LCOV - code coverage report
Current view: top level - src/backend/storage/smgr - md.c (source / functions) Hit Total Coverage
Test: PostgreSQL 17beta1 Lines: 342 449 76.2 %
Date: 2024-05-29 10:10:37 Functions: 33 34 97.1 %
Legend: Lines: hit not hit

          Line data    Source code
       1             : /*-------------------------------------------------------------------------
       2             :  *
       3             :  * md.c
       4             :  *    This code manages relations that reside on magnetic disk.
       5             :  *
       6             :  * Or at least, that was what the Berkeley folk had in mind when they named
       7             :  * this file.  In reality, what this code provides is an interface from
       8             :  * the smgr API to Unix-like filesystem APIs, so it will work with any type
       9             :  * of device for which the operating system provides filesystem support.
      10             :  * It doesn't matter whether the bits are on spinning rust or some other
      11             :  * storage technology.
      12             :  *
      13             :  * Portions Copyright (c) 1996-2024, PostgreSQL Global Development Group
      14             :  * Portions Copyright (c) 1994, Regents of the University of California
      15             :  *
      16             :  *
      17             :  * IDENTIFICATION
      18             :  *    src/backend/storage/smgr/md.c
      19             :  *
      20             :  *-------------------------------------------------------------------------
      21             :  */
      22             : #include "postgres.h"
      23             : 
      24             : #include <unistd.h>
      25             : #include <fcntl.h>
      26             : #include <sys/file.h>
      27             : 
      28             : #include "access/xlogutils.h"
      29             : #include "commands/tablespace.h"
      30             : #include "common/file_utils.h"
      31             : #include "miscadmin.h"
      32             : #include "pg_trace.h"
      33             : #include "pgstat.h"
      34             : #include "storage/bufmgr.h"
      35             : #include "storage/fd.h"
      36             : #include "storage/md.h"
      37             : #include "storage/relfilelocator.h"
      38             : #include "storage/smgr.h"
      39             : #include "storage/sync.h"
      40             : #include "utils/memutils.h"
      41             : 
      42             : /*
      43             :  * The magnetic disk storage manager keeps track of open file
      44             :  * descriptors in its own descriptor pool.  This is done to make it
      45             :  * easier to support relations that are larger than the operating
      46             :  * system's file size limit (often 2GBytes).  In order to do that,
      47             :  * we break relations up into "segment" files that are each shorter than
      48             :  * the OS file size limit.  The segment size is set by the RELSEG_SIZE
      49             :  * configuration constant in pg_config.h.
      50             :  *
      51             :  * On disk, a relation must consist of consecutively numbered segment
      52             :  * files in the pattern
      53             :  *  -- Zero or more full segments of exactly RELSEG_SIZE blocks each
      54             :  *  -- Exactly one partial segment of size 0 <= size < RELSEG_SIZE blocks
      55             :  *  -- Optionally, any number of inactive segments of size 0 blocks.
      56             :  * The full and partial segments are collectively the "active" segments.
      57             :  * Inactive segments are those that once contained data but are currently
      58             :  * not needed because of an mdtruncate() operation.  The reason for leaving
      59             :  * them present at size zero, rather than unlinking them, is that other
      60             :  * backends and/or the checkpointer might be holding open file references to
      61             :  * such segments.  If the relation expands again after mdtruncate(), such
      62             :  * that a deactivated segment becomes active again, it is important that
      63             :  * such file references still be valid --- else data might get written
      64             :  * out to an unlinked old copy of a segment file that will eventually
      65             :  * disappear.
      66             :  *
      67             :  * File descriptors are stored in the per-fork md_seg_fds arrays inside
      68             :  * SMgrRelation. The length of these arrays is stored in md_num_open_segs.
      69             :  * Note that a fork's md_num_open_segs having a specific value does not
      70             :  * necessarily mean the relation doesn't have additional segments; we may
      71             :  * just not have opened the next segment yet.  (We could not have "all
      72             :  * segments are in the array" as an invariant anyway, since another backend
      73             :  * could extend the relation while we aren't looking.)  We do not have
      74             :  * entries for inactive segments, however; as soon as we find a partial
      75             :  * segment, we assume that any subsequent segments are inactive.
      76             :  *
      77             :  * The entire MdfdVec array is palloc'd in the MdCxt memory context.
      78             :  */
      79             : 
      80             : typedef struct _MdfdVec
      81             : {
      82             :     File        mdfd_vfd;       /* fd number in fd.c's pool */
      83             :     BlockNumber mdfd_segno;     /* segment number, from 0 */
      84             : } MdfdVec;
      85             : 
      86             : static MemoryContext MdCxt;     /* context for all MdfdVec objects */
      87             : 
      88             : 
      89             : /* Populate a file tag describing an md.c segment file. */
      90             : #define INIT_MD_FILETAG(a,xx_rlocator,xx_forknum,xx_segno) \
      91             : ( \
      92             :     memset(&(a), 0, sizeof(FileTag)), \
      93             :     (a).handler = SYNC_HANDLER_MD, \
      94             :     (a).rlocator = (xx_rlocator), \
      95             :     (a).forknum = (xx_forknum), \
      96             :     (a).segno = (xx_segno) \
      97             : )
      98             : 
      99             : 
     100             : /*** behavior for mdopen & _mdfd_getseg ***/
     101             : /* ereport if segment not present */
     102             : #define EXTENSION_FAIL              (1 << 0)
     103             : /* return NULL if segment not present */
     104             : #define EXTENSION_RETURN_NULL       (1 << 1)
     105             : /* create new segments as needed */
     106             : #define EXTENSION_CREATE            (1 << 2)
     107             : /* create new segments if needed during recovery */
     108             : #define EXTENSION_CREATE_RECOVERY   (1 << 3)
     109             : /*
     110             :  * Allow opening segments which are preceded by segments smaller than
     111             :  * RELSEG_SIZE, e.g. inactive segments (see above). Note that this breaks
     112             :  * mdnblocks() and related functionality henceforth - which currently is ok,
     113             :  * because this is only required in the checkpointer which never uses
     114             :  * mdnblocks().
     115             :  */
     116             : #define EXTENSION_DONT_CHECK_SIZE   (1 << 4)
     117             : /* don't try to open a segment, if not already open */
     118             : #define EXTENSION_DONT_OPEN         (1 << 5)
     119             : 
     120             : 
     121             : /* local routines */
     122             : static void mdunlinkfork(RelFileLocatorBackend rlocator, ForkNumber forknum,
     123             :                          bool isRedo);
     124             : static MdfdVec *mdopenfork(SMgrRelation reln, ForkNumber forknum, int behavior);
     125             : static void register_dirty_segment(SMgrRelation reln, ForkNumber forknum,
     126             :                                    MdfdVec *seg);
     127             : static void register_unlink_segment(RelFileLocatorBackend rlocator, ForkNumber forknum,
     128             :                                     BlockNumber segno);
     129             : static void register_forget_request(RelFileLocatorBackend rlocator, ForkNumber forknum,
     130             :                                     BlockNumber segno);
     131             : static void _fdvec_resize(SMgrRelation reln,
     132             :                           ForkNumber forknum,
     133             :                           int nseg);
     134             : static char *_mdfd_segpath(SMgrRelation reln, ForkNumber forknum,
     135             :                            BlockNumber segno);
     136             : static MdfdVec *_mdfd_openseg(SMgrRelation reln, ForkNumber forknum,
     137             :                               BlockNumber segno, int oflags);
     138             : static MdfdVec *_mdfd_getseg(SMgrRelation reln, ForkNumber forknum,
     139             :                              BlockNumber blkno, bool skipFsync, int behavior);
     140             : static BlockNumber _mdnblocks(SMgrRelation reln, ForkNumber forknum,
     141             :                               MdfdVec *seg);
     142             : 
     143             : static inline int
     144     2058010 : _mdfd_open_flags(void)
     145             : {
     146     2058010 :     int         flags = O_RDWR | PG_BINARY;
     147             : 
     148     2058010 :     if (io_direct_flags & IO_DIRECT_DATA)
     149         620 :         flags |= PG_O_DIRECT;
     150             : 
     151     2058010 :     return flags;
     152             : }
     153             : 
     154             : /*
     155             :  * mdinit() -- Initialize private state for magnetic disk storage manager.
     156             :  */
     157             : void
     158       29872 : mdinit(void)
     159             : {
     160       29872 :     MdCxt = AllocSetContextCreate(TopMemoryContext,
     161             :                                   "MdSmgr",
     162             :                                   ALLOCSET_DEFAULT_SIZES);
     163       29872 : }
     164             : 
     165             : /*
     166             :  * mdexists() -- Does the physical file exist?
     167             :  *
     168             :  * Note: this will return true for lingering files, with pending deletions
     169             :  */
     170             : bool
     171     1841082 : mdexists(SMgrRelation reln, ForkNumber forknum)
     172             : {
     173             :     /*
     174             :      * Close it first, to ensure that we notice if the fork has been unlinked
     175             :      * since we opened it.  As an optimization, we can skip that in recovery,
     176             :      * which already closes relations when dropping them.
     177             :      */
     178     1841082 :     if (!InRecovery)
     179      840558 :         mdclose(reln, forknum);
     180             : 
     181     1841082 :     return (mdopenfork(reln, forknum, EXTENSION_RETURN_NULL) != NULL);
     182             : }
     183             : 
     184             : /*
     185             :  * mdcreate() -- Create a new relation on magnetic disk.
     186             :  *
     187             :  * If isRedo is true, it's okay for the relation to exist already.
     188             :  */
     189             : void
     190     5487714 : mdcreate(SMgrRelation reln, ForkNumber forknum, bool isRedo)
     191             : {
     192             :     MdfdVec    *mdfd;
     193             :     char       *path;
     194             :     File        fd;
     195             : 
     196     5487714 :     if (isRedo && reln->md_num_open_segs[forknum] > 0)
     197     5208386 :         return;                 /* created and opened already... */
     198             : 
     199             :     Assert(reln->md_num_open_segs[forknum] == 0);
     200             : 
     201             :     /*
     202             :      * We may be using the target table space for the first time in this
     203             :      * database, so create a per-database subdirectory if needed.
     204             :      *
     205             :      * XXX this is a fairly ugly violation of module layering, but this seems
     206             :      * to be the best place to put the check.  Maybe TablespaceCreateDbspace
     207             :      * should be here and not in commands/tablespace.c?  But that would imply
     208             :      * importing a lot of stuff that smgr.c oughtn't know, either.
