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, 2016, 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/kernel.h>
39 #include <linux/string.h>
40 #include <linux/stat.h>
41 #include <linux/errno.h>
42 #include <linux/unistd.h>
43 #include <asm/uaccess.h>
46 #include <linux/migrate.h>
47 #elif defined(HAVE_MIGRATE_MODE_H)
48 #include <linux/migrate_mode.h>
51 #include <linux/buffer_head.h>
52 #include <linux/mpage.h>
53 #include <linux/writeback.h>
54 #include <linux/stat.h>
55 #include <asm/uaccess.h>
57 #include <linux/pagemap.h>
59 #define DEBUG_SUBSYSTEM S_LLITE
61 #include "llite_internal.h"
62 #include <lustre_compat.h>
65 * Implements Linux VM address_space::invalidatepage() method. This method is
66 * called when the page is truncate from a file, either as a result of
67 * explicit truncate, or when inode is removed from memory (as a result of
68 * final iput(), umount, or memory pressure induced icache shrinking).
70 * [0, offset] bytes of the page remain valid (this is for a case of not-page
71 * aligned truncate). Lustre leaves partially truncated page in the cache,
72 * relying on struct inode::i_size to limit further accesses.
74 static void ll_invalidatepage(struct page *vmpage,
75 #ifdef HAVE_INVALIDATE_RANGE
76 unsigned int offset, unsigned int length
85 struct cl_object *obj;
87 LASSERT(PageLocked(vmpage));
88 LASSERT(!PageWriteback(vmpage));
91 * It is safe to not check anything in invalidatepage/releasepage
92 * below because they are run with page locked and all our io is
93 * happening with locked page too
95 #ifdef HAVE_INVALIDATE_RANGE
96 if (offset == 0 && length == PAGE_SIZE) {
100 /* See the comment in ll_releasepage() */
101 env = cl_env_percpu_get();
102 LASSERT(!IS_ERR(env));
104 inode = vmpage->mapping->host;
105 obj = ll_i2info(inode)->lli_clob;
107 page = cl_vmpage_page(vmpage, obj);
109 cl_page_delete(env, page);
110 cl_page_put(env, page);
113 LASSERT(vmpage->private == 0);
115 cl_env_percpu_put(env);
119 #ifdef HAVE_RELEASEPAGE_WITH_INT
120 #define RELEASEPAGE_ARG_TYPE int
122 #define RELEASEPAGE_ARG_TYPE gfp_t
124 static int ll_releasepage(struct page *vmpage, RELEASEPAGE_ARG_TYPE gfp_mask)
127 struct cl_object *obj;
128 struct cl_page *page;
129 struct address_space *mapping;
132 LASSERT(PageLocked(vmpage));
133 if (PageWriteback(vmpage) || PageDirty(vmpage))
136 mapping = vmpage->mapping;
140 obj = ll_i2info(mapping->host)->lli_clob;
144 /* 1 for caller, 1 for cl_page and 1 for page cache */
145 if (page_count(vmpage) > 3)
148 page = cl_vmpage_page(vmpage, obj);
152 env = cl_env_percpu_get();
153 LASSERT(!IS_ERR(env));
155 if (!cl_page_in_use(page)) {
157 cl_page_delete(env, page);
160 /* To use percpu env array, the call path can not be rescheduled;
161 * otherwise percpu array will be messed if ll_releaspage() called
162 * again on the same CPU.
164 * If this page holds the last refc of cl_object, the following
165 * call path may cause reschedule:
166 * cl_page_put -> cl_page_free -> cl_object_put ->
167 * lu_object_put -> lu_object_free -> lov_delete_raid0.
169 * However, the kernel can't get rid of this inode until all pages have
170 * been cleaned up. Now that we hold page lock here, it's pretty safe
171 * that we won't get into object delete path.
