Loading...
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/falloc.h>
11
12#include "super.h"
13#include "mds_client.h"
14#include "cache.h"
15
16/*
17 * Ceph file operations
18 *
19 * Implement basic open/close functionality, and implement
20 * read/write.
21 *
22 * We implement three modes of file I/O:
23 * - buffered uses the generic_file_aio_{read,write} helpers
24 *
25 * - synchronous is used when there is multi-client read/write
26 * sharing, avoids the page cache, and synchronously waits for an
27 * ack from the OSD.
28 *
29 * - direct io takes the variant of the sync path that references
30 * user pages directly.
31 *
32 * fsync() flushes and waits on dirty pages, but just queues metadata
33 * for writeback: since the MDS can recover size and mtime there is no
34 * need to wait for MDS acknowledgement.
35 */
36
37/*
38 * Calculate the length sum of direct io vectors that can
39 * be combined into one page vector.
40 */
41static size_t dio_get_pagev_size(const struct iov_iter *it)
42{
43 const struct iovec *iov = it->iov;
44 const struct iovec *iovend = iov + it->nr_segs;
45 size_t size;
46
47 size = iov->iov_len - it->iov_offset;
48 /*
49 * An iov can be page vectored when both the current tail
50 * and the next base are page aligned.
51 */
52 while (PAGE_ALIGNED((iov->iov_base + iov->iov_len)) &&
53 (++iov < iovend && PAGE_ALIGNED((iov->iov_base)))) {
54 size += iov->iov_len;
55 }
56 dout("dio_get_pagevlen len = %zu\n", size);
57 return size;
58}
59
60/*
61 * Allocate a page vector based on (@it, @nbytes).
62 * The return value is the tuple describing a page vector,
63 * that is (@pages, @page_align, @num_pages).
64 */
65static struct page **
66dio_get_pages_alloc(const struct iov_iter *it, size_t nbytes,
67 size_t *page_align, int *num_pages)
68{
69 struct iov_iter tmp_it = *it;
70 size_t align;
71 struct page **pages;
72 int ret = 0, idx, npages;
73
74 align = (unsigned long)(it->iov->iov_base + it->iov_offset) &
75 (PAGE_SIZE - 1);
76 npages = calc_pages_for(align, nbytes);
77 pages = kmalloc(sizeof(*pages) * npages, GFP_KERNEL);
78 if (!pages) {
79 pages = vmalloc(sizeof(*pages) * npages);
80 if (!pages)
81 return ERR_PTR(-ENOMEM);
82 }
83
84 for (idx = 0; idx < npages; ) {
85 size_t start;
86 ret = iov_iter_get_pages(&tmp_it, pages + idx, nbytes,
87 npages - idx, &start);
88 if (ret < 0)
89 goto fail;
90
91 iov_iter_advance(&tmp_it, ret);
92 nbytes -= ret;
93 idx += (ret + start + PAGE_SIZE - 1) / PAGE_SIZE;
94 }
95
96 BUG_ON(nbytes != 0);
97 *num_pages = npages;
98 *page_align = align;
99 dout("dio_get_pages_alloc: got %d pages align %zu\n", npages, align);
100 return pages;
101fail:
102 ceph_put_page_vector(pages, idx, false);
103 return ERR_PTR(ret);
104}
105
106/*
107 * Prepare an open request. Preallocate ceph_cap to avoid an
108 * inopportune ENOMEM later.
109 */
110static struct ceph_mds_request *
111prepare_open_request(struct super_block *sb, int flags, int create_mode)
112{
113 struct ceph_fs_client *fsc = ceph_sb_to_client(sb);
114 struct ceph_mds_client *mdsc = fsc->mdsc;
115 struct ceph_mds_request *req;
116 int want_auth = USE_ANY_MDS;
117 int op = (flags & O_CREAT) ? CEPH_MDS_OP_CREATE : CEPH_MDS_OP_OPEN;
118
119 if (flags & (O_WRONLY|O_RDWR|O_CREAT|O_TRUNC))
120 want_auth = USE_AUTH_MDS;
121
122 req = ceph_mdsc_create_request(mdsc, op, want_auth);
123 if (IS_ERR(req))
124 goto out;
125 req->r_fmode = ceph_flags_to_mode(flags);
126 req->r_args.open.flags = cpu_to_le32(flags);
127 req->r_args.open.mode = cpu_to_le32(create_mode);
128out:
129 return req;
130}
131
132/*
133 * initialize private struct file data.
134 * if we fail, clean up by dropping fmode reference on the ceph_inode
135 */
136static int ceph_init_file(struct inode *inode, struct file *file, int fmode)
137{
138 struct ceph_file_info *cf;
139 int ret = 0;
140 struct ceph_inode_info *ci = ceph_inode(inode);
141 struct ceph_fs_client *fsc = ceph_sb_to_client(inode->i_sb);
142 struct ceph_mds_client *mdsc = fsc->mdsc;
143
144 switch (inode->i_mode & S_IFMT) {
145 case S_IFREG:
146 /* First file open request creates the cookie, we want to keep
147 * this cookie around for the filetime of the inode as not to
148 * have to worry about fscache register / revoke / operation
149 * races.
150 *
151 * Also, if we know the operation is going to invalidate data
152 * (non readonly) just nuke the cache right away.
153 */
154 ceph_fscache_register_inode_cookie(mdsc->fsc, ci);
155 if ((fmode & CEPH_FILE_MODE_WR))
156 ceph_fscache_invalidate(inode);
157 case S_IFDIR:
158 dout("init_file %p %p 0%o (regular)\n", inode, file,
159 inode->i_mode);
160 cf = kmem_cache_zalloc(ceph_file_cachep, GFP_KERNEL);
161 if (cf == NULL) {
162 ceph_put_fmode(ceph_inode(inode), fmode); /* clean up */
163 return -ENOMEM;
164 }
165 cf->fmode = fmode;
166 cf->next_offset = 2;
167 cf->readdir_cache_idx = -1;
168 file->private_data = cf;
169 BUG_ON(inode->i_fop->release != ceph_release);
170 break;
171
172 case S_IFLNK:
173 dout("init_file %p %p 0%o (symlink)\n", inode, file,
174 inode->i_mode);
175 ceph_put_fmode(ceph_inode(inode), fmode); /* clean up */
176 break;
177
178 default:
179 dout("init_file %p %p 0%o (special)\n", inode, file,
180 inode->i_mode);
181 /*
182 * we need to drop the open ref now, since we don't
183 * have .release set to ceph_release.
184 */
185 ceph_put_fmode(ceph_inode(inode), fmode); /* clean up */
186 BUG_ON(inode->i_fop->release == ceph_release);
187
188 /* call the proper open fop */
189 ret = inode->i_fop->open(inode, file);
190 }
191 return ret;
192}
193
194/*
195 * If we already have the requisite capabilities, we can satisfy
196 * the open request locally (no need to request new caps from the
197 * MDS). We do, however, need to inform the MDS (asynchronously)
198 * if our wanted caps set expands.
199 */
200int ceph_open(struct inode *inode, struct file *file)
201{
202 struct ceph_inode_info *ci = ceph_inode(inode);
203 struct ceph_fs_client *fsc = ceph_sb_to_client(inode->i_sb);
204 struct ceph_mds_client *mdsc = fsc->mdsc;
205 struct ceph_mds_request *req;
206 struct ceph_file_info *cf = file->private_data;
207 int err;
208 int flags, fmode, wanted;
209
210 if (cf) {
211 dout("open file %p is already opened\n", file);
212 return 0;
213 }
214
215 /* filter out O_CREAT|O_EXCL; vfs did that already. yuck. */
216 flags = file->f_flags & ~(O_CREAT|O_EXCL);
217 if (S_ISDIR(inode->i_mode))
218 flags = O_DIRECTORY; /* mds likes to know */
219
220 dout("open inode %p ino %llx.%llx file %p flags %d (%d)\n", inode,
221 ceph_vinop(inode), file, flags, file->f_flags);
222 fmode = ceph_flags_to_mode(flags);
223 wanted = ceph_caps_for_mode(fmode);
224
225 /* snapped files are read-only */
226 if (ceph_snap(inode) != CEPH_NOSNAP && (file->f_mode & FMODE_WRITE))
227 return -EROFS;
228
229 /* trivially open snapdir */
230 if (ceph_snap(inode) == CEPH_SNAPDIR) {
231 spin_lock(&ci->i_ceph_lock);
232 __ceph_get_fmode(ci, fmode);
233 spin_unlock(&ci->i_ceph_lock);
234 return ceph_init_file(inode, file, fmode);
235 }
236
237 /*
238 * No need to block if we have caps on the auth MDS (for
239 * write) or any MDS (for read). Update wanted set
240 * asynchronously.
241 */
242 spin_lock(&ci->i_ceph_lock);
243 if (__ceph_is_any_real_caps(ci) &&
244 (((fmode & CEPH_FILE_MODE_WR) == 0) || ci->i_auth_cap)) {
245 int mds_wanted = __ceph_caps_mds_wanted(ci);
246 int issued = __ceph_caps_issued(ci, NULL);
247
248 dout("open %p fmode %d want %s issued %s using existing\n",
249 inode, fmode, ceph_cap_string(wanted),
250 ceph_cap_string(issued));
251 __ceph_get_fmode(ci, fmode);
252 spin_unlock(&ci->i_ceph_lock);
253
254 /* adjust wanted? */
255 if ((issued & wanted) != wanted &&
256 (mds_wanted & wanted) != wanted &&
257 ceph_snap(inode) != CEPH_SNAPDIR)
258 ceph_check_caps(ci, 0, NULL);
259
260 return ceph_init_file(inode, file, fmode);
261 } else if (ceph_snap(inode) != CEPH_NOSNAP &&
262 (ci->i_snap_caps & wanted) == wanted) {
263 __ceph_get_fmode(ci, fmode);
264 spin_unlock(&ci->i_ceph_lock);
265 return ceph_init_file(inode, file, fmode);
266 }
267
268 spin_unlock(&ci->i_ceph_lock);
269
270 dout("open fmode %d wants %s\n", fmode, ceph_cap_string(wanted));
271 req = prepare_open_request(inode->i_sb, flags, 0);
272 if (IS_ERR(req)) {
273 err = PTR_ERR(req);
274 goto out;
275 }
276 req->r_inode = inode;
277 ihold(inode);
278
279 req->r_num_caps = 1;
280 err = ceph_mdsc_do_request(mdsc, NULL, req);
281 if (!err)
282 err = ceph_init_file(inode, file, req->r_fmode);
283 ceph_mdsc_put_request(req);
284 dout("open result=%d on %llx.%llx\n", err, ceph_vinop(inode));
285out:
286 return err;
287}
288
289
290/*
291 * Do a lookup + open with a single request. If we get a non-existent
292 * file or symlink, return 1 so the VFS can retry.
