switch() instead of ifs in move_pgt_entry(), per Kirill.
  - Update commit message with description of Android
    garbage collector use case for HAVE_MOVE_PUD, as per Joel.
  - Fix build test error reported by kernel test robot in [1].

Changes in v3:
  - Make lines 80 cols or less where they don’t need to be longer,
    per John.
  - Removed unused PATTERN_SIZE in mremap_test
  - Added Reviewed-by tag for patch 1/5 (mremap kselftest patch).
  - Use switch() instead of ifs in get_extent(), per Kirill
  - Add BUILD_BUG() is get_extent() default case.
  - Move get_old_pud() and alloc_new_pud() out of
    #ifdef CONFIG_HAVE_MOVE_PUD, per Kirill.
  - Have get_old_pmd() and alloc_new_pmd() use get_old_pud() and
    alloc_old_pud(), per Kirill.
  - Replace #ifdef CONFIG_HAVE_MOVE_PMD / PUD in move_page_tables()
    with IS_ENABLED(CONFIG_HAVE_MOVE_PMD / PUD), per Kirill.
  - Fold Add set_pud_at() patch into patch 4/5, per Kirill.

[1] https://lists.01.org/hyperkitty/list/kbuild-all@lists.01.org/thread/CKPGL4FH4NG7TGH2CVYX2UX76L25BTA3/

Kalesh Singh (5):
  kselftests: vm: Add mremap tests
  arm64: mremap speedup - Enable HAVE_MOVE_PMD
  mm: Speedup mremap on 1GB or larger regions
  arm64: mremap speedup - Enable HAVE_MOVE_PUD
  x86: mremap speedup - Enable HAVE_MOVE_PUD

 arch/Kconfig                             |   7 +
 arch/arm64/Kconfig                       |   2 +
 arch/arm64/include/asm/pgtable.h         |   1 +
 arch/x86/Kconfig                         |   1 +
 mm/mremap.c                              | 230 ++++++++++++---
 tools/testing/selftests/vm/.gitignore    |   1 +
 tools/testing/selftests/vm/Makefile      |   1 +
 tools/testing/selftests/vm/mremap_test.c | 344 +++++++++++++++++++++++
 tools/testing/selftests/vm/run_vmtests   |  11 +
 9 files changed, 558 insertions(+), 40 deletions(-)
 create mode 100644 tools/testing/selftests/vm/mremap_test.c


base-commit: 549738f15da0e5a00275977623be199fbbf7df50
-- 
2.28.0.806.g8561365e88-goog

[PATCH v3 1/5] kselftests: vm: Add mremap tests

[Kalesh Singh]

Test mremap on regions of various sizes and alignments and validate
data after remapping. Also provide total time for remapping
the region which is useful for performance comparison of the mremap
optimizations that move pages at the PMD/PUD levels if HAVE_MOVE_PMD
and/or HAVE_MOVE_PUD are enabled.

Signed-off-by: Kalesh Singh <kaleshsingh@google.com>
Reviewed-by: John Hubbard <jhubbard@nvidia.com>
---
Changes in v2:
  - Reduce test time by only validating a certain threshold of the
    remapped region (4MB by default). The -t flag can be used to
    set a custom threshold in MB or no threshold by passing 0. (-t0).
    mremap time is not provided in stdout for only partially validated
    regions. This time is only applicable for comparison if the entire
    mapped region was faulted in.
  - Use a random pattern for validating the remapped region. The -p
    flag can be used to run the tests with a specified seed for the
    random pattern.
  - Print test configs (threshold_mb and pattern_seed) to stdout.
  - Remove MAKE_SIMPLE_TEST macro.
  - Define named flags instead of 0 / 1.
  - Add comments for destination address' align_mask and offset.

Changes in v3:
  - Remove unused PATTERN_SIZE definition.
  - Make lines 80 cols or less where they don’t need to be longer.
  - Add John Hubbard’s Reviewed-by tag.

 tools/testing/selftests/vm/.gitignore    |   1 +
 tools/testing/selftests/vm/Makefile      |   1 +
 tools/testing/selftests/vm/mremap_test.c | 344 +++++++++++++++++++++++
 tools/testing/selftests/vm/run_vmtests   |  11 +
 4 files changed, 357 insertions(+)
 create mode 100644 tools/testing/selftests/vm/mremap_test.c

