blob: 4add3d5da2c1ba3ee650c0d0945ce4c4ca2f68a5 [file] [log] [blame]
Sage Weil124e68e2009-10-06 11:31:08 -07001#include "ceph_debug.h"
2
3#include <linux/sched.h>
Tejun Heo5a0e3ad2010-03-24 17:04:11 +09004#include <linux/slab.h>
Sage Weil124e68e2009-10-06 11:31:08 -07005#include <linux/file.h>
6#include <linux/namei.h>
7#include <linux/writeback.h>
8
9#include "super.h"
10#include "mds_client.h"
11
12/*
13 * Ceph file operations
14 *
15 * Implement basic open/close functionality, and implement
16 * read/write.
17 *
18 * We implement three modes of file I/O:
19 * - buffered uses the generic_file_aio_{read,write} helpers
20 *
21 * - synchronous is used when there is multi-client read/write
22 * sharing, avoids the page cache, and synchronously waits for an
23 * ack from the OSD.
24 *
25 * - direct io takes the variant of the sync path that references
26 * user pages directly.
27 *
28 * fsync() flushes and waits on dirty pages, but just queues metadata
29 * for writeback: since the MDS can recover size and mtime there is no
30 * need to wait for MDS acknowledgement.
31 */
32
33
34/*
35 * Prepare an open request. Preallocate ceph_cap to avoid an
36 * inopportune ENOMEM later.
37 */
38static struct ceph_mds_request *
39prepare_open_request(struct super_block *sb, int flags, int create_mode)
40{
41 struct ceph_client *client = ceph_sb_to_client(sb);
42 struct ceph_mds_client *mdsc = &client->mdsc;
43 struct ceph_mds_request *req;
44 int want_auth = USE_ANY_MDS;
45 int op = (flags & O_CREAT) ? CEPH_MDS_OP_CREATE : CEPH_MDS_OP_OPEN;
46
47 if (flags & (O_WRONLY|O_RDWR|O_CREAT|O_TRUNC))
48 want_auth = USE_AUTH_MDS;
49
50 req = ceph_mdsc_create_request(mdsc, op, want_auth);
51 if (IS_ERR(req))
52 goto out;
53 req->r_fmode = ceph_flags_to_mode(flags);
54 req->r_args.open.flags = cpu_to_le32(flags);
55 req->r_args.open.mode = cpu_to_le32(create_mode);
Sage Weil6a18be12009-11-04 11:40:05 -080056 req->r_args.open.preferred = cpu_to_le32(-1);
Sage Weil124e68e2009-10-06 11:31:08 -070057out:
58 return req;
59}
60
61/*
62 * initialize private struct file data.
63 * if we fail, clean up by dropping fmode reference on the ceph_inode
64 */
65static int ceph_init_file(struct inode *inode, struct file *file, int fmode)
66{
67 struct ceph_file_info *cf;
68 int ret = 0;
69
70 switch (inode->i_mode & S_IFMT) {
71 case S_IFREG:
72 case S_IFDIR:
73 dout("init_file %p %p 0%o (regular)\n", inode, file,
74 inode->i_mode);
75 cf = kmem_cache_alloc(ceph_file_cachep, GFP_NOFS | __GFP_ZERO);
76 if (cf == NULL) {
77 ceph_put_fmode(ceph_inode(inode), fmode); /* clean up */
78 return -ENOMEM;
79 }
80 cf->fmode = fmode;
81 cf->next_offset = 2;
82 file->private_data = cf;
83 BUG_ON(inode->i_fop->release != ceph_release);
84 break;
85
86 case S_IFLNK:
87 dout("init_file %p %p 0%o (symlink)\n", inode, file,
88 inode->i_mode);
89 ceph_put_fmode(ceph_inode(inode), fmode); /* clean up */
90 break;
91
92 default:
93 dout("init_file %p %p 0%o (special)\n", inode, file,
94 inode->i_mode);
95 /*
96 * we need to drop the open ref now, since we don't
97 * have .release set to ceph_release.
98 */
99 ceph_put_fmode(ceph_inode(inode), fmode); /* clean up */
100 BUG_ON(inode->i_fop->release == ceph_release);
101
102 /* call the proper open fop */
103 ret = inode->i_fop->open(inode, file);
104 }
105 return ret;
106}
107
108/*
109 * If the filp already has private_data, that means the file was
110 * already opened by intent during lookup, and we do nothing.
111 *
112 * If we already have the requisite capabilities, we can satisfy
113 * the open request locally (no need to request new caps from the
114 * MDS). We do, however, need to inform the MDS (asynchronously)
115 * if our wanted caps set expands.
