1 #include <linux/ceph/ceph_debug.h>
3 #include <linux/backing-dev.h>
6 #include <linux/pagemap.h>
7 #include <linux/writeback.h> /* generic_writepages */
8 #include <linux/slab.h>
9 #include <linux/pagevec.h>
10 #include <linux/task_io_accounting_ops.h>
13 #include "mds_client.h"
14 #include <linux/ceph/osd_client.h>
17 * Ceph address space ops.
19 * There are a few funny things going on here.
21 * The page->private field is used to reference a struct
22 * ceph_snap_context for _every_ dirty page. This indicates which
23 * snapshot the page was logically dirtied in, and thus which snap
24 * context needs to be associated with the osd write during writeback.
26 * Similarly, struct ceph_inode_info maintains a set of counters to
27 * count dirty pages on the inode. In the absence of snapshots,
28 * i_wrbuffer_ref == i_wrbuffer_ref_head == the dirty page count.
30 * When a snapshot is taken (that is, when the client receives
31 * notification that a snapshot was taken), each inode with caps and
32 * with dirty pages (dirty pages implies there is a cap) gets a new
33 * ceph_cap_snap in the i_cap_snaps list (which is sorted in ascending
34 * order, new snaps go to the tail). The i_wrbuffer_ref_head count is
35 * moved to capsnap->dirty. (Unless a sync write is currently in
36 * progress. In that case, the capsnap is said to be "pending", new
37 * writes cannot start, and the capsnap isn't "finalized" until the
38 * write completes (or fails) and a final size/mtime for the inode for
39 * that snap can be settled upon.) i_wrbuffer_ref_head is reset to 0.
41 * On writeback, we must submit writes to the osd IN SNAP ORDER. So,
42 * we look for the first capsnap in i_cap_snaps and write out pages in
43 * that snap context _only_. Then we move on to the next capsnap,
44 * eventually reaching the "live" or "head" context (i.e., pages that
45 * are not yet snapped) and are writing the most recently dirtied
48 * Invalidate and so forth must take care to ensure the dirty page
49 * accounting is preserved.
52 #define CONGESTION_ON_THRESH(congestion_kb) (congestion_kb >> (PAGE_SHIFT-10))
53 #define CONGESTION_OFF_THRESH(congestion_kb) \
54 (CONGESTION_ON_THRESH(congestion_kb) - \
55 (CONGESTION_ON_THRESH(congestion_kb) >> 2))
57 static inline struct ceph_snap_context *page_snap_context(struct page *page)
59 if (PagePrivate(page))
60 return (void *)page->private;
65 * Dirty a page. Optimistically adjust accounting, on the assumption
66 * that we won't race with invalidate. If we do, readjust.
68 static int ceph_set_page_dirty(struct page *page)
70 struct address_space *mapping = page->mapping;
72 struct ceph_inode_info *ci;
74 struct ceph_snap_context *snapc;
76 if (unlikely(!mapping))
77 return !TestSetPageDirty(page);
79 if (TestSetPageDirty(page)) {
80 dout("%p set_page_dirty %p idx %lu -- already dirty\n",
81 mapping->host, page, page->index);
85 inode = mapping->host;
86 ci = ceph_inode(inode);
89 * Note that we're grabbing a snapc ref here without holding
92 snapc = ceph_get_snap_context(ci->i_snap_realm->cached_context);
95 spin_lock(&ci->i_ceph_lock);
96 if (ci->i_head_snapc == NULL)
97 ci->i_head_snapc = ceph_get_snap_context(snapc);
98 ++ci->i_wrbuffer_ref_head;
99 if (ci->i_wrbuffer_ref == 0)
101 ++ci->i_wrbuffer_ref;
102 dout("%p set_page_dirty %p idx %lu head %d/%d -> %d/%d "
103 "snapc %p seq %lld (%d snaps)\n",
104 mapping->host, page, page->index,
105 ci->i_wrbuffer_ref-1, ci->i_wrbuffer_ref_head-1,
106 ci->i_wrbuffer_ref, ci->i_wrbuffer_ref_head,
107 snapc, snapc->seq, snapc->num_snaps);
108 spin_unlock(&ci->i_ceph_lock);
110 /* now adjust page */
111 spin_lock_irq(&mapping->tree_lock);
112 if (page->mapping) { /* Race with truncate? */
113 WARN_ON_ONCE(!PageUptodate(page));
114 account_page_dirtied(page, page->mapping);
115 radix_tree_tag_set(&mapping->page_tree,
116 page_index(page), PAGECACHE_TAG_DIRTY);
119 * Reference snap context in page->private. Also set
120 * PagePrivate so that we get invalidatepage callback.
