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  * Lustre is a trademark of Sun Microsystems, Inc.
  31  */
  32
  33 #include <linux/errno.h>
  34 #include <linux/delay.h>
  35 #include <linux/kernel.h>
  36 #include <linux/mm.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_sem must have been held by caller. */
  61         LASSERT(!down_write_trylock(&mm->mmap_sem));
  62
  63         for(vma = find_vma(mm, addr);
  64             vma != NULL && vma->vm_start < (addr + count); vma = vma->vm_next) {
  65                 if (vma->vm_ops && vma->vm_ops == &ll_file_vm_ops &&
  66                     vma->vm_flags & VM_SHARED) {
  67                         ret = vma;
  68                         break;
  69                 }
  70         }
  71         RETURN(ret);
  72 }
  73
  74 /**
  75  * API independent part for page fault initialization.
  76  * \param env - corespondent lu_env to processing
  77  * \param vma - virtual memory area addressed to page fault
  78  * \param index - page index corespondent to fault.
  79  * \parm ra_flags - vma readahead flags.
  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,
  85                  pgoff_t index, unsigned long *ra_flags)
  86 {
  87         struct file            *file = vma->vm_file;
  88         struct inode           *inode = file_inode(file);
  89         struct cl_io           *io;
  90         struct cl_fault_io     *fio;
  91         int                     rc;
  92         ENTRY;
  93
  94         if (ll_file_nolock(file))
  95                 RETURN(ERR_PTR(-EOPNOTSUPP));
  96
  97 restart:
  98         io = vvp_env_thread_io(env);
  99         io->ci_obj = ll_i2info(inode)->lli_clob;
 100         LASSERT(io->ci_obj != NULL);
 101
 102         fio = &io->u.ci_fault;
 103         fio->ft_index      = index;
 104         fio->ft_executable = vma->vm_flags&VM_EXEC;
 105
 106         /*
 107          * disable VM_SEQ_READ and use VM_RAND_READ to make sure that
 108          * the kernel will not read other pages not covered by ldlm in
 109          * filemap_nopage. we do our readahead in ll_readpage.
 110          */
 111         if (ra_flags != NULL)
 112                 *ra_flags = vma->vm_flags & (VM_RAND_READ|VM_SEQ_READ);
 113         vma->vm_flags &= ~VM_SEQ_READ;
 114         vma->vm_flags |= VM_RAND_READ;
 115
 116         CDEBUG(D_MMAP, "vm_flags: %lx (%lu %d)\n", vma->vm_flags,
 117                fio->ft_index, fio->ft_executable);
 118
 119         rc = cl_io_init(env, io, CIT_FAULT, io->ci_obj);
 120         if (rc == 0) {
 121                 struct vvp_io *vio = vvp_env_io(env);
 122                 struct ll_file_data *fd = file->private_data;
 123
 124                 LASSERT(vio->vui_cl.cis_io == io);
 125
 126                 /* mmap lock must be MANDATORY it has to cache
 127                  * pages. */
 128                 io->ci_lockreq = CILR_MANDATORY;
 129                 vio->vui_fd = fd;
 130         } else {
 131                 LASSERT(rc < 0);
 132                 cl_io_fini(env, io);
 133                 if (io->ci_need_restart)
 134                         goto restart;
 135
 136                 io = ERR_PTR(rc);
 137         }
 138
 139         RETURN(io);
 140 }
 141
 142 /* Sharing code of page_mkwrite method for rhel5 and rhel6 */
 143 static int ll_page_mkwrite0(struct vm_area_struct *vma, struct page *vmpage,
 144                             bool *retry)
 145 {
 146         struct lu_env           *env;
 147         struct cl_io            *io;
 148         struct vvp_io           *vio;
 149         int                      result;
 150         __u16                    refcheck;
 151         sigset_t old, new;
 152         struct inode             *inode = NULL;
 153         struct ll_inode_info     *lli;
 154         ENTRY;
 155
 156         LASSERT(vmpage != NULL);
