tctx.c

  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
 15static 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
 39	/* Do QD, or 4 * CPUS, whatever is smallest */
 40	concurrency = min(ctx->sq_entries, 4 * num_online_cpus());
 41
 42	return io_wq_create(concurrency, &data);
 43}
 44
 45void __io_uring_free(struct task_struct *tsk)
 46{
 47	struct io_uring_task *tctx = tsk->io_uring;
 48	struct io_tctx_node *node;
 49	unsigned long index;
 50
 51	/*
 52	 * Fault injection forcing allocation errors in the xa_store() path
 53	 * can lead to xa_empty() returning false, even though no actual
 54	 * node is stored in the xarray. Until that gets sorted out, attempt
 55	 * an iteration here and warn if any entries are found.
 56	 */
 57	xa_for_each(&tctx->xa, index, node) {
 58		WARN_ON_ONCE(1);
 59		break;
 60	}
 61	WARN_ON_ONCE(tctx->io_wq);
 62	WARN_ON_ONCE(tctx->cached_refs);
 63
 64	percpu_counter_destroy(&tctx->inflight);
 65	kfree(tctx);
 66	tsk->io_uring = NULL;
 67}
 68
 69__cold int io_uring_alloc_task_context(struct task_struct *task,
 70				       struct io_ring_ctx *ctx)
 71{
 72	struct io_uring_task *tctx;
 73	int ret;
 74
 75	tctx = kzalloc(sizeof(*tctx), GFP_KERNEL);
 76	if (unlikely(!tctx))
 77		return -ENOMEM;
 78
 79	ret = percpu_counter_init(&tctx->inflight, 0, GFP_KERNEL);
 80	if (unlikely(ret)) {
 81		kfree(tctx);
 82		return ret;
 83	}
 84
 85	tctx->io_wq = io_init_wq_offload(ctx, task);
 86	if (IS_ERR(tctx->io_wq)) {
 87		ret = PTR_ERR(tctx->io_wq);
 88		percpu_counter_destroy(&tctx->inflight);
 89		kfree(tctx);
 90		return ret;
 91	}
 92
 93	tctx->task = task;
 94	xa_init(&tctx->xa);
 95	init_waitqueue_head(&tctx->wait);
 96	atomic_set(&tctx->in_cancel, 0);
 97	atomic_set(&tctx->inflight_tracked, 0);
 98	task->io_uring = tctx;
 99	init_llist_head(&tctx->task_list);
100	init_task_work(&tctx->task_work, tctx_task_work);
101	return 0;
102}
103
104int __io_uring_add_tctx_node(struct io_ring_ctx *ctx)
105{
106	struct io_uring_task *tctx = current->io_uring;
107	struct io_tctx_node *node;
108	int ret;
109
110	if (unlikely(!tctx)) {
111		ret = io_uring_alloc_task_context(current, ctx);
112		if (unlikely(ret))
113			return ret;
114
115		tctx = current->io_uring;
116		if (ctx->iowq_limits_set) {
117			unsigned int limits[2] = { ctx->iowq_limits[0],
118						   ctx->iowq_limits[1], };
119
120			ret = io_wq_max_workers(tctx->io_wq, limits);
121			if (ret)
122				return ret;
123		}
124	}
125
126	/*
127	 * Re-activate io-wq keepalive on any new io_uring usage. The wq may have
128	 * been marked for idle-exit when the task temporarily had no active
129	 * io_uring instances.
130	 */
131	if (tctx->io_wq)
132		io_wq_set_exit_on_idle(tctx->io_wq, false);
133	if (!xa_load(&tctx->xa, (unsigned long)ctx)) {
134		node = kmalloc(sizeof(*node), GFP_KERNEL);
135		if (!node)
136			return -ENOMEM;
137		node->ctx = ctx;
138		node->task = current;
139
140		ret = xa_err(xa_store(&tctx->xa, (unsigned long)ctx,
141					node, GFP_KERNEL));
142		if (ret) {
143			kfree(node);
144			return ret;
145		}
146
147		mutex_lock(&ctx->tctx_lock);
148		list_add(&node->ctx_node, &ctx->tctx_list);
149		mutex_unlock(&ctx->tctx_lock);
150	}
151	return 0;
152}
153
154int __io_uring_add_tctx_node_from_submit(struct io_ring_ctx *ctx)
155{
156	int ret;
157
158	if (ctx->flags & IORING_SETUP_SINGLE_ISSUER
159	    && ctx->submitter_task != current)
160		return -EEXIST;
161
162	ret = __io_uring_add_tctx_node(ctx);
163	if (ret)
164		return ret;
165
166	current->io_uring->last = ctx;
167	return 0;
168}
169
170/*
171 * Remove this io_uring_file -> task mapping.
172 */
173__cold void io_uring_del_tctx_node(unsigned long index)
174{
175	struct io_uring_task *tctx = current->io_uring;
176	struct io_tctx_node *node;
177
178	if (!tctx)
179		return;
180	node = xa_erase(&tctx->xa, index);
181	if (!node)
182		return;
183
184	WARN_ON_ONCE(current != node->task);
185	WARN_ON_ONCE(list_empty(&node->ctx_node));
186
187	mutex_lock(&node->ctx->tctx_lock);
188	list_del(&node->ctx_node);
189	mutex_unlock(&node->ctx->tctx_lock);
190
191	if (tctx->last == node->ctx)
192		tctx->last = NULL;
193	kfree(node);
194
195	if (xa_empty(&tctx->xa) && tctx->io_wq)
196		io_wq_set_exit_on_idle(tctx->io_wq, true);
197}
198
199__cold void io_uring_clean_tctx(struct io_uring_task *tctx)
200{
201	struct io_wq *wq = tctx->io_wq;
202	struct io_tctx_node *node;
203	unsigned long index;
204
205	xa_for_each(&tctx->xa, index, node) {
206		io_uring_del_tctx_node(index);
207		cond_resched();
208	}
209	if (wq) {
210		/*
211		 * Must be after io_uring_del_tctx_node() (removes nodes under
212		 * uring_lock) to avoid race with io_uring_try_cancel_iowq().