293 */
294int ceph_atomic_open(struct inode *dir, struct dentry *dentry,
295 struct file *file, unsigned flags, umode_t mode,
296 int *opened)
297{
298 struct ceph_fs_client *fsc = ceph_sb_to_client(dir->i_sb);
299 struct ceph_mds_client *mdsc = fsc->mdsc;
300 struct ceph_mds_request *req;
301 struct dentry *dn;
302 struct ceph_acls_info acls = {};
303 int mask;
304 int err;
305
306 dout("atomic_open %p dentry %p '%pd' %s flags %d mode 0%o\n",
307 dir, dentry, dentry,
308 d_unhashed(dentry) ? "unhashed" : "hashed", flags, mode);
309
310 if (dentry->d_name.len > NAME_MAX)
311 return -ENAMETOOLONG;
312
313 err = ceph_init_dentry(dentry);
314 if (err < 0)
315 return err;
316
317 if (flags & O_CREAT) {
318 err = ceph_pre_init_acls(dir, &mode, &acls);
319 if (err < 0)
320 return err;
321 }
322
323 /* do the open */
324 req = prepare_open_request(dir->i_sb, flags, mode);
325 if (IS_ERR(req)) {
326 err = PTR_ERR(req);
327 goto out_acl;
328 }
329 req->r_dentry = dget(dentry);
330 req->r_num_caps = 2;
331 if (flags & O_CREAT) {
332 req->r_dentry_drop = CEPH_CAP_FILE_SHARED;
333 req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
334 if (acls.pagelist) {
335 req->r_pagelist = acls.pagelist;
336 acls.pagelist = NULL;
337 }
338 }
339
340 mask = CEPH_STAT_CAP_INODE | CEPH_CAP_AUTH_SHARED;
341 if (ceph_security_xattr_wanted(dir))
342 mask |= CEPH_CAP_XATTR_SHARED;
343 req->r_args.open.mask = cpu_to_le32(mask);
344
345 req->r_locked_dir = dir; /* caller holds dir->i_mutex */
346 err = ceph_mdsc_do_request(mdsc,
347 (flags & (O_CREAT|O_TRUNC)) ? dir : NULL,
348 req);
349 err = ceph_handle_snapdir(req, dentry, err);
350 if (err)
351 goto out_req;
352
353 if ((flags & O_CREAT) && !req->r_reply_info.head->is_dentry)
354 err = ceph_handle_notrace_create(dir, dentry);
355
356 if (d_unhashed(dentry)) {
357 dn = ceph_finish_lookup(req, dentry, err);
358 if (IS_ERR(dn))
359 err = PTR_ERR(dn);
360 } else {
361 /* we were given a hashed negative dentry */
362 dn = NULL;
363 }
364 if (err)
365 goto out_req;
366 if (dn || d_really_is_negative(dentry) || d_is_symlink(dentry)) {
367 /* make vfs retry on splice, ENOENT, or symlink */
368 dout("atomic_open finish_no_open on dn %p\n", dn);
369 err = finish_no_open(file, dn);
370 } else {
371 dout("atomic_open finish_open on dn %p\n", dn);
372 if (req->r_op == CEPH_MDS_OP_CREATE && req->r_reply_info.has_create_ino) {
373 ceph_init_inode_acls(d_inode(dentry), &acls);
374 *opened |= FILE_CREATED;
375 }
376 err = finish_open(file, dentry, ceph_open, opened);
377 }
378out_req:
379 if (!req->r_err && req->r_target_inode)
380 ceph_put_fmode(ceph_inode(req->r_target_inode), req->r_fmode);
381 ceph_mdsc_put_request(req);
382out_acl:
383 ceph_release_acls_info(&acls);
384 dout("atomic_open result=%d\n", err);
385 return err;
386}
387
388int ceph_release(struct inode *inode, struct file *file)
389{
390 struct ceph_inode_info *ci = ceph_inode(inode);
391 struct ceph_file_info *cf = file->private_data;
392
393 dout("release inode %p file %p\n", inode, file);
394 ceph_put_fmode(ci, cf->fmode);
395 if (cf->last_readdir)
396 ceph_mdsc_put_request(cf->last_readdir);
397 kfree(cf->last_name);
398 kfree(cf->dir_info);
399 kmem_cache_free(ceph_file_cachep, cf);
400
401 /* wake up anyone waiting for caps on this inode */
402 wake_up_all(&ci->i_cap_wq);
403 return 0;
404}
405
406enum {
407 HAVE_RETRIED = 1,
408 CHECK_EOF = 2,
409 READ_INLINE = 3,
410};
411
412/*
413 * Read a range of bytes striped over one or more objects. Iterate over
414 * objects we stripe over. (That's not atomic, but good enough for now.)
415 *
416 * If we get a short result from the OSD, check against i_size; we need to
417 * only return a short read to the caller if we hit EOF.
418 */
419static int striped_read(struct inode *inode,
420 u64 off, u64 len,
421 struct page **pages, int num_pages,
422 int *checkeof)
423{
424 struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
425 struct ceph_inode_info *ci = ceph_inode(inode);
426 u64 pos, this_len, left;
427 loff_t i_size;
428 int page_align, pages_left;
429 int read, ret;
430 struct page **page_pos;
431 bool hit_stripe, was_short;
432
433 /*
434 * we may need to do multiple reads. not atomic, unfortunately.
435 */
436 pos = off;
437 left = len;
438 page_pos = pages;
439 pages_left = num_pages;
440 read = 0;
441
442more:
443 page_align = pos & ~PAGE_MASK;
444 this_len = left;
445 ret = ceph_osdc_readpages(&fsc->client->osdc, ceph_vino(inode),
446 &ci->i_layout, pos, &this_len,
447 ci->i_truncate_seq,
448 ci->i_truncate_size,
449 page_pos, pages_left, page_align);
450 if (ret == -ENOENT)
451 ret = 0;
452 hit_stripe = this_len < left;
453 was_short = ret >= 0 && ret < this_len;
454 dout("striped_read %llu~%llu (read %u) got %d%s%s\n", pos, left, read,
455 ret, hit_stripe ? " HITSTRIPE" : "", was_short ? " SHORT" : "");
456
457 i_size = i_size_read(inode);
458 if (ret >= 0) {
459 int didpages;
460 if (was_short && (pos + ret < i_size)) {
461 int zlen = min(this_len - ret, i_size - pos - ret);
462 int zoff = (off & ~PAGE_MASK) + read + ret;
463 dout(" zero gap %llu to %llu\n",
464 pos + ret, pos + ret + zlen);
465 ceph_zero_page_vector_range(zoff, zlen, pages);
466 ret += zlen;
467 }
468
469 didpages = (page_align + ret) >> PAGE_SHIFT;
470 pos += ret;
471 read = pos - off;
472 left -= ret;
473 page_pos += didpages;
474 pages_left -= didpages;
475
476 /* hit stripe and need continue*/
477 if (left && hit_stripe && pos < i_size)
478 goto more;
479 }
480
481 if (read > 0) {
482 ret = read;
483 /* did we bounce off eof? */
484 if (pos + left > i_size)
485 *checkeof = CHECK_EOF;
486 }
487
488 dout("striped_read returns %d\n", ret);
489 return ret;
490}
491
492/*
493 * Completely synchronous read and write methods. Direct from __user
494 * buffer to osd, or directly to user pages (if O_DIRECT).
495 *
496 * If the read spans object boundary, just do multiple reads.
497 */
498static ssize_t ceph_sync_read(struct kiocb *iocb, struct iov_iter *i,
499 int *checkeof)
500{
501 struct file *file = iocb->ki_filp;
502 struct inode *inode = file_inode(file);
503 struct page **pages;
504 u64 off = iocb->ki_pos;
505 int num_pages, ret;
506 size_t len = iov_iter_count(i);
507
508 dout("sync_read on file %p %llu~%u %s\n", file, off,
509 (unsigned)len,
510 (file->f_flags & O_DIRECT) ? "O_DIRECT" : "");
511
512 if (!len)
513 return 0;
514 /*
515 * flush any page cache pages in this range. this
516 * will make concurrent normal and sync io slow,
517 * but it will at least behave sensibly when they are
518 * in sequence.
519 */
520 ret = filemap_write_and_wait_range(inode->i_mapping, off,
521 off + len);
522 if (ret < 0)
523 return ret;
524
525 num_pages = calc_pages_for(off, len);
526 pages = ceph_alloc_page_vector(num_pages, GFP_KERNEL);
527 if (IS_ERR(pages))
528 return PTR_ERR(pages);
529 ret = striped_read(inode, off, len, pages,
530 num_pages, checkeof);
531 if (ret > 0) {
532 int l, k = 0;
533 size_t left = ret;
534
535 while (left) {
536 size_t page_off = off & ~PAGE_MASK;
537 size_t copy = min_t(size_t, left,
538 PAGE_SIZE - page_off);
539 l = copy_page_to_iter(pages[k++], page_off, copy, i);
540 off += l;
541 left -= l;
542 if (l < copy)
543 break;
544 }
545 }
546 ceph_release_page_vector(pages, num_pages);
547
548 if (off > iocb->ki_pos) {
549 ret = off - iocb->ki_pos;
550 iocb->ki_pos = off;
551 }
552
553 dout("sync_read result %d\n", ret);
554 return ret;
555}
556
557struct ceph_aio_request {
558 struct kiocb *iocb;
559 size_t total_len;
560 int write;
561 int error;
562 struct list_head osd_reqs;
563 unsigned num_reqs;
564 atomic_t pending_reqs;
565 struct timespec mtime;
566 struct ceph_cap_flush *prealloc_cf;
567};
568
569struct ceph_aio_work {
570 struct work_struct work;
571 struct ceph_osd_request *req;
572};
573
574static void ceph_aio_retry_work(struct work_struct *work);
575
576static void ceph_aio_complete(struct inode *inode,
577 struct ceph_aio_request *aio_req)
578{
579 struct ceph_inode_info *ci = ceph_inode(inode);
580 int ret;
581
582 if (!atomic_dec_and_test(&aio_req->pending_reqs))
583 return;
584
585 ret = aio_req->error;
586 if (!ret)
587 ret = aio_req->total_len;
588
589 dout("ceph_aio_complete %p rc %d\n", inode, ret);
590
591 if (ret >= 0 && aio_req->write) {
592 int dirty;
593
594 loff_t endoff = aio_req->iocb->ki_pos + aio_req->total_len;
595 if (endoff > i_size_read(inode)) {
596 if (ceph_inode_set_size(inode, endoff))
597 ceph_check_caps(ci, CHECK_CAPS_AUTHONLY, NULL);
598 }
599
600 spin_lock(&ci->i_ceph_lock);
601 ci->i_inline_version = CEPH_INLINE_NONE;
602 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR,
603 &aio_req->prealloc_cf);
604 spin_unlock(&ci->i_ceph_lock);
605 if (dirty)
606 __mark_inode_dirty(inode, dirty);
607
608 }
609
610 ceph_put_cap_refs(ci, (aio_req->write ? CEPH_CAP_FILE_WR :
611 CEPH_CAP_FILE_RD));
612
613 aio_req->iocb->ki_complete(aio_req->iocb, ret, 0);
614
615 ceph_free_cap_flush(aio_req->prealloc_cf);
616 kfree(aio_req);
617}
618
619static void ceph_aio_complete_req(struct ceph_osd_request *req,
620 struct ceph_msg *msg)
621{
622 int rc = req->r_result;
623 struct inode *inode = req->r_inode;
624 struct ceph_aio_request *aio_req = req->r_priv;
625 struct ceph_osd_data *osd_data = osd_req_op_extent_osd_data(req, 0);
626 int num_pages = calc_pages_for((u64)osd_data->alignment,
627 osd_data->length);
628
629 dout("ceph_aio_complete_req %p rc %d bytes %llu\n",
630 inode, rc, osd_data->length);
631
632 if (rc == -EOLDSNAPC) {
633 struct ceph_aio_work *aio_work;
634 BUG_ON(!aio_req->write);
635
636 aio_work = kmalloc(sizeof(*aio_work), GFP_NOFS);
637 if (aio_work) {
638 INIT_WORK(&aio_work->work, ceph_aio_retry_work);
639 aio_work->req = req;
640 queue_work(ceph_inode_to_client(inode)->wb_wq,
641 &aio_work->work);
642 return;
643 }
644 rc = -ENOMEM;
645 } else if (!aio_req->write) {
646 if (rc == -ENOENT)
647 rc = 0;
648 if (rc >= 0 && osd_data->length > rc) {
649 int zoff = osd_data->alignment + rc;
650 int zlen = osd_data->length - rc;
651 /*
652 * If read is satisfied by single OSD request,
653 * it can pass EOF. Otherwise read is within
654 * i_size.
