xref: /openbmc/linux/fs/cachefiles/io.c (revision 355f841a)
1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /* kiocb-using read/write
3  *
4  * Copyright (C) 2021 Red Hat, Inc. All Rights Reserved.
5  * Written by David Howells (dhowells@redhat.com)
6  */
7 
8 #include <linux/mount.h>
9 #include <linux/slab.h>
10 #include <linux/file.h>
11 #include <linux/uio.h>
12 #include <linux/falloc.h>
13 #include <linux/sched/mm.h>
14 #include <trace/events/fscache.h>
15 #include "internal.h"
16 
17 struct cachefiles_kiocb {
18 	struct kiocb		iocb;
19 	refcount_t		ki_refcnt;
20 	loff_t			start;
21 	union {
22 		size_t		skipped;
23 		size_t		len;
24 	};
25 	struct cachefiles_object *object;
26 	netfs_io_terminated_t	term_func;
27 	void			*term_func_priv;
28 	bool			was_async;
29 	unsigned int		inval_counter;	/* Copy of cookie->inval_counter */
30 	u64			b_writing;
31 };
32 
33 static inline void cachefiles_put_kiocb(struct cachefiles_kiocb *ki)
34 {
35 	if (refcount_dec_and_test(&ki->ki_refcnt)) {
36 		cachefiles_put_object(ki->object, cachefiles_obj_put_ioreq);
37 		fput(ki->iocb.ki_filp);
38 		kfree(ki);
39 	}
40 }
41 
42 /*
43  * Handle completion of a read from the cache.
44  */
45 static void cachefiles_read_complete(struct kiocb *iocb, long ret)
46 {
47 	struct cachefiles_kiocb *ki = container_of(iocb, struct cachefiles_kiocb, iocb);
48 	struct inode *inode = file_inode(ki->iocb.ki_filp);
49 
50 	_enter("%ld", ret);
51 
52 	if (ret < 0)
53 		trace_cachefiles_io_error(ki->object, inode, ret,
54 					  cachefiles_trace_read_error);
55 
56 	if (ki->term_func) {
57 		if (ret >= 0) {
58 			if (ki->object->cookie->inval_counter == ki->inval_counter)
59 				ki->skipped += ret;
60 			else
61 				ret = -ESTALE;
62 		}
63 
64 		ki->term_func(ki->term_func_priv, ret, ki->was_async);
65 	}
66 
67 	cachefiles_put_kiocb(ki);
68 }
69 
70 /*
71  * Initiate a read from the cache.
72  */
73 static int cachefiles_read(struct netfs_cache_resources *cres,
74 			   loff_t start_pos,
75 			   struct iov_iter *iter,
76 			   enum netfs_read_from_hole read_hole,
77 			   netfs_io_terminated_t term_func,
78 			   void *term_func_priv)
79 {
80 	struct cachefiles_object *object;
81 	struct cachefiles_kiocb *ki;
82 	struct file *file;
83 	unsigned int old_nofs;
84 	ssize_t ret = -ENOBUFS;
85 	size_t len = iov_iter_count(iter), skipped = 0;
86 
87 	if (!fscache_wait_for_operation(cres, FSCACHE_WANT_READ))
88 		goto presubmission_error;
89 
90 	fscache_count_read();
91 	object = cachefiles_cres_object(cres);
92 	file = cachefiles_cres_file(cres);
93 
94 	_enter("%pD,%li,%llx,%zx/%llx",
95 	       file, file_inode(file)->i_ino, start_pos, len,
96 	       i_size_read(file_inode(file)));
97 
98 	/* If the caller asked us to seek for data before doing the read, then
99 	 * we should do that now.  If we find a gap, we fill it with zeros.
100 	 */
101 	if (read_hole != NETFS_READ_HOLE_IGNORE) {
102 		loff_t off = start_pos, off2;
103 
104 		off2 = cachefiles_inject_read_error();
105 		if (off2 == 0)
106 			off2 = vfs_llseek(file, off, SEEK_DATA);
107 		if (off2 < 0 && off2 >= (loff_t)-MAX_ERRNO && off2 != -ENXIO) {
108 			skipped = 0;
109 			ret = off2;
110 			goto presubmission_error;
111 		}
112 
113 		if (off2 == -ENXIO || off2 >= start_pos + len) {
114 			/* The region is beyond the EOF or there's no more data
115 			 * in the region, so clear the rest of the buffer and
116 			 * return success.