diff --git a/tools/testing/selftests/vm/.gitignore b/tools/testing/selftests/vm/.gitignore
index 849e8226395a..b3a183c36cb5 100644
--- a/tools/testing/selftests/vm/.gitignore
+++ b/tools/testing/selftests/vm/.gitignore
@@ -8,6 +8,7 @@ thuge-gen
 compaction_test
 mlock2-tests
 mremap_dontunmap
+mremap_test
 on-fault-limit
 transhuge-stress
 protection_keys
diff --git a/tools/testing/selftests/vm/Makefile b/tools/testing/selftests/vm/Makefile
index a9026706d597..f044808b45fa 100644
--- a/tools/testing/selftests/vm/Makefile
+++ b/tools/testing/selftests/vm/Makefile
@@ -16,6 +16,7 @@ TEST_GEN_FILES += map_populate
 TEST_GEN_FILES += mlock-random-test
 TEST_GEN_FILES += mlock2-tests
 TEST_GEN_FILES += mremap_dontunmap
+TEST_GEN_FILES += mremap_test
 TEST_GEN_FILES += on-fault-limit
 TEST_GEN_FILES += thuge-gen
 TEST_GEN_FILES += transhuge-stress
diff --git a/tools/testing/selftests/vm/mremap_test.c b/tools/testing/selftests/vm/mremap_test.c
new file mode 100644
index 000000000000..9c391d016922
--- /dev/null
+++ b/tools/testing/selftests/vm/mremap_test.c
@@ -0,0 +1,344 @@
+// SPDX-License-Identifier: GPL-2.0
+/*
+ * Copyright 2020 Google LLC
+ */
+#define _GNU_SOURCE
+
+#include <errno.h>
+#include <stdlib.h>
+#include <string.h>
+#include <sys/mman.h>
+#include <time.h>
+
+#include "../kselftest.h"
+
+#define EXPECT_SUCCESS 0
+#define EXPECT_FAILURE 1
+#define NON_OVERLAPPING 0
+#define OVERLAPPING 1
+#define NS_PER_SEC 1000000000ULL
+#define VALIDATION_DEFAULT_THRESHOLD 4	/* 4MB */
+#define VALIDATION_NO_THRESHOLD 0	/* Verify the entire region */
+
+#define ARRAY_SIZE(x) (sizeof(x) / sizeof((x)[0]))
+#define MIN(X, Y) ((X) < (Y) ? (X) : (Y))
+
+struct config {
+	unsigned long long src_alignment;
+	unsigned long long dest_alignment;
+	unsigned long long region_size;
+	int overlapping;
+};
+
+struct test {
+	const char *name;
+	struct config config;
+	int expect_failure;
+};
+
+enum {
+	_1KB = 1ULL << 10,	/* 1KB -> not page aligned */
+	_4KB = 4ULL << 10,
+	_8KB = 8ULL << 10,
+	_1MB = 1ULL << 20,
+	_2MB = 2ULL << 20,
+	_4MB = 4ULL << 20,
+	_1GB = 1ULL << 30,
+	_2GB = 2ULL << 30,
+	PTE = _4KB,
+	PMD = _2MB,
+	PUD = _1GB,
+};
+
+#define MAKE_TEST(source_align, destination_align, size,	\
+		  overlaps, should_fail, test_name)		\
+{								\
+	.name = test_name,					\
+	.config = {						\
+		.src_alignment = source_align,			\
+		.dest_alignment = destination_align,		\
+		.region_size = size,				\
+		.overlapping = overlaps,			\
+	},							\
+	.expect_failure = should_fail				\
+}
+
+/*
+ * Returns the start address of the mapping on success, else returns
+ * NULL on failure.
+ */
+static void *get_source_mapping(struct config c)
+{
+	unsigned long long addr = 0ULL;
+	void *src_addr = NULL;
+retry:
+	addr += c.src_alignment;
+	src_addr = mmap((void *) addr, c.region_size, PROT_READ | PROT_WRITE,
+			MAP_FIXED | MAP_ANONYMOUS | MAP_SHARED, -1, 0);
+	if (src_addr == MAP_FAILED) {
+		if (errno == EPERM)
+			goto retry;
+		goto error;
+	}
+	/*
+	 * Check that the address is aligned to the specified alignment.
+	 * Addresses which have alignments that are multiples of that
+	 * specified are not considered valid. For instance, 1GB address is
+	 * 2MB-aligned, however it will not be considered valid for a
+	 * requested alignment of 2MB. This is done to reduce coincidental
+	 * alignment in the tests.
+	 */
+	if (((unsigned long long) src_addr & (c.src_alignment - 1)) ||
+			!((unsigned long long) src_addr & c.src_alignment))
+		goto retry;
+
+	if (!src_addr)
+		goto error;
+
+	return src_addr;
+error:
+	ksft_print_msg("Failed to map source region: %s\n",
+			strerror(errno));
+	return NULL;
+}
+
+/* Returns the time taken for the remap on success else returns -1. */
+static long long remap_region(struct config c, unsigned int threshold_mb,
+			      char pattern_seed)
+{
+	void *addr, *src_addr, *dest_addr;
+	unsigned long long i;
+	struct timespec t_start = {0, 0}, t_end = {0, 0};
