LCOV - code coverage report
Current view: top level - src/backend/storage/smgr - md.c (source / functions) Hit Total Coverage
Test: PostgreSQL 18devel Lines: 372 497 74.8 %
Date: 2025-04-24 13:15:39 Functions: 37 39 94.9 %
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) 1996-2025, 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/aio.h"
      35             : #include "storage/bufmgr.h"
      36             : #include "storage/fd.h"
      37             : #include "storage/md.h"
      38             : #include "storage/relfilelocator.h"
      39             : #include "storage/smgr.h"
      40             : #include "storage/sync.h"
      41             : #include "utils/memutils.h"
      42             : 
      43             : /*
      44             :  * The magnetic disk storage manager keeps track of open file
      45             :  * descriptors in its own descriptor pool.  This is done to make it
      46             :  * easier to support relations that are larger than the operating
      47             :  * system's file size limit (often 2GBytes).  In order to do that,
      48             :  * we break relations up into "segment" files that are each shorter than
      49             :  * the OS file size limit.  The segment size is set by the RELSEG_SIZE
      50             :  * configuration constant in pg_config.h.
      51             :  *
      52             :  * On disk, a relation must consist of consecutively numbered segment
      53             :  * files in the pattern
      54             :  *  -- Zero or more full segments of exactly RELSEG_SIZE blocks each
      55             :  *  -- Exactly one partial segment of size 0 <= size < RELSEG_SIZE blocks
      56             :  *  -- Optionally, any number of inactive segments of size 0 blocks.
      57             :  * The full and partial segments are collectively the "active" segments.
      58             :  * Inactive segments are those that once contained data but are currently
      59             :  * not needed because of an mdtruncate() operation.  The reason for leaving
      60             :  * them present at size zero, rather than unlinking them, is that other
      61             :  * backends and/or the checkpointer might be holding open file references to
      62             :  * such segments.  If the relation expands again after mdtruncate(), such
      63             :  * that a deactivated segment becomes active again, it is important that
      64             :  * such file references still be valid --- else data might get written
      65             :  * out to an unlinked old copy of a segment file that will eventually
      66             :  * disappear.
      67             :  *
      68             :  * File descriptors are stored in the per-fork md_seg_fds arrays inside
      69             :  * SMgrRelation. The length of these arrays is stored in md_num_open_segs.
      70             :  * Note that a fork's md_num_open_segs having a specific value does not
      71             :  * necessarily mean the relation doesn't have additional segments; we may
      72             :  * just not have opened the next segment yet.  (We could not have "all
      73             :  * segments are in the array" as an invariant anyway, since another backend
      74             :  * could extend the relation while we aren't looking.)  We do not have
      75             :  * entries for inactive segments, however; as soon as we find a partial
      76             :  * segment, we assume that any subsequent segments are inactive.
      77             :  *
      78             :  * The entire MdfdVec array is palloc'd in the MdCxt memory context.
      79             :  */
      80             : 
      81             : typedef struct _MdfdVec
      82             : {
      83             :     File        mdfd_vfd;       /* fd number in fd.c's pool */
      84             :     BlockNumber mdfd_segno;     /* segment number, from 0 */
      85             : } MdfdVec;
      86             : 
      87             : static MemoryContext MdCxt;     /* context for all MdfdVec objects */
      88             : 
      89             : 
      90             : /* Populate a file tag describing an md.c segment file. */
      91             : #define INIT_MD_FILETAG(a,xx_rlocator,xx_forknum,xx_segno) \
      92             : ( \
      93             :     memset(&(a), 0, sizeof(FileTag)), \
      94             :     (a).handler = SYNC_HANDLER_MD, \
      95             :     (a).rlocator = (xx_rlocator), \
      96             :     (a).forknum = (xx_forknum), \
      97             :     (a).segno = (xx_segno) \
      98             : )
      99             : 
     100             : 
     101             : /*** behavior for mdopen & _mdfd_getseg ***/
     102             : /* ereport if segment not present */
     103             : #define EXTENSION_FAIL              (1 << 0)
     104             : /* return NULL if segment not present */
     105             : #define EXTENSION_RETURN_NULL       (1 << 1)
     106             : /* create new segments as needed */
     107             : #define EXTENSION_CREATE            (1 << 2)
     108             : /* create new segments if needed during recovery */
     109             : #define EXTENSION_CREATE_RECOVERY   (1 << 3)
     110             : /* don't try to open a segment, if not already open */
     111             : #define EXTENSION_DONT_OPEN         (1 << 5)
     112             : 
     113             : 
     114             : /*
     115             :  * Fixed-length string to represent paths to files that need to be built by
     116             :  * md.c.
     117             :  *
     118             :  * The maximum number of segments is MaxBlockNumber / RELSEG_SIZE, where
     119             :  * RELSEG_SIZE can be set to 1 (for testing only).
     120             :  */
     121             : #define SEGMENT_CHARS   OIDCHARS
     122             : #define MD_PATH_STR_MAXLEN \
     123             :     (\
     124             :         REL_PATH_STR_MAXLEN \
     125             :         + sizeof((char)'.') \
     126             :         + SEGMENT_CHARS \
     127             :     )
     128             : typedef struct MdPathStr
     129             : {
     130             :     char        str[MD_PATH_STR_MAXLEN + 1];
     131             : } MdPathStr;
     132             : 
     133             : 
     134             : /* local routines */
     135             : static void mdunlinkfork(RelFileLocatorBackend rlocator, ForkNumber forknum,
     136             :                          bool isRedo);
     137             : static MdfdVec *mdopenfork(SMgrRelation reln, ForkNumber forknum, int behavior);
     138             : static void register_dirty_segment(SMgrRelation reln, ForkNumber forknum,
     139             :                                    MdfdVec *seg);
     140             : static void register_unlink_segment(RelFileLocatorBackend rlocator, ForkNumber forknum,
     141             :                                     BlockNumber segno);
     142             : static void register_forget_request(RelFileLocatorBackend rlocator, ForkNumber forknum,
     143             :                                     BlockNumber segno);
     144             : static void _fdvec_resize(SMgrRelation reln,
     145             :                           ForkNumber forknum,
     146             :                           int nseg);
     147             : static MdPathStr _mdfd_segpath(SMgrRelation reln, ForkNumber forknum,
     148             :                                BlockNumber segno);
     149             : static MdfdVec *_mdfd_openseg(SMgrRelation reln, ForkNumber forknum,
     150             :                               BlockNumber segno, int oflags);
     151             : static MdfdVec *_mdfd_getseg(SMgrRelation reln, ForkNumber forknum,
     152             :                              BlockNumber blkno, bool skipFsync, int behavior);
     153             : static BlockNumber _mdnblocks(SMgrRelation reln, ForkNumber forknum,
     154             :                               MdfdVec *seg);
     155             : 
     156             : static PgAioResult md_readv_complete(PgAioHandle *ioh, PgAioResult prior_result, uint8 cb_data);
     157             : static void md_readv_report(PgAioResult result, const PgAioTargetData *td, int elevel);
     158             : 
     159             : const PgAioHandleCallbacks aio_md_readv_cb = {
     160             :     .complete_shared = md_readv_complete,
     161             :     .report = md_readv_report,
     162             : };
     163             : 
     164             : 
     165             : static inline int
     166     2812944 : _mdfd_open_flags(void)
     167             : {
     168     2812944 :     int         flags = O_RDWR | PG_BINARY;
     169             : 
     170     2812944 :     if (io_direct_flags & IO_DIRECT_DATA)
     171         622 :         flags |= PG_O_DIRECT;
     172             : 
     173     2812944 :     return flags;
     174             : }
     175             : 
     176             : /*
     177             :  * mdinit() -- Initialize private state for magnetic disk storage manager.
     178             :  */
     179             : void
     180       43174 : mdinit(void)
     181             : {
     182       43174 :     MdCxt = AllocSetContextCreate(TopMemoryContext,
     183             :                                   "MdSmgr",
     184             :                                   ALLOCSET_DEFAULT_SIZES);
     185       43174 : }
     186             : 
     187             : /*
     188             :  * mdexists() -- Does the physical file exist?
     189             :  *
     190             :  * Note: this will return true for lingering files, with pending deletions
     191             :  */
     192             : bool
     193     1170728 : mdexists(SMgrRelation reln, ForkNumber forknum)
     194             : {
     195             :     /*
     196             :      * Close it first, to ensure that we notice if the fork has been unlinked
     197             :      * since we opened it.  As an optimization, we can skip that in recovery,
     198             :      * which already closes relations when dropping them.
     199             :      */
     200     1170728 :     if (!InRecovery)
     201     1129320 :         mdclose(reln, forknum);
     202             : 
     203     1170728 :     return (mdopenfork(reln, forknum, EXTENSION_RETURN_NULL) != NULL);
     204             : }
     205             : 
     206             : /*
     207             :  * mdcreate() -- Create a new relation on magnetic disk.
     208             :  *
     209             :  * If isRedo is true, it's okay for the relation to exist already.
     210             :  */
     211             : void
     212    11233076 : mdcreate(SMgrRelation reln, ForkNumber forknum, bool isRedo)
     213             : {
     214             :     MdfdVec    *mdfd;
     215             :     RelPathStr  path;
     216             :     File        fd;
     217             : 
     218    11233076 :     if (isRedo && reln->md_num_open_segs[forknum] > 0)
     219    10913004 :         return;                 /* created and opened already... */
     220             : 
     221             :     Assert(reln->md_num_open_segs[forknum] == 0);
     222             : 
     223             :     /*
     224             :      * We may be using the target table space for the first time in this
     225             :      * database, so create a per-database subdirectory if needed.
     226             :      *
     227             :      * XXX this is a fairly ugly violation of module layering, but this seems
     228             :      * to be the best place to put the check.  Maybe TablespaceCreateDbspace
     229             :      * should be here and not in commands/tablespace.c?  But that would imply
     230             :      * importing a lot of stuff that smgr.c oughtn't know, either.
     231             :      */
     232      320072 :     TablespaceCreateDbspace(reln->smgr_rlocator.locator.spcOid,
     233             :                             reln->smgr_rlocator.locator.dbOid,
     234             :                             isRedo);
     235             : 
     236      320072 :     path = relpath(reln->smgr_rlocator, forknum);
     237             : 
     238      320072 :     fd = PathNameOpenFile(path.str, _mdfd_open_flags() | O_CREAT | O_EXCL);
     239             : 
     240      320072 :     if (fd < 0)
     241             :     {
     242        8838 :         int         save_errno = errno;
     243             : 
     244        8838 :         if (isRedo)
     245        8838 :             fd = PathNameOpenFile(path.str, _mdfd_open_flags());
     246        8838 :         if (fd < 0)
     247             :         {
     248             :             /* be sure to report the error reported by create, not open */
     249           0 :             errno = save_errno;
     250           0 :             ereport(ERROR,
     251             :                     (errcode_for_file_access(),
     252             :                      errmsg("could not create file \"%s\": %m", path.str)));
     253             :         }
     254             :     }
     255             : 
     256      320072 :     _fdvec_resize(reln, forknum, 1);
     257      320072 :     mdfd = &reln->md_seg_fds[forknum][0];
     258      320072 :     mdfd->mdfd_vfd = fd;
     259      320072 :     mdfd->mdfd_segno = 0;
     260             : 
     261      320072 :     if (!SmgrIsTemp(reln))
     262      313548 :         register_dirty_segment(reln, forknum, mdfd);
     263             : }
     264             : 
     265             : /*
     266             :  * mdunlink() -- Unlink a relation.
