lustre/llite/rw26.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  * Lustre Lite I/O page cache routines for the 2.5/2.6 kernel version
   5  *
   6  *  Copyright (c) 2001-2003 Cluster File Systems, Inc.
   7  *
   8  *   This file is part of Lustre, http://www.lustre.org.
   9  *
  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.
  13  *
  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.
  18  *
  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.
  22  */
  23 #ifndef AUTOCONF_INCLUDED
  24 #include <linux/config.h>
  25 #endif
  26 #include <linux/kernel.h>
  27 #include <linux/mm.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>
  36
  37 #include <linux/fs.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 #ifdef HAVE_SEGMENT_H
  44 # include <asm/segment.h>
  45 #endif
  46 #include <linux/mm.h>
  47 #include <linux/pagemap.h>
  48 #include <linux/smp_lock.h>
  49
  50 #define DEBUG_SUBSYSTEM S_LLITE
  51
  52 #include <lustre_lite.h>
  53 #include "llite_internal.h"
  54 #include <linux/lustre_compat25.h>
  55
  56 static int ll_writepage_26(struct page *page, struct writeback_control *wbc)
  57 {
  58         return ll_writepage(page);
  59 }
  60
  61 /* It is safe to not check anything in invalidatepage/releasepage below
  62    because they are run with page locked and all our io is happening with
  63    locked page too */
  64 #ifdef HAVE_INVALIDATEPAGE_RETURN_INT
  65 static int ll_invalidatepage(struct page *page, unsigned long offset)
  66 {
  67         if (offset)
  68                 return 0;
  69         if (PagePrivate(page))
  70                 ll_removepage(page);
  71         return 1;
  72 }
  73 #else
  74 static void ll_invalidatepage(struct page *page, unsigned long offset)
  75 {
  76         if (offset)
  77                 return;
  78         if (PagePrivate(page))
  79                 ll_removepage(page);
  80 }
  81 #endif
  82
  83 static int ll_releasepage(struct page *page, gfp_t gfp_mask)
  84 {
  85         if (PagePrivate(page))
  86                 ll_removepage(page);
  87         return 1;
  88 }
  89
  90 #define MAX_DIRECTIO_SIZE 2*1024*1024*1024UL
  91
  92 static inline int ll_get_user_pages(int rw, unsigned long user_addr,
  93                                     size_t size, struct page ***pages)
  94 {
  95         int result = -ENOMEM;
  96         int page_count;
  97
  98         /* set an arbitrary limit to prevent arithmetic overflow */
  99         if (size > MAX_DIRECTIO_SIZE) {
 100                 *pages = NULL;
 101                 return -EFBIG;
 102         }
 103
 104         page_count = ((user_addr + size + CFS_PAGE_SIZE - 1) >> CFS_PAGE_SHIFT)-
 105                       (user_addr >> CFS_PAGE_SHIFT);
 106
 107         OBD_ALLOC_WAIT(*pages, page_count * sizeof(**pages));
 108         if (*pages) {
 109                 down_read(&current->mm->mmap_sem);
 110                 result = get_user_pages(current, current->mm, user_addr,
 111                                         page_count, (rw == READ), 0, *pages,
 112                                         NULL);
 113                 up_read(&current->mm->mmap_sem);
 114                 if (result < 0)
 115                         OBD_FREE(*pages, page_count * sizeof(**pages));
 116         }
 117
 118         return result;
 119 }
 120
 121 /*  ll_free_user_pages - tear down page struct array
 122  *  @pages: array of page struct pointers underlying target buffer */
 123 static void ll_free_user_pages(struct page **pages, int npages, int do_dirty)
 124 {
 125         int i;
 126
 127         for (i = 0; i < npages; i++) {
 128                 if (do_dirty)
 129                         set_page_dirty_lock(pages[i]);
 130                 page_cache_release(pages[i]);
 131         }
 132
 133         OBD_FREE(pages, npages * sizeof(*pages));
 134 }
 135
 136 static ssize_t ll_direct_IO_26_seg(int rw, struct inode *inode,
 137                                    struct address_space *mapping,
 138                                    struct obd_info *oinfo,
 139                                    struct ptlrpc_request_set *set,
 140                                    size_t size, loff_t file_offset,
 141                                    struct page **pages, int page_count)
 142 {
 143         struct brw_page *pga;
 144         int i, rc = 0;
 145         size_t length;
 146         ENTRY;
 147
