| 1 | // SPDX-License-Identifier: GPL-2.0 |
| 2 | #include <linux/kernel.h> |
| 3 | #include <linux/errno.h> |
| 4 | #include <linux/fs.h> |
| 5 | #include <linux/file.h> |
| 6 | #include <linux/io_uring.h> |
| 7 | |
| 8 | #include <uapi/linux/io_uring.h> |
| 9 | |
| 10 | #include "../kernel/futex/futex.h" |
| 11 | #include "io_uring.h" |
| 12 | #include "alloc_cache.h" |
| 13 | #include "futex.h" |
| 14 | |
| 15 | struct io_futex { |
| 16 | struct file *file; |
| 17 | void __user *uaddr; |
| 18 | unsigned long futex_val; |
| 19 | unsigned long futex_mask; |
| 20 | unsigned long futexv_owned; |
| 21 | u32 futex_flags; |
| 22 | unsigned int futex_nr; |
| 23 | bool futexv_unqueued; |
| 24 | }; |
| 25 | |
| 26 | struct io_futex_data { |
| 27 | struct futex_q q; |
| 28 | struct io_kiocb *req; |
| 29 | }; |
| 30 | |
| 31 | #define IO_FUTEX_ALLOC_CACHE_MAX 32 |
| 32 | |
| 33 | bool io_futex_cache_init(struct io_ring_ctx *ctx) |
| 34 | { |
| 35 | return io_alloc_cache_init(cache: &ctx->futex_cache, IO_FUTEX_ALLOC_CACHE_MAX, |
| 36 | size: sizeof(struct io_futex_data), init_bytes: 0); |
| 37 | } |
| 38 | |
| 39 | void io_futex_cache_free(struct io_ring_ctx *ctx) |
| 40 | { |
| 41 | io_alloc_cache_free(cache: &ctx->futex_cache, free: kfree); |
| 42 | } |
| 43 | |
| 44 | static void __io_futex_complete(struct io_kiocb *req, io_tw_token_t tw) |
| 45 | { |
| 46 | hlist_del_init(n: &req->hash_node); |
| 47 | io_req_task_complete(req, tw); |
| 48 | } |
| 49 | |
| 50 | static void io_futex_complete(struct io_kiocb *req, io_tw_token_t tw) |
| 51 | { |
| 52 | struct io_ring_ctx *ctx = req->ctx; |
| 53 | |
| 54 | io_tw_lock(ctx, tw); |
| 55 | io_cache_free(cache: &ctx->futex_cache, obj: req->async_data); |
| 56 | io_req_async_data_clear(req, extra_flags: 0); |
| 57 | __io_futex_complete(req, tw); |
| 58 | } |
| 59 | |
| 60 | static void io_futexv_complete(struct io_kiocb *req, io_tw_token_t tw) |
| 61 | { |
| 62 | struct io_futex *iof = io_kiocb_to_cmd(req, struct io_futex); |
| 63 | struct futex_vector *futexv = req->async_data; |
| 64 | |
| 65 | io_tw_lock(ctx: req->ctx, tw); |
| 66 | |
| 67 | if (!iof->futexv_unqueued) { |
| 68 | int res; |
| 69 | |
| 70 | res = futex_unqueue_multiple(v: futexv, count: iof->futex_nr); |
| 71 | if (res != -1) |
| 72 | io_req_set_res(req, res, cflags: 0); |
| 73 | } |
| 74 | |
| 75 | io_req_async_data_free(req); |
| 76 | __io_futex_complete(req, tw); |
| 77 | } |
| 78 | |
| 79 | static bool io_futexv_claim(struct io_futex *iof) |
| 80 | { |
| 81 | if (test_bit(0, &iof->futexv_owned) || |
| 82 | test_and_set_bit_lock(nr: 0, addr: &iof->futexv_owned)) |
| 83 | return false; |
| 84 | return true; |
| 85 | } |
| 86 | |
| 87 | static bool __io_futex_cancel(struct io_kiocb *req) |
| 88 | { |
| 89 | /* futex wake already done or in progress */ |
| 90 | if (req->opcode == IORING_OP_FUTEX_WAIT) { |
| 91 | struct io_futex_data *ifd = req->async_data; |
| 92 | |