213		 */
214		io_wq_put_and_exit(wq);
215		tctx->io_wq = NULL;
216	}
217}
218
219void io_uring_unreg_ringfd(void)
220{
221	struct io_uring_task *tctx = current->io_uring;
222	int i;
223
224	for (i = 0; i < IO_RINGFD_REG_MAX; i++) {
225		if (tctx->registered_rings[i]) {
226			fput(tctx->registered_rings[i]);
227			tctx->registered_rings[i] = NULL;
228		}
229	}
230}
231
232int io_ring_add_registered_file(struct io_uring_task *tctx, struct file *file,
233				     int start, int end)
234{
235	int offset;
236	for (offset = start; offset < end; offset++) {
237		offset = array_index_nospec(offset, IO_RINGFD_REG_MAX);
238		if (tctx->registered_rings[offset])
239			continue;
240
241		tctx->registered_rings[offset] = file;
242		return offset;
243	}
244	return -EBUSY;
245}
246
247static int io_ring_add_registered_fd(struct io_uring_task *tctx, int fd,
248				     int start, int end)
249{
250	struct file *file;
251	int offset;
252
253	file = fget(fd);
254	if (!file) {
255		return -EBADF;
256	} else if (!io_is_uring_fops(file)) {
257		fput(file);
258		return -EOPNOTSUPP;
259	}
260	offset = io_ring_add_registered_file(tctx, file, start, end);
261	if (offset < 0)
262		fput(file);
263	return offset;
264}
265
266/*
267 * Register a ring fd to avoid fdget/fdput for each io_uring_enter()
268 * invocation. User passes in an array of struct io_uring_rsrc_update
269 * with ->data set to the ring_fd, and ->offset given for the desired
270 * index. If no index is desired, application may set ->offset == -1U
271 * and we'll find an available index. Returns number of entries
272 * successfully processed, or < 0 on error if none were processed.
273 */
274int io_ringfd_register(struct io_ring_ctx *ctx, void __user *__arg,
275		       unsigned nr_args)
276{
277	struct io_uring_rsrc_update __user *arg = __arg;
278	struct io_uring_rsrc_update reg;
279	struct io_uring_task *tctx;
280	int ret, i;
281
282	if (!nr_args || nr_args > IO_RINGFD_REG_MAX)
283		return -EINVAL;
284
285	mutex_unlock(&ctx->uring_lock);
286	ret = __io_uring_add_tctx_node(ctx);
287	mutex_lock(&ctx->uring_lock);
288	if (ret)
289		return ret;
290
291	tctx = current->io_uring;
292	for (i = 0; i < nr_args; i++) {
293		int start, end;
294
295		if (copy_from_user(&reg, &arg[i], sizeof(reg))) {
296			ret = -EFAULT;
297			break;
298		}
299
300		if (reg.resv) {
301			ret = -EINVAL;
302			break;
303		}
304
305		if (reg.offset == -1U) {
306			start = 0;
307			end = IO_RINGFD_REG_MAX;
308		} else {
309			if (reg.offset >= IO_RINGFD_REG_MAX) {
310				ret = -EINVAL;
311				break;
312			}
313			start = reg.offset;
314			end = start + 1;
315		}
316
317		ret = io_ring_add_registered_fd(tctx, reg.data, start, end);
318		if (ret < 0)
319			break;
320
321		reg.offset = ret;
322		if (copy_to_user(&arg[i], &reg, sizeof(reg))) {
323			fput(tctx->registered_rings[reg.offset]);
324			tctx->registered_rings[reg.offset] = NULL;
325			ret = -EFAULT;
326			break;
327		}
328	}
329
330	return i ? i : ret;
331}
332
333int io_ringfd_unregister(struct io_ring_ctx *ctx, void __user *__arg,
334			 unsigned nr_args)
335{
336	struct io_uring_rsrc_update __user *arg = __arg;
337	struct io_uring_task *tctx = current->io_uring;
338	struct io_uring_rsrc_update reg;
339	int ret = 0, i;
340
341	if (!nr_args || nr_args > IO_RINGFD_REG_MAX)
342		return -EINVAL;
343	if (!tctx)
344		return 0;
345
346	for (i = 0; i < nr_args; i++) {
347		if (copy_from_user(&reg, &arg[i], sizeof(reg))) {
348			ret = -EFAULT;
349			break;
350		}
351		if (reg.resv || reg.data || reg.offset >= IO_RINGFD_REG_MAX) {
352			ret = -EINVAL;
353			break;
354		}
355
356		reg.offset = array_index_nospec(reg.offset, IO_RINGFD_REG_MAX);
357		if (tctx->registered_rings[reg.offset]) {
358			fput(tctx->registered_rings[reg.offset]);
359			tctx->registered_rings[reg.offset] = NULL;
360		}
361	}
362
363	return i ? i : ret;
364}