4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2011, 2017, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
32 * lustre/lustre/llite/rw26.c
34 * Lustre Lite I/O page cache routines for the 2.5/2.6 kernel version
37 #include <linux/buffer_head.h>
38 #include <linux/errno.h>
40 #include <linux/kernel.h>
42 #include <linux/mpage.h>
43 #include <linux/pagemap.h>
44 #include <linux/string.h>
45 #include <linux/unistd.h>
46 #include <linux/writeback.h>
47 #include <linux/migrate.h>
49 #define DEBUG_SUBSYSTEM S_LLITE
51 #include "llite_internal.h"
52 #include <lustre_compat.h>
55 * Implements Linux VM address_space::invalidatepage() method. This method is
56 * called when the page is truncate from a file, either as a result of
57 * explicit truncate, or when inode is removed from memory (as a result of
58 * final iput(), umount, or memory pressure induced icache shrinking).
60 * [0, offset] bytes of the page remain valid (this is for a case of not-page
61 * aligned truncate). Lustre leaves partially truncated page in the cache,
62 * relying on struct inode::i_size to limit further accesses.
64 static void ll_invalidatepage(struct page *vmpage,
65 #ifdef HAVE_INVALIDATE_RANGE
66 unsigned int offset, unsigned int length
75 struct cl_object *obj;
77 LASSERT(PageLocked(vmpage));
78 LASSERT(!PageWriteback(vmpage));
81 * It is safe to not check anything in invalidatepage/releasepage
82 * below because they are run with page locked and all our io is
83 * happening with locked page too
85 #ifdef HAVE_INVALIDATE_RANGE
86 if (offset == 0 && length == PAGE_SIZE) {
90 /* See the comment in ll_releasepage() */
91 env = cl_env_percpu_get();
92 LASSERT(!IS_ERR(env));
94 inode = vmpage->mapping->host;
95 obj = ll_i2info(inode)->lli_clob;
97 page = cl_vmpage_page(vmpage, obj);
99 cl_page_delete(env, page);
100 cl_page_put(env, page);
103 LASSERT(vmpage->private == 0);
105 cl_env_percpu_put(env);
109 #ifdef HAVE_RELEASEPAGE_WITH_INT
110 #define RELEASEPAGE_ARG_TYPE int
112 #define RELEASEPAGE_ARG_TYPE gfp_t
114 static int ll_releasepage(struct page *vmpage, RELEASEPAGE_ARG_TYPE gfp_mask)
117 struct cl_object *obj;
118 struct cl_page *page;
119 struct address_space *mapping;
122 LASSERT(PageLocked(vmpage));
123 if (PageWriteback(vmpage) || PageDirty(vmpage))
126 mapping = vmpage->mapping;
130 obj = ll_i2info(mapping->host)->lli_clob;
134 page = cl_vmpage_page(vmpage, obj);
138 env = cl_env_percpu_get();
139 LASSERT(!IS_ERR(env));
141 if (!cl_page_in_use(page)) {
143 cl_page_delete(env, page);
146 /* To use percpu env array, the call path can not be rescheduled;
147 * otherwise percpu array will be messed if ll_releaspage() called
148 * again on the same CPU.
150 * If this page holds the last refc of cl_object, the following
151 * call path may cause reschedule:
152 * cl_page_put -> cl_page_free -> cl_object_put ->
153 * lu_object_put -> lu_object_free -> lov_delete_raid0.
155 * However, the kernel can't get rid of this inode until all pages have
156 * been cleaned up. Now that we hold page lock here, it's pretty safe
157 * that we won't get into object delete path.