+	long long  start_ns, end_ns, align_mask, ret, offset;
+	unsigned long long threshold;
+
+	if (threshold_mb == VALIDATION_NO_THRESHOLD)
+		threshold = c.region_size;
+	else
+		threshold = MIN(threshold_mb * _1MB, c.region_size);
+
+	src_addr = get_source_mapping(c);
+	if (!src_addr) {
+		ret = -1;
+		goto out;
+	}
+
+	/* Set byte pattern */
+	srand(pattern_seed);
+	for (i = 0; i < threshold; i++)
+		memset((char *) src_addr + i, (char) rand(), 1);
+
+	/* Mask to zero out lower bits of address for alignment */
+	align_mask = ~(c.dest_alignment - 1);
+	/* Offset of destination address from the end of the source region */
+	offset = (c.overlapping) ? -c.dest_alignment : c.dest_alignment;
+	addr = (void *) (((unsigned long long) src_addr + c.region_size
+			  + offset) & align_mask);
+
+	/* See comment in get_source_mapping() */
+	if (!((unsigned long long) addr & c.dest_alignment))
+		addr = (void *) ((unsigned long long) addr | c.dest_alignment);
+
+	clock_gettime(CLOCK_MONOTONIC, &t_start);
+	dest_addr = mremap(src_addr, c.region_size, c.region_size,
+			MREMAP_MAYMOVE|MREMAP_FIXED, (char *) addr);
+	clock_gettime(CLOCK_MONOTONIC, &t_end);
+
+	if (dest_addr == MAP_FAILED) {
+		ksft_print_msg("mremap failed: %s\n", strerror(errno));
+		ret = -1;
+		goto clean_up_src;
+	}
+
+	/* Verify byte pattern after remapping */
+	srand(pattern_seed);
+	for (i = 0; i < threshold; i++) {
+		char c = (char) rand();
+
+		if (((char *) dest_addr)[i] != c) {
+			ksft_print_msg("Data after remap doesn't match at offset %d\n",
+				       i);
+			ksft_print_msg("Expected: %#x\t Got: %#x\n", c & 0xff,
+					((char *) dest_addr)[i] & 0xff);
+			ret = -1;
+			goto clean_up_dest;
+		}
+	}
+
+	start_ns = t_start.tv_sec * NS_PER_SEC + t_start.tv_nsec;
+	end_ns = t_end.tv_sec * NS_PER_SEC + t_end.tv_nsec;
+	ret = end_ns - start_ns;
+
+/*
+ * Since the destination address is specified using MREMAP_FIXED, subsequent
+ * mremap will unmap any previous mapping at the address range specified by
+ * dest_addr and region_size. This significantly affects the remap time of
+ * subsequent tests. So we clean up mappings after each test.
+ */
+clean_up_dest:
+	munmap(dest_addr, c.region_size);
+clean_up_src:
+	munmap(src_addr, c.region_size);
+out:
+	return ret;
+}
+
+static void run_mremap_test_case(struct test test_case, int *failures,
+				 unsigned int threshold_mb,
+				 unsigned int pattern_seed)
+{
+	long long remap_time = remap_region(test_case.config, threshold_mb,
+					    pattern_seed);
+
+	if (remap_time < 0) {
+		if (test_case.expect_failure)
+			ksft_test_result_pass("%s\n\tExpected mremap failure\n",
+					      test_case.name);
+		else {
+			ksft_test_result_fail("%s\n", test_case.name);
+			*failures += 1;
+		}
+	} else {
+		/*
+		 * Comparing mremap time is only applicable if entire region
+		 * was faulted in.
+		 */
+		if (threshold_mb == VALIDATION_NO_THRESHOLD ||
+		    test_case.config.region_size <= threshold_mb * _1MB)
+			ksft_test_result_pass("%s\n\tmremap time: %12lldns\n",
+					      test_case.name, remap_time);
+		else
+			ksft_test_result_pass("%s\n", test_case.name);
+	}
+}
+
+static void usage(const char *cmd)
+{
+	fprintf(stderr,
+		"Usage: %s [[-t <threshold_mb>] [-p <pattern_seed>]]\n"
+		"-t\t only validate threshold_mb of the remapped region\n"
+		"  \t if 0 is supplied no threshold is used; all tests\n"
+		"  \t are run and remapped regions validated fully.\n"
+		"  \t The default threshold used is 4MB.\n"
+		"-p\t provide a seed to generate the random pattern for\n"
+		"  \t validating the remapped region.\n", cmd);
+}
+
+static int parse_args(int argc, char **argv, unsigned int *threshold_mb,
+		      unsigned int *pattern_seed)
+{
+	const char *optstr = "t:p:";
+	int opt;
+
+	while ((opt = getopt(argc, argv, optstr)) != -1) {
+		switch (opt) {
+		case 't':
+			*threshold_mb = atoi(optarg);
+			break;
+		case 'p':
+			*pattern_seed = atoi(optarg);
