xref: /openbmc/linux/fs/ceph/file.c (revision 5104d265)
1 #include <linux/ceph/ceph_debug.h>
2 
3 #include <linux/module.h>
4 #include <linux/sched.h>
5 #include <linux/slab.h>
6 #include <linux/file.h>
7 #include <linux/mount.h>
8 #include <linux/namei.h>
9 #include <linux/writeback.h>
10 #include <linux/aio.h>
11 
12 #include "super.h"
13 #include "mds_client.h"
14 
15 /*
16  * Ceph file operations
17  *
18  * Implement basic open/close functionality, and implement
19  * read/write.
20  *
21  * We implement three modes of file I/O:
22  *  - buffered uses the generic_file_aio_{read,write} helpers
23  *
24  *  - synchronous is used when there is multi-client read/write
25  *    sharing, avoids the page cache, and synchronously waits for an
26  *    ack from the OSD.
27  *
28  *  - direct io takes the variant of the sync path that references
29  *    user pages directly.
30  *
31  * fsync() flushes and waits on dirty pages, but just queues metadata
32  * for writeback: since the MDS can recover size and mtime there is no
33  * need to wait for MDS acknowledgement.
34  */
35 
36 
37 /*
38  * Prepare an open request.  Preallocate ceph_cap to avoid an
39  * inopportune ENOMEM later.
40  */
41 static struct ceph_mds_request *
42 prepare_open_request(struct super_block *sb, int flags, int create_mode)
43 {
44 	struct ceph_fs_client *fsc = ceph_sb_to_client(sb);
45 	struct ceph_mds_client *mdsc = fsc->mdsc;
46 	struct ceph_mds_request *req;
47 	int want_auth = USE_ANY_MDS;
48 	int op = (flags & O_CREAT) ? CEPH_MDS_OP_CREATE : CEPH_MDS_OP_OPEN;
49 
50 	if (flags & (O_WRONLY|O_RDWR|O_CREAT|O_TRUNC))
51 		want_auth = USE_AUTH_MDS;
52 
53 	req = ceph_mdsc_create_request(mdsc, op, want_auth);
54 	if (IS_ERR(req))
55 		goto out;
56 	req->r_fmode = ceph_flags_to_mode(flags);
57 	req->r_args.open.flags = cpu_to_le32(flags);
58 	req->r_args.open.mode = cpu_to_le32(create_mode);
59 out:
60 	return req;
61 }
62 
63 /*
64  * initialize private struct file data.
65  * if we fail, clean up by dropping fmode reference on the ceph_inode
66  */
67 static int ceph_init_file(struct inode *inode, struct file *file, int fmode)
68 {
69 	struct ceph_file_info *cf;
70 	int ret = 0;
71 
72 	switch (inode->i_mode & S_IFMT) {
73 	case S_IFREG:
74 	case S_IFDIR:
75 		dout("init_file %p %p 0%o (regular)\n", inode, file,
76 		     inode->i_mode);
77 		cf = kmem_cache_alloc(ceph_file_cachep, GFP_NOFS | __GFP_ZERO);
78 		if (cf == NULL) {
79 			ceph_put_fmode(ceph_inode(inode), fmode); /* clean up */
80 			return -ENOMEM;
81 		}
82 		cf->fmode = fmode;
83 		cf->next_offset = 2;
84 		file->private_data = cf;
85 		BUG_ON(inode->i_fop->release != ceph_release);
86 		break;
87 
88 	case S_IFLNK:
89 		dout("init_file %p %p 0%o (symlink)\n", inode, file,
90 		     inode->i_mode);
91 		ceph_put_fmode(ceph_inode(inode), fmode); /* clean up */
92 		break;
93 
94 	default:
95 		dout("init_file %p %p 0%o (special)\n", inode, file,
96 		     inode->i_mode);
97 		/*
98 		 * we need to drop the open ref now, since we don't
99 		 * have .release set to ceph_release.
100 		 */
101 		ceph_put_fmode(ceph_inode(inode), fmode); /* clean up */
102 		BUG_ON(inode->i_fop->release == ceph_release);
103 
104 		/* call the proper open fop */
105 		ret = inode->i_fop->open(inode, file);
106 	}
107 	return ret;
108 }
109 
110 /*
111  * If we already have the requisite capabilities, we can satisfy
112  * the open request locally (no need to request new caps from the
113  * MDS).  We do, however, need to inform the MDS (asynchronously)
114  * if our wanted caps set expands.