| 93 | if (!futex_unqueue(q: &ifd->q)) |
| 94 | return false; |
| 95 | req->io_task_work.func = io_futex_complete; |
| 96 | } else { |
| 97 | struct io_futex *iof = io_kiocb_to_cmd(req, struct io_futex); |
| 98 | |
| 99 | if (!io_futexv_claim(iof)) |
| 100 | return false; |
| 101 | req->io_task_work.func = io_futexv_complete; |
| 102 | } |
| 103 | |
| 104 | hlist_del_init(n: &req->hash_node); |
| 105 | io_req_set_res(req, res: -ECANCELED, cflags: 0); |
| 106 | io_req_task_work_add(req); |
| 107 | return true; |
| 108 | } |
| 109 | |
| 110 | int io_futex_cancel(struct io_ring_ctx *ctx, struct io_cancel_data *cd, |
| 111 | unsigned int issue_flags) |
| 112 | { |
| 113 | return io_cancel_remove(ctx, cd, issue_flags, list: &ctx->futex_list, cancel: __io_futex_cancel); |
| 114 | } |
| 115 | |
| 116 | bool io_futex_remove_all(struct io_ring_ctx *ctx, struct io_uring_task *tctx, |
| 117 | bool cancel_all) |
| 118 | { |
| 119 | return io_cancel_remove_all(ctx, tctx, list: &ctx->futex_list, cancel_all, cancel: __io_futex_cancel); |
| 120 | } |
| 121 | |
| 122 | int io_futex_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
| 123 | { |
| 124 | struct io_futex *iof = io_kiocb_to_cmd(req, struct io_futex); |
| 125 | u32 flags; |
| 126 | |
| 127 | if (unlikely(sqe->len || sqe->futex_flags || sqe->buf_index || |
| 128 | sqe->file_index)) |
| 129 | return -EINVAL; |
| 130 | |
| 131 | iof->uaddr = u64_to_user_ptr(READ_ONCE(sqe->addr)); |
| 132 | iof->futex_val = READ_ONCE(sqe->addr2); |
| 133 | iof->futex_mask = READ_ONCE(sqe->addr3); |
| 134 | flags = READ_ONCE(sqe->fd); |
| 135 | |
| 136 | if (flags & ~FUTEX2_VALID_MASK) |
| 137 | return -EINVAL; |
| 138 | |
| 139 | iof->futex_flags = futex2_to_flags(flags2: flags); |
| 140 | if (!futex_flags_valid(flags: iof->futex_flags)) |
| 141 | return -EINVAL; |
| 142 | |
| 143 | if (!futex_validate_input(flags: iof->futex_flags, val: iof->futex_val) || |
| 144 | !futex_validate_input(flags: iof->futex_flags, val: iof->futex_mask)) |
| 145 | return -EINVAL; |
| 146 | |
| 147 | /* Mark as inflight, so file exit cancelation will find it */ |
| 148 | io_req_track_inflight(req); |
| 149 | return 0; |
| 150 | } |
| 151 | |
| 152 | static void io_futex_wakev_fn(struct wake_q_head *wake_q, struct futex_q *q) |
| 153 | { |
| 154 | struct io_kiocb *req = q->wake_data; |
| 155 | struct io_futex *iof = io_kiocb_to_cmd(req, struct io_futex); |
| 156 | |
| 157 | if (!io_futexv_claim(iof)) |
| 158 | return; |
| 159 | if (unlikely(!__futex_wake_mark(q))) |
| 160 | return; |
| 161 | |
| 162 | io_req_set_res(req, res: 0, cflags: 0); |
| 163 | req->io_task_work.func = io_futexv_complete; |
| 164 | io_req_task_work_add(req); |
| 165 | } |
| 166 | |
| 167 | int io_futexv_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) |
| 168 | { |
| 169 | struct io_futex *iof = io_kiocb_to_cmd(req, struct io_futex); |
| 170 | struct futex_vector *futexv; |
| 171 | int ret; |
| 172 | |