+			break;
+		default:
+			usage(argv[0]);
+			return -1;
+		}
+	}
+
+	if (optind < argc) {
+		usage(argv[0]);
+		return -1;
+	}
+
+	return 0;
+}
+
+int main(int argc, char **argv)
+{
+	int failures = 0;
+	int i, run_perf_tests;
+	unsigned int threshold_mb = VALIDATION_DEFAULT_THRESHOLD;
+	unsigned int pattern_seed;
+	time_t t;
+
+	pattern_seed = (unsigned int) time(&t);
+
+	if (parse_args(argc, argv, &threshold_mb, &pattern_seed) < 0)
+		exit(EXIT_FAILURE);
+
+	ksft_print_msg("Test configs:\n\tthreshold_mb=%u\n\tpattern_seed=%u\n\n",
+		       threshold_mb, pattern_seed);
+
+	struct test test_cases[] = {
+		/* Expected mremap failures */
+		MAKE_TEST(_4KB, _4KB, _4KB, OVERLAPPING, EXPECT_FAILURE,
+		  "mremap - Source and Destination Regions Overlapping"),
+		MAKE_TEST(_4KB, _1KB, _4KB, NON_OVERLAPPING, EXPECT_FAILURE,
+		  "mremap - Destination Address Misaligned (1KB-aligned)"),
+		MAKE_TEST(_1KB, _4KB, _4KB, NON_OVERLAPPING, EXPECT_FAILURE,
+		  "mremap - Source Address Misaligned (1KB-aligned)"),
+
+		/* Src addr PTE aligned */
+		MAKE_TEST(PTE, PTE, _8KB, NON_OVERLAPPING, EXPECT_SUCCESS,
+		  "8KB mremap - Source PTE-aligned, Destination PTE-aligned"),
+
+		/* Src addr 1MB aligned */
+		MAKE_TEST(_1MB, PTE, _2MB, NON_OVERLAPPING, EXPECT_SUCCESS,
+		  "2MB mremap - Source 1MB-aligned, Destination PTE-aligned"),
+		MAKE_TEST(_1MB, _1MB, _2MB, NON_OVERLAPPING, EXPECT_SUCCESS,
+		  "2MB mremap - Source 1MB-aligned, Destination 1MB-aligned"),
+
+		/* Src addr PMD aligned */
+		MAKE_TEST(PMD, PTE, _4MB, NON_OVERLAPPING, EXPECT_SUCCESS,
+		  "4MB mremap - Source PMD-aligned, Destination PTE-aligned"),
+		MAKE_TEST(PMD, _1MB, _4MB, NON_OVERLAPPING, EXPECT_SUCCESS,
+		  "4MB mremap - Source PMD-aligned, Destination 1MB-aligned"),
+		MAKE_TEST(PMD, PMD, _4MB, NON_OVERLAPPING, EXPECT_SUCCESS,
+		  "4MB mremap - Source PMD-aligned, Destination PMD-aligned"),
+
+		/* Src addr PUD aligned */
+		MAKE_TEST(PUD, PTE, _2GB, NON_OVERLAPPING, EXPECT_SUCCESS,
+		  "2GB mremap - Source PUD-aligned, Destination PTE-aligned"),
+		MAKE_TEST(PUD, _1MB, _2GB, NON_OVERLAPPING, EXPECT_SUCCESS,
+		  "2GB mremap - Source PUD-aligned, Destination 1MB-aligned"),
+		MAKE_TEST(PUD, PMD, _2GB, NON_OVERLAPPING, EXPECT_SUCCESS,
+		  "2GB mremap - Source PUD-aligned, Destination PMD-aligned"),
+		MAKE_TEST(PUD, PUD, _2GB, NON_OVERLAPPING, EXPECT_SUCCESS,
+		  "2GB mremap - Source PUD-aligned, Destination PUD-aligned"),
+	};
+
+	struct test perf_test_cases[] = {
+		/*
+		 * mremap 1GB region - Page table level aligned time
+		 * comparison.
+		 */
+		MAKE_TEST(PTE, PTE, _1GB, NON_OVERLAPPING, EXPECT_SUCCESS,
+		  "1GB mremap - Source PTE-aligned, Destination PTE-aligned"),
+		MAKE_TEST(PMD, PMD, _1GB, NON_OVERLAPPING, EXPECT_SUCCESS,
+		  "1GB mremap - Source PMD-aligned, Destination PMD-aligned"),
+		MAKE_TEST(PUD, PUD, _1GB, NON_OVERLAPPING, EXPECT_SUCCESS,
+		  "1GB mremap - Source PUD-aligned, Destination PUD-aligned"),
+	};
+
+	run_perf_tests =  (threshold_mb == VALIDATION_NO_THRESHOLD) ||
+				(threshold_mb * _1MB >= _1GB);
+
+	ksft_set_plan(ARRAY_SIZE(test_cases) + (run_perf_tests ?
+		      ARRAY_SIZE(perf_test_cases) : 0));
+
+	for (i = 0; i < ARRAY_SIZE(test_cases); i++)
+		run_mremap_test_case(test_cases[i], &failures, threshold_mb,
+				     pattern_seed);
+
+	if (run_perf_tests) {
+		ksft_print_msg("\n%s\n",
+		 "mremap HAVE_MOVE_PMD/PUD optimization time comparison for 1GB region:");
+		for (i = 0; i < ARRAY_SIZE(perf_test_cases); i++)
+			run_mremap_test_case(perf_test_cases[i], &failures,
+					     threshold_mb, pattern_seed);
+	}
+
+	if (failures > 0)
+		ksft_exit_fail();
+	else
+		ksft_exit_pass();
+}
diff --git a/tools/testing/selftests/vm/run_vmtests b/tools/testing/selftests/vm/run_vmtests
index a3f4f30f0a2e..d578ad831813 100755
--- a/tools/testing/selftests/vm/run_vmtests
+++ b/tools/testing/selftests/vm/run_vmtests
@@ -241,6 +241,17 @@ else
 	echo "[PASS]"
 fi
 