 148         OBD_ALLOC(pga, sizeof(*pga) * page_count);
 149         if (!pga) {
 150                 CDEBUG(D_VFSTRACE, "sizeof(*pga) = %u page_count = %u\n",
 151                       (int)sizeof(*pga), page_count);
 152                 RETURN(-ENOMEM);
 153         }
 154
 155         for (i = 0, length = size; length > 0;
 156              length -=pga[i].count, file_offset +=pga[i].count,i++) {/*i last!*/
 157                 pga[i].pg = pages[i];
 158                 pga[i].off = file_offset;
 159                 /* To the end of the page, or the length, whatever is less */
 160                 pga[i].count = min_t(int, CFS_PAGE_SIZE -(file_offset & ~CFS_PAGE_MASK),
 161                                      length);
 162                 pga[i].flag = 0;
 163                 if (rw == READ)
 164                         POISON_PAGE(pages[i], 0x0d);
 165         }
 166
 167         rc = obd_brw_async(rw == WRITE ? OBD_BRW_WRITE : OBD_BRW_READ,
 168                            ll_i2obdexp(inode), oinfo, page_count,
 169                            pga, NULL, set);
 170         if (rc == 0)
 171                 rc = size;
 172
 173         OBD_FREE(pga, sizeof(*pga) * page_count);
 174         RETURN(rc);
 175 }
 176
 177 /* This is the maximum size of a single O_DIRECT request, based on a 128kB
 178  * kmalloc limit.  We need to fit all of the brw_page structs, each one
 179  * representing PAGE_SIZE worth of user data, into a single buffer, and
 180  * then truncate this to be a full-sized RPC.  This is 22MB for 4kB pages. */
 181 #define MAX_DIO_SIZE ((128 * 1024 / sizeof(struct brw_page) * CFS_PAGE_SIZE) & \
 182                       ~(PTLRPC_MAX_BRW_SIZE - 1))
 183 static ssize_t ll_direct_IO_26(int rw, struct kiocb *iocb,
 184                                const struct iovec *iov, loff_t file_offset,
 185                                unsigned long nr_segs)
 186 {
 187         struct file *file = iocb->ki_filp;
 188         struct inode *inode = file->f_mapping->host;
 189         ssize_t count = iov_length(iov, nr_segs), tot_bytes = 0;
 190         struct ll_inode_info *lli = ll_i2info(inode);
 191         struct lov_stripe_md *lsm = lli->lli_smd;
 192         struct ptlrpc_request_set *set;
 193         struct obd_info oinfo;
 194         struct obdo oa;
 195         unsigned long seg;
 196         size_t size = MAX_DIO_SIZE;
 197         ENTRY;
 198
 199         if (!lli->lli_smd || !lli->lli_smd->lsm_object_id)
 200                 RETURN(-EBADF);
 201
 202         /* FIXME: io smaller than CFS_PAGE_SIZE is broken on ia64 ??? */
 203         if ((file_offset & (~CFS_PAGE_MASK)) || (count & ~CFS_PAGE_MASK))
 204                 RETURN(-EINVAL);
 205
 206         CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p), size="LPSZ" (max %lu), "
 207                "offset=%lld=%llx, pages "LPSZ" (max %lu)\n",
 208                inode->i_ino, inode->i_generation, inode, count, MAX_DIO_SIZE,
 209                file_offset, file_offset, count >> CFS_PAGE_SHIFT,
 210                MAX_DIO_SIZE >> CFS_PAGE_SHIFT);
 211
 212         if (rw == WRITE)
 213                 ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_DIRECT_WRITE, count);
 214         else
 215                 ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_DIRECT_READ, count);
 216
 217         /* Check that all user buffers are aligned as well */
 218         for (seg = 0; seg < nr_segs; seg++) {
 219                 if (((unsigned long)iov[seg].iov_base & ~CFS_PAGE_MASK) ||
 220                     (iov[seg].iov_len & ~CFS_PAGE_MASK))
 221                         RETURN(-EINVAL);
 222         }
 223
 224         set = ptlrpc_prep_set();
 225         if (set == NULL)
 226                 RETURN(-ENOMEM);
 227
 228         ll_inode_fill_obdo(inode, rw == WRITE ? OBD_BRW_WRITE : OBD_BRW_READ, &oa);
 229         oinfo.oi_oa = &oa;
 230         oinfo.oi_md = lsm;
 231
 232         /* need locking between buffered and direct access. and race with
 233          *size changing by concurrent truncates and writes. */
 234         if (rw == READ)
 235                 LOCK_INODE_MUTEX(inode);
 236
 237         for (seg = 0; seg < nr_segs; seg++) {
 238                 size_t iov_left = iov[seg].iov_len;
 239                 unsigned long user_addr = (unsigned long)iov[seg].iov_base;
 240
 241                 if (rw == READ) {
 242                         if (file_offset >= inode->i_size)
