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 2008 Sun Microsystems, Inc. All rights reserved
30 * Use is subject to license terms.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
38 * Lustre Lite I/O page cache routines for the 2.5/2.6 kernel version
41 #ifndef AUTOCONF_INCLUDED
42 #include <linux/config.h>
44 #include <linux/kernel.h>
46 #include <linux/string.h>
47 #include <linux/stat.h>
48 #include <linux/errno.h>
49 #include <linux/smp_lock.h>
50 #include <linux/unistd.h>
51 #include <linux/version.h>
52 #include <asm/system.h>
53 #include <asm/uaccess.h>
56 #include <linux/buffer_head.h>
57 #include <linux/writeback.h>
58 #include <linux/stat.h>
59 #include <asm/uaccess.h>
61 # include <asm/segment.h>
64 #include <linux/pagemap.h>
65 #include <linux/smp_lock.h>
67 #define DEBUG_SUBSYSTEM S_LLITE
69 #include <lustre_lite.h>
70 #include "llite_internal.h"
71 #include <linux/lustre_compat25.h>
73 static int ll_writepage_26(struct page *page, struct writeback_control *wbc)
75 return ll_writepage(page);
78 /* It is safe to not check anything in invalidatepage/releasepage below
79 because they are run with page locked and all our io is happening with
81 #ifdef HAVE_INVALIDATEPAGE_RETURN_INT
82 static int ll_invalidatepage(struct page *page, unsigned long offset)
86 if (PagePrivate(page))
91 static void ll_invalidatepage(struct page *page, unsigned long offset)
95 if (PagePrivate(page))
100 #ifdef HAVE_RELEASEPAGE_WITH_GFP
101 #define RELEASEPAGE_ARG_TYPE gfp_t
103 #define RELEASEPAGE_ARG_TYPE int
105 static int ll_releasepage(struct page *page, RELEASEPAGE_ARG_TYPE gfp_mask)
107 if (PagePrivate(page))
112 #define MAX_DIRECTIO_SIZE 2*1024*1024*1024UL
114 static inline int ll_get_user_pages(int rw, unsigned long user_addr,
115 size_t size, struct page ***pages,
118 int result = -ENOMEM;
120 /* set an arbitrary limit to prevent arithmetic overflow */
121 if (size > MAX_DIRECTIO_SIZE) {
126 *max_pages = ((user_addr + size + CFS_PAGE_SIZE - 1) >> CFS_PAGE_SHIFT)-
127 (user_addr >> CFS_PAGE_SHIFT);
129 OBD_ALLOC_WAIT(*pages, *max_pages * sizeof(**pages));
131 down_read(¤t->mm->mmap_sem);
132 result = get_user_pages(current, current->mm, user_addr,
133 *max_pages, (rw == READ), 0, *pages,
135 up_read(¤t->mm->mmap_sem);
136 if (unlikely(result <= 0))
137 OBD_FREE(*pages, *max_pages * sizeof(**pages));
143 /* ll_free_user_pages - tear down page struct array
144 * @pages: array of page struct pointers underlying target buffer */
145 static void ll_free_user_pages(struct page **pages, int npages, int do_dirty)
149 for (i = 0; i < npages; i++) {
150 if (pages[i] == NULL)
153 set_page_dirty_lock(pages[i]);
154 page_cache_release(pages[i]);
156 OBD_FREE(pages, npages * sizeof(*pages));
159 static ssize_t ll_direct_IO_26_seg(int rw, struct inode *inode,
160 struct address_space *mapping,
161 struct obd_info *oinfo,
162 struct ptlrpc_request_set *set,
163 size_t size, loff_t file_offset,
164 struct page **pages, int page_count,
165 unsigned long user_addr, int locked)
167 struct brw_page *pga;
168 int i, rc = 0, pshift;
172 OBD_ALLOC(pga, sizeof(*pga) * page_count);
174 CDEBUG(D_VFSTRACE, "sizeof(*pga) = %u page_count = %u\n",
175 (int)sizeof(*pga), page_count);
180 * pshift is something we'll add to ->off to get the in-memory offset,
181 * also see the OSC_FILE2MEM_OFF macro
183 pshift = (user_addr & ~CFS_PAGE_MASK) - (file_offset & ~CFS_PAGE_MASK);
185 for (i = 0, length = size; length > 0; i++) {/*i last!*/
186 LASSERT(i < page_count);
188 pga[i].pg = pages[i];
