1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
4 * Lustre Lite I/O page cache routines shared by different kernel revs
6 * Copyright (c) 2001-2003 Cluster File Systems, Inc.
8 * This file is part of Lustre, http://www.lustre.org.
10 * Lustre is free software; you can redistribute it and/or
11 * modify it under the terms of version 2 of the GNU General Public
12 * License as published by the Free Software Foundation.
14 * Lustre is distributed in the hope that it will be useful,
15 * but WITHOUT ANY WARRANTY; without even the implied warranty of
16 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
17 * GNU General Public License for more details.
19 * You should have received a copy of the GNU General Public License
20 * along with Lustre; if not, write to the Free Software
21 * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
24 #include <linux/config.h>
25 #include <linux/kernel.h>
27 #include <linux/string.h>
28 #include <linux/stat.h>
29 #include <linux/errno.h>
30 #include <linux/smp_lock.h>
31 #include <linux/unistd.h>
32 #include <linux/version.h>
33 #include <asm/system.h>
34 #include <asm/uaccess.h>
37 #include <linux/stat.h>
38 #include <asm/uaccess.h>
39 #include <asm/segment.h>
41 #include <linux/pagemap.h>
42 #include <linux/smp_lock.h>
44 #define DEBUG_SUBSYSTEM S_LLITE
46 #include <linux/lustre_mds.h>
47 #include <linux/lustre_lite.h>
48 #include "llite_internal.h"
49 #include <linux/lustre_compat25.h>
51 #ifndef list_for_each_prev_safe
52 #define list_for_each_prev_safe(pos, n, head) \
53 for (pos = (head)->prev, n = pos->prev; pos != (head); \
54 pos = n, n = pos->prev )
57 /* SYNCHRONOUS I/O to object storage for an inode */
58 static int ll_brw(int cmd, struct inode *inode, struct obdo *oa,
59 struct page *page, int flags)
61 struct ll_inode_info *lli = ll_i2info(inode);
62 struct lov_stripe_md *lsm = lli->lli_smd;
68 pg.off = ((obd_off)page->index) << PAGE_SHIFT;
70 if (cmd == OBD_BRW_WRITE && (pg.off + PAGE_SIZE > inode->i_size))
71 pg.count = inode->i_size % PAGE_SIZE;
75 CDEBUG(D_PAGE, "%s %d bytes ino %lu at "LPU64"/"LPX64"\n",
76 cmd & OBD_BRW_WRITE ? "write" : "read", pg.count, inode->i_ino,
79 CERROR("ZERO COUNT: ino %lu: size %p:%Lu(%p:%Lu) idx %lu off "
81 inode->i_ino, inode, inode->i_size, page->mapping->host,
82 page->mapping->host->i_size, page->index, pg.off);
87 if (cmd == OBD_BRW_WRITE)
88 lprocfs_counter_add(ll_i2sbi(inode)->ll_stats,
89 LPROC_LL_BRW_WRITE, pg.count);
91 lprocfs_counter_add(ll_i2sbi(inode)->ll_stats,
92 LPROC_LL_BRW_READ, pg.count);
93 rc = obd_brw(cmd, ll_i2obdexp(inode), oa, lsm, 1, &pg, NULL);
95 obdo_to_inode(inode, oa, OBD_MD_FLBLOCKS);
97 CERROR("error from obd_brw: rc = %d\n", rc);
101 /* this isn't where truncate starts. roughly:
102 * sys_truncate->ll_setattr_raw->vmtruncate->ll_truncate
103 * we grab the lock back in setattr_raw to avoid races. */
104 void ll_truncate(struct inode *inode)
106 struct lov_stripe_md *lsm = ll_i2info(inode)->lli_smd;
110 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p)\n", inode->i_ino,
111 inode->i_generation, inode);
113 /* object not yet allocated - this is handled in ll_setattr_raw */
115 CERROR("truncate on inode %lu with no objects\n", inode->i_ino);
120 oa.o_id = lsm->lsm_object_id;
121 oa.o_valid = OBD_MD_FLID;
122 obdo_from_inode(&oa, inode, OBD_MD_FLTYPE|OBD_MD_FLMODE|OBD_MD_FLATIME|
123 OBD_MD_FLMTIME | OBD_MD_FLCTIME);
125 CDEBUG(D_INFO, "calling punch for "LPX64" (all bytes after %Lu)\n",
126 oa.o_id, inode->i_size);
128 /* truncate == punch from new size to absolute end of file */
129 rc = obd_punch(ll_i2obdexp(inode), &oa, lsm, inode->i_size,
130 OBD_OBJECT_EOF, NULL);