115  */
116 int ceph_open(struct inode *inode, struct file *file)
117 {
118 	struct ceph_inode_info *ci = ceph_inode(inode);
119 	struct ceph_fs_client *fsc = ceph_sb_to_client(inode->i_sb);
120 	struct ceph_mds_client *mdsc = fsc->mdsc;
121 	struct ceph_mds_request *req;
122 	struct ceph_file_info *cf = file->private_data;
123 	struct inode *parent_inode = NULL;
124 	int err;
125 	int flags, fmode, wanted;
126 
127 	if (cf) {
128 		dout("open file %p is already opened\n", file);
129 		return 0;
130 	}
131 
132 	/* filter out O_CREAT|O_EXCL; vfs did that already.  yuck. */
133 	flags = file->f_flags & ~(O_CREAT|O_EXCL);
134 	if (S_ISDIR(inode->i_mode))
135 		flags = O_DIRECTORY;  /* mds likes to know */
136 
137 	dout("open inode %p ino %llx.%llx file %p flags %d (%d)\n", inode,
138 	     ceph_vinop(inode), file, flags, file->f_flags);
139 	fmode = ceph_flags_to_mode(flags);
140 	wanted = ceph_caps_for_mode(fmode);
141 
142 	/* snapped files are read-only */
143 	if (ceph_snap(inode) != CEPH_NOSNAP && (file->f_mode & FMODE_WRITE))
144 		return -EROFS;
145 
146 	/* trivially open snapdir */
147 	if (ceph_snap(inode) == CEPH_SNAPDIR) {
148 		spin_lock(&ci->i_ceph_lock);
149 		__ceph_get_fmode(ci, fmode);
150 		spin_unlock(&ci->i_ceph_lock);
151 		return ceph_init_file(inode, file, fmode);
152 	}
153 
154 	/*
155 	 * No need to block if we have caps on the auth MDS (for
156 	 * write) or any MDS (for read).  Update wanted set
157 	 * asynchronously.
158 	 */
159 	spin_lock(&ci->i_ceph_lock);
160 	if (__ceph_is_any_real_caps(ci) &&
161 	    (((fmode & CEPH_FILE_MODE_WR) == 0) || ci->i_auth_cap)) {
162 		int mds_wanted = __ceph_caps_mds_wanted(ci);
163 		int issued = __ceph_caps_issued(ci, NULL);
164 
165 		dout("open %p fmode %d want %s issued %s using existing\n",
166 		     inode, fmode, ceph_cap_string(wanted),
167 		     ceph_cap_string(issued));
168 		__ceph_get_fmode(ci, fmode);
169 		spin_unlock(&ci->i_ceph_lock);
170 
171 		/* adjust wanted? */
172 		if ((issued & wanted) != wanted &&
173 		    (mds_wanted & wanted) != wanted &&
174 		    ceph_snap(inode) != CEPH_SNAPDIR)
175 			ceph_check_caps(ci, 0, NULL);
176 
177 		return ceph_init_file(inode, file, fmode);
178 	} else if (ceph_snap(inode) != CEPH_NOSNAP &&
179 		   (ci->i_snap_caps & wanted) == wanted) {
180 		__ceph_get_fmode(ci, fmode);
181 		spin_unlock(&ci->i_ceph_lock);
182 		return ceph_init_file(inode, file, fmode);
183 	}
184 	spin_unlock(&ci->i_ceph_lock);
185 
186 	dout("open fmode %d wants %s\n", fmode, ceph_cap_string(wanted));
187 	req = prepare_open_request(inode->i_sb, flags, 0);
188 	if (IS_ERR(req)) {
189 		err = PTR_ERR(req);
190 		goto out;
191 	}
192 	req->r_inode = inode;
193 	ihold(inode);
194 	req->r_num_caps = 1;
195 	if (flags & (O_CREAT|O_TRUNC))
196 		parent_inode = ceph_get_dentry_parent_inode(file->f_dentry);
197 	err = ceph_mdsc_do_request(mdsc, parent_inode, req);
198 	iput(parent_inode);
199 	if (!err)
200 		err = ceph_init_file(inode, file, req->r_fmode);
201 	ceph_mdsc_put_request(req);
202 	dout("open result=%d on %llx.%llx\n", err, ceph_vinop(inode));
203 out:
204 	return err;
205 }
206 
207 
208 /*
209  * Do a lookup + open with a single request.  If we get a non-existent
210  * file or symlink, return 1 so the VFS can retry.
