|
| 1 | +/* SPDX-License-Identifier: MIT */ |
| 2 | +/* |
| 3 | + * Description: Ensure that an io-wq worker created by io_uring doesn't linger |
| 4 | + * in the task thread group after the last ring is closed. |
| 5 | + * |
| 6 | + * This matters for checkpoint/restore of long-running processes which quiesce |
| 7 | + * and close their io_uring instance(s) before dumping process state. |
| 8 | + */ |
| 9 | +#include <dirent.h> |
| 10 | +#include <errno.h> |
| 11 | +#include <fcntl.h> |
| 12 | +#include <limits.h> |
| 13 | +#include <stdio.h> |
| 14 | +#include <stdlib.h> |
| 15 | +#include <string.h> |
| 16 | +#include <sys/time.h> |
| 17 | +#include <unistd.h> |
| 18 | + |
| 19 | +#include "liburing.h" |
| 20 | +#include "helpers.h" |
| 21 | + |
| 22 | +#define IOWQ_COMM_PREFIX "iou-wrk-" |
| 23 | + |
| 24 | +static int count_iowq_workers(void) |
| 25 | +{ |
| 26 | + DIR *dir; |
| 27 | + struct dirent *de; |
| 28 | + int nr = 0; |
| 29 | + |
| 30 | + dir = opendir("/proc/self/task"); |
| 31 | + if (!dir) |
| 32 | + return -errno; |
| 33 | + |
| 34 | + while ((de = readdir(dir)) != NULL) { |
| 35 | + char path[PATH_MAX]; |
| 36 | + char comm[64]; |
| 37 | + int fd, ret; |
| 38 | + |
| 39 | + if (de->d_name[0] == '.') |
| 40 | + continue; |
| 41 | + |
| 42 | + snprintf(path, sizeof(path), "/proc/self/task/%s/comm", de->d_name); |
| 43 | + fd = open(path, O_RDONLY); |
| 44 | + if (fd < 0) |
| 45 | + continue; |
| 46 | + |
| 47 | + ret = read(fd, comm, sizeof(comm) - 1); |
| 48 | + close(fd); |
| 49 | + if (ret <= 0) |
| 50 | + continue; |
| 51 | + |
| 52 | + comm[ret] = '\0'; |
| 53 | + if (!strncmp(comm, IOWQ_COMM_PREFIX, strlen(IOWQ_COMM_PREFIX))) |
| 54 | + nr++; |
| 55 | + } |
| 56 | + |
| 57 | + closedir(dir); |
| 58 | + return nr; |
| 59 | +} |
| 60 | + |
| 61 | +static int wait_for_iowq_workers(bool want_any, unsigned long timeout_ms) |
| 62 | +{ |
| 63 | + struct timeval start; |
| 64 | + int nr = 0; |
| 65 | + |
| 66 | + gettimeofday(&start, NULL); |
| 67 | + do { |
| 68 | + nr = count_iowq_workers(); |
| 69 | + if (nr < 0) { |
| 70 | + fprintf(stderr, "count_iowq_workers: %s\n", strerror(-nr)); |
| 71 | + return T_EXIT_SKIP; |
| 72 | + } |
| 73 | + |
| 74 | + if (want_any) { |
| 75 | + if (nr > 0) |
| 76 | + return T_EXIT_PASS; |
| 77 | + } else { |
| 78 | + if (!nr) |
| 79 | + return T_EXIT_PASS; |
| 80 | + } |
| 81 | + |
| 82 | + usleep(10 * 1000); |
| 83 | + } while (mtime_since_now(&start) < timeout_ms); |
| 84 | + |
| 85 | + fprintf(stderr, "timeout waiting for iou-wrk threads (want_any=%d, nr=%d)\n", |
| 86 | + want_any, nr); |
| 87 | + return T_EXIT_FAIL; |
| 88 | +} |
| 89 | + |
| 90 | +int main(int argc, char *argv[]) |
| 91 | +{ |
| 92 | + struct io_uring ring; |
| 93 | + struct io_uring_sqe *sqe; |
| 94 | + struct io_uring_cqe *cqe; |
| 95 | + bool ring_inited = false; |
| 96 | + bool ring_exited = false; |
| 97 | + char src_fname[PATH_MAX]; |
| 98 | + char dst_fname[PATH_MAX]; |
| 99 | + char buf[32]; |
| 100 | + int pipe_fds[2] = { -1, -1 }; |
| 101 | + int fd_src = -1, fd_dst = -1; |
| 102 | + int ret; |
| 103 | + |
| 104 | + if (argc > 1) |
| 105 | + return T_EXIT_SKIP; |
| 106 | + |
| 107 | + if (t_create_ring(4, &ring, 0) != T_SETUP_OK) |
| 108 | + return T_EXIT_SKIP; |
| 109 | + ring_inited = true; |
| 110 | + |
| 111 | + snprintf(src_fname, sizeof(src_fname), ".iowq_unused_exit.%d.src", getpid()); |
| 112 | + fd_src = open(src_fname, O_RDWR | O_CREAT | O_TRUNC, 0644); |
| 113 | + if (fd_src < 0) { |
| 114 | + perror("open source"); |
