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 #include "bpf_filter.h"
15
io_init_wq_offload(struct io_ring_ctx * ctx,struct task_struct * task)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_obj(*hash);
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
40 /* Do QD, or 4 * CPUS, whatever is smallest */
41 concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
42
43 return io_wq_create(concurrency, &data);
44 }
45
__io_uring_free(struct task_struct * tsk)46 void __io_uring_free(struct task_struct *tsk)
47 {
48 struct io_uring_task *tctx = tsk->io_uring;
49 struct io_tctx_node *node;
50 unsigned long index;
51
52 /*
53 * Fault injection forcing allocation errors in the xa_store() path
54 * can lead to xa_empty() returning false, even though no actual
55 * node is stored in the xarray. Until that gets sorted out, attempt
56 * an iteration here and warn if any entries are found.
57 */
58 if (tctx) {
59 xa_for_each(&tctx->xa, index, node) {
60 WARN_ON_ONCE(1);
61 break;
62 }
63 WARN_ON_ONCE(tctx->io_wq);
64 WARN_ON_ONCE(tctx->cached_refs);
65
66 percpu_counter_destroy(&tctx->inflight);
67 kfree(tctx);
68 tsk->io_uring = NULL;
69 }
70 if (tsk->io_uring_restrict) {
71 io_put_bpf_filters(tsk->io_uring_restrict);
72 kfree(tsk->io_uring_restrict);
73 tsk->io_uring_restrict = NULL;
74 }
75 }
76
io_uring_alloc_task_context(struct task_struct * task,struct io_ring_ctx * ctx)77 __cold struct io_uring_task *io_uring_alloc_task_context(struct task_struct *task,
78 struct io_ring_ctx *ctx)
79 {
80 struct io_uring_task *tctx;
81 int ret;
82
83 tctx = kzalloc_obj(*tctx);
84 if (unlikely(!tctx))
85 return ERR_PTR(-ENOMEM);
86
87 ret = percpu_counter_init(&tctx->inflight, 0, GFP_KERNEL);
88 if (unlikely(ret)) {
89 kfree(tctx);
90 return ERR_PTR(ret);
91 }
92
93 tctx->io_wq = io_init_wq_offload(ctx, task);
94 if (IS_ERR(tctx->io_wq)) {
95 ret = PTR_ERR(tctx->io_wq);
96 percpu_counter_destroy(&tctx->inflight);
97 kfree(tctx);
98 return ERR_PTR(ret);
99 }
100
101 tctx->task = task;
102 xa_init(&tctx->xa);
103 init_waitqueue_head(&tctx->wait);
104 atomic_set(&tctx->in_cancel, 0);
105 atomic_set(&tctx->inflight_tracked, 0);
106 init_llist_head(&tctx->task_list);
107 init_task_work(&tctx->task_work, tctx_task_work);
108 return tctx;
109 }
110
io_tctx_install_node(struct io_ring_ctx * ctx,struct io_uring_task * tctx)111 static int io_tctx_install_node(struct io_ring_ctx *ctx,
112 struct io_uring_task *tctx)
113 {
114 struct io_tctx_node *node;
115 int ret;
116
117 if (xa_load(&tctx->xa, (unsigned long)ctx))
118 return 0;
119
120 node = kmalloc_obj(*node);
121 if (!node)
122 return -ENOMEM;
123 node->ctx = ctx;
124 node->task = current;
125
126 ret = xa_err(xa_store(&tctx->xa, (unsigned long)ctx,
127 node, GFP_KERNEL));
128 if (ret) {
129 kfree(node);
130 return ret;
131 }
132
133 mutex_lock(&ctx->tctx_lock);
134 list_add(&node->ctx_node, &ctx->tctx_list);
135 mutex_unlock(&ctx->tctx_lock);
136 return 0;
137 }
138
__io_uring_add_tctx_node(struct io_ring_ctx * ctx)139 int __io_uring_add_tctx_node(struct io_ring_ctx *ctx)
140 {
141 struct io_uring_task *tctx = current->io_uring;
142 bool new_tctx = false;
143 int ret;
144
145 if (unlikely(!tctx)) {
146 tctx = io_uring_alloc_task_context(current, ctx);
147 if (IS_ERR(tctx))
148 return PTR_ERR(tctx);
149 new_tctx = true;
150
151 if (data_race(ctx->int_flags) & IO_RING_F_IOWQ_LIMITS_SET) {
152 unsigned int limits[2];
153
154 mutex_lock(&ctx->uring_lock);
155 limits[0] = ctx->iowq_limits[0];
156 limits[1] = ctx->iowq_limits[1];
157 mutex_unlock(&ctx->uring_lock);
158
159 ret = io_wq_max_workers(tctx->io_wq, limits);
160 if (ret)
161 goto err_free;
162 }
163 }
164
165 /*
166 * Re-activate io-wq keepalive on any new io_uring usage. The wq may have
167 * been marked for idle-exit when the task temporarily had no active
168 * io_uring instances.
