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,
80 #ifdef HAVE_INVALIDATE_RANGE
81 unsigned int offset, unsigned int length
90 struct cl_object *obj;
94 LASSERT(PageLocked(vmpage));
95 LASSERT(!PageWriteback(vmpage));
98 * It is safe to not check anything in invalidatepage/releasepage
99 * below because they are run with page locked and all our io is
100 * happening with locked page too
102 #ifdef HAVE_INVALIDATE_RANGE
103 if (offset == 0 && length == PAGE_CACHE_SIZE) {
107 env = cl_env_get(&refcheck);
109 inode = vmpage->mapping->host;
110 obj = ll_i2info(inode)->lli_clob;
112 page = cl_vmpage_page(vmpage, obj);
114 cl_page_delete(env, page);
115 cl_page_put(env, page);
118 LASSERT(vmpage->private == 0);
119 cl_env_put(env, &refcheck);
124 #ifdef HAVE_RELEASEPAGE_WITH_INT
125 #define RELEASEPAGE_ARG_TYPE int
127 #define RELEASEPAGE_ARG_TYPE gfp_t
129 static int ll_releasepage(struct page *vmpage, RELEASEPAGE_ARG_TYPE gfp_mask)
133 struct cl_object *obj;
134 struct cl_page *page;
135 struct address_space *mapping;
138 LASSERT(PageLocked(vmpage));
139 if (PageWriteback(vmpage) || PageDirty(vmpage))
142 mapping = vmpage->mapping;
146 obj = ll_i2info(mapping->host)->lli_clob;
150 /* 1 for caller, 1 for cl_page and 1 for page cache */
151 if (page_count(vmpage) > 3)
154 page = cl_vmpage_page(vmpage, obj);
158 cookie = cl_env_reenter();
159 env = cl_env_percpu_get();
160 LASSERT(!IS_ERR(env));
162 if (!cl_page_in_use(page)) {
164 cl_page_delete(env, page);
167 /* To use percpu env array, the call path can not be rescheduled;
168 * otherwise percpu array will be messed if ll_releaspage() called
169 * again on the same CPU.
171 * If this page holds the last refc of cl_object, the following
172 * call path may cause reschedule:
173 * cl_page_put -> cl_page_free -> cl_object_put ->
174 * lu_object_put -> lu_object_free -> lov_delete_raid0 ->
177 * However, the kernel can't get rid of this inode until all pages have
178 * been cleaned up. Now that we hold page lock here, it's pretty safe
179 * that we won't get into object delete path.
181 LASSERT(cl_object_refc(obj) > 1);
182 cl_page_put(env, page);
184 cl_env_percpu_put(env);
185 cl_env_reexit(cookie);
189 #define MAX_DIRECTIO_SIZE 2*1024*1024*1024UL
191 static inline int ll_get_user_pages(int rw, unsigned long user_addr,
192 size_t size, struct page ***pages,
195 int result = -ENOMEM;
197 /* set an arbitrary limit to prevent arithmetic overflow */
198 if (size > MAX_DIRECTIO_SIZE) {
203 *max_pages = (user_addr + size + PAGE_CACHE_SIZE - 1) >>
205 *max_pages -= user_addr >> PAGE_CACHE_SHIFT;
207 OBD_ALLOC_LARGE(*pages, *max_pages * sizeof(**pages));
209 down_read(¤t->mm->mmap_sem);
210 result = get_user_pages(current, current->mm, user_addr,
211 *max_pages, (rw == READ), 0, *pages,
213 up_read(¤t->mm->mmap_sem);
214 if (unlikely(result <= 0))
215 OBD_FREE_LARGE(*pages, *max_pages * sizeof(**pages));
221 /* ll_free_user_pages - tear down page struct array
222 * @pages: array of page struct pointers underlying target buffer */
223 static void ll_free_user_pages(struct page **pages, int npages, int do_dirty)
227 for (i = 0; i < npages; i++) {
228 if (pages[i] == NULL)
231 set_page_dirty_lock(pages[i]);
232 page_cache_release(pages[i]);
235 OBD_FREE_LARGE(pages, npages * sizeof(*pages));
238 ssize_t ll_direct_rw_pages(const struct lu_env *env, struct cl_io *io,
239 int rw, struct inode *inode,
240 struct ll_dio_pages *pv)
243 struct cl_2queue *queue;
244 struct cl_object *obj = io->ci_obj;
247 loff_t file_offset = pv->ldp_start_offset;
248 long size = pv->ldp_size;
249 int page_count = pv->ldp_nr;
250 struct page **pages = pv->ldp_pages;
251 long page_size = cl_page_size(obj);
256 queue = &io->ci_queue;
257 cl_2queue_init(queue);
