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 spin_lock_init(&tctx->task_lock); 90 INIT_WQ_LIST(&tctx->task_list); 91 init_task_work(&tctx->task_work, tctx_task_work); 92 return 0; 93 } 94 95 static int io_register_submitter(struct io_ring_ctx *ctx) 96 { 97 int ret = 0; 98 99 mutex_lock(&ctx->uring_lock); 100 if (!ctx->submitter_task) 101 ctx->submitter_task = get_task_struct(current); 102 else if (ctx->submitter_task != current) 103 ret = -EEXIST; 104 mutex_unlock(&ctx->uring_lock); 105 106 return ret; 107 } 108 109 int __io_uring_add_tctx_node(struct io_ring_ctx *ctx, bool submitter) 110 { 111 struct io_uring_task *tctx = current->io_uring; 112 struct io_tctx_node *node; 113 int ret; 114 115 if ((ctx->flags & IORING_SETUP_SINGLE_ISSUER) && submitter) { 116 ret = io_register_submitter(ctx); 117 if (ret) 118 return ret; 119 } 120 121 if (unlikely(!tctx)) { 122 ret = io_uring_alloc_task_context(current, ctx); 123 if (unlikely(ret)) 124 return ret; 125 126 tctx = current->io_uring; 127 if (ctx->iowq_limits_set) { 128 unsigned int limits[2] = { ctx->iowq_limits[0], 129 ctx->iowq_limits[1], }; 130 131 ret = io_wq_max_workers(tctx->io_wq, limits); 132 if (ret) 133 return ret; 134 } 135 } 136 if (!xa_load(&tctx->xa, (unsigned long)ctx)) { 137 node = kmalloc(sizeof(*node), GFP_KERNEL); 138 if (!node) 139 return -ENOMEM; 140 node->ctx = ctx; 141 node->task = current; 142 143 ret = xa_err(xa_store(&tctx->xa, (unsigned long)ctx, 144 node, GFP_KERNEL)); 145 if (ret) { 146 kfree(node); 147 return ret; 148 } 149 150 mutex_lock(&ctx->uring_lock); 151 list_add(&node->ctx_node, &ctx->tctx_list); 152 mutex_unlock(&ctx->uring_lock); 153 } 154 if (submitter) 155 tctx->last = ctx; 156 return 0; 157 } 158 159 /* 160 * Remove this io_uring_file -> task mapping. 161 */ 162 __cold void io_uring_del_tctx_node(unsigned long index) 163 { 164 struct io_uring_task *tctx = current->io_uring; 165 struct io_tctx_node *node; 166 167 if (!tctx) 168 return; 169 node = xa_erase(&tctx->xa, index); 170 if (!node) 171 return; 172 173 WARN_ON_ONCE(current != node->task); 174 WARN_ON_ONCE(list_empty(&node->ctx_node)); 175 176 mutex_lock(&node->ctx->uring_lock); 177 list_del(&node->ctx_node); 178 mutex_unlock(&node->ctx->uring_lock); 179 180 if (tctx->last == node->ctx) 181 tctx->last = NULL; 182 kfree(node); 183 } 184 185 __cold void io_uring_clean_tctx(struct io_uring_task *tctx) 186 { 187 struct io_wq *wq = tctx->io_wq; 188 struct io_tctx_node *node; 189 unsigned long index; 190 191 xa_for_each(&tctx->xa, index, node) { 192 io_uring_del_tctx_node(index); 193 cond_resched(); 194 } 195 if (wq) { 196 /* 197 * Must be after io_uring_del_tctx_node() (removes nodes under 198 * uring_lock) to avoid race with io_uring_try_cancel_iowq(). 