169 */
170 if (tctx->io_wq)
171 io_wq_set_exit_on_idle(tctx->io_wq, false);
172
173 if (new_tctx)
174 current->io_uring = tctx;
175
176 ret = io_tctx_install_node(ctx, tctx);
177 if (!ret)
178 return 0;
179 err_free:
180 if (new_tctx) {
181 current->io_uring = NULL;
182 if (tctx->io_wq) {
183 io_wq_exit_start(tctx->io_wq);
184 io_wq_put_and_exit(tctx->io_wq);
185 }
186 percpu_counter_destroy(&tctx->inflight);
187 kfree(tctx);
188 }
189 return ret;
190 }
191
__io_uring_add_tctx_node_from_submit(struct io_ring_ctx * ctx)192 int __io_uring_add_tctx_node_from_submit(struct io_ring_ctx *ctx)
193 {
194 int ret;
195
196 if (ctx->flags & IORING_SETUP_SINGLE_ISSUER
197 && ctx->submitter_task != current)
198 return -EEXIST;
199
200 ret = __io_uring_add_tctx_node(ctx);
201 if (ret)
202 return ret;
203
204 current->io_uring->last = ctx;
205 return 0;
206 }
207
208 /*
209 * Remove this io_uring_file -> task mapping.
210 */
io_uring_del_tctx_node(unsigned long index)211 __cold void io_uring_del_tctx_node(unsigned long index)
212 {
213 struct io_uring_task *tctx = current->io_uring;
214 struct io_tctx_node *node;
215
216 if (!tctx)
217 return;
218 node = xa_erase(&tctx->xa, index);
219 if (!node)
220 return;
221
222 WARN_ON_ONCE(current != node->task);
223 WARN_ON_ONCE(list_empty(&node->ctx_node));
224
225 mutex_lock(&node->ctx->tctx_lock);
226 list_del(&node->ctx_node);
227 mutex_unlock(&node->ctx->tctx_lock);
228
229 if (tctx->last == node->ctx)
230 tctx->last = NULL;
231 kfree(node);
232
233 if (xa_empty(&tctx->xa) && tctx->io_wq)
234 io_wq_set_exit_on_idle(tctx->io_wq, true);
235 }
236
io_uring_clean_tctx(struct io_uring_task * tctx)237 __cold void io_uring_clean_tctx(struct io_uring_task *tctx)
238 {
239 struct io_wq *wq = tctx->io_wq;
240 struct io_tctx_node *node;
241 unsigned long index;
242
243 xa_for_each(&tctx->xa, index, node) {
244 io_uring_del_tctx_node(index);
245 cond_resched();
246 }
247 if (wq) {
248 /*
249 * Must be after io_uring_del_tctx_node() (removes nodes under
250 * uring_lock) to avoid race with io_uring_try_cancel_iowq().
251 */
252 io_wq_put_and_exit(wq);
253 tctx->io_wq = NULL;
254 }
255 }
256
io_uring_unreg_ringfd(void)257 void io_uring_unreg_ringfd(void)
258 {
259 struct io_uring_task *tctx = current->io_uring;
260 int i;
261
262 for (i = 0; i < IO_RINGFD_REG_MAX; i++) {
263 if (tctx->registered_rings[i]) {
264 fput(tctx->registered_rings[i]);
265 tctx->registered_rings[i] = NULL;
266 }
267 }
268 }
269
io_ring_add_registered_file(struct io_uring_task * tctx,struct file * file,int start,int end)270 int io_ring_add_registered_file(struct io_uring_task *tctx, struct file *file,
271 int start, int end)
272 {
273 int offset, idx;
274 for (offset = start; offset < end; offset++) {
275 idx = array_index_nospec(offset, IO_RINGFD_REG_MAX);
276 if (tctx->registered_rings[idx])
277 continue;
278
279 tctx->registered_rings[idx] = file;
280 return idx;
281 }
282 return -EBUSY;
283 }
284
io_ring_add_registered_fd(struct io_uring_task * tctx,int fd,int start,int end)285 static int io_ring_add_registered_fd(struct io_uring_task *tctx, int fd,
286 int start, int end)
287 {
288 struct file *file;
289 int offset;
290
291 file = fget(fd);
292 if (!file) {
293 return -EBADF;
294 } else if (!io_is_uring_fops(file)) {
295 fput(file);
296 return -EOPNOTSUPP;
297 }
298 offset = io_ring_add_registered_file(tctx, file, start, end);
299 if (offset < 0)
300 fput(file);
301 return offset;
302 }
303
304 /*
305 * Register a ring fd to avoid fdget/fdput for each io_uring_enter()
306 * invocation. User passes in an array of struct io_uring_rsrc_update
307 * with ->data set to the ring_fd, and ->offset given for the desired
308 * index. If no index is desired, application may set ->offset == -1U
309 * and we'll find an available index. Returns number of entries
310 * successfully processed, or < 0 on error if none were processed.
