xref: /openbmc/linux/fs/ceph/file.c (revision 4d09a938861ef16248334d57ae250e109b0004ae)
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/namei.h>
8 #include <linux/writeback.h>
9 
10 #include "super.h"
11 #include "mds_client.h"
12 
13 /*
14  * Ceph file operations
15  *
16  * Implement basic open/close functionality, and implement
17  * read/write.
18  *
19  * We implement three modes of file I/O:
20  *  - buffered uses the generic_file_aio_{read,write} helpers
21  *
22  *  - synchronous is used when there is multi-client read/write
23  *    sharing, avoids the page cache, and synchronously waits for an
24  *    ack from the OSD.
25  *
26  *  - direct io takes the variant of the sync path that references
27  *    user pages directly.
28  *
29  * fsync() flushes and waits on dirty pages, but just queues metadata
30  * for writeback: since the MDS can recover size and mtime there is no
31  * need to wait for MDS acknowledgement.
32  */
33 
34 
35 /*
36  * Prepare an open request.  Preallocate ceph_cap to avoid an
37  * inopportune ENOMEM later.
38  */
39 static struct ceph_mds_request *
40 prepare_open_request(struct super_block *sb, int flags, int create_mode)
41 {
42 	struct ceph_fs_client *fsc = ceph_sb_to_client(sb);
43 	struct ceph_mds_client *mdsc = fsc->mdsc;
44 	struct ceph_mds_request *req;
45 	int want_auth = USE_ANY_MDS;
46 	int op = (flags & O_CREAT) ? CEPH_MDS_OP_CREATE : CEPH_MDS_OP_OPEN;
47 
48 	if (flags & (O_WRONLY|O_RDWR|O_CREAT|O_TRUNC))
49 		want_auth = USE_AUTH_MDS;
50 
51 	req = ceph_mdsc_create_request(mdsc, op, want_auth);
52 	if (IS_ERR(req))
53 		goto out;
54 	req->r_fmode = ceph_flags_to_mode(flags);
55 	req->r_args.open.flags = cpu_to_le32(flags);
56 	req->r_args.open.mode = cpu_to_le32(create_mode);
57 	req->r_args.open.preferred = cpu_to_le32(-1);
58 out:
59 	return req;
60 }
61 
62 /*
63  * initialize private struct file data.
64  * if we fail, clean up by dropping fmode reference on the ceph_inode
65  */
66 static int ceph_init_file(struct inode *inode, struct file *file, int fmode)
67 {
68 	struct ceph_file_info *cf;
69 	int ret = 0;
70 
71 	switch (inode->i_mode & S_IFMT) {
72 	case S_IFREG:
73 	case S_IFDIR:
74 		dout("init_file %p %p 0%o (regular)\n", inode, file,
75 		     inode->i_mode);
76 		cf = kmem_cache_alloc(ceph_file_cachep, GFP_NOFS | __GFP_ZERO);
77 		if (cf == NULL) {
78 			ceph_put_fmode(ceph_inode(inode), fmode); /* clean up */
79 			return -ENOMEM;
80 		}
81 		cf->fmode = fmode;
82 		cf->next_offset = 2;
83 		file->private_data = cf;
84 		BUG_ON(inode->i_fop->release != ceph_release);
85 		break;
86 
87 	case S_IFLNK:
88 		dout("init_file %p %p 0%o (symlink)\n", inode, file,
89 		     inode->i_mode);
90 		ceph_put_fmode(ceph_inode(inode), fmode); /* clean up */
91 		break;
92 
93 	default:
94 		dout("init_file %p %p 0%o (special)\n", inode, file,
95 		     inode->i_mode);
96 		/*
97 		 * we need to drop the open ref now, since we don't
98 		 * have .release set to ceph_release.
99 		 */
100 		ceph_put_fmode(ceph_inode(inode), fmode); /* clean up */
101 		BUG_ON(inode->i_fop->release == ceph_release);
102 
103 		/* call the proper open fop */
104 		ret = inode->i_fop->open(inode, file);
105 	}
106 	return ret;
107 }
108 
109 /*
110  * If the filp already has private_data, that means the file was
111  * already opened by intent during lookup, and we do nothing.
