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 int offset,
81 struct cl_object *obj;
85 LASSERT(PageLocked(vmpage));
86 LASSERT(!PageWriteback(vmpage));
89 * It is safe to not check anything in invalidatepage/releasepage
90 * below because they are run with page locked and all our io is
91 * happening with locked page too
93 if (offset == 0 && length == PAGE_CACHE_SIZE) {
94 env = cl_env_get(&refcheck);
96 inode = vmpage->mapping->host;
97 obj = ll_i2info(inode)->lli_clob;
99 page = cl_vmpage_page(vmpage, obj);
101 lu_ref_add(&page->cp_reference,
103 cl_page_delete(env, page);
104 lu_ref_del(&page->cp_reference,
106 cl_page_put(env, page);
109 LASSERT(vmpage->private == 0);
110 cl_env_put(env, &refcheck);
115 #ifdef HAVE_RELEASEPAGE_WITH_INT
116 #define RELEASEPAGE_ARG_TYPE int
118 #define RELEASEPAGE_ARG_TYPE gfp_t
120 static int ll_releasepage(struct page *vmpage, RELEASEPAGE_ARG_TYPE gfp_mask)
122 struct cl_env_nest nest;
124 struct cl_object *obj;
125 struct cl_page *page;
126 struct address_space *mapping;
129 LASSERT(PageLocked(vmpage));
130 if (PageWriteback(vmpage) || PageDirty(vmpage))
133 mapping = vmpage->mapping;
137 obj = ll_i2info(mapping->host)->lli_clob;
141 /* 1 for page allocator, 1 for cl_page and 1 for page cache */
142 if (page_count(vmpage) > 3)
145 /* TODO: determine what gfp should be used by @gfp_mask. */
146 env = cl_env_nested_get(&nest);
148 /* If we can't allocate an env we won't call cl_page_put()
149 * later on which further means it's impossible to drop
150 * page refcount by cl_page, so ask kernel to not free
154 page = cl_vmpage_page(vmpage, obj);
155 result = page == NULL;
157 if (!cl_page_in_use(page)) {
159 cl_page_delete(env, page);
161 cl_page_put(env, page);
163 cl_env_nested_put(&nest, env);
167 static int ll_set_page_dirty(struct page *vmpage)
170 struct cl_page *page = vvp_vmpage_page_transient(vmpage);
171 struct vvp_object *obj = cl_inode2vvp(vmpage->mapping->host);
172 struct vvp_page *cpg;
175 * XXX should page method be called here?
177 LASSERT(&obj->co_cl == page->cp_obj);
178 cpg = cl2vvp_page(cl_page_at(page, &vvp_device_type));
180 * XXX cannot do much here, because page is possibly not locked:
182 * ->unmap_page_range()->zap_pte_range()->set_page_dirty().
184 vvp_write_pending(obj, cpg);
186 RETURN(__set_page_dirty_nobuffers(vmpage));
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) >> PAGE_CACHE_SHIFT;
204 *max_pages -= user_addr >> PAGE_CACHE_SHIFT;
206 OBD_ALLOC_LARGE(*pages, *max_pages * sizeof(**pages));
208 down_read(¤t->mm->mmap_sem);
209 result = get_user_pages(current, current->mm, user_addr,
210 *max_pages, (rw == READ), 0, *pages,
212 up_read(¤t->mm->mmap_sem);
213 if (unlikely(result <= 0))
214 OBD_FREE_LARGE(*pages, *max_pages * sizeof(**pages));
220 /* ll_free_user_pages - tear down page struct array
221 * @pages: array of page struct pointers underlying target buffer */
222 static void ll_free_user_pages(struct page **pages, int npages, int do_dirty)
226 for (i = 0; i < npages; i++) {
227 if (pages[i] == NULL)
230 set_page_dirty_lock(pages[i]);
231 page_cache_release(pages[i]);
234 OBD_FREE_LARGE(pages, npages * sizeof(*pages));
237 ssize_t ll_direct_rw_pages(const struct lu_env *env, struct cl_io *io,
238 int rw, struct inode *inode,
239 struct ll_dio_pages *pv)
242 struct cl_2queue *queue;
243 struct cl_object *obj = io->ci_obj;
246 loff_t file_offset = pv->ldp_start_offset;
247 long size = pv->ldp_size;
248 int page_count = pv->ldp_nr;
249 struct page **pages = pv->ldp_pages;
250 long page_size = cl_page_size(obj);
255 queue = &io->ci_queue;
256 cl_2queue_init(queue);
257 for (i = 0; i < page_count; i++) {
259 file_offset = pv->ldp_offsets[i];
261 LASSERT(!(file_offset & (page_size - 1)));
262 clp = cl_page_find(env, obj, cl_index(obj, file_offset),
263 pv->ldp_pages[i], CPT_TRANSIENT);
269 rc = cl_page_own(env, io, clp);
271 LASSERT(clp->cp_state == CPS_FREEING);
272 cl_page_put(env, clp);
278 /* check the page type: if the page is a host page, then do
280 if (clp->cp_type == CPT_CACHEABLE) {
281 struct page *vmpage = cl_page_vmpage(env, clp);
282 struct page *src_page;
283 struct page *dst_page;
287 src_page = (rw == WRITE) ? pages[i] : vmpage;
288 dst_page = (rw == WRITE) ? vmpage : pages[i];
290 src = ll_kmap_atomic(src_page, KM_USER0);
291 dst = ll_kmap_atomic(dst_page, KM_USER1);
292 memcpy(dst, src, min(page_size, size));
293 ll_kunmap_atomic(dst, KM_USER1);
294 ll_kunmap_atomic(src, KM_USER0);
296 /* make sure page will be added to the transfer by
297 * cl_io_submit()->...->vvp_page_prep_write(). */
299 set_page_dirty(vmpage);
302 /* do not issue the page for read, since it
303 * may reread a ra page which has NOT uptodate
305 cl_page_disown(env, io, clp);
311 cl_2queue_add(queue, clp);
314 * Set page clip to tell transfer formation engine
315 * that page has to be sent even if it is beyond KMS.