     209             :      */
     210      279328 :     TablespaceCreateDbspace(reln->smgr_rlocator.locator.spcOid,
     211             :                             reln->smgr_rlocator.locator.dbOid,
     212             :                             isRedo);
     213             : 
     214      279328 :     path = relpath(reln->smgr_rlocator, forknum);
     215             : 
     216      279328 :     fd = PathNameOpenFile(path, _mdfd_open_flags() | O_CREAT | O_EXCL);
     217             : 
     218      279328 :     if (fd < 0)
     219             :     {
     220        8124 :         int         save_errno = errno;
     221             : 
     222        8124 :         if (isRedo)
     223        8124 :             fd = PathNameOpenFile(path, _mdfd_open_flags());
     224        8124 :         if (fd < 0)
     225             :         {
     226             :             /* be sure to report the error reported by create, not open */
     227           0 :             errno = save_errno;
     228           0 :             ereport(ERROR,
     229             :                     (errcode_for_file_access(),
     230             :                      errmsg("could not create file \"%s\": %m", path)));
     231             :         }
     232             :     }
     233             : 
     234      279328 :     pfree(path);
     235             : 
     236      279328 :     _fdvec_resize(reln, forknum, 1);
     237      279328 :     mdfd = &reln->md_seg_fds[forknum][0];
     238      279328 :     mdfd->mdfd_vfd = fd;
     239      279328 :     mdfd->mdfd_segno = 0;
     240             : 
     241      279328 :     if (!SmgrIsTemp(reln))
     242      273148 :         register_dirty_segment(reln, forknum, mdfd);
     243             : }
     244             : 
     245             : /*
     246             :  * mdunlink() -- Unlink a relation.
     247             :  *
     248             :  * Note that we're passed a RelFileLocatorBackend --- by the time this is called,
     249             :  * there won't be an SMgrRelation hashtable entry anymore.
     250             :  *
     251             :  * forknum can be a fork number to delete a specific fork, or InvalidForkNumber
     252             :  * to delete all forks.
     253             :  *
     254             :  * For regular relations, we don't unlink the first segment file of the rel,
     255             :  * but just truncate it to zero length, and record a request to unlink it after
     256             :  * the next checkpoint.  Additional segments can be unlinked immediately,
     257             :  * however.  Leaving the empty file in place prevents that relfilenumber
     258             :  * from being reused.  The scenario this protects us from is:
     259             :  * 1. We delete a relation (and commit, and actually remove its file).
     260             :  * 2. We create a new relation, which by chance gets the same relfilenumber as
     261             :  *    the just-deleted one (OIDs must've wrapped around for that to happen).
     262             :  * 3. We crash before another checkpoint occurs.
     263             :  * During replay, we would delete the file and then recreate it, which is fine
     264             :  * if the contents of the file were repopulated by subsequent WAL entries.
     265             :  * But if we didn't WAL-log insertions, but instead relied on fsyncing the
     266             :  * file after populating it (as we do at wal_level=minimal), the contents of
     267             :  * the file would be lost forever.  By leaving the empty file until after the
     268             :  * next checkpoint, we prevent reassignment of the relfilenumber until it's
     269             :  * safe, because relfilenumber assignment skips over any existing file.
     270             :  *
     271             :  * Additional segments, if any, are truncated and then unlinked.  The reason
     272             :  * for truncating is that other backends may still hold open FDs for these at
     273             :  * the smgr level, so that the kernel can't remove the file yet.  We want to
     274             :  * reclaim the disk space right away despite that.
     275             :  *
     276             :  * We do not need to go through this dance for temp relations, though, because
     277             :  * we never make WAL entries for temp rels, and so a temp rel poses no threat
     278             :  * to the health of a regular rel that has taken over its relfilenumber.
     279             :  * The fact that temp rels and regular rels have different file naming
     280             :  * patterns provides additional safety.  Other backends shouldn't have open
     281             :  * FDs for them, either.
     282             :  *
     283             :  * We also don't do it while performing a binary upgrade.  There is no reuse
     284             :  * hazard in that case, since after a crash or even a simple ERROR, the
     285             :  * upgrade fails and the whole cluster must be recreated from scratch.
     286             :  * Furthermore, it is important to remove the files from disk immediately,
     287             :  * because we may be about to reuse the same relfilenumber.
     288             :  *
     289             :  * All the above applies only to the relation's main fork; other forks can
     290             :  * just be removed immediately, since they are not needed to prevent the
     291             :  * relfilenumber from being recycled.  Also, we do not carefully
     292             :  * track whether other forks have been created or not, but just attempt to
     293             :  * unlink them unconditionally; so we should never complain about ENOENT.
     294             :  *
     295             :  * If isRedo is true, it's unsurprising for the relation to be already gone.
     296             :  * Also, we should remove the file immediately instead of queuing a request
     297             :  * for later, since during redo there's no possibility of creating a
     298             :  * conflicting relation.
     299             :  *
     300             :  * Note: we currently just never warn about ENOENT at all.  We could warn in
     301             :  * the main-fork, non-isRedo case, but it doesn't seem worth the trouble.
     302             :  *
     303             :  * Note: any failure should be reported as WARNING not ERROR, because
     304             :  * we are usually not in a transaction anymore when this is called.
     305             :  */
     306             : void
     307      330208 : mdunlink(RelFileLocatorBackend rlocator, ForkNumber forknum, bool isRedo)
     308             : {
     309             :     /* Now do the per-fork work */
     310      330208 :     if (forknum == InvalidForkNumber)
     311             :     {
     312           0 :         for (forknum = 0; forknum <= MAX_FORKNUM; forknum++)
     313           0 :             mdunlinkfork(rlocator, forknum, isRedo);
     314             :     }
     315             :     else
     316      330208 :         mdunlinkfork(rlocator, forknum, isRedo);
     317      330208 : }
     318             : 
     319             : /*
     320             :  * Truncate a file to release disk space.
     321             :  */
     322             : static int
     323      386496 : do_truncate(const char *path)
     324             : {
     325             :     int         save_errno;
     326             :     int         ret;
     327             : 
     328      386496 :     ret = pg_truncate(path, 0);
     329             : 
     330             :     /* Log a warning here to avoid repetition in callers. */
     331      386496 :     if (ret < 0 && errno != ENOENT)
     332             :     {
     333           0 :         save_errno = errno;
     334           0 :         ereport(WARNING,
     335             :                 (errcode_for_file_access(),
     336             :                  errmsg("could not truncate file \"%s\": %m", path)));
     337           0 :         errno = save_errno;
     338             :     }
     339             : 
     340      386496 :     return ret;
     341             : }
     342             : 
     343             : static void
     344      330208 : mdunlinkfork(RelFileLocatorBackend rlocator, ForkNumber forknum, bool isRedo)
     345             : {
     346             :     char       *path;
     347             :     int         ret;
     348             :     int         save_errno;
     349             : 
     350      330208 :     path = relpath(rlocator, forknum);
     351             : 
     352             :     /*
     353             :      * Truncate and then unlink the first segment, or just register a request
     354             :      * to unlink it later, as described in the comments for mdunlink().
     355             :      */
     356      330208 :     if (isRedo || IsBinaryUpgrade || forknum != MAIN_FORKNUM ||
     357       69270 :         RelFileLocatorBackendIsTemp(rlocator))
     358             :     {
     359      266836 :         if (!RelFileLocatorBackendIsTemp(rlocator))
     360             :         {
     361             :             /* Prevent other backends' fds from holding on to the disk space */
     362      243244 :             ret = do_truncate(path);
     363             : 
     364             :             /* Forget any pending sync requests for the first segment */
     365      243244 :             save_errno = errno;
     366      243244 :             register_forget_request(rlocator, forknum, 0 /* first seg */ );
     367      243244 :             errno = save_errno;
     368             :         }
     369             :         else
     370       23592 :             ret = 0;
     371             : 
     372             :         /* Next unlink the file, unless it was already found to be missing */
     373      266836 :         if (ret >= 0 || errno != ENOENT)
     374             :         {
     375       40108 :             ret = unlink(path);
     376       40108 :             if (ret < 0 && errno != ENOENT)
     377             :             {
     378           0 :                 save_errno = errno;
     379           0 :                 ereport(WARNING,
     380             :                         (errcode_for_file_access(),
     381             :                          errmsg("could not remove file \"%s\": %m", path)));
     382           0 :                 errno = save_errno;
     383             :             }
     384             :         }
     385             :     }
     386             :     else
     387             :     {
     388             :         /* Prevent other backends' fds from holding on to the disk space */
     389       63372 :         ret = do_truncate(path);
     390             : 
     391             :         /* Register request to unlink first segment later */
     392       63372 :         save_errno = errno;
     393       63372 :         register_unlink_segment(rlocator, forknum, 0 /* first seg */ );
     394       63372 :         errno = save_errno;
     395             :     }
     396             : 
     397             :     /*
     398             :      * Delete any additional segments.
     399             :      *
     400             :      * Note that because we loop until getting ENOENT, we will correctly
     401             :      * remove all inactive segments as well as active ones.  Ideally we'd
     402             :      * continue the loop until getting exactly that errno, but that risks an
     403             :      * infinite loop if the problem is directory-wide (for instance, if we
     404             :      * suddenly can't read the data directory itself).  We compromise by
     405             :      * continuing after a non-ENOENT truncate error, but stopping after any
     406             :      * unlink error.  If there is indeed a directory-wide problem, additional
     407             :      * unlink attempts wouldn't work anyway.
     408             :      */
     409      330208 :     if (ret >= 0 || errno != ENOENT)
     410             :     {
     411       86060 :         char       *segpath = (char *) palloc(strlen(path) + 12);
     412             :         BlockNumber segno;
     413             : 
     414       86060 :         for (segno = 1;; segno++)
     415             :         {
     416       86060 :             sprintf(segpath, "%s.%u", path, segno);
     417             : 
     418       86060 :             if (!RelFileLocatorBackendIsTemp(rlocator))
     419             :             {
     420             :                 /*
     421             :                  * Prevent other backends' fds from holding on to the disk
     422             :                  * space.  We're done if we see ENOENT, though.
     423             :                  */
     424       79880 :                 if (do_truncate(segpath) < 0 && errno == ENOENT)
     425       79880 :                     break;
     426             : 
     427             :                 /*
     428             :                  * Forget any pending sync requests for this segment before we
     429             :                  * try to unlink.
     430             :                  */
     431           0 :                 register_forget_request(rlocator, forknum, segno);
     432             :             }
     433             : 
     434        6180 :             if (unlink(segpath) < 0)
     435             :             {
     436             :                 /* ENOENT is expected after the last segment... */
     437        6180 :                 if (errno != ENOENT)
     438           0 :                     ereport(WARNING,
     439             :                             (errcode_for_file_access(),
     440             :                              errmsg("could not remove file \"%s\": %m", segpath)));
     441        6180 :                 break;
     442             :             }
     443             :         }
     444       86060 :         pfree(segpath);
     445             :     }
     446             : 
     447      330208 :     pfree(path);
     448      330208 : }
     449             : 
     450             : /*
     451             :  * mdextend() -- Add a block to the specified relation.
     452             :  *
     453             :  * The semantics are nearly the same as mdwrite(): write at the
     454             :  * specified position.  However, this is to be used for the case of
     455             :  * extending a relation (i.e., blocknum is at or beyond the current
     456             :  * EOF).  Note that we assume writing a block beyond current EOF
     457             :  * causes intervening file space to become filled with zeroes.
