lustre/llite/rw26.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  * GPL HEADER START
   5  *
   6  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   7  *
   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.
  11  *
  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).
  17  *
  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
  21  *
  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
  24  * have any questions.
  25  *
  26  * GPL HEADER END
  27  */
  28 /*
  29  * Copyright  2008 Sun Microsystems, Inc. All rights reserved
  30  * Use is subject to license terms.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  *
  36  * lustre/lustre/llite/rw26.c
  37  *
  38  * Lustre Lite I/O page cache routines for the 2.5/2.6 kernel version
  39  */
  40
  41 #ifndef AUTOCONF_INCLUDED
  42 #include <linux/config.h>
  43 #endif
  44 #include <linux/kernel.h>
  45 #include <linux/mm.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>
  54
  55 #include <linux/fs.h>
  56 #include <linux/buffer_head.h>
  57 #include <linux/writeback.h>
  58 #include <linux/stat.h>
  59 #include <asm/uaccess.h>
  60 #include <linux/mm.h>
  61 #include <linux/pagemap.h>
  62 #include <linux/smp_lock.h>
  63
  64 #define DEBUG_SUBSYSTEM S_LLITE
  65
  66 //#include <lustre_mdc.h>
  67 #include <lustre_lite.h>
  68 #include "llite_internal.h"
  69 #include <linux/lustre_compat25.h>
  70
  71 static int ll_writepage_26(struct page *page, struct writeback_control *wbc)
  72 {
  73         return ll_writepage(page);
  74 }
  75
  76 /* It is safe to not check anything in invalidatepage/releasepage below
  77    because they are run with page locked and all our io is happening with
  78    locked page too */
  79 #ifdef HAVE_INVALIDATEPAGE_RETURN_INT
  80 static int ll_invalidatepage(struct page *page, unsigned long offset)
  81 {
  82         if (offset)
  83                 return 0;
  84         if (PagePrivate(page))
  85                 ll_removepage(page);
  86         return 1;
  87 }
  88 #else
  89 static void ll_invalidatepage(struct page *page, unsigned long offset)
  90 {
  91         if (offset == 0 && PagePrivate(page))
  92                 ll_removepage(page);
  93 }
  94 #endif
  95
  96 #ifdef HAVE_RELEASEPAGE_WITH_INT
  97 #define RELEASEPAGE_ARG_TYPE int
  98 #else
  99 #define RELEASEPAGE_ARG_TYPE gfp_t
 100 #endif
 101 static int ll_releasepage(struct page *page, RELEASEPAGE_ARG_TYPE gfp_mask)
 102 {
 103         if (PagePrivate(page))
 104                 ll_removepage(page);
 105         return 1;
 106 }
 107
 108 static int ll_set_page_dirty(struct page *page)
 109 {
 110         struct ll_async_page *llap;
 111         ENTRY;
 112
 113         llap = llap_from_page(page, LLAP_ORIGIN_UNKNOWN);
 114         if (IS_ERR(llap))
 115                 RETURN(PTR_ERR(llap));
 116
 117         llap_write_pending(page->mapping->host, llap);
 118         RETURN(__set_page_dirty_nobuffers(page));
 119 }
 120
 121 #define MAX_DIRECTIO_SIZE 2*1024*1024*1024UL
 122
 123 static inline int ll_get_user_pages(int rw, unsigned long user_addr,
 124                                     size_t size, struct page ***pages)
 125 {
 126         int result = -ENOMEM;
 127         int page_count;
 128
 129         /* set an arbitrary limit to prevent arithmetic overflow */
 130         if (size > MAX_DIRECTIO_SIZE) {
 131                 *pages = NULL;
 132                 return -EFBIG;
 133         }
 134
 135         page_count = (user_addr + size + CFS_PAGE_SIZE - 1) >> CFS_PAGE_SHIFT;
 136         page_count -= user_addr >> CFS_PAGE_SHIFT;
 137
 138         OBD_ALLOC_WAIT(*pages, page_count * sizeof(**pages));
 139         if (*pages) {
 140                 down_read(&current->mm->mmap_sem);
 141                 result = get_user_pages(current, current->mm, user_addr,
 142                                         page_count, (rw == READ), 0, *pages,
 143                                         NULL);
 144                 up_read(&current->mm->mmap_sem);
 145                 if (result < 0)
 146                         OBD_FREE(*pages, page_count * sizeof(**pages));
 147         }
 148
 149         return result;
 150 }
 151
 152 /*  ll_free_user_pages - tear down page struct array
 153  *  @pages: array of page struct pointers underlying target buffer */
 154 static void ll_free_user_pages(struct page **pages, int npages, int do_dirty)
 155 {
 156         int i;
 157
 158         for (i = 0; i < npages; i++) {
 159                 if (do_dirty)
 160                         set_page_dirty_lock(pages[i]);
 161                 page_cache_release(pages[i]);
 162         }
 163
 164         OBD_FREE(pages, npages * sizeof(*pages));
 165 }
 166