173 LASSERT(cl_object_refc(obj) > 1);
174 cl_page_put(env, page);
176 cl_env_percpu_put(env);
180 #define MAX_DIRECTIO_SIZE 2*1024*1024*1024UL
183 ll_direct_IO_seg(const struct lu_env *env, struct cl_io *io, int rw,
184 struct inode *inode, size_t size, loff_t file_offset,
185 struct page **pages, int page_count)
188 struct cl_2queue *queue;
189 struct cl_object *obj = io->ci_obj;
192 size_t page_size = cl_page_size(obj);
193 size_t orig_size = size;
198 queue = &io->ci_queue;
199 cl_2queue_init(queue);
200 for (i = 0; i < page_count; i++) {
201 LASSERT(!(file_offset & (page_size - 1)));
202 clp = cl_page_find(env, obj, cl_index(obj, file_offset),
203 pages[i], CPT_TRANSIENT);
209 rc = cl_page_own(env, io, clp);
211 LASSERT(clp->cp_state == CPS_FREEING);
212 cl_page_put(env, clp);
218 /* check the page type: if the page is a host page, then do
221 if (clp->cp_type == CPT_CACHEABLE) {
222 struct page *vmpage = cl_page_vmpage(clp);
223 struct page *src_page;
224 struct page *dst_page;
228 src_page = (rw == WRITE) ? pages[i] : vmpage;
229 dst_page = (rw == WRITE) ? vmpage : pages[i];
231 src = ll_kmap_atomic(src_page, KM_USER0);
232 dst = ll_kmap_atomic(dst_page, KM_USER1);
233 memcpy(dst, src, min(page_size, size));
234 ll_kunmap_atomic(dst, KM_USER1);
235 ll_kunmap_atomic(src, KM_USER0);
237 /* make sure page will be added to the transfer by
238 * cl_io_submit()->...->vvp_page_prep_write().
241 set_page_dirty(vmpage);
244 /* do not issue the page for read, since it
245 * may reread a ra page which has NOT uptodate
248 cl_page_disown(env, io, clp);
254 cl_2queue_add(queue, clp);
257 * Set page clip to tell transfer formation engine
258 * that page has to be sent even if it is beyond KMS.
260 cl_page_clip(env, clp, 0, min(size, page_size));
265 /* drop the reference count for cl_page_find */
266 cl_page_put(env, clp);
268 file_offset += page_size;
271 if (rc == 0 && io_pages) {
272 rc = cl_io_submit_sync(env, io,
273 rw == READ ? CRT_READ : CRT_WRITE,
279 cl_2queue_discard(env, io, queue);
280 cl_2queue_disown(env, io, queue);
281 cl_2queue_fini(env, queue);
285 /* ll_free_user_pages - tear down page struct array
286 * @pages: array of page struct pointers underlying target buffer */
287 static void ll_free_user_pages(struct page **pages, int npages, int do_dirty)
291 for (i = 0; i < npages; i++) {
292 if (pages[i] == NULL)
295 set_page_dirty_lock(pages[i]);
299 #if defined(HAVE_DIRECTIO_ITER) || defined(HAVE_IOV_ITER_RW)
302 OBD_FREE_LARGE(pages, npages * sizeof(*pages));
306 #ifdef KMALLOC_MAX_SIZE
307 #define MAX_MALLOC KMALLOC_MAX_SIZE
309 #define MAX_MALLOC (128 * 1024)
312 /* This is the maximum size of a single O_DIRECT request, based on the
313 * kmalloc limit. We need to fit all of the brw_page structs, each one
314 * representing PAGE_SIZE worth of user data, into a single buffer, and
315 * then truncate this to be a full-sized RPC. For 4kB PAGE_SIZE this is
316 * up to 22MB for 128kB kmalloc and up to 682MB for 4MB kmalloc. */
317 #define MAX_DIO_SIZE ((MAX_MALLOC / sizeof(struct brw_page) * PAGE_SIZE) & \
318 ~(DT_MAX_BRW_SIZE - 1))
320 #ifndef HAVE_IOV_ITER_RW
321 # define iov_iter_rw(iter) rw
324 #if defined(HAVE_DIRECTIO_ITER) || defined(HAVE_IOV_ITER_RW)
327 # ifndef HAVE_IOV_ITER_RW
330 struct kiocb *iocb, struct iov_iter *iter
331 # ifndef HAVE_DIRECTIO_2ARGS
336 #ifdef HAVE_DIRECTIO_2ARGS
337 loff_t file_offset = iocb->ki_pos;
339 struct ll_cl_context *lcc;
340 const struct lu_env *env;
342 struct file *file = iocb->ki_filp;
343 struct inode *inode = file->f_mapping->host;
344 ssize_t count = iov_iter_count(iter);
345 ssize_t tot_bytes = 0, result = 0;
346 size_t size = MAX_DIO_SIZE;
348 /* Check EOF by ourselves */
349 if (iov_iter_rw(iter) == READ && file_offset >= i_size_read(inode))
351 /* FIXME: io smaller than PAGE_SIZE is broken on ia64 ??? */
352 if ((file_offset & ~PAGE_MASK) || (count & ~PAGE_MASK))
355 CDEBUG(D_VFSTRACE, "VFS Op:inode="DFID"(%p), size=%zd (max %lu), "
356 "offset=%lld=%llx, pages %zd (max %lu)\n",
357 PFID(ll_inode2fid(inode)), inode, count, MAX_DIO_SIZE,
358 file_offset, file_offset, count >> PAGE_SHIFT,
359 MAX_DIO_SIZE >> PAGE_SHIFT);
361 /* Check that all user buffers are aligned as well */
362 if (iov_iter_alignment(iter) & ~PAGE_MASK)
365 lcc = ll_cl_find(file);
370 LASSERT(!IS_ERR(env));
374 /* 0. Need locking between buffered and direct access. and race with
375 * size changing by concurrent truncates and writes.
