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.sun.com/software/products/lustre/docs/GPLv2.pdf
20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21 * CA 95054 USA or visit www.sun.com if you need additional information or
27 * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2011, 2012, Intel Corporation.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
36 * lustre/lustre/llite/rw26.c
38 * Lustre Lite I/O page cache routines for the 2.5/2.6 kernel version
41 #include <linux/kernel.h>
43 #include <linux/string.h>
44 #include <linux/stat.h>
45 #include <linux/errno.h>
46 #include <linux/unistd.h>
47 #include <asm/uaccess.h>
50 #include <linux/migrate.h>
51 #elif defined(HAVE_MIGRATE_MODE_H)
52 #include <linux/migrate_mode.h>
55 #include <linux/buffer_head.h>
56 #include <linux/mpage.h>
57 #include <linux/writeback.h>
58 #include <linux/stat.h>
59 #include <asm/uaccess.h>
61 #include <linux/pagemap.h>
63 #define DEBUG_SUBSYSTEM S_LLITE
65 #include <lustre_lite.h>
66 #include "llite_internal.h"
67 #include <linux/lustre_compat25.h>
70 * Implements Linux VM address_space::invalidatepage() method. This method is
71 * called when the page is truncate from a file, either as a result of
72 * explicit truncate, or when inode is removed from memory (as a result of
73 * final iput(), umount, or memory pressure induced icache shrinking).
75 * [0, offset] bytes of the page remain valid (this is for a case of not-page
76 * aligned truncate). Lustre leaves partially truncated page in the cache,
77 * relying on struct inode::i_size to limit further accesses.
79 static void ll_invalidatepage(struct page *vmpage, unsigned long offset)
84 struct cl_object *obj;
88 LASSERT(PageLocked(vmpage));
89 LASSERT(!PageWriteback(vmpage));
92 * It is safe to not check anything in invalidatepage/releasepage
93 * below because they are run with page locked and all our io is
94 * happening with locked page too
97 env = cl_env_get(&refcheck);
99 inode = vmpage->mapping->host;
100 obj = ll_i2info(inode)->lli_clob;
102 page = cl_vmpage_page(vmpage, obj);
104 cl_page_delete(env, page);
105 cl_page_put(env, page);
108 LASSERT(vmpage->private == 0);
109 cl_env_put(env, &refcheck);
114 #ifdef HAVE_RELEASEPAGE_WITH_INT
115 #define RELEASEPAGE_ARG_TYPE int
117 #define RELEASEPAGE_ARG_TYPE gfp_t
119 static int ll_releasepage(struct page *vmpage, RELEASEPAGE_ARG_TYPE gfp_mask)
123 struct cl_object *obj;
124 struct cl_page *page;
125 struct address_space *mapping;
128 LASSERT(PageLocked(vmpage));
129 if (PageWriteback(vmpage) || PageDirty(vmpage))
132 mapping = vmpage->mapping;
136 obj = ll_i2info(mapping->host)->lli_clob;
140 /* 1 for caller, 1 for cl_page and 1 for page cache */
141 if (page_count(vmpage) > 3)
144 page = cl_vmpage_page(vmpage, obj);
148 cookie = cl_env_reenter();
149 env = cl_env_percpu_get();
150 LASSERT(!IS_ERR(env));
152 if (!cl_page_in_use(page)) {
154 cl_page_delete(env, page);
157 /* To use percpu env array, the call path can not be rescheduled;
158 * otherwise percpu array will be messed if ll_releaspage() called
159 * again on the same CPU.
161 * If this page holds the last refc of cl_object, the following
162 * call path may cause reschedule:
163 * cl_page_put -> cl_page_free -> cl_object_put ->
164 * lu_object_put -> lu_object_free -> lov_delete_raid0 ->
167 * However, the kernel can't get rid of this inode until all pages have
168 * been cleaned up. Now that we hold page lock here, it's pretty safe
169 * that we won't get into object delete path.
171 LASSERT(cl_object_refc(obj) > 1);
172 cl_page_put(env, page);
174 cl_env_percpu_put(env);
175 cl_env_reexit(cookie);
179 static int ll_set_page_dirty(struct page *vmpage)
182 struct cl_page *page = vvp_vmpage_page_transient(vmpage);
183 struct vvp_object *obj = cl_inode2vvp(vmpage->mapping->host);
184 struct vvp_page *cpg;
187 * XXX should page method be called here?
