| 1 | /* |
| 2 | * Copyright (C) 2023 Michael Jeanson <mjeanson@efficios.com> |
| 3 | * |
| 4 | * SPDX-License-Identifier: LGPL-2.1-only |
| 5 | * |
| 6 | */ |
| 7 | |
| 8 | #include <common/error.hpp> |
| 9 | #include <common/io-hint.hpp> |
| 10 | #include <common/scope-exit.hpp> |
| 11 | |
| 12 | #include <cinttypes> |
| 13 | #include <fcntl.h> |
| 14 | #include <sys/mman.h> |
| 15 | #include <unistd.h> |
| 16 | |
| 17 | /* |
| 18 | * Use sync_file_range when available. |
| 19 | */ |
| 20 | #ifdef HAVE_SYNC_FILE_RANGE |
| 21 | |
| 22 | #include <fcntl.h> |
| 23 | |
| 24 | namespace { |
| 25 | int flush_range(int fd, off_t offset, off_t nbytes, unsigned int flags) |
| 26 | { |
| 27 | int ret; |
| 28 | |
| 29 | ret = sync_file_range(fd, offset, nbytes, flags); |
| 30 | if (ret) { |
| 31 | PERROR("Failed to sync file range: fd=%i, offset=%" PRIu64 ", nbytes=%" PRIu64 |
| 32 | ", flags=%i", |
| 33 | fd, |
| 34 | static_cast<uint64_t>(offset), |
| 35 | static_cast<uint64_t>(nbytes), |
| 36 | flags); |
| 37 | } |
| 38 | |
| 39 | return ret; |
| 40 | } |
| 41 | |
| 42 | int flush_range_sync(int fd, off_t offset, off_t nbytes) |
| 43 | { |
| 44 | return flush_range(fd, |
| 45 | offset, |
| 46 | nbytes, |
| 47 | SYNC_FILE_RANGE_WAIT_BEFORE | SYNC_FILE_RANGE_WRITE | |
| 48 | SYNC_FILE_RANGE_WAIT_AFTER); |
| 49 | } |
| 50 | |
| 51 | int flush_range_async(int fd, off_t offset, off_t nbytes) |
| 52 | { |
| 53 | return flush_range(fd, offset, nbytes, SYNC_FILE_RANGE_WRITE); |
| 54 | } |
| 55 | } /* namespace */ |
| 56 | |
| 57 | #else |
| 58 | |
| 59 | namespace { |
| 60 | /* |
| 61 | * Use a memory mapping with msync() to emulate sync_file_range(). |
| 62 | */ |
| 63 | int flush_range(int fd, off_t offset, off_t nbytes, int flags) |
| 64 | { |
| 65 | void *mapped_region = mmap(NULL, nbytes, PROT_READ | PROT_WRITE, MAP_SHARED, fd, offset); |
| 66 | if (mapped_region == MAP_FAILED) { |
| 67 | PERROR("Failed to mmap region to flush range: fd=%i, offset=%" PRIu64 |
| 68 | ", nbytes=%" PRIu64 ", flags=%i", |
| 69 | fd, |
| 70 | static_cast<uint64_t>(offset), |
| 71 | static_cast<uint64_t>(nbytes), |
| 72 | flags); |
| 73 | return -1; |
| 74 | } |
| 75 | |
| 76 | const auto munmap_on_exit = lttng::make_scope_exit([&]() noexcept { |
| 77 | const auto munmap_ret = munmap(mapped_region, nbytes); |
| 78 | if (munmap_ret) { |
| 79 | PERROR("Failed to munmap region while flushing range: fd=%i, offset=%" PRIu64 |
| 80 | ", nbytes=%" PRIu64 ", flags=%i", |
| 81 | fd, |
| 82 | static_cast<uint64_t>(offset), |
| 83 | static_cast<uint64_t>(nbytes), |
| 84 | flags); |
| 85 | } |
| 86 | }); |
| 87 | |
| 88 | const auto msync_ret = msync(mapped_region, nbytes, flags); |
| 89 | if (msync_ret) { |
| 90 | PERROR("Failed to msync region while flushing range: fd=%i, offset=%" PRIu64 |
| 91 | ", nbytes=%" PRIu64 ", flags=%i", |
| 92 | fd, |
| 93 | static_cast<uint64_t>(offset), |
| 94 | static_cast<uint64_t>(nbytes), |
| 95 | flags); |
| 96 | return -1; |
| 97 | } |
| 98 | |
| 99 | return 0; |
| 100 | } |
| 101 | |
| 102 | int flush_range_sync(int fd, off_t offset, off_t nbytes) |
| 103 | { |
| 104 | return flush_range(fd, offset, nbytes, MS_SYNC); |
| 105 | } |
| 106 | |
| 107 | int flush_range_async(int fd, off_t offset, off_t nbytes) |
| 108 | { |
| 109 | return flush_range(fd, offset, nbytes, MS_ASYNC); |
| 110 | } |
| 111 | } /* namespace */ |
| 112 | #endif |
| 113 | |
| 114 | /* |
| 115 | * Give a hint to the kernel that we won't need the data at the specified range |
| 116 | * so it can be dropped from the page cache and wait for it to be flushed to |
| 117 | * disk. |
| 118 | */ |
| 119 | void lttng::io::hint_flush_range_dont_need_sync(int fd, off_t offset, off_t nbytes) |
| 120 | { |
| 121 | /* Waited for the page writeback to complete. */ |
| 122 | flush_range_sync(fd, offset, nbytes); |
| 123 | |
| 124 | /* |
| 125 | * Give hints to the kernel about how we access the file: |
| 126 | * POSIX_FADV_DONTNEED : we won't re-access data in a near future after |
| 127 | * we write it. |
| 128 | * |
| 129 | * We need to call fadvise again after the file grows because the |
| 130 | * kernel does not seem to apply fadvise to non-existing parts of the |
| 131 | * file. |
| 132 | * |
| 133 | * Call fadvise _after_ having waited for the page writeback to |
| 134 | * complete because the dirty page writeback semantic is not well |
| 135 | * defined. So it can be expected to lead to lower throughput in |
| 136 | * streaming. |
| 137 | */ |
| 138 | const int ret = posix_fadvise(fd, offset, nbytes, POSIX_FADV_DONTNEED); |
| 139 | if (ret && ret != -ENOSYS) { |
| 140 | PERROR("Failed to mark region as DONTNEED with posix_fadvise: fd=%i, offset=%" PRIu64 |
| 141 | ", nbytes=%" PRIu64, |
| 142 | fd, |
| 143 | static_cast<uint64_t>(offset), |
| 144 | static_cast<uint64_t>(nbytes)); |
| 145 | errno = ret; |
| 146 | } |
| 147 | } |
| 148 | |
| 149 | /* |
| 150 | * Give a hint to the kernel that the data at the specified range should be |
| 151 | * flushed to disk and wait for it to complete. |
| 152 | */ |
| 153 | void lttng::io::hint_flush_range_sync(int fd, off_t offset, off_t nbytes) |
| 154 | { |
| 155 | flush_range_sync(fd, offset, nbytes); |
| 156 | } |
| 157 | |
| 158 | /* |
| 159 | * Give a hint to the kernel that the data at the specified range should be |
| 160 | * flushed to disk and return immediatly. |
| 161 | */ |
| 162 | void lttng::io::hint_flush_range_async(int fd, off_t offset, off_t nbytes) |
| 163 | { |
| 164 | flush_range_async(fd, offset, nbytes); |
| 165 | } |