112  *
113  * If we already have the requisite capabilities, we can satisfy
114  * the open request locally (no need to request new caps from the
115  * MDS).  We do, however, need to inform the MDS (asynchronously)
116  * if our wanted caps set expands.
117  */
118 int ceph_open(struct inode *inode, struct file *file)
119 {
120 	struct ceph_inode_info *ci = ceph_inode(inode);
121 	struct ceph_fs_client *fsc = ceph_sb_to_client(inode->i_sb);
122 	struct ceph_mds_client *mdsc = fsc->mdsc;
123 	struct ceph_mds_request *req;
124 	struct ceph_file_info *cf = file->private_data;
125 	struct inode *parent_inode = file->f_dentry->d_parent->d_inode;
126 	int err;
127 	int flags, fmode, wanted;
128 
129 	if (cf) {
130 		dout("open file %p is already opened\n", file);
131 		return 0;
132 	}
133 
134 	/* filter out O_CREAT|O_EXCL; vfs did that already.  yuck. */
135 	flags = file->f_flags & ~(O_CREAT|O_EXCL);
136 	if (S_ISDIR(inode->i_mode))
137 		flags = O_DIRECTORY;  /* mds likes to know */
138 
139 	dout("open inode %p ino %llx.%llx file %p flags %d (%d)\n", inode,
140 	     ceph_vinop(inode), file, flags, file->f_flags);
141 	fmode = ceph_flags_to_mode(flags);
142 	wanted = ceph_caps_for_mode(fmode);
143 
144 	/* snapped files are read-only */
145 	if (ceph_snap(inode) != CEPH_NOSNAP && (file->f_mode & FMODE_WRITE))
146 		return -EROFS;
147 
148 	/* trivially open snapdir */
149 	if (ceph_snap(inode) == CEPH_SNAPDIR) {
150 		spin_lock(&inode->i_lock);
151 		__ceph_get_fmode(ci, fmode);
152 		spin_unlock(&inode->i_lock);
153 		return ceph_init_file(inode, file, fmode);
154 	}
155 
156 	/*
157 	 * No need to block if we have caps on the auth MDS (for
158 	 * write) or any MDS (for read).  Update wanted set
159 	 * asynchronously.
160 	 */
161 	spin_lock(&inode->i_lock);
162 	if (__ceph_is_any_real_caps(ci) &&
163 	    (((fmode & CEPH_FILE_MODE_WR) == 0) || ci->i_auth_cap)) {
164 		int mds_wanted = __ceph_caps_mds_wanted(ci);
165 		int issued = __ceph_caps_issued(ci, NULL);
166 
167 		dout("open %p fmode %d want %s issued %s using existing\n",
168 		     inode, fmode, ceph_cap_string(wanted),
169 		     ceph_cap_string(issued));
170 		__ceph_get_fmode(ci, fmode);
171 		spin_unlock(&inode->i_lock);
172 
173 		/* adjust wanted? */
174 		if ((issued & wanted) != wanted &&
175 		    (mds_wanted & wanted) != wanted &&
176 		    ceph_snap(inode) != CEPH_SNAPDIR)
177 			ceph_check_caps(ci, 0, NULL);
178 
179 		return ceph_init_file(inode, file, fmode);
180 	} else if (ceph_snap(inode) != CEPH_NOSNAP &&
181 		   (ci->i_snap_caps & wanted) == wanted) {
182 		__ceph_get_fmode(ci, fmode);
183 		spin_unlock(&inode->i_lock);
184 		return ceph_init_file(inode, file, fmode);
185 	}
186 	spin_unlock(&inode->i_lock);
187 
188 	dout("open fmode %d wants %s\n", fmode, ceph_cap_string(wanted));
189 	req = prepare_open_request(inode->i_sb, flags, 0);
190 	if (IS_ERR(req)) {
191 		err = PTR_ERR(req);
192 		goto out;
193 	}
194 	req->r_inode = inode;
195 	ihold(inode);
196 	req->r_num_caps = 1;
197 	err = ceph_mdsc_do_request(mdsc, parent_inode, req);
198 	if (!err)
199 		err = ceph_init_file(inode, file, req->r_fmode);
200 	ceph_mdsc_put_request(req);
201 	dout("open result=%d on %llx.%llx\n", err, ceph_vinop(inode));
202 out:
203 	return err;
204 }
205 
206 
207 /*
208  * Do a lookup + open with a single request.
