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  * \param mkwrite - whether it is mmap write.
  81  *
  82  * \return error codes from cl_io_init.
  83  */
  84 static struct cl_io *
  85 ll_fault_io_init(struct lu_env *env, struct vm_area_struct *vma,
  86                 pgoff_t index, bool mkwrite)
  87 {
  88         struct file            *file = vma->vm_file;
  89         struct inode           *inode = file_inode(file);
  90         struct cl_io           *io;
  91         struct cl_fault_io     *fio;
  92         int                     rc;
  93         ENTRY;
  94
  95         if (ll_file_nolock(file))
  96                 RETURN(ERR_PTR(-EOPNOTSUPP));
  97
  98 restart:
  99         io = vvp_env_thread_io(env);
 100         io->ci_obj = ll_i2info(inode)->lli_clob;
 101         LASSERT(io->ci_obj != NULL);
 102
 103         fio = &io->u.ci_fault;
 104         fio->ft_index = index;
 105         fio->ft_executable = vma->vm_flags & VM_EXEC;
 106
 107         if (mkwrite) {
 108                 fio->ft_mkwrite = 1;
 109                 fio->ft_writable = 1;
 110         }
 111
 112         CDEBUG(D_MMAP,
 113                DFID": vma=%p start=%#lx end=%#lx vm_flags=%#lx idx=%lu\n",
 114                PFID(&ll_i2info(inode)->lli_fid), vma, vma->vm_start,
 115                vma->vm_end, vma->vm_flags, fio->ft_index);
 116
 117         if (vma->vm_flags & VM_SEQ_READ)
 118                 io->ci_seq_read = 1;
 119         else if (vma->vm_flags & VM_RAND_READ)
 120                 io->ci_rand_read = 1;
 121
 122         rc = cl_io_init(env, io, CIT_FAULT, io->ci_obj);
 123         if (rc == 0) {
 124                 struct vvp_io *vio = vvp_env_io(env);
 125                 struct ll_file_data *fd = file->private_data;
 126
 127                 LASSERT(vio->vui_cl.cis_io == io);
 128
 129                 /* mmap lock must be MANDATORY it has to cache
 130                  * pages. */
 131                 io->ci_lockreq = CILR_MANDATORY;
 132                 vio->vui_fd = fd;
 133         } else {
 134                 cl_io_fini(env, io);
 135                 if (io->ci_need_restart)
 136                         goto restart;
 137
 138                 io = ERR_PTR(rc);
 139         }
 140
 141         RETURN(io);
 142 }
 143
 144 /* Sharing code of page_mkwrite method for rhel5 and rhel6 */
 145 static int ll_page_mkwrite0(struct vm_area_struct *vma, struct page *vmpage,
 146                             bool *retry)
 147 {
 148         struct lu_env           *env;
 149         struct cl_io            *io;
 150         struct vvp_io           *vio;
 151         int                      result;
 152         __u16                    refcheck;
 153         sigset_t old, new;
 154         struct inode             *inode = NULL;
 155         struct ll_inode_info     *lli;
 156         ENTRY;
 157
 158         LASSERT(vmpage != NULL);
 159         env = cl_env_get(&refcheck);
 160         if (IS_ERR(env))
 161                 RETURN(PTR_ERR(env));
 162
 163         io = ll_fault_io_init(env, vma, vmpage->index, true);
 164         if (IS_ERR(io))
 165                 GOTO(out, result = PTR_ERR(io));
 166
 167         result = io->ci_result;
 168         if (result < 0)
 169                 GOTO(out_io, result);
 170
 171         vio = vvp_env_io(env);
 172         vio->u.fault.ft_vma    = vma;
 173         vio->u.fault.ft_vmpage = vmpage;
 174
 175         siginitsetinv(&new, sigmask(SIGKILL) | sigmask(SIGTERM));
 176         sigprocmask(SIG_BLOCK, &new, &old);
 177
 178         inode = vvp_object_inode(io->ci_obj);
 179         lli = ll_i2info(inode);
 180
 181         result = cl_io_loop(env, io);
 182
 183         sigprocmask(SIG_SETMASK, &old, NULL);
 184
 185         if (result == 0) {
 186                 lock_page(vmpage);
 187                 if (vmpage->mapping == NULL) {
 188                         unlock_page(vmpage);
 189
 190                         /* page was truncated and lock was cancelled, return
 191                          * ENODATA so that VM_FAULT_NOPAGE will be returned
 192                          * to handle_mm_fault(). */
 193                         if (result == 0)
 194                                 result = -ENODATA;
 195                 } else if (!PageDirty(vmpage)) {
 196                         /* race, the page has been cleaned by ptlrpcd after
 197                          * it was unlocked, it has to be added into dirty
 198                          * cache again otherwise this soon-to-dirty page won't
 199                          * consume any grants, even worse if this page is being
 200                          * transferred because it will break RPC checksum.