116 */
117int ceph_open(struct inode *inode, struct file *file)
118{
119 struct ceph_inode_info *ci = ceph_inode(inode);
120 struct ceph_client *client = ceph_sb_to_client(inode->i_sb);
121 struct ceph_mds_client *mdsc = &client->mdsc;
122 struct ceph_mds_request *req;
123 struct ceph_file_info *cf = file->private_data;
124 struct inode *parent_inode = file->f_dentry->d_parent->d_inode;
125 int err;
126 int flags, fmode, wanted;
127
128 if (cf) {
129 dout("open file %p is already opened\n", file);
130 return 0;
131 }
132
133 /* filter out O_CREAT|O_EXCL; vfs did that already. yuck. */
134 flags = file->f_flags & ~(O_CREAT|O_EXCL);
135 if (S_ISDIR(inode->i_mode))
136 flags = O_DIRECTORY; /* mds likes to know */
137
138 dout("open inode %p ino %llx.%llx file %p flags %d (%d)\n", inode,
139 ceph_vinop(inode), file, flags, file->f_flags);
140 fmode = ceph_flags_to_mode(flags);
141 wanted = ceph_caps_for_mode(fmode);
142
143 /* snapped files are read-only */
144 if (ceph_snap(inode) != CEPH_NOSNAP && (file->f_mode & FMODE_WRITE))
145 return -EROFS;
146
147 /* trivially open snapdir */
148 if (ceph_snap(inode) == CEPH_SNAPDIR) {
149 spin_lock(&inode->i_lock);
150 __ceph_get_fmode(ci, fmode);
151 spin_unlock(&inode->i_lock);
152 return ceph_init_file(inode, file, fmode);
153 }
154
155 /*
156 * No need to block if we have any caps. Update wanted set
157 * asynchronously.
158 */
159 spin_lock(&inode->i_lock);
160 if (__ceph_is_any_real_caps(ci)) {
161 int mds_wanted = __ceph_caps_mds_wanted(ci);
162 int issued = __ceph_caps_issued(ci, NULL);
163
164 dout("open %p fmode %d want %s issued %s using existing\n",
165 inode, fmode, ceph_cap_string(wanted),
166 ceph_cap_string(issued));
167 __ceph_get_fmode(ci, fmode);
168 spin_unlock(&inode->i_lock);
169
170 /* adjust wanted? */
171 if ((issued & wanted) != wanted &&
172 (mds_wanted & wanted) != wanted &&
173 ceph_snap(inode) != CEPH_SNAPDIR)
174 ceph_check_caps(ci, 0, NULL);
175
176 return ceph_init_file(inode, file, fmode);
177 } else if (ceph_snap(inode) != CEPH_NOSNAP &&
178 (ci->i_snap_caps & wanted) == wanted) {
179 __ceph_get_fmode(ci, fmode);
180 spin_unlock(&inode->i_lock);
181 return ceph_init_file(inode, file, fmode);
182 }
183 spin_unlock(&inode->i_lock);
184
185 dout("open fmode %d wants %s\n", fmode, ceph_cap_string(wanted));
186 req = prepare_open_request(inode->i_sb, flags, 0);
187 if (IS_ERR(req)) {
188 err = PTR_ERR(req);
189 goto out;
190 }
191 req->r_inode = igrab(inode);
192 req->r_num_caps = 1;
193 err = ceph_mdsc_do_request(mdsc, parent_inode, req);
194 if (!err)
195 err = ceph_init_file(inode, file, req->r_fmode);
196 ceph_mdsc_put_request(req);
197 dout("open result=%d on %llx.%llx\n", err, ceph_vinop(inode));
198out:
199 return err;
200}
201
202
203/*
204 * Do a lookup + open with a single request.
205 *
206 * If this succeeds, but some subsequent check in the vfs
207 * may_open() fails, the struct *file gets cleaned up (i.e.
208 * ceph_release gets called). So fear not!
