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);
311 pgoff_t index = offset >> PAGE_SHIFT;
316 cl_2queue_init(queue);
317 for (i = 0; i < pv->ldp_count; i++) {
318 LASSERT(!(offset & (PAGE_SIZE - 1)));
319 page = cl_page_find(env, obj, cl_index(obj, offset),
320 pv->ldp_pages[i], CPT_TRANSIENT);
325 LASSERT(page->cp_type == CPT_TRANSIENT);
326 rc = cl_page_own(env, io, page);
328 cl_page_put(env, page);
332 page->cp_sync_io = anchor;
333 if (inode && IS_ENCRYPTED(inode)) {
334 struct page *vmpage = cl_page_vmpage(page);
336 /* In case of Direct IO on encrypted file, we need to
337 * set the correct page index, and add a reference to
338 * the mapping. This is required by llcrypt to proceed
339 * to encryption/decryption, because each block is
340 * encrypted independently, and each block's IV is set
341 * to the logical block number within the file.
342 * This is safe because we know these pages are private
343 * to the thread doing the Direct IO, and despite
344 * setting a mapping on the pages, cached lookups will
346 * Set PageChecked to detect special case of Direct IO
347 * in osc_brw_fini_request().
348 * Reference to the mapping and PageChecked flag are
349 * removed in cl_aio_end().
351 vmpage->index = index++;
352 vmpage->mapping = inode->i_mapping;
353 SetPageChecked(vmpage);
355 cl_2queue_add(queue, page);
357 * Set page clip to tell transfer formation engine
358 * that page has to be sent even if it is beyond KMS.
360 cl_page_clip(env, page, 0, min(size, page_size));
363 /* drop the reference count for cl_page_find */
364 cl_page_put(env, page);
368 if (rc == 0 && io_pages > 0) {
369 int iot = rw == READ ? CRT_READ : CRT_WRITE;
371 atomic_add(io_pages, &anchor->csi_sync_nr);
373 * Avoid out-of-order execution of adding inflight
374 * modifications count and io submit.
377 rc = cl_io_submit_rw(env, io, iot, queue);
379 cl_page_list_splice(&queue->c2_qout,
380 &pv->ldp_aio->cda_pages);
382 atomic_add(-queue->c2_qin.pl_nr,
383 &anchor->csi_sync_nr);
384 cl_page_list_for_each(page, &queue->c2_qin)
385 page->cp_sync_io = NULL;
387 /* handle partially submitted reqs */
388 if (queue->c2_qin.pl_nr > 0) {
389 CERROR(DFID " failed to submit %d dio pages: %zd\n",
390 PFID(lu_object_fid(&obj->co_lu)),
391 queue->c2_qin.pl_nr, rc);
397 cl_2queue_discard(env, io, queue);
398 cl_2queue_disown(env, io, queue);
399 cl_2queue_fini(env, queue);
403 #ifdef KMALLOC_MAX_SIZE
404 #define MAX_MALLOC KMALLOC_MAX_SIZE
406 #define MAX_MALLOC (128 * 1024)
409 /* This is the maximum size of a single O_DIRECT request, based on the
410 * kmalloc limit. We need to fit all of the brw_page structs, each one
411 * representing PAGE_SIZE worth of user data, into a single buffer, and
412 * then truncate this to be a full-sized RPC. For 4kB PAGE_SIZE this is
413 * up to 22MB for 128kB kmalloc and up to 682MB for 4MB kmalloc. */
414 #define MAX_DIO_SIZE ((MAX_MALLOC / sizeof(struct brw_page) * PAGE_SIZE) & \
415 ~((size_t)DT_MAX_BRW_SIZE - 1))
418 ll_direct_IO_impl(struct kiocb *iocb, struct iov_iter *iter, int rw)
420 struct ll_cl_context *lcc;
421 const struct lu_env *env;
423 struct file *file = iocb->ki_filp;
424 struct inode *inode = file->f_mapping->host;
425 struct cl_dio_aio *aio;
426 size_t count = iov_iter_count(iter);
427 ssize_t tot_bytes = 0, result = 0;
428 loff_t file_offset = iocb->ki_pos;