117 			 */
118 			ret = -ENODATA;
119 			if (read_hole == NETFS_READ_HOLE_FAIL)
120 				goto presubmission_error;
121 
122 			iov_iter_zero(len, iter);
123 			skipped = len;
124 			ret = 0;
125 			goto presubmission_error;
126 		}
127 
128 		skipped = off2 - off;
129 		iov_iter_zero(skipped, iter);
130 	}
131 
132 	ret = -ENOMEM;
133 	ki = kzalloc(sizeof(struct cachefiles_kiocb), GFP_KERNEL);
134 	if (!ki)
135 		goto presubmission_error;
136 
137 	refcount_set(&ki->ki_refcnt, 2);
138 	ki->iocb.ki_filp	= file;
139 	ki->iocb.ki_pos		= start_pos + skipped;
140 	ki->iocb.ki_flags	= IOCB_DIRECT;
141 	ki->iocb.ki_hint	= ki_hint_validate(file_write_hint(file));
142 	ki->iocb.ki_ioprio	= get_current_ioprio();
143 	ki->skipped		= skipped;
144 	ki->object		= object;
145 	ki->inval_counter	= cres->inval_counter;
146 	ki->term_func		= term_func;
147 	ki->term_func_priv	= term_func_priv;
148 	ki->was_async		= true;
149 
150 	if (ki->term_func)
151 		ki->iocb.ki_complete = cachefiles_read_complete;
152 
153 	get_file(ki->iocb.ki_filp);
154 	cachefiles_grab_object(object, cachefiles_obj_get_ioreq);
155 
156 	trace_cachefiles_read(object, file_inode(file), ki->iocb.ki_pos, len - skipped);
157 	old_nofs = memalloc_nofs_save();
158 	ret = cachefiles_inject_read_error();
159 	if (ret == 0)
160 		ret = vfs_iocb_iter_read(file, &ki->iocb, iter);
161 	memalloc_nofs_restore(old_nofs);
162 	switch (ret) {
163 	case -EIOCBQUEUED:
164 		goto in_progress;
165 
166 	case -ERESTARTSYS:
167 	case -ERESTARTNOINTR:
168 	case -ERESTARTNOHAND:
169 	case -ERESTART_RESTARTBLOCK:
170 		/* There's no easy way to restart the syscall since other AIO's
171 		 * may be already running. Just fail this IO with EINTR.
172 		 */
173 		ret = -EINTR;
174 		fallthrough;
175 	default:
176 		ki->was_async = false;
177 		cachefiles_read_complete(&ki->iocb, ret);
178 		if (ret > 0)
179 			ret = 0;
180 		break;
181 	}
182 
183 in_progress:
184 	cachefiles_put_kiocb(ki);
185 	_leave(" = %zd", ret);
186 	return ret;
187 
188 presubmission_error:
189 	if (term_func)
190 		term_func(term_func_priv, ret < 0 ? ret : skipped, false);
191 	return ret;
192 }
193 
194 /*
195  * Handle completion of a write to the cache.
196  */
197 static void cachefiles_write_complete(struct kiocb *iocb, long ret)
198 {
199 	struct cachefiles_kiocb *ki = container_of(iocb, struct cachefiles_kiocb, iocb);
200 	struct cachefiles_object *object = ki->object;
201 	struct inode *inode = file_inode(ki->iocb.ki_filp);
202 
203 	_enter("%ld", ret);
204 
205 	/* Tell lockdep we inherited freeze protection from submission thread */
206 	__sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE);
207 	__sb_end_write(inode->i_sb, SB_FREEZE_WRITE);
208 
209 	if (ret < 0)
210 		trace_cachefiles_io_error(object, inode, ret,
211 					  cachefiles_trace_write_error);
212 
213 	atomic_long_sub(ki->b_writing, &object->volume->cache->b_writing);
214 	set_bit(FSCACHE_COOKIE_HAVE_DATA, &object->cookie->flags);
215 	if (ki->term_func)
216 		ki->term_func(ki->term_func_priv, ret, ki->was_async);
217 	cachefiles_put_kiocb(ki);
218 }
219 
220 /*
221  * Initiate a write to the cache.