     458             :  */
     459             : void
     460      214186 : mdextend(SMgrRelation reln, ForkNumber forknum, BlockNumber blocknum,
     461             :          const void *buffer, bool skipFsync)
     462             : {
     463             :     off_t       seekpos;
     464             :     int         nbytes;
     465             :     MdfdVec    *v;
     466             : 
     467             :     /* If this build supports direct I/O, the buffer must be I/O aligned. */
     468             :     if (PG_O_DIRECT != 0 && PG_IO_ALIGN_SIZE <= BLCKSZ)
     469             :         Assert((uintptr_t) buffer == TYPEALIGN(PG_IO_ALIGN_SIZE, buffer));
     470             : 
     471             :     /* This assert is too expensive to have on normally ... */
     472             : #ifdef CHECK_WRITE_VS_EXTEND
     473             :     Assert(blocknum >= mdnblocks(reln, forknum));
     474             : #endif
     475             : 
     476             :     /*
     477             :      * If a relation manages to grow to 2^32-1 blocks, refuse to extend it any
     478             :      * more --- we mustn't create a block whose number actually is
     479             :      * InvalidBlockNumber.  (Note that this failure should be unreachable
     480             :      * because of upstream checks in bufmgr.c.)
     481             :      */
     482      214186 :     if (blocknum == InvalidBlockNumber)
     483           0 :         ereport(ERROR,
     484             :                 (errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED),
     485             :                  errmsg("cannot extend file \"%s\" beyond %u blocks",
     486             :                         relpath(reln->smgr_rlocator, forknum),
     487             :                         InvalidBlockNumber)));
     488             : 
     489      214186 :     v = _mdfd_getseg(reln, forknum, blocknum, skipFsync, EXTENSION_CREATE);
     490             : 
     491      214186 :     seekpos = (off_t) BLCKSZ * (blocknum % ((BlockNumber) RELSEG_SIZE));
     492             : 
     493             :     Assert(seekpos < (off_t) BLCKSZ * RELSEG_SIZE);
     494             : 
     495      214186 :     if ((nbytes = FileWrite(v->mdfd_vfd, buffer, BLCKSZ, seekpos, WAIT_EVENT_DATA_FILE_EXTEND)) != BLCKSZ)
     496             :     {
     497           0 :         if (nbytes < 0)
     498           0 :             ereport(ERROR,
     499             :                     (errcode_for_file_access(),
     500             :                      errmsg("could not extend file \"%s\": %m",
     501             :                             FilePathName(v->mdfd_vfd)),
     502             :                      errhint("Check free disk space.")));
     503             :         /* short write: complain appropriately */
     504           0 :         ereport(ERROR,
     505             :                 (errcode(ERRCODE_DISK_FULL),
     506             :                  errmsg("could not extend file \"%s\": wrote only %d of %d bytes at block %u",
     507             :                         FilePathName(v->mdfd_vfd),
     508             :                         nbytes, BLCKSZ, blocknum),
     509             :                  errhint("Check free disk space.")));
     510             :     }
     511             : 
     512      214186 :     if (!skipFsync && !SmgrIsTemp(reln))
     513          58 :         register_dirty_segment(reln, forknum, v);
     514             : 
     515             :     Assert(_mdnblocks(reln, forknum, v) <= ((BlockNumber) RELSEG_SIZE));
     516      214186 : }
     517             : 
     518             : /*
     519             :  * mdzeroextend() -- Add new zeroed out blocks to the specified relation.
     520             :  *
     521             :  * Similar to mdextend(), except the relation can be extended by multiple
     522             :  * blocks at once and the added blocks will be filled with zeroes.
     523             :  */
     524             : void
     525      370878 : mdzeroextend(SMgrRelation reln, ForkNumber forknum,
     526             :              BlockNumber blocknum, int nblocks, bool skipFsync)
     527             : {
     528             :     MdfdVec    *v;
     529      370878 :     BlockNumber curblocknum = blocknum;
     530      370878 :     int         remblocks = nblocks;
     531             : 
     532             :     Assert(nblocks > 0);
     533             : 
     534             :     /* This assert is too expensive to have on normally ... */
     535             : #ifdef CHECK_WRITE_VS_EXTEND
     536             :     Assert(blocknum >= mdnblocks(reln, forknum));
     537             : #endif
     538             : 
     539             :     /*
     540             :      * If a relation manages to grow to 2^32-1 blocks, refuse to extend it any
     541             :      * more --- we mustn't create a block whose number actually is
     542             :      * InvalidBlockNumber or larger.
     543             :      */
     544      370878 :     if ((uint64) blocknum + nblocks >= (uint64) InvalidBlockNumber)
     545           0 :         ereport(ERROR,
     546             :                 (errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED),
     547             :                  errmsg("cannot extend file \"%s\" beyond %u blocks",
     548             :                         relpath(reln->smgr_rlocator, forknum),
     549             :                         InvalidBlockNumber)));
     550             : 
     551      741756 :     while (remblocks > 0)
     552             :     {
     553      370878 :         BlockNumber segstartblock = curblocknum % ((BlockNumber) RELSEG_SIZE);
     554      370878 :         off_t       seekpos = (off_t) BLCKSZ * segstartblock;
     555             :         int         numblocks;
     556             : 
     557      370878 :         if (segstartblock + remblocks > RELSEG_SIZE)
     558           0 :             numblocks = RELSEG_SIZE - segstartblock;
     559             :         else
     560      370878 :             numblocks = remblocks;
     561             : 
     562      370878 :         v = _mdfd_getseg(reln, forknum, curblocknum, skipFsync, EXTENSION_CREATE);
     563             : 
     564             :         Assert(segstartblock < RELSEG_SIZE);
     565             :         Assert(segstartblock + numblocks <= RELSEG_SIZE);
     566             : 
     567             :         /*
     568             :          * If available and useful, use posix_fallocate() (via
     569             :          * FileFallocate()) to extend the relation. That's often more
     570             :          * efficient than using write(), as it commonly won't cause the kernel
     571             :          * to allocate page cache space for the extended pages.
     572             :          *
     573             :          * However, we don't use FileFallocate() for small extensions, as it
     574             :          * defeats delayed allocation on some filesystems. Not clear where
     575             :          * that decision should be made though? For now just use a cutoff of
     576             :          * 8, anything between 4 and 8 worked OK in some local testing.
     577             :          */
     578      370878 :         if (numblocks > 8)
     579             :         {
     580             :             int         ret;
     581             : 
     582        1002 :             ret = FileFallocate(v->mdfd_vfd,
     583             :                                 seekpos, (off_t) BLCKSZ * numblocks,
     584             :                                 WAIT_EVENT_DATA_FILE_EXTEND);
     585        1002 :             if (ret != 0)
     586             :             {
     587           0 :                 ereport(ERROR,
     588             :                         errcode_for_file_access(),
     589             :                         errmsg("could not extend file \"%s\" with FileFallocate(): %m",
     590             :                                FilePathName(v->mdfd_vfd)),
     591             :                         errhint("Check free disk space."));
     592             :             }
     593             :         }
     594             :         else
     595             :         {
     596             :             int         ret;
     597             : 
     598             :             /*
     599             :              * Even if we don't want to use fallocate, we can still extend a
     600             :              * bit more efficiently than writing each 8kB block individually.
     601             :              * pg_pwrite_zeros() (via FileZero()) uses pg_pwritev_with_retry()
     602             :              * to avoid multiple writes or needing a zeroed buffer for the
     603             :              * whole length of the extension.
     604             :              */
     605      369876 :             ret = FileZero(v->mdfd_vfd,
     606             :                            seekpos, (off_t) BLCKSZ * numblocks,
     607             :                            WAIT_EVENT_DATA_FILE_EXTEND);
     608      369876 :             if (ret < 0)
     609           0 :                 ereport(ERROR,
     610             :                         errcode_for_file_access(),
     611             :                         errmsg("could not extend file \"%s\": %m",
     612             :                                FilePathName(v->mdfd_vfd)),
     613             :                         errhint("Check free disk space."));
     614             :         }
     615             : 
     616      370878 :         if (!skipFsync && !SmgrIsTemp(reln))
     617      353338 :             register_dirty_segment(reln, forknum, v);
     618             : 
     619             :         Assert(_mdnblocks(reln, forknum, v) <= ((BlockNumber) RELSEG_SIZE));
     620             : 
     621      370878 :         remblocks -= numblocks;
     622      370878 :         curblocknum += numblocks;
     623             :     }
     624      370878 : }
     625             : 
     626             : /*
     627             :  * mdopenfork() -- Open one fork of the specified relation.
     628             :  *
     629             :  * Note we only open the first segment, when there are multiple segments.
     630             :  *
     631             :  * If first segment is not present, either ereport or return NULL according
     632             :  * to "behavior".  We treat EXTENSION_CREATE the same as EXTENSION_FAIL;
     633             :  * EXTENSION_CREATE means it's OK to extend an existing relation, not to
     634             :  * invent one out of whole cloth.
     635             :  */
     636             : static MdfdVec *
     637     5665860 : mdopenfork(SMgrRelation reln, ForkNumber forknum, int behavior)
     638             : {
     639             :     MdfdVec    *mdfd;
     640             :     char       *path;
     641             :     File        fd;
     642             : 
     643             :     /* No work if already open */
     644     5665860 :     if (reln->md_num_open_segs[forknum] > 0)
     645     3939220 :         return &reln->md_seg_fds[forknum][0];
     646             : 
     647     1726640 :     path = relpath(reln->smgr_rlocator, forknum);
     648             : 
     649     1726640 :     fd = PathNameOpenFile(path, _mdfd_open_flags());
     650             : 
     651     1726640 :     if (fd < 0)
     652             :     {
     653      625384 :         if ((behavior & EXTENSION_RETURN_NULL) &&
     654      625340 :             FILE_POSSIBLY_DELETED(errno))
     655             :         {
     656      625340 :             pfree(path);
     657      625340 :             return NULL;
     658             :         }
     659          44 :         ereport(ERROR,
     660             :                 (errcode_for_file_access(),
     661             :                  errmsg("could not open file \"%s\": %m", path)));
     662             :     }
     663             : 
     664     1101256 :     pfree(path);
     665             : 
     666     1101256 :     _fdvec_resize(reln, forknum, 1);
     667     1101256 :     mdfd = &reln->md_seg_fds[forknum][0];
     668     1101256 :     mdfd->mdfd_vfd = fd;
     669     1101256 :     mdfd->mdfd_segno = 0;
     670             : 
     671             :     Assert(_mdnblocks(reln, forknum, mdfd) <= ((BlockNumber) RELSEG_SIZE));
     672             : 
     673     1101256 :     return mdfd;
     674             : }
     675             : 
     676             : /*
     677             :  * mdopen() -- Initialize newly-opened relation.