376 * 1. Need inode mutex to operate transient pages.
378 if (iov_iter_rw(iter) == READ)
381 while (iov_iter_count(iter)) {
385 count = min_t(size_t, iov_iter_count(iter), size);
386 if (iov_iter_rw(iter) == READ) {
387 if (file_offset >= i_size_read(inode))
390 if (file_offset + count > i_size_read(inode))
391 count = i_size_read(inode) - file_offset;
394 result = iov_iter_get_pages_alloc(iter, &pages, count, &offs);
395 if (likely(result > 0)) {
396 int n = DIV_ROUND_UP(result + offs, PAGE_SIZE);
398 result = ll_direct_IO_seg(env, io, iov_iter_rw(iter),
399 inode, result, file_offset,
401 ll_free_user_pages(pages, n,
402 iov_iter_rw(iter) == READ);
405 if (unlikely(result <= 0)) {
406 /* If we can't allocate a large enough buffer
407 * for the request, shrink it to a smaller
408 * PAGE_SIZE multiple and try again.
409 * We should always be able to kmalloc for a
410 * page worth of page pointers = 4MB on i386. */
411 if (result == -ENOMEM &&
412 size > (PAGE_SIZE / sizeof(*pages)) *
414 size = ((((size / 2) - 1) |
415 ~PAGE_MASK) + 1) & PAGE_MASK;
416 CDEBUG(D_VFSTRACE, "DIO size now %zu\n",
424 iov_iter_advance(iter, result);
426 file_offset += result;
429 if (iov_iter_rw(iter) == READ)
433 struct vvp_io *vio = vvp_env_io(env);
435 /* no commit async for direct IO */
436 vio->u.write.vui_written += tot_bytes;
439 return tot_bytes ? : result;
441 #else /* !HAVE_DIRECTIO_ITER && !HAVE_IOV_ITER_RW */
443 static inline int ll_get_user_pages(int rw, unsigned long user_addr,
444 size_t size, struct page ***pages,
447 int result = -ENOMEM;
449 /* set an arbitrary limit to prevent arithmetic overflow */
450 if (size > MAX_DIRECTIO_SIZE) {
455 *max_pages = (user_addr + size + PAGE_SIZE - 1) >>
457 *max_pages -= user_addr >> PAGE_SHIFT;
459 OBD_ALLOC_LARGE(*pages, *max_pages * sizeof(**pages));
461 down_read(¤t->mm->mmap_sem);
462 result = get_user_pages(current, current->mm, user_addr,
463 *max_pages, (rw == READ), 0, *pages,
465 up_read(¤t->mm->mmap_sem);
466 if (unlikely(result <= 0))
467 OBD_FREE_LARGE(*pages, *max_pages * sizeof(**pages));
474 ll_direct_IO(int rw, struct kiocb *iocb, const struct iovec *iov,
475 loff_t file_offset, unsigned long nr_segs)
477 struct ll_cl_context *lcc;
478 const struct lu_env *env;
480 struct file *file = iocb->ki_filp;
481 struct inode *inode = file->f_mapping->host;
482 ssize_t count = iov_length(iov, nr_segs);
483 ssize_t tot_bytes = 0, result = 0;
484 unsigned long seg = 0;
485 size_t size = MAX_DIO_SIZE;
488 /* FIXME: io smaller than PAGE_SIZE is broken on ia64 ??? */
489 if ((file_offset & ~PAGE_MASK) || (count & ~PAGE_MASK))
492 CDEBUG(D_VFSTRACE, "VFS Op:inode="DFID"(%p), size=%zd (max %lu), "
493 "offset=%lld=%llx, pages %zd (max %lu)\n",
494 PFID(ll_inode2fid(inode)), inode, count, MAX_DIO_SIZE,
495 file_offset, file_offset, count >> PAGE_SHIFT,
496 MAX_DIO_SIZE >> PAGE_SHIFT);
498 /* Check that all user buffers are aligned as well */
499 for (seg = 0; seg < nr_segs; seg++) {
500 if (((unsigned long)iov[seg].iov_base & ~PAGE_MASK) ||
501 (iov[seg].iov_len & ~PAGE_MASK))
505 lcc = ll_cl_find(file);
510 LASSERT(!IS_ERR(env));
514 for (seg = 0; seg < nr_segs; seg++) {
515 size_t iov_left = iov[seg].iov_len;
516 unsigned long user_addr = (unsigned long)iov[seg].iov_base;
519 if (file_offset >= i_size_read(inode))
521 if (file_offset + iov_left > i_size_read(inode))
522 iov_left = i_size_read(inode) - file_offset;