122 page->private = (unsigned long)snapc;
123 SetPagePrivate(page);
125 dout("ANON set_page_dirty %p (raced truncate?)\n", page);
129 spin_unlock_irq(&mapping->tree_lock);
132 /* whoops, we failed to dirty the page */
133 ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
135 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
137 BUG_ON(!PageDirty(page));
142 * If we are truncating the full page (i.e. offset == 0), adjust the
143 * dirty page counters appropriately. Only called if there is private
146 static void ceph_invalidatepage(struct page *page, unsigned long offset)
149 struct ceph_inode_info *ci;
150 struct ceph_snap_context *snapc = page_snap_context(page);
152 BUG_ON(!PageLocked(page));
153 BUG_ON(!PagePrivate(page));
154 BUG_ON(!page->mapping);
156 inode = page->mapping->host;
159 * We can get non-dirty pages here due to races between
160 * set_page_dirty and truncate_complete_page; just spit out a
161 * warning, in case we end up with accounting problems later.
163 if (!PageDirty(page))
164 pr_err("%p invalidatepage %p page not dirty\n", inode, page);
167 ClearPageChecked(page);
169 ci = ceph_inode(inode);
171 dout("%p invalidatepage %p idx %lu full dirty page %lu\n",
172 inode, page, page->index, offset);
173 ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
174 ceph_put_snap_context(snapc);
176 ClearPagePrivate(page);
178 dout("%p invalidatepage %p idx %lu partial dirty page\n",
179 inode, page, page->index);
183 /* just a sanity check */
184 static int ceph_releasepage(struct page *page, gfp_t g)
186 struct inode *inode = page->mapping ? page->mapping->host : NULL;
187 dout("%p releasepage %p idx %lu\n", inode, page, page->index);
188 WARN_ON(PageDirty(page));
189 WARN_ON(PagePrivate(page));
194 * read a single page, without unlocking it.
196 static int readpage_nounlock(struct file *filp, struct page *page)
198 struct inode *inode = file_inode(filp);
199 struct ceph_inode_info *ci = ceph_inode(inode);
200 struct ceph_osd_client *osdc =
201 &ceph_inode_to_client(inode)->client->osdc;
203 u64 len = PAGE_CACHE_SIZE;
205 dout("readpage inode %p file %p page %p index %lu\n",
206 inode, filp, page, page->index);
207 err = ceph_osdc_readpages(osdc, ceph_vino(inode), &ci->i_layout,
208 (u64) page_offset(page), &len,
209 ci->i_truncate_seq, ci->i_truncate_size,
216 } else if (err < PAGE_CACHE_SIZE) {
217 /* zero fill remainder of page */
218 zero_user_segment(page, err, PAGE_CACHE_SIZE);
220 SetPageUptodate(page);
223 return err < 0 ? err : 0;
226 static int ceph_readpage(struct file *filp, struct page *page)
228 int r = readpage_nounlock(filp, page);
234 * Finish an async read(ahead) op.
236 static void finish_read(struct ceph_osd_request *req, struct ceph_msg *msg)
238 struct inode *inode = req->r_inode;
239 int rc = req->r_result;
240 int bytes = le32_to_cpu(msg->hdr.data_len);
244 dout("finish_read %p req %p rc %d bytes %d\n", inode, req, rc, bytes);
246 /* unlock all pages, zeroing any data we didn't read */
247 BUG_ON(req->r_data_in.type != CEPH_OSD_DATA_TYPE_PAGES);
248 num_pages = calc_pages_for((u64)req->r_data_in.alignment,
249 (u64)req->r_data_in.length);
250 for (i = 0; i < num_pages; i++) {
251 struct page *page = req->r_data_in.pages[i];
253 if (bytes < (int)PAGE_CACHE_SIZE) {
254 /* zero (remainder of) page */
255 int s = bytes < 0 ? 0 : bytes;
256 zero_user_segment(page, s, PAGE_CACHE_SIZE);
258 dout("finish_read %p uptodate %p idx %lu\n", inode, page,
260 flush_dcache_page(page);
261 SetPageUptodate(page);
263 page_cache_release(page);
264 bytes -= PAGE_CACHE_SIZE;
266 kfree(req->r_data_in.pages);
269 static void ceph_unlock_page_vector(struct page **pages, int num_pages)
273 for (i = 0; i < num_pages; i++)
274 unlock_page(pages[i]);
278 * start an async read(ahead) operation. return nr_pages we submitted
279 * a read for on success, or negative error code.
