1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
6 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
8 * This program is free software; you can redistribute it and/or modify
9 * it under the terms of the GNU General Public License version 2 only,
10 * as published by the Free Software Foundation.
12 * This program is distributed in the hope that it will be useful, but
13 * WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * General Public License version 2 for more details (a copy is included
16 * in the LICENSE file that accompanied this code).
18 * You should have received a copy of the GNU General Public License
19 * version 2 along with this program; If not, see
20 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
22 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
23 * CA 95054 USA or visit www.sun.com if you need additional information or
29 * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
30 * Use is subject to license terms.
32 * Copyright (c) 2011, 2012, Whamcloud, Inc.
35 * This file is part of Lustre, http://www.lustre.org/
36 * Lustre is a trademark of Sun Microsystems, Inc.
38 * lustre/lustre/llite/rw26.c
40 * Lustre Lite I/O page cache routines for the 2.5/2.6 kernel version
43 #ifndef AUTOCONF_INCLUDED
44 #include <linux/config.h>
46 #include <linux/kernel.h>
48 #include <linux/string.h>
49 #include <linux/stat.h>
50 #include <linux/errno.h>
51 #include <linux/smp_lock.h>
52 #include <linux/unistd.h>
53 #include <linux/version.h>
54 #include <asm/system.h>
55 #include <asm/uaccess.h>
58 #include <linux/buffer_head.h>
59 #include <linux/writeback.h>
60 #include <linux/stat.h>
61 #include <asm/uaccess.h>
63 #include <linux/pagemap.h>
64 #include <linux/smp_lock.h>
66 #define DEBUG_SUBSYSTEM S_LLITE
68 #include <lustre_lite.h>
69 #include "llite_internal.h"
70 #include <linux/lustre_compat25.h>
73 * Implements Linux VM address_space::invalidatepage() method. This method is
74 * called when the page is truncate from a file, either as a result of
75 * explicit truncate, or when inode is removed from memory (as a result of
76 * final iput(), umount, or memory pressure induced icache shrinking).
78 * [0, offset] bytes of the page remain valid (this is for a case of not-page
79 * aligned truncate). Lustre leaves partially truncated page in the cache,
80 * relying on struct inode::i_size to limit further accesses.
82 static int cl_invalidatepage(struct page *vmpage, unsigned long offset)
87 struct cl_object *obj;
92 LASSERT(PageLocked(vmpage));
93 LASSERT(!PageWriteback(vmpage));
96 * It is safe to not check anything in invalidatepage/releasepage
97 * below because they are run with page locked and all our io is
98 * happening with locked page too
102 env = cl_env_get(&refcheck);
104 inode = vmpage->mapping->host;
105 obj = ll_i2info(inode)->lli_clob;
107 page = cl_vmpage_page(vmpage, obj);
109 lu_ref_add(&page->cp_reference,
111 cl_page_delete(env, page);
113 lu_ref_del(&page->cp_reference,
115 cl_page_put(env, page);
118 LASSERT(vmpage->private == 0);
119 cl_env_put(env, &refcheck);
125 #ifdef HAVE_INVALIDATEPAGE_RETURN_INT
126 static int ll_invalidatepage(struct page *page, unsigned long offset)
128 return cl_invalidatepage(page, offset);
130 #else /* !HAVE_INVALIDATEPAGE_RETURN_INT */
131 static void ll_invalidatepage(struct page *page, unsigned long offset)
133 cl_invalidatepage(page, offset);
137 #ifdef HAVE_RELEASEPAGE_WITH_INT
138 #define RELEASEPAGE_ARG_TYPE int
140 #define RELEASEPAGE_ARG_TYPE gfp_t
142 static int ll_releasepage(struct page *page, RELEASEPAGE_ARG_TYPE gfp_mask)
146 cookie = cl_env_reenter();
147 ll_invalidatepage(page, 0);
148 cl_env_reexit(cookie);
152 static int ll_set_page_dirty(struct page *vmpage)
155 struct cl_page *page = vvp_vmpage_page_transient(vmpage);
156 struct vvp_object *obj = cl_inode2vvp(vmpage->mapping->host);
157 struct vvp_page *cpg;
160 * XXX should page method be called here?