     267             :  *
     268             :  * Note that we're passed a RelFileLocatorBackend --- by the time this is called,
     269             :  * there won't be an SMgrRelation hashtable entry anymore.
     270             :  *
     271             :  * forknum can be a fork number to delete a specific fork, or InvalidForkNumber
     272             :  * to delete all forks.
     273             :  *
     274             :  * For regular relations, we don't unlink the first segment file of the rel,
     275             :  * but just truncate it to zero length, and record a request to unlink it after
     276             :  * the next checkpoint.  Additional segments can be unlinked immediately,
     277             :  * however.  Leaving the empty file in place prevents that relfilenumber
     278             :  * from being reused.  The scenario this protects us from is:
     279             :  * 1. We delete a relation (and commit, and actually remove its file).
     280             :  * 2. We create a new relation, which by chance gets the same relfilenumber as
     281             :  *    the just-deleted one (OIDs must've wrapped around for that to happen).
     282             :  * 3. We crash before another checkpoint occurs.
     283             :  * During replay, we would delete the file and then recreate it, which is fine
     284             :  * if the contents of the file were repopulated by subsequent WAL entries.
     285             :  * But if we didn't WAL-log insertions, but instead relied on fsyncing the
     286             :  * file after populating it (as we do at wal_level=minimal), the contents of
     287             :  * the file would be lost forever.  By leaving the empty file until after the
     288             :  * next checkpoint, we prevent reassignment of the relfilenumber until it's
     289             :  * safe, because relfilenumber assignment skips over any existing file.
     290             :  *
     291             :  * Additional segments, if any, are truncated and then unlinked.  The reason
     292             :  * for truncating is that other backends may still hold open FDs for these at
     293             :  * the smgr level, so that the kernel can't remove the file yet.  We want to
     294             :  * reclaim the disk space right away despite that.
     295             :  *
     296             :  * We do not need to go through this dance for temp relations, though, because
     297             :  * we never make WAL entries for temp rels, and so a temp rel poses no threat
     298             :  * to the health of a regular rel that has taken over its relfilenumber.
     299             :  * The fact that temp rels and regular rels have different file naming
     300             :  * patterns provides additional safety.  Other backends shouldn't have open
     301             :  * FDs for them, either.
     302             :  *
     303             :  * We also don't do it while performing a binary upgrade.  There is no reuse
     304             :  * hazard in that case, since after a crash or even a simple ERROR, the
     305             :  * upgrade fails and the whole cluster must be recreated from scratch.
     306             :  * Furthermore, it is important to remove the files from disk immediately,
     307             :  * because we may be about to reuse the same relfilenumber.
     308             :  *
     309             :  * All the above applies only to the relation's main fork; other forks can
     310             :  * just be removed immediately, since they are not needed to prevent the
     311             :  * relfilenumber from being recycled.  Also, we do not carefully
     312             :  * track whether other forks have been created or not, but just attempt to
     313             :  * unlink them unconditionally; so we should never complain about ENOENT.
     314             :  *
     315             :  * If isRedo is true, it's unsurprising for the relation to be already gone.
     316             :  * Also, we should remove the file immediately instead of queuing a request
     317             :  * for later, since during redo there's no possibility of creating a
     318             :  * conflicting relation.
     319             :  *
     320             :  * Note: we currently just never warn about ENOENT at all.  We could warn in
     321             :  * the main-fork, non-isRedo case, but it doesn't seem worth the trouble.
     322             :  *
     323             :  * Note: any failure should be reported as WARNING not ERROR, because
     324             :  * we are usually not in a transaction anymore when this is called.
     325             :  */
     326             : void
     327      370504 : mdunlink(RelFileLocatorBackend rlocator, ForkNumber forknum, bool isRedo)
     328             : {
     329             :     /* Now do the per-fork work */
     330      370504 :     if (forknum == InvalidForkNumber)
     331             :     {
     332           0 :         for (forknum = 0; forknum <= MAX_FORKNUM; forknum++)
     333           0 :             mdunlinkfork(rlocator, forknum, isRedo);
     334             :     }
     335             :     else
     336      370504 :         mdunlinkfork(rlocator, forknum, isRedo);
     337      370504 : }
     338             : 
     339             : /*
     340             :  * Truncate a file to release disk space.
     341             :  */
     342             : static int
     343      435834 : do_truncate(const char *path)
     344             : {
     345             :     int         save_errno;
     346             :     int         ret;
     347             : 
     348      435834 :     ret = pg_truncate(path, 0);
     349             : 
     350             :     /* Log a warning here to avoid repetition in callers. */
     351      435834 :     if (ret < 0 && errno != ENOENT)
     352             :     {
     353           0 :         save_errno = errno;
     354           0 :         ereport(WARNING,
     355             :                 (errcode_for_file_access(),
     356             :                  errmsg("could not truncate file \"%s\": %m", path)));
     357           0 :         errno = save_errno;
     358             :     }
     359             : 
     360      435834 :     return ret;
     361             : }
     362             : 
     363             : static void
     364      370504 : mdunlinkfork(RelFileLocatorBackend rlocator, ForkNumber forknum, bool isRedo)
     365             : {
     366             :     RelPathStr  path;
     367             :     int         ret;
     368             :     int         save_errno;
     369             : 
     370      370504 :     path = relpath(rlocator, forknum);
     371             : 
     372             :     /*
     373             :      * Truncate and then unlink the first segment, or just register a request
     374             :      * to unlink it later, as described in the comments for mdunlink().
     375             :      */
     376      370504 :     if (isRedo || IsBinaryUpgrade || forknum != MAIN_FORKNUM ||
     377       77584 :         RelFileLocatorBackendIsTemp(rlocator))
     378             :     {
     379      299114 :         if (!RelFileLocatorBackendIsTemp(rlocator))
     380             :         {
     381             :             /* Prevent other backends' fds from holding on to the disk space */
     382      274338 :             ret = do_truncate(path.str);
     383             : 
     384             :             /* Forget any pending sync requests for the first segment */
     385      274338 :             save_errno = errno;
     386      274338 :             register_forget_request(rlocator, forknum, 0 /* first seg */ );
     387      274338 :             errno = save_errno;
     388             :         }
     389             :         else
     390       24776 :             ret = 0;
     391             : 
     392             :         /* Next unlink the file, unless it was already found to be missing */
     393      299114 :         if (ret >= 0 || errno != ENOENT)
     394             :         {
     395       43500 :             ret = unlink(path.str);
     396       43500 :             if (ret < 0 && errno != ENOENT)
     397             :             {
     398           0 :                 save_errno = errno;
     399           0 :                 ereport(WARNING,
     400             :                         (errcode_for_file_access(),
     401             :                          errmsg("could not remove file \"%s\": %m", path.str)));
     402           0 :                 errno = save_errno;
     403             :             }
     404             :         }
     405             :     }
     406             :     else
     407             :     {
     408             :         /* Prevent other backends' fds from holding on to the disk space */
     409       71390 :         ret = do_truncate(path.str);
     410             : 
     411             :         /* Register request to unlink first segment later */
     412       71390 :         save_errno = errno;
     413       71390 :         register_unlink_segment(rlocator, forknum, 0 /* first seg */ );
     414       71390 :         errno = save_errno;
     415             :     }
     416             : 
     417             :     /*
     418             :      * Delete any additional segments.
     419             :      *
     420             :      * Note that because we loop until getting ENOENT, we will correctly
     421             :      * remove all inactive segments as well as active ones.  Ideally we'd
     422             :      * continue the loop until getting exactly that errno, but that risks an
     423             :      * infinite loop if the problem is directory-wide (for instance, if we
     424             :      * suddenly can't read the data directory itself).  We compromise by
     425             :      * continuing after a non-ENOENT truncate error, but stopping after any
     426             :      * unlink error.  If there is indeed a directory-wide problem, additional
     427             :      * unlink attempts wouldn't work anyway.
     428             :      */
     429      370504 :     if (ret >= 0 || errno != ENOENT)
     430             :     {
     431             :         MdPathStr   segpath;
     432             :         BlockNumber segno;
     433             : 
     434       96630 :         for (segno = 1;; segno++)
     435             :         {
     436       96630 :             sprintf(segpath.str, "%s.%u", path.str, segno);
     437             : 
     438       96630 :             if (!RelFileLocatorBackendIsTemp(rlocator))
     439             :             {
     440             :                 /*
     441             :                  * Prevent other backends' fds from holding on to the disk
     442             :                  * space.  We're done if we see ENOENT, though.
     443             :                  */
     444       90106 :                 if (do_truncate(segpath.str) < 0 && errno == ENOENT)
     445       90106 :                     break;
     446             : 
     447             :                 /*
     448             :                  * Forget any pending sync requests for this segment before we
     449             :                  * try to unlink.
     450             :                  */
     451           0 :                 register_forget_request(rlocator, forknum, segno);
     452             :             }
     453             : 
     454        6524 :             if (unlink(segpath.str) < 0)
     455             :             {
     456             :                 /* ENOENT is expected after the last segment... */
     457        6524 :                 if (errno != ENOENT)
     458           0 :                     ereport(WARNING,
     459             :                             (errcode_for_file_access(),
     460             :                              errmsg("could not remove file \"%s\": %m", segpath.str)));
     461        6524 :                 break;
     462             :             }
     463             :         }
     464             :     }
     465      370504 : }
     466             : 
     467             : /*
     468             :  * mdextend() -- Add a block to the specified relation.
     469             :  *
     470             :  * The semantics are nearly the same as mdwrite(): write at the
     471             :  * specified position.  However, this is to be used for the case of
     472             :  * extending a relation (i.e., blocknum is at or beyond the current
     473             :  * EOF).  Note that we assume writing a block beyond current EOF
     474             :  * causes intervening file space to become filled with zeroes.
     475             :  */
     476             : void
     477      240658 : mdextend(SMgrRelation reln, ForkNumber forknum, BlockNumber blocknum,
     478             :          const void *buffer, bool skipFsync)
     479             : {
     480             :     off_t       seekpos;
     481             :     int         nbytes;
     482             :     MdfdVec    *v;
     483             : 
     484             :     /* If this build supports direct I/O, the buffer must be I/O aligned. */
     485             :     if (PG_O_DIRECT != 0 && PG_IO_ALIGN_SIZE <= BLCKSZ)
     486             :         Assert((uintptr_t) buffer == TYPEALIGN(PG_IO_ALIGN_SIZE, buffer));
     487             : 
     488             :     /* This assert is too expensive to have on normally ... */
     489             : #ifdef CHECK_WRITE_VS_EXTEND
     490             :     Assert(blocknum >= mdnblocks(reln, forknum));
     491             : #endif
     492             : 
     493             :     /*
     494             :      * If a relation manages to grow to 2^32-1 blocks, refuse to extend it any
     495             :      * more --- we mustn't create a block whose number actually is
     496             :      * InvalidBlockNumber.  (Note that this failure should be unreachable
     497             :      * because of upstream checks in bufmgr.c.)