281 static int start_read(struct inode *inode, struct list_head *page_list, int max)
283 struct ceph_osd_client *osdc =
284 &ceph_inode_to_client(inode)->client->osdc;
285 struct ceph_inode_info *ci = ceph_inode(inode);
286 struct page *page = list_entry(page_list->prev, struct page, lru);
287 struct ceph_vino vino;
288 struct ceph_osd_request *req;
289 struct ceph_osd_req_op op;
298 off = (u64) page_offset(page);
301 next_index = page->index;
302 list_for_each_entry_reverse(page, page_list, lru) {
303 if (page->index != next_index)
307 if (max && nr_pages == max)
310 len = nr_pages << PAGE_CACHE_SHIFT;
311 dout("start_read %p nr_pages %d is %lld~%lld\n", inode, nr_pages,
313 vino = ceph_vino(inode);
314 req = ceph_osdc_new_request(osdc, &ci->i_layout, vino, off, &len,
315 1, &op, CEPH_OSD_OP_READ,
316 CEPH_OSD_FLAG_READ, NULL,
317 ci->i_truncate_seq, ci->i_truncate_size,
322 /* build page vector */
323 nr_pages = calc_pages_for(0, len);
324 pages = kmalloc(sizeof(*pages) * nr_pages, GFP_NOFS);
328 for (i = 0; i < nr_pages; ++i) {
329 page = list_entry(page_list->prev, struct page, lru);
330 BUG_ON(PageLocked(page));
331 list_del(&page->lru);
333 dout("start_read %p adding %p idx %lu\n", inode, page,
335 if (add_to_page_cache_lru(page, &inode->i_data, page->index,
337 page_cache_release(page);
338 dout("start_read %p add_to_page_cache failed %p\n",
345 req->r_data_in.type = CEPH_OSD_DATA_TYPE_PAGES;
346 req->r_data_in.pages = pages;
347 req->r_data_in.length = len;
348 req->r_data_in.alignment = 0;
349 req->r_callback = finish_read;
350 req->r_inode = inode;
352 ceph_osdc_build_request(req, off, 1, &op, NULL, vino.snap, NULL);
354 dout("start_read %p starting %p %lld~%lld\n", inode, req, off, len);
355 ret = ceph_osdc_start_request(osdc, req, false);
358 ceph_osdc_put_request(req);
362 ceph_unlock_page_vector(pages, nr_pages);
363 ceph_release_page_vector(pages, nr_pages);
365 ceph_osdc_put_request(req);
371 * Read multiple pages. Leave pages we don't read + unlock in page_list;
372 * the caller (VM) cleans them up.
374 static int ceph_readpages(struct file *file, struct address_space *mapping,
375 struct list_head *page_list, unsigned nr_pages)
377 struct inode *inode = file_inode(file);
378 struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
382 if (fsc->mount_options->rsize >= PAGE_CACHE_SIZE)
383 max = (fsc->mount_options->rsize + PAGE_CACHE_SIZE - 1)
386 dout("readpages %p file %p nr_pages %d max %d\n", inode,
389 while (!list_empty(page_list)) {
390 rc = start_read(inode, page_list, max);
396 dout("readpages %p file %p ret %d\n", inode, file, rc);
401 * Get ref for the oldest snapc for an inode with dirty data... that is, the
402 * only snap context we are allowed to write back.
404 static struct ceph_snap_context *get_oldest_context(struct inode *inode,
407 struct ceph_inode_info *ci = ceph_inode(inode);
408 struct ceph_snap_context *snapc = NULL;
409 struct ceph_cap_snap *capsnap = NULL;
411 spin_lock(&ci->i_ceph_lock);
412 list_for_each_entry(capsnap, &ci->i_cap_snaps, ci_item) {
413 dout(" cap_snap %p snapc %p has %d dirty pages\n", capsnap,
414 capsnap->context, capsnap->dirty_pages);
415 if (capsnap->dirty_pages) {
416 snapc = ceph_get_snap_context(capsnap->context);
418 *snap_size = capsnap->size;
422 if (!snapc && ci->i_wrbuffer_ref_head) {
423 snapc = ceph_get_snap_context(ci->i_head_snapc);
424 dout(" head snapc %p has %d dirty pages\n",
425 snapc, ci->i_wrbuffer_ref_head);
427 spin_unlock(&ci->i_ceph_lock);
432 * Write a single page, but leave the page locked.
434 * If we get a write error, set the page error bit, but still adjust the
435 * dirty page accounting (i.e., page is no longer dirty).