162 LASSERT(&obj->co_cl == page->cp_obj);
163 cpg = cl2vvp_page(cl_page_at(page, &vvp_device_type));
165 * XXX cannot do much here, because page is possibly not locked:
167 * ->unmap_page_range()->zap_pte_range()->set_page_dirty().
169 vvp_write_pending(obj, cpg);
171 RETURN(__set_page_dirty_nobuffers(vmpage));
174 #define MAX_DIRECTIO_SIZE 2*1024*1024*1024UL
176 static inline int ll_get_user_pages(int rw, unsigned long user_addr,
177 size_t size, struct page ***pages,
180 int result = -ENOMEM;
182 /* set an arbitrary limit to prevent arithmetic overflow */
183 if (size > MAX_DIRECTIO_SIZE) {
188 *max_pages = (user_addr + size + CFS_PAGE_SIZE - 1) >> CFS_PAGE_SHIFT;
189 *max_pages -= user_addr >> CFS_PAGE_SHIFT;
191 OBD_ALLOC_LARGE(*pages, *max_pages * sizeof(**pages));
193 down_read(¤t->mm->mmap_sem);
194 result = get_user_pages(current, current->mm, user_addr,
195 *max_pages, (rw == READ), 0, *pages,
197 up_read(¤t->mm->mmap_sem);
198 if (unlikely(result <= 0))
199 OBD_FREE_LARGE(*pages, *max_pages * sizeof(**pages));
205 /* ll_free_user_pages - tear down page struct array
206 * @pages: array of page struct pointers underlying target buffer */
207 static void ll_free_user_pages(struct page **pages, int npages, int do_dirty)
211 for (i = 0; i < npages; i++) {
212 if (pages[i] == NULL)
215 set_page_dirty_lock(pages[i]);
216 page_cache_release(pages[i]);
219 OBD_FREE_LARGE(pages, npages * sizeof(*pages));
222 ssize_t ll_direct_rw_pages(const struct lu_env *env, struct cl_io *io,
223 int rw, struct inode *inode,
224 struct ll_dio_pages *pv)
227 struct cl_2queue *queue;
228 struct cl_object *obj = io->ci_obj;
231 loff_t file_offset = pv->ldp_start_offset;
232 long size = pv->ldp_size;
233 int page_count = pv->ldp_nr;
234 struct page **pages = pv->ldp_pages;
235 long page_size = cl_page_size(obj);
240 queue = &io->ci_queue;
241 cl_2queue_init(queue);
242 for (i = 0; i < page_count; i++) {
244 file_offset = pv->ldp_offsets[i];
246 LASSERT(!(file_offset & (page_size - 1)));
247 clp = cl_page_find(env, obj, cl_index(obj, file_offset),
248 pv->ldp_pages[i], CPT_TRANSIENT);
254 rc = cl_page_own(env, io, clp);
256 LASSERT(clp->cp_state == CPS_FREEING);
257 cl_page_put(env, clp);
263 /* check the page type: if the page is a host page, then do
265 if (clp->cp_type == CPT_CACHEABLE) {
266 cfs_page_t *vmpage = cl_page_vmpage(env, clp);
267 cfs_page_t *src_page;
268 cfs_page_t *dst_page;
272 src_page = (rw == WRITE) ? pages[i] : vmpage;
273 dst_page = (rw == WRITE) ? vmpage : pages[i];
275 src = kmap_atomic(src_page, KM_USER0);
276 dst = kmap_atomic(dst_page, KM_USER1);
277 memcpy(dst, src, min(page_size, size));
278 kunmap_atomic(dst, KM_USER1);
279 kunmap_atomic(src, KM_USER0);
281 /* make sure page will be added to the transfer by
282 * cl_io_submit()->...->vvp_page_prep_write(). */
284 set_page_dirty(vmpage);
287 /* do not issue the page for read, since it
288 * may reread a ra page which has NOT uptodate
290 cl_page_disown(env, io, clp);
296 cl_2queue_add(queue, clp);
299 * Set page clip to tell transfer formation engine
300 * that page has to be sent even if it is beyond KMS.