209  *
210  * If this succeeds, but some subsequent check in the vfs
211  * may_open() fails, the struct *file gets cleaned up (i.e.
212  * ceph_release gets called).  So fear not!
213  */
214 /*
215  * flags
216  *  path_lookup_open   -> LOOKUP_OPEN
217  *  path_lookup_create -> LOOKUP_OPEN|LOOKUP_CREATE
218  */
219 struct dentry *ceph_lookup_open(struct inode *dir, struct dentry *dentry,
220 				struct nameidata *nd, int mode,
221 				int locked_dir)
222 {
223 	struct ceph_fs_client *fsc = ceph_sb_to_client(dir->i_sb);
224 	struct ceph_mds_client *mdsc = fsc->mdsc;
225 	struct file *file = nd->intent.open.file;
226 	struct inode *parent_inode = get_dentry_parent_inode(file->f_dentry);
227 	struct ceph_mds_request *req;
228 	int err;
229 	int flags = nd->intent.open.flags - 1;  /* silly vfs! */
230 
231 	dout("ceph_lookup_open dentry %p '%.*s' flags %d mode 0%o\n",
232 	     dentry, dentry->d_name.len, dentry->d_name.name, flags, mode);
233 
234 	/* do the open */
235 	req = prepare_open_request(dir->i_sb, flags, mode);
236 	if (IS_ERR(req))
237 		return ERR_CAST(req);
238 	req->r_dentry = dget(dentry);
239 	req->r_num_caps = 2;
240 	if (flags & O_CREAT) {
241 		req->r_dentry_drop = CEPH_CAP_FILE_SHARED;
242 		req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
243 	}
244 	req->r_locked_dir = dir;           /* caller holds dir->i_mutex */
245 	err = ceph_mdsc_do_request(mdsc, parent_inode, req);
246 	dentry = ceph_finish_lookup(req, dentry, err);
247 	if (!err && (flags & O_CREAT) && !req->r_reply_info.head->is_dentry)
248 		err = ceph_handle_notrace_create(dir, dentry);
249 	if (!err)
250 		err = ceph_init_file(req->r_dentry->d_inode, file,
251 				     req->r_fmode);
252 	ceph_mdsc_put_request(req);
253 	dout("ceph_lookup_open result=%p\n", dentry);
254 	return dentry;
255 }
256 
257 int ceph_release(struct inode *inode, struct file *file)
258 {
259 	struct ceph_inode_info *ci = ceph_inode(inode);
260 	struct ceph_file_info *cf = file->private_data;
261 
262 	dout("release inode %p file %p\n", inode, file);
263 	ceph_put_fmode(ci, cf->fmode);
264 	if (cf->last_readdir)
265 		ceph_mdsc_put_request(cf->last_readdir);
266 	kfree(cf->last_name);
267 	kfree(cf->dir_info);
268 	dput(cf->dentry);
269 	kmem_cache_free(ceph_file_cachep, cf);
270 
271 	/* wake up anyone waiting for caps on this inode */
272 	wake_up_all(&ci->i_cap_wq);
273 	return 0;
274 }
275 
276 /*
277  * Read a range of bytes striped over one or more objects.  Iterate over
278  * objects we stripe over.  (That's not atomic, but good enough for now.)
279  *
280  * If we get a short result from the OSD, check against i_size; we need to
281  * only return a short read to the caller if we hit EOF.
282  */
283 static int striped_read(struct inode *inode,
284 			u64 off, u64 len,
285 			struct page **pages, int num_pages,
286 			int *checkeof, bool o_direct,
287 			unsigned long buf_align)
288 {
289 	struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
290 	struct ceph_inode_info *ci = ceph_inode(inode);
291 	u64 pos, this_len;
292 	int io_align, page_align;
293 	int left, pages_left;
294 	int read;
295 	struct page **page_pos;
296 	int ret;
297 	bool hit_stripe, was_short;
298 
299 	/*
300 	 * we may need to do multiple reads.  not atomic, unfortunately.