437 static int writepage_nounlock(struct page *page, struct writeback_control *wbc)
440 struct ceph_inode_info *ci;
441 struct ceph_fs_client *fsc;
442 struct ceph_osd_client *osdc;
443 loff_t page_off = page_offset(page);
444 int len = PAGE_CACHE_SIZE;
447 struct ceph_snap_context *snapc, *oldest;
451 dout("writepage %p idx %lu\n", page, page->index);
453 if (!page->mapping || !page->mapping->host) {
454 dout("writepage %p - no mapping\n", page);
457 inode = page->mapping->host;
458 ci = ceph_inode(inode);
459 fsc = ceph_inode_to_client(inode);
460 osdc = &fsc->client->osdc;
462 /* verify this is a writeable snap context */
463 snapc = page_snap_context(page);
465 dout("writepage %p page %p not dirty?\n", inode, page);
468 oldest = get_oldest_context(inode, &snap_size);
469 if (snapc->seq > oldest->seq) {
470 dout("writepage %p page %p snapc %p not writeable - noop\n",
472 /* we should only noop if called by kswapd */
473 WARN_ON((current->flags & PF_MEMALLOC) == 0);
474 ceph_put_snap_context(oldest);
477 ceph_put_snap_context(oldest);
479 /* is this a partial page at end of file? */
483 i_size = i_size_read(inode);
484 if (i_size < page_off + len)
485 len = i_size - page_off;
487 dout("writepage %p page %p index %lu on %llu~%u snapc %p\n",
488 inode, page, page->index, page_off, len, snapc);
490 writeback_stat = atomic_long_inc_return(&fsc->writeback_count);
492 CONGESTION_ON_THRESH(fsc->mount_options->congestion_kb))
493 set_bdi_congested(&fsc->backing_dev_info, BLK_RW_ASYNC);
495 set_page_writeback(page);
496 err = ceph_osdc_writepages(osdc, ceph_vino(inode),
497 &ci->i_layout, snapc,
499 ci->i_truncate_seq, ci->i_truncate_size,
500 &inode->i_mtime, &page, 1);
502 dout("writepage setting page/mapping error %d %p\n", err, page);
504 mapping_set_error(&inode->i_data, err);
506 wbc->pages_skipped++;
508 dout("writepage cleaned page %p\n", page);
509 err = 0; /* vfs expects us to return 0 */
512 ClearPagePrivate(page);
513 end_page_writeback(page);
514 ceph_put_wrbuffer_cap_refs(ci, 1, snapc);
515 ceph_put_snap_context(snapc); /* page's reference */
520 static int ceph_writepage(struct page *page, struct writeback_control *wbc)
523 struct inode *inode = page->mapping->host;
526 err = writepage_nounlock(page, wbc);
534 * lame release_pages helper. release_pages() isn't exported to
537 static void ceph_release_pages(struct page **pages, int num)
542 pagevec_init(&pvec, 0);
543 for (i = 0; i < num; i++) {
544 if (pagevec_add(&pvec, pages[i]) == 0)
545 pagevec_release(&pvec);
547 pagevec_release(&pvec);
552 * async writeback completion handler.
554 * If we get an error, set the mapping error bit, but not the individual
557 static void writepages_finish(struct ceph_osd_request *req,
558 struct ceph_msg *msg)
560 struct inode *inode = req->r_inode;
561 struct ceph_inode_info *ci = ceph_inode(inode);
566 struct ceph_snap_context *snapc = req->r_snapc;
567 struct address_space *mapping = inode->i_mapping;
568 int rc = req->r_result;
569 u64 bytes = le64_to_cpu(req->r_request_ops[0].extent.length);
570 struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
572 unsigned issued = ceph_caps_issued(ci);
574 BUG_ON(req->r_data_out.type != CEPH_OSD_DATA_TYPE_PAGES);
575 num_pages = calc_pages_for((u64)req->r_data_out.alignment,
576 (u64)req->r_data_out.length);
579 * Assume we wrote the pages we originally sent. The
580 * osd might reply with fewer pages if our writeback
581 * raced with a truncation and was adjusted at the osd,
582 * so don't believe the reply.