525 while (iov_left > 0) {
527 int page_count, max_pages = 0;
530 bytes = min(size, iov_left);
531 page_count = ll_get_user_pages(rw, user_addr, bytes,
533 if (likely(page_count > 0)) {
534 if (unlikely(page_count < max_pages))
535 bytes = page_count << PAGE_SHIFT;
536 result = ll_direct_IO_seg(env, io, rw, inode,
539 ll_free_user_pages(pages, max_pages, rw==READ);
540 } else if (page_count == 0) {
541 GOTO(out, result = -EFAULT);
545 if (unlikely(result <= 0)) {
546 /* If we can't allocate a large enough buffer
547 * for the request, shrink it to a smaller
548 * PAGE_SIZE multiple and try again.
549 * We should always be able to kmalloc for a
550 * page worth of page pointers = 4MB on i386. */
551 if (result == -ENOMEM &&
552 size > (PAGE_SIZE / sizeof(*pages)) *
554 size = ((((size / 2) - 1) |
557 CDEBUG(D_VFSTRACE, "DIO size now %zu\n",
566 file_offset += result;
573 struct vvp_io *vio = vvp_env_io(env);
575 /* no commit async for direct IO */
576 vio->u.write.vui_written += tot_bytes;
579 RETURN(tot_bytes ? tot_bytes : result);
581 #endif /* HAVE_DIRECTIO_ITER || HAVE_IOV_ITER_RW */
584 * Prepare partially written-to page for a write.
586 static int ll_prepare_partial_page(const struct lu_env *env, struct cl_io *io,
589 struct cl_attr *attr = vvp_env_thread_attr(env);
590 struct cl_object *obj = io->ci_obj;
591 struct vvp_page *vpg = cl_object_page_slice(obj, pg);
592 loff_t offset = cl_offset(obj, vvp_index(vpg));
595 cl_object_attr_lock(obj);
596 result = cl_object_attr_get(env, obj, attr);
597 cl_object_attr_unlock(obj);
600 * If are writing to a new page, no need to read old data.
601 * The extent locking will have updated the KMS, and for our
602 * purposes here we can treat it like i_size.
604 if (attr->cat_kms <= offset) {
605 char *kaddr = ll_kmap_atomic(vpg->vpg_page, KM_USER0);
607 memset(kaddr, 0, cl_page_size(obj));
608 ll_kunmap_atomic(kaddr, KM_USER0);
609 } else if (vpg->vpg_defer_uptodate)
610 vpg->vpg_ra_used = 1;
612 result = ll_page_sync_io(env, io, pg, CRT_READ);
617 static int ll_write_begin(struct file *file, struct address_space *mapping,
618 loff_t pos, unsigned len, unsigned flags,
619 struct page **pagep, void **fsdata)
621 struct ll_cl_context *lcc;
622 const struct lu_env *env = NULL;
624 struct cl_page *page = NULL;
626 struct cl_object *clob = ll_i2info(mapping->host)->lli_clob;
627 pgoff_t index = pos >> PAGE_SHIFT;
628 struct page *vmpage = NULL;
629 unsigned from = pos & (PAGE_SIZE - 1);
630 unsigned to = from + len;
634 CDEBUG(D_PAGE, "Writing %lu of %d to %d bytes\n", index, from, len);
636 lcc = ll_cl_find(file);
639 GOTO(out, result = -EIO);
645 if (file->f_flags & O_DIRECT && io->ci_designated_mirror > 0) {
646 /* direct IO failed because it couldn't clean up cached pages,
647 * this causes a problem for mirror write because the cached
648 * page may belong to another mirror, which will result in
649 * problem submitting the I/O. */
650 GOTO(out, result = -EBUSY);
653 /* To avoid deadlock, try to lock page first. */
654 vmpage = grab_cache_page_nowait(mapping, index);
656 if (unlikely(vmpage == NULL ||
657 PageDirty(vmpage) || PageWriteback(vmpage))) {
658 struct vvp_io *vio = vvp_env_io(env);
659 struct cl_page_list *plist = &vio->u.write.vui_queue;
661 /* if the page is already in dirty cache, we have to commit
662 * the pages right now; otherwise, it may cause deadlock
663 * because it holds page lock of a dirty page and request for