| 115 | + ret = T_EXIT_SKIP; |
| 116 | + goto out; |
| 117 | + } |
| 118 | + |
| 119 | + memset(buf, 0xab, sizeof(buf)); |
| 120 | + if (write(fd_src, buf, sizeof(buf)) != sizeof(buf)) { |
| 121 | + perror("write source"); |
| 122 | + ret = T_EXIT_FAIL; |
| 123 | + goto out; |
| 124 | + } |
| 125 | + |
| 126 | + snprintf(dst_fname, sizeof(dst_fname), ".iowq_unused_exit.%d.dst", getpid()); |
| 127 | + fd_dst = open(dst_fname, O_RDWR | O_CREAT | O_TRUNC, 0644); |
| 128 | + if (fd_dst < 0) { |
| 129 | + perror("open destination"); |
| 130 | + ret = T_EXIT_SKIP; |
| 131 | + goto out; |
| 132 | + } |
| 133 | + |
| 134 | + if (pipe(pipe_fds) < 0) { |
| 135 | + perror("pipe"); |
| 136 | + ret = T_EXIT_FAIL; |
| 137 | + goto out; |
| 138 | + } |
| 139 | + |
| 140 | + sqe = io_uring_get_sqe(&ring); |
| 141 | + io_uring_prep_splice(sqe, fd_src, 0, pipe_fds[1], -1, sizeof(buf), 0); |
| 142 | + sqe->user_data = 1; |
| 143 | + |
| 144 | + sqe = io_uring_get_sqe(&ring); |
| 145 | + io_uring_prep_splice(sqe, pipe_fds[0], -1, fd_dst, 0, sizeof(buf), 0); |
| 146 | + sqe->user_data = 2; |
| 147 | + |
| 148 | + ret = io_uring_submit(&ring); |
| 149 | + if (ret < 0) { |
| 150 | + fprintf(stderr, "io_uring_submit: %d\n", ret); |
| 151 | + ret = T_EXIT_FAIL; |
| 152 | + goto out; |
| 153 | + } |
| 154 | + if (ret != 2) { |
| 155 | + fprintf(stderr, "io_uring_submit: submitted %d (expected 2)\n", ret); |
| 156 | + ret = T_EXIT_FAIL; |
| 157 | + goto out; |
| 158 | + } |
| 159 | + |
| 160 | + ret = io_uring_wait_cqe(&ring, &cqe); |
| 161 | + if (ret) { |
| 162 | + fprintf(stderr, "wait cqe: %d\n", ret); |
| 163 | + ret = T_EXIT_FAIL; |
| 164 | + goto out; |
| 165 | + } |
| 166 | + if (cqe->res != sizeof(buf)) { |
| 167 | + fprintf(stderr, "splice cqe res %d (expected %zu)\n", cqe->res, |
| 168 | + sizeof(buf)); |
| 169 | + ret = T_EXIT_FAIL; |
| 170 | + goto out; |
| 171 | + } |
| 172 | + io_uring_cqe_seen(&ring, cqe); |
| 173 | + |
| 174 | + ret = io_uring_wait_cqe(&ring, &cqe); |
| 175 | + if (ret) { |
| 176 | + fprintf(stderr, "wait cqe: %d\n", ret); |
| 177 | + ret = T_EXIT_FAIL; |
| 178 | + goto out; |
| 179 | + } |
| 180 | + if (cqe->res != sizeof(buf)) { |
| 181 | + fprintf(stderr, "splice cqe res %d (expected %zu)\n", cqe->res, |
| 182 | + sizeof(buf)); |
| 183 | + ret = T_EXIT_FAIL; |
| 184 | + goto out; |
| 185 | + } |
| 186 | + io_uring_cqe_seen(&ring, cqe); |
| 187 | + |
| 188 | + /* |
| 189 | + * Ensure we actually created io-wq worker(s) before checking they exit |
| 190 | + * after closing the last ring. |
| 191 | + */ |
| 192 | + ret = wait_for_iowq_workers(true, 1000); |
| 193 | + if (ret != T_EXIT_PASS) { |
| 194 | + if (ret == T_EXIT_FAIL) |
| 195 | + ret = T_EXIT_SKIP; |
| 196 | + goto out; |
| 197 | + } |
| 198 | + |
| 199 | + io_uring_queue_exit(&ring); |
| 200 | + ring_exited = true; |
| 201 | + |
| 202 | + /* Now the worker(s) should exit quickly. */ |
| 203 | + ret = wait_for_iowq_workers(false, 2000); |
| 204 | + |
| 205 | +out: |
| 206 | + if (ring_inited && !ring_exited) |
| 207 | + io_uring_queue_exit(&ring); |
| 208 | + if (pipe_fds[0] != -1) |
| 209 | + close(pipe_fds[0]); |
| 210 | + if (pipe_fds[1] != -1) |
| 211 | + close(pipe_fds[1]); |
| 212 | + if (fd_src != -1) |
| 213 | + close(fd_src); |
| 214 | + if (fd_dst != -1) |
| 215 | + close(fd_dst); |
| 216 | + if (fd_src != -1) |
| 217 | + unlink(src_fname); |
| 218 | + if (fd_dst != -1) |
| 219 | + unlink(dst_fname); |
| 220 | + return ret; |
| 221 | +} |
0 commit comments