301 	 */
302 	pos = off;
303 	left = len;
304 	page_pos = pages;
305 	pages_left = num_pages;
306 	read = 0;
307 	io_align = off & ~PAGE_MASK;
308 
309 more:
310 	if (o_direct)
311 		page_align = (pos - io_align + buf_align) & ~PAGE_MASK;
312 	else
313 		page_align = pos & ~PAGE_MASK;
314 	this_len = left;
315 	ret = ceph_osdc_readpages(&fsc->client->osdc, ceph_vino(inode),
316 				  &ci->i_layout, pos, &this_len,
317 				  ci->i_truncate_seq,
318 				  ci->i_truncate_size,
319 				  page_pos, pages_left, page_align);
320 	if (ret == -ENOENT)
321 		ret = 0;
322 	hit_stripe = this_len < left;
323 	was_short = ret >= 0 && ret < this_len;
324 	dout("striped_read %llu~%u (read %u) got %d%s%s\n", pos, left, read,
325 	     ret, hit_stripe ? " HITSTRIPE" : "", was_short ? " SHORT" : "");
326 
327 	if (ret > 0) {
328 		int didpages = (page_align + ret) >> PAGE_CACHE_SHIFT;
329 
330 		if (read < pos - off) {
331 			dout(" zero gap %llu to %llu\n", off + read, pos);
332 			ceph_zero_page_vector_range(page_align + read,
333 						    pos - off - read, pages);
334 		}
335 		pos += ret;
336 		read = pos - off;
337 		left -= ret;
338 		page_pos += didpages;
339 		pages_left -= didpages;
340 
341 		/* hit stripe? */
342 		if (left && hit_stripe)
343 			goto more;
344 	}
345 
346 	if (was_short) {
347 		/* did we bounce off eof? */
348 		if (pos + left > inode->i_size)
349 			*checkeof = 1;
350 
351 		/* zero trailing bytes (inside i_size) */
352 		if (left > 0 && pos < inode->i_size) {
353 			if (pos + left > inode->i_size)
354 				left = inode->i_size - pos;
355 
356 			dout("zero tail %d\n", left);
357 			ceph_zero_page_vector_range(page_align + read, left,
358 						    pages);
359 			read += left;
360 		}
361 	}
362 
363 	if (ret >= 0)
364 		ret = read;
365 	dout("striped_read returns %d\n", ret);
366 	return ret;
367 }
368 
369 /*
370  * Completely synchronous read and write methods.  Direct from __user
371  * buffer to osd, or directly to user pages (if O_DIRECT).
372  *
373  * If the read spans object boundary, just do multiple reads.
374  */
375 static ssize_t ceph_sync_read(struct file *file, char __user *data,
376 			      unsigned len, loff_t *poff, int *checkeof)
377 {
378 	struct inode *inode = file->f_dentry->d_inode;
379 	struct page **pages;
380 	u64 off = *poff;
381 	int num_pages, ret;
382 
383 	dout("sync_read on file %p %llu~%u %s\n", file, off, len,
384 	     (file->f_flags & O_DIRECT) ? "O_DIRECT" : "");
385 
386 	if (file->f_flags & O_DIRECT) {
387 		num_pages = calc_pages_for((unsigned long)data, len);
388 		pages = ceph_get_direct_page_vector(data, num_pages, true);
389 	} else {
390 		num_pages = calc_pages_for(off, len);
391 		pages = ceph_alloc_page_vector(num_pages, GFP_NOFS);
392 	}
393 	if (IS_ERR(pages))
394 		return PTR_ERR(pages);
395 
396 	/*
397 	 * flush any page cache pages in this range.  this
398 	 * will make concurrent normal and sync io slow,
399 	 * but it will at least behave sensibly when they are
400 	 * in sequence.
401 	 */
402 	ret = filemap_write_and_wait(inode->i_mapping);
403 	if (ret < 0)
404 		goto done;
405 
406 	ret = striped_read(inode, off, len, pages, num_pages, checkeof,
407 			   file->f_flags & O_DIRECT,
408 			   (unsigned long)data & ~PAGE_MASK);
409 
410 	if (ret >= 0 && (file->f_flags & O_DIRECT) == 0)
411 		ret = ceph_copy_page_vector_to_user(pages, data, off, ret);
412 	if (ret >= 0)
413 		*poff = off + ret;
414 
415 done:
416 	if (file->f_flags & O_DIRECT)
417 		ceph_put_page_vector(pages, num_pages, true);
418 	else
419 		ceph_release_page_vector(pages, num_pages);
420 	dout("sync_read result %d\n", ret);
421 	return ret;
422 }
423 
424 /*
425  * Write commit callback, called if we requested both an ACK and
426  * ONDISK commit reply from the OSD.