587 mapping_set_error(mapping, rc);
589 dout("writepages_finish %p rc %d bytes %llu wrote %d (pages)\n",
590 inode, rc, bytes, wrote);
592 /* clean all pages */
593 for (i = 0; i < num_pages; i++) {
594 page = req->r_data_out.pages[i];
596 WARN_ON(!PageUptodate(page));
599 atomic_long_dec_return(&fsc->writeback_count);
601 CONGESTION_OFF_THRESH(fsc->mount_options->congestion_kb))
602 clear_bdi_congested(&fsc->backing_dev_info,
605 ceph_put_snap_context(page_snap_context(page));
607 ClearPagePrivate(page);
608 dout("unlocking %d %p\n", i, page);
609 end_page_writeback(page);
612 * We lost the cache cap, need to truncate the page before
613 * it is unlocked, otherwise we'd truncate it later in the
614 * page truncation thread, possibly losing some data that
617 if ((issued & (CEPH_CAP_FILE_CACHE|CEPH_CAP_FILE_LAZYIO)) == 0)
618 generic_error_remove_page(inode->i_mapping, page);
622 dout("%p wrote+cleaned %d pages\n", inode, wrote);
623 ceph_put_wrbuffer_cap_refs(ci, num_pages, snapc);
625 ceph_release_pages(req->r_data_out.pages, num_pages);
626 if (req->r_data_out.pages_from_pool)
627 mempool_free(req->r_data_out.pages,
628 ceph_sb_to_client(inode->i_sb)->wb_pagevec_pool);
630 kfree(req->r_data_out.pages);
631 ceph_osdc_put_request(req);
634 static struct ceph_osd_request *
635 ceph_writepages_osd_request(struct inode *inode, u64 offset, u64 *len,
636 struct ceph_snap_context *snapc,
637 int num_ops, struct ceph_osd_req_op *ops)
639 struct ceph_fs_client *fsc;
640 struct ceph_inode_info *ci;
641 struct ceph_vino vino;
643 fsc = ceph_inode_to_client(inode);
644 ci = ceph_inode(inode);
645 vino = ceph_vino(inode);
646 /* BUG_ON(vino.snap != CEPH_NOSNAP); */
648 return ceph_osdc_new_request(&fsc->client->osdc, &ci->i_layout,
649 vino, offset, len, num_ops, ops, CEPH_OSD_OP_WRITE,
650 CEPH_OSD_FLAG_WRITE|CEPH_OSD_FLAG_ONDISK,
651 snapc, ci->i_truncate_seq, ci->i_truncate_size, true);
655 * initiate async writeback
657 static int ceph_writepages_start(struct address_space *mapping,
658 struct writeback_control *wbc)
660 struct inode *inode = mapping->host;
661 struct ceph_inode_info *ci = ceph_inode(inode);
662 struct ceph_fs_client *fsc;
663 pgoff_t index, start, end;
666 pgoff_t max_pages = 0, max_pages_ever = 0;
667 struct ceph_snap_context *snapc = NULL, *last_snapc = NULL, *pgsnapc;
671 unsigned wsize = 1 << inode->i_blkbits;
672 struct ceph_osd_request *req = NULL;
677 * Include a 'sync' in the OSD request if this is a data
678 * integrity write (e.g., O_SYNC write or fsync()), or if our
679 * cap is being revoked.
681 do_sync = wbc->sync_mode == WB_SYNC_ALL;
682 if (ceph_caps_revoking(ci, CEPH_CAP_FILE_BUFFER))
684 dout("writepages_start %p dosync=%d (mode=%s)\n",
686 wbc->sync_mode == WB_SYNC_NONE ? "NONE" :
687 (wbc->sync_mode == WB_SYNC_ALL ? "ALL" : "HOLD"));
689 fsc = ceph_inode_to_client(inode);
690 if (fsc->mount_state == CEPH_MOUNT_SHUTDOWN) {
691 pr_warning("writepage_start %p on forced umount\n", inode);
692 return -EIO; /* we're in a forced umount, don't write! */
694 if (fsc->mount_options->wsize && fsc->mount_options->wsize < wsize)
695 wsize = fsc->mount_options->wsize;
696 if (wsize < PAGE_CACHE_SIZE)
697 wsize = PAGE_CACHE_SIZE;
698 max_pages_ever = wsize >> PAGE_CACHE_SHIFT;
700 pagevec_init(&pvec, 0);
702 /* where to start/end? */
703 if (wbc->range_cyclic) {
704 start = mapping->writeback_index; /* Start from prev offset */
706 dout(" cyclic, start at %lu\n", start);
708 start = wbc->range_start >> PAGE_CACHE_SHIFT;
709 end = wbc->range_end >> PAGE_CACHE_SHIFT;
710 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
713 dout(" not cyclic, %lu to %lu\n", start, end);
718 /* find oldest snap context with dirty data */
719 ceph_put_snap_context(snapc);
720 snapc = get_oldest_context(inode, &snap_size);
722 /* hmm, why does writepages get called when there
724 dout(" no snap context with dirty data?\n");
727 dout(" oldest snapc is %p seq %lld (%d snaps)\n",
728 snapc, snapc->seq, snapc->num_snaps);
729 if (last_snapc && snapc != last_snapc) {
730 /* if we switched to a newer snapc, restart our scan at the
731 * start of the original file range. */
732 dout(" snapc differs from last pass, restarting at %lu\n",
738 while (!done && index <= end) {
739 struct ceph_osd_req_op ops[2];
740 int num_ops = do_sync ? 2 : 1;
741 struct ceph_vino vino;
745 int pvec_pages, locked_pages;
746 struct page **pages = NULL;
747 mempool_t *pool = NULL; /* Becomes non-null if mempool used */
755 max_pages = max_pages_ever;