209 */
210/*
211 * flags
212 * path_lookup_open -> LOOKUP_OPEN
213 * path_lookup_create -> LOOKUP_OPEN|LOOKUP_CREATE
214 */
215struct dentry *ceph_lookup_open(struct inode *dir, struct dentry *dentry,
216 struct nameidata *nd, int mode,
217 int locked_dir)
218{
219 struct ceph_client *client = ceph_sb_to_client(dir->i_sb);
220 struct ceph_mds_client *mdsc = &client->mdsc;
221 struct file *file = nd->intent.open.file;
222 struct inode *parent_inode = get_dentry_parent_inode(file->f_dentry);
223 struct ceph_mds_request *req;
224 int err;
225 int flags = nd->intent.open.flags - 1; /* silly vfs! */
226
227 dout("ceph_lookup_open dentry %p '%.*s' flags %d mode 0%o\n",
228 dentry, dentry->d_name.len, dentry->d_name.name, flags, mode);
229
230 /* do the open */
231 req = prepare_open_request(dir->i_sb, flags, mode);
232 if (IS_ERR(req))
233 return ERR_PTR(PTR_ERR(req));
234 req->r_dentry = dget(dentry);
235 req->r_num_caps = 2;
236 if (flags & O_CREAT) {
237 req->r_dentry_drop = CEPH_CAP_FILE_SHARED;
238 req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
239 }
240 req->r_locked_dir = dir; /* caller holds dir->i_mutex */
241 err = ceph_mdsc_do_request(mdsc, parent_inode, req);
242 dentry = ceph_finish_lookup(req, dentry, err);
243 if (!err && (flags & O_CREAT) && !req->r_reply_info.head->is_dentry)
244 err = ceph_handle_notrace_create(dir, dentry);
245 if (!err)
246 err = ceph_init_file(req->r_dentry->d_inode, file,
247 req->r_fmode);
248 ceph_mdsc_put_request(req);
249 dout("ceph_lookup_open result=%p\n", dentry);
250 return dentry;
251}
252
253int ceph_release(struct inode *inode, struct file *file)
254{
255 struct ceph_inode_info *ci = ceph_inode(inode);
256 struct ceph_file_info *cf = file->private_data;
257
258 dout("release inode %p file %p\n", inode, file);
259 ceph_put_fmode(ci, cf->fmode);
260 if (cf->last_readdir)
261 ceph_mdsc_put_request(cf->last_readdir);
262 kfree(cf->last_name);
263 kfree(cf->dir_info);
264 dput(cf->dentry);
265 kmem_cache_free(ceph_file_cachep, cf);
Sage Weil195d3ce2010-03-01 09:57:54 -0800266
267 /* wake up anyone waiting for caps on this inode */
268 wake_up(&ci->i_cap_wq);
Sage Weil124e68e2009-10-06 11:31:08 -0700269 return 0;
270}
271
272/*
273 * build a vector of user pages
274 */
275static struct page **get_direct_page_vector(const char __user *data,
276 int num_pages,
277 loff_t off, size_t len)
278{
279 struct page **pages;
280 int rc;
281
282 pages = kmalloc(sizeof(*pages) * num_pages, GFP_NOFS);
283 if (!pages)
284 return ERR_PTR(-ENOMEM);
285
286 down_read(&current->mm->mmap_sem);
287 rc = get_user_pages(current, current->mm, (unsigned long)data,
288 num_pages, 0, 0, pages, NULL);
289 up_read(&current->mm->mmap_sem);
290 if (rc < 0)
291 goto fail;
292 return pages;
293
294fail:
295 kfree(pages);
296 return ERR_PTR(rc);
297}
298
299static void put_page_vector(struct page **pages, int num_pages)
300{
301 int i;
302
303 for (i = 0; i < num_pages; i++)
304 put_page(pages[i]);
305 kfree(pages);
306}
307
308void ceph_release_page_vector(struct page **pages, int num_pages)
309{
310 int i;
311
312 for (i = 0; i < num_pages; i++)
313 __free_pages(pages[i], 0);
314 kfree(pages);
315}
316
317/*
318 * allocate a vector new pages
319 */
320static struct page **alloc_page_vector(int num_pages)
321{
322 struct page **pages;
323 int i;
324
325 pages = kmalloc(sizeof(*pages) * num_pages, GFP_NOFS);
326 if (!pages)
327 return ERR_PTR(-ENOMEM);
328 for (i = 0; i < num_pages; i++) {
329 pages[i] = alloc_page(GFP_NOFS);
330 if (pages[i] == NULL) {
331 ceph_release_page_vector(pages, i);
332 return ERR_PTR(-ENOMEM);
333 }
334 }
335 return pages;
336}
337
338/*
339 * copy user data into a page vector
340 */
341static int copy_user_to_page_vector(struct page **pages,
342 const char __user *data,
343 loff_t off, size_t len)
344{
345 int i = 0;
346 int po = off & ~PAGE_CACHE_MASK;
347 int left = len;
348 int l, bad;
349
350 while (left > 0) {
351 l = min_t(int, PAGE_CACHE_SIZE-po, left);
352 bad = copy_from_user(page_address(pages[i]) + po, data, l);
353 if (bad == l)
354 return -EFAULT;
355 data += l - bad;
356 left -= l - bad;
Yehuda Sadeh6a4ef482009-12-31 12:04:58 -0800357 po += l - bad;
358 if (po == PAGE_CACHE_SIZE) {
359 po = 0;
360 i++;
Sage Weil124e68e2009-10-06 11:31:08 -0700361 }
362 }
363 return len;
364}
365
366/*
367 * copy user data from a page vector into a user pointer
368 */
369static int copy_page_vector_to_user(struct page **pages, char __user *data,
370 loff_t off, size_t len)
371{
372 int i = 0;
373 int po = off & ~PAGE_CACHE_MASK;
374 int left = len;
375 int l, bad;
376
377 while (left > 0) {
378 l = min_t(int, left, PAGE_CACHE_SIZE-po);
379 bad = copy_to_user(data, page_address(pages[i]) + po, l);
380 if (bad == l)
381 return -EFAULT;
382 data += l - bad;
383 left -= l - bad;
384 if (po) {
385 po += l - bad;
386 if (po == PAGE_CACHE_SIZE)
387 po = 0;
388 }
389 i++;
390 }
391 return len;
392}
393
394/*
395 * Zero an extent within a page vector. Offset is relative to the
396 * start of the first page.