+echo "-------------------"
+echo "running mremap_test"
+echo "-------------------"
+./mremap_test
+if [ $? -ne 0 ]; then
+	echo "[FAIL]"
+	exitcode=1
+else
+	echo "[PASS]"
+fi
+
 echo "-----------------"
 echo "running thuge-gen"
 echo "-----------------"
-- 
2.28.0.806.g8561365e88-goog

[PATCH v3 2/5] arm64: mremap speedup - Enable HAVE_MOVE_PMD

[Kalesh Singh]

HAVE_MOVE_PMD enables remapping pages at the PMD level if both the
source and destination addresses are PMD-aligned.

HAVE_MOVE_PMD is already enabled on x86. The original patch [1] that
introduced this config did not enable it on arm64 at the time because
of performance issues with flushing the TLB on every PMD move. These
issues have since been addressed in more recent releases with
improvements to the arm64 TLB invalidation and core mmu_gather code as
Will Deacon mentioned in [2].

From the data below, it can be inferred that there is approximately
8x improvement in performance when HAVE_MOVE_PMD is enabled on arm64.

--------- Test Results ----------

The following results were obtained on an arm64 device running a 5.4
kernel, by remapping a PMD-aligned, 1GB sized region to a PMD-aligned
destination. The results from 10 iterations of the test are given below.
All times are in nanoseconds.

Control    HAVE_MOVE_PMD

9220833    1247761
9002552    1219896
9254115    1094792
8725885    1227760
9308646    1043698
9001667    1101771
8793385    1159896
8774636    1143594
9553125    1025833
9374010    1078125

9100885.4  1134312.6    <-- Mean Time in nanoseconds

Total mremap time for a 1GB sized PMD-aligned region drops from
~9.1 milliseconds to ~1.1 milliseconds. (~8x speedup).

[1] https://lore.kernel.org/r/20181108181201.88826-3-joelaf@google.com
[2] https://www.mail-archive.com/linuxppc-dev@lists.ozlabs.org/msg140837.html

Signed-off-by: Kalesh Singh <kaleshsingh@google.com>
---
 arch/arm64/Kconfig | 1 +
 1 file changed, 1 insertion(+)

diff --git a/arch/arm64/Kconfig b/arch/arm64/Kconfig
index 6d232837cbee..844d089668e3 100644
--- a/arch/arm64/Kconfig
+++ b/arch/arm64/Kconfig
@@ -121,6 +121,7 @@ config ARM64
 	select GENERIC_VDSO_TIME_NS
 	select HANDLE_DOMAIN_IRQ
 	select HARDIRQS_SW_RESEND
+	select HAVE_MOVE_PMD
 	select HAVE_PCI
 	select HAVE_ACPI_APEI if (ACPI && EFI)
 	select HAVE_ALIGNED_STRUCT_PAGE if SLUB
-- 
2.28.0.806.g8561365e88-goog

[PATCH v3 3/5] mm: Speedup mremap on 1GB or larger regions

[Kalesh Singh]

Android needs to move large memory regions for garbage collection.
The GC requires moving physical pages of multi-gigabyte heap
using mremap. During this move, the application threads have to
be paused for correctness. It is critical to keep this pause as
short as possible to avoid jitters during user interaction.

Optimize mremap for >= 1GB-sized regions by moving at the PUD/PGD
level if the source and destination addresses are PUD-aligned.
For CONFIG_PGTABLE_LEVELS == 3, moving at the PUD level in effect moves
PGD entries, since the PUD entry is “folded back” onto the PGD entry.
Add HAVE_MOVE_PUD so that architectures where moving at the PUD level
isn't supported/tested can turn this off by not selecting the config.

Fix build test error from v1 of this series reported by
kernel test robot in [1].

[1] https://lists.01.org/hyperkitty/list/kbuild-all@lists.01.org/thread/CKPGL4FH4NG7TGH2CVYX2UX76L25BTA3/

Signed-off-by: Kalesh Singh <kaleshsingh@google.com>
Reported-by: kernel test robot <lkp@intel.com>
---
Changes in v2:
  - Update commit message with description of Android GC's use case.
  - Move set_pud_at() to a separate patch.
  - Use switch() instead of ifs in move_pgt_entry()
  - Fix build test error reported by kernel test robot on x86_64 in [1].
    Guard move_huge_pmd() with IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE),
    since this section doesn't get optimized out in the kernel test
    robot's build test when HAVE_MOVE_PUD is enabled.
  - Keep WARN_ON_ONCE(1) instead of BUILD_BUG() for the aforementioned
    reason.

Changes in v3:
  - Move get_old_pud() and alloc_new_pud() out of
    #ifdef CONFIG_HAVE_MOVE_PUD.
  - Have get_old_pmd() and alloc_new_pmd() use get_old_pud() and
    alloc_old_pud().
  - Use switch() in get_extent() instead of ifs.
  - Add BUILD_BUG() to default case of get_extent().
  - Replace #ifdef CONFIG_HAVE_MOVE_PMD/PUD in move_page_tables() with
    IS_ENABLED(CONFIG_HAVE_MOVE_PMD/PUD).
  - Make lines 80 cols or less, where they don’t need to be longer.
  - s/=  /= /g (Fixed double spaces after '=').

 arch/Kconfig |   7 ++
 mm/mremap.c  | 230 ++++++++++++++++++++++++++++++++++++++++++---------
 2 files changed, 197 insertions(+), 40 deletions(-)

diff --git a/arch/Kconfig b/arch/Kconfig
index af14a567b493..5eabaa00bf9b 100644
--- a/arch/Kconfig
+++ b/arch/Kconfig
@@ -602,6 +602,13 @@ config HAVE_IRQ_TIME_ACCOUNTING
 	  Archs need to ensure they use a high enough resolution clock to
 	  support irq time accounting and then call enable_sched_clock_irqtime().
 
