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