     678             :  */
     679             : void
     680     1551828 : mdopen(SMgrRelation reln)
     681             : {
     682             :     /* mark it not open */
     683     7759140 :     for (int forknum = 0; forknum <= MAX_FORKNUM; forknum++)
     684     6207312 :         reln->md_num_open_segs[forknum] = 0;
     685     1551828 : }
     686             : 
     687             : /*
     688             :  * mdclose() -- Close the specified relation, if it isn't closed already.
     689             :  */
     690             : void
     691     6096750 : mdclose(SMgrRelation reln, ForkNumber forknum)
     692             : {
     693     6096750 :     int         nopensegs = reln->md_num_open_segs[forknum];
     694             : 
     695             :     /* No work if already closed */
     696     6096750 :     if (nopensegs == 0)
     697     5243320 :         return;
     698             : 
     699             :     /* close segments starting from the end */
     700     1706860 :     while (nopensegs > 0)
     701             :     {
     702      853430 :         MdfdVec    *v = &reln->md_seg_fds[forknum][nopensegs - 1];
     703             : 
     704      853430 :         FileClose(v->mdfd_vfd);
     705      853430 :         _fdvec_resize(reln, forknum, nopensegs - 1);
     706      853430 :         nopensegs--;
     707             :     }
     708             : }
     709             : 
     710             : /*
     711             :  * mdprefetch() -- Initiate asynchronous read of the specified blocks of a relation
     712             :  */
     713             : bool
     714      241114 : mdprefetch(SMgrRelation reln, ForkNumber forknum, BlockNumber blocknum,
     715             :            int nblocks)
     716             : {
     717             : #ifdef USE_PREFETCH
     718             : 
     719             :     Assert((io_direct_flags & IO_DIRECT_DATA) == 0);
     720             : 
     721      241114 :     if ((uint64) blocknum + nblocks > (uint64) MaxBlockNumber + 1)
     722           0 :         return false;
     723             : 
     724      482228 :     while (nblocks > 0)
     725             :     {
     726             :         off_t       seekpos;
     727             :         MdfdVec    *v;
     728             :         int         nblocks_this_segment;
     729             : 
     730      241114 :         v = _mdfd_getseg(reln, forknum, blocknum, false,
     731      241114 :                          InRecovery ? EXTENSION_RETURN_NULL : EXTENSION_FAIL);
     732      241114 :         if (v == NULL)
     733           0 :             return false;
     734             : 
     735      241114 :         seekpos = (off_t) BLCKSZ * (blocknum % ((BlockNumber) RELSEG_SIZE));
     736             : 
     737             :         Assert(seekpos < (off_t) BLCKSZ * RELSEG_SIZE);
     738             : 
     739      241114 :         nblocks_this_segment =
     740      241114 :             Min(nblocks,
     741             :                 RELSEG_SIZE - (blocknum % ((BlockNumber) RELSEG_SIZE)));
     742             : 
     743      241114 :         (void) FilePrefetch(v->mdfd_vfd, seekpos, BLCKSZ * nblocks_this_segment,
     744             :                             WAIT_EVENT_DATA_FILE_PREFETCH);
     745             : 
     746      241114 :         blocknum += nblocks_this_segment;
     747      241114 :         nblocks -= nblocks_this_segment;
     748             :     }
     749             : #endif                          /* USE_PREFETCH */
     750             : 
     751      241114 :     return true;
     752             : }
     753             : 
     754             : /*
     755             :  * Convert an array of buffer address into an array of iovec objects, and
     756             :  * return the number that were required.  'iov' must have enough space for up
     757             :  * to 'nblocks' elements, but the number used may be less depending on
     758             :  * merging.  In the case of a run of fully contiguous buffers, a single iovec
     759             :  * will be populated that can be handled as a plain non-vectored I/O.
     760             :  */
     761             : static int
     762     2988568 : buffers_to_iovec(struct iovec *iov, void **buffers, int nblocks)
     763             : {
     764             :     struct iovec *iovp;
     765             :     int         iovcnt;
     766             : 
     767             :     Assert(nblocks >= 1);
     768             : 
     769             :     /* If this build supports direct I/O, buffers must be I/O aligned. */
     770     6051096 :     for (int i = 0; i < nblocks; ++i)
     771             :     {
     772             :         if (PG_O_DIRECT != 0 && PG_IO_ALIGN_SIZE <= BLCKSZ)
     773             :             Assert((uintptr_t) buffers[i] ==
     774             :                    TYPEALIGN(PG_IO_ALIGN_SIZE, buffers[i]));
     775             :     }
     776             : 
     777             :     /* Start the first iovec off with the first buffer. */
     778     2988568 :     iovp = &iov[0];
     779     2988568 :     iovp->iov_base = buffers[0];
     780     2988568 :     iovp->iov_len = BLCKSZ;
     781     2988568 :     iovcnt = 1;
     782             : 
     783             :     /* Try to merge the rest. */
     784     3062528 :     for (int i = 1; i < nblocks; ++i)
     785             :     {
     786       73960 :         void       *buffer = buffers[i];
     787             : 
     788       73960 :         if (((char *) iovp->iov_base + iovp->iov_len) == buffer)
     789             :         {
     790             :             /* Contiguous with the last iovec. */
     791       70772 :             iovp->iov_len += BLCKSZ;
     792             :         }
     793             :         else
     794             :         {
     795             :             /* Need a new iovec. */
     796        3188 :             iovp++;
     797        3188 :             iovp->iov_base = buffer;
     798        3188 :             iovp->iov_len = BLCKSZ;
     799        3188 :             iovcnt++;
     800             :         }
     801             :     }
     802             : 
     803     2988568 :     return iovcnt;
     804             : }
     805             : 
     806             : /*
     807             :  * mdreadv() -- Read the specified blocks from a relation.
     808             :  */
     809             : void
     810     2100444 : mdreadv(SMgrRelation reln, ForkNumber forknum, BlockNumber blocknum,
     811             :         void **buffers, BlockNumber nblocks)
     812             : {
     813     4200858 :     while (nblocks > 0)
     814             :     {
     815             :         struct iovec iov[PG_IOV_MAX];
     816             :         int         iovcnt;
     817             :         off_t       seekpos;
     818             :         int         nbytes;
     819             :         MdfdVec    *v;
     820             :         BlockNumber nblocks_this_segment;
     821             :         size_t      transferred_this_segment;
     822             :         size_t      size_this_segment;
     823             : 
     824     2100444 :         v = _mdfd_getseg(reln, forknum, blocknum, false,
     825             :                          EXTENSION_FAIL | EXTENSION_CREATE_RECOVERY);
     826             : 
     827     2100414 :         seekpos = (off_t) BLCKSZ * (blocknum % ((BlockNumber) RELSEG_SIZE));
     828             : 
     829             :         Assert(seekpos < (off_t) BLCKSZ * RELSEG_SIZE);
     830             : 
     831     2100414 :         nblocks_this_segment =
     832     2100414 :             Min(nblocks,
     833             :                 RELSEG_SIZE - (blocknum % ((BlockNumber) RELSEG_SIZE)));
     834     2100414 :         nblocks_this_segment = Min(nblocks_this_segment, lengthof(iov));
     835             : 
     836     2100414 :         iovcnt = buffers_to_iovec(iov, buffers, nblocks_this_segment);
     837     2100414 :         size_this_segment = nblocks_this_segment * BLCKSZ;
     838     2100414 :         transferred_this_segment = 0;
     839             : 
     840             :         /*
     841             :          * Inner loop to continue after a short read.  We'll keep going until
     842             :          * we hit EOF rather than assuming that a short read means we hit the
     843             :          * end.
     844             :          */
     845             :         for (;;)
     846             :         {
     847             :             TRACE_POSTGRESQL_SMGR_MD_READ_START(forknum, blocknum,
     848             :                                                 reln->smgr_rlocator.locator.spcOid,
     849             :                                                 reln->smgr_rlocator.locator.dbOid,
     850             :                                                 reln->smgr_rlocator.locator.relNumber,
     851             :                                                 reln->smgr_rlocator.backend);
     852     2100414 :             nbytes = FileReadV(v->mdfd_vfd, iov, iovcnt, seekpos,
     853             :                                WAIT_EVENT_DATA_FILE_READ);
     854             :             TRACE_POSTGRESQL_SMGR_MD_READ_DONE(forknum, blocknum,
     855             :                                                reln->smgr_rlocator.locator.spcOid,
     856             :                                                reln->smgr_rlocator.locator.dbOid,
     857             :                                                reln->smgr_rlocator.locator.relNumber,
     858             :                                                reln->smgr_rlocator.backend,
     859             :                                                nbytes,
     860             :                                                size_this_segment - transferred_this_segment);
     861             : 
     862             : #ifdef SIMULATE_SHORT_READ
     863             :             nbytes = Min(nbytes, 4096);
     864             : #endif
     865             : 
     866     2100414 :             if (nbytes < 0)
     867           0 :                 ereport(ERROR,
     868             :                         (errcode_for_file_access(),
     869             :                          errmsg("could not read blocks %u..%u in file \"%s\": %m",
     870             :                                 blocknum,
     871             :                                 blocknum + nblocks_this_segment - 1,
     872             :                                 FilePathName(v->mdfd_vfd))));
     873             : 
     874     2100414 :             if (nbytes == 0)
     875             :             {
     876             :                 /*
     877             :                  * We are at or past EOF, or we read a partial block at EOF.
     878             :                  * Normally this is an error; upper levels should never try to
     879             :                  * read a nonexistent block.  However, if zero_damaged_pages
     880             :                  * is ON or we are InRecovery, we should instead return zeroes
     881             :                  * without complaining.  This allows, for example, the case of
     882             :                  * trying to update a block that was later truncated away.
     883             :                  */
     884           0 :                 if (zero_damaged_pages || InRecovery)
     885             :                 {
     886           0 :                     for (BlockNumber i = transferred_this_segment / BLCKSZ;
     887             :                          i < nblocks_this_segment;
     888           0 :                          ++i)
     889           0 :                         memset(buffers[i], 0, BLCKSZ);
     890           0 :                     break;
     891             :                 }
     892             :                 else
     893           0 :                     ereport(ERROR,
     894             :                             (errcode(ERRCODE_DATA_CORRUPTED),
     895             :                              errmsg("could not read blocks %u..%u in file \"%s\": read only %zu of %zu bytes",
     896             :                                     blocknum,
     897             :                                     blocknum + nblocks_this_segment - 1,
     898             :                                     FilePathName(v->mdfd_vfd),
     899             :                                     transferred_this_segment,
     900             :                                     size_this_segment)));
     901             :             }
     902             : 
     903             :             /* One loop should usually be enough. */
     904     2100414 :             transferred_this_segment += nbytes;
     905             :             Assert(transferred_this_segment <= size_this_segment);
     906     2100414 :             if (transferred_this_segment == size_this_segment)
     907     2100414 :                 break;
     908             : 
     909             :             /* Adjust position and vectors after a short read. */
     910           0 :             seekpos += nbytes;
     911           0 :             iovcnt = compute_remaining_iovec(iov, iov, iovcnt, nbytes);
     912             :         }
     913             : 
     914     2100414 :         nblocks -= nblocks_this_segment;
     915     2100414 :         buffers += nblocks_this_segment;
     916     2100414 :         blocknum += nblocks_this_segment;
     917             :     }
     918     2100414 : }
     919             : 
     920             : /*
     921             :  * mdwritev() -- Write the supplied blocks at the appropriate location.