189 pga[i].off = file_offset;
190 /* To the end of the page, or the length, whatever is less */
191 pga[i].count = min_t(int, CFS_PAGE_SIZE -(user_addr & ~CFS_PAGE_MASK),
194 pga[i].flag = OBD_BRW_SYNC;
196 pga[i].flag |= OBD_BRW_SRVLOCK;
199 POISON_PAGE(pages[i], 0x0d);
201 length -= pga[i].count;
202 file_offset += pga[i].count;
203 user_addr += pga[i].count;
206 rc = obd_brw_async(rw == WRITE ? OBD_BRW_WRITE : OBD_BRW_READ,
207 ll_i2obdexp(inode), oinfo, page_count,
208 pga, NULL, set, pshift);
212 OBD_FREE(pga, sizeof(*pga) * page_count);
216 /* This is the maximum size of a single O_DIRECT request, based on a 128kB
217 * kmalloc limit. We need to fit all of the brw_page structs, each one
218 * representing PAGE_SIZE worth of user data, into a single buffer, and
219 * then truncate this to be a full-sized RPC. This is 22MB for 4kB pages. */
220 #define MAX_DIO_SIZE ((128 * 1024 / sizeof(struct brw_page) * CFS_PAGE_SIZE) & \
221 ~(PTLRPC_MAX_BRW_SIZE - 1))
223 ssize_t ll_direct_IO(int rw, struct file *file,
224 const struct iovec *iov, loff_t file_offset,
225 unsigned long nr_segs, int locked)
227 struct inode *inode = file->f_mapping->host;
228 ssize_t count = iov_length(iov, nr_segs);
229 ssize_t tot_bytes = 0, result = 0;
230 struct ll_sb_info *sbi = ll_i2sbi(inode);
231 struct ll_inode_info *lli = ll_i2info(inode);
232 struct lov_stripe_md *lsm = lli->lli_smd;
233 struct ptlrpc_request_set *set;
234 struct obd_info oinfo;
235 struct obdo oa = { 0 };
237 size_t size = MAX_DIO_SIZE;
240 if (!lli->lli_smd || !lli->lli_smd->lsm_object_id)
243 CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p), size="LPSZ" (max %lu), "
244 "offset=%lld=%llx, pages "LPSZ" (max %lu)\n",
245 inode->i_ino, inode->i_generation, inode, count, MAX_DIO_SIZE,
246 file_offset, file_offset, count >> CFS_PAGE_SHIFT,
247 MAX_DIO_SIZE >> CFS_PAGE_SHIFT);
250 ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_DIRECT_WRITE, count);
252 ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_DIRECT_READ, count);
254 set = ptlrpc_prep_set();
258 ll_inode_fill_obdo(inode, rw == WRITE ? OBD_BRW_WRITE : OBD_BRW_READ, &oa);
262 /* need locking between buffered and direct access. and race with
263 *size changing by concurrent truncates and writes. */
265 LOCK_INODE_MUTEX(inode);
266 for (seg = 0; seg < nr_segs; seg++) {
267 size_t iov_left = iov[seg].iov_len;
268 unsigned long user_addr = (unsigned long)iov[seg].iov_base;
271 if (file_offset >= inode->i_size)
273 if (file_offset + iov_left > inode->i_size)
274 iov_left = inode->i_size - file_offset;
277 while (iov_left > 0) {
279 int page_count, max_pages = 0;
282 bytes = min(size,iov_left);
284 /* a dirty hack for non-aligned I/O: avoid filling pgas,
285 * which cross stripe boundaries (20777) */
286 if (user_addr & ~CFS_PAGE_MASK ||
287 file_offset & ~CFS_PAGE_MASK) {
288 obd_off end = file_offset;
290 obd_extent_calc(sbi->ll_osc_exp, lsm,
291 OBD_CALC_STRIPE_END, &end);
293 if (file_offset + bytes > end + 1)
294 bytes = end - file_offset + 1;
297 page_count = ll_get_user_pages(rw, user_addr,
300 if (likely(page_count > 0)) {
301 if (unlikely(page_count < max_pages))
302 bytes = page_count << CFS_PAGE_SHIFT;
303 result = ll_direct_IO_26_seg(rw, inode,
310 ll_free_user_pages(pages, max_pages, rw==READ);
311 } else if (page_count == 0) {
312 GOTO(out, result = -EFAULT);
316 if (unlikely(result <= 0)) {
317 /* If we can't allocate a large enough buffer
318 * for the request, shrink it to a smaller
319 * PAGE_SIZE multiple and try again.