211  */
212 int ceph_atomic_open(struct inode *dir, struct dentry *dentry,
213 		     struct file *file, unsigned flags, umode_t mode,
214 		     int *opened)
215 {
216 	struct ceph_fs_client *fsc = ceph_sb_to_client(dir->i_sb);
217 	struct ceph_mds_client *mdsc = fsc->mdsc;
218 	struct ceph_mds_request *req;
219 	struct dentry *dn;
220 	int err;
221 
222 	dout("atomic_open %p dentry %p '%.*s' %s flags %d mode 0%o\n",
223 	     dir, dentry, dentry->d_name.len, dentry->d_name.name,
224 	     d_unhashed(dentry) ? "unhashed" : "hashed", flags, mode);
225 
226 	if (dentry->d_name.len > NAME_MAX)
227 		return -ENAMETOOLONG;
228 
229 	err = ceph_init_dentry(dentry);
230 	if (err < 0)
231 		return err;
232 
233 	/* do the open */
234 	req = prepare_open_request(dir->i_sb, flags, mode);
235 	if (IS_ERR(req))
236 		return PTR_ERR(req);
237 	req->r_dentry = dget(dentry);
238 	req->r_num_caps = 2;
239 	if (flags & O_CREAT) {
240 		req->r_dentry_drop = CEPH_CAP_FILE_SHARED;
241 		req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
242 	}
243 	req->r_locked_dir = dir;           /* caller holds dir->i_mutex */
244 	err = ceph_mdsc_do_request(mdsc,
245 				   (flags & (O_CREAT|O_TRUNC)) ? dir : NULL,
246 				   req);
247 	if (err)
248 		goto out_err;
249 
250 	err = ceph_handle_snapdir(req, dentry, err);
251 	if (err == 0 && (flags & O_CREAT) && !req->r_reply_info.head->is_dentry)
252 		err = ceph_handle_notrace_create(dir, dentry);
253 
254 	if (d_unhashed(dentry)) {
255 		dn = ceph_finish_lookup(req, dentry, err);
256 		if (IS_ERR(dn))
257 			err = PTR_ERR(dn);
258 	} else {
259 		/* we were given a hashed negative dentry */
260 		dn = NULL;
261 	}
262 	if (err)
263 		goto out_err;
264 	if (dn || dentry->d_inode == NULL || S_ISLNK(dentry->d_inode->i_mode)) {
265 		/* make vfs retry on splice, ENOENT, or symlink */
266 		dout("atomic_open finish_no_open on dn %p\n", dn);
267 		err = finish_no_open(file, dn);
268 	} else {
269 		dout("atomic_open finish_open on dn %p\n", dn);
270 		if (req->r_op == CEPH_MDS_OP_CREATE && req->r_reply_info.has_create_ino) {
271 			*opened |= FILE_CREATED;
272 		}
273 		err = finish_open(file, dentry, ceph_open, opened);
274 	}
275 
276 out_err:
277 	ceph_mdsc_put_request(req);
278 	dout("atomic_open result=%d\n", err);
279 	return err;
280 }
281 
282 int ceph_release(struct inode *inode, struct file *file)
283 {
284 	struct ceph_inode_info *ci = ceph_inode(inode);
285 	struct ceph_file_info *cf = file->private_data;
286 
287 	dout("release inode %p file %p\n", inode, file);
288 	ceph_put_fmode(ci, cf->fmode);
289 	if (cf->last_readdir)
290 		ceph_mdsc_put_request(cf->last_readdir);
291 	kfree(cf->last_name);
292 	kfree(cf->dir_info);
293 	dput(cf->dentry);
294 	kmem_cache_free(ceph_file_cachep, cf);
295 
296 	/* wake up anyone waiting for caps on this inode */
297 	wake_up_all(&ci->i_cap_wq);
298 	return 0;
299 }
300 
301 /*
302  * Read a range of bytes striped over one or more objects.  Iterate over
303  * objects we stripe over.  (That's not atomic, but good enough for now.)
304  *
305  * If we get a short result from the OSD, check against i_size; we need to
306  * only return a short read to the caller if we hit EOF.
307  */
308 static int striped_read(struct inode *inode,
309 			u64 off, u64 len,
310 			struct page **pages, int num_pages,
311 			int *checkeof, bool o_direct,
312 			unsigned long buf_align)
313 {
314 	struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
315 	struct ceph_inode_info *ci = ceph_inode(inode);
316 	u64 pos, this_len;
317 	int io_align, page_align;
318 	int left, pages_left;
319 	int read;
320 	struct page **page_pos;
321 	int ret;
322 	bool hit_stripe, was_short;
323 
324 	/*
325 	 * we may need to do multiple reads.  not atomic, unfortunately.
