1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
4 * Lustre Lite I/O page cache routines for the 2.5/2.6 kernel version
6 * Copyright (c) 2001-2003 Cluster File Systems, Inc.
8 * This file is part of Lustre, http://www.lustre.org.
10 * Lustre is free software; you can redistribute it and/or
11 * modify it under the terms of version 2 of the GNU General Public
12 * License as published by the Free Software Foundation.
14 * Lustre is distributed in the hope that it will be useful,
15 * but WITHOUT ANY WARRANTY; without even the implied warranty of
16 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
17 * GNU General Public License for more details.
19 * You should have received a copy of the GNU General Public License
20 * along with Lustre; if not, write to the Free Software
21 * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
23 #ifdef HAVE_KERNEL_CONFIG_H
24 #include <linux/config.h>
26 #include <linux/kernel.h>
28 #include <linux/string.h>
29 #include <linux/stat.h>
30 #include <linux/errno.h>
31 #include <linux/smp_lock.h>
32 #include <linux/unistd.h>
33 #include <linux/version.h>
34 #include <asm/system.h>
35 #include <asm/uaccess.h>
38 #include <linux/buffer_head.h>
39 #include <linux/mpage.h>
40 #include <linux/writeback.h>
41 #include <linux/stat.h>
42 #include <asm/uaccess.h>
43 #include <asm/segment.h>
45 #include <linux/pagemap.h>
46 #include <linux/smp_lock.h>
48 #define DEBUG_SUBSYSTEM S_LLITE
50 #include <lustre_lite.h>
51 #include "llite_internal.h"
52 #include <linux/lustre_compat25.h>
54 static int ll_writepage_26(struct page *page, struct writeback_control *wbc)
56 return ll_writepage(page);
59 /* It is safe to not check anything in invalidatepage/releasepage below
60 because they are run with page locked and all our io is happening with
62 #ifdef HAVE_INVALIDATEPAGE_RETURN_INT
63 static int ll_invalidatepage(struct page *page, unsigned long offset)
67 if (PagePrivate(page))
72 static void ll_invalidatepage(struct page *page, unsigned long offset)
76 if (PagePrivate(page))
81 static int ll_releasepage(struct page *page, gfp_t gfp_mask)
83 if (PagePrivate(page))
88 #define MAX_DIRECTIO_SIZE 2*1024*1024*1024UL
90 static inline int ll_get_user_pages(int rw, unsigned long user_addr,
91 size_t size, struct page ***pages)
96 /* set an arbitrary limit to prevent arithmetic overflow */
97 if (size > MAX_DIRECTIO_SIZE) {
102 page_count = ((user_addr + size + CFS_PAGE_SIZE - 1) >> CFS_PAGE_SHIFT)-
103 (user_addr >> CFS_PAGE_SHIFT);
105 OBD_ALLOC_GFP(*pages, page_count * sizeof(**pages), GFP_KERNEL);
107 down_read(¤t->mm->mmap_sem);
108 result = get_user_pages(current, current->mm, user_addr,
109 page_count, (rw == READ), 0, *pages,
111 up_read(¤t->mm->mmap_sem);
113 OBD_FREE(*pages, page_count * sizeof(**pages));
119 /* ll_free_user_pages - tear down page struct array
120 * @pages: array of page struct pointers underlying target buffer */
121 static void ll_free_user_pages(struct page **pages, int npages, int do_dirty)
125 for (i = 0; i < npages; i++) {
127 set_page_dirty_lock(pages[i]);
128 page_cache_release(pages[i]);
131 OBD_FREE(pages, npages * sizeof(*pages));
134 static ssize_t ll_direct_IO_26_seg(int rw, struct inode *inode,
135 struct address_space *mapping,
136 struct lov_stripe_md *lsm,
137 size_t size, loff_t file_offset,
138 struct page **pages, int page_count)
140 struct brw_page *pga;
144 loff_t file_offset_orig = file_offset;
147 OBD_ALLOC(pga, sizeof(*pga) * page_count);
149 CDEBUG(D_VFSTRACE, "sizeof(*pga) = %u page_count = %u\n",
150 (int)sizeof(*pga), page_count);
154 for (i = 0, length = size; length > 0;