431 /* Check EOF by ourselves */
432 if (rw == READ && file_offset >= i_size_read(inode))
435 /* FIXME: io smaller than PAGE_SIZE is broken on ia64 ??? */
436 if ((file_offset & ~PAGE_MASK) || (count & ~PAGE_MASK))
439 CDEBUG(D_VFSTRACE, "VFS Op:inode="DFID"(%p), size=%zd (max %lu), "
440 "offset=%lld=%llx, pages %zd (max %lu)\n",
441 PFID(ll_inode2fid(inode)), inode, count, MAX_DIO_SIZE,
442 file_offset, file_offset, count >> PAGE_SHIFT,
443 MAX_DIO_SIZE >> PAGE_SHIFT);
445 /* Check that all user buffers are aligned as well */
446 if (ll_iov_iter_alignment(iter) & ~PAGE_MASK)
449 lcc = ll_cl_find(file);
454 LASSERT(!IS_ERR(env));
455 vio = vvp_env_io(env);
461 LASSERT(aio->cda_iocb == iocb);
463 while (iov_iter_count(iter)) {
464 struct ll_dio_pages pvec = { .ldp_aio = aio };
467 count = min_t(size_t, iov_iter_count(iter), MAX_DIO_SIZE);
469 if (file_offset >= i_size_read(inode))
472 if (file_offset + count > i_size_read(inode))
473 count = i_size_read(inode) - file_offset;
476 result = ll_get_user_pages(rw, iter, &pages,
477 &pvec.ldp_count, count);
478 if (unlikely(result <= 0))
482 pvec.ldp_file_offset = file_offset;
483 pvec.ldp_pages = pages;
485 result = ll_direct_rw_pages(env, io, count,
487 ll_free_user_pages(pages, pvec.ldp_count);
489 if (unlikely(result < 0))
492 iov_iter_advance(iter, count);
494 file_offset += count;
498 aio->cda_bytes += tot_bytes;
500 if (is_sync_kiocb(iocb)) {
501 struct cl_sync_io *anchor = &aio->cda_sync;
505 * @anchor was inited as 1 to prevent end_io to be
506 * called before we add all pages for IO, so drop
507 * one extra reference to make sure we could wait
510 cl_sync_io_note(env, anchor, result);
512 rc2 = cl_sync_io_wait(env, anchor, 0);
513 if (result == 0 && rc2)
516 * One extra reference again, as if @anchor is
517 * reused we assume it as 1 before using.
519 atomic_add(1, &anchor->csi_sync_nr);
521 /* no commit async for direct IO */
522 vio->u.readwrite.vui_written += tot_bytes;
527 vio->u.readwrite.vui_written += tot_bytes;
529 vio->u.readwrite.vui_read += tot_bytes;
530 result = -EIOCBQUEUED;
536 #if defined(HAVE_DIO_ITER)
537 static ssize_t ll_direct_IO(
538 #ifndef HAVE_IOV_ITER_RW
541 struct kiocb *iocb, struct iov_iter *iter
542 #ifndef HAVE_DIRECTIO_2ARGS
549 #ifndef HAVE_IOV_ITER_RW
552 nrw = iov_iter_rw(iter);
555 return ll_direct_IO_impl(iocb, iter, nrw);
558 #else /* !defined(HAVE_DIO_ITER) */
561 ll_direct_IO(int rw, struct kiocb *iocb, const struct iovec *iov,
562 loff_t file_offset, unsigned long nr_segs)
564 struct iov_iter iter;
566 iov_iter_init(&iter, iov, nr_segs, iov_length(iov, nr_segs), 0);
567 return ll_direct_IO_impl(iocb, &iter, rw);
570 #endif /* !defined(HAVE_DIO_ITER) */
573 * Prepare partially written-to page for a write.
574 * @pg is owned when passed in and disowned when it returns non-zero result to
577 static int ll_prepare_partial_page(const struct lu_env *env, struct cl_io *io,
578 struct cl_page *pg, struct file *file)
580 struct cl_attr *attr = vvp_env_thread_attr(env);
581 struct cl_object *obj = io->ci_obj;
582 struct vvp_page *vpg = cl_object_page_slice(obj, pg);
583 loff_t offset = cl_offset(obj, vvp_index(vpg));
587 cl_object_attr_lock(obj);
588 result = cl_object_attr_get(env, obj, attr);
589 cl_object_attr_unlock(obj);
591 cl_page_disown(env, io, pg);
596 * If are writing to a new page, no need to read old data.