759 want = min(end - index,
760 min((pgoff_t)PAGEVEC_SIZE,
761 max_pages - (pgoff_t)locked_pages) - 1)
763 pvec_pages = pagevec_lookup_tag(&pvec, mapping, &index,
766 dout("pagevec_lookup_tag got %d\n", pvec_pages);
767 if (!pvec_pages && !locked_pages)
769 for (i = 0; i < pvec_pages && locked_pages < max_pages; i++) {
770 page = pvec.pages[i];
771 dout("? %p idx %lu\n", page, page->index);
772 if (locked_pages == 0)
773 lock_page(page); /* first page */
774 else if (!trylock_page(page))
777 /* only dirty pages, or our accounting breaks */
778 if (unlikely(!PageDirty(page)) ||
779 unlikely(page->mapping != mapping)) {
780 dout("!dirty or !mapping %p\n", page);
784 if (!wbc->range_cyclic && page->index > end) {
785 dout("end of range %p\n", page);
790 if (next && (page->index != next)) {
791 dout("not consecutive %p\n", page);
795 if (wbc->sync_mode != WB_SYNC_NONE) {
796 dout("waiting on writeback %p\n", page);
797 wait_on_page_writeback(page);
799 if ((snap_size && page_offset(page) > snap_size) ||
801 page_offset(page) > i_size_read(inode))) {
802 dout("%p page eof %llu\n", page, snap_size ?
803 snap_size : i_size_read(inode));
808 if (PageWriteback(page)) {
809 dout("%p under writeback\n", page);
814 /* only if matching snap context */
815 pgsnapc = page_snap_context(page);
816 if (pgsnapc->seq > snapc->seq) {
817 dout("page snapc %p %lld > oldest %p %lld\n",
818 pgsnapc, pgsnapc->seq, snapc, snapc->seq);
821 continue; /* keep looking for snap */
825 if (!clear_page_dirty_for_io(page)) {
826 dout("%p !clear_page_dirty_for_io\n", page);
832 * We have something to write. If this is
833 * the first locked page this time through,
834 * allocate an osd request and a page array
837 if (locked_pages == 0) {
842 /* prepare async write request */
843 offset = (u64)page_offset(page);
845 req = ceph_writepages_osd_request(inode,
855 req->r_callback = writepages_finish;
856 req->r_inode = inode;
858 max_pages = calc_pages_for(0, (u64)len);
859 size = max_pages * sizeof (*pages);
860 pages = kmalloc(size, GFP_NOFS);
862 pool = fsc->wb_pagevec_pool;
863 pages = mempool_alloc(pool, GFP_NOFS);
868 /* note position of first page in pvec */
871 dout("%p will write page %p idx %lu\n",
872 inode, page, page->index);
875 atomic_long_inc_return(&fsc->writeback_count);
876 if (writeback_stat > CONGESTION_ON_THRESH(
877 fsc->mount_options->congestion_kb)) {
878 set_bdi_congested(&fsc->backing_dev_info,
882 set_page_writeback(page);
883 pages[locked_pages] = page;
885 next = page->index + 1;
888 /* did we get anything? */
890 goto release_pvec_pages;
893 BUG_ON(!locked_pages || first < 0);
895 if (pvec_pages && i == pvec_pages &&
896 locked_pages < max_pages) {
897 dout("reached end pvec, trying for more\n");
898 pagevec_reinit(&pvec);
902 /* shift unused pages over in the pvec... we
903 * will need to release them below. */
904 for (j = i; j < pvec_pages; j++) {
905 dout(" pvec leftover page %p\n",
907 pvec.pages[j-i+first] = pvec.pages[j];
912 /* Format the osd request message and submit the write */
914 offset = page_offset(pages[0]);
915 len = min((snap_size ? snap_size : i_size_read(inode)) - offset,
916 (u64)locked_pages << PAGE_CACHE_SHIFT);
917 dout("writepages got %d pages at %llu~%llu\n",
918 locked_pages, offset, len);
920 req->r_data_out.type = CEPH_OSD_DATA_TYPE_PAGES;
921 req->r_data_out.pages = pages;
922 req->r_data_out.length = len;
923 req->r_data_out.alignment = 0;
924 req->r_data_out.pages_from_pool = !!pool;
926 pages = NULL; /* request message now owns the pages array */
929 /* Update the write op length in case we changed it */
931 osd_req_op_extent_update(&ops[0], len);
933 vino = ceph_vino(inode);
934 ceph_osdc_build_request(req, offset, num_ops, ops,
935 snapc, vino.snap, &inode->i_mtime);
937 rc = ceph_osdc_start_request(&fsc->client->osdc, req, true);
943 wbc->nr_to_write -= locked_pages;
944 if (wbc->nr_to_write <= 0)
948 dout("pagevec_release on %d pages (%p)\n", (int)pvec.nr,
949 pvec.nr ? pvec.pages[0] : NULL);
950 pagevec_release(&pvec);
952 if (locked_pages && !done)
956 if (should_loop && !done) {
957 /* more to do; loop back to beginning of file */
958 dout("writepages looping back to beginning of file\n");
964 if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
965 mapping->writeback_index = index;
969 ceph_osdc_put_request(req);
970 ceph_put_snap_context(snapc);
971 dout("writepages done, rc = %d\n", rc);
978 * See if a given @snapc is either writeable, or already written.