     498             :      */
     499      240658 :     if (blocknum == InvalidBlockNumber)
     500           0 :         ereport(ERROR,
     501             :                 (errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED),
     502             :                  errmsg("cannot extend file \"%s\" beyond %u blocks",
     503             :                         relpath(reln->smgr_rlocator, forknum).str,
     504             :                         InvalidBlockNumber)));
     505             : 
     506      240658 :     v = _mdfd_getseg(reln, forknum, blocknum, skipFsync, EXTENSION_CREATE);
     507             : 
     508      240658 :     seekpos = (off_t) BLCKSZ * (blocknum % ((BlockNumber) RELSEG_SIZE));
     509             : 
     510             :     Assert(seekpos < (off_t) BLCKSZ * RELSEG_SIZE);
     511             : 
     512      240658 :     if ((nbytes = FileWrite(v->mdfd_vfd, buffer, BLCKSZ, seekpos, WAIT_EVENT_DATA_FILE_EXTEND)) != BLCKSZ)
     513             :     {
     514           0 :         if (nbytes < 0)
     515           0 :             ereport(ERROR,
     516             :                     (errcode_for_file_access(),
     517             :                      errmsg("could not extend file \"%s\": %m",
     518             :                             FilePathName(v->mdfd_vfd)),
     519             :                      errhint("Check free disk space.")));
     520             :         /* short write: complain appropriately */
     521           0 :         ereport(ERROR,
     522             :                 (errcode(ERRCODE_DISK_FULL),
     523             :                  errmsg("could not extend file \"%s\": wrote only %d of %d bytes at block %u",
     524             :                         FilePathName(v->mdfd_vfd),
     525             :                         nbytes, BLCKSZ, blocknum),
     526             :                  errhint("Check free disk space.")));
     527             :     }
     528             : 
     529      240658 :     if (!skipFsync && !SmgrIsTemp(reln))
     530          52 :         register_dirty_segment(reln, forknum, v);
     531             : 
     532             :     Assert(_mdnblocks(reln, forknum, v) <= ((BlockNumber) RELSEG_SIZE));
     533      240658 : }
     534             : 
     535             : /*
     536             :  * mdzeroextend() -- Add new zeroed out blocks to the specified relation.
     537             :  *
     538             :  * Similar to mdextend(), except the relation can be extended by multiple
     539             :  * blocks at once and the added blocks will be filled with zeroes.
     540             :  */
     541             : void
     542      423572 : mdzeroextend(SMgrRelation reln, ForkNumber forknum,
     543             :              BlockNumber blocknum, int nblocks, bool skipFsync)
     544             : {
     545             :     MdfdVec    *v;
     546      423572 :     BlockNumber curblocknum = blocknum;
     547      423572 :     int         remblocks = nblocks;
     548             : 
     549             :     Assert(nblocks > 0);
     550             : 
     551             :     /* This assert is too expensive to have on normally ... */
     552             : #ifdef CHECK_WRITE_VS_EXTEND
     553             :     Assert(blocknum >= mdnblocks(reln, forknum));
     554             : #endif
     555             : 
     556             :     /*
     557             :      * If a relation manages to grow to 2^32-1 blocks, refuse to extend it any
     558             :      * more --- we mustn't create a block whose number actually is
     559             :      * InvalidBlockNumber or larger.
     560             :      */
     561      423572 :     if ((uint64) blocknum + nblocks >= (uint64) InvalidBlockNumber)
     562           0 :         ereport(ERROR,
     563             :                 (errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED),
     564             :                  errmsg("cannot extend file \"%s\" beyond %u blocks",
     565             :                         relpath(reln->smgr_rlocator, forknum).str,
     566             :                         InvalidBlockNumber)));
     567             : 
     568      847144 :     while (remblocks > 0)
     569             :     {
     570      423572 :         BlockNumber segstartblock = curblocknum % ((BlockNumber) RELSEG_SIZE);
     571      423572 :         off_t       seekpos = (off_t) BLCKSZ * segstartblock;
     572             :         int         numblocks;
     573             : 
     574      423572 :         if (segstartblock + remblocks > RELSEG_SIZE)
     575           0 :             numblocks = RELSEG_SIZE - segstartblock;
     576             :         else
     577      423572 :             numblocks = remblocks;
     578             : 
     579      423572 :         v = _mdfd_getseg(reln, forknum, curblocknum, skipFsync, EXTENSION_CREATE);
     580             : 
     581             :         Assert(segstartblock < RELSEG_SIZE);
     582             :         Assert(segstartblock + numblocks <= RELSEG_SIZE);
     583             : 
     584             :         /*
     585             :          * If available and useful, use posix_fallocate() (via
     586             :          * FileFallocate()) to extend the relation. That's often more
     587             :          * efficient than using write(), as it commonly won't cause the kernel
     588             :          * to allocate page cache space for the extended pages.
     589             :          *
     590             :          * However, we don't use FileFallocate() for small extensions, as it
     591             :          * defeats delayed allocation on some filesystems. Not clear where
     592             :          * that decision should be made though? For now just use a cutoff of
     593             :          * 8, anything between 4 and 8 worked OK in some local testing.
     594             :          */
     595      423572 :         if (numblocks > 8)
     596             :         {
     597             :             int         ret;
     598             : 
     599        1330 :             ret = FileFallocate(v->mdfd_vfd,
     600             :                                 seekpos, (off_t) BLCKSZ * numblocks,
     601             :                                 WAIT_EVENT_DATA_FILE_EXTEND);
     602        1330 :             if (ret != 0)
     603             :             {
     604           0 :                 ereport(ERROR,
     605             :                         errcode_for_file_access(),
     606             :                         errmsg("could not extend file \"%s\" with FileFallocate(): %m",
     607             :                                FilePathName(v->mdfd_vfd)),
     608             :                         errhint("Check free disk space."));
     609             :             }
     610             :         }
     611             :         else
     612             :         {
     613             :             int         ret;
     614             : 
     615             :             /*
     616             :              * Even if we don't want to use fallocate, we can still extend a
     617             :              * bit more efficiently than writing each 8kB block individually.
     618             :              * pg_pwrite_zeros() (via FileZero()) uses pg_pwritev_with_retry()
     619             :              * to avoid multiple writes or needing a zeroed buffer for the
     620             :              * whole length of the extension.
     621             :              */
     622      422242 :             ret = FileZero(v->mdfd_vfd,
     623             :                            seekpos, (off_t) BLCKSZ * numblocks,
     624             :                            WAIT_EVENT_DATA_FILE_EXTEND);
     625      422242 :             if (ret < 0)
     626           0 :                 ereport(ERROR,
     627             :                         errcode_for_file_access(),
     628             :                         errmsg("could not extend file \"%s\": %m",
     629             :                                FilePathName(v->mdfd_vfd)),
     630             :                         errhint("Check free disk space."));
     631             :         }
     632             : 
     633      423572 :         if (!skipFsync && !SmgrIsTemp(reln))
     634      400842 :             register_dirty_segment(reln, forknum, v);
     635             : 
     636             :         Assert(_mdnblocks(reln, forknum, v) <= ((BlockNumber) RELSEG_SIZE));
     637             : 
     638      423572 :         remblocks -= numblocks;
     639      423572 :         curblocknum += numblocks;
     640             :     }
     641      423572 : }
     642             : 
     643             : /*
     644             :  * mdopenfork() -- Open one fork of the specified relation.
     645             :  *
     646             :  * Note we only open the first segment, when there are multiple segments.
     647             :  *
     648             :  * If first segment is not present, either ereport or return NULL according
     649             :  * to "behavior".  We treat EXTENSION_CREATE the same as EXTENSION_FAIL;
     650             :  * EXTENSION_CREATE means it's OK to extend an existing relation, not to
     651             :  * invent one out of whole cloth.
     652             :  */
     653             : static MdfdVec *
     654     7396324 : mdopenfork(SMgrRelation reln, ForkNumber forknum, int behavior)
     655             : {
     656             :     MdfdVec    *mdfd;
     657             :     RelPathStr  path;
     658             :     File        fd;
     659             : 
     660             :     /* No work if already open */
     661     7396324 :     if (reln->md_num_open_segs[forknum] > 0)
     662     4963248 :         return &reln->md_seg_fds[forknum][0];
     663             : 
     664     2433076 :     path = relpath(reln->smgr_rlocator, forknum);
     665             : 
     666     2433076 :     fd = PathNameOpenFile(path.str, _mdfd_open_flags());
     667             : 
     668     2433076 :     if (fd < 0)
     669             :     {
     670      787468 :         if ((behavior & EXTENSION_RETURN_NULL) &&
     671      787424 :             FILE_POSSIBLY_DELETED(errno))
     672      787424 :             return NULL;
     673          44 :         ereport(ERROR,
     674             :                 (errcode_for_file_access(),
     675             :                  errmsg("could not open file \"%s\": %m", path.str)));
     676             :     }
     677             : 
     678     1645608 :     _fdvec_resize(reln, forknum, 1);
     679     1645608 :     mdfd = &reln->md_seg_fds[forknum][0];
     680     1645608 :     mdfd->mdfd_vfd = fd;
     681     1645608 :     mdfd->mdfd_segno = 0;
     682             : 
     683             :     Assert(_mdnblocks(reln, forknum, mdfd) <= ((BlockNumber) RELSEG_SIZE));
     684             : 
     685     1645608 :     return mdfd;
     686             : }
     687             : 
     688             : /*
     689             :  * mdopen() -- Initialize newly-opened relation.
     690             :  */
     691             : void
     692     2292360 : mdopen(SMgrRelation reln)
     693             : {
     694             :     /* mark it not open */
     695    11461800 :     for (int forknum = 0; forknum <= MAX_FORKNUM; forknum++)
     696     9169440 :         reln->md_num_open_segs[forknum] = 0;
     697     2292360 : }
     698             : 
     699             : /*
     700             :  * mdclose() -- Close the specified relation, if it isn't closed already.