302 cl_page_clip(env, clp, 0, min(size, page_size));
307 /* drop the reference count for cl_page_find */
308 cl_page_put(env, clp);
310 file_offset += page_size;
313 if (rc == 0 && io_pages) {
314 rc = cl_io_submit_sync(env, io,
315 rw == READ ? CRT_READ : CRT_WRITE,
316 queue, CRP_NORMAL, 0);
321 cl_2queue_discard(env, io, queue);
322 cl_2queue_disown(env, io, queue);
323 cl_2queue_fini(env, queue);
326 EXPORT_SYMBOL(ll_direct_rw_pages);
328 static ssize_t ll_direct_IO_26_seg(const struct lu_env *env, struct cl_io *io,
329 int rw, struct inode *inode,
330 struct address_space *mapping,
331 size_t size, loff_t file_offset,
332 struct page **pages, int page_count)
334 struct ll_dio_pages pvec = { .ldp_pages = pages,
335 .ldp_nr = page_count,
338 .ldp_start_offset = file_offset
341 return ll_direct_rw_pages(env, io, rw, inode, &pvec);
344 #ifdef KMALLOC_MAX_SIZE
345 #define MAX_MALLOC KMALLOC_MAX_SIZE
347 #define MAX_MALLOC (128 * 1024)
350 /* This is the maximum size of a single O_DIRECT request, based on the
351 * kmalloc limit. We need to fit all of the brw_page structs, each one
352 * representing PAGE_SIZE worth of user data, into a single buffer, and
353 * then truncate this to be a full-sized RPC. For 4kB PAGE_SIZE this is
354 * up to 22MB for 128kB kmalloc and up to 682MB for 4MB kmalloc. */
355 #define MAX_DIO_SIZE ((MAX_MALLOC / sizeof(struct brw_page) * CFS_PAGE_SIZE) & \
356 ~(PTLRPC_MAX_BRW_SIZE - 1))
357 static ssize_t ll_direct_IO_26(int rw, struct kiocb *iocb,
358 const struct iovec *iov, loff_t file_offset,
359 unsigned long nr_segs)
363 struct file *file = iocb->ki_filp;
364 struct inode *inode = file->f_mapping->host;
365 struct ccc_object *obj = cl_inode2ccc(inode);
366 long count = iov_length(iov, nr_segs);
367 long tot_bytes = 0, result = 0;
368 struct ll_inode_info *lli = ll_i2info(inode);
369 struct lov_stripe_md *lsm = lli->lli_smd;
370 unsigned long seg = 0;
371 long size = MAX_DIO_SIZE;
375 if (!lli->lli_smd || !lli->lli_smd->lsm_object_id)
378 /* FIXME: io smaller than PAGE_SIZE is broken on ia64 ??? */
379 if ((file_offset & ~CFS_PAGE_MASK) || (count & ~CFS_PAGE_MASK))
382 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p), size=%lu (max %lu), "
383 "offset=%lld=%llx, pages %lu (max %lu)\n",
384 inode->i_ino, inode->i_generation, inode, count, MAX_DIO_SIZE,
385 file_offset, file_offset, count >> CFS_PAGE_SHIFT,
386 MAX_DIO_SIZE >> CFS_PAGE_SHIFT);
388 /* Check that all user buffers are aligned as well */
389 for (seg = 0; seg < nr_segs; seg++) {
390 if (((unsigned long)iov[seg].iov_base & ~CFS_PAGE_MASK) ||
391 (iov[seg].iov_len & ~CFS_PAGE_MASK))
395 env = cl_env_get(&refcheck);
396 LASSERT(!IS_ERR(env));
397 io = ccc_env_io(env)->cui_cl.cis_io;
400 /* 0. Need locking between buffered and direct access. and race with
401 * size changing by concurrent truncates and writes.
402 * 1. Need inode sem to operate transient pages. */
404 LOCK_INODE_MUTEX(inode);
406 LASSERT(obj->cob_transient_pages == 0);
407 for (seg = 0; seg < nr_segs; seg++) {
408 long iov_left = iov[seg].iov_len;
409 unsigned long user_addr = (unsigned long)iov[seg].iov_base;
412 if (file_offset >= i_size_read(inode))
414 if (file_offset + iov_left > i_size_read(inode))
415 iov_left = i_size_read(inode) - file_offset;
418 while (iov_left > 0) {
420 int page_count, max_pages = 0;
423 bytes = min(size, iov_left);
424 page_count = ll_get_user_pages(rw, user_addr, bytes,
426 if (likely(page_count > 0)) {
427 if (unlikely(page_count < max_pages))
428 bytes = page_count << CFS_PAGE_SHIFT;
429 result = ll_direct_IO_26_seg(env, io, rw, inode,
433 ll_free_user_pages(pages, max_pages, rw==READ);
434 } else if (page_count == 0) {
435 GOTO(out, result = -EFAULT);
439 if (unlikely(result <= 0)) {
440 /* If we can't allocate a large enough buffer
441 * for the request, shrink it to a smaller
442 * PAGE_SIZE multiple and try again.