159 LASSERT(cl_object_refc(obj) > 1);
160 cl_page_put(env, page);
162 cl_env_percpu_put(env);
166 #if defined(HAVE_DIRECTIO_ITER) || defined(HAVE_IOV_ITER_RW) || \
167 defined(HAVE_DIRECTIO_2ARGS)
168 #define HAVE_DIO_ITER 1
172 * ll_free_user_pages - tear down page struct array
173 * @pages: array of page struct pointers underlying target buffer
175 static void ll_free_user_pages(struct page **pages, int npages)
179 for (i = 0; i < npages; i++) {
185 #if defined(HAVE_DIO_ITER)
188 OBD_FREE_PTR_ARRAY_LARGE(pages, npages);
192 static ssize_t ll_get_user_pages(int rw, struct iov_iter *iter,
193 struct page ***pages, ssize_t *npages,
196 #if defined(HAVE_DIO_ITER)
201 * iov_iter_get_pages_alloc() is introduced in 3.16 similar
204 result = iov_iter_get_pages_alloc(iter, pages, maxsize, &start);
206 *npages = DIV_ROUND_UP(result + start, PAGE_SIZE);
221 addr = (unsigned long)iter->iov->iov_base + iter->iov_offset;
222 if (addr & ~PAGE_MASK)
225 size = min_t(size_t, maxsize, iter->iov->iov_len);
226 page_count = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
227 OBD_ALLOC_PTR_ARRAY_LARGE(*pages, page_count);
231 down_read(¤t->mm->mmap_sem);
232 result = get_user_pages(current, current->mm, addr, page_count,
233 rw == READ, 0, *pages, NULL);
234 up_read(¤t->mm->mmap_sem);
236 if (unlikely(result != page_count)) {
237 ll_free_user_pages(*pages, page_count);
245 *npages = page_count;
251 /* iov_iter_alignment() is introduced in 3.16 similar to HAVE_DIO_ITER */
252 #if defined(HAVE_DIO_ITER)
253 static unsigned long ll_iov_iter_alignment(const struct iov_iter *i)
255 return iov_iter_alignment(i);
257 #else /* copied from alignment_iovec() */
258 static unsigned long ll_iov_iter_alignment(const struct iov_iter *i)
260 const struct iovec *iov = i->iov;
262 size_t size = i->count;
268 res = (unsigned long)iov->iov_base + i->iov_offset;
269 n = iov->iov_len - i->iov_offset;
275 while (size > (++iov)->iov_len) {
276 res |= (unsigned long)iov->iov_base | iov->iov_len;
277 size -= iov->iov_len;
279 res |= (unsigned long)iov->iov_base | size;
285 /** direct IO pages */
286 struct ll_dio_pages {
287 struct cl_dio_aio *ldp_aio;
289 * page array to be written. we don't support
290 * partial pages except the last one.
292 struct page **ldp_pages;
293 /** # of pages in the array. */
295 /* the file offset of the first page. */
296 loff_t ldp_file_offset;
300 ll_direct_rw_pages(const struct lu_env *env, struct cl_io *io, size_t size,
301 int rw, struct inode *inode, struct ll_dio_pages *pv)
303 struct cl_page *page;
304 struct cl_2queue *queue = &io->ci_queue;
305 struct cl_object *obj = io->ci_obj;
306 struct cl_sync_io *anchor = &pv->ldp_aio->cda_sync;
307 loff_t offset = pv->ldp_file_offset;
309 size_t page_size = cl_page_size(obj);
315 cl_2queue_init(queue);
316 for (i = 0; i < pv->ldp_count; i++) {
317 LASSERT(!(offset & (PAGE_SIZE - 1)));
318 page = cl_page_find(env, obj, cl_index(obj, offset),
319 pv->ldp_pages[i], CPT_TRANSIENT);
324 LASSERT(page->cp_type == CPT_TRANSIENT);
325 rc = cl_page_own(env, io, page);
327 cl_page_put(env, page);
331 page->cp_sync_io = anchor;
332 cl_2queue_add(queue, page);
334 * Set page clip to tell transfer formation engine
335 * that page has to be sent even if it is beyond KMS.