| 173 | /* No flags or mask supported for waitv */ |
| 174 | if (unlikely(sqe->fd || sqe->buf_index || sqe->file_index || |
| 175 | sqe->addr2 || sqe->futex_flags || sqe->addr3)) |
| 176 | return -EINVAL; |
| 177 | |
| 178 | iof->uaddr = u64_to_user_ptr(READ_ONCE(sqe->addr)); |
| 179 | iof->futex_nr = READ_ONCE(sqe->len); |
| 180 | if (!iof->futex_nr || iof->futex_nr > FUTEX_WAITV_MAX) |
| 181 | return -EINVAL; |
| 182 | |
| 183 | futexv = kcalloc(iof->futex_nr, sizeof(*futexv), GFP_KERNEL); |
| 184 | if (!futexv) |
| 185 | return -ENOMEM; |
| 186 | |
| 187 | ret = futex_parse_waitv(futexv, uwaitv: iof->uaddr, nr_futexes: iof->futex_nr, |
| 188 | wake: io_futex_wakev_fn, wake_data: req); |
| 189 | if (ret) { |
| 190 | kfree(objp: futexv); |
| 191 | return ret; |
| 192 | } |
| 193 | |
| 194 | /* Mark as inflight, so file exit cancelation will find it */ |
| 195 | io_req_track_inflight(req); |
| 196 | iof->futexv_owned = 0; |
| 197 | iof->futexv_unqueued = 0; |
| 198 | req->flags |= REQ_F_ASYNC_DATA; |
| 199 | req->async_data = futexv; |
| 200 | return 0; |
| 201 | } |
| 202 | |
| 203 | static void io_futex_wake_fn(struct wake_q_head *wake_q, struct futex_q *q) |
| 204 | { |
| 205 | struct io_futex_data *ifd = container_of(q, struct io_futex_data, q); |
| 206 | struct io_kiocb *req = ifd->req; |
| 207 | |
| 208 | if (unlikely(!__futex_wake_mark(q))) |
| 209 | return; |
| 210 | |
| 211 | io_req_set_res(req, res: 0, cflags: 0); |
| 212 | req->io_task_work.func = io_futex_complete; |
| 213 | io_req_task_work_add(req); |
| 214 | } |
| 215 | |
| 216 | int io_futexv_wait(struct io_kiocb *req, unsigned int issue_flags) |
| 217 | { |
| 218 | struct io_futex *iof = io_kiocb_to_cmd(req, struct io_futex); |
| 219 | struct futex_vector *futexv = req->async_data; |
| 220 | struct io_ring_ctx *ctx = req->ctx; |
| 221 | int ret, woken = -1; |
| 222 | |
| 223 | io_ring_submit_lock(ctx, issue_flags); |
| 224 | |
| 225 | ret = futex_wait_multiple_setup(vs: futexv, count: iof->futex_nr, woken: &woken); |
| 226 | |
| 227 | /* |
| 228 | * Error case, ret is < 0. Mark the request as failed. |
| 229 | */ |
| 230 | if (unlikely(ret < 0)) { |
| 231 | io_ring_submit_unlock(ctx, issue_flags); |
| 232 | req_set_fail(req); |
| 233 | io_req_set_res(req, res: ret, cflags: 0); |
| 234 | io_req_async_data_free(req); |
| 235 | return IOU_COMPLETE; |
| 236 | } |
| 237 | |
| 238 | /* |
| 239 | * 0 return means that we successfully setup the waiters, and that |
| 240 | * nobody triggered a wakeup while we were doing so. If the wakeup |
| 241 | * happened post setup, the task_work will be run post this issue and |
| 242 | * under the submission lock. 1 means We got woken while setting up, |
| 243 | * let that side do the completion. Note that |
| 244 | * futex_wait_multiple_setup() will have unqueued all the futexes in |
| 245 | * this case. Mark us as having done that already, since this is |
| 246 | * different from normal wakeup. |
