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