397 */
398static void zero_page_vector_range(int off, int len, struct page **pages)
399{
400 int i = off >> PAGE_CACHE_SHIFT;
401
Yehuda Sadeh972f0d32010-02-04 13:41:41 -0800402 off &= ~PAGE_CACHE_MASK;
403
Sage Weil124e68e2009-10-06 11:31:08 -0700404 dout("zero_page_vector_page %u~%u\n", off, len);
Sage Weil124e68e2009-10-06 11:31:08 -0700405
406 /* leading partial page? */
Yehuda Sadeh972f0d32010-02-04 13:41:41 -0800407 if (off) {
408 int end = min((int)PAGE_CACHE_SIZE, off + len);
Sage Weil124e68e2009-10-06 11:31:08 -0700409 dout("zeroing %d %p head from %d\n", i, pages[i],
Yehuda Sadeh972f0d32010-02-04 13:41:41 -0800410 (int)off);
411 zero_user_segment(pages[i], off, end);
412 len -= (end - off);
Sage Weil124e68e2009-10-06 11:31:08 -0700413 i++;
414 }
415 while (len >= PAGE_CACHE_SIZE) {
Yehuda Sadeh29065a52010-02-09 11:14:41 -0800416 dout("zeroing %d %p len=%d\n", i, pages[i], len);
Sage Weil124e68e2009-10-06 11:31:08 -0700417 zero_user_segment(pages[i], 0, PAGE_CACHE_SIZE);
Sage Weil124e68e2009-10-06 11:31:08 -0700418 len -= PAGE_CACHE_SIZE;
419 i++;
420 }
421 /* trailing partial page? */
422 if (len) {
423 dout("zeroing %d %p tail to %d\n", i, pages[i], (int)len);
424 zero_user_segment(pages[i], 0, len);
425 }
426}
427
428
429/*
430 * Read a range of bytes striped over one or more objects. Iterate over
431 * objects we stripe over. (That's not atomic, but good enough for now.)
432 *
433 * If we get a short result from the OSD, check against i_size; we need to
434 * only return a short read to the caller if we hit EOF.
435 */
436static int striped_read(struct inode *inode,
437 u64 off, u64 len,
Sage Weil6a026582010-02-09 14:04:02 -0800438 struct page **pages, int num_pages,
439 int *checkeof)
Sage Weil124e68e2009-10-06 11:31:08 -0700440{
441 struct ceph_client *client = ceph_inode_to_client(inode);
442 struct ceph_inode_info *ci = ceph_inode(inode);
443 u64 pos, this_len;
Yehuda Sadeh972f0d32010-02-04 13:41:41 -0800444 int page_off = off & ~PAGE_CACHE_MASK; /* first byte's offset in page */
Sage Weil124e68e2009-10-06 11:31:08 -0700445 int left, pages_left;
446 int read;
447 struct page **page_pos;
448 int ret;
449 bool hit_stripe, was_short;
450
451 /*
452 * we may need to do multiple reads. not atomic, unfortunately.