326 	 */
327 	pos = off;
328 	left = len;
329 	page_pos = pages;
330 	pages_left = num_pages;
331 	read = 0;
332 	io_align = off & ~PAGE_MASK;
333 
334 more:
335 	if (o_direct)
336 		page_align = (pos - io_align + buf_align) & ~PAGE_MASK;
337 	else
338 		page_align = pos & ~PAGE_MASK;
339 	this_len = left;
340 	ret = ceph_osdc_readpages(&fsc->client->osdc, ceph_vino(inode),
341 				  &ci->i_layout, pos, &this_len,
342 				  ci->i_truncate_seq,
343 				  ci->i_truncate_size,
344 				  page_pos, pages_left, page_align);
345 	if (ret == -ENOENT)
346 		ret = 0;
347 	hit_stripe = this_len < left;
348 	was_short = ret >= 0 && ret < this_len;
349 	dout("striped_read %llu~%u (read %u) got %d%s%s\n", pos, left, read,
350 	     ret, hit_stripe ? " HITSTRIPE" : "", was_short ? " SHORT" : "");
351 
352 	if (ret > 0) {
353 		int didpages = (page_align + ret) >> PAGE_CACHE_SHIFT;
354 
355 		if (read < pos - off) {
356 			dout(" zero gap %llu to %llu\n", off + read, pos);
357 			ceph_zero_page_vector_range(page_align + read,
358 						    pos - off - read, pages);
359 		}
360 		pos += ret;
361 		read = pos - off;
362 		left -= ret;
363 		page_pos += didpages;
364 		pages_left -= didpages;
365 
366 		/* hit stripe? */
367 		if (left && hit_stripe)
368 			goto more;
369 	}
370 
371 	if (was_short) {
372 		/* did we bounce off eof? */
373 		if (pos + left > inode->i_size)
374 			*checkeof = 1;
375 
376 		/* zero trailing bytes (inside i_size) */
377 		if (left > 0 && pos < inode->i_size) {
378 			if (pos + left > inode->i_size)
379 				left = inode->i_size - pos;
380 
381 			dout("zero tail %d\n", left);
382 			ceph_zero_page_vector_range(page_align + read, left,
383 						    pages);
384 			read += left;
385 		}
386 	}
387 
388 	if (ret >= 0)
389 		ret = read;
390 	dout("striped_read returns %d\n", ret);
391 	return ret;
392 }
393 
394 /*
395  * Completely synchronous read and write methods.  Direct from __user
396  * buffer to osd, or directly to user pages (if O_DIRECT).
397  *
398  * If the read spans object boundary, just do multiple reads.
399  */
400 static ssize_t ceph_sync_read(struct file *file, char __user *data,
401 			      unsigned len, loff_t *poff, int *checkeof)
402 {
403 	struct inode *inode = file_inode(file);
404 	struct page **pages;
405 	u64 off = *poff;
406 	int num_pages, ret;
407 
408 	dout("sync_read on file %p %llu~%u %s\n", file, off, len,
409 	     (file->f_flags & O_DIRECT) ? "O_DIRECT" : "");
410 
411 	if (file->f_flags & O_DIRECT) {
412 		num_pages = calc_pages_for((unsigned long)data, len);
413 		pages = ceph_get_direct_page_vector(data, num_pages, true);
414 	} else {
415 		num_pages = calc_pages_for(off, len);
416 		pages = ceph_alloc_page_vector(num_pages, GFP_NOFS);
417 	}
418 	if (IS_ERR(pages))
419 		return PTR_ERR(pages);
420 
421 	/*
422 	 * flush any page cache pages in this range.  this
423 	 * will make concurrent normal and sync io slow,
424 	 * but it will at least behave sensibly when they are
425 	 * in sequence.
426 	 */
427 	ret = filemap_write_and_wait(inode->i_mapping);
428 	if (ret < 0)
429 		goto done;
430 
431 	ret = striped_read(inode, off, len, pages, num_pages, checkeof,
432 			   file->f_flags & O_DIRECT,
433 			   (unsigned long)data & ~PAGE_MASK);
434 
435 	if (ret >= 0 && (file->f_flags & O_DIRECT) == 0)
436 		ret = ceph_copy_page_vector_to_user(pages, data, off, ret);
437 	if (ret >= 0)
438 		*poff = off + ret;
439 
440 done:
441 	if (file->f_flags & O_DIRECT)
442 		ceph_put_page_vector(pages, num_pages, true);
443 	else
444 		ceph_release_page_vector(pages, num_pages);
445 	dout("sync_read result %d\n", ret);
446 	return ret;
447 }
448 
449 /*
450  * Write commit request unsafe callback, called to tell us when a
451  * request is unsafe (that is, in flight--has been handed to the
452  * messenger to send to its target osd).  It is called again when
453  * we've received a response message indicating the request is
454  * "safe" (its CEPH_OSD_FLAG_ONDISK flag is set), or when a request
455  * is completed early (and unsuccessfully) due to a timeout or
456  * interrupt.