427  */
428 static void sync_write_commit(struct ceph_osd_request *req,
429 			      struct ceph_msg *msg)
430 {
431 	struct ceph_inode_info *ci = ceph_inode(req->r_inode);
432 
433 	dout("sync_write_commit %p tid %llu\n", req, req->r_tid);
434 	spin_lock(&ci->i_unsafe_lock);
435 	list_del_init(&req->r_unsafe_item);
436 	spin_unlock(&ci->i_unsafe_lock);
437 	ceph_put_cap_refs(ci, CEPH_CAP_FILE_WR);
438 }
439 
440 /*
441  * Synchronous write, straight from __user pointer or user pages (if
442  * O_DIRECT).
443  *
444  * If write spans object boundary, just do multiple writes.  (For a
445  * correct atomic write, we should e.g. take write locks on all
446  * objects, rollback on failure, etc.)
447  */
448 static ssize_t ceph_sync_write(struct file *file, const char __user *data,
449 			       size_t left, loff_t *offset)
450 {
451 	struct inode *inode = file->f_dentry->d_inode;
452 	struct ceph_inode_info *ci = ceph_inode(inode);
453 	struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
454 	struct ceph_osd_request *req;
455 	struct page **pages;
456 	int num_pages;
457 	long long unsigned pos;
458 	u64 len;
459 	int written = 0;
460 	int flags;
461 	int do_sync = 0;
462 	int check_caps = 0;
463 	int page_align, io_align;
464 	unsigned long buf_align;
465 	int ret;
466 	struct timespec mtime = CURRENT_TIME;
467 
468 	if (ceph_snap(file->f_dentry->d_inode) != CEPH_NOSNAP)
469 		return -EROFS;
470 
471 	dout("sync_write on file %p %lld~%u %s\n", file, *offset,
472 	     (unsigned)left, (file->f_flags & O_DIRECT) ? "O_DIRECT" : "");
473 
474 	if (file->f_flags & O_APPEND)
475 		pos = i_size_read(inode);
476 	else
477 		pos = *offset;
478 
479 	ret = filemap_write_and_wait_range(inode->i_mapping, pos, pos + left);
480 	if (ret < 0)
481 		return ret;
482 
483 	ret = invalidate_inode_pages2_range(inode->i_mapping,
484 					    pos >> PAGE_CACHE_SHIFT,
485 					    (pos + left) >> PAGE_CACHE_SHIFT);
486 	if (ret < 0)
487 		dout("invalidate_inode_pages2_range returned %d\n", ret);
488 
489 	flags = CEPH_OSD_FLAG_ORDERSNAP |
490 		CEPH_OSD_FLAG_ONDISK |
491 		CEPH_OSD_FLAG_WRITE;
492 	if ((file->f_flags & (O_SYNC|O_DIRECT)) == 0)
493 		flags |= CEPH_OSD_FLAG_ACK;
494 	else
495 		do_sync = 1;
496 
497 	/*
498 	 * we may need to do multiple writes here if we span an object
499 	 * boundary.  this isn't atomic, unfortunately.  :(
500 	 */
501 more:
502 	io_align = pos & ~PAGE_MASK;
503 	buf_align = (unsigned long)data & ~PAGE_MASK;
504 	len = left;
505 	if (file->f_flags & O_DIRECT) {
506 		/* write from beginning of first page, regardless of
507 		   io alignment */
508 		page_align = (pos - io_align + buf_align) & ~PAGE_MASK;
509 		num_pages = calc_pages_for((unsigned long)data, len);
510 	} else {
511 		page_align = pos & ~PAGE_MASK;
512 		num_pages = calc_pages_for(pos, len);
513 	}
514 	req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
515 				    ceph_vino(inode), pos, &len,
516 				    CEPH_OSD_OP_WRITE, flags,
517 				    ci->i_snap_realm->cached_context,
518 				    do_sync,
519 				    ci->i_truncate_seq, ci->i_truncate_size,
520 				    &mtime, false, 2, page_align);
521 	if (!req)
522 		return -ENOMEM;
523 
524 	if (file->f_flags & O_DIRECT) {
525 		pages = ceph_get_direct_page_vector(data, num_pages, false);
526 		if (IS_ERR(pages)) {
527 			ret = PTR_ERR(pages);
528 			goto out;
529 		}
530 
531 		/*
532 		 * throw out any page cache pages in this range. this
533 		 * may block.