453 */
454 pos = off;
455 left = len;
456 page_pos = pages;
457 pages_left = num_pages;
458 read = 0;
459
460more:
461 this_len = left;
462 ret = ceph_osdc_readpages(&client->osdc, ceph_vino(inode),
463 &ci->i_layout, pos, &this_len,
464 ci->i_truncate_seq,
465 ci->i_truncate_size,
466 page_pos, pages_left);
467 hit_stripe = this_len < left;
468 was_short = ret >= 0 && ret < this_len;
469 if (ret == -ENOENT)
470 ret = 0;
471 dout("striped_read %llu~%u (read %u) got %d%s%s\n", pos, left, read,
472 ret, hit_stripe ? " HITSTRIPE" : "", was_short ? " SHORT" : "");
473
474 if (ret > 0) {
475 int didpages =
476 ((pos & ~PAGE_CACHE_MASK) + ret) >> PAGE_CACHE_SHIFT;
477
478 if (read < pos - off) {
479 dout(" zero gap %llu to %llu\n", off + read, pos);
480 zero_page_vector_range(page_off + read,
481 pos - off - read, pages);
482 }
483 pos += ret;
484 read = pos - off;
485 left -= ret;
486 page_pos += didpages;
487 pages_left -= didpages;
488
489 /* hit stripe? */
490 if (left && hit_stripe)
491 goto more;
492 }
493
494 if (was_short) {
495 /* was original extent fully inside i_size? */
496 if (pos + left <= inode->i_size) {
497 dout("zero tail\n");
498 zero_page_vector_range(page_off + read, len - read,
499 pages);
Yehuda Sadeh972f0d32010-02-04 13:41:41 -0800500 read = len;
Sage Weil124e68e2009-10-06 11:31:08 -0700501 goto out;
502 }
503
504 /* check i_size */
Sage Weil6a026582010-02-09 14:04:02 -0800505 *checkeof = 1;
Sage Weil124e68e2009-10-06 11:31:08 -0700506 }
507
508out:
509 if (ret >= 0)
510 ret = read;
511 dout("striped_read returns %d\n", ret);
512 return ret;
513}
514
515/*
516 * Completely synchronous read and write methods. Direct from __user
517 * buffer to osd, or directly to user pages (if O_DIRECT).
518 *
519 * If the read spans object boundary, just do multiple reads.
520 */
521static ssize_t ceph_sync_read(struct file *file, char __user *data,
Sage Weil6a026582010-02-09 14:04:02 -0800522 unsigned len, loff_t *poff, int *checkeof)
Sage Weil124e68e2009-10-06 11:31:08 -0700523{
524 struct inode *inode = file->f_dentry->d_inode;
525 struct page **pages;
526 u64 off = *poff;
527 int num_pages = calc_pages_for(off, len);
528 int ret;
529
530 dout("sync_read on file %p %llu~%u %s\n", file, off, len,
531 (file->f_flags & O_DIRECT) ? "O_DIRECT" : "");
532
533 if (file->f_flags & O_DIRECT) {
534 pages = get_direct_page_vector(data, num_pages, off, len);
535
536 /*
537 * flush any page cache pages in this range. this
538 * will make concurrent normal and O_DIRECT io slow,
539 * but it will at least behave sensibly when they are
540 * in sequence.
541 */
Sage Weil124e68e2009-10-06 11:31:08 -0700542 } else {
543 pages = alloc_page_vector(num_pages);
544 }
545 if (IS_ERR(pages))
546 return PTR_ERR(pages);
547
Yehuda Sadeh29065a52010-02-09 11:14:41 -0800548 ret = filemap_write_and_wait(inode->i_mapping);
549 if (ret < 0)
550 goto done;
551
Sage Weil6a026582010-02-09 14:04:02 -0800552 ret = striped_read(inode, off, len, pages, num_pages, checkeof);
Sage Weil124e68e2009-10-06 11:31:08 -0700553
554 if (ret >= 0 && (file->f_flags & O_DIRECT) == 0)
555 ret = copy_page_vector_to_user(pages, data, off, ret);
556 if (ret >= 0)
557 *poff = off + ret;
558
Yehuda Sadeh29065a52010-02-09 11:14:41 -0800559done:
Sage Weil124e68e2009-10-06 11:31:08 -0700560 if (file->f_flags & O_DIRECT)
561 put_page_vector(pages, num_pages);
562 else
563 ceph_release_page_vector(pages, num_pages);
564 dout("sync_read result %d\n", ret);
565 return ret;
566}
567
568/*
569 * Write commit callback, called if we requested both an ACK and
570 * ONDISK commit reply from the OSD.
571 */
572static void sync_write_commit(struct ceph_osd_request *req,
573 struct ceph_msg *msg)
574{
575 struct ceph_inode_info *ci = ceph_inode(req->r_inode);
576
577 dout("sync_write_commit %p tid %llu\n", req, req->r_tid);
578 spin_lock(&ci->i_unsafe_lock);
579 list_del_init(&req->r_unsafe_item);
580 spin_unlock(&ci->i_unsafe_lock);
581 ceph_put_cap_refs(ci, CEPH_CAP_FILE_WR);
582}
583
584/*
585 * Synchronous write, straight from __user pointer or user pages (if
586 * O_DIRECT).
587 *
588 * If write spans object boundary, just do multiple writes. (For a
589 * correct atomic write, we should e.g. take write locks on all
590 * objects, rollback on failure, etc.)
