~ [ source navigation ] ~ [ diff markup ] ~ [ identifier search ] ~

TOMOYO Linux Cross Reference
Linux/fs/ceph/file.c

Version: ~ [ linux-5.18 ] ~ [ linux-5.17.9 ] ~ [ linux-5.16.20 ] ~ [ linux-5.15.41 ] ~ [ linux-5.14.21 ] ~ [ linux-5.13.19 ] ~ [ linux-5.12.19 ] ~ [ linux-5.11.22 ] ~ [ linux-5.10.117 ] ~ [ linux-5.9.16 ] ~ [ linux-5.8.18 ] ~ [ linux-5.7.19 ] ~ [ linux-5.6.19 ] ~ [ linux-5.5.19 ] ~ [ linux-5.4.195 ] ~ [ linux-5.3.18 ] ~ [ linux-5.2.21 ] ~ [ linux-5.1.21 ] ~ [ linux-5.0.21 ] ~ [ linux-4.20.17 ] ~ [ linux-4.19.244 ] ~ [ linux-4.18.20 ] ~ [ linux-4.17.19 ] ~ [ linux-4.16.18 ] ~ [ linux-4.15.18 ] ~ [ linux-4.14.280 ] ~ [ linux-4.13.16 ] ~ [ linux-4.12.14 ] ~ [ linux-4.11.12 ] ~ [ linux-4.10.17 ] ~ [ linux-4.9.315 ] ~ [ linux-4.8.17 ] ~ [ linux-4.7.10 ] ~ [ linux-4.6.7 ] ~ [ linux-4.5.7 ] ~ [ linux-4.4.302 ] ~ [ linux-4.3.6 ] ~ [ linux-4.2.8 ] ~ [ linux-4.1.52 ] ~ [ linux-4.0.9 ] ~ [ linux-3.10.108 ] ~ [ linux-2.6.32.71 ] ~ [ linux-2.6.0 ] ~ [ linux-2.4.37.11 ] ~ [ unix-v6-master ] ~ [ ccs-tools-1.8.9 ] ~ [ policy-sample ] ~
Architecture: ~ [ i386 ] ~ [ alpha ] ~ [ m68k ] ~ [ mips ] ~ [ ppc ] ~ [ sparc ] ~ [ sparc64 ] ~

  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  */
 41 static 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  */
 65 static struct page **
 66 dio_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;
101 fail:
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  */
110 static struct ceph_mds_request *
111 prepare_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);
128 out:
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  */
136 static 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_alloc(ceph_file_cachep, GFP_KERNEL | __GFP_ZERO);
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  */
200 int 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));
285 out:
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  */
294 int 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 err;
304 
305         dout("atomic_open %p dentry %p '%pd' %s flags %d mode 0%o\n",
306              dir, dentry, dentry,
307              d_unhashed(dentry) ? "unhashed" : "hashed", flags, mode);
308 
309         if (dentry->d_name.len > NAME_MAX)
310                 return -ENAMETOOLONG;
311 
312         err = ceph_init_dentry(dentry);
313         if (err < 0)
314                 return err;
315 
316         if (flags & O_CREAT) {
317                 err = ceph_pre_init_acls(dir, &mode, &acls);
318                 if (err < 0)
319                         return err;
320         }
321 
322         /* do the open */
323         req = prepare_open_request(dir->i_sb, flags, mode);
324         if (IS_ERR(req)) {
325                 err = PTR_ERR(req);
326                 goto out_acl;
327         }
328         req->r_dentry = dget(dentry);
329         req->r_num_caps = 2;
330         if (flags & O_CREAT) {
331                 req->r_dentry_drop = CEPH_CAP_FILE_SHARED;
332                 req->r_dentry_unless = CEPH_CAP_FILE_EXCL;
333                 if (acls.pagelist) {
334                         req->r_pagelist = acls.pagelist;
335                         acls.pagelist = NULL;
336                 }
337         }
338         req->r_locked_dir = dir;           /* caller holds dir->i_mutex */
339         err = ceph_mdsc_do_request(mdsc,
340                                    (flags & (O_CREAT|O_TRUNC)) ? dir : NULL,
341                                    req);
342         err = ceph_handle_snapdir(req, dentry, err);
343         if (err)
344                 goto out_req;
345 
346         if ((flags & O_CREAT) && !req->r_reply_info.head->is_dentry)
347                 err = ceph_handle_notrace_create(dir, dentry);
348 
349         if (d_unhashed(dentry)) {
350                 dn = ceph_finish_lookup(req, dentry, err);
351                 if (IS_ERR(dn))
352                         err = PTR_ERR(dn);
353         } else {
354                 /* we were given a hashed negative dentry */
355                 dn = NULL;
356         }
357         if (err)
358                 goto out_req;
359         if (dn || d_really_is_negative(dentry) || d_is_symlink(dentry)) {
360                 /* make vfs retry on splice, ENOENT, or symlink */
361                 dout("atomic_open finish_no_open on dn %p\n", dn);
362                 err = finish_no_open(file, dn);
363         } else {
364                 dout("atomic_open finish_open on dn %p\n", dn);
365                 if (req->r_op == CEPH_MDS_OP_CREATE && req->r_reply_info.has_create_ino) {
366                         ceph_init_inode_acls(d_inode(dentry), &acls);
367                         *opened |= FILE_CREATED;
368                 }
369                 err = finish_open(file, dentry, ceph_open, opened);
370         }
371 out_req:
372         if (!req->r_err && req->r_target_inode)
373                 ceph_put_fmode(ceph_inode(req->r_target_inode), req->r_fmode);
374         ceph_mdsc_put_request(req);
375 out_acl:
376         ceph_release_acls_info(&acls);
377         dout("atomic_open result=%d\n", err);
378         return err;
379 }
380 
381 int ceph_release(struct inode *inode, struct file *file)
382 {
383         struct ceph_inode_info *ci = ceph_inode(inode);
384         struct ceph_file_info *cf = file->private_data;
385 
386         dout("release inode %p file %p\n", inode, file);
387         ceph_put_fmode(ci, cf->fmode);
388         if (cf->last_readdir)
389                 ceph_mdsc_put_request(cf->last_readdir);
390         kfree(cf->last_name);
391         kfree(cf->dir_info);
392         kmem_cache_free(ceph_file_cachep, cf);
393 
394         /* wake up anyone waiting for caps on this inode */
395         wake_up_all(&ci->i_cap_wq);
396         return 0;
397 }
398 
399 enum {
400         HAVE_RETRIED = 1,
401         CHECK_EOF =    2,
402         READ_INLINE =  3,
403 };
404 
405 /*
406  * Read a range of bytes striped over one or more objects.  Iterate over
407  * objects we stripe over.  (That's not atomic, but good enough for now.)
408  *
409  * If we get a short result from the OSD, check against i_size; we need to
410  * only return a short read to the caller if we hit EOF.