189 LASSERT(&obj->co_cl == page->cp_obj);
190 cpg = cl2vvp_page(cl_page_at(page, &vvp_device_type));
192 * XXX cannot do much here, because page is possibly not locked:
194 * ->unmap_page_range()->zap_pte_range()->set_page_dirty().
196 vvp_write_pending(obj, cpg);
198 RETURN(__set_page_dirty_nobuffers(vmpage));
201 #define MAX_DIRECTIO_SIZE 2*1024*1024*1024UL
203 static inline int ll_get_user_pages(int rw, unsigned long user_addr,
204 size_t size, struct page ***pages,
207 int result = -ENOMEM;
209 /* set an arbitrary limit to prevent arithmetic overflow */
210 if (size > MAX_DIRECTIO_SIZE) {
215 *max_pages = (user_addr + size + PAGE_CACHE_SIZE - 1) >>
217 *max_pages -= user_addr >> PAGE_CACHE_SHIFT;
219 OBD_ALLOC_LARGE(*pages, *max_pages * sizeof(**pages));
221 down_read(¤t->mm->mmap_sem);
222 result = get_user_pages(current, current->mm, user_addr,
223 *max_pages, (rw == READ), 0, *pages,
225 up_read(¤t->mm->mmap_sem);
226 if (unlikely(result <= 0))
227 OBD_FREE_LARGE(*pages, *max_pages * sizeof(**pages));
233 /* ll_free_user_pages - tear down page struct array
234 * @pages: array of page struct pointers underlying target buffer */
235 static void ll_free_user_pages(struct page **pages, int npages, int do_dirty)
239 for (i = 0; i < npages; i++) {
240 if (pages[i] == NULL)
243 set_page_dirty_lock(pages[i]);
244 page_cache_release(pages[i]);
247 OBD_FREE_LARGE(pages, npages * sizeof(*pages));
250 ssize_t ll_direct_rw_pages(const struct lu_env *env, struct cl_io *io,
251 int rw, struct inode *inode,
252 struct ll_dio_pages *pv)
255 struct cl_2queue *queue;
256 struct cl_object *obj = io->ci_obj;
259 loff_t file_offset = pv->ldp_start_offset;
260 long size = pv->ldp_size;
261 int page_count = pv->ldp_nr;
262 struct page **pages = pv->ldp_pages;
263 long page_size = cl_page_size(obj);
268 queue = &io->ci_queue;
269 cl_2queue_init(queue);
270 for (i = 0; i < page_count; i++) {
272 file_offset = pv->ldp_offsets[i];
274 LASSERT(!(file_offset & (page_size - 1)));
275 clp = cl_page_find(env, obj, cl_index(obj, file_offset),
276 pv->ldp_pages[i], CPT_TRANSIENT);
282 rc = cl_page_own(env, io, clp);
284 LASSERT(clp->cp_state == CPS_FREEING);
285 cl_page_put(env, clp);
291 /* check the page type: if the page is a host page, then do
293 if (clp->cp_type == CPT_CACHEABLE) {
294 struct page *vmpage = cl_page_vmpage(env, clp);
295 struct page *src_page;
296 struct page *dst_page;
300 src_page = (rw == WRITE) ? pages[i] : vmpage;
301 dst_page = (rw == WRITE) ? vmpage : pages[i];
303 src = ll_kmap_atomic(src_page, KM_USER0);
304 dst = ll_kmap_atomic(dst_page, KM_USER1);
305 memcpy(dst, src, min(page_size, size));
306 ll_kunmap_atomic(dst, KM_USER1);
307 ll_kunmap_atomic(src, KM_USER0);
309 /* make sure page will be added to the transfer by
310 * cl_io_submit()->...->vvp_page_prep_write(). */
312 set_page_dirty(vmpage);
315 /* do not issue the page for read, since it
316 * may reread a ra page which has NOT uptodate
318 cl_page_disown(env, io, clp);
324 cl_2queue_add(queue, clp);
327 * Set page clip to tell transfer formation engine
328 * that page has to be sent even if it is beyond KMS.