664 * more grants. It's okay for the dirty page to be the first
665 * one in commit page list, though. */
666 if (vmpage != NULL && plist->pl_nr > 0) {
672 /* commit pages and then wait for page lock */
673 result = vvp_io_write_commit(env, io);
677 if (vmpage == NULL) {
678 vmpage = grab_cache_page_write_begin(mapping, index,
681 GOTO(out, result = -ENOMEM);
685 page = cl_page_find(env, clob, vmpage->index, vmpage, CPT_CACHEABLE);
687 GOTO(out, result = PTR_ERR(page));
689 lcc->lcc_page = page;
690 lu_ref_add(&page->cp_reference, "cl_io", io);
692 cl_page_assume(env, io, page);
693 if (!PageUptodate(vmpage)) {
695 * We're completely overwriting an existing page,
696 * so _don't_ set it up to date until commit_write
698 if (from == 0 && to == PAGE_SIZE) {
699 CL_PAGE_HEADER(D_PAGE, env, page, "full page write\n");
700 POISON_PAGE(vmpage, 0x11);
702 /* TODO: can be optimized at OSC layer to check if it
703 * is a lockless IO. In that case, it's not necessary
704 * to read the data. */
705 result = ll_prepare_partial_page(env, io, page);
707 SetPageUptodate(vmpage);
711 cl_page_unassume(env, io, page);
715 if (vmpage != NULL) {
719 if (!IS_ERR_OR_NULL(page)) {
720 lu_ref_del(&page->cp_reference, "cl_io", io);
721 cl_page_put(env, page);
724 io->ci_result = result;
732 static int ll_write_end(struct file *file, struct address_space *mapping,
733 loff_t pos, unsigned len, unsigned copied,
734 struct page *vmpage, void *fsdata)
736 struct ll_cl_context *lcc = fsdata;
737 const struct lu_env *env;
740 struct cl_page *page;
741 unsigned from = pos & (PAGE_SIZE - 1);
748 LASSERT(lcc != NULL);
750 page = lcc->lcc_page;
752 vio = vvp_env_io(env);
754 LASSERT(cl_page_is_owned(page, io));
756 struct cl_page_list *plist = &vio->u.write.vui_queue;
758 lcc->lcc_page = NULL; /* page will be queued */
760 /* Add it into write queue */
761 cl_page_list_add(plist, page);
762 if (plist->pl_nr == 1) /* first page */
763 vio->u.write.vui_from = from;
766 vio->u.write.vui_to = from + copied;
768 /* To address the deadlock in balance_dirty_pages() where
769 * this dirty page may be written back in the same thread. */
770 if (PageDirty(vmpage))
773 /* We may have one full RPC, commit it soon */
774 if (plist->pl_nr >= PTLRPC_MAX_BRW_PAGES)
777 CL_PAGE_DEBUG(D_PAGE, env, page,
778 "queued page: %d.\n", plist->pl_nr);
780 cl_page_disown(env, io, page);
782 lcc->lcc_page = NULL;
783 lu_ref_del(&page->cp_reference, "cl_io", io);
784 cl_page_put(env, page);
786 /* page list is not contiguous now, commit it now */
790 file->f_flags & O_SYNC || IS_SYNC(file_inode(file)))
791 result = vvp_io_write_commit(env, io);
794 io->ci_result = result;
795 RETURN(result >= 0 ? copied : result);
798 #ifdef CONFIG_MIGRATION
799 static int ll_migratepage(struct address_space *mapping,
800 struct page *newpage, struct page *page
801 #ifdef HAVE_MIGRATEPAGE_4ARGS
802 , enum migrate_mode mode
806 /* Always fail page migration until we have a proper implementation */
811 const struct address_space_operations ll_aops = {
812 .readpage = ll_readpage,
813 .direct_IO = ll_direct_IO,
814 .writepage = ll_writepage,
815 .writepages = ll_writepages,
816 .set_page_dirty = __set_page_dirty_nobuffers,
817 .write_begin = ll_write_begin,
818 .write_end = ll_write_end,
819 .invalidatepage = ll_invalidatepage,
820 .releasepage = (void *)ll_releasepage,
821 #ifdef CONFIG_MIGRATION
822 .migratepage = ll_migratepage,