337 cl_page_clip(env, page, 0, min(size, page_size));
340 /* drop the reference count for cl_page_find */
341 cl_page_put(env, page);
345 if (rc == 0 && io_pages > 0) {
346 int iot = rw == READ ? CRT_READ : CRT_WRITE;
348 atomic_add(io_pages, &anchor->csi_sync_nr);
349 rc = cl_io_submit_rw(env, io, iot, queue);
351 cl_page_list_splice(&queue->c2_qout,
352 &pv->ldp_aio->cda_pages);
354 atomic_add(-queue->c2_qin.pl_nr,
355 &anchor->csi_sync_nr);
356 cl_page_list_for_each(page, &queue->c2_qin)
357 page->cp_sync_io = NULL;
359 /* handle partially submitted reqs */
360 if (queue->c2_qin.pl_nr > 0) {
361 CERROR(DFID " failed to submit %d dio pages: %zd\n",
362 PFID(lu_object_fid(&obj->co_lu)),
363 queue->c2_qin.pl_nr, rc);
369 cl_2queue_discard(env, io, queue);
370 cl_2queue_disown(env, io, queue);
371 cl_2queue_fini(env, queue);
375 #ifdef KMALLOC_MAX_SIZE
376 #define MAX_MALLOC KMALLOC_MAX_SIZE
378 #define MAX_MALLOC (128 * 1024)
381 /* This is the maximum size of a single O_DIRECT request, based on the
382 * kmalloc limit. We need to fit all of the brw_page structs, each one
383 * representing PAGE_SIZE worth of user data, into a single buffer, and
384 * then truncate this to be a full-sized RPC. For 4kB PAGE_SIZE this is
385 * up to 22MB for 128kB kmalloc and up to 682MB for 4MB kmalloc. */
386 #define MAX_DIO_SIZE ((MAX_MALLOC / sizeof(struct brw_page) * PAGE_SIZE) & \
387 ~((size_t)DT_MAX_BRW_SIZE - 1))
390 ll_direct_IO_impl(struct kiocb *iocb, struct iov_iter *iter, int rw)
392 struct ll_cl_context *lcc;
393 const struct lu_env *env;
395 struct file *file = iocb->ki_filp;
396 struct inode *inode = file->f_mapping->host;
397 struct cl_dio_aio *aio;
398 size_t count = iov_iter_count(iter);
399 ssize_t tot_bytes = 0, result = 0;
400 loff_t file_offset = iocb->ki_pos;
403 /* if file is encrypted, return 0 so that we fall back to buffered IO */
404 if (IS_ENCRYPTED(inode))
407 /* Check EOF by ourselves */
408 if (rw == READ && file_offset >= i_size_read(inode))
411 /* FIXME: io smaller than PAGE_SIZE is broken on ia64 ??? */
412 if ((file_offset & ~PAGE_MASK) || (count & ~PAGE_MASK))
415 CDEBUG(D_VFSTRACE, "VFS Op:inode="DFID"(%p), size=%zd (max %lu), "
416 "offset=%lld=%llx, pages %zd (max %lu)\n",
417 PFID(ll_inode2fid(inode)), inode, count, MAX_DIO_SIZE,
418 file_offset, file_offset, count >> PAGE_SHIFT,
419 MAX_DIO_SIZE >> PAGE_SHIFT);
421 /* Check that all user buffers are aligned as well */
422 if (ll_iov_iter_alignment(iter) & ~PAGE_MASK)
425 lcc = ll_cl_find(file);
430 LASSERT(!IS_ERR(env));
431 vio = vvp_env_io(env);
437 LASSERT(aio->cda_iocb == iocb);
439 /* 0. Need locking between buffered and direct access. and race with
440 * size changing by concurrent truncates and writes.
441 * 1. Need inode mutex to operate transient pages.