591 */
592static ssize_t ceph_sync_write(struct file *file, const char __user *data,
593 size_t left, loff_t *offset)
594{
595 struct inode *inode = file->f_dentry->d_inode;
596 struct ceph_inode_info *ci = ceph_inode(inode);
597 struct ceph_client *client = ceph_inode_to_client(inode);
598 struct ceph_osd_request *req;
599 struct page **pages;
600 int num_pages;
601 long long unsigned pos;
602 u64 len;
603 int written = 0;
604 int flags;
605 int do_sync = 0;
606 int check_caps = 0;
607 int ret;
608 struct timespec mtime = CURRENT_TIME;
609
610 if (ceph_snap(file->f_dentry->d_inode) != CEPH_NOSNAP)
611 return -EROFS;
612
613 dout("sync_write on file %p %lld~%u %s\n", file, *offset,
614 (unsigned)left, (file->f_flags & O_DIRECT) ? "O_DIRECT" : "");
615
616 if (file->f_flags & O_APPEND)
617 pos = i_size_read(inode);
618 else
619 pos = *offset;
620
Yehuda Sadeh29065a52010-02-09 11:14:41 -0800621 ret = filemap_write_and_wait_range(inode->i_mapping, pos, pos + left);
622 if (ret < 0)
623 return ret;
624
625 ret = invalidate_inode_pages2_range(inode->i_mapping,
626 pos >> PAGE_CACHE_SHIFT,
627 (pos + left) >> PAGE_CACHE_SHIFT);
628 if (ret < 0)
629 dout("invalidate_inode_pages2_range returned %d\n", ret);
630
Sage Weil124e68e2009-10-06 11:31:08 -0700631 flags = CEPH_OSD_FLAG_ORDERSNAP |
632 CEPH_OSD_FLAG_ONDISK |
633 CEPH_OSD_FLAG_WRITE;
634 if ((file->f_flags & (O_SYNC|O_DIRECT)) == 0)
635 flags |= CEPH_OSD_FLAG_ACK;
636 else
637 do_sync = 1;
638
639 /*
640 * we may need to do multiple writes here if we span an object
641 * boundary. this isn't atomic, unfortunately. :(
642 */
643more:
644 len = left;
645 req = ceph_osdc_new_request(&client->osdc, &ci->i_layout,
646 ceph_vino(inode), pos, &len,
647 CEPH_OSD_OP_WRITE, flags,
648 ci->i_snap_realm->cached_context,
649 do_sync,
650 ci->i_truncate_seq, ci->i_truncate_size,
651 &mtime, false, 2);
652 if (IS_ERR(req))
653 return PTR_ERR(req);
654
655 num_pages = calc_pages_for(pos, len);
656
657 if (file->f_flags & O_DIRECT) {
658 pages = get_direct_page_vector(data, num_pages, pos, len);
659 if (IS_ERR(pages)) {
660 ret = PTR_ERR(pages);
661 goto out;
662 }
663
664 /*
665 * throw out any page cache pages in this range. this
666 * may block.
667 */
668 truncate_inode_pages_range(inode->i_mapping, pos, pos+len);
669 } else {
670 pages = alloc_page_vector(num_pages);
671 if (IS_ERR(pages)) {
672 ret = PTR_ERR(pages);
673 goto out;
674 }
675 ret = copy_user_to_page_vector(pages, data, pos, len);
676 if (ret < 0) {
677 ceph_release_page_vector(pages, num_pages);
678 goto out;
679 }
680
681 if ((file->f_flags & O_SYNC) == 0) {
682 /* get a second commit callback */
683 req->r_safe_callback = sync_write_commit;
684 req->r_own_pages = 1;
685 }
686 }
687 req->r_pages = pages;
688 req->r_num_pages = num_pages;
689 req->r_inode = inode;
690
691 ret = ceph_osdc_start_request(&client->osdc, req, false);
692 if (!ret) {
693 if (req->r_safe_callback) {
694 /*
695 * Add to inode unsafe list only after we
696 * start_request so that a tid has been assigned.
697 */
698 spin_lock(&ci->i_unsafe_lock);
699 list_add(&ci->i_unsafe_writes, &req->r_unsafe_item);
700 spin_unlock(&ci->i_unsafe_lock);
701 ceph_get_cap_refs(ci, CEPH_CAP_FILE_WR);
702 }
703 ret = ceph_osdc_wait_request(&client->osdc, req);
704 }
705
706 if (file->f_flags & O_DIRECT)
707 put_page_vector(pages, num_pages);
708 else if (file->f_flags & O_SYNC)
709 ceph_release_page_vector(pages, num_pages);
710
711out:
712 ceph_osdc_put_request(req);
713 if (ret == 0) {
714 pos += len;
715 written += len;
716 left -= len;
717 if (left)
718 goto more;
719
720 ret = written;
721 *offset = pos;
722 if (pos > i_size_read(inode))
723 check_caps = ceph_inode_set_size(inode, pos);
724 if (check_caps)
725 ceph_check_caps(ceph_inode(inode), CHECK_CAPS_AUTHONLY,
726 NULL);
727 }
728 return ret;
729}
730
731/*
732 * Wrap generic_file_aio_read with checks for cap bits on the inode.