655 */
656 if (aio_req->num_reqs == 1) {
657 loff_t i_size = i_size_read(inode);
658 loff_t endoff = aio_req->iocb->ki_pos + rc;
659 if (endoff < i_size)
660 zlen = min_t(size_t, zlen,
661 i_size - endoff);
662 aio_req->total_len = rc + zlen;
663 }
664
665 if (zlen > 0)
666 ceph_zero_page_vector_range(zoff, zlen,
667 osd_data->pages);
668 }
669 }
670
671 ceph_put_page_vector(osd_data->pages, num_pages, false);
672 ceph_osdc_put_request(req);
673
674 if (rc < 0)
675 cmpxchg(&aio_req->error, 0, rc);
676
677 ceph_aio_complete(inode, aio_req);
678 return;
679}
680
681static void ceph_aio_retry_work(struct work_struct *work)
682{
683 struct ceph_aio_work *aio_work =
684 container_of(work, struct ceph_aio_work, work);
685 struct ceph_osd_request *orig_req = aio_work->req;
686 struct ceph_aio_request *aio_req = orig_req->r_priv;
687 struct inode *inode = orig_req->r_inode;
688 struct ceph_inode_info *ci = ceph_inode(inode);
689 struct ceph_snap_context *snapc;
690 struct ceph_osd_request *req;
691 int ret;
692
693 spin_lock(&ci->i_ceph_lock);
694 if (__ceph_have_pending_cap_snap(ci)) {
695 struct ceph_cap_snap *capsnap =
696 list_last_entry(&ci->i_cap_snaps,
697 struct ceph_cap_snap,
698 ci_item);
699 snapc = ceph_get_snap_context(capsnap->context);
700 } else {
701 BUG_ON(!ci->i_head_snapc);
702 snapc = ceph_get_snap_context(ci->i_head_snapc);
703 }
704 spin_unlock(&ci->i_ceph_lock);
705
706 req = ceph_osdc_alloc_request(orig_req->r_osdc, snapc, 2,
707 false, GFP_NOFS);
708 if (!req) {
709 ret = -ENOMEM;
710 req = orig_req;
711 goto out;
712 }
713
714 req->r_flags = CEPH_OSD_FLAG_ORDERSNAP |
715 CEPH_OSD_FLAG_ONDISK |
716 CEPH_OSD_FLAG_WRITE;
717 req->r_base_oloc = orig_req->r_base_oloc;
718 req->r_base_oid = orig_req->r_base_oid;
719
720 req->r_ops[0] = orig_req->r_ops[0];
721 osd_req_op_init(req, 1, CEPH_OSD_OP_STARTSYNC, 0);
722
723 ceph_osdc_build_request(req, req->r_ops[0].extent.offset,
724 snapc, CEPH_NOSNAP, &aio_req->mtime);
725
726 ceph_osdc_put_request(orig_req);
727
728 req->r_callback = ceph_aio_complete_req;
729 req->r_inode = inode;
730 req->r_priv = aio_req;
731
732 ret = ceph_osdc_start_request(req->r_osdc, req, false);
733out:
734 if (ret < 0) {
735 req->r_result = ret;
736 ceph_aio_complete_req(req, NULL);
737 }
738
739 ceph_put_snap_context(snapc);
740 kfree(aio_work);
741}
742
743/*
744 * Write commit request unsafe callback, called to tell us when a
745 * request is unsafe (that is, in flight--has been handed to the
746 * messenger to send to its target osd). It is called again when
747 * we've received a response message indicating the request is
748 * "safe" (its CEPH_OSD_FLAG_ONDISK flag is set), or when a request
749 * is completed early (and unsuccessfully) due to a timeout or
750 * interrupt.
751 *
752 * This is used if we requested both an ACK and ONDISK commit reply
753 * from the OSD.
754 */
755static void ceph_sync_write_unsafe(struct ceph_osd_request *req, bool unsafe)
756{
757 struct ceph_inode_info *ci = ceph_inode(req->r_inode);
758
759 dout("%s %p tid %llu %ssafe\n", __func__, req, req->r_tid,
760 unsafe ? "un" : "");
761 if (unsafe) {
762 ceph_get_cap_refs(ci, CEPH_CAP_FILE_WR);
763 spin_lock(&ci->i_unsafe_lock);
764 list_add_tail(&req->r_unsafe_item,
765 &ci->i_unsafe_writes);
766 spin_unlock(&ci->i_unsafe_lock);
767 } else {
768 spin_lock(&ci->i_unsafe_lock);
769 list_del_init(&req->r_unsafe_item);
770 spin_unlock(&ci->i_unsafe_lock);
771 ceph_put_cap_refs(ci, CEPH_CAP_FILE_WR);
772 }
773}
774
775
776static ssize_t
777ceph_direct_read_write(struct kiocb *iocb, struct iov_iter *iter,
778 struct ceph_snap_context *snapc,
779 struct ceph_cap_flush **pcf)
780{
781 struct file *file = iocb->ki_filp;
782 struct inode *inode = file_inode(file);
783 struct ceph_inode_info *ci = ceph_inode(inode);
784 struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
785 struct ceph_vino vino;
786 struct ceph_osd_request *req;
787 struct page **pages;
788 struct ceph_aio_request *aio_req = NULL;
789 int num_pages = 0;
790 int flags;
791 int ret;
792 struct timespec mtime = current_fs_time(inode->i_sb);
793 size_t count = iov_iter_count(iter);
794 loff_t pos = iocb->ki_pos;
795 bool write = iov_iter_rw(iter) == WRITE;
796
797 if (write && ceph_snap(file_inode(file)) != CEPH_NOSNAP)
798 return -EROFS;
799
800 dout("sync_direct_read_write (%s) on file %p %lld~%u\n",
801 (write ? "write" : "read"), file, pos, (unsigned)count);
802
803 ret = filemap_write_and_wait_range(inode->i_mapping, pos, pos + count);
804 if (ret < 0)
805 return ret;
806
807 if (write) {
808 ret = invalidate_inode_pages2_range(inode->i_mapping,
809 pos >> PAGE_SHIFT,
810 (pos + count) >> PAGE_SHIFT);
811 if (ret < 0)
812 dout("invalidate_inode_pages2_range returned %d\n", ret);
813
814 flags = CEPH_OSD_FLAG_ORDERSNAP |
815 CEPH_OSD_FLAG_ONDISK |
816 CEPH_OSD_FLAG_WRITE;
817 } else {
818 flags = CEPH_OSD_FLAG_READ;
819 }
820
821 while (iov_iter_count(iter) > 0) {
822 u64 size = dio_get_pagev_size(iter);
823 size_t start = 0;
824 ssize_t len;
825
826 vino = ceph_vino(inode);
827 req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
828 vino, pos, &size, 0,
829 /*include a 'startsync' command*/
830 write ? 2 : 1,
831 write ? CEPH_OSD_OP_WRITE :
832 CEPH_OSD_OP_READ,
833 flags, snapc,
834 ci->i_truncate_seq,
835 ci->i_truncate_size,
836 false);
837 if (IS_ERR(req)) {
838 ret = PTR_ERR(req);
839 break;
840 }
841
842 len = size;
843 pages = dio_get_pages_alloc(iter, len, &start, &num_pages);
844 if (IS_ERR(pages)) {
845 ceph_osdc_put_request(req);
846 ret = PTR_ERR(pages);
847 break;
848 }
849
850 /*
851 * To simplify error handling, allow AIO when IO within i_size
852 * or IO can be satisfied by single OSD request.
853 */
854 if (pos == iocb->ki_pos && !is_sync_kiocb(iocb) &&
855 (len == count || pos + count <= i_size_read(inode))) {
856 aio_req = kzalloc(sizeof(*aio_req), GFP_KERNEL);
857 if (aio_req) {
858 aio_req->iocb = iocb;
859 aio_req->write = write;
860 INIT_LIST_HEAD(&aio_req->osd_reqs);
861 if (write) {
862 aio_req->mtime = mtime;
863 swap(aio_req->prealloc_cf, *pcf);
864 }
865 }
866 /* ignore error */
867 }
868
869 if (write) {
870 /*
871 * throw out any page cache pages in this range. this
872 * may block.