222  */
223 static int cachefiles_write(struct netfs_cache_resources *cres,
224 			    loff_t start_pos,
225 			    struct iov_iter *iter,
226 			    netfs_io_terminated_t term_func,
227 			    void *term_func_priv)
228 {
229 	struct cachefiles_object *object;
230 	struct cachefiles_cache *cache;
231 	struct cachefiles_kiocb *ki;
232 	struct inode *inode;
233 	struct file *file;
234 	unsigned int old_nofs;
235 	ssize_t ret = -ENOBUFS;
236 	size_t len = iov_iter_count(iter);
237 
238 	if (!fscache_wait_for_operation(cres, FSCACHE_WANT_WRITE))
239 		goto presubmission_error;
240 	fscache_count_write();
241 	object = cachefiles_cres_object(cres);
242 	cache = object->volume->cache;
243 	file = cachefiles_cres_file(cres);
244 
245 	_enter("%pD,%li,%llx,%zx/%llx",
246 	       file, file_inode(file)->i_ino, start_pos, len,
247 	       i_size_read(file_inode(file)));
248 
249 	ret = -ENOMEM;
250 	ki = kzalloc(sizeof(struct cachefiles_kiocb), GFP_KERNEL);
251 	if (!ki)
252 		goto presubmission_error;
253 
254 	refcount_set(&ki->ki_refcnt, 2);
255 	ki->iocb.ki_filp	= file;
256 	ki->iocb.ki_pos		= start_pos;
257 	ki->iocb.ki_flags	= IOCB_DIRECT | IOCB_WRITE;
258 	ki->iocb.ki_hint	= ki_hint_validate(file_write_hint(file));
259 	ki->iocb.ki_ioprio	= get_current_ioprio();
260 	ki->object		= object;
261 	ki->inval_counter	= cres->inval_counter;
262 	ki->start		= start_pos;
263 	ki->len			= len;
264 	ki->term_func		= term_func;
265 	ki->term_func_priv	= term_func_priv;
266 	ki->was_async		= true;
267 	ki->b_writing		= (len + (1 << cache->bshift) - 1) >> cache->bshift;
268 
269 	if (ki->term_func)
270 		ki->iocb.ki_complete = cachefiles_write_complete;
271 	atomic_long_add(ki->b_writing, &cache->b_writing);
272 
273 	/* Open-code file_start_write here to grab freeze protection, which
274 	 * will be released by another thread in aio_complete_rw().  Fool
275 	 * lockdep by telling it the lock got released so that it doesn't
276 	 * complain about the held lock when we return to userspace.
277 	 */
278 	inode = file_inode(file);
279 	__sb_start_write(inode->i_sb, SB_FREEZE_WRITE);
280 	__sb_writers_release(inode->i_sb, SB_FREEZE_WRITE);
281 
282 	get_file(ki->iocb.ki_filp);
283 	cachefiles_grab_object(object, cachefiles_obj_get_ioreq);
284 
285 	trace_cachefiles_write(object, inode, ki->iocb.ki_pos, len);
286 	old_nofs = memalloc_nofs_save();
287 	ret = cachefiles_inject_write_error();
288 	if (ret == 0)
289 		ret = vfs_iocb_iter_write(file, &ki->iocb, iter);
290 	memalloc_nofs_restore(old_nofs);
291 	switch (ret) {
292 	case -EIOCBQUEUED:
293 		goto in_progress;
294 
295 	case -ERESTARTSYS:
296 	case -ERESTARTNOINTR:
297 	case -ERESTARTNOHAND:
298 	case -ERESTART_RESTARTBLOCK:
299 		/* There's no easy way to restart the syscall since other AIO's
300 		 * may be already running. Just fail this IO with EINTR.
301 		 */
302 		ret = -EINTR;
303 		fallthrough;
304 	default:
305 		ki->was_async = false;
306 		cachefiles_write_complete(&ki->iocb, ret);
307 		if (ret > 0)
308 			ret = 0;
309 		break;
310 	}
311 
312 in_progress:
313 	cachefiles_put_kiocb(ki);
314 	_leave(" = %zd", ret);
315 	return ret;
316 
317 presubmission_error:
318 	if (term_func)
319 		term_func(term_func_priv, ret, false);
320 	return ret;
321 }
322 
323 /*
324  * Prepare a read operation, shortening it to a cached/uncached
325  * boundary as appropriate.