132 CERROR("obd_truncate fails (%d) ino %lu\n", rc, inode->i_ino);
134 obdo_to_inode(inode, &oa, OBD_MD_FLSIZE | OBD_MD_FLBLOCKS |
135 OBD_MD_FLATIME | OBD_MD_FLMTIME |
142 int ll_prepare_write(struct file *file, struct page *page, unsigned from,
145 struct inode *inode = page->mapping->host;
146 struct ll_inode_info *lli = ll_i2info(inode);
147 struct lov_stripe_md *lsm = lli->lli_smd;
148 obd_off offset = ((obd_off)page->index) << PAGE_SHIFT;
154 if (!PageLocked(page))
157 if (PageUptodate(page))
160 /* Check to see if we should return -EIO right away */
163 pg.count = PAGE_SIZE;
165 rc = obd_brw(OBD_BRW_CHECK, ll_i2obdexp(inode), NULL, lsm, 1, &pg,
170 /* We're completely overwriting an existing page, so _don't_ set it up
171 * to date until commit_write */
172 if (from == 0 && to == PAGE_SIZE) {
173 POISON_PAGE(page, 0x11);
177 /* If are writing to a new page, no need to read old data.
178 * the extent locking and getattr procedures in ll_file_write have
179 * guaranteed that i_size is stable enough for our zeroing needs */
180 if (inode->i_size <= offset) {
181 memset(kmap(page), 0, PAGE_SIZE);
183 GOTO(prepare_done, rc = 0);
186 oa.o_id = lsm->lsm_object_id;
187 oa.o_mode = inode->i_mode;
188 oa.o_valid = OBD_MD_FLID | OBD_MD_FLMODE | OBD_MD_FLTYPE;
190 /* XXX could be an async ocp read.. read-ahead? */
191 rc = ll_brw(OBD_BRW_READ, inode, &oa, page, 0);
193 /* bug 1598: don't clobber blksize */
194 oa.o_valid &= ~(OBD_MD_FLSIZE | OBD_MD_FLBLKSZ);
195 obdo_refresh_inode(inode, &oa, oa.o_valid);
201 SetPageUptodate(page);
206 int ll_write_count(struct page *page)
208 struct inode *inode = page->mapping->host;
210 /* catch race with truncate */
211 if (((loff_t)page->index << PAGE_SHIFT) >= inode->i_size)
214 /* catch sub-page write at end of file */
215 if (((loff_t)page->index << PAGE_SHIFT) + PAGE_SIZE > inode->i_size)
216 return inode->i_size % PAGE_SIZE;
221 struct ll_async_page *llap_from_cookie(void *cookie)
223 struct ll_async_page *llap = cookie;
224 if (llap->llap_magic != LLAP_MAGIC)
225 return ERR_PTR(-EINVAL);
229 static int ll_ap_make_ready(void *data, int cmd)
231 struct ll_async_page *llap;
235 /* reads are always locked between queueing and completion,
236 * llite should never queue pages without _READY */
237 LASSERT(cmd != OBD_BRW_READ);
239 llap = llap_from_cookie(data);
243 page = llap->llap_page;
245 if (TryLockPage(page))
248 LL_CDEBUG_PAGE(page, "made ready\n");
249 page_cache_get(page);
251 /* if we left PageDirty we might get another writepage call
252 * in the future. list walkers are bright enough
253 * to check page dirty so we can leave it on whatever list
254 * its on. XXX also, we're called with the cli list so if
255 * we got the page cache list we'd create a lock inversion
256 * with the removepage path which gets the page lock then the
258 clear_page_dirty(page);
262 static int ll_ap_refresh_count(void *data, int cmd)
264 struct ll_async_page *llap;
267 /* readpage queues with _COUNT_STABLE, shouldn't get here. */
268 LASSERT(cmd != OBD_BRW_READ);
270 llap = llap_from_cookie(data);
272 RETURN(PTR_ERR(llap));
274 return ll_write_count(llap->llap_page);
277 void ll_inode_fill_obdo(struct inode *inode, int cmd, struct obdo *oa)
279 struct lov_stripe_md *lsm;
280 obd_flag valid_flags;
282 lsm = ll_i2info(inode)->lli_smd;
284 oa->o_id = lsm->lsm_object_id;
285 oa->o_valid = OBD_MD_FLID;
286 valid_flags = OBD_MD_FLTYPE | OBD_MD_FLATIME;
287 if (cmd == OBD_BRW_WRITE) {
288 oa->o_valid |= OBD_MD_FLIFID | OBD_MD_FLEPOCH;
289 mdc_pack_fid(obdo_fid(oa), inode->i_ino, 0, inode->i_mode);
290 oa->o_easize = ll_i2info(inode)->lli_io_epoch;
292 valid_flags |= OBD_MD_FLMTIME | OBD_MD_FLCTIME;
295 obdo_from_inode(oa, inode, valid_flags);
