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>
49 #include <linux/migrate.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 <lustre_lite.h>
62 #include "llite_internal.h"
63 #include <linux/lustre_compat25.h>
66 * Implements Linux VM address_space::invalidatepage() method. This method is
67 * called when the page is truncate from a file, either as a result of
68 * explicit truncate, or when inode is removed from memory (as a result of
69 * final iput(), umount, or memory pressure induced icache shrinking).
71 * [0, offset] bytes of the page remain valid (this is for a case of not-page
72 * aligned truncate). Lustre leaves partially truncated page in the cache,
73 * relying on struct inode::i_size to limit further accesses.
75 static void ll_invalidatepage(struct page *vmpage, unsigned long offset)
80 struct cl_object *obj;
84 LASSERT(PageLocked(vmpage));
85 LASSERT(!PageWriteback(vmpage));
88 * It is safe to not check anything in invalidatepage/releasepage
89 * below because they are run with page locked and all our io is
90 * happening with locked page too
93 env = cl_env_get(&refcheck);
95 inode = vmpage->mapping->host;
96 obj = ll_i2info(inode)->lli_clob;
98 page = cl_vmpage_page(vmpage, obj);
100 lu_ref_add(&page->cp_reference,
102 cl_page_delete(env, page);
103 lu_ref_del(&page->cp_reference,
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)
121 struct cl_env_nest nest;
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 page allocator, 1 for cl_page and 1 for page cache */
141 if (page_count(vmpage) > 3)
144 /* TODO: determine what gfp should be used by @gfp_mask. */
145 env = cl_env_nested_get(&nest);
147 /* If we can't allocate an env we won't call cl_page_put()
148 * later on which further means it's impossible to drop
149 * page refcount by cl_page, so ask kernel to not free
153 page = cl_vmpage_page(vmpage, obj);
154 result = page == NULL;
156 if (!cl_page_in_use(page)) {
158 cl_page_delete(env, page);
160 cl_page_put(env, page);
162 cl_env_nested_put(&nest, env);
166 static int ll_set_page_dirty(struct page *vmpage)
169 struct cl_page *page = vvp_vmpage_page_transient(vmpage);
170 struct vvp_object *obj = cl_inode2vvp(vmpage->mapping->host);
171 struct vvp_page *cpg;
174 * XXX should page method be called here?
176 LASSERT(&obj->co_cl == page->cp_obj);
177 cpg = cl2vvp_page(cl_page_at(page, &vvp_device_type));
179 * XXX cannot do much here, because page is possibly not locked:
181 * ->unmap_page_range()->zap_pte_range()->set_page_dirty().
183 vvp_write_pending(obj, cpg);
185 RETURN(__set_page_dirty_nobuffers(vmpage));
188 #define MAX_DIRECTIO_SIZE 2*1024*1024*1024UL
190 static inline int ll_get_user_pages(int rw, unsigned long user_addr,
191 size_t size, struct page ***pages,
194 int result = -ENOMEM;
196 /* set an arbitrary limit to prevent arithmetic overflow */
197 if (size > MAX_DIRECTIO_SIZE) {
202 *max_pages = (user_addr + size + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
203 *max_pages -= user_addr >> PAGE_CACHE_SHIFT;
205 OBD_ALLOC_LARGE(*pages, *max_pages * sizeof(**pages));
207 down_read(¤t->mm->mmap_sem);
208 result = get_user_pages(current, current->mm, user_addr,
209 *max_pages, (rw == READ), 0, *pages,
211 up_read(¤t->mm->mmap_sem);
212 if (unlikely(result <= 0))
213 OBD_FREE_LARGE(*pages, *max_pages * sizeof(**pages));
219 /* ll_free_user_pages - tear down page struct array
220 * @pages: array of page struct pointers underlying target buffer */
221 static void ll_free_user_pages(struct page **pages, int npages, int do_dirty)
225 for (i = 0; i < npages; i++) {
226 if (pages[i] == NULL)
229 set_page_dirty_lock(pages[i]);
230 page_cache_release(pages[i]);
233 OBD_FREE_LARGE(pages, npages * sizeof(*pages));
236 ssize_t ll_direct_rw_pages(const struct lu_env *env, struct cl_io *io,
237 int rw, struct inode *inode,
238 struct ll_dio_pages *pv)
241 struct cl_2queue *queue;
242 struct cl_object *obj = io->ci_obj;
245 loff_t file_offset = pv->ldp_start_offset;
246 long size = pv->ldp_size;
247 int page_count = pv->ldp_nr;
248 struct page **pages = pv->ldp_pages;
249 long page_size = cl_page_size(obj);
254 queue = &io->ci_queue;
255 cl_2queue_init(queue);
256 for (i = 0; i < page_count; i++) {
258 file_offset = pv->ldp_offsets[i];
260 LASSERT(!(file_offset & (page_size - 1)));
261 clp = cl_page_find(env, obj, cl_index(obj, file_offset),
262 pv->ldp_pages[i], CPT_TRANSIENT);
268 rc = cl_page_own(env, io, clp);
270 LASSERT(clp->cp_state == CPS_FREEING);
271 cl_page_put(env, clp);
277 /* check the page type: if the page is a host page, then do
279 if (clp->cp_type == CPT_CACHEABLE) {
280 struct page *vmpage = cl_page_vmpage(env, clp);
281 struct page *src_page;
282 struct page *dst_page;
286 src_page = (rw == WRITE) ? pages[i] : vmpage;
287 dst_page = (rw == WRITE) ? vmpage : pages[i];
289 src = ll_kmap_atomic(src_page, KM_USER0);
290 dst = ll_kmap_atomic(dst_page, KM_USER1);
291 memcpy(dst, src, min(page_size, size));
292 ll_kunmap_atomic(dst, KM_USER1);
293 ll_kunmap_atomic(src, KM_USER0);
295 /* make sure page will be added to the transfer by
296 * cl_io_submit()->...->vvp_page_prep_write(). */
298 set_page_dirty(vmpage);
301 /* do not issue the page for read, since it
302 * may reread a ra page which has NOT uptodate
304 cl_page_disown(env, io, clp);
310 cl_2queue_add(queue, clp);
313 * Set page clip to tell transfer formation engine
314 * that page has to be sent even if it is beyond KMS.