873 */
874 truncate_inode_pages_range(inode->i_mapping, pos,
875 (pos+len) | (PAGE_SIZE - 1));
876
877 osd_req_op_init(req, 1, CEPH_OSD_OP_STARTSYNC, 0);
878 }
879
880
881 osd_req_op_extent_osd_data_pages(req, 0, pages, len, start,
882 false, false);
883
884 ceph_osdc_build_request(req, pos, snapc, vino.snap, &mtime);
885
886 if (aio_req) {
887 aio_req->total_len += len;
888 aio_req->num_reqs++;
889 atomic_inc(&aio_req->pending_reqs);
890
891 req->r_callback = ceph_aio_complete_req;
892 req->r_inode = inode;
893 req->r_priv = aio_req;
894 list_add_tail(&req->r_unsafe_item, &aio_req->osd_reqs);
895
896 pos += len;
897 iov_iter_advance(iter, len);
898 continue;
899 }
900
901 ret = ceph_osdc_start_request(req->r_osdc, req, false);
902 if (!ret)
903 ret = ceph_osdc_wait_request(&fsc->client->osdc, req);
904
905 size = i_size_read(inode);
906 if (!write) {
907 if (ret == -ENOENT)
908 ret = 0;
909 if (ret >= 0 && ret < len && pos + ret < size) {
910 int zlen = min_t(size_t, len - ret,
911 size - pos - ret);
912 ceph_zero_page_vector_range(start + ret, zlen,
913 pages);
914 ret += zlen;
915 }
916 if (ret >= 0)
917 len = ret;
918 }
919
920 ceph_put_page_vector(pages, num_pages, false);
921
922 ceph_osdc_put_request(req);
923 if (ret < 0)
924 break;
925
926 pos += len;
927 iov_iter_advance(iter, len);
928
929 if (!write && pos >= size)
930 break;
931
932 if (write && pos > size) {
933 if (ceph_inode_set_size(inode, pos))
934 ceph_check_caps(ceph_inode(inode),
935 CHECK_CAPS_AUTHONLY,
936 NULL);
937 }
938 }
939
940 if (aio_req) {
941 if (aio_req->num_reqs == 0) {
942 kfree(aio_req);
943 return ret;
944 }
945
946 ceph_get_cap_refs(ci, write ? CEPH_CAP_FILE_WR :
947 CEPH_CAP_FILE_RD);
948
949 while (!list_empty(&aio_req->osd_reqs)) {
950 req = list_first_entry(&aio_req->osd_reqs,
951 struct ceph_osd_request,
952 r_unsafe_item);
953 list_del_init(&req->r_unsafe_item);
954 if (ret >= 0)
955 ret = ceph_osdc_start_request(req->r_osdc,
956 req, false);
957 if (ret < 0) {
958 req->r_result = ret;
959 ceph_aio_complete_req(req, NULL);
960 }
961 }
962 return -EIOCBQUEUED;
963 }
964
965 if (ret != -EOLDSNAPC && pos > iocb->ki_pos) {
966 ret = pos - iocb->ki_pos;
967 iocb->ki_pos = pos;
968 }
969 return ret;
970}
971
972/*
973 * Synchronous write, straight from __user pointer or user pages.
974 *
975 * If write spans object boundary, just do multiple writes. (For a
976 * correct atomic write, we should e.g. take write locks on all
977 * objects, rollback on failure, etc.)
978 */
979static ssize_t
980ceph_sync_write(struct kiocb *iocb, struct iov_iter *from, loff_t pos,
981 struct ceph_snap_context *snapc)
982{
983 struct file *file = iocb->ki_filp;
984 struct inode *inode = file_inode(file);
985 struct ceph_inode_info *ci = ceph_inode(inode);
986 struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
987 struct ceph_vino vino;
988 struct ceph_osd_request *req;
989 struct page **pages;
990 u64 len;
991 int num_pages;
992 int written = 0;
993 int flags;
994 int check_caps = 0;
995 int ret;
996 struct timespec mtime = current_fs_time(inode->i_sb);
997 size_t count = iov_iter_count(from);
998
999 if (ceph_snap(file_inode(file)) != CEPH_NOSNAP)
1000 return -EROFS;
1001
1002 dout("sync_write on file %p %lld~%u\n", file, pos, (unsigned)count);
1003
1004 ret = filemap_write_and_wait_range(inode->i_mapping, pos, pos + count);
1005 if (ret < 0)
1006 return ret;
1007
1008 ret = invalidate_inode_pages2_range(inode->i_mapping,
1009 pos >> PAGE_SHIFT,
1010 (pos + count) >> PAGE_SHIFT);
1011 if (ret < 0)
1012 dout("invalidate_inode_pages2_range returned %d\n", ret);
1013
1014 flags = CEPH_OSD_FLAG_ORDERSNAP |
1015 CEPH_OSD_FLAG_ONDISK |
1016 CEPH_OSD_FLAG_WRITE |
1017 CEPH_OSD_FLAG_ACK;
1018
1019 while ((len = iov_iter_count(from)) > 0) {
1020 size_t left;
1021 int n;
1022
1023 vino = ceph_vino(inode);
1024 req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1025 vino, pos, &len, 0, 1,
1026 CEPH_OSD_OP_WRITE, flags, snapc,
1027 ci->i_truncate_seq,
1028 ci->i_truncate_size,
1029 false);
1030 if (IS_ERR(req)) {
1031 ret = PTR_ERR(req);
1032 break;
1033 }
1034
1035 /*
1036 * write from beginning of first page,
1037 * regardless of io alignment
1038 */
1039 num_pages = (len + PAGE_SIZE - 1) >> PAGE_SHIFT;
1040
1041 pages = ceph_alloc_page_vector(num_pages, GFP_KERNEL);
1042 if (IS_ERR(pages)) {
1043 ret = PTR_ERR(pages);
1044 goto out;
1045 }
1046
1047 left = len;
1048 for (n = 0; n < num_pages; n++) {
1049 size_t plen = min_t(size_t, left, PAGE_SIZE);
1050 ret = copy_page_from_iter(pages[n], 0, plen, from);
1051 if (ret != plen) {
1052 ret = -EFAULT;
1053 break;
1054 }
1055 left -= ret;
1056 }
1057
1058 if (ret < 0) {
1059 ceph_release_page_vector(pages, num_pages);
1060 goto out;
1061 }
1062
1063 /* get a second commit callback */
1064 req->r_unsafe_callback = ceph_sync_write_unsafe;
1065 req->r_inode = inode;
1066
1067 osd_req_op_extent_osd_data_pages(req, 0, pages, len, 0,
1068 false, true);
1069
1070 /* BUG_ON(vino.snap != CEPH_NOSNAP); */
1071 ceph_osdc_build_request(req, pos, snapc, vino.snap, &mtime);
1072
1073 ret = ceph_osdc_start_request(&fsc->client->osdc, req, false);
1074 if (!ret)
1075 ret = ceph_osdc_wait_request(&fsc->client->osdc, req);
1076
1077out:
1078 ceph_osdc_put_request(req);
1079 if (ret == 0) {
1080 pos += len;
1081 written += len;
1082
1083 if (pos > i_size_read(inode)) {
1084 check_caps = ceph_inode_set_size(inode, pos);
1085 if (check_caps)
1086 ceph_check_caps(ceph_inode(inode),
1087 CHECK_CAPS_AUTHONLY,
1088 NULL);
1089 }
1090 } else
1091 break;
1092 }
1093
1094 if (ret != -EOLDSNAPC && written > 0) {
1095 ret = written;
1096 iocb->ki_pos = pos;
1097 }
1098 return ret;
1099}
1100
1101/*
1102 * Wrap generic_file_aio_read with checks for cap bits on the inode.
1103 * Atomically grab references, so that those bits are not released
1104 * back to the MDS mid-read.
1105 *
1106 * Hmm, the sync read case isn't actually async... should it be?
1107 */
1108static ssize_t ceph_read_iter(struct kiocb *iocb, struct iov_iter *to)
1109{
1110 struct file *filp = iocb->ki_filp;
1111 struct ceph_file_info *fi = filp->private_data;
1112 size_t len = iov_iter_count(to);
1113 struct inode *inode = file_inode(filp);
1114 struct ceph_inode_info *ci = ceph_inode(inode);
1115 struct page *pinned_page = NULL;
1116 ssize_t ret;
1117 int want, got = 0;
1118 int retry_op = 0, read = 0;
1119
1120again:
1121 dout("aio_read %p %llx.%llx %llu~%u trying to get caps on %p\n",
1122 inode, ceph_vinop(inode), iocb->ki_pos, (unsigned)len, inode);
1123
1124 if (fi->fmode & CEPH_FILE_MODE_LAZY)
1125 want = CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO;
1126 else
1127 want = CEPH_CAP_FILE_CACHE;
1128 ret = ceph_get_caps(ci, CEPH_CAP_FILE_RD, want, -1, &got, &pinned_page);
1129 if (ret < 0)
1130 return ret;
1131
1132 if ((got & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) == 0 ||
1133 (iocb->ki_flags & IOCB_DIRECT) ||
1134 (fi->flags & CEPH_F_SYNC)) {
1135
1136 dout("aio_sync_read %p %llx.%llx %llu~%u got cap refs on %s\n",
1137 inode, ceph_vinop(inode), iocb->ki_pos, (unsigned)len,
1138 ceph_cap_string(got));
1139
1140 if (ci->i_inline_version == CEPH_INLINE_NONE) {
1141 if (!retry_op && (iocb->ki_flags & IOCB_DIRECT)) {
1142 ret = ceph_direct_read_write(iocb, to,
1143 NULL, NULL);
1144 if (ret >= 0 && ret < len)
1145 retry_op = CHECK_EOF;
1146 } else {
1147 ret = ceph_sync_read(iocb, to, &retry_op);
1148 }
1149 } else {
1150 retry_op = READ_INLINE;
1151 }
1152 } else {
1153 dout("aio_read %p %llx.%llx %llu~%u got cap refs on %s\n",
1154 inode, ceph_vinop(inode), iocb->ki_pos, (unsigned)len,
1155 ceph_cap_string(got));
1156
1157 ret = generic_file_read_iter(iocb, to);
1158 }
1159 dout("aio_read %p %llx.%llx dropping cap refs on %s = %d\n",
1160 inode, ceph_vinop(inode), ceph_cap_string(got), (int)ret);
1161 if (pinned_page) {
1162 put_page(pinned_page);
1163 pinned_page = NULL;
1164 }
1165 ceph_put_cap_refs(ci, got);
1166 if (retry_op > HAVE_RETRIED && ret >= 0) {
1167 int statret;
1168 struct page *page = NULL;
1169 loff_t i_size;
1170 if (retry_op == READ_INLINE) {
1171 page = __page_cache_alloc(GFP_KERNEL);
1172 if (!page)
1173 return -ENOMEM;
1174 }
1175
1176 statret = __ceph_do_getattr(inode, page,
1177 CEPH_STAT_CAP_INLINE_DATA, !!page);
1178 if (statret < 0) {
1179 __free_page(page);
1180 if (statret == -ENODATA) {
1181 BUG_ON(retry_op != READ_INLINE);
1182 goto again;
1183 }
1184 return statret;
1185 }
1186
1187 i_size = i_size_read(inode);
1188 if (retry_op == READ_INLINE) {
1189 BUG_ON(ret > 0 || read > 0);
1190 if (iocb->ki_pos < i_size &&
1191 iocb->ki_pos < PAGE_SIZE) {
1192 loff_t end = min_t(loff_t, i_size,
1193 iocb->ki_pos + len);
1194 end = min_t(loff_t, end, PAGE_SIZE);
1195 if (statret < end)
1196 zero_user_segment(page, statret, end);
1197 ret = copy_page_to_iter(page,
1198 iocb->ki_pos & ~PAGE_MASK,
1199 end - iocb->ki_pos, to);
1200 iocb->ki_pos += ret;
1201 read += ret;
1202 }
1203 if (iocb->ki_pos < i_size && read < len) {
1204 size_t zlen = min_t(size_t, len - read,
1205 i_size - iocb->ki_pos);
1206 ret = iov_iter_zero(zlen, to);
1207 iocb->ki_pos += ret;
1208 read += ret;
1209 }
1210 __free_pages(page, 0);
1211 return read;
1212 }
1213
1214 /* hit EOF or hole? */
1215 if (retry_op == CHECK_EOF && iocb->ki_pos < i_size &&
1216 ret < len) {
1217 dout("sync_read hit hole, ppos %lld < size %lld"
1218 ", reading more\n", iocb->ki_pos, i_size);
1219
1220 read += ret;
1221 len -= ret;
1222 retry_op = HAVE_RETRIED;
1223 goto again;
1224 }
1225 }
1226
1227 if (ret >= 0)
1228 ret += read;
1229
1230 return ret;
1231}
1232
1233/*
1234 * Take cap references to avoid releasing caps to MDS mid-write.