 201                          */
 202                         unlock_page(vmpage);
 203
 204                         CDEBUG(D_MMAP, "Race on page_mkwrite %p/%lu, page has "
 205                                "been written out, retry.\n",
 206                                vmpage, vmpage->index);
 207
 208                         *retry = true;
 209                         result = -EAGAIN;
 210                 }
 211
 212                 if (result == 0)
 213                         set_bit(LLIF_DATA_MODIFIED, &lli->lli_flags);
 214         }
 215         EXIT;
 216
 217 out_io:
 218         cl_io_fini(env, io);
 219 out:
 220         cl_env_put(env, &refcheck);
 221         CDEBUG(D_MMAP, "%s mkwrite with %d\n", current->comm, result);
 222         LASSERT(ergo(result == 0, PageLocked(vmpage)));
 223
 224         /* if page has been unmapped, presumably due to lock reclaim for
 225          * concurrent usage, add some delay before retrying to prevent
 226          * entering live-lock situation with competitors
 227          */
 228         if (result == -ENODATA && inode != NULL) {
 229                 CDEBUG(D_MMAP, "delaying new page-fault for inode %p to "
 230                                "prevent live-lock\n", inode);
 231                 msleep(10);
 232         }
 233
 234         return result;
 235 }
 236
 237 static inline int to_fault_error(int result)
 238 {
 239         switch(result) {
 240         case 0:
 241                 result = VM_FAULT_LOCKED;
 242                 break;
 243         case -ENOMEM:
 244                 result = VM_FAULT_OOM;
 245                 break;
 246         default:
 247                 result = VM_FAULT_SIGBUS;
 248                 break;
 249         }
 250         return result;
 251 }
 252
 253 int ll_filemap_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
 254 {
 255         struct inode *inode = file_inode(vma->vm_file);
 256         int ret;
 257         unsigned int seq;
 258
 259         /* this seqlock lets us notice if a page has been deleted on this inode
 260          * during the fault process, allowing us to catch an erroneous SIGBUS
 261          * See LU-16160
 262          */
 263         do {
 264                 seq = read_seqbegin(&ll_i2info(inode)->lli_page_inv_lock);
 265                 ret = __ll_filemap_fault(vma, vmf);
 266         } while (read_seqretry(&ll_i2info(inode)->lli_page_inv_lock, seq) &&
 267                  (ret & VM_FAULT_SIGBUS));
 268
 269         return ret;
 270 }
 271
 272 /**
 273  * Lustre implementation of a vm_operations_struct::fault() method, called by
 274  * VM to server page fault (both in kernel and user space).
 275  *
 276  * \param vma - is virtiual area struct related to page fault
 277  * \param vmf - structure which describe type and address where hit fault
 278  *
 279  * \return allocated and filled _locked_ page for address
 280  * \retval VM_FAULT_ERROR on general error
 281  * \retval NOPAGE_OOM not have memory for allocate new page
 282  */
 283 static vm_fault_t ll_fault0(struct vm_area_struct *vma, struct vm_fault *vmf)
 284 {
 285         struct inode *inode = file_inode(vma->vm_file);
 286         struct lu_env           *env;
 287         struct cl_io            *io;
 288         struct vvp_io           *vio = NULL;
 289         struct page             *vmpage;
 290         int                      result = 0;
 291         int                      fault_ret = 0;
 292         __u16                    refcheck;
 293         ENTRY;
 294
 295         env = cl_env_get(&refcheck);
 296         if (IS_ERR(env))
 297                 RETURN(PTR_ERR(env));
 298
 299         if (ll_sbi_has_fast_read(ll_i2sbi(inode))) {
 300                 /* do fast fault */
 301                 bool allow_retry = vmf->flags & FAULT_FLAG_ALLOW_RETRY;
 302                 bool has_retry = vmf->flags & FAULT_FLAG_RETRY_NOWAIT;
 303
 304                 /* To avoid loops, instruct downstream to not drop mmap_sem */
 305                 /**
 306                  * only need FAULT_FLAG_ALLOW_RETRY prior to Linux 5.1
 307                  * (6b4c9f4469819), where FAULT_FLAG_RETRY_NOWAIT is enough
 308                  * to not drop mmap_sem when failed to lock the page.
