lustre/llite/llite_mmap.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.gnu.org/licenses/gpl-2.0.html
  19  *
  20  * GPL HEADER END
  21  */
  22 /*
  23  * Copyright (c) 2004, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Use is subject to license terms.
  25  *
  26  * Copyright (c) 2011, 2016, Intel Corporation.
  27  */
  28 /*
  29  * This file is part of Lustre, http://www.lustre.org/
  30  */
  31
  32 #include <linux/errno.h>
  33 #include <linux/delay.h>
  34 #include <linux/kernel.h>
  35 #include <linux/mm.h>
  36 #include <linux/file.h>
  37
  38 #define DEBUG_SUBSYSTEM S_LLITE
  39
  40 #include "llite_internal.h"
  41 #include <lustre_compat.h>
  42
  43 static const struct vm_operations_struct ll_file_vm_ops;
  44
  45 void policy_from_vma(union ldlm_policy_data *policy, struct vm_area_struct *vma,
  46                      unsigned long addr, size_t count)
  47 {
  48         policy->l_extent.start = ((addr - vma->vm_start) & PAGE_MASK) +
  49                                  (vma->vm_pgoff << PAGE_SHIFT);
  50         policy->l_extent.end = (policy->l_extent.start + count - 1) |
  51                                ~PAGE_MASK;
  52 }
  53
  54 struct vm_area_struct *our_vma(struct mm_struct *mm, unsigned long addr,
  55                                size_t count)
  56 {
  57         struct vm_area_struct *vma, *ret = NULL;
  58         ENTRY;
  59
  60         /* mmap_lock must have been held by caller. */
  61         LASSERT(!mmap_write_trylock(mm));
  62
  63         for (vma = find_vma(mm, addr);
  64              vma != NULL && vma->vm_start < (addr + count);
  65              vma = vma->vm_next) {
  66                 if (vma->vm_ops && vma->vm_ops == &ll_file_vm_ops &&
  67                     vma->vm_flags & VM_SHARED) {
  68                         ret = vma;
  69                         break;
  70                 }
  71         }
  72         RETURN(ret);
  73 }
  74
  75 /**
  76  * API independent part for page fault initialization.
  77  * \param env - corespondent lu_env to processing
  78  * \param vma - virtual memory area addressed to page fault
  79  * \param index - page index corespondent to fault.
  80  *
  81  * \return error codes from cl_io_init.
  82  */
  83 static struct cl_io *
  84 ll_fault_io_init(struct lu_env *env, struct vm_area_struct *vma, pgoff_t index)
  85 {
  86         struct file            *file = vma->vm_file;
  87         struct inode           *inode = file_inode(file);
  88         struct cl_io           *io;
  89         struct cl_fault_io     *fio;
  90         int                     rc;
  91         ENTRY;
  92
  93         if (ll_file_nolock(file))
  94                 RETURN(ERR_PTR(-EOPNOTSUPP));
  95
  96 restart:
  97         io = vvp_env_thread_io(env);
  98         io->ci_obj = ll_i2info(inode)->lli_clob;
  99         LASSERT(io->ci_obj != NULL);
 100
 101         fio = &io->u.ci_fault;
 102         fio->ft_index = index;
 103         fio->ft_executable = vma->vm_flags & VM_EXEC;
 104
 105         CDEBUG(D_MMAP,
 106                DFID": vma=%p start=%#lx end=%#lx vm_flags=%#lx idx=%lu\n",
 107                PFID(&ll_i2info(inode)->lli_fid), vma, vma->vm_start,
 108                vma->vm_end, vma->vm_flags, fio->ft_index);
 109
 110         if (vma->vm_flags & VM_SEQ_READ)
 111                 io->ci_seq_read = 1;
 112         else if (vma->vm_flags & VM_RAND_READ)
 113                 io->ci_rand_read = 1;
 114
 115         if (vma->vm_flags & VM_WRITE)
 116                 fio->ft_writable = 1;
 117
 118         rc = cl_io_init(env, io, CIT_FAULT, io->ci_obj);
 119         if (rc == 0) {
 120                 struct vvp_io *vio = vvp_env_io(env);
 121                 struct ll_file_data *fd = file->private_data;
 122
 123                 LASSERT(vio->vui_cl.cis_io == io);
 124
 125                 /* mmap lock must be MANDATORY it has to cache
 126                  * pages. */
 127                 io->ci_lockreq = CILR_MANDATORY;
 128                 vio->vui_fd = fd;
 129         } else {
 130                 cl_io_fini(env, io);
 131                 if (io->ci_need_restart)
 132                         goto restart;
 133
 134                 io = ERR_PTR(rc);
 135         }
 136
 137         RETURN(io);
 138 }
 139
 140 /* Sharing code of page_mkwrite method for rhel5 and rhel6 */