330 cl_page_clip(env, clp, 0, min(size, page_size));
335 /* drop the reference count for cl_page_find */
336 cl_page_put(env, clp);
338 file_offset += page_size;
341 if (rc == 0 && io_pages) {
342 rc = cl_io_submit_sync(env, io,
343 rw == READ ? CRT_READ : CRT_WRITE,
349 cl_2queue_discard(env, io, queue);
350 cl_2queue_disown(env, io, queue);
351 cl_2queue_fini(env, queue);
354 EXPORT_SYMBOL(ll_direct_rw_pages);
356 static ssize_t ll_direct_IO_26_seg(const struct lu_env *env, struct cl_io *io,
357 int rw, struct inode *inode,
358 struct address_space *mapping,
359 size_t size, loff_t file_offset,
360 struct page **pages, int page_count)
362 struct ll_dio_pages pvec = { .ldp_pages = pages,
363 .ldp_nr = page_count,
366 .ldp_start_offset = file_offset
369 return ll_direct_rw_pages(env, io, rw, inode, &pvec);
372 #ifdef KMALLOC_MAX_SIZE
373 #define MAX_MALLOC KMALLOC_MAX_SIZE
375 #define MAX_MALLOC (128 * 1024)
378 /* This is the maximum size of a single O_DIRECT request, based on the
379 * kmalloc limit. We need to fit all of the brw_page structs, each one
380 * representing PAGE_SIZE worth of user data, into a single buffer, and
381 * then truncate this to be a full-sized RPC. For 4kB PAGE_SIZE this is
382 * up to 22MB for 128kB kmalloc and up to 682MB for 4MB kmalloc. */
383 #define MAX_DIO_SIZE ((MAX_MALLOC / sizeof(struct brw_page) * PAGE_CACHE_SIZE) & \
384 ~(DT_MAX_BRW_SIZE - 1))
385 static ssize_t ll_direct_IO_26(int rw, struct kiocb *iocb,
386 const struct iovec *iov, loff_t file_offset,
387 unsigned long nr_segs)
391 struct file *file = iocb->ki_filp;
392 struct inode *inode = file->f_mapping->host;
393 struct ccc_object *obj = cl_inode2ccc(inode);
394 long count = iov_length(iov, nr_segs);
395 long tot_bytes = 0, result = 0;
396 struct ll_inode_info *lli = ll_i2info(inode);
397 unsigned long seg = 0;
398 long size = MAX_DIO_SIZE;
402 if (!lli->lli_has_smd)
405 /* FIXME: io smaller than PAGE_SIZE is broken on ia64 ??? */
406 if ((file_offset & ~CFS_PAGE_MASK) || (count & ~CFS_PAGE_MASK))
409 CDEBUG(D_VFSTRACE, "VFS Op:inode="DFID"(%p), size=%lu (max %lu), "
410 "offset=%lld=%llx, pages %lu (max %lu)\n",
411 PFID(ll_inode2fid(inode)), inode, count, MAX_DIO_SIZE,
412 file_offset, file_offset, count >> PAGE_CACHE_SHIFT,
413 MAX_DIO_SIZE >> PAGE_CACHE_SHIFT);
415 /* Check that all user buffers are aligned as well */
416 for (seg = 0; seg < nr_segs; seg++) {
417 if (((unsigned long)iov[seg].iov_base & ~CFS_PAGE_MASK) ||
418 (iov[seg].iov_len & ~CFS_PAGE_MASK))
422 env = cl_env_get(&refcheck);
423 LASSERT(!IS_ERR(env));
424 io = ccc_env_io(env)->cui_cl.cis_io;
427 /* 0. Need locking between buffered and direct access. and race with
428 * size changing by concurrent truncates and writes.
429 * 1. Need inode mutex to operate transient pages.