+config HAVE_MOVE_PUD
+	bool
+	help
+	  Architectures that select this are able to move page tables at the
+	  PUD level. If there are only 3 page table levels, the move effectively
+	  happens at the PGD level.
+
 config HAVE_MOVE_PMD
 	bool
 	help
diff --git a/mm/mremap.c b/mm/mremap.c
index 138abbae4f75..078f731277b6 100644
--- a/mm/mremap.c
+++ b/mm/mremap.c
@@ -30,12 +30,11 @@
 
 #include "internal.h"
 
-static pmd_t *get_old_pmd(struct mm_struct *mm, unsigned long addr)
+static pud_t *get_old_pud(struct mm_struct *mm, unsigned long addr)
 {
 	pgd_t *pgd;
 	p4d_t *p4d;
 	pud_t *pud;
-	pmd_t *pmd;
 
 	pgd = pgd_offset(mm, addr);
 	if (pgd_none_or_clear_bad(pgd))
@@ -49,6 +48,18 @@ static pmd_t *get_old_pmd(struct mm_struct *mm, unsigned long addr)
 	if (pud_none_or_clear_bad(pud))
 		return NULL;
 
+	return pud;
+}
+
+static pmd_t *get_old_pmd(struct mm_struct *mm, unsigned long addr)
+{
+	pud_t *pud;
+	pmd_t *pmd;
+
+	pud = get_old_pud(mm, addr);
+	if (!pud)
+		return NULL;
+
 	pmd = pmd_offset(pud, addr);
 	if (pmd_none(*pmd))
 		return NULL;
@@ -56,19 +67,27 @@ static pmd_t *get_old_pmd(struct mm_struct *mm, unsigned long addr)
 	return pmd;
 }
 
-static pmd_t *alloc_new_pmd(struct mm_struct *mm, struct vm_area_struct *vma,
+static pud_t *alloc_new_pud(struct mm_struct *mm, struct vm_area_struct *vma,
 			    unsigned long addr)
 {
 	pgd_t *pgd;
 	p4d_t *p4d;
-	pud_t *pud;
-	pmd_t *pmd;
 
 	pgd = pgd_offset(mm, addr);
 	p4d = p4d_alloc(mm, pgd, addr);
 	if (!p4d)
 		return NULL;
-	pud = pud_alloc(mm, p4d, addr);
+
+	return pud_alloc(mm, p4d, addr);
+}
+
+static pmd_t *alloc_new_pmd(struct mm_struct *mm, struct vm_area_struct *vma,
+			    unsigned long addr)
+{
+	pud_t *pud;
+	pmd_t *pmd;
+
+	pud = alloc_new_pud(mm, vma, addr);
 	if (!pud)
 		return NULL;
 
@@ -249,14 +268,148 @@ static bool move_normal_pmd(struct vm_area_struct *vma, unsigned long old_addr,
 
 	return true;
 }
+#else
+static inline bool move_normal_pmd(struct vm_area_struct *vma,
+		unsigned long old_addr, unsigned long new_addr, pmd_t *old_pmd,
+		pmd_t *new_pmd)
+{
+	return false;
+}
 #endif
 
+#ifdef CONFIG_HAVE_MOVE_PUD
+static bool move_normal_pud(struct vm_area_struct *vma, unsigned long old_addr,
+		  unsigned long new_addr, pud_t *old_pud, pud_t *new_pud)
+{
+	spinlock_t *old_ptl, *new_ptl;
+	struct mm_struct *mm = vma->vm_mm;
+	pud_t pud;
+
+	/*
+	 * The destination pud shouldn't be established, free_pgtables()
+	 * should have released it.
+	 */
+	if (WARN_ON_ONCE(!pud_none(*new_pud)))
+		return false;
+
+	/*
+	 * We don't have to worry about the ordering of src and dst
+	 * ptlocks because exclusive mmap_lock prevents deadlock.
+	 */
+	old_ptl = pud_lock(vma->vm_mm, old_pud);
+	new_ptl = pud_lockptr(mm, new_pud);
+	if (new_ptl != old_ptl)
+		spin_lock_nested(new_ptl, SINGLE_DEPTH_NESTING);
+
+	/* Clear the pud */
+	pud = *old_pud;
+	pud_clear(old_pud);
+
+	VM_BUG_ON(!pud_none(*new_pud));
+
+	/* Set the new pud */
+	set_pud_at(mm, new_addr, new_pud, pud);
+	flush_tlb_range(vma, old_addr, old_addr + PUD_SIZE);
+	if (new_ptl != old_ptl)
+		spin_unlock(new_ptl);
+	spin_unlock(old_ptl);
+
+	return true;
+}
+#else
+static inline bool move_normal_pud(struct vm_area_struct *vma,
+		unsigned long old_addr, unsigned long new_addr, pud_t *old_pud,
+		pud_t *new_pud)
+{
+	return false;
+}
+#endif
+
+enum pgt_entry {
+	NORMAL_PMD,
+	HPAGE_PMD,
+	NORMAL_PUD,
+};
+
+/*
+ * Returns an extent of the corresponding size for the pgt_entry specified if
+ * valid. Else returns a smaller extent bounded by the end of the source and
+ * destination pgt_entry.
+ */
+static unsigned long get_extent(enum pgt_entry entry, unsigned long old_addr,
+			unsigned long old_end, unsigned long new_addr)
+{
+	unsigned long next, extent, mask, size;
+
+	switch (entry) {
+	case HPAGE_PMD:
+	case NORMAL_PMD:
+		mask = PMD_MASK;
+		size = PMD_SIZE;
+		break;
+	case NORMAL_PUD:
+		mask = PUD_MASK;
+		size = PUD_SIZE;
+		break;
+	default:
+		BUILD_BUG();
+		break;
+	}
+
+	next = (old_addr + size) & mask;
+	/* even if next overflowed, extent below will be ok */
+	extent = (next > old_end) ? old_end - old_addr : next - old_addr;
+	next = (new_addr + size) & mask;
+	if (extent > next - new_addr)
+		extent = next - new_addr;
+	return extent;
+}
+
+/*
+ * Attempts to speedup the move by moving entry at the level corresponding to
+ * pgt_entry. Returns true if the move was successful, else false.
+ */
+static bool move_pgt_entry(enum pgt_entry entry, struct vm_area_struct *vma,
+			unsigned long old_addr, unsigned long new_addr,
+			void *old_entry, void *new_entry, bool need_rmap_locks)
+{
+	bool moved = false;
+
+	/* See comment in move_ptes() */
+	if (need_rmap_locks)
+		take_rmap_locks(vma);
+
+	switch (entry) {
+	case NORMAL_PMD:
+		moved = move_normal_pmd(vma, old_addr, new_addr, old_entry,
+					new_entry);
+		break;
+	case NORMAL_PUD:
+		moved = move_normal_pud(vma, old_addr, new_addr, old_entry,
+					new_entry);
+		break;
+	case HPAGE_PMD:
+		moved = IS_ENABLED(CONFIG_TRANSPARENT_HUGEPAGE) &&
+			move_huge_pmd(vma, old_addr, new_addr, old_entry,
+				      new_entry);
+		break;
+	default:
+		WARN_ON_ONCE(1);
+		break;
+	}
+
+	if (need_rmap_locks)
+		drop_rmap_locks(vma);
+
+	return moved;
+}
+
 unsigned long move_page_tables(struct vm_area_struct *vma,
 		unsigned long old_addr, struct vm_area_struct *new_vma,
 		unsigned long new_addr, unsigned long len,
 		bool need_rmap_locks)
 {
-	unsigned long extent, next, old_end;
+	unsigned long extent, old_end;
 	struct mmu_notifier_range range;
 	pmd_t *old_pmd, *new_pmd;
 
