1 // SPDX-License-Identifier: GPL-2.0 2 #include <linux/kernel.h> 3 #include <linux/errno.h> 4 #include <linux/file.h> 5 #include <linux/mm.h> 6 #include <linux/slab.h> 7 #include <linux/nospec.h> 8 #include <linux/io_uring.h> 9 10 #include <uapi/linux/io_uring.h> 11 12 #include "io_uring.h" 13 #include "tctx.h" 14 15 static struct io_wq *io_init_wq_offload(struct io_ring_ctx *ctx, 16 struct task_struct *task) 17 { 18 struct io_wq_hash *hash; 19 struct io_wq_data data; 20 unsigned int concurrency; 21 22 mutex_lock(&ctx->uring_lock); 23 hash = ctx->hash_map; 24 if (!hash) { 25 hash = kzalloc(sizeof(*hash), GFP_KERNEL); 26 if (!hash) { 27 mutex_unlock(&ctx->uring_lock); 28 return ERR_PTR(-ENOMEM); 29 } 30 refcount_set(&hash->refs, 1); 31 init_waitqueue_head(&hash->wait); 32 ctx->hash_map = hash; 33 } 34 mutex_unlock(&ctx->uring_lock); 35 36 data.hash = hash; 37 data.task = task; 38 data.free_work = io_wq_free_work; 39 data.do_work = io_wq_submit_work; 40 41 /* Do QD, or 4 * CPUS, whatever is smallest */ 42 concurrency = min(ctx->sq_entries, 4 * num_online_cpus()); 43 44 return io_wq_create(concurrency, &data); 45 } 46 47 void __io_uring_free(struct task_struct *tsk) 48 { 49 struct io_uring_task *tctx = tsk->io_uring; 50 51 WARN_ON_ONCE(!xa_empty(&tctx->xa)); 52 WARN_ON_ONCE(tctx->io_wq); 53 WARN_ON_ONCE(tctx->cached_refs); 54 55 percpu_counter_destroy(&tctx->inflight); 56 kfree(tctx); 57 tsk->io_uring = NULL; 58 } 59 60 __cold int io_uring_alloc_task_context(struct task_struct *task, 61 struct io_ring_ctx *ctx) 62 { 63 struct io_uring_task *tctx; 64 int ret; 65 66 tctx = kzalloc(sizeof(*tctx), GFP_KERNEL); 67 if (unlikely(!tctx)) 68 return -ENOMEM; 69 70 ret = percpu_counter_init(&tctx->inflight, 0, GFP_KERNEL); 71 if (unlikely(ret)) { 72 kfree(tctx); 73 return ret; 74 } 75 76 tctx->io_wq = io_init_wq_offload(ctx, task); 77 if (IS_ERR(tctx->io_wq)) { 78 ret = PTR_ERR(tctx->io_wq); 79 percpu_counter_destroy(&tctx->inflight); 80 kfree(tctx); 81 return ret; 82 } 83 84 xa_init(&tctx->xa); 85 init_waitqueue_head(&tctx->wait); 86 atomic_set(&tctx->in_idle, 0); 87 atomic_set(&tctx->inflight_tracked, 0); 88 task->io_uring = tctx; 89 init_llist_head(&tctx->task_list); 90 init_task_work(&tctx->task_work, tctx_task_work); 91 return 0; 92 } 93 94 static int io_register_submitter(struct io_ring_ctx *ctx) 95 { 96 int ret = 0; 97 98 mutex_lock(&ctx->uring_lock); 99 if (!ctx->submitter_task) 100 ctx->submitter_task = get_task_struct(current); 101 else if (ctx->submitter_task != current) 102 ret = -EEXIST; 103 mutex_unlock(&ctx->uring_lock); 104 105 return ret; 106 } 107 108 int __io_uring_add_tctx_node(struct io_ring_ctx *ctx) 109 { 110 struct io_uring_task *tctx = current->io_uring; 111 struct io_tctx_node *node; 112 int ret; 113 114 if (unlikely(!tctx)) { 115 ret = io_uring_alloc_task_context(current, ctx); 116 if (unlikely(ret)) 117 return ret; 118 119 tctx = current->io_uring; 120 if (ctx->iowq_limits_set) { 121 unsigned int limits[2] = { ctx->iowq_limits[0], 122 ctx->iowq_limits[1], }; 123 124 ret = io_wq_max_workers(tctx->io_wq, limits); 125 if (ret) 126 return ret; 127 } 128 } 129 if (!xa_load(&tctx->xa, (unsigned long)ctx)) { 130 node = kmalloc(sizeof(*node), GFP_KERNEL); 131 if (!node) 132 return -ENOMEM; 133 node->ctx = ctx; 134 node->task = current; 135 136 ret = xa_err(xa_store(&tctx->xa, (unsigned long)ctx, 137 node, GFP_KERNEL)); 138 if (ret) { 139 kfree(node); 140 return ret; 141 } 142 143 mutex_lock(&ctx->uring_lock); 144 list_add(&node->ctx_node, &ctx->tctx_list); 145 mutex_unlock(&ctx->uring_lock); 146 } 147 return 0; 148 } 149 150 int __io_uring_add_tctx_node_from_submit(struct io_ring_ctx *ctx) 151 { 152 int ret; 153 154 if (ctx->flags & IORING_SETUP_SINGLE_ISSUER) { 155 ret = io_register_submitter(ctx); 156 if (ret) 157 return ret; 158 } 159 160 ret = __io_uring_add_tctx_node(ctx); 161 if (ret) 162 return ret; 163 164 current->io_uring->last = ctx; 165 return 0; 166 } 167 168 /* 169 * Remove this io_uring_file -> task mapping. 