298 static void ll_ap_fill_obdo(void *data, int cmd, struct obdo *oa)
300 struct ll_async_page *llap;
303 llap = llap_from_cookie(data);
309 ll_inode_fill_obdo(llap->llap_page->mapping->host, cmd, oa);
313 static struct obd_async_page_ops ll_async_page_ops = {
314 .ap_make_ready = ll_ap_make_ready,
315 .ap_refresh_count = ll_ap_refresh_count,
316 .ap_fill_obdo = ll_ap_fill_obdo,
317 .ap_completion = ll_ap_completion,
320 /* XXX have the exp be an argument? */
321 struct ll_async_page *llap_from_page(struct page *page)
323 struct ll_async_page *llap;
324 struct obd_export *exp;
325 struct inode *inode = page->mapping->host;
326 struct ll_sb_info *sbi = ll_i2sbi(inode);
330 if (page->private != 0) {
331 llap = (struct ll_async_page *)page->private;
332 if (llap->llap_magic != LLAP_MAGIC)
333 RETURN(ERR_PTR(-EINVAL));
337 exp = ll_i2obdexp(page->mapping->host);
339 RETURN(ERR_PTR(-EINVAL));
341 OBD_ALLOC(llap, sizeof(*llap));
342 llap->llap_magic = LLAP_MAGIC;
343 rc = obd_prep_async_page(exp, ll_i2info(inode)->lli_smd,
345 (obd_off)page->index << PAGE_SHIFT,
346 &ll_async_page_ops, llap, &llap->llap_cookie);
348 OBD_FREE(llap, sizeof(*llap));
352 CDEBUG(D_CACHE, "llap %p page %p cookie %p obj off "LPU64"\n", llap,
353 page, llap->llap_cookie, (obd_off)page->index << PAGE_SHIFT);
354 page->private = (unsigned long)llap;
355 llap->llap_page = page;
357 spin_lock(&sbi->ll_pglist_lock);
358 sbi->ll_pglist_gen++;
359 list_add_tail(&llap->llap_proc_item, &sbi->ll_pglist);
360 spin_unlock(&sbi->ll_pglist_lock);
365 /* update our write count to account for i_size increases that may have
366 * happened since we've queued the page for io. */
368 /* be careful not to return success without setting the page Uptodate or
369 * the next pass through prepare_write will read in stale data from disk. */
370 int ll_commit_write(struct file *file, struct page *page, unsigned from,
373 struct inode *inode = page->mapping->host;
374 struct ll_inode_info *lli = ll_i2info(inode);
375 struct lov_stripe_md *lsm = lli->lli_smd;
376 struct obd_export *exp = NULL;
377 struct ll_async_page *llap;
382 SIGNAL_MASK_ASSERT(); /* XXX BUG 1511 */
383 LASSERT(inode == file->f_dentry->d_inode);
384 LASSERT(PageLocked(page));
386 CDEBUG(D_INODE, "inode %p is writing page %p from %d to %d at %lu\n",
387 inode, page, from, to, page->index);
389 llap = llap_from_page(page);
391 RETURN(PTR_ERR(llap));
393 /* queue a write for some time in the future the first time we
395 if (!PageDirty(page)) {
396 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats,
397 LPROC_LL_DIRTY_MISSES);
399 exp = ll_i2obdexp(inode);
403 rc = obd_queue_async_io(exp, lsm, NULL, llap->llap_cookie,
404 OBD_BRW_WRITE, 0, 0, 0, 0);
405 if (rc != 0) { /* async failed, try sync.. */
406 struct obd_sync_io_container *osic;
409 rc = obd_queue_sync_io(exp, lsm, NULL, osic,
411 OBD_BRW_WRITE, 0, to, 0);
415 rc = obd_trigger_sync_io(exp, lsm, NULL, osic);
419 rc = osic_wait(osic);
424 LL_CDEBUG_PAGE(page, "write queued\n");
425 llap->llap_queued = 1;
426 //llap_write_pending(inode, llap);
428 lprocfs_counter_incr(ll_i2sbi(inode)->ll_stats,
429 LPROC_LL_DIRTY_HITS);
432 /* put the page in the page cache, from now on ll_removepage is
433 * responsible for cleaning up the llap */
434 set_page_dirty(page);
438 /* XXX needs to be pushed down to the OSC as EOC */
439 size = (((obd_off)page->index) << PAGE_SHIFT) + to;
440 if (size > inode->i_size) {
441 inode->i_size = size;
442 /* see commentary in file.c:ll_inode_getattr() */
443 set_bit(LLI_F_PREFER_EXTENDED_SIZE, &lli->lli_flags);
445 SetPageUptodate(page);
450 /* the kernel calls us here when a page is unhashed from the page cache.