     701             :  */
     702             : void
     703     7296056 : mdclose(SMgrRelation reln, ForkNumber forknum)
     704             : {
     705     7296056 :     int         nopensegs = reln->md_num_open_segs[forknum];
     706             : 
     707             :     /* No work if already closed */
     708     7296056 :     if (nopensegs == 0)
     709     6204552 :         return;
     710             : 
     711             :     /* close segments starting from the end */
     712     2183008 :     while (nopensegs > 0)
     713             :     {
     714     1091504 :         MdfdVec    *v = &reln->md_seg_fds[forknum][nopensegs - 1];
     715             : 
     716     1091504 :         FileClose(v->mdfd_vfd);
     717     1091504 :         _fdvec_resize(reln, forknum, nopensegs - 1);
     718     1091504 :         nopensegs--;
     719             :     }
     720             : }
     721             : 
     722             : /*
     723             :  * mdprefetch() -- Initiate asynchronous read of the specified blocks of a relation
     724             :  */
     725             : bool
     726       16952 : mdprefetch(SMgrRelation reln, ForkNumber forknum, BlockNumber blocknum,
     727             :            int nblocks)
     728             : {
     729             : #ifdef USE_PREFETCH
     730             : 
     731             :     Assert((io_direct_flags & IO_DIRECT_DATA) == 0);
     732             : 
     733       16952 :     if ((uint64) blocknum + nblocks > (uint64) MaxBlockNumber + 1)
     734           0 :         return false;
     735             : 
     736       33904 :     while (nblocks > 0)
     737             :     {
     738             :         off_t       seekpos;
     739             :         MdfdVec    *v;
     740             :         int         nblocks_this_segment;
     741             : 
     742       16952 :         v = _mdfd_getseg(reln, forknum, blocknum, false,
     743       16952 :                          InRecovery ? EXTENSION_RETURN_NULL : EXTENSION_FAIL);
     744       16952 :         if (v == NULL)
     745           0 :             return false;
     746             : 
     747       16952 :         seekpos = (off_t) BLCKSZ * (blocknum % ((BlockNumber) RELSEG_SIZE));
     748             : 
     749             :         Assert(seekpos < (off_t) BLCKSZ * RELSEG_SIZE);
     750             : 
     751       16952 :         nblocks_this_segment =
     752       16952 :             Min(nblocks,
     753             :                 RELSEG_SIZE - (blocknum % ((BlockNumber) RELSEG_SIZE)));
     754             : 
     755       16952 :         (void) FilePrefetch(v->mdfd_vfd, seekpos, BLCKSZ * nblocks_this_segment,
     756             :                             WAIT_EVENT_DATA_FILE_PREFETCH);
     757             : 
     758       16952 :         blocknum += nblocks_this_segment;
     759       16952 :         nblocks -= nblocks_this_segment;
     760             :     }
     761             : #endif                          /* USE_PREFETCH */
     762             : 
     763       16952 :     return true;
     764             : }
     765             : 
     766             : /*
     767             :  * Convert an array of buffer address into an array of iovec objects, and
     768             :  * return the number that were required.  'iov' must have enough space for up
     769             :  * to 'nblocks' elements, but the number used may be less depending on
     770             :  * merging.  In the case of a run of fully contiguous buffers, a single iovec
     771             :  * will be populated that can be handled as a plain non-vectored I/O.
     772             :  */
     773             : static int
     774     3540118 : buffers_to_iovec(struct iovec *iov, void **buffers, int nblocks)
     775             : {
     776             :     struct iovec *iovp;
     777             :     int         iovcnt;
     778             : 
     779             :     Assert(nblocks >= 1);
     780             : 
     781             :     /* If this build supports direct I/O, buffers must be I/O aligned. */
     782     7406322 :     for (int i = 0; i < nblocks; ++i)
     783             :     {
     784             :         if (PG_O_DIRECT != 0 && PG_IO_ALIGN_SIZE <= BLCKSZ)
     785             :             Assert((uintptr_t) buffers[i] ==
     786             :                    TYPEALIGN(PG_IO_ALIGN_SIZE, buffers[i]));
     787             :     }
     788             : 
     789             :     /* Start the first iovec off with the first buffer. */
     790     3540118 :     iovp = &iov[0];
     791     3540118 :     iovp->iov_base = buffers[0];
     792     3540118 :     iovp->iov_len = BLCKSZ;
     793     3540118 :     iovcnt = 1;
     794             : 
     795             :     /* Try to merge the rest. */
     796     3866204 :     for (int i = 1; i < nblocks; ++i)
     797             :     {
     798      326086 :         void       *buffer = buffers[i];
     799             : 
     800      326086 :         if (((char *) iovp->iov_base + iovp->iov_len) == buffer)
     801             :         {
     802             :             /* Contiguous with the last iovec. */
     803      317082 :             iovp->iov_len += BLCKSZ;
     804             :         }
     805             :         else
     806             :         {
     807             :             /* Need a new iovec. */
     808        9004 :             iovp++;
     809        9004 :             iovp->iov_base = buffer;
     810        9004 :             iovp->iov_len = BLCKSZ;
     811        9004 :             iovcnt++;
     812             :         }
     813             :     }
     814             : 
     815     3540118 :     return iovcnt;
     816             : }
     817             : 
     818             : /*
     819             :  * mdmaxcombine() -- Return the maximum number of total blocks that can be
     820             :  *               combined with an IO starting at blocknum.
     821             :  */
     822             : uint32
     823       66268 : mdmaxcombine(SMgrRelation reln, ForkNumber forknum,
     824             :              BlockNumber blocknum)
     825             : {
     826             :     BlockNumber segoff;
     827             : 
     828       66268 :     segoff = blocknum % ((BlockNumber) RELSEG_SIZE);
     829             : 
     830       66268 :     return RELSEG_SIZE - segoff;
     831             : }
     832             : 
     833             : /*
     834             :  * mdreadv() -- Read the specified blocks from a relation.
     835             :  */
     836             : void
     837        1196 : mdreadv(SMgrRelation reln, ForkNumber forknum, BlockNumber blocknum,
     838             :         void **buffers, BlockNumber nblocks)
     839             : {
     840        2392 :     while (nblocks > 0)
     841             :     {
     842             :         struct iovec iov[PG_IOV_MAX];
     843             :         int         iovcnt;
     844             :         off_t       seekpos;
     845             :         int         nbytes;
     846             :         MdfdVec    *v;
     847             :         BlockNumber nblocks_this_segment;
     848             :         size_t      transferred_this_segment;
     849             :         size_t      size_this_segment;
     850             : 
     851        1196 :         v = _mdfd_getseg(reln, forknum, blocknum, false,
     852             :                          EXTENSION_FAIL | EXTENSION_CREATE_RECOVERY);
     853             : 
     854        1196 :         seekpos = (off_t) BLCKSZ * (blocknum % ((BlockNumber) RELSEG_SIZE));
     855             : 
     856             :         Assert(seekpos < (off_t) BLCKSZ * RELSEG_SIZE);
     857             : 
     858        1196 :         nblocks_this_segment =
     859        1196 :             Min(nblocks,
     860             :                 RELSEG_SIZE - (blocknum % ((BlockNumber) RELSEG_SIZE)));
     861        1196 :         nblocks_this_segment = Min(nblocks_this_segment, lengthof(iov));
     862             : 
     863        1196 :         if (nblocks_this_segment != nblocks)
     864           0 :             elog(ERROR, "read crosses segment boundary");
     865             : 
     866        1196 :         iovcnt = buffers_to_iovec(iov, buffers, nblocks_this_segment);
     867        1196 :         size_this_segment = nblocks_this_segment * BLCKSZ;
     868        1196 :         transferred_this_segment = 0;
     869             : 
     870             :         /*
     871             :          * Inner loop to continue after a short read.  We'll keep going until
     872             :          * we hit EOF rather than assuming that a short read means we hit the
     873             :          * end.
     874             :          */
     875             :         for (;;)
     876             :         {
     877             :             TRACE_POSTGRESQL_SMGR_MD_READ_START(forknum, blocknum,
     878             :                                                 reln->smgr_rlocator.locator.spcOid,
     879             :                                                 reln->smgr_rlocator.locator.dbOid,
     880             :                                                 reln->smgr_rlocator.locator.relNumber,
     881             :                                                 reln->smgr_rlocator.backend);
     882        1196 :             nbytes = FileReadV(v->mdfd_vfd, iov, iovcnt, seekpos,
     883             :                                WAIT_EVENT_DATA_FILE_READ);
     884             :             TRACE_POSTGRESQL_SMGR_MD_READ_DONE(forknum, blocknum,
     885             :                                                reln->smgr_rlocator.locator.spcOid,
     886             :                                                reln->smgr_rlocator.locator.dbOid,
     887             :                                                reln->smgr_rlocator.locator.relNumber,
     888             :                                                reln->smgr_rlocator.backend,
     889             :                                                nbytes,
     890             :                                                size_this_segment - transferred_this_segment);
     891             : 
     892             : #ifdef SIMULATE_SHORT_READ
     893             :             nbytes = Min(nbytes, 4096);
     894             : #endif
     895             : 
     896        1196 :             if (nbytes < 0)
     897           0 :                 ereport(ERROR,
     898             :                         (errcode_for_file_access(),
     899             :                          errmsg("could not read blocks %u..%u in file \"%s\": %m",
     900             :                                 blocknum,
     901             :                                 blocknum + nblocks_this_segment - 1,
     902             :                                 FilePathName(v->mdfd_vfd))));
     903             : 
     904        1196 :             if (nbytes == 0)
     905             :             {
     906             :                 /*
     907             :                  * We are at or past EOF, or we read a partial block at EOF.
     908             :                  * Normally this is an error; upper levels should never try to
     909             :                  * read a nonexistent block.  However, if zero_damaged_pages
     910             :                  * is ON or we are InRecovery, we should instead return zeroes
     911             :                  * without complaining.  This allows, for example, the case of
     912             :                  * trying to update a block that was later truncated away.
     913             :                  *
     914             :                  * NB: We think that this codepath is unreachable in recovery
     915             :                  * and incomplete with zero_damaged_pages, as missing segments
     916             :                  * are not created. Putting blocks into the buffer-pool that
     917             :                  * do not exist on disk is rather problematic, as it will not
     918             :                  * be found by scans that rely on smgrnblocks(), as they are
     919             :                  * beyond EOF. It also can cause weird problems with relation
     920             :                  * extension, as relation extension does not expect blocks
     921             :                  * beyond EOF to exist.
     922             :                  *
     923             :                  * Therefore we do not want to copy the logic into
     924             :                  * mdstartreadv(), where it would have to be more complicated
     925             :                  * due to potential differences in the zero_damaged_pages
     926             :                  * setting between the definer and completor of IO.
     927             :                  *
     928             :                  * For PG 18, we are putting an Assert(false) in mdreadv()
     929             :                  * (triggering failures in assertion-enabled builds, but
     930             :                  * continuing to work in production builds). Afterwards we
     931             :                  * plan to remove this code entirely.
