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
  24 #ifndef AUTOCONF_INCLUDED
  25 #include <linux/config.h>
  26 #endif
  27 #include <linux/kernel.h>
  28 #include <linux/mm.h>
  29 #include <linux/string.h>
  30 #include <linux/stat.h>
  31 #include <linux/errno.h>
  32 #include <linux/smp_lock.h>
  33 #include <linux/unistd.h>
  34 #include <linux/version.h>
  35 #include <asm/system.h>
  36 #include <asm/uaccess.h>
  37
  38 #include <linux/fs.h>
  39 #include <linux/buffer_head.h>
  40 #include <linux/mpage.h>
  41 #include <linux/writeback.h>
  42 #include <linux/stat.h>
  43 #include <asm/uaccess.h>
  44 #include <asm/segment.h>
  45 #include <linux/mm.h>
  46 #include <linux/pagemap.h>
  47 #include <linux/smp_lock.h>
  48
  49 #define DEBUG_SUBSYSTEM S_LLITE
  50
  51 //#include <lustre_mdc.h>
  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 == 0 && PagePrivate(page))
  77                 ll_removepage(page);
  78 }
  79 #endif
  80 static int ll_releasepage(struct page *page, int gfp_mask)
  81 {
  82         if (PagePrivate(page))
  83                 ll_removepage(page);
  84         return 1;
  85 }
  86
  87 static int ll_set_page_dirty(struct page *page)
  88 {
  89         struct ll_async_page *llap;
  90         ENTRY;
  91
  92         llap = llap_from_page(page, LLAP_ORIGIN_UNKNOWN);
  93         if (IS_ERR(llap))
  94                 RETURN(PTR_ERR(llap));
  95
  96         llap_write_pending(page->mapping->host, llap);
  97         RETURN(__set_page_dirty_nobuffers(page));
  98 }
  99
 100 #define MAX_DIRECTIO_SIZE 2*1024*1024*1024UL
 101
 102 static inline int ll_get_user_pages(int rw, unsigned long user_addr,
 103                                     size_t size, struct page ***pages)
 104 {
 105         int result = -ENOMEM;
 106         int page_count;
 107
 108         /* set an arbitrary limit to prevent arithmetic overflow */
 109         if (size > MAX_DIRECTIO_SIZE) {
 110                 *pages = NULL;
 111                 return -EFBIG;
 112         }
 113
 114         page_count = (user_addr + size + CFS_PAGE_SIZE - 1) >> CFS_PAGE_SHIFT;
 115         page_count -= user_addr >> CFS_PAGE_SHIFT;
 116
 117         OBD_ALLOC_WAIT(*pages, page_count * sizeof(**pages));
 118         if (*pages) {
 119                 down_read(&current->mm->mmap_sem);
 120                 result = get_user_pages(current, current->mm, user_addr,
 121                                         page_count, (rw == READ), 0, *pages,
 122                                         NULL);
 123                 up_read(&current->mm->mmap_sem);
 124                 if (result < 0)
 125                         OBD_FREE(*pages, page_count * sizeof(**pages));
 126         }
 127
 128         return result;
 129 }
 130
 131 /*  ll_free_user_pages - tear down page struct array
 132  *  @pages: array of page struct pointers underlying target buffer */
 133 static void ll_free_user_pages(struct page **pages, int npages, int do_dirty)
 134 {
 135         int i;
 136
 137         for (i = 0; i < npages; i++) {
 138                 if (do_dirty)
 139                         set_page_dirty_lock(pages[i]);
 140                 page_cache_release(pages[i]);
 141         }
 142
 143         OBD_FREE(pages, npages * sizeof(*pages));
 144 }
 145
 146 static ssize_t ll_direct_IO_26_seg(int rw, struct inode *inode,
 147                                    struct address_space *mapping,
 148                                    struct obd_info *oinfo,
 149                                    struct ptlrpc_request_set *set,
 150                                    size_t size, loff_t file_offset,
 151                                    struct page **pages, int page_count)
 152 {
 153         struct brw_page *pga;
 154         int i, rc = 0;
 155         size_t length;
 156         ENTRY;
 157
 158         OBD_ALLOC(pga, sizeof(*pga) * page_count);
 159         if (!pga) {
 160                 CDEBUG(D_VFSTRACE, "sizeof(*pga) = %u page_count = %u\n",
 161                        (int)sizeof(*pga), page_count);
 162                 RETURN(-ENOMEM);
 163         }
 164
 165         for (i = 0, length = size; length > 0;
 166              length -=pga[i].count, file_offset +=pga[i].count,i++) {/*i last!*/
 167                 pga[i].pg = pages[i];