320 * We should always be able to kmalloc for a
321 * page worth of page pointers = 4MB on i386. */
322 if (result == -ENOMEM &&
323 size > (CFS_PAGE_SIZE / sizeof(*pages)) *
325 size = ((((size / 2) - 1) |
326 ~CFS_PAGE_MASK) + 1) &
328 CDEBUG(D_VFSTRACE, "DIO size now %u\n",
335 file_offset += result;
341 if (likely(tot_bytes > 0)) {
344 rc = ptlrpc_set_wait(set);
345 if (unlikely(rc != 0))
346 GOTO(unlock_mutex, tot_bytes = rc);
347 if (rw == WRITE && locked) {
348 lov_stripe_lock(lsm);
349 obd_adjust_kms(ll_i2obdexp(inode),
350 lsm, file_offset, 0);
351 lov_stripe_unlock(lsm);
358 UNLOCK_INODE_MUTEX(inode);
360 ptlrpc_set_destroy(set);
364 static ssize_t ll_direct_IO_26(int rw, struct kiocb *kiocb,
365 const struct iovec *iov, loff_t file_offset,
366 unsigned long nr_segs)
368 return ll_direct_IO(rw, kiocb->ki_filp, iov, file_offset, nr_segs, 1);
371 #ifdef HAVE_KERNEL_WRITE_BEGIN_END
372 static int ll_write_begin(struct file *file, struct address_space *mapping,
373 loff_t pos, unsigned len, unsigned flags,
374 struct page **pagep, void **fsdata)
376 pgoff_t index = pos >> PAGE_CACHE_SHIFT;
379 unsigned from = pos & (PAGE_CACHE_SIZE - 1);
382 page = grab_cache_page_write_begin(mapping, index, flags);
388 rc = ll_prepare_write(file, page, from, from + len);
391 page_cache_release(page);
396 static int ll_write_end(struct file *file, struct address_space *mapping,
397 loff_t pos, unsigned len, unsigned copied,
398 struct page *page, void *fsdata)
400 unsigned from = pos & (PAGE_CACHE_SIZE - 1);
402 rc = ll_commit_write(file, page, from, from + copied);
405 page_cache_release(page);
410 struct address_space_operations ll_aops = {
411 .readpage = ll_readpage,
412 // .readpages = ll_readpages,
413 .direct_IO = ll_direct_IO_26,
414 .writepage = ll_writepage_26,
415 .writepages = generic_writepages,
416 .set_page_dirty = __set_page_dirty_nobuffers,
418 #ifdef HAVE_KERNEL_WRITE_BEGIN_END
419 .write_begin = ll_write_begin,
420 .write_end = ll_write_end,
422 .prepare_write = ll_prepare_write,
423 .commit_write = ll_commit_write,
425 .invalidatepage = ll_invalidatepage,
426 .releasepage = ll_releasepage,