457  *
458  * This is used if we requested both an ACK and ONDISK commit reply
459  * from the OSD.
460  */
461 static void ceph_sync_write_unsafe(struct ceph_osd_request *req, bool unsafe)
462 {
463 	struct ceph_inode_info *ci = ceph_inode(req->r_inode);
464 
465 	dout("%s %p tid %llu %ssafe\n", __func__, req, req->r_tid,
466 		unsafe ? "un" : "");
467 	if (unsafe) {
468 		ceph_get_cap_refs(ci, CEPH_CAP_FILE_WR);
469 		spin_lock(&ci->i_unsafe_lock);
470 		list_add_tail(&req->r_unsafe_item,
471 			      &ci->i_unsafe_writes);
472 		spin_unlock(&ci->i_unsafe_lock);
473 	} else {
474 		spin_lock(&ci->i_unsafe_lock);
475 		list_del_init(&req->r_unsafe_item);
476 		spin_unlock(&ci->i_unsafe_lock);
477 		ceph_put_cap_refs(ci, CEPH_CAP_FILE_WR);
478 	}
479 }
480 
481 /*
482  * Synchronous write, straight from __user pointer or user pages (if
483  * O_DIRECT).
484  *
485  * If write spans object boundary, just do multiple writes.  (For a
486  * correct atomic write, we should e.g. take write locks on all
487  * objects, rollback on failure, etc.)
488  */
489 static ssize_t ceph_sync_write(struct file *file, const char __user *data,
490 			       size_t left, loff_t pos, loff_t *ppos)
491 {
492 	struct inode *inode = file_inode(file);
493 	struct ceph_inode_info *ci = ceph_inode(inode);
494 	struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
495 	struct ceph_snap_context *snapc;
496 	struct ceph_vino vino;
497 	struct ceph_osd_request *req;
498 	int num_ops = 1;
499 	struct page **pages;
500 	int num_pages;
501 	u64 len;
502 	int written = 0;
503 	int flags;
504 	int check_caps = 0;
505 	int page_align, io_align;
506 	unsigned long buf_align;
507 	int ret;
508 	struct timespec mtime = CURRENT_TIME;
509 	bool own_pages = false;
510 
511 	if (ceph_snap(file_inode(file)) != CEPH_NOSNAP)
512 		return -EROFS;
513 
514 	dout("sync_write on file %p %lld~%u %s\n", file, pos,
515 	     (unsigned)left, (file->f_flags & O_DIRECT) ? "O_DIRECT" : "");
516 
517 	ret = filemap_write_and_wait_range(inode->i_mapping, pos, pos + left);
518 	if (ret < 0)
519 		return ret;
520 
521 	ret = invalidate_inode_pages2_range(inode->i_mapping,
522 					    pos >> PAGE_CACHE_SHIFT,
523 					    (pos + left) >> PAGE_CACHE_SHIFT);
524 	if (ret < 0)
525 		dout("invalidate_inode_pages2_range returned %d\n", ret);
526 
527 	flags = CEPH_OSD_FLAG_ORDERSNAP |
528 		CEPH_OSD_FLAG_ONDISK |
529 		CEPH_OSD_FLAG_WRITE;
530 	if ((file->f_flags & (O_SYNC|O_DIRECT)) == 0)
531 		flags |= CEPH_OSD_FLAG_ACK;
532 	else
533 		num_ops++;	/* Also include a 'startsync' command. */
534 
535 	/*
536 	 * we may need to do multiple writes here if we span an object
537 	 * boundary.  this isn't atomic, unfortunately.  :(
538 	 */
539 more:
540 	io_align = pos & ~PAGE_MASK;
541 	buf_align = (unsigned long)data & ~PAGE_MASK;
542 	len = left;
543 
544 	snapc = ci->i_snap_realm->cached_context;
545 	vino = ceph_vino(inode);
546 	req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
547 				    vino, pos, &len, num_ops,
548 				    CEPH_OSD_OP_WRITE, flags, snapc,
549 				    ci->i_truncate_seq, ci->i_truncate_size,
550 				    false);
551 	if (IS_ERR(req))
552 		return PTR_ERR(req);
553 
554 	/* write from beginning of first page, regardless of io alignment */
555 	page_align = file->f_flags & O_DIRECT ? buf_align : io_align;
556 	num_pages = calc_pages_for(page_align, len);
557 	if (file->f_flags & O_DIRECT) {
558 		pages = ceph_get_direct_page_vector(data, num_pages, false);
559 		if (IS_ERR(pages)) {
560 			ret = PTR_ERR(pages);
561 			goto out;
562 		}
563 
564 		/*
565 		 * throw out any page cache pages in this range. this
566 		 * may block.