     932             :                  */
     933           0 :                 if (zero_damaged_pages || InRecovery)
     934             :                 {
     935             :                     Assert(false);  /* see comment above */
     936             : 
     937           0 :                     for (BlockNumber i = transferred_this_segment / BLCKSZ;
     938             :                          i < nblocks_this_segment;
     939           0 :                          ++i)
     940           0 :                         memset(buffers[i], 0, BLCKSZ);
     941           0 :                     break;
     942             :                 }
     943             :                 else
     944           0 :                     ereport(ERROR,
     945             :                             (errcode(ERRCODE_DATA_CORRUPTED),
     946             :                              errmsg("could not read blocks %u..%u in file \"%s\": read only %zu of %zu bytes",
     947             :                                     blocknum,
     948             :                                     blocknum + nblocks_this_segment - 1,
     949             :                                     FilePathName(v->mdfd_vfd),
     950             :                                     transferred_this_segment,
     951             :                                     size_this_segment)));
     952             :             }
     953             : 
     954             :             /* One loop should usually be enough. */
     955        1196 :             transferred_this_segment += nbytes;
     956             :             Assert(transferred_this_segment <= size_this_segment);
     957        1196 :             if (transferred_this_segment == size_this_segment)
     958        1196 :                 break;
     959             : 
     960             :             /* Adjust position and vectors after a short read. */
     961           0 :             seekpos += nbytes;
     962           0 :             iovcnt = compute_remaining_iovec(iov, iov, iovcnt, nbytes);
     963             :         }
     964             : 
     965        1196 :         nblocks -= nblocks_this_segment;
     966        1196 :         buffers += nblocks_this_segment;
     967        1196 :         blocknum += nblocks_this_segment;
     968             :     }
     969        1196 : }
     970             : 
     971             : /*
     972             :  * mdstartreadv() -- Asynchronous version of mdreadv().
     973             :  */
     974             : void
     975     2466050 : mdstartreadv(PgAioHandle *ioh,
     976             :              SMgrRelation reln, ForkNumber forknum, BlockNumber blocknum,
     977             :              void **buffers, BlockNumber nblocks)
     978             : {
     979             :     off_t       seekpos;
     980             :     MdfdVec    *v;
     981             :     BlockNumber nblocks_this_segment;
     982             :     struct iovec *iov;
     983             :     int         iovcnt;
     984             :     int         ret;
     985             : 
     986     2466050 :     v = _mdfd_getseg(reln, forknum, blocknum, false,
     987             :                      EXTENSION_FAIL | EXTENSION_CREATE_RECOVERY);
     988             : 
     989     2466020 :     seekpos = (off_t) BLCKSZ * (blocknum % ((BlockNumber) RELSEG_SIZE));
     990             : 
     991             :     Assert(seekpos < (off_t) BLCKSZ * RELSEG_SIZE);
     992             : 
     993     2466020 :     nblocks_this_segment =
     994     2466020 :         Min(nblocks,
     995             :             RELSEG_SIZE - (blocknum % ((BlockNumber) RELSEG_SIZE)));
     996             : 
     997     2466020 :     if (nblocks_this_segment != nblocks)
     998           0 :         elog(ERROR, "read crossing segment boundary");
     999             : 
    1000     2466020 :     iovcnt = pgaio_io_get_iovec(ioh, &iov);
    1001             : 
    1002             :     Assert(nblocks <= iovcnt);
    1003             : 
    1004     2466020 :     iovcnt = buffers_to_iovec(iov, buffers, nblocks_this_segment);
    1005             : 
    1006             :     Assert(iovcnt <= nblocks_this_segment);
    1007             : 
    1008     2466020 :     if (!(io_direct_flags & IO_DIRECT_DATA))
    1009     2463248 :         pgaio_io_set_flag(ioh, PGAIO_HF_BUFFERED);
    1010             : 
    1011     2466020 :     pgaio_io_set_target_smgr(ioh,
    1012             :                              reln,
    1013             :                              forknum,
    1014             :                              blocknum,
    1015             :                              nblocks,
    1016             :                              false);
    1017     2466020 :     pgaio_io_register_callbacks(ioh, PGAIO_HCB_MD_READV, 0);
    1018             : 
    1019     2466020 :     ret = FileStartReadV(ioh, v->mdfd_vfd, iovcnt, seekpos, WAIT_EVENT_DATA_FILE_READ);
    1020     2466020 :     if (ret != 0)
    1021           0 :         ereport(ERROR,
    1022             :                 (errcode_for_file_access(),
    1023             :                  errmsg("could not start reading blocks %u..%u in file \"%s\": %m",
    1024             :                         blocknum,
    1025             :                         blocknum + nblocks_this_segment - 1,
    1026             :                         FilePathName(v->mdfd_vfd))));
    1027             : 
    1028             :     /*
    1029             :      * The error checks corresponding to the post-read checks in mdreadv() are
    1030             :      * in md_readv_complete().
    1031             :      *
    1032             :      * However we chose, at least for now, to not implement the
    1033             :      * zero_damaged_pages logic present in mdreadv(). As outlined in mdreadv()
    1034             :      * that logic is rather problematic, and we want to get rid of it. Here
    1035             :      * equivalent logic would have to be more complicated due to potential
    1036             :      * differences in the zero_damaged_pages setting between the definer and
    1037             :      * completor of IO.
    1038             :      */
    1039     2466020 : }
    1040             : 
    1041             : /*
    1042             :  * mdwritev() -- Write the supplied blocks at the appropriate location.
    1043             :  *
    1044             :  * This is to be used only for updating already-existing blocks of a
    1045             :  * relation (ie, those before the current EOF).  To extend a relation,
    1046             :  * use mdextend().
    1047             :  */
    1048             : void
    1049     1072902 : mdwritev(SMgrRelation reln, ForkNumber forknum, BlockNumber blocknum,
    1050             :          const void **buffers, BlockNumber nblocks, bool skipFsync)
    1051             : {
    1052             :     /* This assert is too expensive to have on normally ... */
    1053             : #ifdef CHECK_WRITE_VS_EXTEND
    1054             :     Assert((uint64) blocknum + (uint64) nblocks <= (uint64) mdnblocks(reln, forknum));
    1055             : #endif
    1056             : 
    1057     2145804 :     while (nblocks > 0)
    1058             :     {
    1059             :         struct iovec iov[PG_IOV_MAX];
    1060             :         int         iovcnt;
    1061             :         off_t       seekpos;
    1062             :         int         nbytes;
    1063             :         MdfdVec    *v;
    1064             :         BlockNumber nblocks_this_segment;
    1065             :         size_t      transferred_this_segment;
    1066             :         size_t      size_this_segment;
    1067             : 
    1068     1072902 :         v = _mdfd_getseg(reln, forknum, blocknum, skipFsync,
    1069             :                          EXTENSION_FAIL | EXTENSION_CREATE_RECOVERY);
    1070             : 
    1071     1072902 :         seekpos = (off_t) BLCKSZ * (blocknum % ((BlockNumber) RELSEG_SIZE));
    1072             : 
    1073             :         Assert(seekpos < (off_t) BLCKSZ * RELSEG_SIZE);
    1074             : 
    1075     1072902 :         nblocks_this_segment =
    1076     1072902 :             Min(nblocks,
    1077             :                 RELSEG_SIZE - (blocknum % ((BlockNumber) RELSEG_SIZE)));
    1078     1072902 :         nblocks_this_segment = Min(nblocks_this_segment, lengthof(iov));
    1079             : 
    1080     1072902 :         if (nblocks_this_segment != nblocks)
    1081           0 :             elog(ERROR, "write crosses segment boundary");
    1082             : 
    1083     1072902 :         iovcnt = buffers_to_iovec(iov, (void **) buffers, nblocks_this_segment);
    1084     1072902 :         size_this_segment = nblocks_this_segment * BLCKSZ;
    1085     1072902 :         transferred_this_segment = 0;
    1086             : 
    1087             :         /*
    1088             :          * Inner loop to continue after a short write.  If the reason is that
    1089             :          * we're out of disk space, a future attempt should get an ENOSPC
    1090             :          * error from the kernel.
    1091             :          */
    1092             :         for (;;)
    1093             :         {
    1094             :             TRACE_POSTGRESQL_SMGR_MD_WRITE_START(forknum, blocknum,
    1095             :                                                  reln->smgr_rlocator.locator.spcOid,
    1096             :                                                  reln->smgr_rlocator.locator.dbOid,
    1097             :                                                  reln->smgr_rlocator.locator.relNumber,
    1098             :                                                  reln->smgr_rlocator.backend);
    1099     1072902 :             nbytes = FileWriteV(v->mdfd_vfd, iov, iovcnt, seekpos,
    1100             :                                 WAIT_EVENT_DATA_FILE_WRITE);
    1101             :             TRACE_POSTGRESQL_SMGR_MD_WRITE_DONE(forknum, blocknum,
    1102             :                                                 reln->smgr_rlocator.locator.spcOid,
    1103             :                                                 reln->smgr_rlocator.locator.dbOid,
    1104             :                                                 reln->smgr_rlocator.locator.relNumber,
    1105             :                                                 reln->smgr_rlocator.backend,
    1106             :                                                 nbytes,
    1107             :                                                 size_this_segment - transferred_this_segment);
    1108             : 
    1109             : #ifdef SIMULATE_SHORT_WRITE
    1110             :             nbytes = Min(nbytes, 4096);
    1111             : #endif
    1112             : 
    1113     1072902 :             if (nbytes < 0)
    1114             :             {
    1115           0 :                 bool        enospc = errno == ENOSPC;
    1116             : 
    1117           0 :                 ereport(ERROR,
    1118             :                         (errcode_for_file_access(),
    1119             :                          errmsg("could not write blocks %u..%u in file \"%s\": %m",
    1120             :                                 blocknum,
    1121             :                                 blocknum + nblocks_this_segment - 1,
    1122             :                                 FilePathName(v->mdfd_vfd)),
    1123             :                          enospc ? errhint("Check free disk space.") : 0));
    1124             :             }
    1125             : 
    1126             :             /* One loop should usually be enough. */
    1127     1072902 :             transferred_this_segment += nbytes;
    1128             :             Assert(transferred_this_segment <= size_this_segment);
    1129     1072902 :             if (transferred_this_segment == size_this_segment)
    1130     1072902 :                 break;
    1131             : 
    1132             :             /* Adjust position and iovecs after a short write. */
    1133           0 :             seekpos += nbytes;
    1134           0 :             iovcnt = compute_remaining_iovec(iov, iov, iovcnt, nbytes);
    1135             :         }
    1136             : 
    1137     1072902 :         if (!skipFsync && !SmgrIsTemp(reln))
    1138     1064930 :             register_dirty_segment(reln, forknum, v);
    1139             : 
    1140     1072902 :         nblocks -= nblocks_this_segment;
    1141     1072902 :         buffers += nblocks_this_segment;
    1142     1072902 :         blocknum += nblocks_this_segment;
    1143             :     }
    1144     1072902 : }
    1145             : 
    1146             : 
    1147             : /*
    1148             :  * mdwriteback() -- Tell the kernel to write pages back to storage.
    1149             :  *
    1150             :  * This accepts a range of blocks because flushing several pages at once is
    1151             :  * considerably more efficient than doing so individually.
    1152             :  */
    1153             : void
    1154           0 : mdwriteback(SMgrRelation reln, ForkNumber forknum,
    1155             :             BlockNumber blocknum, BlockNumber nblocks)
    1156             : {
    1157             :     Assert((io_direct_flags & IO_DIRECT_DATA) == 0);
    1158             : 
    1159             :     /*
    1160             :      * Issue flush requests in as few requests as possible; have to split at
    1161             :      * segment boundaries though, since those are actually separate files.