 309                  */
 310                 vmf->flags |= FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_RETRY_NOWAIT;
 311                 ll_cl_add(inode, env, NULL, LCC_MMAP);
 312                 fault_ret = ll_filemap_fault(vma, vmf);
 313                 ll_cl_remove(inode, env);
 314                 if (!has_retry)
 315                         vmf->flags &= ~FAULT_FLAG_RETRY_NOWAIT;
 316                 if (!allow_retry)
 317                         vmf->flags &= ~FAULT_FLAG_ALLOW_RETRY;
 318
 319                 /* - If there is no error, then the page was found in cache and
 320                  *   uptodate;
 321                  * - If VM_FAULT_RETRY is set, the page existed but failed to
 322                  *   lock. We will try slow path to avoid loops.
 323                  * - Otherwise, it should try normal fault under DLM lock. */
 324                 if (!(fault_ret & VM_FAULT_RETRY) &&
 325                     !(fault_ret & VM_FAULT_ERROR))
 326                         GOTO(out, result = 0);
 327
 328                 fault_ret = 0;
 329         }
 330
 331         io = ll_fault_io_init(env, vma, vmf->pgoff, false);
 332         if (IS_ERR(io))
 333                 GOTO(out, result = PTR_ERR(io));
 334
 335         result = io->ci_result;
 336         if (result == 0) {
 337                 vio = vvp_env_io(env);
 338                 vio->u.fault.ft_vma       = vma;
 339                 vio->u.fault.ft_vmpage    = NULL;
 340                 vio->u.fault.ft_vmf = vmf;
 341                 vio->u.fault.ft_flags = 0;
 342                 vio->u.fault.ft_flags_valid = 0;
 343
 344                 /* May call ll_readpage() */
 345                 ll_cl_add(inode, env, io, LCC_MMAP);
 346
 347                 result = cl_io_loop(env, io);
 348
 349                 ll_cl_remove(inode, env);
 350
 351                 /* ft_flags are only valid if we reached
 352                  * the call to filemap_fault */
 353                 if (vio->u.fault.ft_flags_valid)
 354                         fault_ret = vio->u.fault.ft_flags;
 355
 356                 vmpage = vio->u.fault.ft_vmpage;
 357                 if (result != 0 && vmpage != NULL) {
 358                         put_page(vmpage);
 359                         vmf->page = NULL;
 360                 }
 361         }
 362         cl_io_fini(env, io);
 363
 364 out:
 365         cl_env_put(env, &refcheck);
 366         if (result != 0 && !(fault_ret & VM_FAULT_RETRY))
 367                 fault_ret |= to_fault_error(result);
 368
 369         CDEBUG(D_MMAP, "%s fault %d/%d\n", current->comm, fault_ret, result);
 370         RETURN(fault_ret);
 371 }
 372
 373 #ifdef HAVE_VM_OPS_USE_VM_FAULT_ONLY
 374 static vm_fault_t ll_fault(struct vm_fault *vmf)
 375 {
 376         struct vm_area_struct *vma = vmf->vma;
 377 #else
 378 static vm_fault_t ll_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
 379 {
 380 #endif
 381         int count = 0;
 382         bool printed = false;
 383         bool cached;
 384         vm_fault_t result;
 385         ktime_t kstart = ktime_get();
 386         sigset_t old, new;
 387
 388         result = pcc_fault(vma, vmf, &cached);
 389         if (cached)
 390                 goto out;
 391
 392         CDEBUG(D_MMAP|D_IOTRACE,
 393                DFID": vma=%p start=%#lx end=%#lx vm_flags=%#lx idx=%lu\n",
 394                PFID(&ll_i2info(file_inode(vma->vm_file))->lli_fid),
 395                vma, vma->vm_start, vma->vm_end, vma->vm_flags, vmf->pgoff);
 396
 397         /* Only SIGKILL and SIGTERM is allowed for fault/nopage/mkwrite
 398          * so that it can be killed by admin but not cause segfault by
 399          * other signals.