980 static int context_is_writeable_or_written(struct inode *inode,
981 struct ceph_snap_context *snapc)
983 struct ceph_snap_context *oldest = get_oldest_context(inode, NULL);
984 int ret = !oldest || snapc->seq <= oldest->seq;
986 ceph_put_snap_context(oldest);
991 * We are only allowed to write into/dirty the page if the page is
992 * clean, or already dirty within the same snap context.
994 * called with page locked.
995 * return success with page locked,
996 * or any failure (incl -EAGAIN) with page unlocked.
998 static int ceph_update_writeable_page(struct file *file,
999 loff_t pos, unsigned len,
1002 struct inode *inode = file_inode(file);
1003 struct ceph_inode_info *ci = ceph_inode(inode);
1004 struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
1005 loff_t page_off = pos & PAGE_CACHE_MASK;
1006 int pos_in_page = pos & ~PAGE_CACHE_MASK;
1007 int end_in_page = pos_in_page + len;
1010 struct ceph_snap_context *snapc, *oldest;
1013 /* writepages currently holds page lock, but if we change that later, */
1014 wait_on_page_writeback(page);
1016 /* check snap context */
1017 BUG_ON(!ci->i_snap_realm);
1018 down_read(&mdsc->snap_rwsem);
1019 BUG_ON(!ci->i_snap_realm->cached_context);
1020 snapc = page_snap_context(page);
1021 if (snapc && snapc != ci->i_head_snapc) {
1023 * this page is already dirty in another (older) snap
1024 * context! is it writeable now?
1026 oldest = get_oldest_context(inode, NULL);
1027 up_read(&mdsc->snap_rwsem);
1029 if (snapc->seq > oldest->seq) {
1030 ceph_put_snap_context(oldest);
1031 dout(" page %p snapc %p not current or oldest\n",
1034 * queue for writeback, and wait for snapc to
1035 * be writeable or written
1037 snapc = ceph_get_snap_context(snapc);
1039 ceph_queue_writeback(inode);
1040 r = wait_event_interruptible(ci->i_cap_wq,
1041 context_is_writeable_or_written(inode, snapc));
1042 ceph_put_snap_context(snapc);
1043 if (r == -ERESTARTSYS)
1047 ceph_put_snap_context(oldest);
1049 /* yay, writeable, do it now (without dropping page lock) */
1050 dout(" page %p snapc %p not current, but oldest\n",
1052 if (!clear_page_dirty_for_io(page))
1054 r = writepage_nounlock(page, NULL);
1060 if (PageUptodate(page)) {
1061 dout(" page %p already uptodate\n", page);
1066 if (pos_in_page == 0 && len == PAGE_CACHE_SIZE)
1069 /* past end of file? */
1070 i_size = inode->i_size; /* caller holds i_mutex */
1072 if (i_size + len > inode->i_sb->s_maxbytes) {
1073 /* file is too big */
1078 if (page_off >= i_size ||
1079 (pos_in_page == 0 && (pos+len) >= i_size &&
1080 end_in_page - pos_in_page != PAGE_CACHE_SIZE)) {
1081 dout(" zeroing %p 0 - %d and %d - %d\n",
1082 page, pos_in_page, end_in_page, (int)PAGE_CACHE_SIZE);
1083 zero_user_segments(page,
1085 end_in_page, PAGE_CACHE_SIZE);
1089 /* we need to read it. */
1090 up_read(&mdsc->snap_rwsem);
1091 r = readpage_nounlock(file, page);
1097 up_read(&mdsc->snap_rwsem);
1104 * We are only allowed to write into/dirty the page if the page is
1105 * clean, or already dirty within the same snap context.