 157         env = cl_env_get(&refcheck);
 158         if (IS_ERR(env))
 159                 RETURN(PTR_ERR(env));
 160
 161         io = ll_fault_io_init(env, vma, vmpage->index, NULL);
 162         if (IS_ERR(io))
 163                 GOTO(out, result = PTR_ERR(io));
 164
 165         result = io->ci_result;
 166         if (result < 0)
 167                 GOTO(out_io, result);
 168
 169         io->u.ci_fault.ft_mkwrite = 1;
 170         io->u.ci_fault.ft_writable = 1;
 171
 172         vio = vvp_env_io(env);
 173         vio->u.fault.ft_vma    = vma;
 174         vio->u.fault.ft_vmpage = vmpage;
 175
 176         siginitsetinv(&new, sigmask(SIGKILL) | sigmask(SIGTERM));
 177         sigprocmask(SIG_BLOCK, &new, &old);
 178
 179         inode = vvp_object_inode(io->ci_obj);
 180         lli = ll_i2info(inode);
 181
 182         result = cl_io_loop(env, io);
 183
 184         sigprocmask(SIG_SETMASK, &old, NULL);
 185
 186         if (result == 0) {
 187                 lock_page(vmpage);
 188                 if (vmpage->mapping == NULL) {
 189                         unlock_page(vmpage);
 190
 191                         /* page was truncated and lock was cancelled, return
 192                          * ENODATA so that VM_FAULT_NOPAGE will be returned
 193                          * to handle_mm_fault(). */
 194                         if (result == 0)
 195                                 result = -ENODATA;
 196                 } else if (!PageDirty(vmpage)) {
 197                         /* race, the page has been cleaned by ptlrpcd after
 198                          * it was unlocked, it has to be added into dirty
 199                          * cache again otherwise this soon-to-dirty page won't
 200                          * consume any grants, even worse if this page is being
 201                          * transferred because it will break RPC checksum.
 202                          */
 203                         unlock_page(vmpage);
 204
 205                         CDEBUG(D_MMAP, "Race on page_mkwrite %p/%lu, page has "
 206                                "been written out, retry.\n",
 207                                vmpage, vmpage->index);
 208
 209                         *retry = true;
 210                         result = -EAGAIN;
 211                 }
 212
 213                 if (result == 0)
 214                         ll_file_set_flag(lli, LLIF_DATA_MODIFIED);
 215         }
 216         EXIT;
 217
 218 out_io:
 219         cl_io_fini(env, io);
 220 out:
 221         cl_env_put(env, &refcheck);
 222         CDEBUG(D_MMAP, "%s mkwrite with %d\n", current->comm, result);
 223         LASSERT(ergo(result == 0, PageLocked(vmpage)));
 224
 225         /* if page has been unmapped, presumably due to lock reclaim for
 226          * concurrent usage, add some delay before retrying to prevent
 227          * entering live-lock situation with competitors
 228          */
 229         if (result == -ENODATA && inode != NULL) {
 230                 CDEBUG(D_MMAP, "delaying new page-fault for inode %p to "
 231                                "prevent live-lock\n", inode);
 232                 msleep(10);
 233         }
 234
 235         return result;
 236 }
 237
 238 static inline int to_fault_error(int result)
 239 {
 240         switch(result) {
 241         case 0:
 242                 result = VM_FAULT_LOCKED;
 243                 break;
 244         case -ENOMEM:
 245                 result = VM_FAULT_OOM;
 246                 break;
 247         default:
 248                 result = VM_FAULT_SIGBUS;
 249                 break;
 250         }
 251         return result;
 252 }
 253
 254 /**
 255  * Lustre implementation of a vm_operations_struct::fault() method, called by
 256  * VM to server page fault (both in kernel and user space).