316 cl_page_clip(env, clp, 0, min(size, page_size));
321 /* drop the reference count for cl_page_find */
322 cl_page_put(env, clp);
324 file_offset += page_size;
327 if (rc == 0 && io_pages) {
328 rc = cl_io_submit_sync(env, io,
329 rw == READ ? CRT_READ : CRT_WRITE,
335 cl_2queue_discard(env, io, queue);
336 cl_2queue_disown(env, io, queue);
337 cl_2queue_fini(env, queue);
340 EXPORT_SYMBOL(ll_direct_rw_pages);
342 static ssize_t ll_direct_IO_26_seg(const struct lu_env *env, struct cl_io *io,
343 int rw, struct inode *inode,
344 struct address_space *mapping,
345 size_t size, loff_t file_offset,
346 struct page **pages, int page_count)
348 struct ll_dio_pages pvec = { .ldp_pages = pages,
349 .ldp_nr = page_count,
352 .ldp_start_offset = file_offset
355 return ll_direct_rw_pages(env, io, rw, inode, &pvec);
358 #ifdef KMALLOC_MAX_SIZE
359 #define MAX_MALLOC KMALLOC_MAX_SIZE
361 #define MAX_MALLOC (128 * 1024)
364 /* This is the maximum size of a single O_DIRECT request, based on the
365 * kmalloc limit. We need to fit all of the brw_page structs, each one
366 * representing PAGE_SIZE worth of user data, into a single buffer, and
367 * then truncate this to be a full-sized RPC. For 4kB PAGE_SIZE this is
368 * up to 22MB for 128kB kmalloc and up to 682MB for 4MB kmalloc. */
369 #define MAX_DIO_SIZE ((MAX_MALLOC / sizeof(struct brw_page) * PAGE_CACHE_SIZE) & \
370 ~(DT_MAX_BRW_SIZE - 1))
371 static ssize_t ll_direct_IO_26(int rw, struct kiocb *iocb,
372 const struct iovec *iov, loff_t file_offset,
373 unsigned long nr_segs)
377 struct file *file = iocb->ki_filp;
378 struct inode *inode = file->f_mapping->host;
379 struct ccc_object *obj = cl_inode2ccc(inode);
380 long count = iov_length(iov, nr_segs);
381 long tot_bytes = 0, result = 0;
382 struct ll_inode_info *lli = ll_i2info(inode);
383 unsigned long seg = 0;
384 long size = MAX_DIO_SIZE;
388 if (!lli->lli_has_smd)
391 /* FIXME: io smaller than PAGE_SIZE is broken on ia64 ??? */
392 if ((file_offset & ~CFS_PAGE_MASK) || (count & ~CFS_PAGE_MASK))
395 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p), size=%lu (max %lu), "
396 "offset=%lld=%llx, pages %lu (max %lu)\n",
397 inode->i_ino, inode->i_generation, inode, count, MAX_DIO_SIZE,
398 file_offset, file_offset, count >> PAGE_CACHE_SHIFT,
399 MAX_DIO_SIZE >> PAGE_CACHE_SHIFT);
401 /* Check that all user buffers are aligned as well */
402 for (seg = 0; seg < nr_segs; seg++) {
403 if (((unsigned long)iov[seg].iov_base & ~CFS_PAGE_MASK) ||
404 (iov[seg].iov_len & ~CFS_PAGE_MASK))
408 env = cl_env_get(&refcheck);
409 LASSERT(!IS_ERR(env));
410 io = ccc_env_io(env)->cui_cl.cis_io;
413 /* 0. Need locking between buffered and direct access. and race with
414 * size changing by concurrent truncates and writes.