411  */
412 static int striped_read(struct inode *inode,
413                         u64 off, u64 len,
414                         struct page **pages, int num_pages,
415                         int *checkeof)
416 {
417         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
418         struct ceph_inode_info *ci = ceph_inode(inode);
419         u64 pos, this_len, left;
420         loff_t i_size;
421         int page_align, pages_left;
422         int read, ret;
423         struct page **page_pos;
424         bool hit_stripe, was_short;
425 
426         /*
427          * we may need to do multiple reads.  not atomic, unfortunately.
428          */
429         pos = off;
430         left = len;
431         page_pos = pages;
432         pages_left = num_pages;
433         read = 0;
434 
435 more:
436         page_align = pos & ~PAGE_MASK;
437         this_len = left;
438         ret = ceph_osdc_readpages(&fsc->client->osdc, ceph_vino(inode),
439                                   &ci->i_layout, pos, &this_len,
440                                   ci->i_truncate_seq,
441                                   ci->i_truncate_size,
442                                   page_pos, pages_left, page_align);
443         if (ret == -ENOENT)
444                 ret = 0;
445         hit_stripe = this_len < left;
446         was_short = ret >= 0 && ret < this_len;
447         dout("striped_read %llu~%llu (read %u) got %d%s%s\n", pos, left, read,
448              ret, hit_stripe ? " HITSTRIPE" : "", was_short ? " SHORT" : "");
449 
450         i_size = i_size_read(inode);
451         if (ret >= 0) {
452                 int didpages;
453                 if (was_short && (pos + ret < i_size)) {
454                         int zlen = min(this_len - ret, i_size - pos - ret);
455                         int zoff = (off & ~PAGE_MASK) + read + ret;
456                         dout(" zero gap %llu to %llu\n",
457                                 pos + ret, pos + ret + zlen);
458                         ceph_zero_page_vector_range(zoff, zlen, pages);
459                         ret += zlen;
460                 }
461 
462                 didpages = (page_align + ret) >> PAGE_CACHE_SHIFT;
463                 pos += ret;
464                 read = pos - off;
465                 left -= ret;
466                 page_pos += didpages;
467                 pages_left -= didpages;
468 
469                 /* hit stripe and need continue*/
470                 if (left && hit_stripe && pos < i_size)
471                         goto more;
472         }
473 
474         if (read > 0) {
475                 ret = read;
476                 /* did we bounce off eof? */
477                 if (pos + left > i_size)
478                         *checkeof = CHECK_EOF;
479         }
480 
481         dout("striped_read returns %d\n", ret);
482         return ret;
483 }
484 
485 /*
486  * Completely synchronous read and write methods.  Direct from __user
487  * buffer to osd, or directly to user pages (if O_DIRECT).
488  *
489  * If the read spans object boundary, just do multiple reads.
490  */
491 static ssize_t ceph_sync_read(struct kiocb *iocb, struct iov_iter *i,
492                                 int *checkeof)
493 {
494         struct file *file = iocb->ki_filp;
495         struct inode *inode = file_inode(file);
496         struct page **pages;
497         u64 off = iocb->ki_pos;
498         int num_pages, ret;
499         size_t len = iov_iter_count(i);
500 
501         dout("sync_read on file %p %llu~%u %s\n", file, off,
502              (unsigned)len,
503              (file->f_flags & O_DIRECT) ? "O_DIRECT" : "");
504 
505         if (!len)
506                 return 0;
507         /*
508          * flush any page cache pages in this range.  this
509          * will make concurrent normal and sync io slow,
510          * but it will at least behave sensibly when they are
511          * in sequence.
512          */
513         ret = filemap_write_and_wait_range(inode->i_mapping, off,
514                                                 off + len);
515         if (ret < 0)
516                 return ret;
517 
518         num_pages = calc_pages_for(off, len);
519         pages = ceph_alloc_page_vector(num_pages, GFP_KERNEL);
520         if (IS_ERR(pages))
521                 return PTR_ERR(pages);
522         ret = striped_read(inode, off, len, pages,
523                                 num_pages, checkeof);
524         if (ret > 0) {
525                 int l, k = 0;
526                 size_t left = ret;
527 
528                 while (left) {
529                         size_t page_off = off & ~PAGE_MASK;
530                         size_t copy = min_t(size_t, left,
531                                             PAGE_SIZE - page_off);
532                         l = copy_page_to_iter(pages[k++], page_off, copy, i);
533                         off += l;
534                         left -= l;
535                         if (l < copy)
536                                 break;
537                 }
538         }
539         ceph_release_page_vector(pages, num_pages);
540 
541         if (off > iocb->ki_pos) {
542                 ret = off - iocb->ki_pos;
543                 iocb->ki_pos = off;
544         }
545 
546         dout("sync_read result %d\n", ret);
547         return ret;
548 }
549 
550 struct ceph_aio_request {
551         struct kiocb *iocb;
552         size_t total_len;
553         int write;
554         int error;
555         struct list_head osd_reqs;
556         unsigned num_reqs;
557         atomic_t pending_reqs;
558         struct timespec mtime;
559         struct ceph_cap_flush *prealloc_cf;
560 };
561 
562 struct ceph_aio_work {
563         struct work_struct work;
564         struct ceph_osd_request *req;
565 };
566 
567 static void ceph_aio_retry_work(struct work_struct *work);
568 
569 static void ceph_aio_complete(struct inode *inode,
570                               struct ceph_aio_request *aio_req)
571 {
572         struct ceph_inode_info *ci = ceph_inode(inode);
573         int ret;
574 
575         if (!atomic_dec_and_test(&aio_req->pending_reqs))
576                 return;
577 
578         ret = aio_req->error;
579         if (!ret)
580                 ret = aio_req->total_len;
581 
582         dout("ceph_aio_complete %p rc %d\n", inode, ret);
583 
584         if (ret >= 0 && aio_req->write) {
585                 int dirty;
586 
587                 loff_t endoff = aio_req->iocb->ki_pos + aio_req->total_len;
588                 if (endoff > i_size_read(inode)) {
589                         if (ceph_inode_set_size(inode, endoff))
590                                 ceph_check_caps(ci, CHECK_CAPS_AUTHONLY, NULL);
591                 }
592 
593                 spin_lock(&ci->i_ceph_lock);
594                 ci->i_inline_version = CEPH_INLINE_NONE;
595                 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR,
596                                                &aio_req->prealloc_cf);
597                 spin_unlock(&ci->i_ceph_lock);
598                 if (dirty)
599                         __mark_inode_dirty(inode, dirty);
600 
601         }
602 
603         ceph_put_cap_refs(ci, (aio_req->write ? CEPH_CAP_FILE_WR :
604                                                 CEPH_CAP_FILE_RD));
605 
606         aio_req->iocb->ki_complete(aio_req->iocb, ret, 0);
607 
608         ceph_free_cap_flush(aio_req->prealloc_cf);
609         kfree(aio_req);
610 }
611 
612 static void ceph_aio_complete_req(struct ceph_osd_request *req,
613                                   struct ceph_msg *msg)
614 {
615         int rc = req->r_result;
616         struct inode *inode = req->r_inode;
617         struct ceph_aio_request *aio_req = req->r_priv;
618         struct ceph_osd_data *osd_data = osd_req_op_extent_osd_data(req, 0);
619         int num_pages = calc_pages_for((u64)osd_data->alignment,
620                                        osd_data->length);
621 
622         dout("ceph_aio_complete_req %p rc %d bytes %llu\n",
623              inode, rc, osd_data->length);
624 
625         if (rc == -EOLDSNAPC) {
626                 struct ceph_aio_work *aio_work;
627                 BUG_ON(!aio_req->write);
628 
629                 aio_work = kmalloc(sizeof(*aio_work), GFP_NOFS);
630                 if (aio_work) {
631                         INIT_WORK(&aio_work->work, ceph_aio_retry_work);
632                         aio_work->req = req;
633                         queue_work(ceph_inode_to_client(inode)->wb_wq,
634                                    &aio_work->work);
635                         return;
636                 }
637                 rc = -ENOMEM;
638         } else if (!aio_req->write) {
639                 if (rc == -ENOENT)
640                         rc = 0;
641                 if (rc >= 0 && osd_data->length > rc) {
642                         int zoff = osd_data->alignment + rc;
643                         int zlen = osd_data->length - rc;
644                         /*
645                          * If read is satisfied by single OSD request,
646                          * it can pass EOF. Otherwise read is within
647                          * i_size.