 257  *
 258  * \param vma - is virtiual area struct related to page fault
 259  * \param vmf - structure which describe type and address where hit fault
 260  *
 261  * \return allocated and filled _locked_ page for address
 262  * \retval VM_FAULT_ERROR on general error
 263  * \retval NOPAGE_OOM not have memory for allocate new page
 264  */
 265 static vm_fault_t ll_fault0(struct vm_area_struct *vma, struct vm_fault *vmf)
 266 {
 267         struct lu_env           *env;
 268         struct cl_io            *io;
 269         struct vvp_io           *vio = NULL;
 270         struct page             *vmpage;
 271         unsigned long            ra_flags;
 272         int                      result = 0;
 273         int                      fault_ret = 0;
 274         __u16                    refcheck;
 275         ENTRY;
 276
 277         env = cl_env_get(&refcheck);
 278         if (IS_ERR(env))
 279                 RETURN(PTR_ERR(env));
 280
 281         if (ll_sbi_has_fast_read(ll_i2sbi(file_inode(vma->vm_file)))) {
 282                 /* do fast fault */
 283                 bool has_retry = vmf->flags & FAULT_FLAG_RETRY_NOWAIT;
 284
 285                 /* To avoid loops, instruct downstream to not drop mmap_sem */
 286                 vmf->flags |= FAULT_FLAG_RETRY_NOWAIT;
 287                 ll_cl_add(vma->vm_file, env, NULL, LCC_MMAP);
 288                 fault_ret = ll_filemap_fault(vma, vmf);
 289                 ll_cl_remove(vma->vm_file, env);
 290                 if (!has_retry)
 291                         vmf->flags &= ~FAULT_FLAG_RETRY_NOWAIT;
 292
 293                 /* - If there is no error, then the page was found in cache and
 294                  *   uptodate;
 295                  * - If VM_FAULT_RETRY is set, the page existed but failed to
 296                  *   lock. We will try slow path to avoid loops.
 297                  * - Otherwise, it should try normal fault under DLM lock. */
 298                 if (!(fault_ret & VM_FAULT_RETRY) &&
 299                     !(fault_ret & VM_FAULT_ERROR))
 300                         GOTO(out, result = 0);
 301
 302                 fault_ret = 0;
 303         }
 304
 305         io = ll_fault_io_init(env, vma, vmf->pgoff, &ra_flags);
 306         if (IS_ERR(io))
 307                 GOTO(out, result = PTR_ERR(io));
 308
 309         result = io->ci_result;
 310         if (result == 0) {
 311                 vio = vvp_env_io(env);
 312                 vio->u.fault.ft_vma       = vma;
 313                 vio->u.fault.ft_vmpage    = NULL;
 314                 vio->u.fault.ft_vmf = vmf;
 315                 vio->u.fault.ft_flags = 0;
 316                 vio->u.fault.ft_flags_valid = 0;
 317
 318                 /* May call ll_readpage() */
 319                 ll_cl_add(vma->vm_file, env, io, LCC_MMAP);
 320
 321                 result = cl_io_loop(env, io);
 322
 323                 ll_cl_remove(vma->vm_file, env);
 324
 325                 /* ft_flags are only valid if we reached
 326                  * the call to filemap_fault */
 327                 if (vio->u.fault.ft_flags_valid)
 328                         fault_ret = vio->u.fault.ft_flags;
 329
 330                 vmpage = vio->u.fault.ft_vmpage;
 331                 if (result != 0 && vmpage != NULL) {
 332                         put_page(vmpage);
 333                         vmf->page = NULL;
 334                 }
 335         }
 336         cl_io_fini(env, io);
 337
 338         vma->vm_flags |= ra_flags;
 339
 340 out:
 341         cl_env_put(env, &refcheck);
 342         if (result != 0 && !(fault_ret & VM_FAULT_RETRY))
 343                 fault_ret |= to_fault_error(result);
 344
 345         CDEBUG(D_MMAP, "%s fault %d/%d\n", current->comm, fault_ret, result);
 346         RETURN(fault_ret);
 347 }
 348
 349 #ifdef HAVE_VM_OPS_USE_VM_FAULT_ONLY
 350 static vm_fault_t ll_fault(struct vm_fault *vmf)
 351 {
 352         struct vm_area_struct *vma = vmf->vma;
 353 #else
 354 static vm_fault_t ll_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
 355 {
 356 #endif
 357         int count = 0;
 358         bool printed = false;
 359         bool cached;
 360         vm_fault_t result;
 361         ktime_t kstart = ktime_get();
 362         sigset_t old, new;
 363
 364         result = pcc_fault(vma, vmf, &cached);
 365         if (cached)
 366                 goto out;
 367
 368         /* Only SIGKILL and SIGTERM is allowed for fault/nopage/mkwrite
 369          * so that it can be killed by admin but not cause segfault by
 370          * other signals.