567 		 */
568 		truncate_inode_pages_range(inode->i_mapping, pos,
569 					   (pos+len) | (PAGE_CACHE_SIZE-1));
570 	} else {
571 		pages = ceph_alloc_page_vector(num_pages, GFP_NOFS);
572 		if (IS_ERR(pages)) {
573 			ret = PTR_ERR(pages);
574 			goto out;
575 		}
576 		ret = ceph_copy_user_to_page_vector(pages, data, pos, len);
577 		if (ret < 0) {
578 			ceph_release_page_vector(pages, num_pages);
579 			goto out;
580 		}
581 
582 		if ((file->f_flags & O_SYNC) == 0) {
583 			/* get a second commit callback */
584 			req->r_unsafe_callback = ceph_sync_write_unsafe;
585 			req->r_inode = inode;
586 			own_pages = true;
587 		}
588 	}
589 	osd_req_op_extent_osd_data_pages(req, 0, pages, len, page_align,
590 					false, own_pages);
591 
592 	/* BUG_ON(vino.snap != CEPH_NOSNAP); */
593 	ceph_osdc_build_request(req, pos, snapc, vino.snap, &mtime);
594 
595 	ret = ceph_osdc_start_request(&fsc->client->osdc, req, false);
596 	if (!ret)
597 		ret = ceph_osdc_wait_request(&fsc->client->osdc, req);
598 
599 	if (file->f_flags & O_DIRECT)
600 		ceph_put_page_vector(pages, num_pages, false);
601 	else if (file->f_flags & O_SYNC)
602 		ceph_release_page_vector(pages, num_pages);
603 
604 out:
605 	ceph_osdc_put_request(req);
606 	if (ret == 0) {
607 		pos += len;
608 		written += len;
609 		left -= len;
610 		data += len;
611 		if (left)
612 			goto more;
613 
614 		ret = written;
615 		*ppos = pos;
616 		if (pos > i_size_read(inode))
617 			check_caps = ceph_inode_set_size(inode, pos);
618 		if (check_caps)
619 			ceph_check_caps(ceph_inode(inode), CHECK_CAPS_AUTHONLY,
620 					NULL);
621 	}
622 	return ret;
623 }
624 
625 /*
626  * Wrap generic_file_aio_read with checks for cap bits on the inode.
627  * Atomically grab references, so that those bits are not released
628  * back to the MDS mid-read.
629  *
630  * Hmm, the sync read case isn't actually async... should it be?
631  */
632 static ssize_t ceph_aio_read(struct kiocb *iocb, const struct iovec *iov,
633 			     unsigned long nr_segs, loff_t pos)
634 {
635 	struct file *filp = iocb->ki_filp;
636 	struct ceph_file_info *fi = filp->private_data;
637 	loff_t *ppos = &iocb->ki_pos;
638 	size_t len = iov->iov_len;
639 	struct inode *inode = file_inode(filp);
640 	struct ceph_inode_info *ci = ceph_inode(inode);
641 	void __user *base = iov->iov_base;
642 	ssize_t ret;
643 	int want, got = 0;
644 	int checkeof = 0, read = 0;
645 
646 	dout("aio_read %p %llx.%llx %llu~%u trying to get caps on %p\n",
647 	     inode, ceph_vinop(inode), pos, (unsigned)len, inode);
648 again:
649 	if (fi->fmode & CEPH_FILE_MODE_LAZY)
650 		want = CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO;
651 	else
652 		want = CEPH_CAP_FILE_CACHE;
653 	ret = ceph_get_caps(ci, CEPH_CAP_FILE_RD, want, &got, -1);
654 	if (ret < 0)
655 		goto out;
656 	dout("aio_read %p %llx.%llx %llu~%u got cap refs on %s\n",
657 	     inode, ceph_vinop(inode), pos, (unsigned)len,
658 	     ceph_cap_string(got));
659 
660 	if ((got & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) == 0 ||
661 	    (iocb->ki_filp->f_flags & O_DIRECT) ||
662 	    (inode->i_sb->s_flags & MS_SYNCHRONOUS) ||
663 	    (fi->flags & CEPH_F_SYNC))
664 		/* hmm, this isn't really async... */
665 		ret = ceph_sync_read(filp, base, len, ppos, &checkeof);
666 	else