258 for (i = 0; i < page_count; i++) {
260 file_offset = pv->ldp_offsets[i];
262 LASSERT(!(file_offset & (page_size - 1)));
263 clp = cl_page_find(env, obj, cl_index(obj, file_offset),
264 pv->ldp_pages[i], CPT_TRANSIENT);
270 rc = cl_page_own(env, io, clp);
272 LASSERT(clp->cp_state == CPS_FREEING);
273 cl_page_put(env, clp);
279 /* check the page type: if the page is a host page, then do
281 if (clp->cp_type == CPT_CACHEABLE) {
282 struct page *vmpage = cl_page_vmpage(clp);
283 struct page *src_page;
284 struct page *dst_page;
288 src_page = (rw == WRITE) ? pages[i] : vmpage;
289 dst_page = (rw == WRITE) ? vmpage : pages[i];
291 src = ll_kmap_atomic(src_page, KM_USER0);
292 dst = ll_kmap_atomic(dst_page, KM_USER1);
293 memcpy(dst, src, min(page_size, size));
294 ll_kunmap_atomic(dst, KM_USER1);
295 ll_kunmap_atomic(src, KM_USER0);
297 /* make sure page will be added to the transfer by
298 * cl_io_submit()->...->vvp_page_prep_write(). */
300 set_page_dirty(vmpage);
303 /* do not issue the page for read, since it
304 * may reread a ra page which has NOT uptodate
306 cl_page_disown(env, io, clp);
312 cl_2queue_add(queue, clp);
315 * Set page clip to tell transfer formation engine
316 * that page has to be sent even if it is beyond KMS.
318 cl_page_clip(env, clp, 0, min(size, page_size));
323 /* drop the reference count for cl_page_find */
324 cl_page_put(env, clp);
326 file_offset += page_size;
329 if (rc == 0 && io_pages) {
330 rc = cl_io_submit_sync(env, io,
331 rw == READ ? CRT_READ : CRT_WRITE,
337 cl_2queue_discard(env, io, queue);
338 cl_2queue_disown(env, io, queue);
339 cl_2queue_fini(env, queue);
342 EXPORT_SYMBOL(ll_direct_rw_pages);
344 static ssize_t ll_direct_IO_26_seg(const struct lu_env *env, struct cl_io *io,
345 int rw, struct inode *inode,
346 struct address_space *mapping,
347 size_t size, loff_t file_offset,
348 struct page **pages, int page_count)
350 struct ll_dio_pages pvec = { .ldp_pages = pages,
351 .ldp_nr = page_count,
354 .ldp_start_offset = file_offset
357 return ll_direct_rw_pages(env, io, rw, inode, &pvec);
360 #ifdef KMALLOC_MAX_SIZE
361 #define MAX_MALLOC KMALLOC_MAX_SIZE
363 #define MAX_MALLOC (128 * 1024)
366 /* This is the maximum size of a single O_DIRECT request, based on the
367 * kmalloc limit. We need to fit all of the brw_page structs, each one
368 * representing PAGE_SIZE worth of user data, into a single buffer, and
369 * then truncate this to be a full-sized RPC. For 4kB PAGE_SIZE this is
370 * up to 22MB for 128kB kmalloc and up to 682MB for 4MB kmalloc. */
371 #define MAX_DIO_SIZE ((MAX_MALLOC / sizeof(struct brw_page) * PAGE_CACHE_SIZE) & \
372 ~(DT_MAX_BRW_SIZE - 1))
373 static ssize_t ll_direct_IO_26(int rw, struct kiocb *iocb,
374 const struct iovec *iov, loff_t file_offset,
375 unsigned long nr_segs)
379 struct file *file = iocb->ki_filp;
380 struct inode *inode = file->f_mapping->host;
381 struct ccc_object *obj = cl_inode2ccc(inode);
382 long count = iov_length(iov, nr_segs);
383 long tot_bytes = 0, result = 0;
384 struct ll_inode_info *lli = ll_i2info(inode);
385 unsigned long seg = 0;
386 long size = MAX_DIO_SIZE;
390 if (!lli->lli_has_smd)
393 /* FIXME: io smaller than PAGE_SIZE is broken on ia64 ??? */
394 if ((file_offset & ~CFS_PAGE_MASK) || (count & ~CFS_PAGE_MASK))
397 CDEBUG(D_VFSTRACE, "VFS Op:inode="DFID"(%p), size=%lu (max %lu), "
398 "offset=%lld=%llx, pages %lu (max %lu)\n",
399 PFID(ll_inode2fid(inode)), inode, count, MAX_DIO_SIZE,
400 file_offset, file_offset, count >> PAGE_CACHE_SHIFT,
401 MAX_DIO_SIZE >> PAGE_CACHE_SHIFT);
403 /* Check that all user buffers are aligned as well */
404 for (seg = 0; seg < nr_segs; seg++) {
405 if (((unsigned long)iov[seg].iov_base & ~CFS_PAGE_MASK) ||
406 (iov[seg].iov_len & ~CFS_PAGE_MASK))