 371          */
 372         siginitsetinv(&new, sigmask(SIGKILL) | sigmask(SIGTERM));
 373         sigprocmask(SIG_BLOCK, &new, &old);
 374
 375         /* make sure offset is not a negative number */
 376         if (vmf->pgoff > (MAX_LFS_FILESIZE >> PAGE_SHIFT))
 377                 return VM_FAULT_SIGBUS;
 378 restart:
 379         result = ll_fault0(vma, vmf);
 380         if (vmf->page &&
 381             !(result & (VM_FAULT_RETRY | VM_FAULT_ERROR | VM_FAULT_LOCKED))) {
 382                 struct page *vmpage = vmf->page;
 383
 384                 /* check if this page has been truncated */
 385                 lock_page(vmpage);
 386                 if (unlikely(vmpage->mapping == NULL)) { /* unlucky */
 387                         unlock_page(vmpage);
 388                         put_page(vmpage);
 389                         vmf->page = NULL;
 390
 391                         if (!printed && ++count > 16) {
 392                                 CWARN("the page is under heavy contention, maybe your app(%s) needs revising :-)\n",
 393                                       current->comm);
 394                                 printed = true;
 395                         }
 396
 397                         goto restart;
 398                 }
 399
 400                 result |= VM_FAULT_LOCKED;
 401         }
 402         sigprocmask(SIG_SETMASK, &old, NULL);
 403
 404 out:
 405         if (vmf->page && result == VM_FAULT_LOCKED) {
 406                 ll_rw_stats_tally(ll_i2sbi(file_inode(vma->vm_file)),
 407                                   current->pid, vma->vm_file->private_data,
 408                                   cl_offset(NULL, vmf->page->index), PAGE_SIZE,
 409                                   READ);
 410                 ll_stats_ops_tally(ll_i2sbi(file_inode(vma->vm_file)),
 411                                    LPROC_LL_FAULT,
 412                                    ktime_us_delta(ktime_get(), kstart));
 413         }
 414
 415         return result;
 416 }
 417
 418 #ifdef HAVE_VM_OPS_USE_VM_FAULT_ONLY
 419 static vm_fault_t ll_page_mkwrite(struct vm_fault *vmf)
 420 {
 421         struct vm_area_struct *vma = vmf->vma;
 422 #else
 423 static vm_fault_t ll_page_mkwrite(struct vm_area_struct *vma,
 424                                   struct vm_fault *vmf)
 425 {
 426 #endif
 427         int count = 0;
 428         bool printed = false;
 429         bool retry;
 430         bool cached;
 431         ktime_t kstart = ktime_get();
 432         vm_fault_t result;
 433
 434         result = pcc_page_mkwrite(vma, vmf, &cached);
 435         if (cached)
 436                 goto out;
 437
 438         file_update_time(vma->vm_file);
 439         do {
 440                 retry = false;
 441                 result = ll_page_mkwrite0(vma, vmf->page, &retry);
 442
 443                 if (!printed && ++count > 16) {
 444                         const struct dentry *de = file_dentry(vma->vm_file);
 445
 446                         CWARN("app(%s): the page %lu of file "DFID" is under heavy contention\n",
 447                               current->comm, vmf->pgoff,
 448                               PFID(ll_inode2fid(de->d_inode)));
 449                         printed = true;
 450                 }
 451         } while (retry);
 452
 453         switch (result) {
 454         case 0:
 455                 LASSERT(PageLocked(vmf->page));
 456                 result = VM_FAULT_LOCKED;
 457                 break;
 458         case -ENODATA:
 459         case -EFAULT:
 460                 result = VM_FAULT_NOPAGE;
 461                 break;
 462         case -ENOMEM:
 463                 result = VM_FAULT_OOM;
 464                 break;
 465         case -EAGAIN:
 466                 result = VM_FAULT_RETRY;
 467                 break;
 468         default:
 469                 result = VM_FAULT_SIGBUS;
 470                 break;
 471         }
 472
 473 out:
 474         if (result == VM_FAULT_LOCKED) {
 475                 ll_rw_stats_tally(ll_i2sbi(file_inode(vma->vm_file)),
 476                                   current->pid, vma->vm_file->private_data,
 477                                   cl_offset(NULL, vmf->page->index), PAGE_SIZE,
 478                                   WRITE);
 479                 ll_stats_ops_tally(ll_i2sbi(file_inode(vma->vm_file)),
 480                                    LPROC_LL_MKWRITE,
 481                                    ktime_us_delta(ktime_get(), kstart));
 482         }
 483
 484         return result;
 485 }
 486
 487 /**
 488  *  To avoid cancel the locks covering mmapped region for lock cache pressure,
 489  *  we track the mapped vma count in vvp_object::vob_mmap_cnt.