@@ -269,53 +422,50 @@ unsigned long move_page_tables(struct vm_area_struct *vma,
 
 	for (; old_addr < old_end; old_addr += extent, new_addr += extent) {
 		cond_resched();
-		next = (old_addr + PMD_SIZE) & PMD_MASK;
-		/* even if next overflowed, extent below will be ok */
-		extent = next - old_addr;
-		if (extent > old_end - old_addr)
-			extent = old_end - old_addr;
-		next = (new_addr + PMD_SIZE) & PMD_MASK;
-		if (extent > next - new_addr)
-			extent = next - new_addr;
+		/*
+		 * If extent is PUD-sized try to speed up the move by moving at the
+		 * PUD level if possible.
+		 */
+		extent = get_extent(NORMAL_PUD, old_addr, old_end, new_addr);
+		if (IS_ENABLED(CONFIG_HAVE_MOVE_PUD) && extent == PUD_SIZE) {
+			pud_t *old_pud, *new_pud;
+
+			old_pud = get_old_pud(vma->vm_mm, old_addr);
+			if (!old_pud)
+				continue;
+			new_pud = alloc_new_pud(vma->vm_mm, vma, new_addr);
+			if (!new_pud)
+				break;
+			if (move_pgt_entry(NORMAL_PUD, vma, old_addr, new_addr,
+					   old_pud, new_pud, need_rmap_locks))
+				continue;
+		}
+
+		extent = get_extent(NORMAL_PMD, old_addr, old_end, new_addr);
 		old_pmd = get_old_pmd(vma->vm_mm, old_addr);
 		if (!old_pmd)
 			continue;
 		new_pmd = alloc_new_pmd(vma->vm_mm, vma, new_addr);
 		if (!new_pmd)
 			break;
-		if (is_swap_pmd(*old_pmd) || pmd_trans_huge(*old_pmd) || pmd_devmap(*old_pmd)) {
-			if (extent == HPAGE_PMD_SIZE) {
-				bool moved;
-				/* See comment in move_ptes() */
-				if (need_rmap_locks)
-					take_rmap_locks(vma);
-				moved = move_huge_pmd(vma, old_addr, new_addr,
-						      old_pmd, new_pmd);
-				if (need_rmap_locks)
-					drop_rmap_locks(vma);
-				if (moved)
-					continue;
-			}
+		if (is_swap_pmd(*old_pmd) || pmd_trans_huge(*old_pmd) ||
+		    pmd_devmap(*old_pmd)) {
+			if (extent == HPAGE_PMD_SIZE &&
+			    move_pgt_entry(HPAGE_PMD, vma, old_addr, new_addr,
+					   old_pmd, new_pmd, need_rmap_locks))
+				continue;
 			split_huge_pmd(vma, old_pmd, old_addr);
 			if (pmd_trans_unstable(old_pmd))
 				continue;
-		} else if (extent == PMD_SIZE) {
-#ifdef CONFIG_HAVE_MOVE_PMD
+		} else if (IS_ENABLED(CONFIG_HAVE_MOVE_PMD) &&
+			   extent == PMD_SIZE) {
 			/*
 			 * If the extent is PMD-sized, try to speed the move by
 			 * moving at the PMD level if possible.
 			 */
-			bool moved;
-
-			if (need_rmap_locks)
-				take_rmap_locks(vma);
-			moved = move_normal_pmd(vma, old_addr, new_addr,
-						old_pmd, new_pmd);
-			if (need_rmap_locks)
-				drop_rmap_locks(vma);
-			if (moved)
+			if (move_pgt_entry(NORMAL_PMD, vma, old_addr, new_addr,
+					   old_pmd, new_pmd, need_rmap_locks))
 				continue;
-#endif
 		}
 
 		if (pte_alloc(new_vma->vm_mm, new_pmd))
-- 
2.28.0.806.g8561365e88-goog

[PATCH v3 4/5] arm64: mremap speedup - Enable HAVE_MOVE_PUD

[Kalesh Singh]

HAVE_MOVE_PUD enables remapping pages at the PUD level if both the
source and destination addresses are PUD-aligned.