1235 *
1236 * If we are synchronous, and write with an old snap context, the OSD
1237 * may return EOLDSNAPC. In that case, retry the write.. _after_
1238 * dropping our cap refs and allowing the pending snap to logically
1239 * complete _before_ this write occurs.
1240 *
1241 * If we are near ENOSPC, write synchronously.
1242 */
1243static ssize_t ceph_write_iter(struct kiocb *iocb, struct iov_iter *from)
1244{
1245 struct file *file = iocb->ki_filp;
1246 struct ceph_file_info *fi = file->private_data;
1247 struct inode *inode = file_inode(file);
1248 struct ceph_inode_info *ci = ceph_inode(inode);
1249 struct ceph_osd_client *osdc =
1250 &ceph_sb_to_client(inode->i_sb)->client->osdc;
1251 struct ceph_cap_flush *prealloc_cf;
1252 ssize_t count, written = 0;
1253 int err, want, got;
1254 loff_t pos;
1255
1256 if (ceph_snap(inode) != CEPH_NOSNAP)
1257 return -EROFS;
1258
1259 prealloc_cf = ceph_alloc_cap_flush();
1260 if (!prealloc_cf)
1261 return -ENOMEM;
1262
1263 inode_lock(inode);
1264
1265 /* We can write back this queue in page reclaim */
1266 current->backing_dev_info = inode_to_bdi(inode);
1267
1268 if (iocb->ki_flags & IOCB_APPEND) {
1269 err = ceph_do_getattr(inode, CEPH_STAT_CAP_SIZE, false);
1270 if (err < 0)
1271 goto out;
1272 }
1273
1274 err = generic_write_checks(iocb, from);
1275 if (err <= 0)
1276 goto out;
1277
1278 pos = iocb->ki_pos;
1279 count = iov_iter_count(from);
1280 err = file_remove_privs(file);
1281 if (err)
1282 goto out;
1283
1284 err = file_update_time(file);
1285 if (err)
1286 goto out;
1287
1288 if (ci->i_inline_version != CEPH_INLINE_NONE) {
1289 err = ceph_uninline_data(file, NULL);
1290 if (err < 0)
1291 goto out;
1292 }
1293
1294retry_snap:
1295 if (ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_FULL)) {
1296 err = -ENOSPC;
1297 goto out;
1298 }
1299
1300 dout("aio_write %p %llx.%llx %llu~%zd getting caps. i_size %llu\n",
1301 inode, ceph_vinop(inode), pos, count, i_size_read(inode));
1302 if (fi->fmode & CEPH_FILE_MODE_LAZY)
1303 want = CEPH_CAP_FILE_BUFFER | CEPH_CAP_FILE_LAZYIO;
1304 else
1305 want = CEPH_CAP_FILE_BUFFER;
1306 got = 0;
1307 err = ceph_get_caps(ci, CEPH_CAP_FILE_WR, want, pos + count,
1308 &got, NULL);
1309 if (err < 0)
1310 goto out;
1311
1312 dout("aio_write %p %llx.%llx %llu~%zd got cap refs on %s\n",
1313 inode, ceph_vinop(inode), pos, count, ceph_cap_string(got));
1314
1315 if ((got & (CEPH_CAP_FILE_BUFFER|CEPH_CAP_FILE_LAZYIO)) == 0 ||
1316 (iocb->ki_flags & IOCB_DIRECT) || (fi->flags & CEPH_F_SYNC)) {
1317 struct ceph_snap_context *snapc;
1318 struct iov_iter data;
1319 inode_unlock(inode);
1320
1321 spin_lock(&ci->i_ceph_lock);
1322 if (__ceph_have_pending_cap_snap(ci)) {
1323 struct ceph_cap_snap *capsnap =
1324 list_last_entry(&ci->i_cap_snaps,
1325 struct ceph_cap_snap,
1326 ci_item);
1327 snapc = ceph_get_snap_context(capsnap->context);
1328 } else {
1329 BUG_ON(!ci->i_head_snapc);
1330 snapc = ceph_get_snap_context(ci->i_head_snapc);
1331 }
1332 spin_unlock(&ci->i_ceph_lock);
1333
1334 /* we might need to revert back to that point */
1335 data = *from;
1336 if (iocb->ki_flags & IOCB_DIRECT)
1337 written = ceph_direct_read_write(iocb, &data, snapc,
1338 &prealloc_cf);
1339 else
1340 written = ceph_sync_write(iocb, &data, pos, snapc);
1341 if (written == -EOLDSNAPC) {
1342 dout("aio_write %p %llx.%llx %llu~%u"
1343 "got EOLDSNAPC, retrying\n",
1344 inode, ceph_vinop(inode),
1345 pos, (unsigned)count);
1346 inode_lock(inode);
1347 goto retry_snap;
1348 }
1349 if (written > 0)
1350 iov_iter_advance(from, written);
1351 ceph_put_snap_context(snapc);
1352 } else {
1353 loff_t old_size = i_size_read(inode);
1354 /*
1355 * No need to acquire the i_truncate_mutex. Because
1356 * the MDS revokes Fwb caps before sending truncate
1357 * message to us. We can't get Fwb cap while there
1358 * are pending vmtruncate. So write and vmtruncate
1359 * can not run at the same time
1360 */
1361 written = generic_perform_write(file, from, pos);
1362 if (likely(written >= 0))
1363 iocb->ki_pos = pos + written;
1364 if (i_size_read(inode) > old_size)
1365 ceph_fscache_update_objectsize(inode);
1366 inode_unlock(inode);
1367 }
1368
1369 if (written >= 0) {
1370 int dirty;
1371 spin_lock(&ci->i_ceph_lock);
1372 ci->i_inline_version = CEPH_INLINE_NONE;
1373 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR,
1374 &prealloc_cf);
1375 spin_unlock(&ci->i_ceph_lock);
1376 if (dirty)
1377 __mark_inode_dirty(inode, dirty);
1378 }
1379
1380 dout("aio_write %p %llx.%llx %llu~%u dropping cap refs on %s\n",
1381 inode, ceph_vinop(inode), pos, (unsigned)count,
1382 ceph_cap_string(got));
1383 ceph_put_cap_refs(ci, got);
1384
1385 if (written >= 0 &&
1386 ((file->f_flags & O_SYNC) || IS_SYNC(file->f_mapping->host) ||
1387 ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_NEARFULL))) {
1388 err = vfs_fsync_range(file, pos, pos + written - 1, 1);
1389 if (err < 0)
1390 written = err;
1391 }
1392
1393 goto out_unlocked;
1394
1395out:
1396 inode_unlock(inode);
1397out_unlocked:
1398 ceph_free_cap_flush(prealloc_cf);
1399 current->backing_dev_info = NULL;
1400 return written ? written : err;
1401}
1402
1403/*
1404 * llseek. be sure to verify file size on SEEK_END.