451 * the page will be locked and the kernel is holding a spinlock, so
452 * we need to be careful. we're just tearing down our book-keeping
454 void ll_removepage(struct page *page)
456 struct inode *inode = page->mapping->host;
457 struct obd_export *exp;
458 struct ll_async_page *llap;
459 struct ll_sb_info *sbi = ll_i2sbi(inode);
463 LASSERT(!in_interrupt());
465 /* sync pages or failed read pages can leave pages in the page
466 * cache that don't have our data associated with them anymore */
467 if (page->private == 0) {
472 LL_CDEBUG_PAGE(page, "being evicted\n");
474 exp = ll_i2obdexp(inode);
476 CERROR("page %p ind %lu gave null export\n", page,
482 llap = llap_from_page(page);
484 CERROR("page %p ind %lu couldn't find llap: %ld\n", page,
485 page->index, PTR_ERR(llap));
490 //llap_write_complete(inode, llap);
491 rc = obd_teardown_async_page(exp, ll_i2info(inode)->lli_smd, NULL,
494 CERROR("page %p ind %lu failed: %d\n", page, page->index, rc);
496 /* this unconditional free is only safe because the page lock
497 * is providing exclusivity to memory pressure/truncate/writeback..*/
500 spin_lock(&sbi->ll_pglist_lock);
501 if (!list_empty(&llap->llap_proc_item))
502 list_del_init(&llap->llap_proc_item);
503 sbi->ll_pglist_gen++;
504 spin_unlock(&sbi->ll_pglist_lock);
505 OBD_FREE(llap, sizeof(*llap));
509 static int ll_start_readpage(struct obd_export *exp, struct inode *inode,
512 struct ll_async_page *llap;
516 llap = llap_from_page(page);
518 RETURN(PTR_ERR(llap));
520 page_cache_get(page);
522 rc = obd_queue_async_io(exp, ll_i2info(inode)->lli_smd, NULL,
523 llap->llap_cookie, OBD_BRW_READ, 0, PAGE_SIZE,
524 0, ASYNC_READY | ASYNC_URGENT |
526 /* XXX verify that failed pages here will make their way
527 * through ->removepage.. I suspect they will. */
529 page_cache_release(page);
531 llap->llap_queued = 1;
532 LL_CDEBUG_PAGE(page, "read queued\n");
537 static void ll_start_readahead(struct obd_export *exp, struct inode *inode,
538 unsigned long first_index)
540 struct lustre_handle match_lockh = {0};
541 struct ldlm_extent page_extent;
542 unsigned long index, end_index;
544 int flags, matched, rc;
546 /* for good throughput we need to have many 'blksize' rpcs in
547 * flight per stripe, so we try to read-ahead a ridiculous amount
548 * of data. "- 3" for 8 rpcs */
549 end_index = first_index + (inode->i_blksize >> (PAGE_CACHE_SHIFT - 3));
550 if (end_index > (inode->i_size >> PAGE_CACHE_SHIFT))
551 end_index = inode->i_size >> PAGE_CACHE_SHIFT;
553 for (index = first_index + 1; index < end_index; index++) {
554 /* try to get a ref on an existing page or create a new
555 * one. if we find a locked page or lose the race
556 * with another reader we stop trying */
557 page = grab_cache_page_nowait(inode->i_mapping, index);
560 /* make sure we didn't race with other teardown/readers */
561 if (!page->mapping || Page_Uptodate(page)) {
563 page_cache_release(page);
567 /* make sure the page we're about to read is covered
568 * by a lock, stop when we go past the end of the lock */
569 page_extent.start = (__u64)page->index << PAGE_CACHE_SHIFT;
570 page_extent.end = page_extent.start + PAGE_CACHE_SIZE - 1;
571 flags = LDLM_FL_CBPENDING | LDLM_FL_BLOCK_GRANTED;
572 matched = obd_match(ll_i2sbi(inode)->ll_osc_exp,
573 ll_i2info(inode)->lli_smd, LDLM_EXTENT,
574 &page_extent, sizeof(page_extent), LCK_PR,
575 &flags, inode, &match_lockh);
577 LL_CDEBUG_PAGE(page, "lock match failed\n");
579 page_cache_release(page);
583 LL_CDEBUG_PAGE(page, "didn't match a lock\n");
585 page_cache_release(page);
589 /* interestingly, we don't need to hold the lock across the IO.