With HAVE_MOVE_PUD enabled it can be inferred that there is approximately
a 19x improvement in performance on arm64. (See data below).

------- Test Results ---------

The following results were obtained using a 5.4 kernel, by remapping
a PUD-aligned, 1GB sized region to a PUD-aligned destination.
The results from 10 iterations of the test are given below:

Total mremap times for 1GB data on arm64. All times are in nanoseconds.

Control          HAVE_MOVE_PUD

1247761          74271
1219896          46771
1094792          59687
1227760          48385
1043698          76666
1101771          50365
1159896          52500
1143594          75261
1025833          61354
1078125          48697

1134312.6        59395.7    <-- Mean time in nanoseconds

A 1GB mremap completion time drops from ~1.1 milliseconds
to ~59 microseconds on arm64. (~19x speed up).

Signed-off-by: Kalesh Singh <kaleshsingh@google.com>
---
Change in v3:
  - Add set_pud_at() macro - Used by move_normal_pud().

 arch/arm64/Kconfig               | 1 +
 arch/arm64/include/asm/pgtable.h | 1 +
 2 files changed, 2 insertions(+)

diff --git a/arch/arm64/Kconfig b/arch/arm64/Kconfig
index 844d089668e3..4d521f0a5863 100644
--- a/arch/arm64/Kconfig
+++ b/arch/arm64/Kconfig
@@ -122,6 +122,7 @@ config ARM64
 	select HANDLE_DOMAIN_IRQ
 	select HARDIRQS_SW_RESEND
 	select HAVE_MOVE_PMD
+	select HAVE_MOVE_PUD
 	select HAVE_PCI
 	select HAVE_ACPI_APEI if (ACPI && EFI)
 	select HAVE_ALIGNED_STRUCT_PAGE if SLUB
diff --git a/arch/arm64/include/asm/pgtable.h b/arch/arm64/include/asm/pgtable.h
index d5d3fbe73953..8848125e3024 100644
--- a/arch/arm64/include/asm/pgtable.h
+++ b/arch/arm64/include/asm/pgtable.h
@@ -415,6 +415,7 @@ static inline pmd_t pmd_mkdevmap(pmd_t pmd)
 #define pfn_pud(pfn,prot)	__pud(__phys_to_pud_val((phys_addr_t)(pfn) << PAGE_SHIFT) | pgprot_val(prot))
 
 #define set_pmd_at(mm, addr, pmdp, pmd)	set_pte_at(mm, addr, (pte_t *)pmdp, pmd_pte(pmd))
+#define set_pud_at(mm, addr, pudp, pud)	set_pte_at(mm, addr, (pte_t *)pudp, pud_pte(pud))
 
 #define __p4d_to_phys(p4d)	__pte_to_phys(p4d_pte(p4d))
 #define __phys_to_p4d_val(phys)	__phys_to_pte_val(phys)
-- 
2.28.0.806.g8561365e88-goog

[PATCH v3 5/5] x86: mremap speedup - Enable HAVE_MOVE_PUD

[Kalesh Singh]

HAVE_MOVE_PUD enables remapping pages at the PUD level if both the
source and destination addresses are PUD-aligned.

With HAVE_MOVE_PUD enabled it can be inferred that there is approximately
a 13x improvement in performance on x86. (See data below).

------- Test Results ---------

The following results were obtained using a 5.4 kernel, by remapping
a PUD-aligned, 1GB sized region to a PUD-aligned destination.
The results from 10 iterations of the test are given below:

Total mremap times for 1GB data on x86. All times are in nanoseconds.

Control        HAVE_MOVE_PUD

180394         15089
235728         14056
238931         25741
187330         13838
241742         14187
177925         14778
182758         14728
160872         14418
205813         15107
245722         13998

205721.5       15594    <-- Mean time in nanoseconds

A 1GB mremap completion time drops from ~205 microseconds
to ~15 microseconds on x86. (~13x speed up).

Signed-off-by: Kalesh Singh <kaleshsingh@google.com>
---
 arch/x86/Kconfig | 1 +
 1 file changed, 1 insertion(+)

diff --git a/arch/x86/Kconfig b/arch/x86/Kconfig
index 7101ac64bb20..ff6e2755cab8 100644
--- a/arch/x86/Kconfig
+++ b/arch/x86/Kconfig
@@ -198,6 +198,7 @@ config X86
 	select HAVE_MIXED_BREAKPOINTS_REGS
 	select HAVE_MOD_ARCH_SPECIFIC
 	select HAVE_MOVE_PMD
+	select HAVE_MOVE_PUD
 	select HAVE_NMI
 	select HAVE_OPROFILE
 	select HAVE_OPTPROBES
-- 
2.28.0.806.g8561365e88-goog