 141 static int ll_page_mkwrite0(struct vm_area_struct *vma, struct page *vmpage,
 142                             bool *retry)
 143 {
 144         struct lu_env           *env;
 145         struct cl_io            *io;
 146         struct vvp_io           *vio;
 147         int                      result;
 148         __u16                    refcheck;
 149         sigset_t old, new;
 150         struct inode             *inode = NULL;
 151         struct ll_inode_info     *lli;
 152         ENTRY;
 153
 154         LASSERT(vmpage != NULL);
 155         env = cl_env_get(&refcheck);
 156         if (IS_ERR(env))
 157                 RETURN(PTR_ERR(env));
 158
 159         io = ll_fault_io_init(env, vma, vmpage->index);
 160         if (IS_ERR(io))
 161                 GOTO(out, result = PTR_ERR(io));
 162
 163         result = io->ci_result;
 164         if (result < 0)
 165                 GOTO(out_io, result);
 166
 167         io->u.ci_fault.ft_mkwrite = 1;
 168         io->u.ci_fault.ft_writable = 1;
 169
 170         vio = vvp_env_io(env);
 171         vio->u.fault.ft_vma    = vma;
 172         vio->u.fault.ft_vmpage = vmpage;
 173
 174         siginitsetinv(&new, sigmask(SIGKILL) | sigmask(SIGTERM));
 175         sigprocmask(SIG_BLOCK, &new, &old);
 176
 177         inode = vvp_object_inode(io->ci_obj);
 178         lli = ll_i2info(inode);
 179
 180         result = cl_io_loop(env, io);
 181
 182         sigprocmask(SIG_SETMASK, &old, NULL);
 183
 184         if (result == 0) {
 185                 lock_page(vmpage);
 186                 if (vmpage->mapping == NULL) {
 187                         unlock_page(vmpage);
 188
 189                         /* page was truncated and lock was cancelled, return
 190                          * ENODATA so that VM_FAULT_NOPAGE will be returned
 191                          * to handle_mm_fault(). */
 192                         if (result == 0)
 193                                 result = -ENODATA;
 194                 } else if (!PageDirty(vmpage)) {
 195                         /* race, the page has been cleaned by ptlrpcd after
 196                          * it was unlocked, it has to be added into dirty
 197                          * cache again otherwise this soon-to-dirty page won't
 198                          * consume any grants, even worse if this page is being
 199                          * transferred because it will break RPC checksum.
 200                          */
 201                         unlock_page(vmpage);
 202
 203                         CDEBUG(D_MMAP, "Race on page_mkwrite %p/%lu, page has "
 204                                "been written out, retry.\n",
 205                                vmpage, vmpage->index);
 206
 207                         *retry = true;
 208                         result = -EAGAIN;
 209                 }
 210
 211                 if (result == 0)
 212                         set_bit(LLIF_DATA_MODIFIED, &lli->lli_flags);
 213         }
 214         EXIT;
 215
 216 out_io:
 217         cl_io_fini(env, io);
 218 out:
 219         cl_env_put(env, &refcheck);
 220         CDEBUG(D_MMAP, "%s mkwrite with %d\n", current->comm, result);
 221         LASSERT(ergo(result == 0, PageLocked(vmpage)));
 222
 223         /* if page has been unmapped, presumably due to lock reclaim for
 224          * concurrent usage, add some delay before retrying to prevent
 225          * entering live-lock situation with competitors
 226          */
 227         if (result == -ENODATA && inode != NULL) {
 228                 CDEBUG(D_MMAP, "delaying new page-fault for inode %p to "
 229                                "prevent live-lock\n", inode);
 230                 msleep(10);
 231         }
 232
 233         return result;
 234 }
 235
 236 static inline int to_fault_error(int result)
 237 {
 238         switch(result) {
 239         case 0:
 240                 result = VM_FAULT_LOCKED;
 241                 break;
 242         case -ENOMEM:
 243                 result = VM_FAULT_OOM;
 244                 break;
 245         default:
 246                 result = VM_FAULT_SIGBUS;
 247                 break;
 248         }
 249         return result;
 250 }
 251
 252 /**
 253  * Lustre implementation of a vm_operations_struct::fault() method, called by
 254  * VM to server page fault (both in kernel and user space).