 167 static ssize_t ll_direct_IO_26_seg(int rw, struct inode *inode,
 168                                    struct address_space *mapping,
 169                                    struct obd_info *oinfo,
 170                                    struct ptlrpc_request_set *set,
 171                                    size_t size, loff_t file_offset,
 172                                    struct page **pages, int page_count)
 173 {
 174         struct brw_page *pga;
 175         int i, rc = 0;
 176         size_t length;
 177         ENTRY;
 178
 179         OBD_ALLOC(pga, sizeof(*pga) * page_count);
 180         if (!pga) {
 181                 CDEBUG(D_VFSTRACE, "sizeof(*pga) = %u page_count = %u\n",
 182                        (int)sizeof(*pga), page_count);
 183                 RETURN(-ENOMEM);
 184         }
 185
 186         for (i = 0, length = size; length > 0;
 187              length -=pga[i].count, file_offset +=pga[i].count,i++) {/*i last!*/
 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 -
 192                                           (file_offset & ~CFS_PAGE_MASK),
 193                                      length);
 194                 pga[i].flag = 0;
 195                 if (rw == READ)
 196                         POISON_PAGE(pages[i], 0x0d);
 197         }
 198
 199         rc = obd_brw_async(rw == WRITE ? OBD_BRW_WRITE : OBD_BRW_READ,
 200                                 ll_i2dtexp(inode), oinfo, page_count,
 201                                 pga, NULL, set);
 202         if (rc == 0)
 203                 rc = size;
 204
 205         OBD_FREE(pga, sizeof(*pga) * page_count);
 206         RETURN(rc);
 207 }
 208
 209 /* This is the maximum size of a single O_DIRECT request, based on a 128kB
 210  * kmalloc limit.  We need to fit all of the brw_page structs, each one
 211  * representing PAGE_SIZE worth of user data, into a single buffer, and
 212  * then truncate this to be a full-sized RPC.  This is 22MB for 4kB pages. */
 213 #define MAX_DIO_SIZE ((128 * 1024 / sizeof(struct brw_page) * CFS_PAGE_SIZE) & \
 214                       ~(PTLRPC_MAX_BRW_SIZE - 1))
 215 static ssize_t ll_direct_IO_26(int rw, struct kiocb *iocb,
 216                                const struct iovec *iov, loff_t file_offset,
 217                                unsigned long nr_segs)
 218 {
 219         struct file *file = iocb->ki_filp;
 220         struct inode *inode = file->f_mapping->host;
 221         ssize_t count = iov_length(iov, nr_segs), tot_bytes = 0;
 222         struct ll_inode_info *lli = ll_i2info(inode);
 223         struct lov_stripe_md *lsm = lli->lli_smd;
 224         struct ptlrpc_request_set *set;
 225         struct obd_info oinfo;
 226         struct obdo oa;
 227         unsigned long seg = 0;
 228         size_t size = MAX_DIO_SIZE;
 229         int opc;
 230         ENTRY;
 231
 232         if (!lli->lli_smd || !lli->lli_smd->lsm_object_id)
 233                 RETURN(-EBADF);
 234
 235         /* FIXME: io smaller than PAGE_SIZE is broken on ia64 ??? */
 236         if ((file_offset & ~CFS_PAGE_MASK) || (count & ~CFS_PAGE_MASK))
 237                 RETURN(-EINVAL);
 238
 239         CDEBUG(D_VFSTRACE, "VFS Op:inode=%lu/%u(%p), size="LPSZ" (max %lu), "
 240                "offset=%lld=%llx, pages "LPSZ" (max %lu)\n",
 241                inode->i_ino, inode->i_generation, inode, count, MAX_DIO_SIZE,
 242                file_offset, file_offset, count >> CFS_PAGE_SHIFT,
 243                MAX_DIO_SIZE >> CFS_PAGE_SHIFT);
 244
 245         if (rw == WRITE) {
 246                 ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_DIRECT_WRITE, count);
 247                 opc = CAPA_OPC_OSS_WRITE;
 248                 llap_write_pending(inode, NULL);
 249         } else {
 250                 ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_DIRECT_READ, count);
 251                 opc = CAPA_OPC_OSS_RW;
 252         }
 253
 254         /* Check that all user buffers are aligned as well */
 255         for (seg = 0; seg < nr_segs; seg++) {
 256                 if (((unsigned long)iov[seg].iov_base & ~CFS_PAGE_MASK) ||
 257                     (iov[seg].iov_len & ~CFS_PAGE_MASK))
 258                         RETURN(-EINVAL);
 259         }
 260
 261         set = ptlrpc_prep_set();
 262         if (set == NULL)
 263                 RETURN(-ENOMEM);
 264
 265         ll_inode_fill_obdo(inode, rw, &oa);
 266         oinfo.oi_oa = &oa;
 267         oinfo.oi_md = lsm;
 268         oinfo.oi_capa = ll_osscapa_get(inode, opc);
 269
 270         /* need locking between buffered and direct access. and race with
 271          *size changing by concurrent truncates and writes. */
 272         if (rw == READ)