     922             :  *
     923             :  * This is to be used only for updating already-existing blocks of a
     924             :  * relation (ie, those before the current EOF).  To extend a relation,
     925             :  * use mdextend().
     926             :  */
     927             : void
     928      888154 : mdwritev(SMgrRelation reln, ForkNumber forknum, BlockNumber blocknum,
     929             :          const void **buffers, BlockNumber nblocks, bool skipFsync)
     930             : {
     931             :     /* This assert is too expensive to have on normally ... */
     932             : #ifdef CHECK_WRITE_VS_EXTEND
     933             :     Assert(blocknum < mdnblocks(reln, forknum));
     934             : #endif
     935             : 
     936     1776308 :     while (nblocks > 0)
     937             :     {
     938             :         struct iovec iov[PG_IOV_MAX];
     939             :         int         iovcnt;
     940             :         off_t       seekpos;
     941             :         int         nbytes;
     942             :         MdfdVec    *v;
     943             :         BlockNumber nblocks_this_segment;
     944             :         size_t      transferred_this_segment;
     945             :         size_t      size_this_segment;
     946             : 
     947      888154 :         v = _mdfd_getseg(reln, forknum, blocknum, skipFsync,
     948             :                          EXTENSION_FAIL | EXTENSION_CREATE_RECOVERY);
     949             : 
     950      888154 :         seekpos = (off_t) BLCKSZ * (blocknum % ((BlockNumber) RELSEG_SIZE));
     951             : 
     952             :         Assert(seekpos < (off_t) BLCKSZ * RELSEG_SIZE);
     953             : 
     954      888154 :         nblocks_this_segment =
     955      888154 :             Min(nblocks,
     956             :                 RELSEG_SIZE - (blocknum % ((BlockNumber) RELSEG_SIZE)));
     957      888154 :         nblocks_this_segment = Min(nblocks_this_segment, lengthof(iov));
     958             : 
     959      888154 :         iovcnt = buffers_to_iovec(iov, (void **) buffers, nblocks_this_segment);
     960      888154 :         size_this_segment = nblocks_this_segment * BLCKSZ;
     961      888154 :         transferred_this_segment = 0;
     962             : 
     963             :         /*
     964             :          * Inner loop to continue after a short write.  If the reason is that
     965             :          * we're out of disk space, a future attempt should get an ENOSPC
     966             :          * error from the kernel.
     967             :          */
     968             :         for (;;)
     969             :         {
     970             :             TRACE_POSTGRESQL_SMGR_MD_WRITE_START(forknum, blocknum,
     971             :                                                  reln->smgr_rlocator.locator.spcOid,
     972             :                                                  reln->smgr_rlocator.locator.dbOid,
     973             :                                                  reln->smgr_rlocator.locator.relNumber,
     974             :                                                  reln->smgr_rlocator.backend);
     975      888154 :             nbytes = FileWriteV(v->mdfd_vfd, iov, iovcnt, seekpos,
     976             :                                 WAIT_EVENT_DATA_FILE_WRITE);
     977             :             TRACE_POSTGRESQL_SMGR_MD_WRITE_DONE(forknum, blocknum,
     978             :                                                 reln->smgr_rlocator.locator.spcOid,
     979             :                                                 reln->smgr_rlocator.locator.dbOid,
     980             :                                                 reln->smgr_rlocator.locator.relNumber,
     981             :                                                 reln->smgr_rlocator.backend,
     982             :                                                 nbytes,
     983             :                                                 size_this_segment - transferred_this_segment);
     984             : 
     985             : #ifdef SIMULATE_SHORT_WRITE
     986             :             nbytes = Min(nbytes, 4096);
     987             : #endif
     988             : 
     989      888154 :             if (nbytes < 0)
     990             :             {
     991           0 :                 bool        enospc = errno == ENOSPC;
     992             : 
     993           0 :                 ereport(ERROR,
     994             :                         (errcode_for_file_access(),
     995             :                          errmsg("could not write blocks %u..%u in file \"%s\": %m",
     996             :                                 blocknum,
     997             :                                 blocknum + nblocks_this_segment - 1,
     998             :                                 FilePathName(v->mdfd_vfd)),
     999             :                          enospc ? errhint("Check free disk space.") : 0));
    1000             :             }
    1001             : 
    1002             :             /* One loop should usually be enough. */
    1003      888154 :             transferred_this_segment += nbytes;
    1004             :             Assert(transferred_this_segment <= size_this_segment);
    1005      888154 :             if (transferred_this_segment == size_this_segment)
    1006      888154 :                 break;
    1007             : 
    1008             :             /* Adjust position and iovecs after a short write. */
    1009           0 :             seekpos += nbytes;
    1010           0 :             iovcnt = compute_remaining_iovec(iov, iov, iovcnt, nbytes);
    1011             :         }
    1012             : 
    1013      888154 :         if (!skipFsync && !SmgrIsTemp(reln))
    1014      886254 :             register_dirty_segment(reln, forknum, v);
    1015             : 
    1016      888154 :         nblocks -= nblocks_this_segment;
    1017      888154 :         buffers += nblocks_this_segment;
    1018      888154 :         blocknum += nblocks_this_segment;
    1019             :     }
    1020      888154 : }
    1021             : 
    1022             : 
    1023             : /*
    1024             :  * mdwriteback() -- Tell the kernel to write pages back to storage.
    1025             :  *
    1026             :  * This accepts a range of blocks because flushing several pages at once is
    1027             :  * considerably more efficient than doing so individually.
    1028             :  */
    1029             : void
    1030      133484 : mdwriteback(SMgrRelation reln, ForkNumber forknum,
    1031             :             BlockNumber blocknum, BlockNumber nblocks)
    1032             : {
    1033             :     Assert((io_direct_flags & IO_DIRECT_DATA) == 0);
    1034             : 
    1035             :     /*
    1036             :      * Issue flush requests in as few requests as possible; have to split at
    1037             :      * segment boundaries though, since those are actually separate files.
    1038             :      */
    1039      266316 :     while (nblocks > 0)
    1040             :     {
    1041      133484 :         BlockNumber nflush = nblocks;
    1042             :         off_t       seekpos;
    1043             :         MdfdVec    *v;
    1044             :         int         segnum_start,
    1045             :                     segnum_end;
    1046             : 
    1047      133484 :         v = _mdfd_getseg(reln, forknum, blocknum, true /* not used */ ,
    1048             :                          EXTENSION_DONT_OPEN);
    1049             : 
    1050             :         /*
    1051             :          * We might be flushing buffers of already removed relations, that's
    1052             :          * ok, just ignore that case.  If the segment file wasn't open already
    1053             :          * (ie from a recent mdwrite()), then we don't want to re-open it, to
    1054             :          * avoid a race with PROCSIGNAL_BARRIER_SMGRRELEASE that might leave
    1055             :          * us with a descriptor to a file that is about to be unlinked.
    1056             :          */
    1057      133484 :         if (!v)
    1058         652 :             return;
    1059             : 
    1060             :         /* compute offset inside the current segment */
    1061      132832 :         segnum_start = blocknum / RELSEG_SIZE;
    1062             : 
    1063             :         /* compute number of desired writes within the current segment */
    1064      132832 :         segnum_end = (blocknum + nblocks - 1) / RELSEG_SIZE;
    1065      132832 :         if (segnum_start != segnum_end)
    1066           0 :             nflush = RELSEG_SIZE - (blocknum % ((BlockNumber) RELSEG_SIZE));
    1067             : 
    1068             :         Assert(nflush >= 1);
    1069             :         Assert(nflush <= nblocks);
    1070             : 
    1071      132832 :         seekpos = (off_t) BLCKSZ * (blocknum % ((BlockNumber) RELSEG_SIZE));
    1072             : 
    1073      132832 :         FileWriteback(v->mdfd_vfd, seekpos, (off_t) BLCKSZ * nflush, WAIT_EVENT_DATA_FILE_FLUSH);
    1074             : 
    1075      132832 :         nblocks -= nflush;
    1076      132832 :         blocknum += nflush;
    1077             :     }
    1078             : }
    1079             : 
    1080             : /*
    1081             :  * mdnblocks() -- Get the number of blocks stored in a relation.
    1082             :  *
    1083             :  * Important side effect: all active segments of the relation are opened
    1084             :  * and added to the md_seg_fds array.  If this routine has not been
    1085             :  * called, then only segments up to the last one actually touched
    1086             :  * are present in the array.
    1087             :  */
    1088             : BlockNumber
    1089     3484900 : mdnblocks(SMgrRelation reln, ForkNumber forknum)
    1090             : {
    1091             :     MdfdVec    *v;
    1092             :     BlockNumber nblocks;
    1093             :     BlockNumber segno;
    1094             : 
    1095     3484900 :     mdopenfork(reln, forknum, EXTENSION_FAIL);
    1096             : 
    1097             :     /* mdopen has opened the first segment */
    1098             :     Assert(reln->md_num_open_segs[forknum] > 0);
    1099             : 
    1100             :     /*
    1101             :      * Start from the last open segments, to avoid redundant seeks.  We have
    1102             :      * previously verified that these segments are exactly RELSEG_SIZE long,
    1103             :      * and it's useless to recheck that each time.
    1104             :      *
    1105             :      * NOTE: this assumption could only be wrong if another backend has
    1106             :      * truncated the relation.  We rely on higher code levels to handle that
    1107             :      * scenario by closing and re-opening the md fd, which is handled via
    1108             :      * relcache flush.  (Since the checkpointer doesn't participate in
    1109             :      * relcache flush, it could have segment entries for inactive segments;
    1110             :      * that's OK because the checkpointer never needs to compute relation
    1111             :      * size.)
    1112             :      */
    1113     3484862 :     segno = reln->md_num_open_segs[forknum] - 1;
    1114     3484862 :     v = &reln->md_seg_fds[forknum][segno];
    1115             : 
    1116             :     for (;;)
    1117             :     {
    1118     3484862 :         nblocks = _mdnblocks(reln, forknum, v);
    1119     3484862 :         if (nblocks > ((BlockNumber) RELSEG_SIZE))
    1120           0 :             elog(FATAL, "segment too big");
    1121     3484862 :         if (nblocks < ((BlockNumber) RELSEG_SIZE))
    1122     3484862 :             return (segno * ((BlockNumber) RELSEG_SIZE)) + nblocks;
    1123             : 
    1124             :         /*
    1125             :          * If segment is exactly RELSEG_SIZE, advance to next one.