534 		 */
535 		truncate_inode_pages_range(inode->i_mapping, pos,
536 					   (pos+len) | (PAGE_CACHE_SIZE-1));
537 	} else {
538 		pages = ceph_alloc_page_vector(num_pages, GFP_NOFS);
539 		if (IS_ERR(pages)) {
540 			ret = PTR_ERR(pages);
541 			goto out;
542 		}
543 		ret = ceph_copy_user_to_page_vector(pages, data, pos, len);
544 		if (ret < 0) {
545 			ceph_release_page_vector(pages, num_pages);
546 			goto out;
547 		}
548 
549 		if ((file->f_flags & O_SYNC) == 0) {
550 			/* get a second commit callback */
551 			req->r_safe_callback = sync_write_commit;
552 			req->r_own_pages = 1;
553 		}
554 	}
555 	req->r_pages = pages;
556 	req->r_num_pages = num_pages;
557 	req->r_inode = inode;
558 
559 	ret = ceph_osdc_start_request(&fsc->client->osdc, req, false);
560 	if (!ret) {
561 		if (req->r_safe_callback) {
562 			/*
563 			 * Add to inode unsafe list only after we
564 			 * start_request so that a tid has been assigned.
565 			 */
566 			spin_lock(&ci->i_unsafe_lock);
567 			list_add_tail(&req->r_unsafe_item,
568 				      &ci->i_unsafe_writes);
569 			spin_unlock(&ci->i_unsafe_lock);
570 			ceph_get_cap_refs(ci, CEPH_CAP_FILE_WR);
571 		}
572 
573 		ret = ceph_osdc_wait_request(&fsc->client->osdc, req);
574 		if (ret < 0 && req->r_safe_callback) {
575 			spin_lock(&ci->i_unsafe_lock);
576 			list_del_init(&req->r_unsafe_item);
577 			spin_unlock(&ci->i_unsafe_lock);
578 			ceph_put_cap_refs(ci, CEPH_CAP_FILE_WR);
579 		}
580 	}
581 
582 	if (file->f_flags & O_DIRECT)
583 		ceph_put_page_vector(pages, num_pages, false);
584 	else if (file->f_flags & O_SYNC)
585 		ceph_release_page_vector(pages, num_pages);
586 
587 out:
588 	ceph_osdc_put_request(req);
589 	if (ret == 0) {
590 		pos += len;
591 		written += len;
592 		left -= len;
593 		data += written;
594 		if (left)
595 			goto more;
596 
597 		ret = written;
598 		*offset = pos;
599 		if (pos > i_size_read(inode))
600 			check_caps = ceph_inode_set_size(inode, pos);
601 		if (check_caps)
602 			ceph_check_caps(ceph_inode(inode), CHECK_CAPS_AUTHONLY,
603 					NULL);
604 	}
605 	return ret;
606 }
607 
608 /*
609  * Wrap generic_file_aio_read with checks for cap bits on the inode.
610  * Atomically grab references, so that those bits are not released
611  * back to the MDS mid-read.
612  *
613  * Hmm, the sync read case isn't actually async... should it be?