 273                 LOCK_INODE_MUTEX(inode);
 274
 275         for (seg = 0; seg < nr_segs; seg++) {
 276                 size_t iov_left = iov[seg].iov_len;
 277                 unsigned long user_addr = (unsigned long)iov[seg].iov_base;
 278
 279                 if (rw == READ) {
 280                         if (file_offset >= inode->i_size)
 281                                 break;
 282                         if (file_offset + iov_left > inode->i_size)
 283                                 iov_left = inode->i_size - file_offset;
 284                 }
 285
 286                 while (iov_left > 0) {
 287                         struct page **pages;
 288                         int page_count;
 289                         ssize_t result;
 290
 291                         page_count = ll_get_user_pages(rw, user_addr,
 292                                                        min(size, iov_left),
 293                                                        &pages);
 294                         LASSERT(page_count != 0);
 295                         if (page_count > 0) {
 296                                 result = ll_direct_IO_26_seg(rw, inode,
 297                                                              file->f_mapping,
 298                                                              &oinfo, set,
 299                                                              min(size,iov_left),
 300                                                              file_offset, pages,
 301                                                              page_count);
 302                                 ll_free_user_pages(pages, page_count, rw==READ);
 303                         } else {
 304                                 result = 0;
 305                         }
 306                         if (page_count < 0 || result <= 0) {
 307                                 /* If we can't allocate a large enough buffer
 308                                  * for the request, shrink it to a smaller
 309                                  * PAGE_SIZE multiple and try again.
 310                                  * We should always be able to kmalloc for a
 311                                  * page worth of page pointers = 4MB on i386. */
 312                                 if ((page_count == -ENOMEM||result == -ENOMEM)&&
 313                                     size > (CFS_PAGE_SIZE / sizeof(*pages)) *
 314                                            CFS_PAGE_SIZE) {
 315                                         size = ((((size / 2) - 1) |
 316                                                  ~CFS_PAGE_MASK) + 1) &
 317                                                 CFS_PAGE_MASK;
 318                                         CDEBUG(D_VFSTRACE, "DIO size now %u\n",
 319                                                (int)size);
 320                                         continue;
 321                                 }
 322
 323                                 if (tot_bytes <= 0)
 324                                         tot_bytes = page_count < 0 ? page_count : result;
 325                                 GOTO(out, tot_bytes);
 326                         }
 327
 328                         tot_bytes += result;
 329                         file_offset += result;
 330                         iov_left -= result;
 331                         user_addr += result;
 332                 }
 333         }
 334 out:
 335         if (rw == READ)
 336                 UNLOCK_INODE_MUTEX(inode);
 337
 338         if (tot_bytes > 0) {
 339                 int rc;
 340
 341                 rc = ptlrpc_set_wait(set);
 342                 if (rc) {
 343                         tot_bytes = rc;
 344                 } else if (rw == WRITE) {
 345                         lov_stripe_lock(lsm);
 346                         obd_adjust_kms(ll_i2dtexp(inode), lsm, file_offset, 0);
 347                         lov_stripe_unlock(lsm);
 348                 }
 349         }
 350
 351         capa_put(oinfo.oi_capa);
 352         ptlrpc_set_destroy(set);
 353         RETURN(tot_bytes);
 354 }
 355
 356 struct address_space_operations ll_aops = {
 357         .readpage       = ll_readpage,
 358 //        .readpages      = ll_readpages,
 359         .direct_IO      = ll_direct_IO_26,
 360         .writepage      = ll_writepage_26,
 361         .writepages     = generic_writepages,
 362         .set_page_dirty = ll_set_page_dirty,
 363         .sync_page      = NULL,
 364         .prepare_write  = ll_prepare_write,
 365         .commit_write   = ll_commit_write,
 366         .invalidatepage = ll_invalidatepage,
 367         .releasepage    = ll_releasepage,
 368         .bmap           = NULL
 369 };