    1126             :          */
    1127           0 :         segno++;
    1128             : 
    1129             :         /*
    1130             :          * We used to pass O_CREAT here, but that has the disadvantage that it
    1131             :          * might create a segment which has vanished through some operating
    1132             :          * system misadventure.  In such a case, creating the segment here
    1133             :          * undermines _mdfd_getseg's attempts to notice and report an error
    1134             :          * upon access to a missing segment.
    1135             :          */
    1136           0 :         v = _mdfd_openseg(reln, forknum, segno, 0);
    1137           0 :         if (v == NULL)
    1138           0 :             return segno * ((BlockNumber) RELSEG_SIZE);
    1139             :     }
    1140             : }
    1141             : 
    1142             : /*
    1143             :  * mdtruncate() -- Truncate relation to specified number of blocks.
    1144             :  */
    1145             : void
    1146        1514 : mdtruncate(SMgrRelation reln, ForkNumber forknum, BlockNumber nblocks)
    1147             : {
    1148             :     BlockNumber curnblk;
    1149             :     BlockNumber priorblocks;
    1150             :     int         curopensegs;
    1151             : 
    1152             :     /*
    1153             :      * NOTE: mdnblocks makes sure we have opened all active segments, so that
    1154             :      * truncation loop will get them all!
    1155             :      */
    1156        1514 :     curnblk = mdnblocks(reln, forknum);
    1157        1514 :     if (nblocks > curnblk)
    1158             :     {
    1159             :         /* Bogus request ... but no complaint if InRecovery */
    1160           0 :         if (InRecovery)
    1161           0 :             return;
    1162           0 :         ereport(ERROR,
    1163             :                 (errmsg("could not truncate file \"%s\" to %u blocks: it's only %u blocks now",
    1164             :                         relpath(reln->smgr_rlocator, forknum),
    1165             :                         nblocks, curnblk)));
    1166             :     }
    1167        1514 :     if (nblocks == curnblk)
    1168         624 :         return;                 /* no work */
    1169             : 
    1170             :     /*
    1171             :      * Truncate segments, starting at the last one. Starting at the end makes
    1172             :      * managing the memory for the fd array easier, should there be errors.
    1173             :      */
    1174         890 :     curopensegs = reln->md_num_open_segs[forknum];
    1175        1780 :     while (curopensegs > 0)
    1176             :     {
    1177             :         MdfdVec    *v;
    1178             : 
    1179         890 :         priorblocks = (curopensegs - 1) * RELSEG_SIZE;
    1180             : 
    1181         890 :         v = &reln->md_seg_fds[forknum][curopensegs - 1];
    1182             : 
    1183         890 :         if (priorblocks > nblocks)
    1184             :         {
    1185             :             /*
    1186             :              * This segment is no longer active. We truncate the file, but do
    1187             :              * not delete it, for reasons explained in the header comments.
    1188             :              */
    1189           0 :             if (FileTruncate(v->mdfd_vfd, 0, WAIT_EVENT_DATA_FILE_TRUNCATE) < 0)
    1190           0 :                 ereport(ERROR,
    1191             :                         (errcode_for_file_access(),
    1192             :                          errmsg("could not truncate file \"%s\": %m",
    1193             :                                 FilePathName(v->mdfd_vfd))));
    1194             : 
    1195           0 :             if (!SmgrIsTemp(reln))
    1196           0 :                 register_dirty_segment(reln, forknum, v);
    1197             : 
    1198             :             /* we never drop the 1st segment */
    1199             :             Assert(v != &reln->md_seg_fds[forknum][0]);
    1200             : 
    1201           0 :             FileClose(v->mdfd_vfd);
    1202           0 :             _fdvec_resize(reln, forknum, curopensegs - 1);
    1203             :         }
    1204         890 :         else if (priorblocks + ((BlockNumber) RELSEG_SIZE) > nblocks)
    1205             :         {
    1206             :             /*
    1207             :              * This is the last segment we want to keep. Truncate the file to
    1208             :              * the right length. NOTE: if nblocks is exactly a multiple K of
    1209             :              * RELSEG_SIZE, we will truncate the K+1st segment to 0 length but
    1210             :              * keep it. This adheres to the invariant given in the header
    1211             :              * comments.
    1212             :              */
    1213         890 :             BlockNumber lastsegblocks = nblocks - priorblocks;
    1214             : 
    1215         890 :             if (FileTruncate(v->mdfd_vfd, (off_t) lastsegblocks * BLCKSZ, WAIT_EVENT_DATA_FILE_TRUNCATE) < 0)
    1216           0 :                 ereport(ERROR,
    1217             :                         (errcode_for_file_access(),
    1218             :                          errmsg("could not truncate file \"%s\" to %u blocks: %m",
    1219             :                                 FilePathName(v->mdfd_vfd),
    1220             :                                 nblocks)));
    1221         890 :             if (!SmgrIsTemp(reln))
    1222         616 :                 register_dirty_segment(reln, forknum, v);
    1223             :         }
    1224             :         else
    1225             :         {
    1226             :             /*
    1227             :              * We still need this segment, so nothing to do for this and any
    1228             :              * earlier segment.
    1229             :              */
    1230           0 :             break;
    1231             :         }
    1232         890 :         curopensegs--;
    1233             :     }
    1234             : }
    1235             : 
    1236             : /*
    1237             :  * mdregistersync() -- Mark whole relation as needing fsync
    1238             :  */
    1239             : void
    1240       43880 : mdregistersync(SMgrRelation reln, ForkNumber forknum)
    1241             : {
    1242             :     int         segno;
    1243             :     int         min_inactive_seg;
    1244             : 
    1245             :     /*
    1246             :      * NOTE: mdnblocks makes sure we have opened all active segments, so that
    1247             :      * the loop below will get them all!
    1248             :      */
    1249       43880 :     mdnblocks(reln, forknum);
    1250             : 
    1251       43880 :     min_inactive_seg = segno = reln->md_num_open_segs[forknum];
    1252             : 
    1253             :     /*
    1254             :      * Temporarily open inactive segments, then close them after sync.  There
    1255             :      * may be some inactive segments left opened after error, but that is
    1256             :      * harmless.  We don't bother to clean them up and take a risk of further
    1257             :      * trouble.  The next mdclose() will soon close them.
    1258             :      */
    1259       43880 :     while (_mdfd_openseg(reln, forknum, segno, 0) != NULL)
    1260           0 :         segno++;
    1261             : 
    1262       87760 :     while (segno > 0)
    1263             :     {
    1264       43880 :         MdfdVec    *v = &reln->md_seg_fds[forknum][segno - 1];
    1265             : 
    1266       43880 :         register_dirty_segment(reln, forknum, v);
    1267             : 
    1268             :         /* Close inactive segments immediately */
    1269       43880 :         if (segno > min_inactive_seg)
    1270             :         {
    1271           0 :             FileClose(v->mdfd_vfd);
    1272           0 :             _fdvec_resize(reln, forknum, segno - 1);
    1273             :         }
    1274             : 
    1275       43880 :         segno--;
    1276             :     }
    1277       43880 : }
    1278             : 
    1279             : /*
    1280             :  * mdimmedsync() -- Immediately sync a relation to stable storage.
    1281             :  *
    1282             :  * Note that only writes already issued are synced; this routine knows
    1283             :  * nothing of dirty buffers that may exist inside the buffer manager.  We
    1284             :  * sync active and inactive segments; smgrDoPendingSyncs() relies on this.
    1285             :  * Consider a relation skipping WAL.  Suppose a checkpoint syncs blocks of
    1286             :  * some segment, then mdtruncate() renders that segment inactive.  If we
    1287             :  * crash before the next checkpoint syncs the newly-inactive segment, that
    1288             :  * segment may survive recovery, reintroducing unwanted data into the table.
    1289             :  */
    1290             : void
    1291          38 : mdimmedsync(SMgrRelation reln, ForkNumber forknum)
    1292             : {
    1293             :     int         segno;
    1294             :     int         min_inactive_seg;
    1295             : 
    1296             :     /*
    1297             :      * NOTE: mdnblocks makes sure we have opened all active segments, so that
    1298             :      * the loop below will get them all!
    1299             :      */
    1300          38 :     mdnblocks(reln, forknum);
    1301             : 
    1302          38 :     min_inactive_seg = segno = reln->md_num_open_segs[forknum];
    1303             : 
    1304             :     /*
    1305             :      * Temporarily open inactive segments, then close them after sync.  There
    1306             :      * may be some inactive segments left opened after fsync() error, but that
    1307             :      * is harmless.  We don't bother to clean them up and take a risk of
    1308             :      * further trouble.  The next mdclose() will soon close them.
    1309             :      */
    1310          38 :     while (_mdfd_openseg(reln, forknum, segno, 0) != NULL)
    1311           0 :         segno++;
    1312             : 
    1313          76 :     while (segno > 0)
    1314             :     {
    1315          38 :         MdfdVec    *v = &reln->md_seg_fds[forknum][segno - 1];
    1316             : 
    1317             :         /*
    1318             :          * fsyncs done through mdimmedsync() should be tracked in a separate
    1319             :          * IOContext than those done through mdsyncfiletag() to differentiate
    1320             :          * between unavoidable client backend fsyncs (e.g. those done during
    1321             :          * index build) and those which ideally would have been done by the
    1322             :          * checkpointer. Since other IO operations bypassing the buffer
    1323             :          * manager could also be tracked in such an IOContext, wait until
    1324             :          * these are also tracked to track immediate fsyncs.
    1325             :          */
    1326          38 :         if (FileSync(v->mdfd_vfd, WAIT_EVENT_DATA_FILE_IMMEDIATE_SYNC) < 0)
    1327           0 :             ereport(data_sync_elevel(ERROR),
    1328             :                     (errcode_for_file_access(),
    1329             :                      errmsg("could not fsync file \"%s\": %m",
    1330             :                             FilePathName(v->mdfd_vfd))));
    1331             : 
    1332             :         /* Close inactive segments immediately */
    1333          38 :         if (segno > min_inactive_seg)
    1334             :         {
    1335           0 :             FileClose(v->mdfd_vfd);
    1336           0 :             _fdvec_resize(reln, forknum, segno - 1);
    1337             :         }
    1338             : 
    1339          38 :         segno--;
    1340             :     }
    1341          38 : }
    1342             : 
    1343             : /*
    1344             :  * register_dirty_segment() -- Mark a relation segment as needing fsync
    1345             :  *
    1346             :  * If there is a local pending-ops table, just make an entry in it for
    1347             :  * ProcessSyncRequests to process later.  Otherwise, try to pass off the
    1348             :  * fsync request to the checkpointer process.  If that fails, just do the
    1349             :  * fsync locally before returning (we hope this will not happen often
    1350             :  * enough to be a performance problem).