    1162             :      */
    1163           0 :     while (nblocks > 0)
    1164             :     {
    1165           0 :         BlockNumber nflush = nblocks;
    1166             :         off_t       seekpos;
    1167             :         MdfdVec    *v;
    1168             :         int         segnum_start,
    1169             :                     segnum_end;
    1170             : 
    1171           0 :         v = _mdfd_getseg(reln, forknum, blocknum, true /* not used */ ,
    1172             :                          EXTENSION_DONT_OPEN);
    1173             : 
    1174             :         /*
    1175             :          * We might be flushing buffers of already removed relations, that's
    1176             :          * ok, just ignore that case.  If the segment file wasn't open already
    1177             :          * (ie from a recent mdwrite()), then we don't want to re-open it, to
    1178             :          * avoid a race with PROCSIGNAL_BARRIER_SMGRRELEASE that might leave
    1179             :          * us with a descriptor to a file that is about to be unlinked.
    1180             :          */
    1181           0 :         if (!v)
    1182           0 :             return;
    1183             : 
    1184             :         /* compute offset inside the current segment */
    1185           0 :         segnum_start = blocknum / RELSEG_SIZE;
    1186             : 
    1187             :         /* compute number of desired writes within the current segment */
    1188           0 :         segnum_end = (blocknum + nblocks - 1) / RELSEG_SIZE;
    1189           0 :         if (segnum_start != segnum_end)
    1190           0 :             nflush = RELSEG_SIZE - (blocknum % ((BlockNumber) RELSEG_SIZE));
    1191             : 
    1192             :         Assert(nflush >= 1);
    1193             :         Assert(nflush <= nblocks);
    1194             : 
    1195           0 :         seekpos = (off_t) BLCKSZ * (blocknum % ((BlockNumber) RELSEG_SIZE));
    1196             : 
    1197           0 :         FileWriteback(v->mdfd_vfd, seekpos, (off_t) BLCKSZ * nflush, WAIT_EVENT_DATA_FILE_FLUSH);
    1198             : 
    1199           0 :         nblocks -= nflush;
    1200           0 :         blocknum += nflush;
    1201             :     }
    1202             : }
    1203             : 
    1204             : /*
    1205             :  * mdnblocks() -- Get the number of blocks stored in a relation.
    1206             :  *
    1207             :  * Important side effect: all active segments of the relation are opened
    1208             :  * and added to the md_seg_fds array.  If this routine has not been
    1209             :  * called, then only segments up to the last one actually touched
    1210             :  * are present in the array.
    1211             :  */
    1212             : BlockNumber
    1213     4842538 : mdnblocks(SMgrRelation reln, ForkNumber forknum)
    1214             : {
    1215             :     MdfdVec    *v;
    1216             :     BlockNumber nblocks;
    1217             :     BlockNumber segno;
    1218             : 
    1219     4842538 :     mdopenfork(reln, forknum, EXTENSION_FAIL);
    1220             : 
    1221             :     /* mdopen has opened the first segment */
    1222             :     Assert(reln->md_num_open_segs[forknum] > 0);
    1223             : 
    1224             :     /*
    1225             :      * Start from the last open segments, to avoid redundant seeks.  We have
    1226             :      * previously verified that these segments are exactly RELSEG_SIZE long,
    1227             :      * and it's useless to recheck that each time.
    1228             :      *
    1229             :      * NOTE: this assumption could only be wrong if another backend has
    1230             :      * truncated the relation.  We rely on higher code levels to handle that
    1231             :      * scenario by closing and re-opening the md fd, which is handled via
    1232             :      * relcache flush.  (Since the checkpointer doesn't participate in
    1233             :      * relcache flush, it could have segment entries for inactive segments;
    1234             :      * that's OK because the checkpointer never needs to compute relation
    1235             :      * size.)
    1236             :      */
    1237     4842500 :     segno = reln->md_num_open_segs[forknum] - 1;
    1238     4842500 :     v = &reln->md_seg_fds[forknum][segno];
    1239             : 
    1240             :     for (;;)
    1241             :     {
    1242     4842500 :         nblocks = _mdnblocks(reln, forknum, v);
    1243     4842500 :         if (nblocks > ((BlockNumber) RELSEG_SIZE))
    1244           0 :             elog(FATAL, "segment too big");
    1245     4842500 :         if (nblocks < ((BlockNumber) RELSEG_SIZE))
    1246     4842500 :             return (segno * ((BlockNumber) RELSEG_SIZE)) + nblocks;
    1247             : 
    1248             :         /*
    1249             :          * If segment is exactly RELSEG_SIZE, advance to next one.
    1250             :          */
    1251           0 :         segno++;
    1252             : 
    1253             :         /*
    1254             :          * We used to pass O_CREAT here, but that has the disadvantage that it
    1255             :          * might create a segment which has vanished through some operating
    1256             :          * system misadventure.  In such a case, creating the segment here
    1257             :          * undermines _mdfd_getseg's attempts to notice and report an error
    1258             :          * upon access to a missing segment.
    1259             :          */
    1260           0 :         v = _mdfd_openseg(reln, forknum, segno, 0);
    1261           0 :         if (v == NULL)
    1262           0 :             return segno * ((BlockNumber) RELSEG_SIZE);
    1263             :     }
    1264             : }
    1265             : 
    1266             : /*
    1267             :  * mdtruncate() -- Truncate relation to specified number of blocks.
    1268             :  *
    1269             :  * Guaranteed not to allocate memory, so it can be used in a critical section.
    1270             :  * Caller must have called smgrnblocks() to obtain curnblk while holding a
    1271             :  * sufficient lock to prevent a change in relation size, and not used any smgr
    1272             :  * functions for this relation or handled interrupts in between.  This makes
    1273             :  * sure we have opened all active segments, so that truncate loop will get
    1274             :  * them all!
    1275             :  */
    1276             : void
    1277        1832 : mdtruncate(SMgrRelation reln, ForkNumber forknum,
    1278             :            BlockNumber curnblk, BlockNumber nblocks)
    1279             : {
    1280             :     BlockNumber priorblocks;
    1281             :     int         curopensegs;
    1282             : 
    1283        1832 :     if (nblocks > curnblk)
    1284             :     {
    1285             :         /* Bogus request ... but no complaint if InRecovery */
    1286           0 :         if (InRecovery)
    1287           0 :             return;
    1288           0 :         ereport(ERROR,
    1289             :                 (errmsg("could not truncate file \"%s\" to %u blocks: it's only %u blocks now",
    1290             :                         relpath(reln->smgr_rlocator, forknum).str,
    1291             :                         nblocks, curnblk)));
    1292             :     }
    1293        1832 :     if (nblocks == curnblk)
    1294         714 :         return;                 /* no work */
    1295             : 
    1296             :     /*
    1297             :      * Truncate segments, starting at the last one. Starting at the end makes
    1298             :      * managing the memory for the fd array easier, should there be errors.
    1299             :      */
    1300        1118 :     curopensegs = reln->md_num_open_segs[forknum];
    1301        2236 :     while (curopensegs > 0)
    1302             :     {
    1303             :         MdfdVec    *v;
    1304             : 
    1305        1118 :         priorblocks = (curopensegs - 1) * RELSEG_SIZE;
    1306             : 
    1307        1118 :         v = &reln->md_seg_fds[forknum][curopensegs - 1];
    1308             : 
    1309        1118 :         if (priorblocks > nblocks)
    1310             :         {
    1311             :             /*
    1312             :              * This segment is no longer active. We truncate the file, but do
    1313             :              * not delete it, for reasons explained in the header comments.
    1314             :              */
    1315           0 :             if (FileTruncate(v->mdfd_vfd, 0, WAIT_EVENT_DATA_FILE_TRUNCATE) < 0)
    1316           0 :                 ereport(ERROR,
    1317             :                         (errcode_for_file_access(),
    1318             :                          errmsg("could not truncate file \"%s\": %m",
    1319             :                                 FilePathName(v->mdfd_vfd))));
    1320             : 
    1321           0 :             if (!SmgrIsTemp(reln))
    1322           0 :                 register_dirty_segment(reln, forknum, v);
    1323             : 
    1324             :             /* we never drop the 1st segment */
    1325             :             Assert(v != &reln->md_seg_fds[forknum][0]);
    1326             : 
    1327           0 :             FileClose(v->mdfd_vfd);
    1328           0 :             _fdvec_resize(reln, forknum, curopensegs - 1);
    1329             :         }
    1330        1118 :         else if (priorblocks + ((BlockNumber) RELSEG_SIZE) > nblocks)
    1331             :         {
    1332             :             /*
    1333             :              * This is the last segment we want to keep. Truncate the file to
    1334             :              * the right length. NOTE: if nblocks is exactly a multiple K of
    1335             :              * RELSEG_SIZE, we will truncate the K+1st segment to 0 length but
    1336             :              * keep it. This adheres to the invariant given in the header
    1337             :              * comments.
    1338             :              */
    1339        1118 :             BlockNumber lastsegblocks = nblocks - priorblocks;
    1340             : 
    1341        1118 :             if (FileTruncate(v->mdfd_vfd, (off_t) lastsegblocks * BLCKSZ, WAIT_EVENT_DATA_FILE_TRUNCATE) < 0)
    1342           0 :                 ereport(ERROR,
    1343             :                         (errcode_for_file_access(),
    1344             :                          errmsg("could not truncate file \"%s\" to %u blocks: %m",
    1345             :                                 FilePathName(v->mdfd_vfd),
    1346             :                                 nblocks)));
    1347        1118 :             if (!SmgrIsTemp(reln))
    1348         790 :                 register_dirty_segment(reln, forknum, v);
    1349             :         }
    1350             :         else
    1351             :         {
    1352             :             /*
    1353             :              * We still need this segment, so nothing to do for this and any
    1354             :              * earlier segment.
    1355             :              */
    1356           0 :             break;
    1357             :         }
    1358        1118 :         curopensegs--;
    1359             :     }
    1360             : }
    1361             : 
    1362             : /*
    1363             :  * mdregistersync() -- Mark whole relation as needing fsync
    1364             :  */
    1365             : void
    1366       50904 : mdregistersync(SMgrRelation reln, ForkNumber forknum)
    1367             : {
    1368             :     int         segno;
    1369             :     int         min_inactive_seg;
    1370             : 
    1371             :     /*
    1372             :      * NOTE: mdnblocks makes sure we have opened all active segments, so that
    1373             :      * the loop below will get them all!
    1374             :      */
    1375       50904 :     mdnblocks(reln, forknum);
    1376             : 
    1377       50904 :     min_inactive_seg = segno = reln->md_num_open_segs[forknum];
    1378             : 
    1379             :     /*
    1380             :      * Temporarily open inactive segments, then close them after sync.  There
    1381             :      * may be some inactive segments left opened after error, but that is
    1382             :      * harmless.  We don't bother to clean them up and take a risk of further
    1383             :      * trouble.  The next mdclose() will soon close them.