 490  */
 491 static void ll_vm_open(struct vm_area_struct * vma)
 492 {
 493         struct inode *inode    = file_inode(vma->vm_file);
 494         struct vvp_object *vob = cl_inode2vvp(inode);
 495
 496         ENTRY;
 497         LASSERT(atomic_read(&vob->vob_mmap_cnt) >= 0);
 498         atomic_inc(&vob->vob_mmap_cnt);
 499         pcc_vm_open(vma);
 500         EXIT;
 501 }
 502
 503 /**
 504  * Dual to ll_vm_open().
 505  */
 506 static void ll_vm_close(struct vm_area_struct *vma)
 507 {
 508         struct inode      *inode = file_inode(vma->vm_file);
 509         struct vvp_object *vob   = cl_inode2vvp(inode);
 510
 511         ENTRY;
 512         atomic_dec(&vob->vob_mmap_cnt);
 513         LASSERT(atomic_read(&vob->vob_mmap_cnt) >= 0);
 514         pcc_vm_close(vma);
 515         EXIT;
 516 }
 517
 518 /* XXX put nice comment here.  talk about __free_pte -> dirty pages and
 519  * nopage's reference passing to the pte */
 520 int ll_teardown_mmaps(struct address_space *mapping, __u64 first, __u64 last)
 521 {
 522         int rc = -ENOENT;
 523         ENTRY;
 524
 525         LASSERTF(last > first, "last %llu first %llu\n", last, first);
 526         if (mapping_mapped(mapping)) {
 527                 rc = 0;
 528                 unmap_mapping_range(mapping, first + PAGE_SIZE - 1,
 529                                     last - first + 1, 0);
 530         }
 531
 532         RETURN(rc);
 533 }
 534
 535 static const struct vm_operations_struct ll_file_vm_ops = {
 536         .fault                  = ll_fault,
 537         .page_mkwrite           = ll_page_mkwrite,
 538         .open                   = ll_vm_open,
 539         .close                  = ll_vm_close,
 540 };
 541
 542 int ll_file_mmap(struct file *file, struct vm_area_struct * vma)
 543 {
 544         struct inode *inode = file_inode(file);
 545         ktime_t kstart = ktime_get();
 546         bool cached;
 547         int rc;
 548
 549         ENTRY;
 550
 551         if (ll_file_nolock(file))
 552                 RETURN(-EOPNOTSUPP);
 553
 554         rc = pcc_file_mmap(file, vma, &cached);
 555         if (cached && rc != 0)
 556                 RETURN(rc);
 557
 558         rc = generic_file_mmap(file, vma);
 559         if (rc == 0) {
 560                 vma->vm_ops = &ll_file_vm_ops;
 561                 vma->vm_ops->open(vma);
 562                 /* update the inode's size and mtime */
 563                 if (!cached)
 564                         rc = ll_glimpse_size(inode);
 565         }
 566
 567         if (!rc)
 568                 ll_stats_ops_tally(ll_i2sbi(inode), LPROC_LL_MMAP,
 569                                    ktime_us_delta(ktime_get(), kstart));
 570
 571         RETURN(rc);
 572 }