    1351             :  */
    1352             : static void
    1353     1557294 : register_dirty_segment(SMgrRelation reln, ForkNumber forknum, MdfdVec *seg)
    1354             : {
    1355             :     FileTag     tag;
    1356             : 
    1357     1557294 :     INIT_MD_FILETAG(tag, reln->smgr_rlocator.locator, forknum, seg->mdfd_segno);
    1358             : 
    1359             :     /* Temp relations should never be fsync'd */
    1360             :     Assert(!SmgrIsTemp(reln));
    1361             : 
    1362     1557294 :     if (!RegisterSyncRequest(&tag, SYNC_REQUEST, false /* retryOnError */ ))
    1363             :     {
    1364             :         instr_time  io_start;
    1365             : 
    1366          54 :         ereport(DEBUG1,
    1367             :                 (errmsg_internal("could not forward fsync request because request queue is full")));
    1368             : 
    1369          54 :         io_start = pgstat_prepare_io_time(track_io_timing);
    1370             : 
    1371          54 :         if (FileSync(seg->mdfd_vfd, WAIT_EVENT_DATA_FILE_SYNC) < 0)
    1372           0 :             ereport(data_sync_elevel(ERROR),
    1373             :                     (errcode_for_file_access(),
    1374             :                      errmsg("could not fsync file \"%s\": %m",
    1375             :                             FilePathName(seg->mdfd_vfd))));
    1376             : 
    1377             :         /*
    1378             :          * We have no way of knowing if the current IOContext is
    1379             :          * IOCONTEXT_NORMAL or IOCONTEXT_[BULKREAD, BULKWRITE, VACUUM] at this
    1380             :          * point, so count the fsync as being in the IOCONTEXT_NORMAL
    1381             :          * IOContext. This is probably okay, because the number of backend
    1382             :          * fsyncs doesn't say anything about the efficacy of the
    1383             :          * BufferAccessStrategy. And counting both fsyncs done in
    1384             :          * IOCONTEXT_NORMAL and IOCONTEXT_[BULKREAD, BULKWRITE, VACUUM] under
    1385             :          * IOCONTEXT_NORMAL is likely clearer when investigating the number of
    1386             :          * backend fsyncs.
    1387             :          */
    1388          54 :         pgstat_count_io_op_time(IOOBJECT_RELATION, IOCONTEXT_NORMAL,
    1389             :                                 IOOP_FSYNC, io_start, 1);
    1390             :     }
    1391     1557294 : }
    1392             : 
    1393             : /*
    1394             :  * register_unlink_segment() -- Schedule a file to be deleted after next checkpoint
    1395             :  */
    1396             : static void
    1397       63372 : register_unlink_segment(RelFileLocatorBackend rlocator, ForkNumber forknum,
    1398             :                         BlockNumber segno)
    1399             : {
    1400             :     FileTag     tag;
    1401             : 
    1402       63372 :     INIT_MD_FILETAG(tag, rlocator.locator, forknum, segno);
    1403             : 
    1404             :     /* Should never be used with temp relations */
    1405             :     Assert(!RelFileLocatorBackendIsTemp(rlocator));
    1406             : 
    1407       63372 :     RegisterSyncRequest(&tag, SYNC_UNLINK_REQUEST, true /* retryOnError */ );
    1408       63372 : }
    1409             : 
    1410             : /*
    1411             :  * register_forget_request() -- forget any fsyncs for a relation fork's segment
    1412             :  */
    1413             : static void
    1414      243244 : register_forget_request(RelFileLocatorBackend rlocator, ForkNumber forknum,
    1415             :                         BlockNumber segno)
    1416             : {
    1417             :     FileTag     tag;
    1418             : 
    1419      243244 :     INIT_MD_FILETAG(tag, rlocator.locator, forknum, segno);
    1420             : 
    1421      243244 :     RegisterSyncRequest(&tag, SYNC_FORGET_REQUEST, true /* retryOnError */ );
    1422      243244 : }
    1423             : 
    1424             : /*
    1425             :  * ForgetDatabaseSyncRequests -- forget any fsyncs and unlinks for a DB
    1426             :  */
    1427             : void
    1428          84 : ForgetDatabaseSyncRequests(Oid dbid)
    1429             : {
    1430             :     FileTag     tag;
    1431             :     RelFileLocator rlocator;
    1432             : 
    1433          84 :     rlocator.dbOid = dbid;
    1434          84 :     rlocator.spcOid = 0;
    1435          84 :     rlocator.relNumber = 0;
    1436             : 
    1437          84 :     INIT_MD_FILETAG(tag, rlocator, InvalidForkNumber, InvalidBlockNumber);
    1438             : 
    1439          84 :     RegisterSyncRequest(&tag, SYNC_FILTER_REQUEST, true /* retryOnError */ );
    1440          84 : }
    1441             : 
    1442             : /*
    1443             :  * DropRelationFiles -- drop files of all given relations
    1444             :  */
    1445             : void
    1446        4800 : DropRelationFiles(RelFileLocator *delrels, int ndelrels, bool isRedo)
    1447             : {
    1448             :     SMgrRelation *srels;
    1449             :     int         i;
    1450             : 
    1451        4800 :     srels = palloc(sizeof(SMgrRelation) * ndelrels);
    1452       18090 :     for (i = 0; i < ndelrels; i++)
    1453             :     {
    1454       13290 :         SMgrRelation srel = smgropen(delrels[i], INVALID_PROC_NUMBER);
    1455             : 
    1456       13290 :         if (isRedo)
    1457             :         {
    1458             :             ForkNumber  fork;
    1459             : 
    1460       66210 :             for (fork = 0; fork <= MAX_FORKNUM; fork++)
    1461       52968 :                 XLogDropRelation(delrels[i], fork);
    1462             :         }
    1463       13290 :         srels[i] = srel;
    1464             :     }
    1465             : 
    1466        4800 :     smgrdounlinkall(srels, ndelrels, isRedo);
    1467             : 
    1468       18090 :     for (i = 0; i < ndelrels; i++)
    1469       13290 :         smgrclose(srels[i]);
    1470        4800 :     pfree(srels);
    1471        4800 : }
    1472             : 
    1473             : 
    1474             : /*
    1475             :  * _fdvec_resize() -- Resize the fork's open segments array
    1476             :  */
    1477             : static void
    1478     2234014 : _fdvec_resize(SMgrRelation reln,
    1479             :               ForkNumber forknum,
    1480             :               int nseg)
    1481             : {
    1482     2234014 :     if (nseg == 0)
    1483             :     {
    1484      853430 :         if (reln->md_num_open_segs[forknum] > 0)
    1485             :         {
    1486      853430 :             pfree(reln->md_seg_fds[forknum]);
    1487      853430 :             reln->md_seg_fds[forknum] = NULL;
    1488             :         }
    1489             :     }
    1490     1380584 :     else if (reln->md_num_open_segs[forknum] == 0)
    1491             :     {
    1492     1380584 :         reln->md_seg_fds[forknum] =
    1493     1380584 :             MemoryContextAlloc(MdCxt, sizeof(MdfdVec) * nseg);
    1494             :     }
    1495             :     else
    1496             :     {
    1497             :         /*
    1498             :          * It doesn't seem worthwhile complicating the code to amortize
    1499             :          * repalloc() calls.  Those are far faster than PathNameOpenFile() or
    1500             :          * FileClose(), and the memory context internally will sometimes avoid
    1501             :          * doing an actual reallocation.
    1502             :          */
    1503           0 :         reln->md_seg_fds[forknum] =
    1504           0 :             repalloc(reln->md_seg_fds[forknum],
    1505             :                      sizeof(MdfdVec) * nseg);
    1506             :     }
    1507             : 
    1508     2234014 :     reln->md_num_open_segs[forknum] = nseg;
    1509     2234014 : }
    1510             : 
    1511             : /*
    1512             :  * Return the filename for the specified segment of the relation. The
    1513             :  * returned string is palloc'd.
    1514             :  */
    1515             : static char *
    1516       43942 : _mdfd_segpath(SMgrRelation reln, ForkNumber forknum, BlockNumber segno)
    1517             : {
    1518             :     char       *path,
    1519             :                *fullpath;
    1520             : 
    1521       43942 :     path = relpath(reln->smgr_rlocator, forknum);
    1522             : 
    1523       43942 :     if (segno > 0)
    1524             :     {
    1525       43942 :         fullpath = psprintf("%s.%u", path, segno);
    1526       43942 :         pfree(path);
    1527             :     }
    1528             :     else
    1529           0 :         fullpath = path;
    1530             : 
    1531       43942 :     return fullpath;
    1532             : }
    1533             : 
    1534             : /*
    1535             :  * Open the specified segment of the relation,
    1536             :  * and make a MdfdVec object for it.  Returns NULL on failure.
    1537             :  */
    1538             : static MdfdVec *
    1539       43918 : _mdfd_openseg(SMgrRelation reln, ForkNumber forknum, BlockNumber segno,
    1540             :               int oflags)
    1541             : {
    1542             :     MdfdVec    *v;
    1543             :     File        fd;
    1544             :     char       *fullpath;
    1545             : 
    1546       43918 :     fullpath = _mdfd_segpath(reln, forknum, segno);
    1547             : 
    1548             :     /* open the file */
    1549       43918 :     fd = PathNameOpenFile(fullpath, _mdfd_open_flags() | oflags);
    1550             : 
    1551       43918 :     pfree(fullpath);
    1552             : 
    1553       43918 :     if (fd < 0)
    1554       43918 :         return NULL;
    1555             : 
    1556             :     /*
    1557             :      * Segments are always opened in order from lowest to highest, so we must
    1558             :      * be adding a new one at the end.
    1559             :      */
    1560             :     Assert(segno == reln->md_num_open_segs[forknum]);
    1561             : 
    1562           0 :     _fdvec_resize(reln, forknum, segno + 1);
    1563             : 
    1564             :     /* fill the entry */
    1565           0 :     v = &reln->md_seg_fds[forknum][segno];
    1566           0 :     v->mdfd_vfd = fd;
    1567           0 :     v->mdfd_segno = segno;
    1568             : 
    1569             :     Assert(_mdnblocks(reln, forknum, v) <= ((BlockNumber) RELSEG_SIZE));
    1570             : 
    1571             :     /* all done */
    1572           0 :     return v;
    1573             : }
    1574             : 
    1575             : /*
    1576             :  * _mdfd_getseg() -- Find the segment of the relation holding the
    1577             :  *                   specified block.
    1578             :  *
    1579             :  * If the segment doesn't exist, we ereport, return NULL, or create the
    1580             :  * segment, according to "behavior".  Note: skipFsync is only used in the
    1581             :  * EXTENSION_CREATE case.