614  */
615 static ssize_t ceph_aio_read(struct kiocb *iocb, const struct iovec *iov,
616 			     unsigned long nr_segs, loff_t pos)
617 {
618 	struct file *filp = iocb->ki_filp;
619 	struct ceph_file_info *fi = filp->private_data;
620 	loff_t *ppos = &iocb->ki_pos;
621 	size_t len = iov->iov_len;
622 	struct inode *inode = filp->f_dentry->d_inode;
623 	struct ceph_inode_info *ci = ceph_inode(inode);
624 	void __user *base = iov->iov_base;
625 	ssize_t ret;
626 	int want, got = 0;
627 	int checkeof = 0, read = 0;
628 
629 	dout("aio_read %p %llx.%llx %llu~%u trying to get caps on %p\n",
630 	     inode, ceph_vinop(inode), pos, (unsigned)len, inode);
631 again:
632 	__ceph_do_pending_vmtruncate(inode);
633 	if (fi->fmode & CEPH_FILE_MODE_LAZY)
634 		want = CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO;
635 	else
636 		want = CEPH_CAP_FILE_CACHE;
637 	ret = ceph_get_caps(ci, CEPH_CAP_FILE_RD, want, &got, -1);
638 	if (ret < 0)
639 		goto out;
640 	dout("aio_read %p %llx.%llx %llu~%u got cap refs on %s\n",
641 	     inode, ceph_vinop(inode), pos, (unsigned)len,
642 	     ceph_cap_string(got));
643 
644 	if ((got & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) == 0 ||
645 	    (iocb->ki_filp->f_flags & O_DIRECT) ||
646 	    (inode->i_sb->s_flags & MS_SYNCHRONOUS))
647 		/* hmm, this isn't really async... */
648 		ret = ceph_sync_read(filp, base, len, ppos, &checkeof);
649 	else
650 		ret = generic_file_aio_read(iocb, iov, nr_segs, pos);
651 
652 out:
653 	dout("aio_read %p %llx.%llx dropping cap refs on %s = %d\n",
654 	     inode, ceph_vinop(inode), ceph_cap_string(got), (int)ret);
655 	ceph_put_cap_refs(ci, got);
656 
657 	if (checkeof && ret >= 0) {
658 		int statret = ceph_do_getattr(inode, CEPH_STAT_CAP_SIZE);
659 
660 		/* hit EOF or hole? */
661 		if (statret == 0 && *ppos < inode->i_size) {
662 			dout("aio_read sync_read hit hole, ppos %lld < size %lld, reading more\n", *ppos, inode->i_size);
663 			read += ret;
664 			base += ret;
665 			len -= ret;
666 			checkeof = 0;
667 			goto again;
668 		}
669 	}
670 	if (ret >= 0)
671 		ret += read;
672 
673 	return ret;
674 }
675 
676 /*
677  * Take cap references to avoid releasing caps to MDS mid-write.
678  *
679  * If we are synchronous, and write with an old snap context, the OSD
680  * may return EOLDSNAPC.  In that case, retry the write.. _after_
681  * dropping our cap refs and allowing the pending snap to logically
682  * complete _before_ this write occurs.
683  *
684  * If we are near ENOSPC, write synchronously.
685  */
686 static ssize_t ceph_aio_write(struct kiocb *iocb, const struct iovec *iov,
687 		       unsigned long nr_segs, loff_t pos)
688 {
689 	struct file *file = iocb->ki_filp;
690 	struct ceph_file_info *fi = file->private_data;
691 	struct inode *inode = file->f_dentry->d_inode;
692 	struct ceph_inode_info *ci = ceph_inode(inode);
693 	struct ceph_osd_client *osdc =
694 		&ceph_sb_to_client(inode->i_sb)->client->osdc;
695 	loff_t endoff = pos + iov->iov_len;
696 	int want, got = 0;
697 	int ret, err;
698 
699 	if (ceph_snap(inode) != CEPH_NOSNAP)
700 		return -EROFS;
701 
702 retry_snap:
703 	if (ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_FULL))
704 		return -ENOSPC;
705 	__ceph_do_pending_vmtruncate(inode);
706 	dout("aio_write %p %llx.%llx %llu~%u getting caps. i_size %llu\n",
707 	     inode, ceph_vinop(inode), pos, (unsigned)iov->iov_len,
708 	     inode->i_size);
709 	if (fi->fmode & CEPH_FILE_MODE_LAZY)