410 env = cl_env_get(&refcheck);
411 LASSERT(!IS_ERR(env));
412 io = ccc_env_io(env)->cui_cl.cis_io;
415 LASSERT(obj->cob_transient_pages == 0);
416 for (seg = 0; seg < nr_segs; seg++) {
417 long iov_left = iov[seg].iov_len;
418 unsigned long user_addr = (unsigned long)iov[seg].iov_base;
421 if (file_offset >= i_size_read(inode))
423 if (file_offset + iov_left > i_size_read(inode))
424 iov_left = i_size_read(inode) - file_offset;
427 while (iov_left > 0) {
429 int page_count, max_pages = 0;
432 bytes = min(size, iov_left);
433 page_count = ll_get_user_pages(rw, user_addr, bytes,
435 if (likely(page_count > 0)) {
436 if (unlikely(page_count < max_pages))
437 bytes = page_count << PAGE_CACHE_SHIFT;
438 result = ll_direct_IO_26_seg(env, io, rw, inode,
442 ll_free_user_pages(pages, max_pages, rw==READ);
443 } else if (page_count == 0) {
444 GOTO(out, result = -EFAULT);
448 if (unlikely(result <= 0)) {
449 /* If we can't allocate a large enough buffer
450 * for the request, shrink it to a smaller
451 * PAGE_SIZE multiple and try again.
452 * We should always be able to kmalloc for a
453 * page worth of page pointers = 4MB on i386. */
454 if (result == -ENOMEM &&
455 size > (PAGE_CACHE_SIZE / sizeof(*pages)) *
457 size = ((((size / 2) - 1) |
458 ~CFS_PAGE_MASK) + 1) &
460 CDEBUG(D_VFSTRACE,"DIO size now %lu\n",
469 file_offset += result;
475 LASSERT(obj->cob_transient_pages == 0);
478 struct ccc_io *cio = ccc_env_io(env);
480 /* no commit async for direct IO */
481 cio->u.write.cui_written += tot_bytes;
484 cl_env_put(env, &refcheck);
485 RETURN(tot_bytes ? tot_bytes : result);
489 * Prepare partially written-to page for a write.
491 static int ll_prepare_partial_page(const struct lu_env *env, struct cl_io *io,
494 struct cl_attr *attr = ccc_env_thread_attr(env);
495 struct cl_object *obj = io->ci_obj;
496 struct ccc_page *cp = cl_object_page_slice(obj, pg);
497 loff_t offset = cl_offset(obj, ccc_index(cp));
500 cl_object_attr_lock(obj);
501 result = cl_object_attr_get(env, obj, attr);
502 cl_object_attr_unlock(obj);
505 * If are writing to a new page, no need to read old data.
506 * The extent locking will have updated the KMS, and for our
507 * purposes here we can treat it like i_size.
509 if (attr->cat_kms <= offset) {
510 char *kaddr = ll_kmap_atomic(cp->cpg_page, KM_USER0);
512 memset(kaddr, 0, cl_page_size(obj));
513 ll_kunmap_atomic(kaddr, KM_USER0);
514 } else if (cp->cpg_defer_uptodate)
517 result = ll_page_sync_io(env, io, pg, CRT_READ);
522 static int ll_write_begin(struct file *file, struct address_space *mapping,
523 loff_t pos, unsigned len, unsigned flags,
524 struct page **pagep, void **fsdata)
526 struct ll_cl_context *lcc;
527 const struct lu_env *env;
529 struct cl_page *page;
531 struct cl_object *clob = ll_i2info(mapping->host)->lli_clob;
532 pgoff_t index = pos >> PAGE_CACHE_SHIFT;
533 struct page *vmpage = NULL;
534 unsigned from = pos & (PAGE_CACHE_SIZE - 1);
535 unsigned to = from + len;
539 CDEBUG(D_VFSTRACE, "Writing %lu of %d to %d bytes\n", index, from, len);
541 lcc = ll_cl_find(file);
543 GOTO(out, result = -EIO);
548 /* To avoid deadlock, try to lock page first. */
549 vmpage = grab_cache_page_nowait(mapping, index);
550 if (unlikely(vmpage == NULL ||
551 PageDirty(vmpage) || PageWriteback(vmpage))) {
552 struct ccc_io *cio = ccc_env_io(env);
553 struct cl_page_list *plist = &cio->u.write.cui_queue;
555 /* if the page is already in dirty cache, we have to commit
556 * the pages right now; otherwise, it may cause deadlock
557 * because it holds page lock of a dirty page and request for
558 * more grants. It's okay for the dirty page to be the first
559 * one in commit page list, though. */
560 if (vmpage != NULL && plist->pl_nr > 0) {