    1384             :      */
    1385       50904 :     while (_mdfd_openseg(reln, forknum, segno, 0) != NULL)
    1386           0 :         segno++;
    1387             : 
    1388      101808 :     while (segno > 0)
    1389             :     {
    1390       50904 :         MdfdVec    *v = &reln->md_seg_fds[forknum][segno - 1];
    1391             : 
    1392       50904 :         register_dirty_segment(reln, forknum, v);
    1393             : 
    1394             :         /* Close inactive segments immediately */
    1395       50904 :         if (segno > min_inactive_seg)
    1396             :         {
    1397           0 :             FileClose(v->mdfd_vfd);
    1398           0 :             _fdvec_resize(reln, forknum, segno - 1);
    1399             :         }
    1400             : 
    1401       50904 :         segno--;
    1402             :     }
    1403       50904 : }
    1404             : 
    1405             : /*
    1406             :  * mdimmedsync() -- Immediately sync a relation to stable storage.
    1407             :  *
    1408             :  * Note that only writes already issued are synced; this routine knows
    1409             :  * nothing of dirty buffers that may exist inside the buffer manager.  We
    1410             :  * sync active and inactive segments; smgrDoPendingSyncs() relies on this.
    1411             :  * Consider a relation skipping WAL.  Suppose a checkpoint syncs blocks of
    1412             :  * some segment, then mdtruncate() renders that segment inactive.  If we
    1413             :  * crash before the next checkpoint syncs the newly-inactive segment, that
    1414             :  * segment may survive recovery, reintroducing unwanted data into the table.
    1415             :  */
    1416             : void
    1417          54 : mdimmedsync(SMgrRelation reln, ForkNumber forknum)
    1418             : {
    1419             :     int         segno;
    1420             :     int         min_inactive_seg;
    1421             : 
    1422             :     /*
    1423             :      * NOTE: mdnblocks makes sure we have opened all active segments, so that
    1424             :      * the loop below will get them all!
    1425             :      */
    1426          54 :     mdnblocks(reln, forknum);
    1427             : 
    1428          54 :     min_inactive_seg = segno = reln->md_num_open_segs[forknum];
    1429             : 
    1430             :     /*
    1431             :      * Temporarily open inactive segments, then close them after sync.  There
    1432             :      * may be some inactive segments left opened after fsync() error, but that
    1433             :      * is harmless.  We don't bother to clean them up and take a risk of
    1434             :      * further trouble.  The next mdclose() will soon close them.
    1435             :      */
    1436          54 :     while (_mdfd_openseg(reln, forknum, segno, 0) != NULL)
    1437           0 :         segno++;
    1438             : 
    1439         108 :     while (segno > 0)
    1440             :     {
    1441          54 :         MdfdVec    *v = &reln->md_seg_fds[forknum][segno - 1];
    1442             : 
    1443             :         /*
    1444             :          * fsyncs done through mdimmedsync() should be tracked in a separate
    1445             :          * IOContext than those done through mdsyncfiletag() to differentiate
    1446             :          * between unavoidable client backend fsyncs (e.g. those done during
    1447             :          * index build) and those which ideally would have been done by the
    1448             :          * checkpointer. Since other IO operations bypassing the buffer
    1449             :          * manager could also be tracked in such an IOContext, wait until
    1450             :          * these are also tracked to track immediate fsyncs.
    1451             :          */
    1452          54 :         if (FileSync(v->mdfd_vfd, WAIT_EVENT_DATA_FILE_IMMEDIATE_SYNC) < 0)
    1453           0 :             ereport(data_sync_elevel(ERROR),
    1454             :                     (errcode_for_file_access(),
    1455             :                      errmsg("could not fsync file \"%s\": %m",
    1456             :                             FilePathName(v->mdfd_vfd))));
    1457             : 
    1458             :         /* Close inactive segments immediately */
    1459          54 :         if (segno > min_inactive_seg)
    1460             :         {
    1461           0 :             FileClose(v->mdfd_vfd);
    1462           0 :             _fdvec_resize(reln, forknum, segno - 1);
    1463             :         }
    1464             : 
    1465          54 :         segno--;
    1466             :     }
    1467          54 : }
    1468             : 
    1469             : int
    1470      941894 : mdfd(SMgrRelation reln, ForkNumber forknum, BlockNumber blocknum, uint32 *off)
    1471             : {
    1472      941894 :     MdfdVec    *v = mdopenfork(reln, forknum, EXTENSION_FAIL);
    1473             : 
    1474      941894 :     v = _mdfd_getseg(reln, forknum, blocknum, false,
    1475             :                      EXTENSION_FAIL);
    1476             : 
    1477      941894 :     *off = (off_t) BLCKSZ * (blocknum % ((BlockNumber) RELSEG_SIZE));
    1478             : 
    1479             :     Assert(*off < (off_t) BLCKSZ * RELSEG_SIZE);
    1480             : 
    1481      941894 :     return FileGetRawDesc(v->mdfd_vfd);
    1482             : }
    1483             : 
    1484             : /*
    1485             :  * register_dirty_segment() -- Mark a relation segment as needing fsync
    1486             :  *
    1487             :  * If there is a local pending-ops table, just make an entry in it for
    1488             :  * ProcessSyncRequests to process later.  Otherwise, try to pass off the
    1489             :  * fsync request to the checkpointer process.  If that fails, just do the
    1490             :  * fsync locally before returning (we hope this will not happen often
    1491             :  * enough to be a performance problem).
    1492             :  */
    1493             : static void
    1494     1831066 : register_dirty_segment(SMgrRelation reln, ForkNumber forknum, MdfdVec *seg)
    1495             : {
    1496             :     FileTag     tag;
    1497             : 
    1498     1831066 :     INIT_MD_FILETAG(tag, reln->smgr_rlocator.locator, forknum, seg->mdfd_segno);
    1499             : 
    1500             :     /* Temp relations should never be fsync'd */
    1501             :     Assert(!SmgrIsTemp(reln));
    1502             : 
    1503     1831066 :     if (!RegisterSyncRequest(&tag, SYNC_REQUEST, false /* retryOnError */ ))
    1504             :     {
    1505             :         instr_time  io_start;
    1506             : 
    1507         622 :         ereport(DEBUG1,
    1508             :                 (errmsg_internal("could not forward fsync request because request queue is full")));
    1509             : 
    1510         622 :         io_start = pgstat_prepare_io_time(track_io_timing);
    1511             : 
    1512         622 :         if (FileSync(seg->mdfd_vfd, WAIT_EVENT_DATA_FILE_SYNC) < 0)
    1513           0 :             ereport(data_sync_elevel(ERROR),
    1514             :                     (errcode_for_file_access(),
    1515             :                      errmsg("could not fsync file \"%s\": %m",
    1516             :                             FilePathName(seg->mdfd_vfd))));
    1517             : 
    1518             :         /*
    1519             :          * We have no way of knowing if the current IOContext is
    1520             :          * IOCONTEXT_NORMAL or IOCONTEXT_[BULKREAD, BULKWRITE, VACUUM] at this
    1521             :          * point, so count the fsync as being in the IOCONTEXT_NORMAL
    1522             :          * IOContext. This is probably okay, because the number of backend
    1523             :          * fsyncs doesn't say anything about the efficacy of the
    1524             :          * BufferAccessStrategy. And counting both fsyncs done in
    1525             :          * IOCONTEXT_NORMAL and IOCONTEXT_[BULKREAD, BULKWRITE, VACUUM] under
    1526             :          * IOCONTEXT_NORMAL is likely clearer when investigating the number of
    1527             :          * backend fsyncs.
    1528             :          */
    1529         622 :         pgstat_count_io_op_time(IOOBJECT_RELATION, IOCONTEXT_NORMAL,
    1530             :                                 IOOP_FSYNC, io_start, 1, 0);
    1531             :     }
    1532     1831066 : }
    1533             : 
    1534             : /*
    1535             :  * register_unlink_segment() -- Schedule a file to be deleted after next checkpoint
    1536             :  */
    1537             : static void
    1538       71390 : register_unlink_segment(RelFileLocatorBackend rlocator, ForkNumber forknum,
    1539             :                         BlockNumber segno)
    1540             : {
    1541             :     FileTag     tag;
    1542             : 
    1543       71390 :     INIT_MD_FILETAG(tag, rlocator.locator, forknum, segno);
    1544             : 
    1545             :     /* Should never be used with temp relations */
    1546             :     Assert(!RelFileLocatorBackendIsTemp(rlocator));
    1547             : 
    1548       71390 :     RegisterSyncRequest(&tag, SYNC_UNLINK_REQUEST, true /* retryOnError */ );
    1549       71390 : }
    1550             : 
    1551             : /*
    1552             :  * register_forget_request() -- forget any fsyncs for a relation fork's segment
    1553             :  */
    1554             : static void
    1555      274338 : register_forget_request(RelFileLocatorBackend rlocator, ForkNumber forknum,
    1556             :                         BlockNumber segno)
    1557             : {
    1558             :     FileTag     tag;
    1559             : 
    1560      274338 :     INIT_MD_FILETAG(tag, rlocator.locator, forknum, segno);
    1561             : 
    1562      274338 :     RegisterSyncRequest(&tag, SYNC_FORGET_REQUEST, true /* retryOnError */ );
    1563      274338 : }
    1564             : 
    1565             : /*
    1566             :  * ForgetDatabaseSyncRequests -- forget any fsyncs and unlinks for a DB
    1567             :  */
    1568             : void
    1569         120 : ForgetDatabaseSyncRequests(Oid dbid)
    1570             : {
    1571             :     FileTag     tag;
    1572             :     RelFileLocator rlocator;
    1573             : 
    1574         120 :     rlocator.dbOid = dbid;
    1575         120 :     rlocator.spcOid = 0;
    1576         120 :     rlocator.relNumber = 0;
    1577             : 
    1578         120 :     INIT_MD_FILETAG(tag, rlocator, InvalidForkNumber, InvalidBlockNumber);
    1579             : 
    1580         120 :     RegisterSyncRequest(&tag, SYNC_FILTER_REQUEST, true /* retryOnError */ );
    1581         120 : }
    1582             : 
    1583             : /*
    1584             :  * DropRelationFiles -- drop files of all given relations
    1585             :  */
    1586             : void
    1587        5180 : DropRelationFiles(RelFileLocator *delrels, int ndelrels, bool isRedo)
    1588             : {
    1589             :     SMgrRelation *srels;
    1590             :     int         i;
    1591             : 
    1592        5180 :     srels = palloc(sizeof(SMgrRelation) * ndelrels);
    1593       20174 :     for (i = 0; i < ndelrels; i++)
    1594             :     {
    1595       14994 :         SMgrRelation srel = smgropen(delrels[i], INVALID_PROC_NUMBER);
    1596             : 
    1597       14994 :         if (isRedo)
    1598             :         {
    1599             :             ForkNumber  fork;
    1600             : 
    1601       74730 :             for (fork = 0; fork <= MAX_FORKNUM; fork++)
    1602       59784 :                 XLogDropRelation(delrels[i], fork);
    1603             :         }
    1604       14994 :         srels[i] = srel;
    1605             :     }
    1606             : 
    1607        5180 :     smgrdounlinkall(srels, ndelrels, isRedo);
    1608             : 
    1609       20174 :     for (i = 0; i < ndelrels; i++)
    1610       14994 :         smgrclose(srels[i]);
    1611        5180 :     pfree(srels);
    1612        5180 : }
    1613             : 
    1614             : 
    1615             : /*
    1616             :  * _fdvec_resize() -- Resize the fork's open segments array
    1617             :  */
    1618             : static void
    1619     3057184 : _fdvec_resize(SMgrRelation reln,
    1620             :               ForkNumber forknum,
    1621             :               int nseg)
    1622             : {
    1623     3057184 :     if (nseg == 0)
    1624             :     {
    1625     1091504 :         if (reln->md_num_open_segs[forknum] > 0)
    1626             :         {
    1627     1091504 :             pfree(reln->md_seg_fds[forknum]);
    1628     1091504 :             reln->md_seg_fds[forknum] = NULL;
    1629             :         }
    1630             :     }
    1631     1965680 :     else if (reln->md_num_open_segs[forknum] == 0)
    1632             :     {
    1633     1965680 :         reln->md_seg_fds[forknum] =
    1634     1965680 :             MemoryContextAlloc(MdCxt, sizeof(MdfdVec) * nseg);
    1635             :     }
    1636           0 :     else if (nseg > reln->md_num_open_segs[forknum])
    1637             :     {
    1638             :         /*
    1639             :          * It doesn't seem worthwhile complicating the code to amortize
    1640             :          * repalloc() calls.  Those are far faster than PathNameOpenFile() or
    1641             :          * FileClose(), and the memory context internally will sometimes avoid
    1642             :          * doing an actual reallocation.