155 length -=pga[i].count, file_offset +=pga[i].count,i++) {/*i last!*/
156 pga[i].pg = pages[i];
157 pga[i].off = file_offset;
158 /* To the end of the page, or the length, whatever is less */
159 pga[i].count = min_t(int, CFS_PAGE_SIZE -(file_offset & ~CFS_PAGE_MASK),
163 POISON_PAGE(pages[i], 0x0d);
166 ll_inode_fill_obdo(inode, rw, &oa);
168 rc = obd_brw_rqset(rw == WRITE ? OBD_BRW_WRITE : OBD_BRW_READ,
169 ll_i2obdexp(inode), &oa, lsm, page_count, pga, NULL);
170 if ((rc > 0) && (rw == WRITE)) {
171 lov_stripe_lock(lsm);
172 obd_adjust_kms(ll_i2obdexp(inode), lsm, file_offset_orig + rc, 0);
173 lov_stripe_unlock(lsm);
176 OBD_FREE(pga, sizeof(*pga) * page_count);
180 /* This is the maximum size of a single O_DIRECT request, based on a 128kB
181 * kmalloc limit. We need to fit all of the brw_page structs, each one
182 * representing PAGE_SIZE worth of user data, into a single buffer, and
183 * then truncate this to be a full-sized RPC. This is 22MB for 4kB pages. */
184 #define MAX_DIO_SIZE ((128 * 1024 / sizeof(struct brw_page) * CFS_PAGE_SIZE) & \
185 ~(PTLRPC_MAX_BRW_SIZE - 1))
186 static ssize_t ll_direct_IO_26(int rw, struct kiocb *iocb,
187 const struct iovec *iov, loff_t file_offset,
188 unsigned long nr_segs)
190 struct file *file = iocb->ki_filp;
191 struct inode *inode = file->f_mapping->host;
192 ssize_t count = iov_length(iov, nr_segs), tot_bytes = 0;
193 struct ll_inode_info *lli = ll_i2info(inode);
195 size_t size = MAX_DIO_SIZE;
198 if (!lli->lli_smd || !lli->lli_smd->lsm_object_id)
201 /* FIXME: io smaller than CFS_PAGE_SIZE is broken on ia64 ??? */
202 if ((file_offset & (~CFS_PAGE_MASK)) || (count & ~CFS_PAGE_MASK))
205 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p), size="LPSZ" (max %lu), "
206 "offset=%lld=%llx, pages "LPSZ" (max %lu)\n",
207 inode->i_ino, inode->i_generation, inode, count, MAX_DIO_SIZE,
208 file_offset, file_offset, count >> CFS_PAGE_SHIFT,
209 MAX_DIO_SIZE >> CFS_PAGE_SHIFT);
212 ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_DIRECT_WRITE, count);
214 ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_DIRECT_READ, count);
216 /* Check that all user buffers are aligned as well */
217 for (seg = 0; seg < nr_segs; seg++) {
218 if (((unsigned long)iov[seg].iov_base & ~CFS_PAGE_MASK) ||
219 (iov[seg].iov_len & ~CFS_PAGE_MASK))
223 for (seg = 0; seg < nr_segs; seg++) {
224 size_t iov_left = iov[seg].iov_len;
225 unsigned long user_addr = (unsigned long)iov[seg].iov_base;
227 while (iov_left > 0) {
232 page_count = ll_get_user_pages(rw, user_addr,
235 LASSERT(page_count != 0);
236 if (page_count > 0) {
237 result = ll_direct_IO_26_seg(rw, inode,
243 ll_free_user_pages(pages, page_count, rw==READ);
247 if (page_count < 0 || result <= 0) {
248 /* If we can't allocate a large enough buffer
249 * for the request, shrink it to a smaller
250 * PAGE_SIZE multiple and try again.
251 * We should always be able to kmalloc for a
252 * page worth of page pointers = 4MB on i386. */
253 if ((page_count == -ENOMEM||result == -ENOMEM)&&
254 size > (CFS_PAGE_SIZE / sizeof(*pages)) *
256 size = ((((size / 2) - 1) |
257 ~CFS_PAGE_MASK) + 1) &
259 CDEBUG(D_VFSTRACE, "DIO size now %u\n",
265 RETURN(page_count < 0 ? page_count : result);
269 file_offset += result;
277 struct address_space_operations ll_aops = {
278 .readpage = ll_readpage,
279 // .readpages = ll_readpages,
280 .direct_IO = ll_direct_IO_26,
281 .writepage = ll_writepage_26,
282 .writepages = generic_writepages,
283 .set_page_dirty = __set_page_dirty_nobuffers,
285 .prepare_write = ll_prepare_write,
286 .commit_write = ll_commit_write,
287 .invalidatepage = ll_invalidatepage,
288 .releasepage = ll_releasepage,