590 * As long as we match the lock while the page is locked in the
591 * page cache we know that the lock's cancelation will wait for
592 * the page to be unlocked. XXX this should transition to
593 * proper association of pages and locks in the future */
594 obd_cancel(ll_i2sbi(inode)->ll_osc_exp,
595 ll_i2info(inode)->lli_smd, LCK_PR, &match_lockh);
597 rc = ll_start_readpage(exp, inode, page);
600 page_cache_release(page);
603 page_cache_release(page);
607 * for now we do our readpage the same on both 2.4 and 2.5. The kernel's
608 * read-ahead assumes it is valid to issue readpage all the way up to
609 * i_size, but our dlm locks make that not the case. We disable the
610 * kernel's read-ahead and do our own by walking ahead in the page cache
611 * checking for dlm lock coverage. the main difference between 2.4 and
612 * 2.6 is how read-ahead gets batched and issued, but we're using our own,
613 * so they look the same.
615 int ll_readpage(struct file *file, struct page *page)
617 struct inode *inode = page->mapping->host;
618 struct lustre_handle match_lockh = {0};
619 struct obd_export *exp;
620 struct ldlm_extent page_extent;
621 int flags, rc = 0, matched;
622 struct ll_sb_info *sbi = ll_i2sbi(inode);
625 LASSERT(PageLocked(page));
626 LASSERT(!PageUptodate(page));
627 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p),offset="LPX64"\n",
628 inode->i_ino, inode->i_generation, inode,
629 (((obd_off)page->index) << PAGE_SHIFT));
630 LASSERT(atomic_read(&file->f_dentry->d_inode->i_count) > 0);
632 if (inode->i_size <= ((obd_off)page->index) << PAGE_SHIFT) {
633 CERROR("reading beyond EOF\n");
634 memset(kmap(page), 0, PAGE_SIZE);
636 SetPageUptodate(page);
640 exp = ll_i2obdexp(inode);
642 GOTO(out, rc = -EINVAL);
644 page_extent.start = (__u64)page->index << PAGE_CACHE_SHIFT;
645 page_extent.end = page_extent.start + PAGE_CACHE_SIZE - 1;
646 flags = LDLM_FL_CBPENDING | LDLM_FL_BLOCK_GRANTED;
647 matched = obd_match(sbi->ll_osc_exp, ll_i2info(inode)->lli_smd,
648 LDLM_EXTENT, &page_extent, sizeof(page_extent),
649 LCK_PR, &flags, inode, &match_lockh);
651 GOTO(out, rc = matched);
654 static unsigned long next_print;
655 CDEBUG(D_INODE, "didn't match a lock");
656 if (time_after(jiffies, next_print)) {
657 next_print = jiffies + 30 * HZ;
658 CERROR("not covered by a lock (mmap?). check debug "
663 rc = ll_start_readpage(exp, inode, page);
664 if (rc == 0 && (sbi->ll_flags & LL_SBI_READAHEAD))
665 ll_start_readahead(exp, inode, page->index);
668 obd_cancel(ll_i2sbi(inode)->ll_osc_exp,
669 ll_i2info(inode)->lli_smd, LCK_PR, &match_lockh);