 243                                 break;
 244                         if (file_offset + iov_left > inode->i_size)
 245                                 iov_left = inode->i_size - file_offset;
 246                 }
 247
 248                 while (iov_left > 0) {
 249                         struct page **pages;
 250                         int page_count;
 251                         ssize_t result;
 252
 253                         page_count = ll_get_user_pages(rw, user_addr,
 254                                                        min(size, iov_left),
 255                                                        &pages);
 256                         LASSERT(page_count != 0);
 257                         if (page_count > 0) {
 258                                 result = ll_direct_IO_26_seg(rw, inode,
 259                                                              file->f_mapping,
 260                                                              &oinfo, set,
 261                                                              min(size,iov_left),
 262                                                              file_offset, pages,
 263                                                              page_count);
 264                                 ll_free_user_pages(pages, page_count, rw==READ);
 265                         } else {
 266                                 result = 0;
 267                         }
 268                         if (page_count < 0 || result <= 0) {
 269                                 /* If we can't allocate a large enough buffer
 270                                  * for the request, shrink it to a smaller
 271                                  * PAGE_SIZE multiple and try again.
 272                                  * We should always be able to kmalloc for a
 273                                  * page worth of page pointers = 4MB on i386. */
 274                                 if ((page_count == -ENOMEM||result == -ENOMEM)&&
 275                                     size > (CFS_PAGE_SIZE / sizeof(*pages)) *
 276                                            CFS_PAGE_SIZE) {
 277                                         size = ((((size / 2) - 1) |
 278                                                  ~CFS_PAGE_MASK) + 1) &
 279                                                 CFS_PAGE_MASK;
 280                                         CDEBUG(D_VFSTRACE, "DIO size now %u\n",
 281                                                (int)size);
 282                                         continue;
 283                                 }
 284                                 if (tot_bytes > 0)
 285                                         GOTO(wait_io, tot_bytes);
 286                                 GOTO(out, tot_bytes = page_count < 0 ? page_count : result);
 287                         }
 288
 289                         tot_bytes += result;
 290                         file_offset += result;
 291                         iov_left -= result;
 292                         user_addr += result;
 293                 }
 294         }
 295
 296         if (tot_bytes > 0) {
 297                 int rc;
 298         wait_io:
 299                 rc = ptlrpc_set_wait(set);
 300                 if (rc)
 301                         GOTO(out, tot_bytes = rc);
 302                 if (rw == WRITE) {
 303                         lov_stripe_lock(lsm);
 304                         obd_adjust_kms(ll_i2obdexp(inode), lsm, file_offset, 0);
 305                         lov_stripe_unlock(lsm);
 306                 }
 307         }
 308 out:
 309         if (rw == READ)
 310                 UNLOCK_INODE_MUTEX(inode);
 311
 312         ptlrpc_set_destroy(set);
 313         RETURN(tot_bytes);
 314 }
 315
 316 struct address_space_operations ll_aops = {
 317         .readpage       = ll_readpage,
 318 //        .readpages      = ll_readpages,
 319         .direct_IO      = ll_direct_IO_26,
 320         .writepage      = ll_writepage_26,
 321         .writepages     = generic_writepages,
 322         .set_page_dirty = __set_page_dirty_nobuffers,
 323         .sync_page      = NULL,
 324         .prepare_write  = ll_prepare_write,
 325         .commit_write   = ll_commit_write,
 326         .invalidatepage = ll_invalidatepage,
 327         .releasepage    = ll_releasepage,
 328         .bmap           = NULL
 329 };