648                          */
649                         if (aio_req->num_reqs == 1) {
650                                 loff_t i_size = i_size_read(inode);
651                                 loff_t endoff = aio_req->iocb->ki_pos + rc;
652                                 if (endoff < i_size)
653                                         zlen = min_t(size_t, zlen,
654                                                      i_size - endoff);
655                                 aio_req->total_len = rc + zlen;
656                         }
657 
658                         if (zlen > 0)
659                                 ceph_zero_page_vector_range(zoff, zlen,
660                                                             osd_data->pages);
661                 }
662         }
663 
664         ceph_put_page_vector(osd_data->pages, num_pages, false);
665         ceph_osdc_put_request(req);
666 
667         if (rc < 0)
668                 cmpxchg(&aio_req->error, 0, rc);
669 
670         ceph_aio_complete(inode, aio_req);
671         return;
672 }
673 
674 static void ceph_aio_retry_work(struct work_struct *work)
675 {
676         struct ceph_aio_work *aio_work =
677                 container_of(work, struct ceph_aio_work, work);
678         struct ceph_osd_request *orig_req = aio_work->req;
679         struct ceph_aio_request *aio_req = orig_req->r_priv;
680         struct inode *inode = orig_req->r_inode;
681         struct ceph_inode_info *ci = ceph_inode(inode);
682         struct ceph_snap_context *snapc;
683         struct ceph_osd_request *req;
684         int ret;
685 
686         spin_lock(&ci->i_ceph_lock);
687         if (__ceph_have_pending_cap_snap(ci)) {
688                 struct ceph_cap_snap *capsnap =
689                         list_last_entry(&ci->i_cap_snaps,
690                                         struct ceph_cap_snap,
691                                         ci_item);
692                 snapc = ceph_get_snap_context(capsnap->context);
693         } else {
694                 BUG_ON(!ci->i_head_snapc);
695                 snapc = ceph_get_snap_context(ci->i_head_snapc);
696         }
697         spin_unlock(&ci->i_ceph_lock);
698 
699         req = ceph_osdc_alloc_request(orig_req->r_osdc, snapc, 2,
700                         false, GFP_NOFS);
701         if (!req) {
702                 ret = -ENOMEM;
703                 req = orig_req;
704                 goto out;
705         }
706 
707         req->r_flags =  CEPH_OSD_FLAG_ORDERSNAP |
708                         CEPH_OSD_FLAG_ONDISK |
709                         CEPH_OSD_FLAG_WRITE;
710         req->r_base_oloc = orig_req->r_base_oloc;
711         req->r_base_oid = orig_req->r_base_oid;
712 
713         req->r_ops[0] = orig_req->r_ops[0];
714         osd_req_op_init(req, 1, CEPH_OSD_OP_STARTSYNC, 0);
715 
716         ceph_osdc_build_request(req, req->r_ops[0].extent.offset,
717                                 snapc, CEPH_NOSNAP, &aio_req->mtime);
718 
719         ceph_osdc_put_request(orig_req);
720 
721         req->r_callback = ceph_aio_complete_req;
722         req->r_inode = inode;
723         req->r_priv = aio_req;
724 
725         ret = ceph_osdc_start_request(req->r_osdc, req, false);
726 out:
727         if (ret < 0) {
728                 BUG_ON(ret == -EOLDSNAPC);
729                 req->r_result = ret;
730                 ceph_aio_complete_req(req, NULL);
731         }
732 
733         ceph_put_snap_context(snapc);
734         kfree(aio_work);
735 }
736 
737 /*
738  * Write commit request unsafe callback, called to tell us when a
739  * request is unsafe (that is, in flight--has been handed to the
740  * messenger to send to its target osd).  It is called again when
741  * we've received a response message indicating the request is
742  * "safe" (its CEPH_OSD_FLAG_ONDISK flag is set), or when a request
743  * is completed early (and unsuccessfully) due to a timeout or
744  * interrupt.
745  *
746  * This is used if we requested both an ACK and ONDISK commit reply
747  * from the OSD.