 255  *
 256  * \param vma - is virtiual area struct related to page fault
 257  * \param vmf - structure which describe type and address where hit fault
 258  *
 259  * \return allocated and filled _locked_ page for address
 260  * \retval VM_FAULT_ERROR on general error
 261  * \retval NOPAGE_OOM not have memory for allocate new page
 262  */
 263 static vm_fault_t ll_fault0(struct vm_area_struct *vma, struct vm_fault *vmf)
 264 {
 265         struct lu_env           *env;
 266         struct cl_io            *io;
 267         struct vvp_io           *vio = NULL;
 268         struct page             *vmpage;
 269         int                      result = 0;
 270         int                      fault_ret = 0;
 271         __u16                    refcheck;
 272         ENTRY;
 273
 274         env = cl_env_get(&refcheck);
 275         if (IS_ERR(env))
 276                 RETURN(PTR_ERR(env));
 277
 278         if (ll_sbi_has_fast_read(ll_i2sbi(file_inode(vma->vm_file)))) {
 279                 /* do fast fault */
 280                 bool has_retry = vmf->flags & FAULT_FLAG_RETRY_NOWAIT;
 281
 282                 /* To avoid loops, instruct downstream to not drop mmap_sem */
 283                 vmf->flags |= FAULT_FLAG_RETRY_NOWAIT;
 284                 ll_cl_add(vma->vm_file, env, NULL, LCC_MMAP);
 285                 fault_ret = ll_filemap_fault(vma, vmf);
 286                 ll_cl_remove(vma->vm_file, env);
 287                 if (!has_retry)
 288                         vmf->flags &= ~FAULT_FLAG_RETRY_NOWAIT;
 289
 290                 /* - If there is no error, then the page was found in cache and
 291                  *   uptodate;
 292                  * - If VM_FAULT_RETRY is set, the page existed but failed to
 293                  *   lock. We will try slow path to avoid loops.
 294                  * - Otherwise, it should try normal fault under DLM lock. */
 295                 if (!(fault_ret & VM_FAULT_RETRY) &&
 296                     !(fault_ret & VM_FAULT_ERROR))
 297                         GOTO(out, result = 0);
 298
 299                 fault_ret = 0;
 300         }
 301
 302         io = ll_fault_io_init(env, vma, vmf->pgoff);
 303         if (IS_ERR(io))
 304                 GOTO(out, result = PTR_ERR(io));
 305
 306         result = io->ci_result;
 307         if (result == 0) {
 308                 struct file *vm_file = vma->vm_file;
 309
 310                 vio = vvp_env_io(env);
 311                 vio->u.fault.ft_vma       = vma;
 312                 vio->u.fault.ft_vmpage    = NULL;
 313                 vio->u.fault.ft_vmf = vmf;
 314                 vio->u.fault.ft_flags = 0;
 315                 vio->u.fault.ft_flags_valid = 0;
 316
 317                 get_file(vm_file);
 318
 319                 /* May call ll_readpage() */
 320                 ll_cl_add(vm_file, env, io, LCC_MMAP);
 321
 322                 result = cl_io_loop(env, io);
 323
 324                 ll_cl_remove(vm_file, env);
 325                 fput(vm_file);
 326                 /* ft_flags are only valid if we reached
 327                  * the call to filemap_fault */
 328                 if (vio->u.fault.ft_flags_valid)
 329                         fault_ret = vio->u.fault.ft_flags;
 330
 331                 vmpage = vio->u.fault.ft_vmpage;
 332                 if (result != 0 && vmpage != NULL) {
 333                         put_page(vmpage);
 334                         vmf->page = NULL;
 335                 }
 336         }
 337         cl_io_fini(env, io);
 338
 339 out:
 340         cl_env_put(env, &refcheck);
 341         if (result != 0 && !(fault_ret & VM_FAULT_RETRY))
 342                 fault_ret |= to_fault_error(result);
 343
 344         CDEBUG(D_MMAP, "%s fault %d/%d\n", current->comm, fault_ret, result);
 345         RETURN(fault_ret);
 346 }
 347
 348 #ifdef HAVE_VM_OPS_USE_VM_FAULT_ONLY
 349 static vm_fault_t ll_fault(struct vm_fault *vmf)
 350 {
 351         struct vm_area_struct *vma = vmf->vma;
 352 #else
 353 static vm_fault_t ll_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
 354 {
 355 #endif
 356         int count = 0;
 357         bool printed = false;
 358         bool cached;
 359         vm_fault_t result;
 360         ktime_t kstart = ktime_get();
 361         sigset_t old, new;
 362
 363         result = pcc_fault(vma, vmf, &cached);
 364         if (cached)
 365                 goto out;
 366
 367         CDEBUG(D_MMAP, DFID": vma=%p start=%#lx end=%#lx vm_flags=%#lx\n",
 368                PFID(&ll_i2info(file_inode(vma->vm_file))->lli_fid),
 369                vma, vma->vm_start, vma->vm_end, vma->vm_flags);
 370
 371         /* Only SIGKILL and SIGTERM is allowed for fault/nopage/mkwrite
 372          * so that it can be killed by admin but not cause segfault by
 373          * other signals.