597 * The extent locking will have updated the KMS, and for our
598 * purposes here we can treat it like i_size.
600 if (attr->cat_kms <= offset) {
601 char *kaddr = kmap_atomic(vpg->vpg_page);
603 memset(kaddr, 0, cl_page_size(obj));
604 kunmap_atomic(kaddr);
605 GOTO(out, result = 0);
608 if (vpg->vpg_defer_uptodate) {
609 vpg->vpg_ra_used = 1;
610 GOTO(out, result = 0);
613 result = ll_io_read_page(env, io, pg, file);
617 /* ll_io_read_page() disowns the page */
618 result = cl_page_own(env, io, pg);
620 if (!PageUptodate(cl_page_vmpage(pg))) {
621 cl_page_disown(env, io, pg);
624 } else if (result == -ENOENT) {
625 /* page was truncated */
634 static int ll_tiny_write_begin(struct page *vmpage, struct address_space *mapping)
636 /* Page must be present, up to date, dirty, and not in writeback. */
637 if (!vmpage || !PageUptodate(vmpage) || !PageDirty(vmpage) ||
638 PageWriteback(vmpage) || vmpage->mapping != mapping)
644 static int ll_write_begin(struct file *file, struct address_space *mapping,
645 loff_t pos, unsigned len, unsigned flags,
646 struct page **pagep, void **fsdata)
648 struct ll_cl_context *lcc = NULL;
649 const struct lu_env *env = NULL;
650 struct cl_io *io = NULL;
651 struct cl_page *page = NULL;
653 struct cl_object *clob = ll_i2info(mapping->host)->lli_clob;
654 pgoff_t index = pos >> PAGE_SHIFT;
655 struct page *vmpage = NULL;
656 unsigned from = pos & (PAGE_SIZE - 1);
657 unsigned to = from + len;
661 CDEBUG(D_VFSTRACE, "Writing %lu of %d to %d bytes\n", index, from, len);
663 lcc = ll_cl_find(file);
665 vmpage = grab_cache_page_nowait(mapping, index);
666 result = ll_tiny_write_begin(vmpage, mapping);
673 if (file->f_flags & O_DIRECT) {
674 /* direct IO failed because it couldn't clean up cached pages,
675 * this causes a problem for mirror write because the cached
676 * page may belong to another mirror, which will result in
677 * problem submitting the I/O. */
678 if (io->ci_designated_mirror > 0)
679 GOTO(out, result = -EBUSY);
682 * Direct read can fall back to buffered read, but DIO is done
683 * with lockless i/o, and buffered requires LDLM locking, so
684 * in this case we must restart without lockless.
686 if (!io->ci_ignore_lockless) {
687 io->ci_ignore_lockless = 1;
688 io->ci_need_restart = 1;
689 GOTO(out, result = -ENOLCK);
693 /* To avoid deadlock, try to lock page first. */
694 vmpage = grab_cache_page_nowait(mapping, index);
696 if (unlikely(vmpage == NULL ||
697 PageDirty(vmpage) || PageWriteback(vmpage))) {
698 struct vvp_io *vio = vvp_env_io(env);
699 struct cl_page_list *plist = &vio->u.readwrite.vui_queue;
701 /* if the page is already in dirty cache, we have to commit
702 * the pages right now; otherwise, it may cause deadlock
703 * because it holds page lock of a dirty page and request for
704 * more grants. It's okay for the dirty page to be the first
705 * one in commit page list, though. */
706 if (vmpage != NULL && plist->pl_nr > 0) {
712 /* commit pages and then wait for page lock */
713 result = vvp_io_write_commit(env, io);
717 if (vmpage == NULL) {
718 vmpage = grab_cache_page_write_begin(mapping, index,
721 GOTO(out, result = -ENOMEM);
725 /* page was truncated */
726 if (mapping != vmpage->mapping) {
727 CDEBUG(D_VFSTRACE, "page: %lu was truncated\n", index);
734 page = cl_page_find(env, clob, vmpage->index, vmpage, CPT_CACHEABLE);
736 GOTO(out, result = PTR_ERR(page));
738 lcc->lcc_page = page;
739 lu_ref_add(&page->cp_reference, "cl_io", io);
741 cl_page_assume(env, io, page);
742 if (!PageUptodate(vmpage)) {
744 * We're completely overwriting an existing page,