748  */
749 static void ceph_sync_write_unsafe(struct ceph_osd_request *req, bool unsafe)
750 {
751         struct ceph_inode_info *ci = ceph_inode(req->r_inode);
752 
753         dout("%s %p tid %llu %ssafe\n", __func__, req, req->r_tid,
754                 unsafe ? "un" : "");
755         if (unsafe) {
756                 ceph_get_cap_refs(ci, CEPH_CAP_FILE_WR);
757                 spin_lock(&ci->i_unsafe_lock);
758                 list_add_tail(&req->r_unsafe_item,
759                               &ci->i_unsafe_writes);
760                 spin_unlock(&ci->i_unsafe_lock);
761         } else {
762                 spin_lock(&ci->i_unsafe_lock);
763                 list_del_init(&req->r_unsafe_item);
764                 spin_unlock(&ci->i_unsafe_lock);
765                 ceph_put_cap_refs(ci, CEPH_CAP_FILE_WR);
766         }
767 }
768 
769 
770 static ssize_t
771 ceph_direct_read_write(struct kiocb *iocb, struct iov_iter *iter,
772                        struct ceph_snap_context *snapc,
773                        struct ceph_cap_flush **pcf)
774 {
775         struct file *file = iocb->ki_filp;
776         struct inode *inode = file_inode(file);
777         struct ceph_inode_info *ci = ceph_inode(inode);
778         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
779         struct ceph_vino vino;
780         struct ceph_osd_request *req;
781         struct page **pages;
782         struct ceph_aio_request *aio_req = NULL;
783         int num_pages = 0;
784         int flags;
785         int ret;
786         struct timespec mtime = CURRENT_TIME;
787         size_t count = iov_iter_count(iter);
788         loff_t pos = iocb->ki_pos;
789         bool write = iov_iter_rw(iter) == WRITE;
790 
791         if (write && ceph_snap(file_inode(file)) != CEPH_NOSNAP)
792                 return -EROFS;
793 
794         dout("sync_direct_read_write (%s) on file %p %lld~%u\n",
795              (write ? "write" : "read"), file, pos, (unsigned)count);
796 
797         ret = filemap_write_and_wait_range(inode->i_mapping, pos, pos + count);
798         if (ret < 0)
799                 return ret;
800 
801         if (write) {
802                 ret = invalidate_inode_pages2_range(inode->i_mapping,
803                                         pos >> PAGE_CACHE_SHIFT,
804                                         (pos + count) >> PAGE_CACHE_SHIFT);
805                 if (ret < 0)
806                         dout("invalidate_inode_pages2_range returned %d\n", ret);
807 
808                 flags = CEPH_OSD_FLAG_ORDERSNAP |
809                         CEPH_OSD_FLAG_ONDISK |
810                         CEPH_OSD_FLAG_WRITE;
811         } else {
812                 flags = CEPH_OSD_FLAG_READ;
813         }
814 
815         while (iov_iter_count(iter) > 0) {
816                 u64 size = dio_get_pagev_size(iter);
817                 size_t start = 0;
818                 ssize_t len;
819 
820                 vino = ceph_vino(inode);
821                 req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
822                                             vino, pos, &size, 0,
823                                             /*include a 'startsync' command*/
824                                             write ? 2 : 1,
825                                             write ? CEPH_OSD_OP_WRITE :
826                                                     CEPH_OSD_OP_READ,
827                                             flags, snapc,
828                                             ci->i_truncate_seq,
829                                             ci->i_truncate_size,
830                                             false);
831                 if (IS_ERR(req)) {
832                         ret = PTR_ERR(req);
833                         break;
834                 }
835 
836                 len = size;
837                 pages = dio_get_pages_alloc(iter, len, &start, &num_pages);
838                 if (IS_ERR(pages)) {
839                         ceph_osdc_put_request(req);
840                         ret = PTR_ERR(pages);
841                         break;
842                 }
843 
844                 /*
845                  * To simplify error handling, allow AIO when IO within i_size
846                  * or IO can be satisfied by single OSD request.
847                  */
848                 if (pos == iocb->ki_pos && !is_sync_kiocb(iocb) &&
849                     (len == count || pos + count <= i_size_read(inode))) {
850                         aio_req = kzalloc(sizeof(*aio_req), GFP_KERNEL);
851                         if (aio_req) {
852                                 aio_req->iocb = iocb;
853                                 aio_req->write = write;
854                                 INIT_LIST_HEAD(&aio_req->osd_reqs);
855                                 if (write) {
856                                         aio_req->mtime = mtime;
857                                         swap(aio_req->prealloc_cf, *pcf);
858                                 }
859                         }
860                         /* ignore error */
861                 }
862 
863                 if (write) {
864                         /*
865                          * throw out any page cache pages in this range. this
866                          * may block.
867                          */
868                         truncate_inode_pages_range(inode->i_mapping, pos,
869                                         (pos+len) | (PAGE_CACHE_SIZE - 1));
870 
871                         osd_req_op_init(req, 1, CEPH_OSD_OP_STARTSYNC, 0);
872                 }
873 
874 
875                 osd_req_op_extent_osd_data_pages(req, 0, pages, len, start,
876                                                  false, false);
877 
878                 ceph_osdc_build_request(req, pos, snapc, vino.snap, &mtime);
879 
880                 if (aio_req) {
881                         aio_req->total_len += len;
882                         aio_req->num_reqs++;
883                         atomic_inc(&aio_req->pending_reqs);
884 
885                         req->r_callback = ceph_aio_complete_req;
886                         req->r_inode = inode;
887                         req->r_priv = aio_req;
888                         list_add_tail(&req->r_unsafe_item, &aio_req->osd_reqs);
889 
890                         pos += len;
891                         iov_iter_advance(iter, len);
892                         continue;
893                 }
894 
895                 ret = ceph_osdc_start_request(req->r_osdc, req, false);
896                 if (!ret)
897                         ret = ceph_osdc_wait_request(&fsc->client->osdc, req);
898 
899                 size = i_size_read(inode);
900                 if (!write) {
901                         if (ret == -ENOENT)
902                                 ret = 0;
903                         if (ret >= 0 && ret < len && pos + ret < size) {
904                                 int zlen = min_t(size_t, len - ret,
905                                                  size - pos - ret);
906                                 ceph_zero_page_vector_range(start + ret, zlen,
907                                                             pages);
908                                 ret += zlen;
909                         }
910                         if (ret >= 0)
911                                 len = ret;
912                 }
913 
914                 ceph_put_page_vector(pages, num_pages, false);
915 
916                 ceph_osdc_put_request(req);
917                 if (ret < 0)
918                         break;
919 
920                 pos += len;
921                 iov_iter_advance(iter, len);
922 
923                 if (!write && pos >= size)
924                         break;
925 
926                 if (write && pos > size) {
927                         if (ceph_inode_set_size(inode, pos))
928                                 ceph_check_caps(ceph_inode(inode),
929                                                 CHECK_CAPS_AUTHONLY,
930                                                 NULL);
931                 }
932         }
933 
934         if (aio_req) {
935                 if (aio_req->num_reqs == 0) {
936                         kfree(aio_req);
937                         return ret;
938                 }
939 
940                 ceph_get_cap_refs(ci, write ? CEPH_CAP_FILE_WR :
941                                               CEPH_CAP_FILE_RD);
942 
943                 while (!list_empty(&aio_req->osd_reqs)) {
944                         req = list_first_entry(&aio_req->osd_reqs,
945                                                struct ceph_osd_request,
946                                                r_unsafe_item);
947                         list_del_init(&req->r_unsafe_item);
948                         if (ret >= 0)
949                                 ret = ceph_osdc_start_request(req->r_osdc,
950                                                               req, false);
951                         if (ret < 0) {
952                                 BUG_ON(ret == -EOLDSNAPC);
953                                 req->r_result = ret;
954                                 ceph_aio_complete_req(req, NULL);
955                         }
956                 }
957                 return -EIOCBQUEUED;
958         }
959 
960         if (ret != -EOLDSNAPC && pos > iocb->ki_pos) {
961                 ret = pos - iocb->ki_pos;
962                 iocb->ki_pos = pos;
963         }
964         return ret;
965 }
966 
967 /*
968  * Synchronous write, straight from __user pointer or user pages.
969  *
970  * If write spans object boundary, just do multiple writes.  (For a
971  * correct atomic write, we should e.g. take write locks on all
972  * objects, rollback on failure, etc.)