326  */
327 static enum netfs_read_source cachefiles_prepare_read(struct netfs_read_subrequest *subreq,
328 						      loff_t i_size)
329 {
330 	enum cachefiles_prepare_read_trace why;
331 	struct netfs_read_request *rreq = subreq->rreq;
332 	struct netfs_cache_resources *cres = &rreq->cache_resources;
333 	struct cachefiles_object *object;
334 	struct cachefiles_cache *cache;
335 	struct fscache_cookie *cookie = fscache_cres_cookie(cres);
336 	const struct cred *saved_cred;
337 	struct file *file = cachefiles_cres_file(cres);
338 	enum netfs_read_source ret = NETFS_DOWNLOAD_FROM_SERVER;
339 	loff_t off, to;
340 	ino_t ino = file ? file_inode(file)->i_ino : 0;
341 
342 	_enter("%zx @%llx/%llx", subreq->len, subreq->start, i_size);
343 
344 	if (subreq->start >= i_size) {
345 		ret = NETFS_FILL_WITH_ZEROES;
346 		why = cachefiles_trace_read_after_eof;
347 		goto out_no_object;
348 	}
349 
350 	if (test_bit(FSCACHE_COOKIE_NO_DATA_TO_READ, &cookie->flags)) {
351 		__set_bit(NETFS_SREQ_WRITE_TO_CACHE, &subreq->flags);
352 		why = cachefiles_trace_read_no_data;
353 		goto out_no_object;
354 	}
355 
356 	/* The object and the file may be being created in the background. */
357 	if (!file) {
358 		why = cachefiles_trace_read_no_file;
359 		if (!fscache_wait_for_operation(cres, FSCACHE_WANT_READ))
360 			goto out_no_object;
361 		file = cachefiles_cres_file(cres);
362 		if (!file)
363 			goto out_no_object;
364 		ino = file_inode(file)->i_ino;
365 	}
366 
367 	object = cachefiles_cres_object(cres);
368 	cache = object->volume->cache;
369 	cachefiles_begin_secure(cache, &saved_cred);
370 
371 	off = cachefiles_inject_read_error();
372 	if (off == 0)
373 		off = vfs_llseek(file, subreq->start, SEEK_DATA);
374 	if (off < 0 && off >= (loff_t)-MAX_ERRNO) {
375 		if (off == (loff_t)-ENXIO) {
376 			why = cachefiles_trace_read_seek_nxio;
377 			goto download_and_store;
378 		}
379 		trace_cachefiles_io_error(object, file_inode(file), off,
380 					  cachefiles_trace_seek_error);
381 		why = cachefiles_trace_read_seek_error;
382 		goto out;
383 	}
384 
385 	if (off >= subreq->start + subreq->len) {
386 		why = cachefiles_trace_read_found_hole;
387 		goto download_and_store;
388 	}
389 
390 	if (off > subreq->start) {
391 		off = round_up(off, cache->bsize);
392 		subreq->len = off - subreq->start;
393 		why = cachefiles_trace_read_found_part;
394 		goto download_and_store;
395 	}
396 
397 	to = cachefiles_inject_read_error();
398 	if (to == 0)
399 		to = vfs_llseek(file, subreq->start, SEEK_HOLE);
400 	if (to < 0 && to >= (loff_t)-MAX_ERRNO) {
401 		trace_cachefiles_io_error(object, file_inode(file), to,
402 					  cachefiles_trace_seek_error);
403 		why = cachefiles_trace_read_seek_error;
404 		goto out;
405 	}
406 
407 	if (to < subreq->start + subreq->len) {
408 		if (subreq->start + subreq->len >= i_size)
409 			to = round_up(to, cache->bsize);
410 		else
411 			to = round_down(to, cache->bsize);
412 		subreq->len = to - subreq->start;
413 	}
414 
415 	why = cachefiles_trace_read_have_data;
416 	ret = NETFS_READ_FROM_CACHE;
417 	goto out;
418 
419 download_and_store:
420 	__set_bit(NETFS_SREQ_WRITE_TO_CACHE, &subreq->flags);
421 out:
422 	cachefiles_end_secure(cache, saved_cred);
423 out_no_object:
424 	trace_cachefiles_prep_read(subreq, ret, why, ino);
425 	return ret;
426 }
427 
428 /*
429  * Prepare for a write to occur.
430  */
431 static int __cachefiles_prepare_write(struct netfs_cache_resources *cres,
432 				      loff_t *_start, size_t *_len, loff_t i_size,
433 				      bool no_space_allocated_yet)
434 {
435 	struct cachefiles_object *object = cachefiles_cres_object(cres);
436 	struct cachefiles_cache *cache = object->volume->cache;
437 	struct file *file = cachefiles_cres_file(cres);
438 	loff_t start = *_start, pos;
439 	size_t len = *_len, down;
440 	int ret;
441 
442 	/* Round to DIO size */
443 	down = start - round_down(start, PAGE_SIZE);
444 	*_start = start - down;
445 	*_len = round_up(down + len, PAGE_SIZE);
446 
447 	/* We need to work out whether there's sufficient disk space to perform
448 	 * the write - but we can skip that check if we have space already
449 	 * allocated.