1405 */
1406static loff_t ceph_llseek(struct file *file, loff_t offset, int whence)
1407{
1408 struct inode *inode = file->f_mapping->host;
1409 loff_t i_size;
1410 int ret;
1411
1412 inode_lock(inode);
1413
1414 if (whence == SEEK_END || whence == SEEK_DATA || whence == SEEK_HOLE) {
1415 ret = ceph_do_getattr(inode, CEPH_STAT_CAP_SIZE, false);
1416 if (ret < 0) {
1417 offset = ret;
1418 goto out;
1419 }
1420 }
1421
1422 i_size = i_size_read(inode);
1423 switch (whence) {
1424 case SEEK_END:
1425 offset += i_size;
1426 break;
1427 case SEEK_CUR:
1428 /*
1429 * Here we special-case the lseek(fd, 0, SEEK_CUR)
1430 * position-querying operation. Avoid rewriting the "same"
1431 * f_pos value back to the file because a concurrent read(),
1432 * write() or lseek() might have altered it
1433 */
1434 if (offset == 0) {
1435 offset = file->f_pos;
1436 goto out;
1437 }
1438 offset += file->f_pos;
1439 break;
1440 case SEEK_DATA:
1441 if (offset >= i_size) {
1442 ret = -ENXIO;
1443 goto out;
1444 }
1445 break;
1446 case SEEK_HOLE:
1447 if (offset >= i_size) {
1448 ret = -ENXIO;
1449 goto out;
1450 }
1451 offset = i_size;
1452 break;
1453 }
1454
1455 offset = vfs_setpos(file, offset, inode->i_sb->s_maxbytes);
1456
1457out:
1458 inode_unlock(inode);
1459 return offset;
1460}
1461
1462static inline void ceph_zero_partial_page(
1463 struct inode *inode, loff_t offset, unsigned size)
1464{
1465 struct page *page;
1466 pgoff_t index = offset >> PAGE_SHIFT;
1467
1468 page = find_lock_page(inode->i_mapping, index);
1469 if (page) {
1470 wait_on_page_writeback(page);
1471 zero_user(page, offset & (PAGE_SIZE - 1), size);
1472 unlock_page(page);
1473 put_page(page);
1474 }
1475}
1476
1477static void ceph_zero_pagecache_range(struct inode *inode, loff_t offset,
1478 loff_t length)
1479{
1480 loff_t nearly = round_up(offset, PAGE_SIZE);
1481 if (offset < nearly) {
1482 loff_t size = nearly - offset;
1483 if (length < size)
1484 size = length;
1485 ceph_zero_partial_page(inode, offset, size);
1486 offset += size;
1487 length -= size;
1488 }
1489 if (length >= PAGE_SIZE) {
1490 loff_t size = round_down(length, PAGE_SIZE);
1491 truncate_pagecache_range(inode, offset, offset + size - 1);
1492 offset += size;
1493 length -= size;
1494 }
1495 if (length)
1496 ceph_zero_partial_page(inode, offset, length);
1497}
1498
1499static int ceph_zero_partial_object(struct inode *inode,
1500 loff_t offset, loff_t *length)
1501{
1502 struct ceph_inode_info *ci = ceph_inode(inode);
1503 struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1504 struct ceph_osd_request *req;
1505 int ret = 0;
1506 loff_t zero = 0;
1507 int op;
1508
1509 if (!length) {
1510 op = offset ? CEPH_OSD_OP_DELETE : CEPH_OSD_OP_TRUNCATE;
1511 length = &zero;
1512 } else {
1513 op = CEPH_OSD_OP_ZERO;
1514 }
1515
1516 req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1517 ceph_vino(inode),
1518 offset, length,
1519 0, 1, op,
1520 CEPH_OSD_FLAG_WRITE |
1521 CEPH_OSD_FLAG_ONDISK,
1522 NULL, 0, 0, false);
1523 if (IS_ERR(req)) {
1524 ret = PTR_ERR(req);
1525 goto out;
1526 }
1527
1528 ceph_osdc_build_request(req, offset, NULL, ceph_vino(inode).snap,
1529 &inode->i_mtime);
1530
1531 ret = ceph_osdc_start_request(&fsc->client->osdc, req, false);
1532 if (!ret) {
1533 ret = ceph_osdc_wait_request(&fsc->client->osdc, req);
1534 if (ret == -ENOENT)
1535 ret = 0;
1536 }
1537 ceph_osdc_put_request(req);
1538
1539out:
1540 return ret;
1541}
1542
1543static int ceph_zero_objects(struct inode *inode, loff_t offset, loff_t length)
1544{
1545 int ret = 0;
1546 struct ceph_inode_info *ci = ceph_inode(inode);
1547 s32 stripe_unit = ceph_file_layout_su(ci->i_layout);
1548 s32 stripe_count = ceph_file_layout_stripe_count(ci->i_layout);
1549 s32 object_size = ceph_file_layout_object_size(ci->i_layout);
1550 u64 object_set_size = object_size * stripe_count;
1551 u64 nearly, t;
1552
1553 /* round offset up to next period boundary */
1554 nearly = offset + object_set_size - 1;
1555 t = nearly;
1556 nearly -= do_div(t, object_set_size);
1557
1558 while (length && offset < nearly) {
1559 loff_t size = length;
1560 ret = ceph_zero_partial_object(inode, offset, &size);
1561 if (ret < 0)
1562 return ret;
1563 offset += size;
1564 length -= size;
1565 }
1566 while (length >= object_set_size) {
1567 int i;
1568 loff_t pos = offset;
1569 for (i = 0; i < stripe_count; ++i) {
1570 ret = ceph_zero_partial_object(inode, pos, NULL);
1571 if (ret < 0)
1572 return ret;
1573 pos += stripe_unit;
1574 }
1575 offset += object_set_size;
1576 length -= object_set_size;
1577 }
1578 while (length) {
1579 loff_t size = length;
1580 ret = ceph_zero_partial_object(inode, offset, &size);
1581 if (ret < 0)
1582 return ret;
1583 offset += size;
1584 length -= size;
1585 }
1586 return ret;
1587}
1588
1589static long ceph_fallocate(struct file *file, int mode,
1590 loff_t offset, loff_t length)
1591{
1592 struct ceph_file_info *fi = file->private_data;
1593 struct inode *inode = file_inode(file);
1594 struct ceph_inode_info *ci = ceph_inode(inode);
1595 struct ceph_osd_client *osdc =
1596 &ceph_inode_to_client(inode)->client->osdc;
1597 struct ceph_cap_flush *prealloc_cf;
1598 int want, got = 0;
1599 int dirty;
1600 int ret = 0;
1601 loff_t endoff = 0;
1602 loff_t size;
1603
1604 if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
1605 return -EOPNOTSUPP;
1606
1607 if (!S_ISREG(inode->i_mode))
1608 return -EOPNOTSUPP;
1609
1610 prealloc_cf = ceph_alloc_cap_flush();
1611 if (!prealloc_cf)
1612 return -ENOMEM;
1613
1614 inode_lock(inode);
1615
1616 if (ceph_snap(inode) != CEPH_NOSNAP) {
1617 ret = -EROFS;
1618 goto unlock;
1619 }
1620
1621 if (ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_FULL) &&
1622 !(mode & FALLOC_FL_PUNCH_HOLE)) {
1623 ret = -ENOSPC;
1624 goto unlock;
1625 }
1626
1627 if (ci->i_inline_version != CEPH_INLINE_NONE) {
1628 ret = ceph_uninline_data(file, NULL);
1629 if (ret < 0)
1630 goto unlock;
1631 }
1632
1633 size = i_size_read(inode);
1634 if (!(mode & FALLOC_FL_KEEP_SIZE))
1635 endoff = offset + length;
1636
1637 if (fi->fmode & CEPH_FILE_MODE_LAZY)
1638 want = CEPH_CAP_FILE_BUFFER | CEPH_CAP_FILE_LAZYIO;
1639 else
1640 want = CEPH_CAP_FILE_BUFFER;
1641
1642 ret = ceph_get_caps(ci, CEPH_CAP_FILE_WR, want, endoff, &got, NULL);
1643 if (ret < 0)
1644 goto unlock;
1645
1646 if (mode & FALLOC_FL_PUNCH_HOLE) {
1647 if (offset < size)
1648 ceph_zero_pagecache_range(inode, offset, length);
1649 ret = ceph_zero_objects(inode, offset, length);
1650 } else if (endoff > size) {
1651 truncate_pagecache_range(inode, size, -1);
1652 if (ceph_inode_set_size(inode, endoff))
1653 ceph_check_caps(ceph_inode(inode),
1654 CHECK_CAPS_AUTHONLY, NULL);
1655 }
1656
1657 if (!ret) {
1658 spin_lock(&ci->i_ceph_lock);
1659 ci->i_inline_version = CEPH_INLINE_NONE;
1660 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR,
1661 &prealloc_cf);
1662 spin_unlock(&ci->i_ceph_lock);
1663 if (dirty)
1664 __mark_inode_dirty(inode, dirty);
1665 }
1666
1667 ceph_put_cap_refs(ci, got);
1668unlock:
1669 inode_unlock(inode);
1670 ceph_free_cap_flush(prealloc_cf);
1671 return ret;
1672}
1673
1674const struct file_operations ceph_file_fops = {
1675 .open = ceph_open,
1676 .release = ceph_release,
1677 .llseek = ceph_llseek,
1678 .read_iter = ceph_read_iter,
1679 .write_iter = ceph_write_iter,
1680 .mmap = ceph_mmap,
1681 .fsync = ceph_fsync,
1682 .lock = ceph_lock,
1683 .flock = ceph_flock,
1684 .splice_read = generic_file_splice_read,
1685 .splice_write = iter_file_splice_write,
1686 .unlocked_ioctl = ceph_ioctl,
1687 .compat_ioctl = ceph_ioctl,
1688 .fallocate = ceph_fallocate,
1689};
1690
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 */
39static struct ceph_mds_request *
40prepare_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);
58out:
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 */
66static 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 */
118int 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 = NULL;
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 if (flags & (O_CREAT|O_TRUNC))
198 parent_inode = ceph_get_dentry_parent_inode(file->f_dentry);
199 err = ceph_mdsc_do_request(mdsc, parent_inode, req);
200 iput(parent_inode);
201 if (!err)
202 err = ceph_init_file(inode, file, req->r_fmode);
203 ceph_mdsc_put_request(req);
204 dout("open result=%d on %llx.%llx\n", err, ceph_vinop(inode));
205out:
206 return err;
207}
208
209
210/*
211 * Do a lookup + open with a single request.
212 *
213 * If this succeeds, but some subsequent check in the vfs
214 * may_open() fails, the struct *file gets cleaned up (i.e.
215 * ceph_release gets called). So fear not!
216 */
217/*
218 * flags
219 * path_lookup_open -> LOOKUP_OPEN
220 * path_lookup_create -> LOOKUP_OPEN|LOOKUP_CREATE
221 */
222struct dentry *ceph_lookup_open(struct inode *dir, struct dentry *dentry,
223 struct nameidata *nd, int mode,
224 int locked_dir)
225{
226 struct ceph_fs_client *fsc = ceph_sb_to_client(dir->i_sb);
227 struct ceph_mds_client *mdsc = fsc->mdsc;
228 struct file *file;
229 struct ceph_mds_request *req;
230 struct dentry *ret;
231 int err;
232 int flags = nd->intent.open.flags;
233
234 dout("ceph_lookup_open dentry %p '%.*s' flags %d mode 0%o\n",
235 dentry, dentry->d_name.len, dentry->d_name.name, flags, mode);
236
237 /* do the open */
238 req = prepare_open_request(dir->i_sb, flags, mode);
239 if (IS_ERR(req))
240 return ERR_CAST(req);
241 req->r_dentry = dget(dentry);
242 req->r_num_caps = 2;
243 if (flags & O_CREAT) {
244 req->r_dentry_drop = CEPH_CAP_FILE_SHARED;
245 req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
246 }
247 req->r_locked_dir = dir; /* caller holds dir->i_mutex */
248 err = ceph_mdsc_do_request(mdsc,
249 (flags & (O_CREAT|O_TRUNC)) ? dir : NULL,
250 req);
251 err = ceph_handle_snapdir(req, dentry, err);
252 if (err)
253 goto out;
254 if ((flags & O_CREAT) && !req->r_reply_info.head->is_dentry)
255 err = ceph_handle_notrace_create(dir, dentry);
256 if (err)
257 goto out;
258 file = lookup_instantiate_filp(nd, req->r_dentry, ceph_open);
259 if (IS_ERR(file))
260 err = PTR_ERR(file);
261out:
262 ret = ceph_finish_lookup(req, dentry, err);
263 ceph_mdsc_put_request(req);
264 dout("ceph_lookup_open result=%p\n", ret);
265 return ret;
266}
267
268int ceph_release(struct inode *inode, struct file *file)
269{
270 struct ceph_inode_info *ci = ceph_inode(inode);
271 struct ceph_file_info *cf = file->private_data;
272
273 dout("release inode %p file %p\n", inode, file);
274 ceph_put_fmode(ci, cf->fmode);
275 if (cf->last_readdir)
276 ceph_mdsc_put_request(cf->last_readdir);
277 kfree(cf->last_name);
278 kfree(cf->dir_info);
279 dput(cf->dentry);
280 kmem_cache_free(ceph_file_cachep, cf);
281
282 /* wake up anyone waiting for caps on this inode */
283 wake_up_all(&ci->i_cap_wq);
284 return 0;
285}
286
287/*
288 * Read a range of bytes striped over one or more objects. Iterate over
289 * objects we stripe over. (That's not atomic, but good enough for now.)