973  */
974 static ssize_t
975 ceph_sync_write(struct kiocb *iocb, struct iov_iter *from, loff_t pos,
976                 struct ceph_snap_context *snapc)
977 {
978         struct file *file = iocb->ki_filp;
979         struct inode *inode = file_inode(file);
980         struct ceph_inode_info *ci = ceph_inode(inode);
981         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
982         struct ceph_vino vino;
983         struct ceph_osd_request *req;
984         struct page **pages;
985         u64 len;
986         int num_pages;
987         int written = 0;
988         int flags;
989         int check_caps = 0;
990         int ret;
991         struct timespec mtime = CURRENT_TIME;
992         size_t count = iov_iter_count(from);
993 
994         if (ceph_snap(file_inode(file)) != CEPH_NOSNAP)
995                 return -EROFS;
996 
997         dout("sync_write on file %p %lld~%u\n", file, pos, (unsigned)count);
998 
999         ret = filemap_write_and_wait_range(inode->i_mapping, pos, pos + count);
1000         if (ret < 0)
1001                 return ret;
1002 
1003         ret = invalidate_inode_pages2_range(inode->i_mapping,
1004                                             pos >> PAGE_CACHE_SHIFT,
1005                                             (pos + count) >> PAGE_CACHE_SHIFT);
1006         if (ret < 0)
1007                 dout("invalidate_inode_pages2_range returned %d\n", ret);
1008 
1009         flags = CEPH_OSD_FLAG_ORDERSNAP |
1010                 CEPH_OSD_FLAG_ONDISK |
1011                 CEPH_OSD_FLAG_WRITE |
1012                 CEPH_OSD_FLAG_ACK;
1013 
1014         while ((len = iov_iter_count(from)) > 0) {
1015                 size_t left;
1016                 int n;
1017 
1018                 vino = ceph_vino(inode);
1019                 req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1020                                             vino, pos, &len, 0, 1,
1021                                             CEPH_OSD_OP_WRITE, flags, snapc,
1022                                             ci->i_truncate_seq,
1023                                             ci->i_truncate_size,
1024                                             false);
1025                 if (IS_ERR(req)) {
1026                         ret = PTR_ERR(req);
1027                         break;
1028                 }
1029 
1030                 /*
1031                  * write from beginning of first page,
1032                  * regardless of io alignment
1033                  */
1034                 num_pages = (len + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1035 
1036                 pages = ceph_alloc_page_vector(num_pages, GFP_KERNEL);
1037                 if (IS_ERR(pages)) {
1038                         ret = PTR_ERR(pages);
1039                         goto out;
1040                 }
1041 
1042                 left = len;
1043                 for (n = 0; n < num_pages; n++) {
1044                         size_t plen = min_t(size_t, left, PAGE_SIZE);
1045                         ret = copy_page_from_iter(pages[n], 0, plen, from);
1046                         if (ret != plen) {
1047                                 ret = -EFAULT;
1048                                 break;
1049                         }
1050                         left -= ret;
1051                 }
1052 
1053                 if (ret < 0) {
1054                         ceph_release_page_vector(pages, num_pages);
1055                         goto out;
1056                 }
1057 
1058                 /* get a second commit callback */
1059                 req->r_unsafe_callback = ceph_sync_write_unsafe;
1060                 req->r_inode = inode;
1061 
1062                 osd_req_op_extent_osd_data_pages(req, 0, pages, len, 0,
1063                                                 false, true);
1064 
1065                 /* BUG_ON(vino.snap != CEPH_NOSNAP); */
1066                 ceph_osdc_build_request(req, pos, snapc, vino.snap, &mtime);
1067 
1068                 ret = ceph_osdc_start_request(&fsc->client->osdc, req, false);
1069                 if (!ret)
1070                         ret = ceph_osdc_wait_request(&fsc->client->osdc, req);
1071 
1072 out:
1073                 ceph_osdc_put_request(req);
1074                 if (ret == 0) {
1075                         pos += len;
1076                         written += len;
1077 
1078                         if (pos > i_size_read(inode)) {
1079                                 check_caps = ceph_inode_set_size(inode, pos);
1080                                 if (check_caps)
1081                                         ceph_check_caps(ceph_inode(inode),
1082                                                         CHECK_CAPS_AUTHONLY,
1083                                                         NULL);
1084                         }
1085                 } else
1086                         break;
1087         }
1088 
1089         if (ret != -EOLDSNAPC && written > 0) {
1090                 ret = written;
1091                 iocb->ki_pos = pos;
1092         }
1093         return ret;
1094 }
1095 
1096 /*
1097  * Wrap generic_file_aio_read with checks for cap bits on the inode.
1098  * Atomically grab references, so that those bits are not released
1099  * back to the MDS mid-read.
1100  *
1101  * Hmm, the sync read case isn't actually async... should it be?
1102  */
1103 static ssize_t ceph_read_iter(struct kiocb *iocb, struct iov_iter *to)
1104 {
1105         struct file *filp = iocb->ki_filp;
1106         struct ceph_file_info *fi = filp->private_data;
1107         size_t len = iov_iter_count(to);
1108         struct inode *inode = file_inode(filp);
1109         struct ceph_inode_info *ci = ceph_inode(inode);
1110         struct page *pinned_page = NULL;
1111         ssize_t ret;
1112         int want, got = 0;
1113         int retry_op = 0, read = 0;
1114 
1115 again:
1116         dout("aio_read %p %llx.%llx %llu~%u trying to get caps on %p\n",
1117              inode, ceph_vinop(inode), iocb->ki_pos, (unsigned)len, inode);
1118 
1119         if (fi->fmode & CEPH_FILE_MODE_LAZY)
1120                 want = CEPH_CAP_FILE_CACHE | CEPH_CAP_FILE_LAZYIO;
1121         else
1122                 want = CEPH_CAP_FILE_CACHE;
1123         ret = ceph_get_caps(ci, CEPH_CAP_FILE_RD, want, -1, &got, &pinned_page);
1124         if (ret < 0)
1125                 return ret;
1126 
1127         if ((got & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) == 0 ||
1128             (iocb->ki_flags & IOCB_DIRECT) ||
1129             (fi->flags & CEPH_F_SYNC)) {
1130 
1131                 dout("aio_sync_read %p %llx.%llx %llu~%u got cap refs on %s\n",
1132                      inode, ceph_vinop(inode), iocb->ki_pos, (unsigned)len,
1133                      ceph_cap_string(got));
1134 
1135                 if (ci->i_inline_version == CEPH_INLINE_NONE) {
1136                         if (!retry_op && (iocb->ki_flags & IOCB_DIRECT)) {
1137                                 ret = ceph_direct_read_write(iocb, to,
1138                                                              NULL, NULL);
1139                                 if (ret >= 0 && ret < len)
1140                                         retry_op = CHECK_EOF;
1141                         } else {
1142                                 ret = ceph_sync_read(iocb, to, &retry_op);
1143                         }
1144                 } else {
1145                         retry_op = READ_INLINE;