    1582             :  */
    1583             : static MdfdVec *
    1584     3948260 : _mdfd_getseg(SMgrRelation reln, ForkNumber forknum, BlockNumber blkno,
    1585             :              bool skipFsync, int behavior)
    1586             : {
    1587             :     MdfdVec    *v;
    1588             :     BlockNumber targetseg;
    1589             :     BlockNumber nextsegno;
    1590             : 
    1591             :     /* some way to handle non-existent segments needs to be specified */
    1592             :     Assert(behavior &
    1593             :            (EXTENSION_FAIL | EXTENSION_CREATE | EXTENSION_RETURN_NULL |
    1594             :             EXTENSION_DONT_OPEN));
    1595             : 
    1596     3948260 :     targetseg = blkno / ((BlockNumber) RELSEG_SIZE);
    1597             : 
    1598             :     /* if an existing and opened segment, we're done */
    1599     3948260 :     if (targetseg < reln->md_num_open_segs[forknum])
    1600             :     {
    1601     3607706 :         v = &reln->md_seg_fds[forknum][targetseg];
    1602     3607706 :         return v;
    1603             :     }
    1604             : 
    1605             :     /* The caller only wants the segment if we already had it open. */
    1606      340554 :     if (behavior & EXTENSION_DONT_OPEN)
    1607         652 :         return NULL;
    1608             : 
    1609             :     /*
    1610             :      * The target segment is not yet open. Iterate over all the segments
    1611             :      * between the last opened and the target segment. This way missing
    1612             :      * segments either raise an error, or get created (according to
    1613             :      * 'behavior'). Start with either the last opened, or the first segment if
    1614             :      * none was opened before.
    1615             :      */
    1616      339902 :     if (reln->md_num_open_segs[forknum] > 0)
    1617          24 :         v = &reln->md_seg_fds[forknum][reln->md_num_open_segs[forknum] - 1];
    1618             :     else
    1619             :     {
    1620      339878 :         v = mdopenfork(reln, forknum, behavior);
    1621      339872 :         if (!v)
    1622           0 :             return NULL;        /* if behavior & EXTENSION_RETURN_NULL */
    1623             :     }
    1624             : 
    1625      339896 :     for (nextsegno = reln->md_num_open_segs[forknum];
    1626           0 :          nextsegno <= targetseg; nextsegno++)
    1627             :     {
    1628          24 :         BlockNumber nblocks = _mdnblocks(reln, forknum, v);
    1629          24 :         int         flags = 0;
    1630             : 
    1631             :         Assert(nextsegno == v->mdfd_segno + 1);
    1632             : 
    1633          24 :         if (nblocks > ((BlockNumber) RELSEG_SIZE))
    1634           0 :             elog(FATAL, "segment too big");
    1635             : 
    1636          24 :         if ((behavior & EXTENSION_CREATE) ||
    1637          24 :             (InRecovery && (behavior & EXTENSION_CREATE_RECOVERY)))
    1638             :         {
    1639             :             /*
    1640             :              * Normally we will create new segments only if authorized by the
    1641             :              * caller (i.e., we are doing mdextend()).  But when doing WAL
    1642             :              * recovery, create segments anyway; this allows cases such as
    1643             :              * replaying WAL data that has a write into a high-numbered
    1644             :              * segment of a relation that was later deleted. We want to go
    1645             :              * ahead and create the segments so we can finish out the replay.
    1646             :              *
    1647             :              * We have to maintain the invariant that segments before the last
    1648             :              * active segment are of size RELSEG_SIZE; therefore, if
    1649             :              * extending, pad them out with zeroes if needed.  (This only
    1650             :              * matters if in recovery, or if the caller is extending the
    1651             :              * relation discontiguously, but that can happen in hash indexes.)
    1652             :              */
    1653           0 :             if (nblocks < ((BlockNumber) RELSEG_SIZE))
    1654             :             {
    1655           0 :                 char       *zerobuf = palloc_aligned(BLCKSZ, PG_IO_ALIGN_SIZE,
    1656             :                                                      MCXT_ALLOC_ZERO);
    1657             : 
    1658           0 :                 mdextend(reln, forknum,
    1659           0 :                          nextsegno * ((BlockNumber) RELSEG_SIZE) - 1,
    1660             :                          zerobuf, skipFsync);
    1661           0 :                 pfree(zerobuf);
    1662             :             }
    1663           0 :             flags = O_CREAT;
    1664             :         }
    1665          24 :         else if (!(behavior & EXTENSION_DONT_CHECK_SIZE) &&
    1666             :                  nblocks < ((BlockNumber) RELSEG_SIZE))
    1667             :         {
    1668             :             /*
    1669             :              * When not extending (or explicitly including truncated
    1670             :              * segments), only open the next segment if the current one is
    1671             :              * exactly RELSEG_SIZE.  If not (this branch), either return NULL
    1672             :              * or fail.
    1673             :              */
    1674          24 :             if (behavior & EXTENSION_RETURN_NULL)
    1675             :             {
    1676             :                 /*
    1677             :                  * Some callers discern between reasons for _mdfd_getseg()
    1678             :                  * returning NULL based on errno. As there's no failing
    1679             :                  * syscall involved in this case, explicitly set errno to
    1680             :                  * ENOENT, as that seems the closest interpretation.
    1681             :                  */
    1682           0 :                 errno = ENOENT;
    1683           0 :                 return NULL;
    1684             :             }
    1685             : 
    1686          24 :             ereport(ERROR,
    1687             :                     (errcode_for_file_access(),
    1688             :                      errmsg("could not open file \"%s\" (target block %u): previous segment is only %u blocks",
    1689             :                             _mdfd_segpath(reln, forknum, nextsegno),
    1690             :                             blkno, nblocks)));
    1691             :         }
    1692             : 
    1693           0 :         v = _mdfd_openseg(reln, forknum, nextsegno, flags);
    1694             : 
    1695           0 :         if (v == NULL)
    1696             :         {
    1697           0 :             if ((behavior & EXTENSION_RETURN_NULL) &&
    1698           0 :                 FILE_POSSIBLY_DELETED(errno))
    1699           0 :                 return NULL;
    1700           0 :             ereport(ERROR,
    1701             :                     (errcode_for_file_access(),
    1702             :                      errmsg("could not open file \"%s\" (target block %u): %m",
    1703             :                             _mdfd_segpath(reln, forknum, nextsegno),
    1704             :                             blkno)));
    1705             :         }
    1706             :     }
    1707             : 
    1708      339872 :     return v;
    1709             : }
    1710             : 
    1711             : /*
    1712             :  * Get number of blocks present in a single disk file
    1713             :  */
    1714             : static BlockNumber
    1715     3484886 : _mdnblocks(SMgrRelation reln, ForkNumber forknum, MdfdVec *seg)
    1716             : {
    1717             :     off_t       len;
    1718             : 
    1719     3484886 :     len = FileSize(seg->mdfd_vfd);
    1720     3484886 :     if (len < 0)
    1721           0 :         ereport(ERROR,
    1722             :                 (errcode_for_file_access(),
    1723             :                  errmsg("could not seek to end of file \"%s\": %m",
    1724             :                         FilePathName(seg->mdfd_vfd))));
    1725             :     /* note that this calculation will ignore any partial block at EOF */
    1726     3484886 :     return (BlockNumber) (len / BLCKSZ);
    1727             : }
    1728             : 
    1729             : /*
    1730             :  * Sync a file to disk, given a file tag.  Write the path into an output
    1731             :  * buffer so the caller can use it in error messages.
    1732             :  *
    1733             :  * Return 0 on success, -1 on failure, with errno set.
    1734             :  */
    1735             : int
    1736           0 : mdsyncfiletag(const FileTag *ftag, char *path)
    1737             : {
    1738           0 :     SMgrRelation reln = smgropen(ftag->rlocator, INVALID_PROC_NUMBER);
    1739             :     File        file;
    1740             :     instr_time  io_start;
    1741             :     bool        need_to_close;
    1742             :     int         result,
    1743             :                 save_errno;
    1744             : 
    1745             :     /* See if we already have the file open, or need to open it. */
    1746           0 :     if (ftag->segno < reln->md_num_open_segs[ftag->forknum])
    1747             :     {
    1748           0 :         file = reln->md_seg_fds[ftag->forknum][ftag->segno].mdfd_vfd;
    1749           0 :         strlcpy(path, FilePathName(file), MAXPGPATH);
    1750           0 :         need_to_close = false;
    1751             :     }
    1752             :     else
    1753             :     {
    1754             :         char       *p;
    1755             : 
    1756           0 :         p = _mdfd_segpath(reln, ftag->forknum, ftag->segno);
    1757           0 :         strlcpy(path, p, MAXPGPATH);
    1758           0 :         pfree(p);
    1759             : 
    1760           0 :         file = PathNameOpenFile(path, _mdfd_open_flags());
    1761           0 :         if (file < 0)
    1762           0 :             return -1;
    1763           0 :         need_to_close = true;
    1764             :     }
    1765             : 
    1766           0 :     io_start = pgstat_prepare_io_time(track_io_timing);
    1767             : 
    1768             :     /* Sync the file. */
    1769           0 :     result = FileSync(file, WAIT_EVENT_DATA_FILE_SYNC);
    1770           0 :     save_errno = errno;
    1771             : 
    1772           0 :     if (need_to_close)
    1773           0 :         FileClose(file);
    1774             : 
    1775           0 :     pgstat_count_io_op_time(IOOBJECT_RELATION, IOCONTEXT_NORMAL,
    1776             :                             IOOP_FSYNC, io_start, 1);
    1777             : 
    1778           0 :     errno = save_errno;
    1779           0 :     return result;
    1780             : }
    1781             : 
    1782             : /*
    1783             :  * Unlink a file, given a file tag.  Write the path into an output
    1784             :  * buffer so the caller can use it in error messages.
    1785             :  *
    1786             :  * Return 0 on success, -1 on failure, with errno set.
    1787             :  */
    1788             : int
    1789       59248 : mdunlinkfiletag(const FileTag *ftag, char *path)
    1790             : {
    1791             :     char       *p;
    1792             : 
    1793             :     /* Compute the path. */
    1794       59248 :     p = relpathperm(ftag->rlocator, MAIN_FORKNUM);
    1795       59248 :     strlcpy(path, p, MAXPGPATH);
    1796       59248 :     pfree(p);
    1797             : 
    1798             :     /* Try to unlink the file. */
    1799       59248 :     return unlink(path);
    1800             : }
    1801             : 
    1802             : /*
    1803             :  * Check if a given candidate request matches a given tag, when processing
    1804             :  * a SYNC_FILTER_REQUEST request.  This will be called for all pending
    1805             :  * requests to find out whether to forget them.
    1806             :  */
    1807             : bool
    1808       13524 : mdfiletagmatches(const FileTag *ftag, const FileTag *candidate)
    1809             : {
    1810             :     /*
    1811             :      * For now we only use filter requests as a way to drop all scheduled
    1812             :      * callbacks relating to a given database, when dropping the database.
    1813             :      * We'll return true for all candidates that have the same database OID as
    1814             :      * the ftag from the SYNC_FILTER_REQUEST request, so they're forgotten.
    1815             :      */
    1816       13524 :     return ftag->rlocator.dbOid == candidate->rlocator.dbOid;
    1817             : }

Generated by: LCOV version 1.14