290 *
291 * If we get a short result from the OSD, check against i_size; we need to
292 * only return a short read to the caller if we hit EOF.
293 */
294static int striped_read(struct inode *inode,
295 u64 off, u64 len,
296 struct page **pages, int num_pages,
297 int *checkeof, bool o_direct,
298 unsigned long buf_align)
299{
300 struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
301 struct ceph_inode_info *ci = ceph_inode(inode);
302 u64 pos, this_len;
303 int io_align, page_align;
304 int left, pages_left;
305 int read;
306 struct page **page_pos;
307 int ret;
308 bool hit_stripe, was_short;
309
310 /*
311 * we may need to do multiple reads. not atomic, unfortunately.
312 */
313 pos = off;
314 left = len;
315 page_pos = pages;
316 pages_left = num_pages;
317 read = 0;
318 io_align = off & ~PAGE_MASK;
319
320more:
321 if (o_direct)
322 page_align = (pos - io_align + buf_align) & ~PAGE_MASK;
323 else
324 page_align = pos & ~PAGE_MASK;
325 this_len = left;
326 ret = ceph_osdc_readpages(&fsc->client->osdc, ceph_vino(inode),
327 &ci->i_layout, pos, &this_len,
328 ci->i_truncate_seq,
329 ci->i_truncate_size,
330 page_pos, pages_left, page_align);
331 if (ret == -ENOENT)
332 ret = 0;
333 hit_stripe = this_len < left;
334 was_short = ret >= 0 && ret < this_len;
335 dout("striped_read %llu~%u (read %u) got %d%s%s\n", pos, left, read,
336 ret, hit_stripe ? " HITSTRIPE" : "", was_short ? " SHORT" : "");
337
338 if (ret > 0) {
339 int didpages = (page_align + ret) >> PAGE_CACHE_SHIFT;
340
341 if (read < pos - off) {
342 dout(" zero gap %llu to %llu\n", off + read, pos);
343 ceph_zero_page_vector_range(page_align + read,
344 pos - off - read, pages);
345 }
346 pos += ret;
347 read = pos - off;
348 left -= ret;
349 page_pos += didpages;
350 pages_left -= didpages;
351
352 /* hit stripe? */
353 if (left && hit_stripe)
354 goto more;
355 }
356
357 if (was_short) {
358 /* did we bounce off eof? */
359 if (pos + left > inode->i_size)
360 *checkeof = 1;
361
362 /* zero trailing bytes (inside i_size) */
363 if (left > 0 && pos < inode->i_size) {
364 if (pos + left > inode->i_size)
365 left = inode->i_size - pos;
366
367 dout("zero tail %d\n", left);
368 ceph_zero_page_vector_range(page_align + read, left,
369 pages);
370 read += left;
371 }
372 }
373
374 if (ret >= 0)
375 ret = read;
376 dout("striped_read returns %d\n", ret);
377 return ret;
378}
379
380/*
381 * Completely synchronous read and write methods. Direct from __user
382 * buffer to osd, or directly to user pages (if O_DIRECT).
383 *
384 * If the read spans object boundary, just do multiple reads.
385 */
386static ssize_t ceph_sync_read(struct file *file, char __user *data,
387 unsigned len, loff_t *poff, int *checkeof)
388{
389 struct inode *inode = file->f_dentry->d_inode;
390 struct page **pages;
391 u64 off = *poff;
392 int num_pages, ret;
393
394 dout("sync_read on file %p %llu~%u %s\n", file, off, len,
395 (file->f_flags & O_DIRECT) ? "O_DIRECT" : "");
396
397 if (file->f_flags & O_DIRECT) {
398 num_pages = calc_pages_for((unsigned long)data, len);
399 pages = ceph_get_direct_page_vector(data, num_pages, true);
400 } else {
401 num_pages = calc_pages_for(off, len);
402 pages = ceph_alloc_page_vector(num_pages, GFP_NOFS);
403 }
404 if (IS_ERR(pages))
405 return PTR_ERR(pages);
406
407 /*
408 * flush any page cache pages in this range. this
409 * will make concurrent normal and sync io slow,
410 * but it will at least behave sensibly when they are
411 * in sequence.
412 */
413 ret = filemap_write_and_wait(inode->i_mapping);
414 if (ret < 0)
415 goto done;
416
417 ret = striped_read(inode, off, len, pages, num_pages, checkeof,
418 file->f_flags & O_DIRECT,
419 (unsigned long)data & ~PAGE_MASK);
420
421 if (ret >= 0 && (file->f_flags & O_DIRECT) == 0)
422 ret = ceph_copy_page_vector_to_user(pages, data, off, ret);
423 if (ret >= 0)
424 *poff = off + ret;
425
426done:
427 if (file->f_flags & O_DIRECT)
428 ceph_put_page_vector(pages, num_pages, true);
429 else
430 ceph_release_page_vector(pages, num_pages);
431 dout("sync_read result %d\n", ret);
432 return ret;
433}
434
435/*
436 * Write commit callback, called if we requested both an ACK and
437 * ONDISK commit reply from the OSD.
438 */
439static void sync_write_commit(struct ceph_osd_request *req,
440 struct ceph_msg *msg)
441{
442 struct ceph_inode_info *ci = ceph_inode(req->r_inode);
443
444 dout("sync_write_commit %p tid %llu\n", req, req->r_tid);
445 spin_lock(&ci->i_unsafe_lock);
446 list_del_init(&req->r_unsafe_item);
447 spin_unlock(&ci->i_unsafe_lock);
448 ceph_put_cap_refs(ci, CEPH_CAP_FILE_WR);
449}
450
451/*
452 * Synchronous write, straight from __user pointer or user pages (if
453 * O_DIRECT).
454 *
455 * If write spans object boundary, just do multiple writes. (For a
456 * correct atomic write, we should e.g. take write locks on all
457 * objects, rollback on failure, etc.)
458 */
459static ssize_t ceph_sync_write(struct file *file, const char __user *data,
460 size_t left, loff_t *offset)
461{
462 struct inode *inode = file->f_dentry->d_inode;
463 struct ceph_inode_info *ci = ceph_inode(inode);
464 struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
465 struct ceph_osd_request *req;
466 struct page **pages;
467 int num_pages;
468 long long unsigned pos;
469 u64 len;
470 int written = 0;
471 int flags;
472 int do_sync = 0;
473 int check_caps = 0;
474 int page_align, io_align;
475 unsigned long buf_align;
476 int ret;
477 struct timespec mtime = CURRENT_TIME;
478
479 if (ceph_snap(file->f_dentry->d_inode) != CEPH_NOSNAP)
480 return -EROFS;
481
482 dout("sync_write on file %p %lld~%u %s\n", file, *offset,
483 (unsigned)left, (file->f_flags & O_DIRECT) ? "O_DIRECT" : "");
484
485 if (file->f_flags & O_APPEND)
486 pos = i_size_read(inode);
487 else
488 pos = *offset;
489
490 ret = filemap_write_and_wait_range(inode->i_mapping, pos, pos + left);
491 if (ret < 0)
492 return ret;
493
494 ret = invalidate_inode_pages2_range(inode->i_mapping,
495 pos >> PAGE_CACHE_SHIFT,
496 (pos + left) >> PAGE_CACHE_SHIFT);
497 if (ret < 0)
498 dout("invalidate_inode_pages2_range returned %d\n", ret);
499
500 flags = CEPH_OSD_FLAG_ORDERSNAP |
501 CEPH_OSD_FLAG_ONDISK |
502 CEPH_OSD_FLAG_WRITE;
503 if ((file->f_flags & (O_SYNC|O_DIRECT)) == 0)
504 flags |= CEPH_OSD_FLAG_ACK;
505 else
506 do_sync = 1;
507
508 /*
509 * we may need to do multiple writes here if we span an object
510 * boundary. this isn't atomic, unfortunately. :(
511 */
512more:
513 io_align = pos & ~PAGE_MASK;
514 buf_align = (unsigned long)data & ~PAGE_MASK;
515 len = left;
516 if (file->f_flags & O_DIRECT) {
517 /* write from beginning of first page, regardless of
518 io alignment */
519 page_align = (pos - io_align + buf_align) & ~PAGE_MASK;
520 num_pages = calc_pages_for((unsigned long)data, len);
521 } else {
522 page_align = pos & ~PAGE_MASK;
523 num_pages = calc_pages_for(pos, len);
524 }
525 req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
526 ceph_vino(inode), pos, &len,
527 CEPH_OSD_OP_WRITE, flags,
528 ci->i_snap_realm->cached_context,
529 do_sync,
530 ci->i_truncate_seq, ci->i_truncate_size,
531 &mtime, false, 2, page_align);
532 if (!req)
533 return -ENOMEM;
534
535 if (file->f_flags & O_DIRECT) {
536 pages = ceph_get_direct_page_vector(data, num_pages, false);
537 if (IS_ERR(pages)) {
538 ret = PTR_ERR(pages);
539 goto out;
540 }
541
542 /*
543 * throw out any page cache pages in this range. this
544 * may block.
545 */
546 truncate_inode_pages_range(inode->i_mapping, pos,
547 (pos+len) | (PAGE_CACHE_SIZE-1));
548 } else {
549 pages = ceph_alloc_page_vector(num_pages, GFP_NOFS);
550 if (IS_ERR(pages)) {
551 ret = PTR_ERR(pages);
552 goto out;
553 }
554 ret = ceph_copy_user_to_page_vector(pages, data, pos, len);
555 if (ret < 0) {
556 ceph_release_page_vector(pages, num_pages);
557 goto out;
558 }
559
560 if ((file->f_flags & O_SYNC) == 0) {
561 /* get a second commit callback */
562 req->r_safe_callback = sync_write_commit;
563 req->r_own_pages = 1;
564 }
565 }
566 req->r_pages = pages;
567 req->r_num_pages = num_pages;
568 req->r_inode = inode;
569
570 ret = ceph_osdc_start_request(&fsc->client->osdc, req, false);
571 if (!ret) {
572 if (req->r_safe_callback) {
573 /*
574 * Add to inode unsafe list only after we
575 * start_request so that a tid has been assigned.