199 */ 200 io_wq_put_and_exit(wq); 201 tctx->io_wq = NULL; 202 } 203 } 204 205 void io_uring_unreg_ringfd(void) 206 { 207 struct io_uring_task *tctx = current->io_uring; 208 int i; 209 210 for (i = 0; i < IO_RINGFD_REG_MAX; i++) { 211 if (tctx->registered_rings[i]) { 212 fput(tctx->registered_rings[i]); 213 tctx->registered_rings[i] = NULL; 214 } 215 } 216 } 217 218 static int io_ring_add_registered_fd(struct io_uring_task *tctx, int fd, 219 int start, int end) 220 { 221 struct file *file; 222 int offset; 223 224 for (offset = start; offset < end; offset++) { 225 offset = array_index_nospec(offset, IO_RINGFD_REG_MAX); 226 if (tctx->registered_rings[offset]) 227 continue; 228 229 file = fget(fd); 230 if (!file) { 231 return -EBADF; 232 } else if (!io_is_uring_fops(file)) { 233 fput(file); 234 return -EOPNOTSUPP; 235 } 236 tctx->registered_rings[offset] = file; 237 return offset; 238 } 239 240 return -EBUSY; 241 } 242 243 /* 244 * Register a ring fd to avoid fdget/fdput for each io_uring_enter() 245 * invocation. User passes in an array of struct io_uring_rsrc_update 246 * with ->data set to the ring_fd, and ->offset given for the desired 247 * index. If no index is desired, application may set ->offset == -1U 248 * and we'll find an available index. Returns number of entries 249 * successfully processed, or < 0 on error if none were processed. 250 */ 251 int io_ringfd_register(struct io_ring_ctx *ctx, void __user *__arg, 252 unsigned nr_args) 253 { 254 struct io_uring_rsrc_update __user *arg = __arg; 255 struct io_uring_rsrc_update reg; 256 struct io_uring_task *tctx; 257 int ret, i; 258 259 if (!nr_args || nr_args > IO_RINGFD_REG_MAX) 260 return -EINVAL; 261 262 mutex_unlock(&ctx->uring_lock); 263 ret = __io_uring_add_tctx_node(ctx, false); 264 mutex_lock(&ctx->uring_lock); 265 if (ret) 266 return ret; 267 268 tctx = current->io_uring; 269 for (i = 0; i < nr_args; i++) { 270 int start, end; 271 272 if (copy_from_user(®, &arg[i], sizeof(reg))) { 273 ret = -EFAULT; 274 break; 275 } 276 277 if (reg.resv) { 278 ret = -EINVAL; 279 break; 280 } 281 282 if (reg.offset == -1U) { 283 start = 0; 284 end = IO_RINGFD_REG_MAX; 285 } else { 286 if (reg.offset >= IO_RINGFD_REG_MAX) { 287 ret = -EINVAL; 288 break; 289 } 290 start = reg.offset; 291 end = start + 1; 292 } 293 294 ret = io_ring_add_registered_fd(tctx, reg.data, start, end); 295 if (ret < 0) 296 break; 297 298 reg.offset = ret; 299 if (copy_to_user(&arg[i], ®, sizeof(reg))) { 300 fput(tctx->registered_rings[reg.offset]); 301 tctx->registered_rings[reg.offset] = NULL; 302 ret = -EFAULT; 303 break; 304 } 305 } 306 307 return i ? i : ret; 308 } 309 310 int io_ringfd_unregister(struct io_ring_ctx *ctx, void __user *__arg, 311 unsigned nr_args) 312 { 313 struct io_uring_rsrc_update __user *arg = __arg; 314 struct io_uring_task *tctx = current->io_uring; 315 struct io_uring_rsrc_update reg; 316 int ret = 0, i; 317 318 if (!nr_args || nr_args > IO_RINGFD_REG_MAX) 319 return -EINVAL; 320 if (!tctx) 321 return 0; 322 323 for (i = 0; i < nr_args; i++) { 324 if (copy_from_user(®, &arg[i], sizeof(reg))) { 325 ret = -EFAULT; 326 break; 327 } 328 if (reg.resv || reg.data || reg.offset >= IO_RINGFD_REG_MAX) { 329 ret = -EINVAL; 330 break; 331 } 332 333 reg.offset = array_index_nospec(reg.offset, IO_RINGFD_REG_MAX); 334 if (tctx->registered_rings[reg.offset]) { 335 fput(tctx->registered_rings[reg.offset]); 336 tctx->registered_rings[reg.offset] = NULL; 337 } 338 } 339 340 return i ? i : ret; 341 } 342