667 		ret = generic_file_aio_read(iocb, iov, nr_segs, pos);
668 
669 out:
670 	dout("aio_read %p %llx.%llx dropping cap refs on %s = %d\n",
671 	     inode, ceph_vinop(inode), ceph_cap_string(got), (int)ret);
672 	ceph_put_cap_refs(ci, got);
673 
674 	if (checkeof && ret >= 0) {
675 		int statret = ceph_do_getattr(inode, CEPH_STAT_CAP_SIZE);
676 
677 		/* hit EOF or hole? */
678 		if (statret == 0 && *ppos < inode->i_size) {
679 			dout("aio_read sync_read hit hole, ppos %lld < size %lld, reading more\n", *ppos, inode->i_size);
680 			read += ret;
681 			base += ret;
682 			len -= ret;
683 			checkeof = 0;
684 			goto again;
685 		}
686 	}
687 	if (ret >= 0)
688 		ret += read;
689 
690 	return ret;
691 }
692 
693 /*
694  * Take cap references to avoid releasing caps to MDS mid-write.
695  *
696  * If we are synchronous, and write with an old snap context, the OSD
697  * may return EOLDSNAPC.  In that case, retry the write.. _after_
698  * dropping our cap refs and allowing the pending snap to logically
699  * complete _before_ this write occurs.
700  *
701  * If we are near ENOSPC, write synchronously.
702  */
703 static ssize_t ceph_aio_write(struct kiocb *iocb, const struct iovec *iov,
704 		       unsigned long nr_segs, loff_t pos)
705 {
706 	struct file *file = iocb->ki_filp;
707 	struct ceph_file_info *fi = file->private_data;
708 	struct inode *inode = file_inode(file);
709 	struct ceph_inode_info *ci = ceph_inode(inode);
710 	struct ceph_osd_client *osdc =
711 		&ceph_sb_to_client(inode->i_sb)->client->osdc;
712 	ssize_t count, written = 0;
713 	int err, want, got;
714 	bool hold_mutex;
715 
716 	if (ceph_snap(inode) != CEPH_NOSNAP)
717 		return -EROFS;
718 
719 	mutex_lock(&inode->i_mutex);
720 	hold_mutex = true;
721 
722 	err = generic_segment_checks(iov, &nr_segs, &count, VERIFY_READ);
723 	if (err)
724 		goto out;
725 
726 	/* We can write back this queue in page reclaim */
727 	current->backing_dev_info = file->f_mapping->backing_dev_info;
728 
729 	err = generic_write_checks(file, &pos, &count, S_ISBLK(inode->i_mode));
730 	if (err)
731 		goto out;
732 
733 	if (count == 0)
734 		goto out;
735 
736 	err = file_remove_suid(file);
737 	if (err)
738 		goto out;
739 
740 	err = file_update_time(file);
741 	if (err)
742 		goto out;
743 
744 retry_snap:
745 	if (ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_FULL)) {
746 		err = -ENOSPC;
747 		goto out;
748 	}
749 
750 	dout("aio_write %p %llx.%llx %llu~%zd getting caps. i_size %llu\n",
751 	     inode, ceph_vinop(inode), pos, count, inode->i_size);
752 	if (fi->fmode & CEPH_FILE_MODE_LAZY)
753 		want = CEPH_CAP_FILE_BUFFER | CEPH_CAP_FILE_LAZYIO;
754 	else
755 		want = CEPH_CAP_FILE_BUFFER;
756 	got = 0;
757 	err = ceph_get_caps(ci, CEPH_CAP_FILE_WR, want, &got, pos + count);
758 	if (err < 0)
759 		goto out;
760 
761 	dout("aio_write %p %llx.%llx %llu~%zd got cap refs on %s\n",
762 	     inode, ceph_vinop(inode), pos, count, ceph_cap_string(got));
763 
764 	if ((got & (CEPH_CAP_FILE_BUFFER|CEPH_CAP_FILE_LAZYIO)) == 0 ||
765 	    (iocb->ki_filp->f_flags & O_DIRECT) ||
766 	    (inode->i_sb->s_flags & MS_SYNCHRONOUS) ||
767 	    (fi->flags & CEPH_F_SYNC)) {
768 		mutex_unlock(&inode->i_mutex);
769 		written = ceph_sync_write(file, iov->iov_base, count,
770 					  pos, &iocb->ki_pos);
771 	} else {
