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(size: sizeof(*hash), GFP_KERNEL); |
26 | if (!hash) { |
27 | mutex_unlock(lock: &ctx->uring_lock); |
28 | return ERR_PTR(error: -ENOMEM); |
29 | } |
30 | refcount_set(r: &hash->refs, n: 1); |
31 | init_waitqueue_head(&hash->wait); |
32 | ctx->hash_map = hash; |
33 | } |
34 | mutex_unlock(lock: &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(bounded: concurrency, data: &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(fbc: &tctx->inflight); |
56 | kfree(objp: 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(size: 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(objp: tctx); |
73 | return ret; |
74 | } |
75 | |
76 | tctx->io_wq = io_init_wq_offload(ctx, task); |
77 | if (IS_ERR(ptr: tctx->io_wq)) { |
78 | ret = PTR_ERR(ptr: tctx->io_wq); |
79 | percpu_counter_destroy(fbc: &tctx->inflight); |
80 | kfree(objp: tctx); |
81 | return ret; |
82 | } |
83 | |
84 | xa_init(xa: &tctx->xa); |
85 | init_waitqueue_head(&tctx->wait); |
86 | atomic_set(v: &tctx->in_cancel, i: 0); |
87 | atomic_set(v: &tctx->inflight_tracked, i: 0); |
88 | task->io_uring = tctx; |
89 | init_llist_head(list: &tctx->task_list); |
90 | init_task_work(twork: &tctx->task_work, func: tctx_task_work); |
91 | return 0; |
92 | } |
93 | |
94 | int __io_uring_add_tctx_node(struct io_ring_ctx *ctx) |
95 | { |
96 | struct io_uring_task *tctx = current->io_uring; |
97 | struct io_tctx_node *node; |
98 | int ret; |
99 | |
100 | if (unlikely(!tctx)) { |
101 | ret = io_uring_alloc_task_context(current, ctx); |
102 | if (unlikely(ret)) |
103 | return ret; |
104 | |
105 | tctx = current->io_uring; |
106 | if (ctx->iowq_limits_set) { |
107 | unsigned int limits[2] = { ctx->iowq_limits[0], |
108 | ctx->iowq_limits[1], }; |
109 | |
110 | ret = io_wq_max_workers(wq: tctx->io_wq, new_count: limits); |
111 | if (ret) |
112 | return ret; |
113 | } |
114 | } |
115 | if (!xa_load(&tctx->xa, index: (unsigned long)ctx)) { |
116 | node = kmalloc(size: sizeof(*node), GFP_KERNEL); |
117 | if (!node) |
118 | return -ENOMEM; |
119 | node->ctx = ctx; |
120 | node->task = current; |
121 | |
122 | ret = xa_err(entry: xa_store(&tctx->xa, index: (unsigned long)ctx, |
123 | entry: node, GFP_KERNEL)); |
124 | if (ret) { |
125 | kfree(objp: node); |
126 | return ret; |
127 | } |
128 | |
129 | mutex_lock(&ctx->uring_lock); |
130 | list_add(new: &node->ctx_node, head: &ctx->tctx_list); |
131 | mutex_unlock(lock: &ctx->uring_lock); |
132 | } |
133 | return 0; |
134 | } |
135 | |
136 | int __io_uring_add_tctx_node_from_submit(struct io_ring_ctx *ctx) |
137 | { |
138 | int ret; |
139 | |
140 | if (ctx->flags & IORING_SETUP_SINGLE_ISSUER |
141 | && ctx->submitter_task != current) |
142 | return -EEXIST; |
143 | |
144 | ret = __io_uring_add_tctx_node(ctx); |
145 | if (ret) |
146 | return ret; |
147 | |
148 | current->io_uring->last = ctx; |
149 | return 0; |
150 | } |
151 | |
152 | /* |
153 | * Remove this io_uring_file -> task mapping. |
154 | */ |
155 | __cold void io_uring_del_tctx_node(unsigned long index) |
156 | { |
157 | struct io_uring_task *tctx = current->io_uring; |
158 | struct io_tctx_node *node; |
159 | |
160 | if (!tctx) |
161 | return; |
162 | node = xa_erase(&tctx->xa, index); |
163 | if (!node) |
164 | return; |
165 | |
166 | WARN_ON_ONCE(current != node->task); |
167 | WARN_ON_ONCE(list_empty(&node->ctx_node)); |
168 | |
169 | mutex_lock(&node->ctx->uring_lock); |
170 | list_del(entry: &node->ctx_node); |
171 | mutex_unlock(lock: &node->ctx->uring_lock); |
172 | |
173 | if (tctx->last == node->ctx) |
174 | tctx->last = NULL; |
175 | kfree(objp: node); |
176 | } |
177 | |
178 | __cold void io_uring_clean_tctx(struct io_uring_task *tctx) |
179 | { |
180 | struct io_wq *wq = tctx->io_wq; |
181 | struct io_tctx_node *node; |
182 | unsigned long index; |
183 | |
184 | xa_for_each(&tctx->xa, index, node) { |
185 | io_uring_del_tctx_node(index); |
186 | cond_resched(); |
187 | } |
188 | if (wq) { |
189 | /* |
190 | * Must be after io_uring_del_tctx_node() (removes nodes under |
191 | * uring_lock) to avoid race with io_uring_try_cancel_iowq(). |
192 | */ |
193 | io_wq_put_and_exit(wq); |
194 | tctx->io_wq = NULL; |
195 | } |
196 | } |
197 | |
198 | void io_uring_unreg_ringfd(void) |
199 | { |
200 | struct io_uring_task *tctx = current->io_uring; |
201 | int i; |
202 | |
203 | for (i = 0; i < IO_RINGFD_REG_MAX; i++) { |
204 | if (tctx->registered_rings[i]) { |
205 | fput(tctx->registered_rings[i]); |
206 | tctx->registered_rings[i] = NULL; |
207 | } |
208 | } |
209 | } |
210 | |
211 | int io_ring_add_registered_file(struct io_uring_task *tctx, struct file *file, |
212 | int start, int end) |
213 | { |
214 | int offset; |
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 | tctx->registered_rings[offset] = file; |
221 | return offset; |
222 | } |
223 | return -EBUSY; |
224 | } |
225 | |
226 | static int io_ring_add_registered_fd(struct io_uring_task *tctx, int fd, |
227 | int start, int end) |
228 | { |
229 | struct file *file; |
230 | int offset; |
231 | |
232 | file = fget(fd); |
233 | if (!file) { |
234 | return -EBADF; |
235 | } else if (!io_is_uring_fops(file)) { |
236 | fput(file); |
237 | return -EOPNOTSUPP; |
238 | } |
239 | offset = io_ring_add_registered_file(tctx, file, start, end); |
240 | if (offset < 0) |
241 | fput(file); |
242 | return offset; |
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(lock: &ctx->uring_lock); |
265 | ret = __io_uring_add_tctx_node(ctx); |
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(to: ®, from: &arg[i], n: 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, fd: reg.data, start, end); |
297 | if (ret < 0) |
298 | break; |
299 | |
300 | reg.offset = ret; |
301 | if (copy_to_user(to: &arg[i], from: ®, n: 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(to: ®, from: &arg[i], n: 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 | |