1146                 }
1147         } else {
1148                 dout("aio_read %p %llx.%llx %llu~%u got cap refs on %s\n",
1149                      inode, ceph_vinop(inode), iocb->ki_pos, (unsigned)len,
1150                      ceph_cap_string(got));
1151 
1152                 ret = generic_file_read_iter(iocb, to);
1153         }
1154         dout("aio_read %p %llx.%llx dropping cap refs on %s = %d\n",
1155              inode, ceph_vinop(inode), ceph_cap_string(got), (int)ret);
1156         if (pinned_page) {
1157                 page_cache_release(pinned_page);
1158                 pinned_page = NULL;
1159         }
1160         ceph_put_cap_refs(ci, got);
1161         if (retry_op > HAVE_RETRIED && ret >= 0) {
1162                 int statret;
1163                 struct page *page = NULL;
1164                 loff_t i_size;
1165                 if (retry_op == READ_INLINE) {
1166                         page = __page_cache_alloc(GFP_KERNEL);
1167                         if (!page)
1168                                 return -ENOMEM;
1169                 }
1170 
1171                 statret = __ceph_do_getattr(inode, page,
1172                                             CEPH_STAT_CAP_INLINE_DATA, !!page);
1173                 if (statret < 0) {
1174                          __free_page(page);
1175                         if (statret == -ENODATA) {
1176                                 BUG_ON(retry_op != READ_INLINE);
1177                                 goto again;
1178                         }
1179                         return statret;
1180                 }
1181 
1182                 i_size = i_size_read(inode);
1183                 if (retry_op == READ_INLINE) {
1184                         BUG_ON(ret > 0 || read > 0);
1185                         if (iocb->ki_pos < i_size &&
1186                             iocb->ki_pos < PAGE_CACHE_SIZE) {
1187                                 loff_t end = min_t(loff_t, i_size,
1188                                                    iocb->ki_pos + len);
1189                                 end = min_t(loff_t, end, PAGE_CACHE_SIZE);
1190                                 if (statret < end)
1191                                         zero_user_segment(page, statret, end);
1192                                 ret = copy_page_to_iter(page,
1193                                                 iocb->ki_pos & ~PAGE_MASK,
1194                                                 end - iocb->ki_pos, to);
1195                                 iocb->ki_pos += ret;
1196                                 read += ret;
1197                         }
1198                         if (iocb->ki_pos < i_size && read < len) {
1199                                 size_t zlen = min_t(size_t, len - read,
1200                                                     i_size - iocb->ki_pos);
1201                                 ret = iov_iter_zero(zlen, to);
1202                                 iocb->ki_pos += ret;
1203                                 read += ret;
1204                         }
1205                         __free_pages(page, 0);
1206                         return read;
1207                 }
1208 
1209                 /* hit EOF or hole? */
1210                 if (retry_op == CHECK_EOF && iocb->ki_pos < i_size &&
1211                     ret < len) {
1212                         dout("sync_read hit hole, ppos %lld < size %lld"
1213                              ", reading more\n", iocb->ki_pos, i_size);
1214 
1215                         read += ret;
1216                         len -= ret;
1217                         retry_op = HAVE_RETRIED;
1218                         goto again;
1219                 }
1220         }
1221 
1222         if (ret >= 0)
1223                 ret += read;
1224 
1225         return ret;
1226 }
1227 
1228 /*
1229  * Take cap references to avoid releasing caps to MDS mid-write.
1230  *
1231  * If we are synchronous, and write with an old snap context, the OSD
1232  * may return EOLDSNAPC.  In that case, retry the write.. _after_
1233  * dropping our cap refs and allowing the pending snap to logically
1234  * complete _before_ this write occurs.
1235  *
1236  * If we are near ENOSPC, write synchronously.
1237  */
1238 static ssize_t ceph_write_iter(struct kiocb *iocb, struct iov_iter *from)
1239 {
1240         struct file *file = iocb->ki_filp;
1241         struct ceph_file_info *fi = file->private_data;
1242         struct inode *inode = file_inode(file);
1243         struct ceph_inode_info *ci = ceph_inode(inode);
1244         struct ceph_osd_client *osdc =
1245                 &ceph_sb_to_client(inode->i_sb)->client->osdc;
1246         struct ceph_cap_flush *prealloc_cf;
1247         ssize_t count, written = 0;
1248         int err, want, got;
1249         loff_t pos;
1250 
1251         if (ceph_snap(inode) != CEPH_NOSNAP)
1252                 return -EROFS;
1253 
1254         prealloc_cf = ceph_alloc_cap_flush();
1255         if (!prealloc_cf)
1256                 return -ENOMEM;
1257 
1258         inode_lock(inode);
1259 
1260         /* We can write back this queue in page reclaim */
1261         current->backing_dev_info = inode_to_bdi(inode);
1262 
1263         if (iocb->ki_flags & IOCB_APPEND) {
1264                 err = ceph_do_getattr(inode, CEPH_STAT_CAP_SIZE, false);
1265                 if (err < 0)
1266                         goto out;
1267         }
1268 
1269         err = generic_write_checks(iocb, from);
1270         if (err <= 0)
1271                 goto out;
1272 
1273         pos = iocb->ki_pos;
1274         count = iov_iter_count(from);
1275         err = file_remove_privs(file);
1276         if (err)
1277                 goto out;
1278 
1279         err = file_update_time(file);
1280         if (err)
1281                 goto out;
1282 
1283         if (ci->i_inline_version != CEPH_INLINE_NONE) {
1284                 err = ceph_uninline_data(file, NULL);
1285                 if (err < 0)
1286                         goto out;
1287         }
1288 
1289 retry_snap:
1290         if (ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_FULL)) {
1291                 err = -ENOSPC;
1292                 goto out;
1293         }
1294 
1295         dout("aio_write %p %llx.%llx %llu~%zd getting caps. i_size %llu\n",
1296              inode, ceph_vinop(inode), pos, count, i_size_read(inode));
1297         if (fi->fmode & CEPH_FILE_MODE_LAZY)
1298                 want = CEPH_CAP_FILE_BUFFER | CEPH_CAP_FILE_LAZYIO;
1299         else
1300                 want = CEPH_CAP_FILE_BUFFER;
1301         got = 0;
1302         err = ceph_get_caps(ci, CEPH_CAP_FILE_WR, want, pos + count,
1303                             &got, NULL);
1304         if (err < 0)
1305                 goto out;
1306 
1307         dout("aio_write %p %llx.%llx %llu~%zd got cap refs on %s\n",
1308              inode, ceph_vinop(inode), pos, count, ceph_cap_string(got));
1309 
1310         if ((got & (CEPH_CAP_FILE_BUFFER|CEPH_CAP_FILE_LAZYIO)) == 0 ||
1311             (iocb->ki_flags & IOCB_DIRECT) || (fi->flags & CEPH_F_SYNC)) {
1312                 struct ceph_snap_context *snapc;
1313                 struct iov_iter data;
1314                 inode_unlock(inode);
1315 
1316                 spin_lock(&ci->i_ceph_lock);
1317                 if (__ceph_have_pending_cap_snap(ci)) {
1318                         struct ceph_cap_snap *capsnap =