432 mutex_lock(&inode->i_mutex);
434 LASSERT(obj->cob_transient_pages == 0);
435 for (seg = 0; seg < nr_segs; seg++) {
436 long iov_left = iov[seg].iov_len;
437 unsigned long user_addr = (unsigned long)iov[seg].iov_base;
440 if (file_offset >= i_size_read(inode))
442 if (file_offset + iov_left > i_size_read(inode))
443 iov_left = i_size_read(inode) - file_offset;
446 while (iov_left > 0) {
448 int page_count, max_pages = 0;
451 bytes = min(size, iov_left);
452 page_count = ll_get_user_pages(rw, user_addr, bytes,
454 if (likely(page_count > 0)) {
455 if (unlikely(page_count < max_pages))
456 bytes = page_count << PAGE_CACHE_SHIFT;
457 result = ll_direct_IO_26_seg(env, io, rw, inode,
461 ll_free_user_pages(pages, max_pages, rw==READ);
462 } else if (page_count == 0) {
463 GOTO(out, result = -EFAULT);
467 if (unlikely(result <= 0)) {
468 /* If we can't allocate a large enough buffer
469 * for the request, shrink it to a smaller
470 * PAGE_SIZE multiple and try again.
471 * We should always be able to kmalloc for a
472 * page worth of page pointers = 4MB on i386. */
473 if (result == -ENOMEM &&
474 size > (PAGE_CACHE_SIZE / sizeof(*pages)) *
476 size = ((((size / 2) - 1) |
477 ~CFS_PAGE_MASK) + 1) &
479 CDEBUG(D_VFSTRACE,"DIO size now %lu\n",
488 file_offset += result;
494 LASSERT(obj->cob_transient_pages == 0);
496 mutex_unlock(&inode->i_mutex);
499 struct ccc_io *cio = ccc_env_io(env);
501 /* no commit async for direct IO */
502 cio->u.write.cui_written += tot_bytes;
505 cl_env_put(env, &refcheck);
506 RETURN(tot_bytes ? tot_bytes : result);
510 * Prepare partially written-to page for a write.
512 static int ll_prepare_partial_page(const struct lu_env *env, struct cl_io *io,
515 struct cl_object *obj = io->ci_obj;
516 struct cl_attr *attr = ccc_env_thread_attr(env);
517 loff_t offset = cl_offset(obj, pg->cp_index);
520 cl_object_attr_lock(obj);
521 result = cl_object_attr_get(env, obj, attr);
522 cl_object_attr_unlock(obj);
526 cp = cl2ccc_page(cl_page_at(pg, &vvp_device_type));
529 * If are writing to a new page, no need to read old data.
530 * The extent locking will have updated the KMS, and for our
531 * purposes here we can treat it like i_size.
533 if (attr->cat_kms <= offset) {
534 char *kaddr = ll_kmap_atomic(cp->cpg_page, KM_USER0);
536 memset(kaddr, 0, cl_page_size(obj));
537 ll_kunmap_atomic(kaddr, KM_USER0);
538 } else if (cp->cpg_defer_uptodate)
541 result = ll_page_sync_io(env, io, pg, CRT_READ);
546 static int ll_write_begin(struct file *file, struct address_space *mapping,
547 loff_t pos, unsigned len, unsigned flags,
548 struct page **pagep, void **fsdata)
550 struct ll_cl_context *lcc;
553 struct cl_page *page;
555 struct cl_object *clob = ll_i2info(mapping->host)->lli_clob;
556 pgoff_t index = pos >> PAGE_CACHE_SHIFT;
557 struct page *vmpage = NULL;
558 unsigned from = pos & (PAGE_CACHE_SIZE - 1);
559 unsigned to = from + len;
563 CDEBUG(D_VFSTRACE, "Writing %lu of %d to %d bytes\n", index, from, len);
565 lcc = ll_cl_init(file, NULL);
567 GOTO(out, result = PTR_ERR(lcc));
572 /* To avoid deadlock, try to lock page first. */
573 vmpage = grab_cache_page_nowait(mapping, index);
574 if (unlikely(vmpage == NULL || PageDirty(vmpage))) {
575 struct ccc_io *cio = ccc_env_io(env);
576 struct cl_page_list *plist = &cio->u.write.cui_queue;
578 /* if the page is already in dirty cache, we have to commit
579 * the pages right now; otherwise, it may cause deadlock
580 * because it holds page lock of a dirty page and request for
581 * more grants. It's okay for the dirty page to be the first
