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