446 while (iov_iter_count(iter)) {
447 struct ll_dio_pages pvec = { .ldp_aio = aio };
450 count = min_t(size_t, iov_iter_count(iter), MAX_DIO_SIZE);
452 if (file_offset >= i_size_read(inode))
455 if (file_offset + count > i_size_read(inode))
456 count = i_size_read(inode) - file_offset;
459 result = ll_get_user_pages(rw, iter, &pages,
460 &pvec.ldp_count, count);
461 if (unlikely(result <= 0))
465 pvec.ldp_file_offset = file_offset;
466 pvec.ldp_pages = pages;
468 result = ll_direct_rw_pages(env, io, count,
470 ll_free_user_pages(pages, pvec.ldp_count);
472 if (unlikely(result < 0))
475 iov_iter_advance(iter, count);
477 file_offset += count;
481 aio->cda_bytes += tot_bytes;
483 if (is_sync_kiocb(iocb)) {
484 struct cl_sync_io *anchor = &aio->cda_sync;
488 * @anchor was inited as 1 to prevent end_io to be
489 * called before we add all pages for IO, so drop
490 * one extra reference to make sure we could wait
493 cl_sync_io_note(env, anchor, result);
495 rc2 = cl_sync_io_wait(env, anchor, 0);
496 if (result == 0 && rc2)
499 * One extra reference again, as if @anchor is
500 * reused we assume it as 1 before using.
502 atomic_add(1, &anchor->csi_sync_nr);
504 /* no commit async for direct IO */
505 vio->u.readwrite.vui_written += tot_bytes;
510 vio->u.readwrite.vui_written += tot_bytes;
512 vio->u.readwrite.vui_read += tot_bytes;
513 result = -EIOCBQUEUED;
522 #if defined(HAVE_DIO_ITER)
523 static ssize_t ll_direct_IO(
524 #ifndef HAVE_IOV_ITER_RW
527 struct kiocb *iocb, struct iov_iter *iter
528 #ifndef HAVE_DIRECTIO_2ARGS
535 #ifndef HAVE_IOV_ITER_RW
538 nrw = iov_iter_rw(iter);
541 return ll_direct_IO_impl(iocb, iter, nrw);
544 #else /* !defined(HAVE_DIO_ITER) */
547 ll_direct_IO(int rw, struct kiocb *iocb, const struct iovec *iov,
548 loff_t file_offset, unsigned long nr_segs)
550 struct iov_iter iter;
552 iov_iter_init(&iter, iov, nr_segs, iov_length(iov, nr_segs), 0);
553 return ll_direct_IO_impl(iocb, &iter, rw);
556 #endif /* !defined(HAVE_DIO_ITER) */
559 * Prepare partially written-to page for a write.
560 * @pg is owned when passed in and disowned when it returns non-zero result to
563 static int ll_prepare_partial_page(const struct lu_env *env, struct cl_io *io,
564 struct cl_page *pg, struct file *file)
566 struct cl_attr *attr = vvp_env_thread_attr(env);
567 struct cl_object *obj = io->ci_obj;
568 struct vvp_page *vpg = cl_object_page_slice(obj, pg);
569 loff_t offset = cl_offset(obj, vvp_index(vpg));
573 cl_object_attr_lock(obj);
574 result = cl_object_attr_get(env, obj, attr);
575 cl_object_attr_unlock(obj);
577 cl_page_disown(env, io, pg);
582 * If are writing to a new page, no need to read old data.
583 * The extent locking will have updated the KMS, and for our
584 * purposes here we can treat it like i_size.