733 * Atomically grab references, so that those bits are not released
734 * back to the MDS mid-read.
735 *
736 * Hmm, the sync read case isn't actually async... should it be?
737 */
738static ssize_t ceph_aio_read(struct kiocb *iocb, const struct iovec *iov,
739 unsigned long nr_segs, loff_t pos)
740{
741 struct file *filp = iocb->ki_filp;
742 loff_t *ppos = &iocb->ki_pos;
743 size_t len = iov->iov_len;
744 struct inode *inode = filp->f_dentry->d_inode;
745 struct ceph_inode_info *ci = ceph_inode(inode);
Sage Weil6a026582010-02-09 14:04:02 -0800746 void *base = iov->iov_base;
Sage Weil124e68e2009-10-06 11:31:08 -0700747 ssize_t ret;
748 int got = 0;
Sage Weil6a026582010-02-09 14:04:02 -0800749 int checkeof = 0, read = 0;
Sage Weil124e68e2009-10-06 11:31:08 -0700750
751 dout("aio_read %p %llx.%llx %llu~%u trying to get caps on %p\n",
752 inode, ceph_vinop(inode), pos, (unsigned)len, inode);
Sage Weil6a026582010-02-09 14:04:02 -0800753again:
Sage Weil124e68e2009-10-06 11:31:08 -0700754 __ceph_do_pending_vmtruncate(inode);
755 ret = ceph_get_caps(ci, CEPH_CAP_FILE_RD, CEPH_CAP_FILE_CACHE,
756 &got, -1);
757 if (ret < 0)
758 goto out;
759 dout("aio_read %p %llx.%llx %llu~%u got cap refs on %s\n",
760 inode, ceph_vinop(inode), pos, (unsigned)len,
761 ceph_cap_string(got));
762
763 if ((got & CEPH_CAP_FILE_CACHE) == 0 ||
764 (iocb->ki_filp->f_flags & O_DIRECT) ||
765 (inode->i_sb->s_flags & MS_SYNCHRONOUS))
766 /* hmm, this isn't really async... */
Sage Weil6a026582010-02-09 14:04:02 -0800767 ret = ceph_sync_read(filp, base, len, ppos, &checkeof);
Sage Weil124e68e2009-10-06 11:31:08 -0700768 else
769 ret = generic_file_aio_read(iocb, iov, nr_segs, pos);
770
771out:
772 dout("aio_read %p %llx.%llx dropping cap refs on %s = %d\n",
773 inode, ceph_vinop(inode), ceph_cap_string(got), (int)ret);
774 ceph_put_cap_refs(ci, got);
Sage Weil6a026582010-02-09 14:04:02 -0800775
776 if (checkeof && ret >= 0) {
777 int statret = ceph_do_getattr(inode, CEPH_STAT_CAP_SIZE);
778
779 /* hit EOF or hole? */
780 if (statret == 0 && *ppos < inode->i_size) {
781 dout("aio_read sync_read hit hole, reading more\n");
782 read += ret;
783 base += ret;
784 len -= ret;
785 checkeof = 0;
786 goto again;
787 }
788 }
789 if (ret >= 0)
790 ret += read;
791
Sage Weil124e68e2009-10-06 11:31:08 -0700792 return ret;
793}
794
795/*
796 * Take cap references to avoid releasing caps to MDS mid-write.
797 *
798 * If we are synchronous, and write with an old snap context, the OSD
799 * may return EOLDSNAPC. In that case, retry the write.. _after_
800 * dropping our cap refs and allowing the pending snap to logically
801 * complete _before_ this write occurs.
802 *
803 * If we are near ENOSPC, write synchronously.