 374          */
 375         siginitsetinv(&new, sigmask(SIGKILL) | sigmask(SIGTERM));
 376         sigprocmask(SIG_BLOCK, &new, &old);
 377
 378         /* make sure offset is not a negative number */
 379         if (vmf->pgoff > (MAX_LFS_FILESIZE >> PAGE_SHIFT))
 380                 return VM_FAULT_SIGBUS;
 381
 382 restart:
 383         result = ll_fault0(vma, vmf);
 384         if (vmf->page &&
 385             !(result & (VM_FAULT_RETRY | VM_FAULT_ERROR | VM_FAULT_LOCKED))) {
 386                 struct page *vmpage = vmf->page;
 387
 388                 /* check if this page has been truncated */
 389                 lock_page(vmpage);
 390                 if (unlikely(vmpage->mapping == NULL)) { /* unlucky */
 391                         unlock_page(vmpage);
 392                         put_page(vmpage);
 393                         vmf->page = NULL;
 394
 395                         if (!printed && ++count > 16) {
 396                                 CWARN("the page is under heavy contention, maybe your app(%s) needs revising :-)\n",
 397                                       current->comm);
 398                                 printed = true;
 399                         }
 400
 401                         goto restart;
 402                 }
 403
 404                 result |= VM_FAULT_LOCKED;
 405         }
 406         sigprocmask(SIG_SETMASK, &old, NULL);
 407
 408 out:
 409         if (vmf->page && result == VM_FAULT_LOCKED) {
 410                 ll_rw_stats_tally(ll_i2sbi(file_inode(vma->vm_file)),
 411                                   current->pid, vma->vm_file->private_data,
 412                                   cl_offset(NULL, vmf->page->index), PAGE_SIZE,
 413                                   READ);
 414                 ll_stats_ops_tally(ll_i2sbi(file_inode(vma->vm_file)),
 415                                    LPROC_LL_FAULT,
 416                                    ktime_us_delta(ktime_get(), kstart));
 417         }
 418
 419         return result;
 420 }
 421
 422 #ifdef HAVE_VM_OPS_USE_VM_FAULT_ONLY
 423 static vm_fault_t ll_page_mkwrite(struct vm_fault *vmf)
 424 {
 425         struct vm_area_struct *vma = vmf->vma;
 426 #else
 427 static vm_fault_t ll_page_mkwrite(struct vm_area_struct *vma,
 428                                   struct vm_fault *vmf)
 429 {
 430 #endif
 431         int count = 0;
 432         bool printed = false;
 433         bool retry;
 434         bool cached;
 435         ktime_t kstart = ktime_get();
 436         vm_fault_t result;
 437
 438         result = pcc_page_mkwrite(vma, vmf, &cached);
 439         if (cached)
 440                 goto out;
 441
 442         file_update_time(vma->vm_file);
 443         do {
 444                 retry = false;
 445                 result = ll_page_mkwrite0(vma, vmf->page, &retry);
 446
 447                 if (!printed && ++count > 16) {
 448                         const struct dentry *de = file_dentry(vma->vm_file);
 449
 450                         CWARN("app(%s): the page %lu of file "DFID" is under heavy contention\n",
 451                               current->comm, vmf->pgoff,
 452                               PFID(ll_inode2fid(de->d_inode)));
 453                         printed = true;
 454                 }
 455         } while (retry);
 456
 457         switch (result) {
 458         case 0:
 459                 LASSERT(PageLocked(vmf->page));
 460                 result = VM_FAULT_LOCKED;
 461                 break;
 462         case -ENODATA:
 463         case -EFAULT:
 464                 result = VM_FAULT_NOPAGE;
 465                 break;
 466         case -ENOMEM:
 467                 result = VM_FAULT_OOM;
 468                 break;
 469         case -EAGAIN:
 470                 result = VM_FAULT_RETRY;
 471                 break;
 472         default:
 473                 result = VM_FAULT_SIGBUS;
 474                 break;
 475         }
 476
 477 out:
 478         if (result == VM_FAULT_LOCKED) {
 479                 ll_rw_stats_tally(ll_i2sbi(file_inode(vma->vm_file)),
 480                                   current->pid, vma->vm_file->private_data,
 481                                   cl_offset(NULL, vmf->page->index), PAGE_SIZE,
 482                                   WRITE);
 483                 ll_stats_ops_tally(ll_i2sbi(file_inode(vma->vm_file)),
 484                                    LPROC_LL_MKWRITE,
 485                                    ktime_us_delta(ktime_get(), kstart));
 486         }
 487
 488         return result;
 489 }
 490
 491 /**
 492  *  To avoid cancel the locks covering mmapped region for lock cache pressure,
 493  *  we track the mapped vma count in vvp_object::vob_mmap_cnt.
 494  */
 495 static void ll_vm_open(struct vm_area_struct * vma)
 496 {
 497         struct inode *inode    = file_inode(vma->vm_file);
 498         struct vvp_object *vob = cl_inode2vvp(inode);
 499
 500         ENTRY;
 501         LASSERT(atomic_read(&vob->vob_mmap_cnt) >= 0);
 502         atomic_inc(&vob->vob_mmap_cnt);
 503         pcc_vm_open(vma);
 504         EXIT;
 505 }
 506
 507 /**
 508  * Dual to ll_vm_open().
 509  */
 510 static void ll_vm_close(struct vm_area_struct *vma)
 511 {
 512         struct inode      *inode = file_inode(vma->vm_file);
 513         struct vvp_object *vob   = cl_inode2vvp(inode);
 514
 515         ENTRY;
 516         atomic_dec(&vob->vob_mmap_cnt);
 517         LASSERT(atomic_read(&vob->vob_mmap_cnt) >= 0);
 518         pcc_vm_close(vma);
 519         EXIT;
 520 }
 521
 522 static const struct vm_operations_struct ll_file_vm_ops = {
 523         .fault                  = ll_fault,
 524         .page_mkwrite           = ll_page_mkwrite,
 525         .open                   = ll_vm_open,
 526         .close                  = ll_vm_close,
 527 };
 528
 529 int ll_file_mmap(struct file *file, struct vm_area_struct * vma)
 530 {
 531         struct inode *inode = file_inode(file);
 532         ktime_t kstart = ktime_get();
 533         bool cached;
 534         int rc;
 535
 536         ENTRY;
 537         CDEBUG(D_VFSTRACE | D_MMAP,
 538                "VFS_Op: fid="DFID" vma=%p start=%#lx end=%#lx vm_flags=%#lx\n",
 539                PFID(&ll_i2info(inode)->lli_fid),
 540                vma, vma->vm_start, vma->vm_end, vma->vm_flags);
 541
 542         if (ll_file_nolock(file))
 543                 RETURN(-EOPNOTSUPP);
 544
 545         rc = pcc_file_mmap(file, vma, &cached);
 546         if (cached && rc != 0)
 547                 RETURN(rc);
 548
 549         rc = generic_file_mmap(file, vma);
 550         if (rc == 0) {
 551                 vma->vm_ops = &ll_file_vm_ops;
 552                 vma->vm_ops->open(vma);
 553                 /* update the inode's size and mtime */
 554                 if (!cached)
 555                         rc = ll_glimpse_size(inode);
 556         }
 557
 558         if (!rc)
 559                 ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_MMAP,
 560                                    ktime_us_delta(ktime_get(), kstart));
 561
 562         RETURN(rc);
 563 }