586 if (attr->cat_kms <= offset) {
587 char *kaddr = kmap_atomic(vpg->vpg_page);
589 memset(kaddr, 0, cl_page_size(obj));
590 kunmap_atomic(kaddr);
591 GOTO(out, result = 0);
594 if (vpg->vpg_defer_uptodate) {
595 vpg->vpg_ra_used = 1;
596 GOTO(out, result = 0);
599 result = ll_io_read_page(env, io, pg, file);
603 /* ll_io_read_page() disowns the page */
604 result = cl_page_own(env, io, pg);
606 if (!PageUptodate(cl_page_vmpage(pg))) {
607 cl_page_disown(env, io, pg);
610 } else if (result == -ENOENT) {
611 /* page was truncated */
620 static int ll_tiny_write_begin(struct page *vmpage, struct address_space *mapping)
622 /* Page must be present, up to date, dirty, and not in writeback. */
623 if (!vmpage || !PageUptodate(vmpage) || !PageDirty(vmpage) ||
624 PageWriteback(vmpage) || vmpage->mapping != mapping)
630 static int ll_write_begin(struct file *file, struct address_space *mapping,
631 loff_t pos, unsigned len, unsigned flags,
632 struct page **pagep, void **fsdata)
634 struct ll_cl_context *lcc = NULL;
635 const struct lu_env *env = NULL;
636 struct cl_io *io = NULL;
637 struct cl_page *page = NULL;
639 struct cl_object *clob = ll_i2info(mapping->host)->lli_clob;
640 pgoff_t index = pos >> PAGE_SHIFT;
641 struct page *vmpage = NULL;
642 unsigned from = pos & (PAGE_SIZE - 1);
643 unsigned to = from + len;
647 CDEBUG(D_VFSTRACE, "Writing %lu of %d to %d bytes\n", index, from, len);
649 lcc = ll_cl_find(file);
651 vmpage = grab_cache_page_nowait(mapping, index);
652 result = ll_tiny_write_begin(vmpage, mapping);
659 if (file->f_flags & O_DIRECT) {
660 /* direct IO failed because it couldn't clean up cached pages,
661 * this causes a problem for mirror write because the cached
662 * page may belong to another mirror, which will result in
663 * problem submitting the I/O. */
664 if (io->ci_designated_mirror > 0)
665 GOTO(out, result = -EBUSY);
668 * Direct read can fall back to buffered read, but DIO is done
669 * with lockless i/o, and buffered requires LDLM locking, so
670 * in this case we must restart without lockless.
672 if (!io->ci_ignore_lockless) {
673 io->ci_ignore_lockless = 1;
674 io->ci_need_restart = 1;
675 GOTO(out, result = -ENOLCK);
679 /* To avoid deadlock, try to lock page first. */
680 vmpage = grab_cache_page_nowait(mapping, index);
682 if (unlikely(vmpage == NULL ||
683 PageDirty(vmpage) || PageWriteback(vmpage))) {
684 struct vvp_io *vio = vvp_env_io(env);
685 struct cl_page_list *plist = &vio->u.readwrite.vui_queue;
687 /* if the page is already in dirty cache, we have to commit
688 * the pages right now; otherwise, it may cause deadlock
689 * because it holds page lock of a dirty page and request for
690 * more grants. It's okay for the dirty page to be the first
691 * one in commit page list, though. */
692 if (vmpage != NULL && plist->pl_nr > 0) {
698 /* commit pages and then wait for page lock */
699 result = vvp_io_write_commit(env, io);
703 if (vmpage == NULL) {
704 vmpage = grab_cache_page_write_begin(mapping, index,
707 GOTO(out, result = -ENOMEM);
711 /* page was truncated */
712 if (mapping != vmpage->mapping) {
713 CDEBUG(D_VFSTRACE, "page: %lu was truncated\n", index);
720 page = cl_page_find(env, clob, vmpage->index, vmpage, CPT_CACHEABLE);
722 GOTO(out, result = PTR_ERR(page));
724 lcc->lcc_page = page;
725 lu_ref_add(&page->cp_reference, "cl_io", io);
727 cl_page_assume(env, io, page);
728 if (!PageUptodate(vmpage)) {
730 * We're completely overwriting an existing page,
731 * so _don't_ set it up to date until commit_write
733 if (from == 0 && to == PAGE_SIZE) {
734 CL_PAGE_HEADER(D_PAGE, env, page, "full page write\n");
735 POISON_PAGE(vmpage, 0x11);
737 /* TODO: can be optimized at OSC layer to check if it
738 * is a lockless IO. In that case, it's not necessary
739 * to read the data. */
740 result = ll_prepare_partial_page(env, io, page, file);
742 /* vmpage should have been unlocked */
746 if (result == -EAGAIN)
755 if (vmpage != NULL) {
759 /* On tiny_write failure, page and io are always null. */
760 if (!IS_ERR_OR_NULL(page)) {
761 lu_ref_del(&page->cp_reference, "cl_io", io);
762 cl_page_put(env, page);
765 io->ci_result = result;
773 static int ll_tiny_write_end(struct file *file, struct address_space *mapping,
774 loff_t pos, unsigned int len, unsigned int copied,
777 struct cl_page *clpage = (struct cl_page *) vmpage->private;
778 loff_t kms = pos+copied;
779 loff_t to = kms & (PAGE_SIZE-1) ? kms & (PAGE_SIZE-1) : PAGE_SIZE;
781 struct lu_env *env = cl_env_get(&refcheck);
791 /* This page is dirty in cache, so it should have a cl_page pointer
792 * set in vmpage->private.