562 page_cache_release(vmpage);
566 /* commit pages and then wait for page lock */
567 result = vvp_io_write_commit(env, io);
571 if (vmpage == NULL) {
572 vmpage = grab_cache_page_write_begin(mapping, index,
575 GOTO(out, result = -ENOMEM);
579 page = cl_page_find(env, clob, vmpage->index, vmpage, CPT_CACHEABLE);
581 GOTO(out, result = PTR_ERR(page));
583 lcc->lcc_page = page;
584 lu_ref_add(&page->cp_reference, "cl_io", io);
586 cl_page_assume(env, io, page);
587 if (!PageUptodate(vmpage)) {
589 * We're completely overwriting an existing page,
590 * so _don't_ set it up to date until commit_write
592 if (from == 0 && to == PAGE_SIZE) {
593 CL_PAGE_HEADER(D_PAGE, env, page, "full page write\n");
594 POISON_PAGE(vmpage, 0x11);
596 /* TODO: can be optimized at OSC layer to check if it
597 * is a lockless IO. In that case, it's not necessary
598 * to read the data. */
599 result = ll_prepare_partial_page(env, io, page);
601 SetPageUptodate(vmpage);
605 cl_page_unassume(env, io, page);
609 if (vmpage != NULL) {
611 page_cache_release(vmpage);
620 static int ll_write_end(struct file *file, struct address_space *mapping,
621 loff_t pos, unsigned len, unsigned copied,
622 struct page *vmpage, void *fsdata)
624 struct ll_cl_context *lcc = fsdata;
625 const struct lu_env *env;
628 struct cl_page *page;
629 unsigned from = pos & (PAGE_CACHE_SIZE - 1);
634 page_cache_release(vmpage);
636 LASSERT(lcc != NULL);
638 page = lcc->lcc_page;
640 cio = ccc_env_io(env);
642 LASSERT(cl_page_is_owned(page, io));
644 struct cl_page_list *plist = &cio->u.write.cui_queue;
646 lcc->lcc_page = NULL; /* page will be queued */
648 /* Add it into write queue */
649 cl_page_list_add(plist, page);
650 if (plist->pl_nr == 1) /* first page */
651 cio->u.write.cui_from = from;
654 cio->u.write.cui_to = from + copied;
656 /* To address the deadlock in balance_dirty_pages() where
657 * this dirty page may be written back in the same thread. */
658 if (PageDirty(vmpage))
661 /* We may have one full RPC, commit it soon */
662 if (plist->pl_nr >= PTLRPC_MAX_BRW_PAGES)
665 CL_PAGE_DEBUG(D_VFSTRACE, env, page,
666 "queued page: %d.\n", plist->pl_nr);
668 cl_page_disown(env, io, page);
670 lcc->lcc_page = NULL;
671 lu_ref_del(&page->cp_reference, "cl_io", io);
672 cl_page_put(env, page);
674 /* page list is not contiguous now, commit it now */
678 file->f_flags & O_SYNC || IS_SYNC(file->f_dentry->d_inode))
679 result = vvp_io_write_commit(env, io);
681 RETURN(result >= 0 ? copied : result);
684 #ifdef CONFIG_MIGRATION
685 static int ll_migratepage(struct address_space *mapping,
686 struct page *newpage, struct page *page
687 #ifdef HAVE_MIGRATEPAGE_4ARGS
688 , enum migrate_mode mode
692 /* Always fail page migration until we have a proper implementation */
697 #ifndef MS_HAS_NEW_AOPS
698 const struct address_space_operations ll_aops = {
699 .readpage = ll_readpage,
700 .direct_IO = ll_direct_IO_26,
701 .writepage = ll_writepage,
702 .writepages = ll_writepages,
703 .set_page_dirty = __set_page_dirty_nobuffers,
704 .write_begin = ll_write_begin,
705 .write_end = ll_write_end,
706 .invalidatepage = ll_invalidatepage,
707 .releasepage = (void *)ll_releasepage,
708 #ifdef CONFIG_MIGRATION
709 .migratepage = ll_migratepage,
713 const struct address_space_operations_ext ll_aops = {
714 .orig_aops.readpage = ll_readpage,
715 .orig_aops.direct_IO = ll_direct_IO_26,
716 .orig_aops.writepage = ll_writepage,
717 .orig_aops.writepages = ll_writepages,
718 .orig_aops.set_page_dirty = __set_page_dirty_nobuffers,
719 .orig_aops.invalidatepage = ll_invalidatepage,
720 .orig_aops.releasepage = ll_releasepage,
721 #ifdef CONFIG_MIGRATION
722 .orig_aops.migratepage = ll_migratepage,
724 .write_begin = ll_write_begin,
725 .write_end = ll_write_end