576 */
577 spin_lock(&ci->i_unsafe_lock);
578 list_add_tail(&req->r_unsafe_item,
579 &ci->i_unsafe_writes);
580 spin_unlock(&ci->i_unsafe_lock);
581 ceph_get_cap_refs(ci, CEPH_CAP_FILE_WR);
582 }
583
584 ret = ceph_osdc_wait_request(&fsc->client->osdc, req);
585 if (ret < 0 && req->r_safe_callback) {
586 spin_lock(&ci->i_unsafe_lock);
587 list_del_init(&req->r_unsafe_item);
588 spin_unlock(&ci->i_unsafe_lock);
589 ceph_put_cap_refs(ci, CEPH_CAP_FILE_WR);
590 }
591 }
592
593 if (file->f_flags & O_DIRECT)
594 ceph_put_page_vector(pages, num_pages, false);
595 else if (file->f_flags & O_SYNC)
596 ceph_release_page_vector(pages, num_pages);
597
598out:
599 ceph_osdc_put_request(req);
600 if (ret == 0) {
601 pos += len;
602 written += len;
603 left -= len;
604 data += written;
605 if (left)
606 goto more;
607
608 ret = written;
609 *offset = pos;
610 if (pos > i_size_read(inode))
611 check_caps = ceph_inode_set_size(inode, pos);
612 if (check_caps)
613 ceph_check_caps(ceph_inode(inode), CHECK_CAPS_AUTHONLY,
614 NULL);
615 }
616 return ret;
617}
618
619/*
620 * Wrap generic_file_aio_read with checks for cap bits on the inode.
621 * Atomically grab references, so that those bits are not released
622 * back to the MDS mid-read.
623 *
624 * Hmm, the sync read case isn't actually async... should it be?
625 */
626static ssize_t ceph_aio_read(struct kiocb *iocb, const struct iovec *iov,
627 unsigned long nr_segs, loff_t pos)
628{
629 struct file *filp = iocb->ki_filp;
630 struct ceph_file_info *fi = filp->private_data;
631 loff_t *ppos = &iocb->ki_pos;
632 size_t len = iov->iov_len;
633 struct inode *inode = filp->f_dentry->d_inode;
634 struct ceph_inode_info *ci = ceph_inode(inode);
635 void __user *base = iov->iov_base;
636 ssize_t ret;
637 int want, got = 0;
638 int checkeof = 0, read = 0;
639
640 dout("aio_read %p %llx.%llx %llu~%u trying to get caps on %p\n",
641 inode, ceph_vinop(inode), pos, (unsigned)len, inode);
642again:
643 __ceph_do_pending_vmtruncate(inode);
644 if (fi->fmode & CEPH_FILE_MODE_LAZY)
645 want = CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO;
646 else
647 want = CEPH_CAP_FILE_CACHE;
648 ret = ceph_get_caps(ci, CEPH_CAP_FILE_RD, want, &got, -1);
649 if (ret < 0)
650 goto out;
651 dout("aio_read %p %llx.%llx %llu~%u got cap refs on %s\n",
652 inode, ceph_vinop(inode), pos, (unsigned)len,
653 ceph_cap_string(got));
654
655 if ((got & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) == 0 ||
656 (iocb->ki_filp->f_flags & O_DIRECT) ||
657 (inode->i_sb->s_flags & MS_SYNCHRONOUS) ||
658 (fi->flags & CEPH_F_SYNC))
659 /* hmm, this isn't really async... */
660 ret = ceph_sync_read(filp, base, len, ppos, &checkeof);
661 else
662 ret = generic_file_aio_read(iocb, iov, nr_segs, pos);
663
664out:
665 dout("aio_read %p %llx.%llx dropping cap refs on %s = %d\n",
666 inode, ceph_vinop(inode), ceph_cap_string(got), (int)ret);
667 ceph_put_cap_refs(ci, got);
668
669 if (checkeof && ret >= 0) {
670 int statret = ceph_do_getattr(inode, CEPH_STAT_CAP_SIZE);
671
672 /* hit EOF or hole? */
673 if (statret == 0 && *ppos < inode->i_size) {
674 dout("aio_read sync_read hit hole, ppos %lld < size %lld, reading more\n", *ppos, inode->i_size);
675 read += ret;
676 base += ret;
677 len -= ret;
678 checkeof = 0;
679 goto again;
680 }
681 }
682 if (ret >= 0)
683 ret += read;
684
685 return ret;
686}
687
688/*
689 * Take cap references to avoid releasing caps to MDS mid-write.
690 *
691 * If we are synchronous, and write with an old snap context, the OSD
692 * may return EOLDSNAPC. In that case, retry the write.. _after_
693 * dropping our cap refs and allowing the pending snap to logically
694 * complete _before_ this write occurs.
695 *
696 * If we are near ENOSPC, write synchronously.
697 */
698static ssize_t ceph_aio_write(struct kiocb *iocb, const struct iovec *iov,
699 unsigned long nr_segs, loff_t pos)
700{
701 struct file *file = iocb->ki_filp;
702 struct ceph_file_info *fi = file->private_data;
703 struct inode *inode = file->f_dentry->d_inode;
704 struct ceph_inode_info *ci = ceph_inode(inode);
705 struct ceph_osd_client *osdc =
706 &ceph_sb_to_client(inode->i_sb)->client->osdc;
707 loff_t endoff = pos + iov->iov_len;
708 int want, got = 0;
709 int ret, err;
710
711 if (ceph_snap(inode) != CEPH_NOSNAP)
712 return -EROFS;
713
714retry_snap:
715 if (ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_FULL))
716 return -ENOSPC;
717 __ceph_do_pending_vmtruncate(inode);
718 dout("aio_write %p %llx.%llx %llu~%u getting caps. i_size %llu\n",
719 inode, ceph_vinop(inode), pos, (unsigned)iov->iov_len,
720 inode->i_size);
721 if (fi->fmode & CEPH_FILE_MODE_LAZY)
722 want = CEPH_CAP_FILE_BUFFER | CEPH_CAP_FILE_LAZYIO;
723 else
724 want = CEPH_CAP_FILE_BUFFER;
725 ret = ceph_get_caps(ci, CEPH_CAP_FILE_WR, want, &got, endoff);
726 if (ret < 0)
727 goto out_put;
728
729 dout("aio_write %p %llx.%llx %llu~%u got cap refs on %s\n",
730 inode, ceph_vinop(inode), pos, (unsigned)iov->iov_len,
731 ceph_cap_string(got));
732
733 if ((got & (CEPH_CAP_FILE_BUFFER|CEPH_CAP_FILE_LAZYIO)) == 0 ||
734 (iocb->ki_filp->f_flags & O_DIRECT) ||
735 (inode->i_sb->s_flags & MS_SYNCHRONOUS) ||
736 (fi->flags & CEPH_F_SYNC)) {
737 ret = ceph_sync_write(file, iov->iov_base, iov->iov_len,
738 &iocb->ki_pos);
739 } else {
740 /*
741 * buffered write; drop Fw early to avoid slow
742 * revocation if we get stuck on balance_dirty_pages
743 */
744 int dirty;
745
746 spin_lock(&inode->i_lock);
747 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR);
748 spin_unlock(&inode->i_lock);
749 ceph_put_cap_refs(ci, got);
750
751 ret = generic_file_aio_write(iocb, iov, nr_segs, pos);
752 if ((ret >= 0 || ret == -EIOCBQUEUED) &&
753 ((file->f_flags & O_SYNC) || IS_SYNC(file->f_mapping->host)
754 || ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_NEARFULL))) {
755 err = vfs_fsync_range(file, pos, pos + ret - 1, 1);
756 if (err < 0)
757 ret = err;
758 }
759
760 if (dirty)
761 __mark_inode_dirty(inode, dirty);
762 goto out;
763 }
764
765 if (ret >= 0) {
766 int dirty;
767 spin_lock(&inode->i_lock);
768 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR);
769 spin_unlock(&inode->i_lock);
770 if (dirty)
771 __mark_inode_dirty(inode, dirty);
772 }
773
774out_put:
775 dout("aio_write %p %llx.%llx %llu~%u dropping cap refs on %s\n",
776 inode, ceph_vinop(inode), pos, (unsigned)iov->iov_len,
777 ceph_cap_string(got));
778 ceph_put_cap_refs(ci, got);
779
780out:
781 if (ret == -EOLDSNAPC) {
782 dout("aio_write %p %llx.%llx %llu~%u got EOLDSNAPC, retrying\n",
783 inode, ceph_vinop(inode), pos, (unsigned)iov->iov_len);
784 goto retry_snap;
785 }
786
787 return ret;
788}
789
790/*
791 * llseek. be sure to verify file size on SEEK_END.
792 */
793static loff_t ceph_llseek(struct file *file, loff_t offset, int origin)
794{
795 struct inode *inode = file->f_mapping->host;
796 int ret;
797
798 mutex_lock(&inode->i_mutex);
799 __ceph_do_pending_vmtruncate(inode);
800 if (origin != SEEK_CUR || origin != SEEK_SET) {
801 ret = ceph_do_getattr(inode, CEPH_STAT_CAP_SIZE);
802 if (ret < 0) {
803 offset = ret;
804 goto out;
805 }
806 }
807
808 switch (origin) {
809 case SEEK_END:
810 offset += inode->i_size;
811 break;
812 case SEEK_CUR:
813 /*
814 * Here we special-case the lseek(fd, 0, SEEK_CUR)
815 * position-querying operation. Avoid rewriting the "same"
816 * f_pos value back to the file because a concurrent read(),
817 * write() or lseek() might have altered it
818 */
819 if (offset == 0) {
820 offset = file->f_pos;
821 goto out;
822 }
823 offset += file->f_pos;
824 break;
825 case SEEK_DATA:
826 if (offset >= inode->i_size) {
827 ret = -ENXIO;
828 goto out;
829 }
830 break;
831 case SEEK_HOLE:
832 if (offset >= inode->i_size) {
833 ret = -ENXIO;
834 goto out;
835 }
836 offset = inode->i_size;
837 break;
838 }
839
840 if (offset < 0 || offset > inode->i_sb->s_maxbytes) {
841 offset = -EINVAL;
842 goto out;
843 }
844
845 /* Special lock needed here? */
846 if (offset != file->f_pos) {
847 file->f_pos = offset;
848 file->f_version = 0;
849 }
850
851out:
852 mutex_unlock(&inode->i_mutex);
853 return offset;
854}
855
856const struct file_operations ceph_file_fops = {
857 .open = ceph_open,
858 .release = ceph_release,
859 .llseek = ceph_llseek,
860 .read = do_sync_read,
861 .write = do_sync_write,
862 .aio_read = ceph_aio_read,
863 .aio_write = ceph_aio_write,
864 .mmap = ceph_mmap,
865 .fsync = ceph_fsync,
866 .lock = ceph_lock,
867 .flock = ceph_flock,
868 .splice_read = generic_file_splice_read,
869 .splice_write = generic_file_splice_write,
870 .unlocked_ioctl = ceph_ioctl,
871 .compat_ioctl = ceph_ioctl,
872};
873