 400          */
 401         siginitsetinv(&new, sigmask(SIGKILL) | sigmask(SIGTERM));
 402         sigprocmask(SIG_BLOCK, &new, &old);
 403
 404         /* make sure offset is not a negative number */
 405         if (vmf->pgoff > (MAX_LFS_FILESIZE >> PAGE_SHIFT))
 406                 return VM_FAULT_SIGBUS;
 407
 408 restart:
 409         result = ll_fault0(vma, vmf);
 410         if (vmf->page &&
 411             !(result & (VM_FAULT_RETRY | VM_FAULT_ERROR | VM_FAULT_LOCKED))) {
 412                 struct page *vmpage = vmf->page;
 413
 414                 /* check if this page has been truncated */
 415                 lock_page(vmpage);
 416                 if (unlikely(vmpage->mapping == NULL)) { /* unlucky */
 417                         unlock_page(vmpage);
 418                         put_page(vmpage);
 419                         vmf->page = NULL;
 420
 421                         if (!printed && ++count > 16) {
 422                                 CWARN("the page is under heavy contention, maybe your app(%s) needs revising :-)\n",
 423                                       current->comm);
 424                                 printed = true;
 425                         }
 426
 427                         goto restart;
 428                 }
 429
 430                 result |= VM_FAULT_LOCKED;
 431         }
 432         sigprocmask(SIG_SETMASK, &old, NULL);
 433
 434 out:
 435         if (vmf->page && result == VM_FAULT_LOCKED) {
 436                 ll_rw_stats_tally(ll_i2sbi(file_inode(vma->vm_file)),
 437                                   current->pid, vma->vm_file->private_data,
 438                                   cl_offset(NULL, vmf->page->index), PAGE_SIZE,
 439                                   READ);
 440                 ll_stats_ops_tally(ll_i2sbi(file_inode(vma->vm_file)),
 441                                    LPROC_LL_FAULT,
 442                                    ktime_us_delta(ktime_get(), kstart));
 443         }
 444
 445         CDEBUG(D_IOTRACE,
 446                "COMPLETED: "DFID": vma=%p start=%#lx end=%#lx vm_flags=%#lx idx=%lu\n",
 447                PFID(&ll_i2info(file_inode(vma->vm_file))->lli_fid),
 448                vma, vma->vm_start, vma->vm_end, vma->vm_flags, vmf->pgoff);
 449
 450         return result;
 451 }
 452
 453 #ifdef HAVE_VM_OPS_USE_VM_FAULT_ONLY
 454 static vm_fault_t ll_page_mkwrite(struct vm_fault *vmf)
 455 {
 456         struct vm_area_struct *vma = vmf->vma;
 457 #else
 458 static vm_fault_t ll_page_mkwrite(struct vm_area_struct *vma,
 459                                   struct vm_fault *vmf)
 460 {
 461 #endif
 462         int count = 0;
 463         bool printed = false;
 464         bool retry;
 465         bool cached;
 466         ktime_t kstart = ktime_get();
 467         vm_fault_t result;
 468
 469         CDEBUG(D_MMAP|D_IOTRACE,
 470                DFID": vma=%p start=%#lx end=%#lx vm_flags=%#lx idx=%lu\n",
 471                PFID(&ll_i2info(file_inode(vma->vm_file))->lli_fid),
 472                vma, vma->vm_start, vma->vm_end, vma->vm_flags,
 473                vmf->page->index);
 474
 475         result = pcc_page_mkwrite(vma, vmf, &cached);
 476         if (cached)
 477                 goto out;
 478
 479         file_update_time(vma->vm_file);
 480         do {
 481                 retry = false;
 482                 result = ll_page_mkwrite0(vma, vmf->page, &retry);
 483
 484                 if (!printed && ++count > 16) {
 485                         const struct dentry *de = file_dentry(vma->vm_file);
 486
 487                         CWARN("app(%s): the page %lu of file "DFID" is under heavy contention\n",
 488                               current->comm, vmf->pgoff,
 489                               PFID(ll_inode2fid(de->d_inode)));
 490                         printed = true;
 491                 }
 492         } while (retry);
 493
 494         switch (result) {
 495         case 0:
 496                 LASSERT(PageLocked(vmf->page));