415 * 1. Need inode mutex to operate transient pages.
418 mutex_lock(&inode->i_mutex);
420 LASSERT(obj->cob_transient_pages == 0);
421 for (seg = 0; seg < nr_segs; seg++) {
422 long iov_left = iov[seg].iov_len;
423 unsigned long user_addr = (unsigned long)iov[seg].iov_base;
426 if (file_offset >= i_size_read(inode))
428 if (file_offset + iov_left > i_size_read(inode))
429 iov_left = i_size_read(inode) - file_offset;
432 while (iov_left > 0) {
434 int page_count, max_pages = 0;
437 bytes = min(size, iov_left);
438 page_count = ll_get_user_pages(rw, user_addr, bytes,
440 if (likely(page_count > 0)) {
441 if (unlikely(page_count < max_pages))
442 bytes = page_count << PAGE_CACHE_SHIFT;
443 result = ll_direct_IO_26_seg(env, io, rw, inode,
447 ll_free_user_pages(pages, max_pages, rw==READ);
448 } else if (page_count == 0) {
449 GOTO(out, result = -EFAULT);
453 if (unlikely(result <= 0)) {
454 /* If we can't allocate a large enough buffer
455 * for the request, shrink it to a smaller
456 * PAGE_SIZE multiple and try again.
457 * We should always be able to kmalloc for a
458 * page worth of page pointers = 4MB on i386. */
459 if (result == -ENOMEM &&
460 size > (PAGE_CACHE_SIZE / sizeof(*pages)) *
462 size = ((((size / 2) - 1) |
463 ~CFS_PAGE_MASK) + 1) &
465 CDEBUG(D_VFSTRACE,"DIO size now %lu\n",
474 file_offset += result;
480 LASSERT(obj->cob_transient_pages == 0);
482 mutex_unlock(&inode->i_mutex);
486 struct lov_stripe_md *lsm;
488 lsm = ccc_inode_lsm_get(inode);
489 LASSERT(lsm != NULL);
490 lov_stripe_lock(lsm);
491 obd_adjust_kms(ll_i2dtexp(inode), lsm, file_offset, 0);
492 lov_stripe_unlock(lsm);
493 ccc_inode_lsm_put(inode, lsm);
497 cl_env_put(env, &refcheck);
498 RETURN(tot_bytes ? : result);
501 static int ll_write_begin(struct file *file, struct address_space *mapping,
502 loff_t pos, unsigned len, unsigned flags,
503 struct page **pagep, void **fsdata)
505 pgoff_t index = pos >> PAGE_CACHE_SHIFT;
508 unsigned from = pos & (PAGE_CACHE_SIZE - 1);
511 page = grab_cache_page_write_begin(mapping, index, flags);
517 rc = ll_prepare_write(file, page, from, from + len);
520 page_cache_release(page);
525 static int ll_write_end(struct file *file, struct address_space *mapping,
526 loff_t pos, unsigned len, unsigned copied,
527 struct page *page, void *fsdata)
529 unsigned from = pos & (PAGE_CACHE_SIZE - 1);
532 rc = ll_commit_write(file, page, from, from + copied);
534 page_cache_release(page);
539 #ifdef CONFIG_MIGRATION
540 int ll_migratepage(struct address_space *mapping,
541 struct page *newpage, struct page *page
542 , enum migrate_mode mode
545 /* Always fail page migration until we have a proper implementation */
550 #ifndef MS_HAS_NEW_AOPS
551 struct address_space_operations ll_aops = {
552 .readpage = ll_readpage,
553 // .readpages = ll_readpages,
554 .direct_IO = ll_direct_IO_26,
555 .writepage = ll_writepage,
556 .writepages = ll_writepages,
557 .set_page_dirty = ll_set_page_dirty,
558 .write_begin = ll_write_begin,
559 .write_end = ll_write_end,
560 .invalidatepage = ll_invalidatepage,
561 .releasepage = (void *)ll_releasepage,
562 #ifdef CONFIG_MIGRATION
563 .migratepage = ll_migratepage,
568 struct address_space_operations_ext ll_aops = {
569 .orig_aops.readpage = ll_readpage,
570 // .orig_aops.readpages = ll_readpages,
571 .orig_aops.direct_IO = ll_direct_IO_26,
572 .orig_aops.writepage = ll_writepage,
573 .orig_aops.writepages = ll_writepages,
574 .orig_aops.set_page_dirty = ll_set_page_dirty,
575 .orig_aops.prepare_write = ll_prepare_write,
576 .orig_aops.commit_write = ll_commit_write,
577 .orig_aops.invalidatepage = ll_invalidatepage,
578 .orig_aops.releasepage = ll_releasepage,
579 #ifdef CONFIG_MIGRATION
580 .orig_aops.migratepage = ll_migratepage,
582 .orig_aops.bmap = NULL,
583 .write_begin = ll_write_begin,
584 .write_end = ll_write_end