804 */
805static ssize_t ceph_aio_write(struct kiocb *iocb, const struct iovec *iov,
806 unsigned long nr_segs, loff_t pos)
807{
808 struct file *file = iocb->ki_filp;
809 struct inode *inode = file->f_dentry->d_inode;
810 struct ceph_inode_info *ci = ceph_inode(inode);
811 struct ceph_osd_client *osdc = &ceph_client(inode->i_sb)->osdc;
812 loff_t endoff = pos + iov->iov_len;
813 int got = 0;
Yehuda Sadeh88d892a2010-02-23 18:16:23 +0000814 int ret, err;
Sage Weil124e68e2009-10-06 11:31:08 -0700815
816 if (ceph_snap(inode) != CEPH_NOSNAP)
817 return -EROFS;
818
819retry_snap:
820 if (ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_FULL))
821 return -ENOSPC;
822 __ceph_do_pending_vmtruncate(inode);
823 dout("aio_write %p %llx.%llx %llu~%u getting caps. i_size %llu\n",
824 inode, ceph_vinop(inode), pos, (unsigned)iov->iov_len,
825 inode->i_size);
826 ret = ceph_get_caps(ci, CEPH_CAP_FILE_WR, CEPH_CAP_FILE_BUFFER,
827 &got, endoff);
828 if (ret < 0)
829 goto out;
830
831 dout("aio_write %p %llx.%llx %llu~%u got cap refs on %s\n",
832 inode, ceph_vinop(inode), pos, (unsigned)iov->iov_len,
833 ceph_cap_string(got));
834
835 if ((got & CEPH_CAP_FILE_BUFFER) == 0 ||
836 (iocb->ki_filp->f_flags & O_DIRECT) ||
837 (inode->i_sb->s_flags & MS_SYNCHRONOUS)) {
838 ret = ceph_sync_write(file, iov->iov_base, iov->iov_len,
839 &iocb->ki_pos);
840 } else {
841 ret = generic_file_aio_write(iocb, iov, nr_segs, pos);
842
843 if ((ret >= 0 || ret == -EIOCBQUEUED) &&
844 ((file->f_flags & O_SYNC) || IS_SYNC(file->f_mapping->host)
Yehuda Sadeh88d892a2010-02-23 18:16:23 +0000845 || ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_NEARFULL))) {
846 err = vfs_fsync_range(file, file->f_path.dentry,
Sage Weil124e68e2009-10-06 11:31:08 -0700847 pos, pos + ret - 1, 1);
Yehuda Sadeh88d892a2010-02-23 18:16:23 +0000848 if (err < 0)
849 ret = err;
850 }
Sage Weil124e68e2009-10-06 11:31:08 -0700851 }
852 if (ret >= 0) {
853 spin_lock(&inode->i_lock);
854 __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR);
855 spin_unlock(&inode->i_lock);
856 }
857
858out:
859 dout("aio_write %p %llx.%llx %llu~%u dropping cap refs on %s\n",
860 inode, ceph_vinop(inode), pos, (unsigned)iov->iov_len,
861 ceph_cap_string(got));
862 ceph_put_cap_refs(ci, got);
863
864 if (ret == -EOLDSNAPC) {
865 dout("aio_write %p %llx.%llx %llu~%u got EOLDSNAPC, retrying\n",
866 inode, ceph_vinop(inode), pos, (unsigned)iov->iov_len);
867 goto retry_snap;
868 }
869
870 return ret;
871}
872
873/*
874 * llseek. be sure to verify file size on SEEK_END.
875 */
876static loff_t ceph_llseek(struct file *file, loff_t offset, int origin)
877{
878 struct inode *inode = file->f_mapping->host;
879 int ret;
880
881 mutex_lock(&inode->i_mutex);
882 __ceph_do_pending_vmtruncate(inode);
883 switch (origin) {
884 case SEEK_END:
885 ret = ceph_do_getattr(inode, CEPH_STAT_CAP_SIZE);
886 if (ret < 0) {
887 offset = ret;
888 goto out;
889 }
890 offset += inode->i_size;
891 break;
892 case SEEK_CUR:
893 /*
894 * Here we special-case the lseek(fd, 0, SEEK_CUR)
895 * position-querying operation. Avoid rewriting the "same"
896 * f_pos value back to the file because a concurrent read(),
897 * write() or lseek() might have altered it
898 */
899 if (offset == 0) {
900 offset = file->f_pos;
901 goto out;
902 }
903 offset += file->f_pos;
904 break;
905 }
906
907 if (offset < 0 || offset > inode->i_sb->s_maxbytes) {
908 offset = -EINVAL;
909 goto out;
910 }
911
912 /* Special lock needed here? */
913 if (offset != file->f_pos) {
914 file->f_pos = offset;
915 file->f_version = 0;
916 }
917
918out:
919 mutex_unlock(&inode->i_mutex);
920 return offset;
921}
922
923const struct file_operations ceph_file_fops = {
924 .open = ceph_open,
925 .release = ceph_release,
926 .llseek = ceph_llseek,
927 .read = do_sync_read,
928 .write = do_sync_write,
929 .aio_read = ceph_aio_read,
930 .aio_write = ceph_aio_write,
931 .mmap = ceph_mmap,
932 .fsync = ceph_fsync,
933 .splice_read = generic_file_splice_read,
934 .splice_write = generic_file_splice_write,
935 .unlocked_ioctl = ceph_ioctl,
936 .compat_ioctl = ceph_ioctl,
937};
938