1107 static int ceph_write_begin(struct file *file, struct address_space *mapping,
1108 loff_t pos, unsigned len, unsigned flags,
1109 struct page **pagep, void **fsdata)
1111 struct inode *inode = file_inode(file);
1113 pgoff_t index = pos >> PAGE_CACHE_SHIFT;
1118 page = grab_cache_page_write_begin(mapping, index, 0);
1123 dout("write_begin file %p inode %p page %p %d~%d\n", file,
1124 inode, page, (int)pos, (int)len);
1126 r = ceph_update_writeable_page(file, pos, len, page);
1127 } while (r == -EAGAIN);
1133 * we don't do anything in here that simple_write_end doesn't do
1134 * except adjust dirty page accounting and drop read lock on
1137 static int ceph_write_end(struct file *file, struct address_space *mapping,
1138 loff_t pos, unsigned len, unsigned copied,
1139 struct page *page, void *fsdata)
1141 struct inode *inode = file_inode(file);
1142 struct ceph_fs_client *fsc = ceph_inode_to_client(inode);
1143 struct ceph_mds_client *mdsc = fsc->mdsc;
1144 unsigned from = pos & (PAGE_CACHE_SIZE - 1);
1147 dout("write_end file %p inode %p page %p %d~%d (%d)\n", file,
1148 inode, page, (int)pos, (int)copied, (int)len);
1150 /* zero the stale part of the page if we did a short copy */
1152 zero_user_segment(page, from+copied, len);
1154 /* did file size increase? */
1155 /* (no need for i_size_read(); we caller holds i_mutex */
1156 if (pos+copied > inode->i_size)
1157 check_cap = ceph_inode_set_size(inode, pos+copied);
1159 if (!PageUptodate(page))
1160 SetPageUptodate(page);
1162 set_page_dirty(page);
1165 up_read(&mdsc->snap_rwsem);
1166 page_cache_release(page);
1169 ceph_check_caps(ceph_inode(inode), CHECK_CAPS_AUTHONLY, NULL);
1175 * we set .direct_IO to indicate direct io is supported, but since we
1176 * intercept O_DIRECT reads and writes early, this function should
1179 static ssize_t ceph_direct_io(int rw, struct kiocb *iocb,
1180 const struct iovec *iov,
1181 loff_t pos, unsigned long nr_segs)
1187 const struct address_space_operations ceph_aops = {
1188 .readpage = ceph_readpage,
1189 .readpages = ceph_readpages,
1190 .writepage = ceph_writepage,
1191 .writepages = ceph_writepages_start,
1192 .write_begin = ceph_write_begin,
1193 .write_end = ceph_write_end,
1194 .set_page_dirty = ceph_set_page_dirty,
1195 .invalidatepage = ceph_invalidatepage,
1196 .releasepage = ceph_releasepage,
1197 .direct_IO = ceph_direct_io,
1206 * Reuse write_begin here for simplicity.
1208 static int ceph_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
1210 struct inode *inode = file_inode(vma->vm_file);
1211 struct page *page = vmf->page;
1212 struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
1213 loff_t off = page_offset(page);
1217 /* Update time before taking page lock */
1218 file_update_time(vma->vm_file);
1220 size = i_size_read(inode);
1221 if (off + PAGE_CACHE_SIZE <= size)
1222 len = PAGE_CACHE_SIZE;
1224 len = size & ~PAGE_CACHE_MASK;
1226 dout("page_mkwrite %p %llu~%llu page %p idx %lu\n", inode,
1227 off, len, page, page->index);
1231 ret = VM_FAULT_NOPAGE;
1233 (page->mapping != inode->i_mapping))
1236 ret = ceph_update_writeable_page(vma->vm_file, off, len, page);
1238 /* success. we'll keep the page locked. */
1239 set_page_dirty(page);
1240 up_read(&mdsc->snap_rwsem);
1241 ret = VM_FAULT_LOCKED;
1246 ret = VM_FAULT_SIGBUS;
1249 dout("page_mkwrite %p %llu~%llu = %d\n", inode, off, len, ret);
1250 if (ret != VM_FAULT_LOCKED)
1255 static struct vm_operations_struct ceph_vmops = {
1256 .fault = filemap_fault,
1257 .page_mkwrite = ceph_page_mkwrite,
1258 .remap_pages = generic_file_remap_pages,
1261 int ceph_mmap(struct file *file, struct vm_area_struct *vma)
1263 struct address_space *mapping = file->f_mapping;
1265 if (!mapping->a_ops->readpage)
1267 file_accessed(file);
1268 vma->vm_ops = &ceph_vmops;