317 cl_page_clip(env, clp, 0, min(size, page_size));
322 /* drop the reference count for cl_page_find */
323 cl_page_put(env, clp);
325 file_offset += page_size;
328 if (rc == 0 && io_pages) {
329 rc = cl_io_submit_sync(env, io,
330 rw == READ ? CRT_READ : CRT_WRITE,
336 cl_2queue_discard(env, io, queue);
337 cl_2queue_disown(env, io, queue);
338 cl_2queue_fini(env, queue);
341 EXPORT_SYMBOL(ll_direct_rw_pages);
343 static ssize_t ll_direct_IO_26_seg(const struct lu_env *env, struct cl_io *io,
344 int rw, struct inode *inode,
345 struct address_space *mapping,
346 size_t size, loff_t file_offset,
347 struct page **pages, int page_count)
349 struct ll_dio_pages pvec = { .ldp_pages = pages,
350 .ldp_nr = page_count,
353 .ldp_start_offset = file_offset
356 return ll_direct_rw_pages(env, io, rw, inode, &pvec);
359 #ifdef KMALLOC_MAX_SIZE
360 #define MAX_MALLOC KMALLOC_MAX_SIZE
362 #define MAX_MALLOC (128 * 1024)
365 /* This is the maximum size of a single O_DIRECT request, based on the
366 * kmalloc limit. We need to fit all of the brw_page structs, each one
367 * representing PAGE_SIZE worth of user data, into a single buffer, and
368 * then truncate this to be a full-sized RPC. For 4kB PAGE_SIZE this is
369 * up to 22MB for 128kB kmalloc and up to 682MB for 4MB kmalloc. */
370 #define MAX_DIO_SIZE ((MAX_MALLOC / sizeof(struct brw_page) * PAGE_CACHE_SIZE) & \
371 ~(DT_MAX_BRW_SIZE - 1))
372 static ssize_t ll_direct_IO_26(int rw, struct kiocb *iocb,
373 const struct iovec *iov, loff_t file_offset,
374 unsigned long nr_segs)
378 struct file *file = iocb->ki_filp;
379 struct inode *inode = file->f_mapping->host;
380 struct ccc_object *obj = cl_inode2ccc(inode);
381 long count = iov_length(iov, nr_segs);
382 long tot_bytes = 0, result = 0;
383 struct ll_inode_info *lli = ll_i2info(inode);
384 unsigned long seg = 0;
385 long size = MAX_DIO_SIZE;
389 if (!lli->lli_has_smd)
392 /* FIXME: io smaller than PAGE_SIZE is broken on ia64 ??? */
393 if ((file_offset & ~CFS_PAGE_MASK) || (count & ~CFS_PAGE_MASK))
396 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p), size=%lu (max %lu), "
397 "offset=%lld=%llx, pages %lu (max %lu)\n",
398 inode->i_ino, inode->i_generation, inode, count, MAX_DIO_SIZE,
399 file_offset, file_offset, count >> PAGE_CACHE_SHIFT,
400 MAX_DIO_SIZE >> PAGE_CACHE_SHIFT);
402 /* Check that all user buffers are aligned as well */
403 for (seg = 0; seg < nr_segs; seg++) {
404 if (((unsigned long)iov[seg].iov_base & ~CFS_PAGE_MASK) ||
405 (iov[seg].iov_len & ~CFS_PAGE_MASK))
409 env = cl_env_get(&refcheck);
410 LASSERT(!IS_ERR(env));
411 io = ccc_env_io(env)->cui_cl.cis_io;
414 /* 0. Need locking between buffered and direct access. and race with
415 * size changing by concurrent truncates and writes.