450 	 */
451 	if (no_space_allocated_yet)
452 		goto check_space;
453 
454 	pos = cachefiles_inject_read_error();
455 	if (pos == 0)
456 		pos = vfs_llseek(file, *_start, SEEK_DATA);
457 	if (pos < 0 && pos >= (loff_t)-MAX_ERRNO) {
458 		if (pos == -ENXIO)
459 			goto check_space; /* Unallocated tail */
460 		trace_cachefiles_io_error(object, file_inode(file), pos,
461 					  cachefiles_trace_seek_error);
462 		return pos;
463 	}
464 	if ((u64)pos >= (u64)*_start + *_len)
465 		goto check_space; /* Unallocated region */
466 
467 	/* We have a block that's at least partially filled - if we're low on
468 	 * space, we need to see if it's fully allocated.  If it's not, we may
469 	 * want to cull it.
470 	 */
471 	if (cachefiles_has_space(cache, 0, *_len / PAGE_SIZE,
472 				 cachefiles_has_space_check) == 0)
473 		return 0; /* Enough space to simply overwrite the whole block */
474 
475 	pos = cachefiles_inject_read_error();
476 	if (pos == 0)
477 		pos = vfs_llseek(file, *_start, SEEK_HOLE);
478 	if (pos < 0 && pos >= (loff_t)-MAX_ERRNO) {
479 		trace_cachefiles_io_error(object, file_inode(file), pos,
480 					  cachefiles_trace_seek_error);
481 		return pos;
482 	}
483 	if ((u64)pos >= (u64)*_start + *_len)
484 		return 0; /* Fully allocated */
485 
486 	/* Partially allocated, but insufficient space: cull. */
487 	fscache_count_no_write_space();
488 	ret = cachefiles_inject_remove_error();
489 	if (ret == 0)
490 		ret = vfs_fallocate(file, FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE,
491 				    *_start, *_len);
492 	if (ret < 0) {
493 		trace_cachefiles_io_error(object, file_inode(file), ret,
494 					  cachefiles_trace_fallocate_error);
495 		cachefiles_io_error_obj(object,
496 					"CacheFiles: fallocate failed (%d)\n", ret);
497 		ret = -EIO;
498 	}
499 
500 	return ret;
501 
502 check_space:
503 	return cachefiles_has_space(cache, 0, *_len / PAGE_SIZE,
504 				    cachefiles_has_space_for_write);
505 }
506 
507 static int cachefiles_prepare_write(struct netfs_cache_resources *cres,
508 				    loff_t *_start, size_t *_len, loff_t i_size,
509 				    bool no_space_allocated_yet)
510 {
511 	struct cachefiles_object *object = cachefiles_cres_object(cres);
512 	struct cachefiles_cache *cache = object->volume->cache;
513 	const struct cred *saved_cred;
514 	int ret;
515 
516 	if (!cachefiles_cres_file(cres)) {
517 		if (!fscache_wait_for_operation(cres, FSCACHE_WANT_WRITE))
518 			return -ENOBUFS;
519 		if (!cachefiles_cres_file(cres))
520 			return -ENOBUFS;
521 	}
522 
523 	cachefiles_begin_secure(cache, &saved_cred);
524 	ret = __cachefiles_prepare_write(cres, _start, _len, i_size,
525 					 no_space_allocated_yet);
526 	cachefiles_end_secure(cache, saved_cred);
527 	return ret;
528 }
529 
530 /*
531  * Clean up an operation.
532  */
533 static void cachefiles_end_operation(struct netfs_cache_resources *cres)
534 {
535 	struct file *file = cachefiles_cres_file(cres);
536 
537 	if (file)
538 		fput(file);
539 	fscache_end_cookie_access(fscache_cres_cookie(cres), fscache_access_io_end);
540 }
541 
542 static const struct netfs_cache_ops cachefiles_netfs_cache_ops = {
543 	.end_operation		= cachefiles_end_operation,
544 	.read			= cachefiles_read,
545 	.write			= cachefiles_write,
546 	.prepare_read		= cachefiles_prepare_read,
547 	.prepare_write		= cachefiles_prepare_write,
548 };
549 
550 /*
551  * Open the cache file when beginning a cache operation.
552  */
553 bool cachefiles_begin_operation(struct netfs_cache_resources *cres,
554 				enum fscache_want_state want_state)
555 {
556 	struct cachefiles_object *object = cachefiles_cres_object(cres);
557 
558 	if (!cachefiles_cres_file(cres)) {
559 		cres->ops = &cachefiles_netfs_cache_ops;
560 		if (object->file) {
561 			spin_lock(&object->lock);
562 			if (!cres->cache_priv2 && object->file)
563 				cres->cache_priv2 = get_file(object->file);
564 			spin_unlock(&object->lock);
565 		}
566 	}
567 
568 	if (!cachefiles_cres_file(cres) && want_state != FSCACHE_WANT_PARAMS) {
569 		pr_err("failed to get cres->file\n");
570 		return false;
571 	}
572 
573 	return true;
574 }
575