170 */ 171 __cold void io_uring_del_tctx_node(unsigned long index) 172 { 173 struct io_uring_task *tctx = current->io_uring; 174 struct io_tctx_node *node; 175 176 if (!tctx) 177 return; 178 node = xa_erase(&tctx->xa, index); 179 if (!node) 180 return; 181 182 WARN_ON_ONCE(current != node->task); 183 WARN_ON_ONCE(list_empty(&node->ctx_node)); 184 185 mutex_lock(&node->ctx->uring_lock); 186 list_del(&node->ctx_node); 187 mutex_unlock(&node->ctx->uring_lock); 188 189 if (tctx->last == node->ctx) 190 tctx->last = NULL; 191 kfree(node); 192 } 193 194 __cold void io_uring_clean_tctx(struct io_uring_task *tctx) 195 { 196 struct io_wq *wq = tctx->io_wq; 197 struct io_tctx_node *node; 198 unsigned long index; 199 200 xa_for_each(&tctx->xa, index, node) { 201 io_uring_del_tctx_node(index); 202 cond_resched(); 203 } 204 if (wq) { 205 /* 206 * Must be after io_uring_del_tctx_node() (removes nodes under 207 * uring_lock) to avoid race with io_uring_try_cancel_iowq(). 208 */ 209 io_wq_put_and_exit(wq); 210 tctx->io_wq = NULL; 211 } 212 } 213 214 void io_uring_unreg_ringfd(void) 215 { 216 struct io_uring_task *tctx = current->io_uring; 217 int i; 218 219 for (i = 0; i < IO_RINGFD_REG_MAX; i++) { 220 if (tctx->registered_rings[i]) { 221 fput(tctx->registered_rings[i]); 222 tctx->registered_rings[i] = NULL; 223 } 224 } 225 } 226 227 static int io_ring_add_registered_fd(struct io_uring_task *tctx, int fd, 228 int start, int end) 229 { 230 struct file *file; 231 int offset; 232 233 for (offset = start; offset < end; offset++) { 234 offset = array_index_nospec(offset, IO_RINGFD_REG_MAX); 235 if (tctx->registered_rings[offset]) 236 continue; 237 238 file = fget(fd); 239 if (!file) { 240 return -EBADF; 241 } else if (!io_is_uring_fops(file)) { 242 fput(file); 243 return -EOPNOTSUPP; 244 } 245 tctx->registered_rings[offset] = file; 246 return offset; 247 } 248 249 return -EBUSY; 250 } 251 252 /* 253 * Register a ring fd to avoid fdget/fdput for each io_uring_enter() 254 * invocation. User passes in an array of struct io_uring_rsrc_update 255 * with ->data set to the ring_fd, and ->offset given for the desired 256 * index. If no index is desired, application may set ->offset == -1U 257 * and we'll find an available index. Returns number of entries 258 * successfully processed, or < 0 on error if none were processed. 259 */ 260 int io_ringfd_register(struct io_ring_ctx *ctx, void __user *__arg, 261 unsigned nr_args) 262 { 263 struct io_uring_rsrc_update __user *arg = __arg; 264 struct io_uring_rsrc_update reg; 265 struct io_uring_task *tctx; 266 int ret, i; 267 268 if (!nr_args || nr_args > IO_RINGFD_REG_MAX) 269 return -EINVAL; 270 271 mutex_unlock(&ctx->uring_lock); 272 ret = __io_uring_add_tctx_node(ctx); 273 mutex_lock(&ctx->uring_lock); 274 if (ret) 275 return ret; 276 277 tctx = current->io_uring; 278 for (i = 0; i < nr_args; i++) { 279 int start, end; 280 281 if (copy_from_user(®, &arg[i], sizeof(reg))) { 282 ret = -EFAULT; 283 break; 284 } 285 286 if (reg.resv) { 287 ret = -EINVAL; 288 break; 289 } 290 291 if (reg.offset == -1U) { 292 start = 0; 293 end = IO_RINGFD_REG_MAX; 294 } else { 295 if (reg.offset >= IO_RINGFD_REG_MAX) { 296 ret = -EINVAL; 297 break; 298 } 299 start = reg.offset; 300 end = start + 1; 301 } 302 303 ret = io_ring_add_registered_fd(tctx, reg.data, start, end); 304 if (ret < 0) 305 break; 306 307 reg.offset = ret; 308 if (copy_to_user(&arg[i], ®, sizeof(reg))) { 309 fput(tctx->registered_rings[reg.offset]); 310 tctx->registered_rings[reg.offset] = NULL; 311 ret = -EFAULT; 312 break; 313 } 314 } 315 316 return i ? i : ret; 317 } 318 319 int io_ringfd_unregister(struct io_ring_ctx *ctx, void __user *__arg, 320 unsigned nr_args) 321 { 322 struct io_uring_rsrc_update __user *arg = __arg; 323 struct io_uring_task *tctx = current->io_uring; 324 struct io_uring_rsrc_update reg; 325 int ret = 0, i; 326 327 if (!nr_args || nr_args > IO_RINGFD_REG_MAX) 328 return -EINVAL; 329 if (!tctx) 330 return 0; 331 332 for (i = 0; i < nr_args; i++) { 333 if (copy_from_user(®, &arg[i], sizeof(reg))) { 334 ret = -EFAULT; 335 break; 336 } 337 if (reg.resv || reg.data || reg.offset >= IO_RINGFD_REG_MAX) { 338 ret = -EINVAL; 339 break; 340 } 341 342 reg.offset = array_index_nospec(reg.offset, IO_RINGFD_REG_MAX); 343 if (tctx->registered_rings[reg.offset]) { 344 fput(tctx->registered_rings[reg.offset]); 345 tctx->registered_rings[reg.offset] = NULL; 346 } 347 } 348 349 return i ? i : ret; 350 } 351