 168                 pga[i].off = file_offset;
 169                 /* To the end of the page, or the length, whatever is less */
 170                 pga[i].count = min_t(int, CFS_PAGE_SIZE -
 171                                           (file_offset & ~CFS_PAGE_MASK),
 172                                      length);
 173                 pga[i].flag = 0;
 174                 if (rw == READ)
 175                         POISON_PAGE(pages[i], 0x0d);
 176         }
 177
 178         rc = obd_brw_async(rw == WRITE ? OBD_BRW_WRITE : OBD_BRW_READ,
 179                                 ll_i2dtexp(inode), oinfo, page_count,
 180                                 pga, NULL, set);
 181         if (rc == 0)
 182                 rc = size;
 183
 184         OBD_FREE(pga, sizeof(*pga) * page_count);
 185         RETURN(rc);
 186 }
 187
 188 /* This is the maximum size of a single O_DIRECT request, based on a 128kB
 189  * kmalloc limit.  We need to fit all of the brw_page structs, each one
 190  * representing PAGE_SIZE worth of user data, into a single buffer, and
 191  * then truncate this to be a full-sized RPC.  This is 22MB for 4kB pages. */
 192 #define MAX_DIO_SIZE ((128 * 1024 / sizeof(struct brw_page) * CFS_PAGE_SIZE) & \
 193                       ~(PTLRPC_MAX_BRW_SIZE - 1))
 194 static ssize_t ll_direct_IO_26(int rw, struct kiocb *iocb,
 195                                const struct iovec *iov, loff_t file_offset,
 196                                unsigned long nr_segs)
 197 {
 198         struct file *file = iocb->ki_filp;
 199         struct inode *inode = file->f_mapping->host;
 200         ssize_t count = iov_length(iov, nr_segs), tot_bytes = 0;
 201         struct ll_inode_info *lli = ll_i2info(inode);
 202         struct lov_stripe_md *lsm = lli->lli_smd;
 203         struct ptlrpc_request_set *set;
 204         struct obd_info oinfo;
 205         struct obdo oa;
 206         unsigned long seg = 0;
 207         size_t size = MAX_DIO_SIZE;
 208         int opc;
 209         ENTRY;
 210
 211         if (!lli->lli_smd || !lli->lli_smd->lsm_object_id)
 212                 RETURN(-EBADF);
 213
 214         /* FIXME: io smaller than PAGE_SIZE is broken on ia64 ??? */
 215         if ((file_offset & ~CFS_PAGE_MASK) || (count & ~CFS_PAGE_MASK))
 216                 RETURN(-EINVAL);
 217
 218         CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p), size="LPSZ" (max %lu), "
 219                "offset=%lld=%llx, pages "LPSZ" (max %lu)\n",
 220                inode->i_ino, inode->i_generation, inode, count, MAX_DIO_SIZE,
 221                file_offset, file_offset, count >> CFS_PAGE_SHIFT,
 222                MAX_DIO_SIZE >> CFS_PAGE_SHIFT);
 223
 224         if (rw == WRITE) {
 225                 ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_DIRECT_WRITE, count);
 226                 opc = CAPA_OPC_OSS_WRITE;
 227                 llap_write_pending(inode, NULL);
 228         } else {
 229                 ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_DIRECT_READ, count);
 230                 opc = CAPA_OPC_OSS_RW;
 231         }
 232
 233         /* Check that all user buffers are aligned as well */
 234         for (seg = 0; seg < nr_segs; seg++) {
 235                 if (((unsigned long)iov[seg].iov_base & ~CFS_PAGE_MASK) ||
 236                     (iov[seg].iov_len & ~CFS_PAGE_MASK))
 237                         RETURN(-EINVAL);
 238         }
 239
 240         set = ptlrpc_prep_set();
 241         if (set == NULL)
 242                 RETURN(-ENOMEM);
 243
 244         ll_inode_fill_obdo(inode, rw, &oa);
 245         oinfo.oi_oa = &oa;
 246         oinfo.oi_md = lsm;
 247         oinfo.oi_capa = ll_osscapa_get(inode, opc);
 248
 249         /* need locking between buffered and direct access. and race with
 250          *size changing by concurrent truncates and writes. */
 251         if (rw == READ)
 252                 LOCK_INODE_MUTEX(inode);
 253
 254         for (seg = 0; seg < nr_segs; seg++) {
 255                 size_t iov_left = iov[seg].iov_len;
 256                 unsigned long user_addr = (unsigned long)iov[seg].iov_base;
 257
 258                 if (rw == READ) {
 259                         if (file_offset >= inode->i_size)
 260                                 break;