710 		want = CEPH_CAP_FILE_BUFFER | CEPH_CAP_FILE_LAZYIO;
711 	else
712 		want = CEPH_CAP_FILE_BUFFER;
713 	ret = ceph_get_caps(ci, CEPH_CAP_FILE_WR, want, &got, endoff);
714 	if (ret < 0)
715 		goto out;
716 
717 	dout("aio_write %p %llx.%llx %llu~%u  got cap refs on %s\n",
718 	     inode, ceph_vinop(inode), pos, (unsigned)iov->iov_len,
719 	     ceph_cap_string(got));
720 
721 	if ((got & (CEPH_CAP_FILE_BUFFER|CEPH_CAP_FILE_LAZYIO)) == 0 ||
722 	    (iocb->ki_filp->f_flags & O_DIRECT) ||
723 	    (inode->i_sb->s_flags & MS_SYNCHRONOUS)) {
724 		ret = ceph_sync_write(file, iov->iov_base, iov->iov_len,
725 			&iocb->ki_pos);
726 	} else {
727 		ret = generic_file_aio_write(iocb, iov, nr_segs, pos);
728 
729 		if ((ret >= 0 || ret == -EIOCBQUEUED) &&
730 		    ((file->f_flags & O_SYNC) || IS_SYNC(file->f_mapping->host)
731 		     || ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_NEARFULL))) {
732 			err = vfs_fsync_range(file, pos, pos + ret - 1, 1);
733 			if (err < 0)
734 				ret = err;
735 		}
736 	}
737 	if (ret >= 0) {
738 		int dirty;
739 		spin_lock(&inode->i_lock);
740 		dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR);
741 		spin_unlock(&inode->i_lock);
742 		if (dirty)
743 			__mark_inode_dirty(inode, dirty);
744 	}
745 
746 out:
747 	dout("aio_write %p %llx.%llx %llu~%u  dropping cap refs on %s\n",
748 	     inode, ceph_vinop(inode), pos, (unsigned)iov->iov_len,
749 	     ceph_cap_string(got));
750 	ceph_put_cap_refs(ci, got);
751 
752 	if (ret == -EOLDSNAPC) {
753 		dout("aio_write %p %llx.%llx %llu~%u got EOLDSNAPC, retrying\n",
754 		     inode, ceph_vinop(inode), pos, (unsigned)iov->iov_len);
755 		goto retry_snap;
756 	}
757 
758 	return ret;
759 }
760 
761 /*
762  * llseek.  be sure to verify file size on SEEK_END.
763  */
764 static loff_t ceph_llseek(struct file *file, loff_t offset, int origin)
765 {
766 	struct inode *inode = file->f_mapping->host;
767 	int ret;
768 
769 	mutex_lock(&inode->i_mutex);
770 	__ceph_do_pending_vmtruncate(inode);
771 	switch (origin) {
772 	case SEEK_END:
773 		ret = ceph_do_getattr(inode, CEPH_STAT_CAP_SIZE);
774 		if (ret < 0) {
775 			offset = ret;
776 			goto out;
777 		}
778 		offset += inode->i_size;
779 		break;
780 	case SEEK_CUR:
781 		/*
782 		 * Here we special-case the lseek(fd, 0, SEEK_CUR)
783 		 * position-querying operation.  Avoid rewriting the "same"
784 		 * f_pos value back to the file because a concurrent read(),
785 		 * write() or lseek() might have altered it
786 		 */
787 		if (offset == 0) {
788 			offset = file->f_pos;
789 			goto out;
790 		}
791 		offset += file->f_pos;
792 		break;
793 	}
794 
795 	if (offset < 0 || offset > inode->i_sb->s_maxbytes) {
796 		offset = -EINVAL;
797 		goto out;
798 	}
799 
800 	/* Special lock needed here? */
801 	if (offset != file->f_pos) {
802 		file->f_pos = offset;
803 		file->f_version = 0;
804 	}
805 
806 out:
807 	mutex_unlock(&inode->i_mutex);
808 	return offset;
809 }
810 
811 const struct file_operations ceph_file_fops = {
812 	.open = ceph_open,
813 	.release = ceph_release,
814 	.llseek = ceph_llseek,
815 	.read = do_sync_read,
816 	.write = do_sync_write,
817 	.aio_read = ceph_aio_read,
818 	.aio_write = ceph_aio_write,
819 	.mmap = ceph_mmap,
820 	.fsync = ceph_fsync,
821 	.lock = ceph_lock,
822 	.flock = ceph_flock,
823 	.splice_read = generic_file_splice_read,
824 	.splice_write = generic_file_splice_write,
825 	.unlocked_ioctl = ceph_ioctl,
826 	.compat_ioctl	= ceph_ioctl,
827 };
828 
829