745 * so _don't_ set it up to date until commit_write
747 if (from == 0 && to == PAGE_SIZE) {
748 CL_PAGE_HEADER(D_PAGE, env, page, "full page write\n");
749 POISON_PAGE(vmpage, 0x11);
751 /* TODO: can be optimized at OSC layer to check if it
752 * is a lockless IO. In that case, it's not necessary
753 * to read the data. */
754 result = ll_prepare_partial_page(env, io, page, file);
756 /* vmpage should have been unlocked */
760 if (result == -EAGAIN)
769 if (vmpage != NULL) {
773 /* On tiny_write failure, page and io are always null. */
774 if (!IS_ERR_OR_NULL(page)) {
775 lu_ref_del(&page->cp_reference, "cl_io", io);
776 cl_page_put(env, page);
779 io->ci_result = result;
787 static int ll_tiny_write_end(struct file *file, struct address_space *mapping,
788 loff_t pos, unsigned int len, unsigned int copied,
791 struct cl_page *clpage = (struct cl_page *) vmpage->private;
792 loff_t kms = pos+copied;
793 loff_t to = kms & (PAGE_SIZE-1) ? kms & (PAGE_SIZE-1) : PAGE_SIZE;
795 struct lu_env *env = cl_env_get(&refcheck);
805 /* This page is dirty in cache, so it should have a cl_page pointer
806 * set in vmpage->private.
808 LASSERT(clpage != NULL);
813 /* Update the underlying size information in the OSC/LOV objects this
816 cl_page_touch(env, clpage, to);
819 cl_env_put(env, &refcheck);
822 /* Must return page unlocked. */
828 static int ll_write_end(struct file *file, struct address_space *mapping,
829 loff_t pos, unsigned len, unsigned copied,
830 struct page *vmpage, void *fsdata)
832 struct ll_cl_context *lcc = fsdata;
833 const struct lu_env *env;
836 struct cl_page *page;
837 unsigned from = pos & (PAGE_SIZE - 1);
844 CDEBUG(D_VFSTRACE, "pos %llu, len %u, copied %u\n", pos, len, copied);
847 result = ll_tiny_write_end(file, mapping, pos, len, copied,
852 LASSERT(lcc != NULL);
854 page = lcc->lcc_page;
856 vio = vvp_env_io(env);
858 LASSERT(cl_page_is_owned(page, io));
860 struct cl_page_list *plist = &vio->u.readwrite.vui_queue;
862 lcc->lcc_page = NULL; /* page will be queued */
864 /* Add it into write queue */
865 cl_page_list_add(plist, page);
866 if (plist->pl_nr == 1) /* first page */
867 vio->u.readwrite.vui_from = from;
870 vio->u.readwrite.vui_to = from + copied;
872 /* To address the deadlock in balance_dirty_pages() where
873 * this dirty page may be written back in the same thread. */
874 if (PageDirty(vmpage))
877 /* We may have one full RPC, commit it soon */
878 if (plist->pl_nr >= PTLRPC_MAX_BRW_PAGES)
881 CL_PAGE_DEBUG(D_VFSTRACE, env, page,
882 "queued page: %d.\n", plist->pl_nr);
884 cl_page_disown(env, io, page);
886 lcc->lcc_page = NULL;
887 lu_ref_del(&page->cp_reference, "cl_io", io);
888 cl_page_put(env, page);
890 /* page list is not contiguous now, commit it now */
893 if (unplug || io->u.ci_wr.wr_sync)
894 result = vvp_io_write_commit(env, io);
897 io->ci_result = result;
901 RETURN(result >= 0 ? copied : result);
904 #ifdef CONFIG_MIGRATION
905 static int ll_migratepage(struct address_space *mapping,
906 struct page *newpage, struct page *page,
907 enum migrate_mode mode)
909 /* Always fail page migration until we have a proper implementation */
914 const struct address_space_operations ll_aops = {
915 .readpage = ll_readpage,
916 .direct_IO = ll_direct_IO,
917 .writepage = ll_writepage,
918 .writepages = ll_writepages,
919 .set_page_dirty = __set_page_dirty_nobuffers,
920 .write_begin = ll_write_begin,
921 .write_end = ll_write_end,
922 .invalidatepage = ll_invalidatepage,
923 .releasepage = (void *)ll_releasepage,
924 #ifdef CONFIG_MIGRATION
925 .migratepage = ll_migratepage,