1319                                         list_last_entry(&ci->i_cap_snaps,
1320                                                         struct ceph_cap_snap,
1321                                                         ci_item);
1322                         snapc = ceph_get_snap_context(capsnap->context);
1323                 } else {
1324                         BUG_ON(!ci->i_head_snapc);
1325                         snapc = ceph_get_snap_context(ci->i_head_snapc);
1326                 }
1327                 spin_unlock(&ci->i_ceph_lock);
1328 
1329                 /* we might need to revert back to that point */
1330                 data = *from;
1331                 if (iocb->ki_flags & IOCB_DIRECT)
1332                         written = ceph_direct_read_write(iocb, &data, snapc,
1333                                                          &prealloc_cf);
1334                 else
1335                         written = ceph_sync_write(iocb, &data, pos, snapc);
1336                 if (written == -EOLDSNAPC) {
1337                         dout("aio_write %p %llx.%llx %llu~%u"
1338                                 "got EOLDSNAPC, retrying\n",
1339                                 inode, ceph_vinop(inode),
1340                                 pos, (unsigned)count);
1341                         inode_lock(inode);
1342                         goto retry_snap;
1343                 }
1344                 if (written > 0)
1345                         iov_iter_advance(from, written);
1346                 ceph_put_snap_context(snapc);
1347         } else {
1348                 loff_t old_size = i_size_read(inode);
1349                 /*
1350                  * No need to acquire the i_truncate_mutex. Because
1351                  * the MDS revokes Fwb caps before sending truncate
1352                  * message to us. We can't get Fwb cap while there
1353                  * are pending vmtruncate. So write and vmtruncate
1354                  * can not run at the same time
1355                  */
1356                 written = generic_perform_write(file, from, pos);
1357                 if (likely(written >= 0))
1358                         iocb->ki_pos = pos + written;
1359                 if (i_size_read(inode) > old_size)
1360                         ceph_fscache_update_objectsize(inode);
1361                 inode_unlock(inode);
1362         }
1363 
1364         if (written >= 0) {
1365                 int dirty;
1366                 spin_lock(&ci->i_ceph_lock);
1367                 ci->i_inline_version = CEPH_INLINE_NONE;
1368                 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR,
1369                                                &prealloc_cf);
1370                 spin_unlock(&ci->i_ceph_lock);
1371                 if (dirty)
1372                         __mark_inode_dirty(inode, dirty);
1373         }
1374 
1375         dout("aio_write %p %llx.%llx %llu~%u  dropping cap refs on %s\n",
1376              inode, ceph_vinop(inode), pos, (unsigned)count,
1377              ceph_cap_string(got));
1378         ceph_put_cap_refs(ci, got);
1379 
1380         if (written >= 0 &&
1381             ((file->f_flags & O_SYNC) || IS_SYNC(file->f_mapping->host) ||
1382              ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_NEARFULL))) {
1383                 err = vfs_fsync_range(file, pos, pos + written - 1, 1);
1384                 if (err < 0)
1385                         written = err;
1386         }
1387 
1388         goto out_unlocked;
1389 
1390 out:
1391         inode_unlock(inode);
1392 out_unlocked:
1393         ceph_free_cap_flush(prealloc_cf);
1394         current->backing_dev_info = NULL;
1395         return written ? written : err;
1396 }
1397 
1398 /*
1399  * llseek.  be sure to verify file size on SEEK_END.
1400  */
1401 static loff_t ceph_llseek(struct file *file, loff_t offset, int whence)
1402 {
1403         struct inode *inode = file->f_mapping->host;
1404         loff_t i_size;
1405         int ret;
1406 
1407         inode_lock(inode);
1408 
1409         if (whence == SEEK_END || whence == SEEK_DATA || whence == SEEK_HOLE) {
1410                 ret = ceph_do_getattr(inode, CEPH_STAT_CAP_SIZE, false);
1411                 if (ret < 0) {
1412                         offset = ret;
1413                         goto out;
1414                 }
1415         }
1416 
1417         i_size = i_size_read(inode);
1418         switch (whence) {
1419         case SEEK_END:
1420                 offset += i_size;
1421                 break;
1422         case SEEK_CUR:
1423                 /*
1424                  * Here we special-case the lseek(fd, 0, SEEK_CUR)
1425                  * position-querying operation.  Avoid rewriting the "same"
1426                  * f_pos value back to the file because a concurrent read(),
1427                  * write() or lseek() might have altered it
1428                  */
1429                 if (offset == 0) {
1430                         offset = file->f_pos;
1431                         goto out;
1432                 }
1433                 offset += file->f_pos;
1434                 break;
1435         case SEEK_DATA:
1436                 if (offset >= i_size) {
1437                         ret = -ENXIO;
1438                         goto out;
1439                 }
1440                 break;
1441         case SEEK_HOLE:
1442                 if (offset >= i_size) {
1443                         ret = -ENXIO;
1444                         goto out;
1445                 }
1446                 offset = i_size;
1447                 break;
1448         }
1449 
1450         offset = vfs_setpos(file, offset, inode->i_sb->s_maxbytes);
1451 
1452 out:
1453         inode_unlock(inode);
1454         return offset;
1455 }
1456 
1457 static inline void ceph_zero_partial_page(
1458         struct inode *inode, loff_t offset, unsigned size)
1459 {
1460         struct page *page;
1461         pgoff_t index = offset >> PAGE_CACHE_SHIFT;
1462 
1463         page = find_lock_page(inode->i_mapping, index);
1464         if (page) {
1465                 wait_on_page_writeback(page);
1466                 zero_user(page, offset & (PAGE_CACHE_SIZE - 1), size);
1467                 unlock_page(page);
1468                 page_cache_release(page);
1469         }
1470 }
1471 
1472 static void ceph_zero_pagecache_range(struct inode *inode, loff_t offset,
1473                                       loff_t length)
1474 {
1475         loff_t nearly = round_up(offset, PAGE_CACHE_SIZE);
1476         if (offset < nearly) {
1477                 loff_t size = nearly - offset;
1478                 if (length < size)
1479                         size = length;
1480                 ceph_zero_partial_page(inode, offset, size);
1481                 offset += size;
1482                 length -= size;
1483         }
1484         if (length >= PAGE_CACHE_SIZE) {
1485                 loff_t size = round_down(length, PAGE_CACHE_SIZE);
1486                 truncate_pagecache_range(inode, offset, offset + size - 1);
1487                 offset += size;
1488                 length -= size;
1489         }
1490         if (length)
1491                 ceph_zero_partial_page(inode, offset, length);
1492 }
1493 
1494 static int ceph_zero_partial_object(struct inode *inode,
1495                                     loff_t offset, loff_t *length)
1496 {
1497         struct ceph_inode_info *ci = ceph_inode(inode);