 497                 result = VM_FAULT_LOCKED;
 498                 break;
 499         case -ENODATA:
 500         case -EFAULT:
 501                 result = VM_FAULT_NOPAGE;
 502                 break;
 503         case -ENOMEM:
 504                 result = VM_FAULT_OOM;
 505                 break;
 506         case -EAGAIN:
 507                 result = VM_FAULT_RETRY;
 508                 break;
 509         default:
 510                 result = VM_FAULT_SIGBUS;
 511                 break;
 512         }
 513
 514 out:
 515         if (result == VM_FAULT_LOCKED) {
 516                 ll_rw_stats_tally(ll_i2sbi(file_inode(vma->vm_file)),
 517                                   current->pid, vma->vm_file->private_data,
 518                                   cl_offset(NULL, vmf->page->index), PAGE_SIZE,
 519                                   WRITE);
 520                 ll_stats_ops_tally(ll_i2sbi(file_inode(vma->vm_file)),
 521                                    LPROC_LL_MKWRITE,
 522                                    ktime_us_delta(ktime_get(), kstart));
 523         }
 524
 525         CDEBUG(D_IOTRACE,
 526                "COMPLETED: "DFID": vma=%p start=%#lx end=%#lx vm_flags=%#lx idx=%lu\n",
 527                PFID(&ll_i2info(file_inode(vma->vm_file))->lli_fid),
 528                vma, vma->vm_start, vma->vm_end, vma->vm_flags,
 529                vmf->page->index);
 530         return result;
 531 }
 532
 533 /**
 534  *  To avoid cancel the locks covering mmapped region for lock cache pressure,
 535  *  we track the mapped vma count in vvp_object::vob_mmap_cnt.
 536  */
 537 static void ll_vm_open(struct vm_area_struct * vma)
 538 {
 539         struct inode *inode    = file_inode(vma->vm_file);
 540         struct vvp_object *vob = cl_inode2vvp(inode);
 541
 542         ENTRY;
 543         LASSERT(atomic_read(&vob->vob_mmap_cnt) >= 0);
 544         atomic_inc(&vob->vob_mmap_cnt);
 545         pcc_vm_open(vma);
 546         EXIT;
 547 }
 548
 549 /**
 550  * Dual to ll_vm_open().
 551  */
 552 static void ll_vm_close(struct vm_area_struct *vma)
 553 {
 554         struct inode      *inode = file_inode(vma->vm_file);
 555         struct vvp_object *vob   = cl_inode2vvp(inode);
 556
 557         ENTRY;
 558         atomic_dec(&vob->vob_mmap_cnt);
 559         LASSERT(atomic_read(&vob->vob_mmap_cnt) >= 0);
 560         pcc_vm_close(vma);
 561         EXIT;
 562 }
 563
 564 static const struct vm_operations_struct ll_file_vm_ops = {
 565         .fault                  = ll_fault,
 566         .page_mkwrite           = ll_page_mkwrite,
 567         .open                   = ll_vm_open,
 568         .close                  = ll_vm_close,
 569 };
 570
 571 int ll_file_mmap(struct file *file, struct vm_area_struct * vma)
 572 {
 573         struct inode *inode = file_inode(file);
 574         ktime_t kstart = ktime_get();
 575         bool cached;
 576         int rc;
 577
 578         ENTRY;
 579         CDEBUG(D_VFSTRACE | D_MMAP,
 580                "VFS_Op: fid="DFID" vma=%p start=%#lx end=%#lx vm_flags=%#lx\n",
 581                PFID(&ll_i2info(inode)->lli_fid),
 582                vma, vma->vm_start, vma->vm_end, vma->vm_flags);
 583
 584         if (ll_file_nolock(file))
 585                 RETURN(-EOPNOTSUPP);
 586
 587         rc = pcc_file_mmap(file, vma, &cached);
 588         if (cached && rc != 0)
 589                 RETURN(rc);
 590
 591         rc = generic_file_mmap(file, vma);
 592         if (rc == 0) {
 593                 vma->vm_ops = &ll_file_vm_ops;
 594                 vma->vm_ops->open(vma);
 595                 /* update the inode's size and mtime */
 596                 if (!cached)
 597                         rc = ll_glimpse_size(inode);
 598         }
 599
 600         if (!rc)
 601                 ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_MMAP,
 602                                    ktime_us_delta(ktime_get(), kstart));
 603
 604         RETURN(rc);
 605 }