311 */
io_ringfd_register(struct io_ring_ctx * ctx,void __user * __arg,unsigned nr_args)312 int io_ringfd_register(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_rsrc_update reg;
317 struct io_uring_task *tctx;
318 int ret, i;
319
320 if (!nr_args || nr_args > IO_RINGFD_REG_MAX)
321 return -EINVAL;
322
323 mutex_unlock(&ctx->uring_lock);
324 ret = __io_uring_add_tctx_node(ctx);
325 mutex_lock(&ctx->uring_lock);
326 if (ret)
327 return ret;
328
329 tctx = current->io_uring;
330 for (i = 0; i < nr_args; i++) {
331 int start, end;
332
333 if (copy_from_user(®, &arg[i], sizeof(reg))) {
334 ret = -EFAULT;
335 break;
336 }
337
338 if (reg.resv) {
339 ret = -EINVAL;
340 break;
341 }
342
343 if (reg.offset == -1U) {
344 start = 0;
345 end = IO_RINGFD_REG_MAX;
346 } else {
347 if (reg.offset >= IO_RINGFD_REG_MAX) {
348 ret = -EINVAL;
349 break;
350 }
351 start = reg.offset;
352 end = start + 1;
353 }
354
355 ret = io_ring_add_registered_fd(tctx, reg.data, start, end);
356 if (ret < 0)
357 break;
358
359 reg.offset = ret;
360 if (copy_to_user(&arg[i], ®, sizeof(reg))) {
361 fput(tctx->registered_rings[reg.offset]);
362 tctx->registered_rings[reg.offset] = NULL;
363 ret = -EFAULT;
364 break;
365 }
366 }
367
368 return i ? i : ret;
369 }
370
io_ringfd_unregister(struct io_ring_ctx * ctx,void __user * __arg,unsigned nr_args)371 int io_ringfd_unregister(struct io_ring_ctx *ctx, void __user *__arg,
372 unsigned nr_args)
373 {
374 struct io_uring_rsrc_update __user *arg = __arg;
375 struct io_uring_task *tctx = current->io_uring;
376 struct io_uring_rsrc_update reg;
377 int ret = 0, i;
378
379 if (!nr_args || nr_args > IO_RINGFD_REG_MAX)
380 return -EINVAL;
381 if (!tctx)
382 return 0;
383
384 for (i = 0; i < nr_args; i++) {
385 if (copy_from_user(®, &arg[i], sizeof(reg))) {
386 ret = -EFAULT;
387 break;
388 }
389 if (reg.resv || reg.data || reg.offset >= IO_RINGFD_REG_MAX) {
390 ret = -EINVAL;
391 break;
392 }
393
394 reg.offset = array_index_nospec(reg.offset, IO_RINGFD_REG_MAX);
395 if (tctx->registered_rings[reg.offset]) {
396 fput(tctx->registered_rings[reg.offset]);
397 tctx->registered_rings[reg.offset] = NULL;
398 }
399 }
400
401 return i ? i : ret;
402 }
403
__io_uring_fork(struct task_struct * tsk)404 int __io_uring_fork(struct task_struct *tsk)
405 {
406 struct io_restriction *res, *src = tsk->io_uring_restrict;
407
408 /* Don't leave it dangling on error */
409 tsk->io_uring_restrict = NULL;
410
411 res = kzalloc_obj(*res, GFP_KERNEL_ACCOUNT);
412 if (!res)
413 return -ENOMEM;
414
415 tsk->io_uring_restrict = res;
416 io_restriction_clone(res, src);
417 return 0;
418 }
419