1498         struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1499         struct ceph_osd_request *req;
1500         int ret = 0;
1501         loff_t zero = 0;
1502         int op;
1503 
1504         if (!length) {
1505                 op = offset ? CEPH_OSD_OP_DELETE : CEPH_OSD_OP_TRUNCATE;
1506                 length = &zero;
1507         } else {
1508                 op = CEPH_OSD_OP_ZERO;
1509         }
1510 
1511         req = ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
1512                                         ceph_vino(inode),
1513                                         offset, length,
1514                                         0, 1, op,
1515                                         CEPH_OSD_FLAG_WRITE |
1516                                         CEPH_OSD_FLAG_ONDISK,
1517                                         NULL, 0, 0, false);
1518         if (IS_ERR(req)) {
1519                 ret = PTR_ERR(req);
1520                 goto out;
1521         }
1522 
1523         ceph_osdc_build_request(req, offset, NULL, ceph_vino(inode).snap,
1524                                 &inode->i_mtime);
1525 
1526         ret = ceph_osdc_start_request(&fsc->client->osdc, req, false);
1527         if (!ret) {
1528                 ret = ceph_osdc_wait_request(&fsc->client->osdc, req);
1529                 if (ret == -ENOENT)
1530                         ret = 0;
1531         }
1532         ceph_osdc_put_request(req);
1533 
1534 out:
1535         return ret;
1536 }
1537 
1538 static int ceph_zero_objects(struct inode *inode, loff_t offset, loff_t length)
1539 {
1540         int ret = 0;
1541         struct ceph_inode_info *ci = ceph_inode(inode);
1542         s32 stripe_unit = ceph_file_layout_su(ci->i_layout);
1543         s32 stripe_count = ceph_file_layout_stripe_count(ci->i_layout);
1544         s32 object_size = ceph_file_layout_object_size(ci->i_layout);
1545         u64 object_set_size = object_size * stripe_count;
1546         u64 nearly, t;
1547 
1548         /* round offset up to next period boundary */
1549         nearly = offset + object_set_size - 1;
1550         t = nearly;
1551         nearly -= do_div(t, object_set_size);
1552 
1553         while (length && offset < nearly) {
1554                 loff_t size = length;
1555                 ret = ceph_zero_partial_object(inode, offset, &size);
1556                 if (ret < 0)
1557                         return ret;
1558                 offset += size;
1559                 length -= size;
1560         }
1561         while (length >= object_set_size) {
1562                 int i;
1563                 loff_t pos = offset;
1564                 for (i = 0; i < stripe_count; ++i) {
1565                         ret = ceph_zero_partial_object(inode, pos, NULL);
1566                         if (ret < 0)
1567                                 return ret;
1568                         pos += stripe_unit;
1569                 }
1570                 offset += object_set_size;
1571                 length -= object_set_size;
1572         }
1573         while (length) {
1574                 loff_t size = length;
1575                 ret = ceph_zero_partial_object(inode, offset, &size);
1576                 if (ret < 0)
1577                         return ret;
1578                 offset += size;
1579                 length -= size;
1580         }
1581         return ret;
1582 }
1583 
1584 static long ceph_fallocate(struct file *file, int mode,
1585                                 loff_t offset, loff_t length)
1586 {
1587         struct ceph_file_info *fi = file->private_data;
1588         struct inode *inode = file_inode(file);
1589         struct ceph_inode_info *ci = ceph_inode(inode);
1590         struct ceph_osd_client *osdc =
1591                 &ceph_inode_to_client(inode)->client->osdc;
1592         struct ceph_cap_flush *prealloc_cf;
1593         int want, got = 0;
1594         int dirty;
1595         int ret = 0;
1596         loff_t endoff = 0;
1597         loff_t size;
1598 
1599         if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
1600                 return -EOPNOTSUPP;
1601 
1602         if (!S_ISREG(inode->i_mode))
1603                 return -EOPNOTSUPP;
1604 
1605         prealloc_cf = ceph_alloc_cap_flush();
1606         if (!prealloc_cf)
1607                 return -ENOMEM;
1608 
1609         inode_lock(inode);
1610 
1611         if (ceph_snap(inode) != CEPH_NOSNAP) {
1612                 ret = -EROFS;
1613                 goto unlock;
1614         }
1615 
1616         if (ceph_osdmap_flag(osdc->osdmap, CEPH_OSDMAP_FULL) &&
1617                 !(mode & FALLOC_FL_PUNCH_HOLE)) {
1618                 ret = -ENOSPC;
1619                 goto unlock;
1620         }
1621 
1622         if (ci->i_inline_version != CEPH_INLINE_NONE) {
1623                 ret = ceph_uninline_data(file, NULL);
1624                 if (ret < 0)
1625                         goto unlock;
1626         }
1627 
1628         size = i_size_read(inode);
1629         if (!(mode & FALLOC_FL_KEEP_SIZE))
1630                 endoff = offset + length;
1631 
1632         if (fi->fmode & CEPH_FILE_MODE_LAZY)
1633                 want = CEPH_CAP_FILE_BUFFER | CEPH_CAP_FILE_LAZYIO;
1634         else
1635                 want = CEPH_CAP_FILE_BUFFER;
1636 
1637         ret = ceph_get_caps(ci, CEPH_CAP_FILE_WR, want, endoff, &got, NULL);
1638         if (ret < 0)
1639                 goto unlock;
1640 
1641         if (mode & FALLOC_FL_PUNCH_HOLE) {
1642                 if (offset < size)
1643                         ceph_zero_pagecache_range(inode, offset, length);
1644                 ret = ceph_zero_objects(inode, offset, length);
1645         } else if (endoff > size) {
1646                 truncate_pagecache_range(inode, size, -1);
1647                 if (ceph_inode_set_size(inode, endoff))
1648                         ceph_check_caps(ceph_inode(inode),
1649                                 CHECK_CAPS_AUTHONLY, NULL);
1650         }
1651 
1652         if (!ret) {
1653                 spin_lock(&ci->i_ceph_lock);
1654                 ci->i_inline_version = CEPH_INLINE_NONE;
1655                 dirty = __ceph_mark_dirty_caps(ci, CEPH_CAP_FILE_WR,
1656                                                &prealloc_cf);
1657                 spin_unlock(&ci->i_ceph_lock);
1658                 if (dirty)
1659                         __mark_inode_dirty(inode, dirty);
1660         }
1661 
1662         ceph_put_cap_refs(ci, got);
1663 unlock:
1664         inode_unlock(inode);
1665         ceph_free_cap_flush(prealloc_cf);
1666         return ret;
1667 }
1668 
1669 const struct file_operations ceph_file_fops = {
1670         .open = ceph_open,
1671         .release = ceph_release,
1672         .llseek = ceph_llseek,
1673         .read_iter = ceph_read_iter,
1674         .write_iter = ceph_write_iter,
1675         .mmap = ceph_mmap,
1676         .fsync = ceph_fsync,
1677         .lock = ceph_lock,
1678         .flock = ceph_flock,
1679         .splice_read = generic_file_splice_read,
1680         .splice_write = iter_file_splice_write,
1681         .unlocked_ioctl = ceph_ioctl,
1682         .compat_ioctl   = ceph_ioctl,
1683         .fallocate      = ceph_fallocate,
1684 };
1685 
1686 

~ [ source navigation ] ~ [ diff markup ] ~ [ identifier search ] ~

kernel.org | git.kernel.org | LWN.net | Project Home | Wiki (Japanese) | Wiki (English) | SVN repository | Mail admin

Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.

osdn.jp