 261                         if (file_offset + iov_left > inode->i_size)
 262                                 iov_left = inode->i_size - file_offset;
 263                 }
 264
 265                 while (iov_left > 0) {
 266                         struct page **pages;
 267                         int page_count;
 268                         ssize_t result;
 269
 270                         page_count = ll_get_user_pages(rw, user_addr,
 271                                                        min(size, iov_left),
 272                                                        &pages);
 273                         LASSERT(page_count != 0);
 274                         if (page_count > 0) {
 275                                 result = ll_direct_IO_26_seg(rw, inode,
 276                                                              file->f_mapping,
 277                                                              &oinfo, set,
 278                                                              min(size,iov_left),
 279                                                              file_offset, pages,
 280                                                              page_count);
 281                                 ll_free_user_pages(pages, page_count, rw==READ);
 282                         } else {
 283                                 result = 0;
 284                         }
 285                         if (page_count < 0 || result <= 0) {
 286                                 /* If we can't allocate a large enough buffer
 287                                  * for the request, shrink it to a smaller
 288                                  * PAGE_SIZE multiple and try again.
 289                                  * We should always be able to kmalloc for a
 290                                  * page worth of page pointers = 4MB on i386. */
 291                                 if ((page_count == -ENOMEM||result == -ENOMEM)&&
 292                                     size > (CFS_PAGE_SIZE / sizeof(*pages)) *
 293                                            CFS_PAGE_SIZE) {
 294                                         size = ((((size / 2) - 1) |
 295                                                  ~CFS_PAGE_MASK) + 1) &
 296                                                 CFS_PAGE_MASK;
 297                                         CDEBUG(D_VFSTRACE, "DIO size now %u\n",
 298                                                (int)size);
 299                                         continue;
 300                                 }
 301
 302                                 if (tot_bytes <= 0)
 303                                         tot_bytes = page_count < 0 ? page_count : result;
 304                                 GOTO(out, tot_bytes);
 305                         }
 306
 307                         tot_bytes += result;
 308                         file_offset += result;
 309                         iov_left -= result;
 310                         user_addr += result;
 311                 }
 312         }
 313 out:
 314         if (rw == READ)
 315                 UNLOCK_INODE_MUTEX(inode);
 316
 317         if (tot_bytes > 0) {
 318                 int rc;
 319
 320                 rc = ptlrpc_set_wait(set);
 321                 if (rc) {
 322                         tot_bytes = rc;
 323                 } else if (rw == WRITE) {
 324                         lov_stripe_lock(lsm);
 325                         obd_adjust_kms(ll_i2dtexp(inode), lsm, file_offset, 0);
 326                         lov_stripe_unlock(lsm);
 327                 }
 328         }
 329
 330         capa_put(oinfo.oi_capa);
 331         ptlrpc_set_destroy(set);
 332         RETURN(tot_bytes);
 333 }
 334
 335 struct address_space_operations ll_aops = {
 336         .readpage       = ll_readpage,
 337 //        .readpages      = ll_readpages,
 338         .direct_IO      = ll_direct_IO_26,
 339         .writepage      = ll_writepage_26,
 340         .writepages     = generic_writepages,
 341         .set_page_dirty = ll_set_page_dirty,
 342         .sync_page      = NULL,
 343         .prepare_write  = ll_prepare_write,
 344         .commit_write   = ll_commit_write,
 345         .invalidatepage = ll_invalidatepage,
 346         .releasepage    = ll_releasepage,
 347         .bmap           = NULL
 348 };