416 * 1. Need inode mutex to operate transient pages.
419 mutex_lock(&inode->i_mutex);
421 LASSERT(obj->cob_transient_pages == 0);
422 for (seg = 0; seg < nr_segs; seg++) {
423 long iov_left = iov[seg].iov_len;
424 unsigned long user_addr = (unsigned long)iov[seg].iov_base;
427 if (file_offset >= i_size_read(inode))
429 if (file_offset + iov_left > i_size_read(inode))
430 iov_left = i_size_read(inode) - file_offset;
433 while (iov_left > 0) {
435 int page_count, max_pages = 0;
438 bytes = min(size, iov_left);
439 page_count = ll_get_user_pages(rw, user_addr, bytes,
441 if (likely(page_count > 0)) {
442 if (unlikely(page_count < max_pages))
443 bytes = page_count << PAGE_CACHE_SHIFT;
444 result = ll_direct_IO_26_seg(env, io, rw, inode,
448 ll_free_user_pages(pages, max_pages, rw==READ);
449 } else if (page_count == 0) {
450 GOTO(out, result = -EFAULT);
454 if (unlikely(result <= 0)) {
455 /* If we can't allocate a large enough buffer
456 * for the request, shrink it to a smaller
457 * PAGE_SIZE multiple and try again.
458 * We should always be able to kmalloc for a
459 * page worth of page pointers = 4MB on i386. */
460 if (result == -ENOMEM &&
461 size > (PAGE_CACHE_SIZE / sizeof(*pages)) *
463 size = ((((size / 2) - 1) |
464 ~CFS_PAGE_MASK) + 1) &
466 CDEBUG(D_VFSTRACE,"DIO size now %lu\n",
475 file_offset += result;
481 LASSERT(obj->cob_transient_pages == 0);
483 mutex_unlock(&inode->i_mutex);
487 struct lov_stripe_md *lsm;
489 lsm = ccc_inode_lsm_get(inode);
490 LASSERT(lsm != NULL);
491 lov_stripe_lock(lsm);
492 obd_adjust_kms(ll_i2dtexp(inode), lsm, file_offset, 0);
493 lov_stripe_unlock(lsm);
494 ccc_inode_lsm_put(inode, lsm);
498 cl_env_put(env, &refcheck);
499 RETURN(tot_bytes ? : result);
502 static int ll_write_begin(struct file *file, struct address_space *mapping,
503 loff_t pos, unsigned len, unsigned flags,
504 struct page **pagep, void **fsdata)
506 pgoff_t index = pos >> PAGE_CACHE_SHIFT;
509 unsigned from = pos & (PAGE_CACHE_SIZE - 1);
512 page = grab_cache_page_write_begin(mapping, index, flags);
518 rc = ll_prepare_write(file, page, from, from + len);
521 page_cache_release(page);
526 static int ll_write_end(struct file *file, struct address_space *mapping,
527 loff_t pos, unsigned len, unsigned copied,
528 struct page *page, void *fsdata)
530 unsigned from = pos & (PAGE_CACHE_SIZE - 1);
533 rc = ll_commit_write(file, page, from, from + copied);
535 page_cache_release(page);
540 #ifdef CONFIG_MIGRATION
541 int ll_migratepage(struct address_space *mapping,
542 struct page *newpage, struct page *page
543 , enum migrate_mode mode
546 /* Always fail page migration until we have a proper implementation */
551 #ifndef MS_HAS_NEW_AOPS
552 struct address_space_operations ll_aops = {
553 .readpage = ll_readpage,
554 // .readpages = ll_readpages,
555 .direct_IO = ll_direct_IO_26,
556 .writepage = ll_writepage,
557 .writepages = ll_writepages,
558 .set_page_dirty = ll_set_page_dirty,
559 .write_begin = ll_write_begin,
560 .write_end = ll_write_end,
561 .invalidatepage = ll_invalidatepage,
562 .releasepage = (void *)ll_releasepage,
563 #ifdef CONFIG_MIGRATION
564 .migratepage = ll_migratepage,
569 struct address_space_operations_ext ll_aops = {
570 .orig_aops.readpage = ll_readpage,
571 // .orig_aops.readpages = ll_readpages,
572 .orig_aops.direct_IO = ll_direct_IO_26,
573 .orig_aops.writepage = ll_writepage,
574 .orig_aops.writepages = ll_writepages,
575 .orig_aops.set_page_dirty = ll_set_page_dirty,
576 .orig_aops.prepare_write = ll_prepare_write,
577 .orig_aops.commit_write = ll_commit_write,
578 .orig_aops.invalidatepage = ll_invalidatepage,
579 .orig_aops.releasepage = ll_releasepage,
580 #ifdef CONFIG_MIGRATION
581 .orig_aops.migratepage = ll_migratepage,
583 .orig_aops.bmap = NULL,
584 .write_begin = ll_write_begin,
585 .write_end = ll_write_end