| 247 | */ |
| 248 | if (!ret) { |
| 249 | /* |
| 250 | * If futex_wait_multiple_setup() returns 0 for a |
| 251 | * successful setup, then the task state will not be |
| 252 | * runnable. This is fine for the sync syscall, as |
| 253 | * it'll be blocking unless we already got one of the |
| 254 | * futexes woken, but it obviously won't work for an |
| 255 | * async invocation. Mark us runnable again. |
| 256 | */ |
| 257 | __set_current_state(TASK_RUNNING); |
| 258 | hlist_add_head(n: &req->hash_node, h: &ctx->futex_list); |
| 259 | } else { |
| 260 | iof->futexv_unqueued = 1; |
| 261 | if (woken != -1) |
| 262 | io_req_set_res(req, res: woken, cflags: 0); |
| 263 | } |
| 264 | |
| 265 | io_ring_submit_unlock(ctx, issue_flags); |
| 266 | return IOU_ISSUE_SKIP_COMPLETE; |
| 267 | } |
| 268 | |
| 269 | int io_futex_wait(struct io_kiocb *req, unsigned int issue_flags) |
| 270 | { |
| 271 | struct io_futex *iof = io_kiocb_to_cmd(req, struct io_futex); |
| 272 | struct io_ring_ctx *ctx = req->ctx; |
| 273 | struct io_futex_data *ifd = NULL; |
| 274 | int ret; |
| 275 | |
| 276 | if (!iof->futex_mask) { |
| 277 | ret = -EINVAL; |
| 278 | goto done; |
| 279 | } |
| 280 | |
| 281 | io_ring_submit_lock(ctx, issue_flags); |
| 282 | ifd = io_cache_alloc(cache: &ctx->futex_cache, GFP_NOWAIT); |
| 283 | if (!ifd) { |
| 284 | ret = -ENOMEM; |
| 285 | goto done_unlock; |
| 286 | } |
| 287 | |
| 288 | req->flags |= REQ_F_ASYNC_DATA; |
| 289 | req->async_data = ifd; |
| 290 | ifd->q = futex_q_init; |
| 291 | ifd->q.bitset = iof->futex_mask; |
| 292 | ifd->q.wake = io_futex_wake_fn; |
| 293 | ifd->req = req; |
| 294 | |
| 295 | ret = futex_wait_setup(uaddr: iof->uaddr, val: iof->futex_val, flags: iof->futex_flags, |
| 296 | q: &ifd->q, NULL, NULL); |
| 297 | if (!ret) { |
| 298 | hlist_add_head(n: &req->hash_node, h: &ctx->futex_list); |
| 299 | io_ring_submit_unlock(ctx, issue_flags); |
| 300 | |
| 301 | return IOU_ISSUE_SKIP_COMPLETE; |
| 302 | } |
| 303 | |
| 304 | done_unlock: |
| 305 | io_ring_submit_unlock(ctx, issue_flags); |
| 306 | done: |
| 307 | if (ret < 0) |
| 308 | req_set_fail(req); |
| 309 | io_req_set_res(req, res: ret, cflags: 0); |
| 310 | io_req_async_data_free(req); |
| 311 | return IOU_COMPLETE; |
| 312 | } |
| 313 | |
| 314 | int io_futex_wake(struct io_kiocb *req, unsigned int issue_flags) |
| 315 | { |
| 316 | struct io_futex *iof = io_kiocb_to_cmd(req, struct io_futex); |
| 317 | int ret; |
| 318 | |
| 319 | /* |
| 320 | * Strict flags - ensure that waking 0 futexes yields a 0 result. |
| 321 | * See commit 43adf8449510 ("futex: FLAGS_STRICT") for details. |
| 322 | */ |
| 323 | ret = futex_wake(uaddr: iof->uaddr, FLAGS_STRICT | iof->futex_flags, |
| 324 | nr_wake: iof->futex_val, bitset: iof->futex_mask); |
| 325 | if (ret < 0) |
| 326 | req_set_fail(req); |
| 327 | io_req_set_res(req, res: ret, cflags: 0); |
| 328 | return IOU_COMPLETE; |
| 329 | } |
| 330 | |