794 LASSERT(clpage != NULL);
799 /* Update the underlying size information in the OSC/LOV objects this
802 cl_page_touch(env, clpage, to);
805 cl_env_put(env, &refcheck);
808 /* Must return page unlocked. */
814 static int ll_write_end(struct file *file, struct address_space *mapping,
815 loff_t pos, unsigned len, unsigned copied,
816 struct page *vmpage, void *fsdata)
818 struct ll_cl_context *lcc = fsdata;
819 const struct lu_env *env;
822 struct cl_page *page;
823 unsigned from = pos & (PAGE_SIZE - 1);
830 CDEBUG(D_VFSTRACE, "pos %llu, len %u, copied %u\n", pos, len, copied);
833 result = ll_tiny_write_end(file, mapping, pos, len, copied,
838 LASSERT(lcc != NULL);
840 page = lcc->lcc_page;
842 vio = vvp_env_io(env);
844 LASSERT(cl_page_is_owned(page, io));
846 struct cl_page_list *plist = &vio->u.readwrite.vui_queue;
848 lcc->lcc_page = NULL; /* page will be queued */
850 /* Add it into write queue */
851 cl_page_list_add(plist, page);
852 if (plist->pl_nr == 1) /* first page */
853 vio->u.readwrite.vui_from = from;
856 vio->u.readwrite.vui_to = from + copied;
858 /* To address the deadlock in balance_dirty_pages() where
859 * this dirty page may be written back in the same thread. */
860 if (PageDirty(vmpage))
863 /* We may have one full RPC, commit it soon */
864 if (plist->pl_nr >= PTLRPC_MAX_BRW_PAGES)
867 CL_PAGE_DEBUG(D_VFSTRACE, env, page,
868 "queued page: %d.\n", plist->pl_nr);
870 cl_page_disown(env, io, page);
872 lcc->lcc_page = NULL;
873 lu_ref_del(&page->cp_reference, "cl_io", io);
874 cl_page_put(env, page);
876 /* page list is not contiguous now, commit it now */
879 if (unplug || io->u.ci_wr.wr_sync)
880 result = vvp_io_write_commit(env, io);
883 io->ci_result = result;
887 RETURN(result >= 0 ? copied : result);
890 #ifdef CONFIG_MIGRATION
891 static int ll_migratepage(struct address_space *mapping,
892 struct page *newpage, struct page *page,
893 enum migrate_mode mode)
895 /* Always fail page migration until we have a proper implementation */
900 const struct address_space_operations ll_aops = {
901 .readpage = ll_readpage,
902 .direct_IO = ll_direct_IO,
903 .writepage = ll_writepage,
904 .writepages = ll_writepages,
905 .set_page_dirty = __set_page_dirty_nobuffers,
906 .write_begin = ll_write_begin,
907 .write_end = ll_write_end,
908 .invalidatepage = ll_invalidatepage,
909 .releasepage = (void *)ll_releasepage,
910 #ifdef CONFIG_MIGRATION
911 .migratepage = ll_migratepage,