582 * one in commit page list, though. */
583 if (vmpage != NULL && PageDirty(vmpage) && plist->pl_nr > 0) {
585 page_cache_release(vmpage);
589 /* commit pages and then wait for page lock */
590 result = vvp_io_write_commit(env, io);
594 if (vmpage == NULL) {
595 vmpage = grab_cache_page_write_begin(mapping, index,
598 GOTO(out, result = -ENOMEM);
602 page = cl_page_find(env, clob, vmpage->index, vmpage, CPT_CACHEABLE);
604 GOTO(out, result = PTR_ERR(page));
606 lcc->lcc_page = page;
607 lu_ref_add(&page->cp_reference, "cl_io", io);
609 cl_page_assume(env, io, page);
610 if (!PageUptodate(vmpage)) {
612 * We're completely overwriting an existing page,
613 * so _don't_ set it up to date until commit_write
615 if (from == 0 && to == PAGE_SIZE) {
616 CL_PAGE_HEADER(D_PAGE, env, page, "full page write\n");
617 POISON_PAGE(vmpage, 0x11);
619 /* TODO: can be optimized at OSC layer to check if it
620 * is a lockless IO. In that case, it's not necessary
621 * to read the data. */
622 result = ll_prepare_partial_page(env, io, page);
624 SetPageUptodate(vmpage);
628 cl_page_unassume(env, io, page);
632 if (vmpage != NULL) {
634 page_cache_release(vmpage);
645 static int ll_write_end(struct file *file, struct address_space *mapping,
646 loff_t pos, unsigned len, unsigned copied,
647 struct page *vmpage, void *fsdata)
649 struct ll_cl_context *lcc = fsdata;
653 struct cl_page *page;
654 unsigned from = pos & (PAGE_CACHE_SIZE - 1);
659 page_cache_release(vmpage);
661 LASSERT(lcc != NULL);
663 page = lcc->lcc_page;
665 cio = ccc_env_io(env);
667 LASSERT(cl_page_is_owned(page, io));
669 struct cl_page_list *plist = &cio->u.write.cui_queue;
671 lcc->lcc_page = NULL; /* page will be queued */
673 /* Add it into write queue */
674 cl_page_list_add(plist, page);
675 if (plist->pl_nr == 1) /* first page */
676 cio->u.write.cui_from = from;
679 cio->u.write.cui_to = from + copied;
681 /* We may have one full RPC, commit it soon */
682 if (plist->pl_nr >= PTLRPC_MAX_BRW_PAGES)
685 CL_PAGE_DEBUG(D_VFSTRACE, env, page,
686 "queued page: %d.\n", plist->pl_nr);
688 cl_page_disown(env, io, page);
690 /* page list is not contiguous now, commit it now */
695 file->f_flags & O_SYNC || IS_SYNC(file->f_dentry->d_inode))
696 result = vvp_io_write_commit(env, io);
699 RETURN(result >= 0 ? copied : result);
702 #ifdef CONFIG_MIGRATION
703 int ll_migratepage(struct address_space *mapping,
704 struct page *newpage, struct page *page
705 #ifdef HAVE_MIGRATEPAGE_4ARGS
706 , enum migrate_mode mode
710 /* Always fail page migration until we have a proper implementation */
715 #ifndef MS_HAS_NEW_AOPS
716 struct address_space_operations ll_aops = {
717 .readpage = ll_readpage,
718 // .readpages = ll_readpages,
719 .direct_IO = ll_direct_IO_26,
720 .writepage = ll_writepage,
721 .writepages = ll_writepages,
722 .set_page_dirty = ll_set_page_dirty,
723 .write_begin = ll_write_begin,
724 .write_end = ll_write_end,
725 .invalidatepage = ll_invalidatepage,
726 .releasepage = (void *)ll_releasepage,
727 #ifdef CONFIG_MIGRATION
728 .migratepage = ll_migratepage,
733 struct address_space_operations_ext ll_aops = {
734 .orig_aops.readpage = ll_readpage,
735 .orig_aops.direct_IO = ll_direct_IO_26,
736 .orig_aops.writepage = ll_writepage,
737 .orig_aops.writepages = ll_writepages,
738 .orig_aops.set_page_dirty = ll_set_page_dirty,
739 .orig_aops.invalidatepage = ll_invalidatepage,
740 .orig_aops.releasepage = ll_releasepage,
741 #ifdef CONFIG_MIGRATION
742 .orig_aops.migratepage = ll_migratepage,
744 .orig_aops.bmap = NULL,
745 .write_begin = ll_write_begin,
746 .write_end = ll_write_end