443 * We should always be able to kmalloc for a
444 * page worth of page pointers = 4MB on i386. */
445 if (result == -ENOMEM &&
446 size > (CFS_PAGE_SIZE / sizeof(*pages)) *
448 size = ((((size / 2) - 1) |
449 ~CFS_PAGE_MASK) + 1) &
451 CDEBUG(D_VFSTRACE,"DIO size now %lu\n",
460 file_offset += result;
466 LASSERT(obj->cob_transient_pages == 0);
468 UNLOCK_INODE_MUTEX(inode);
472 lov_stripe_lock(lsm);
473 obd_adjust_kms(ll_i2dtexp(inode), lsm, file_offset, 0);
474 lov_stripe_unlock(lsm);
478 cl_env_put(env, &refcheck);
479 RETURN(tot_bytes ? : result);
482 #if defined(HAVE_KERNEL_WRITE_BEGIN_END) || defined(MS_HAS_NEW_AOPS)
483 static int ll_write_begin(struct file *file, struct address_space *mapping,
484 loff_t pos, unsigned len, unsigned flags,
485 struct page **pagep, void **fsdata)
487 pgoff_t index = pos >> PAGE_CACHE_SHIFT;
490 unsigned from = pos & (PAGE_CACHE_SIZE - 1);
493 page = grab_cache_page_write_begin(mapping, index, flags);
499 rc = ll_prepare_write(file, page, from, from + len);
502 page_cache_release(page);
507 static int ll_write_end(struct file *file, struct address_space *mapping,
508 loff_t pos, unsigned len, unsigned copied,
509 struct page *page, void *fsdata)
511 unsigned from = pos & (PAGE_CACHE_SIZE - 1);
514 rc = ll_commit_write(file, page, from, from + copied);
516 page_cache_release(page);
522 #ifdef CONFIG_MIGRATION
523 int ll_migratepage(struct address_space *mapping,
524 struct page *newpage, struct page *page)
526 /* Always fail page migration until we have a proper implementation */
531 #ifndef MS_HAS_NEW_AOPS
532 struct address_space_operations ll_aops = {
533 .readpage = ll_readpage,
534 // .readpages = ll_readpages,
535 .direct_IO = ll_direct_IO_26,
536 .writepage = ll_writepage,
537 .writepages = generic_writepages,
538 .set_page_dirty = ll_set_page_dirty,
540 #ifdef HAVE_KERNEL_WRITE_BEGIN_END
541 .write_begin = ll_write_begin,
542 .write_end = ll_write_end,
544 .prepare_write = ll_prepare_write,
545 .commit_write = ll_commit_write,
547 .invalidatepage = ll_invalidatepage,
548 .releasepage = (void *)ll_releasepage,
549 #ifdef CONFIG_MIGRATION
550 .migratepage = ll_migratepage,
555 struct address_space_operations_ext ll_aops = {
556 .orig_aops.readpage = ll_readpage,
557 // .orig_aops.readpages = ll_readpages,
558 .orig_aops.direct_IO = ll_direct_IO_26,
559 .orig_aops.writepage = ll_writepage,
560 .orig_aops.writepages = generic_writepages,
561 .orig_aops.set_page_dirty = ll_set_page_dirty,
562 .orig_aops.sync_page = NULL,
563 .orig_aops.prepare_write = ll_prepare_write,
564 .orig_aops.commit_write = ll_commit_write,
565 .orig_aops.invalidatepage = ll_invalidatepage,
566 .orig_aops.releasepage = ll_releasepage,
567 #ifdef CONFIG_MIGRATION
568 .orig_aops.migratepage = ll_migratepage,
570 .orig_aops.bmap = NULL,
571 .write_begin = ll_write_begin,
572 .write_end = ll_write_end