772 		written = generic_file_buffered_write(iocb, iov, nr_segs,
773 						      pos, &iocb->ki_pos,
774 						      count, 0);
775 		mutex_unlock(&inode->i_mutex);
776 	}
777 	hold_mutex = false;
778 
779 	if (written >= 0) {
780 		int dirty;
781 		spin_lock(&ci->i_ceph_lock);
782 		dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR);
783 		spin_unlock(&ci->i_ceph_lock);
784 		if (dirty)
785 			__mark_inode_dirty(inode, dirty);
786 	}
787 
788 	dout("aio_write %p %llx.%llx %llu~%u  dropping cap refs on %s\n",
789 	     inode, ceph_vinop(inode), pos, (unsigned)iov->iov_len,
790 	     ceph_cap_string(got));
791 	ceph_put_cap_refs(ci, got);
792 
793 	if (written >= 0 &&
794 	    ((file->f_flags & O_SYNC) || IS_SYNC(file->f_mapping->host) ||
795 	     ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_NEARFULL))) {
796 		err = vfs_fsync_range(file, pos, pos + written - 1, 1);
797 		if (err < 0)
798 			written = err;
799 	}
800 
801 	if (written == -EOLDSNAPC) {
802 		dout("aio_write %p %llx.%llx %llu~%u got EOLDSNAPC, retrying\n",
803 		     inode, ceph_vinop(inode), pos, (unsigned)iov->iov_len);
804 		mutex_lock(&inode->i_mutex);
805 		hold_mutex = true;
806 		goto retry_snap;
807 	}
808 out:
809 	if (hold_mutex)
810 		mutex_unlock(&inode->i_mutex);
811 	current->backing_dev_info = NULL;
812 
813 	return written ? written : err;
814 }
815 
816 /*
817  * llseek.  be sure to verify file size on SEEK_END.
818  */
819 static loff_t ceph_llseek(struct file *file, loff_t offset, int whence)
820 {
821 	struct inode *inode = file->f_mapping->host;
822 	int ret;
823 
824 	mutex_lock(&inode->i_mutex);
825 	__ceph_do_pending_vmtruncate(inode);
826 
827 	if (whence == SEEK_END || whence == SEEK_DATA || whence == SEEK_HOLE) {
828 		ret = ceph_do_getattr(inode, CEPH_STAT_CAP_SIZE);
829 		if (ret < 0) {
830 			offset = ret;
831 			goto out;
832 		}
833 	}
834 
835 	switch (whence) {
836 	case SEEK_END:
837 		offset += inode->i_size;
838 		break;
839 	case SEEK_CUR:
840 		/*
841 		 * Here we special-case the lseek(fd, 0, SEEK_CUR)
842 		 * position-querying operation.  Avoid rewriting the "same"
843 		 * f_pos value back to the file because a concurrent read(),
844 		 * write() or lseek() might have altered it
845 		 */
846 		if (offset == 0) {
847 			offset = file->f_pos;
848 			goto out;
849 		}
850 		offset += file->f_pos;
851 		break;
852 	case SEEK_DATA:
853 		if (offset >= inode->i_size) {
854 			ret = -ENXIO;
855 			goto out;
856 		}
857 		break;
858 	case SEEK_HOLE:
859 		if (offset >= inode->i_size) {
860 			ret = -ENXIO;
861 			goto out;
862 		}
863 		offset = inode->i_size;
864 		break;
865 	}
866 
867 	offset = vfs_setpos(file, offset, inode->i_sb->s_maxbytes);
868 
869 out:
870 	mutex_unlock(&inode->i_mutex);
871 	return offset;
872 }
873 
874 const struct file_operations ceph_file_fops = {
875 	.open = ceph_open,
876 	.release = ceph_release,
877 	.llseek = ceph_llseek,
878 	.read = do_sync_read,
879 	.write = do_sync_write,
880 	.aio_read = ceph_aio_read,
881 	.aio_write = ceph_aio_write,
882 	.mmap = ceph_mmap,
883 	.fsync = ceph_fsync,
884 	.lock = ceph_lock,
885 	.flock = ceph_flock,
886 	.splice_read = generic_file_splice_read,
887 	.splice_write = generic_file_splice_write,
888 	.unlocked_ioctl = ceph_ioctl,
889 	.compat_ioctl	= ceph_ioctl,
890 };
891 
892