    1643             :          */
    1644           0 :         reln->md_seg_fds[forknum] =
    1645           0 :             repalloc(reln->md_seg_fds[forknum],
    1646             :                      sizeof(MdfdVec) * nseg);
    1647             :     }
    1648             :     else
    1649             :     {
    1650             :         /*
    1651             :          * We don't reallocate a smaller array, because we want mdtruncate()
    1652             :          * to be able to promise that it won't allocate memory, so that it is
    1653             :          * allowed in a critical section.  This means that a bit of space in
    1654             :          * the array is now wasted, until the next time we add a segment and
    1655             :          * reallocate.
    1656             :          */
    1657             :     }
    1658             : 
    1659     3057184 :     reln->md_num_open_segs[forknum] = nseg;
    1660     3057184 : }
    1661             : 
    1662             : /*
    1663             :  * Return the filename for the specified segment of the relation. The
    1664             :  * returned string is palloc'd.
    1665             :  */
    1666             : static MdPathStr
    1667       50982 : _mdfd_segpath(SMgrRelation reln, ForkNumber forknum, BlockNumber segno)
    1668             : {
    1669             :     RelPathStr  path;
    1670             :     MdPathStr   fullpath;
    1671             : 
    1672       50982 :     path = relpath(reln->smgr_rlocator, forknum);
    1673             : 
    1674       50982 :     if (segno > 0)
    1675       50982 :         sprintf(fullpath.str, "%s.%u", path.str, segno);
    1676             :     else
    1677           0 :         strcpy(fullpath.str, path.str);
    1678             : 
    1679       50982 :     return fullpath;
    1680             : }
    1681             : 
    1682             : /*
    1683             :  * Open the specified segment of the relation,
    1684             :  * and make a MdfdVec object for it.  Returns NULL on failure.
    1685             :  */
    1686             : static MdfdVec *
    1687       50958 : _mdfd_openseg(SMgrRelation reln, ForkNumber forknum, BlockNumber segno,
    1688             :               int oflags)
    1689             : {
    1690             :     MdfdVec    *v;
    1691             :     File        fd;
    1692             :     MdPathStr   fullpath;
    1693             : 
    1694       50958 :     fullpath = _mdfd_segpath(reln, forknum, segno);
    1695             : 
    1696             :     /* open the file */
    1697       50958 :     fd = PathNameOpenFile(fullpath.str, _mdfd_open_flags() | oflags);
    1698             : 
    1699       50958 :     if (fd < 0)
    1700       50958 :         return NULL;
    1701             : 
    1702             :     /*
    1703             :      * Segments are always opened in order from lowest to highest, so we must
    1704             :      * be adding a new one at the end.
    1705             :      */
    1706             :     Assert(segno == reln->md_num_open_segs[forknum]);
    1707             : 
    1708           0 :     _fdvec_resize(reln, forknum, segno + 1);
    1709             : 
    1710             :     /* fill the entry */
    1711           0 :     v = &reln->md_seg_fds[forknum][segno];
    1712           0 :     v->mdfd_vfd = fd;
    1713           0 :     v->mdfd_segno = segno;
    1714             : 
    1715             :     Assert(_mdnblocks(reln, forknum, v) <= ((BlockNumber) RELSEG_SIZE));
    1716             : 
    1717             :     /* all done */
    1718           0 :     return v;
    1719             : }
    1720             : 
    1721             : /*
    1722             :  * _mdfd_getseg() -- Find the segment of the relation holding the
    1723             :  *                   specified block.
    1724             :  *
    1725             :  * If the segment doesn't exist, we ereport, return NULL, or create the
    1726             :  * segment, according to "behavior".  Note: skipFsync is only used in the
    1727             :  * EXTENSION_CREATE case.
    1728             :  */
    1729             : static MdfdVec *
    1730     5163224 : _mdfd_getseg(SMgrRelation reln, ForkNumber forknum, BlockNumber blkno,
    1731             :              bool skipFsync, int behavior)
    1732             : {
    1733             :     MdfdVec    *v;
    1734             :     BlockNumber targetseg;
    1735             :     BlockNumber nextsegno;
    1736             : 
    1737             :     /* some way to handle non-existent segments needs to be specified */
    1738             :     Assert(behavior &
    1739             :            (EXTENSION_FAIL | EXTENSION_CREATE | EXTENSION_RETURN_NULL |
    1740             :             EXTENSION_DONT_OPEN));
    1741             : 
    1742     5163224 :     targetseg = blkno / ((BlockNumber) RELSEG_SIZE);
    1743             : 
    1744             :     /* if an existing and opened segment, we're done */
    1745     5163224 :     if (targetseg < reln->md_num_open_segs[forknum])
    1746             :     {
    1747     4722036 :         v = &reln->md_seg_fds[forknum][targetseg];
    1748     4722036 :         return v;
    1749             :     }
    1750             : 
    1751             :     /* The caller only wants the segment if we already had it open. */
    1752      441188 :     if (behavior & EXTENSION_DONT_OPEN)
    1753           0 :         return NULL;
    1754             : 
    1755             :     /*
    1756             :      * The target segment is not yet open. Iterate over all the segments
    1757             :      * between the last opened and the target segment. This way missing
    1758             :      * segments either raise an error, or get created (according to
    1759             :      * 'behavior'). Start with either the last opened, or the first segment if
    1760             :      * none was opened before.
    1761             :      */
    1762      441188 :     if (reln->md_num_open_segs[forknum] > 0)
    1763          24 :         v = &reln->md_seg_fds[forknum][reln->md_num_open_segs[forknum] - 1];
    1764             :     else
    1765             :     {
    1766      441164 :         v = mdopenfork(reln, forknum, behavior);
    1767      441158 :         if (!v)
    1768           0 :             return NULL;        /* if behavior & EXTENSION_RETURN_NULL */
    1769             :     }
    1770             : 
    1771      441182 :     for (nextsegno = reln->md_num_open_segs[forknum];
    1772           0 :          nextsegno <= targetseg; nextsegno++)
    1773             :     {
    1774          24 :         BlockNumber nblocks = _mdnblocks(reln, forknum, v);
    1775          24 :         int         flags = 0;
    1776             : 
    1777             :         Assert(nextsegno == v->mdfd_segno + 1);
    1778             : 
    1779          24 :         if (nblocks > ((BlockNumber) RELSEG_SIZE))
    1780           0 :             elog(FATAL, "segment too big");
    1781             : 
    1782          24 :         if ((behavior & EXTENSION_CREATE) ||
    1783          24 :             (InRecovery && (behavior & EXTENSION_CREATE_RECOVERY)))
    1784             :         {
    1785             :             /*
    1786             :              * Normally we will create new segments only if authorized by the
    1787             :              * caller (i.e., we are doing mdextend()).  But when doing WAL
    1788             :              * recovery, create segments anyway; this allows cases such as
    1789             :              * replaying WAL data that has a write into a high-numbered
    1790             :              * segment of a relation that was later deleted. We want to go
    1791             :              * ahead and create the segments so we can finish out the replay.
    1792             :              *
    1793             :              * We have to maintain the invariant that segments before the last
    1794             :              * active segment are of size RELSEG_SIZE; therefore, if
    1795             :              * extending, pad them out with zeroes if needed.  (This only
    1796             :              * matters if in recovery, or if the caller is extending the
    1797             :              * relation discontiguously, but that can happen in hash indexes.)
    1798             :              */
    1799           0 :             if (nblocks < ((BlockNumber) RELSEG_SIZE))
    1800             :             {
    1801           0 :                 char       *zerobuf = palloc_aligned(BLCKSZ, PG_IO_ALIGN_SIZE,
    1802             :                                                      MCXT_ALLOC_ZERO);
    1803             : 
    1804           0 :                 mdextend(reln, forknum,
    1805           0 :                          nextsegno * ((BlockNumber) RELSEG_SIZE) - 1,
    1806             :                          zerobuf, skipFsync);
    1807           0 :                 pfree(zerobuf);
    1808             :             }
    1809           0 :             flags = O_CREAT;
    1810             :         }
    1811          24 :         else if (nblocks < ((BlockNumber) RELSEG_SIZE))
    1812             :         {
    1813             :             /*
    1814             :              * When not extending, only open the next segment if the current
    1815             :              * one is exactly RELSEG_SIZE.  If not (this branch), either
    1816             :              * return NULL or fail.
    1817             :              */
    1818          24 :             if (behavior & EXTENSION_RETURN_NULL)
    1819             :             {
    1820             :                 /*
    1821             :                  * Some callers discern between reasons for _mdfd_getseg()
    1822             :                  * returning NULL based on errno. As there's no failing
    1823             :                  * syscall involved in this case, explicitly set errno to
    1824             :                  * ENOENT, as that seems the closest interpretation.
    1825             :                  */
    1826           0 :                 errno = ENOENT;
    1827           0 :                 return NULL;
    1828             :             }
    1829             : 
    1830          24 :             ereport(ERROR,
    1831             :                     (errcode_for_file_access(),
    1832             :                      errmsg("could not open file \"%s\" (target block %u): previous segment is only %u blocks",
    1833             :                             _mdfd_segpath(reln, forknum, nextsegno).str,
    1834             :                             blkno, nblocks)));
    1835             :         }
    1836             : 
    1837           0 :         v = _mdfd_openseg(reln, forknum, nextsegno, flags);
    1838             : 
    1839           0 :         if (v == NULL)
    1840             :         {
    1841           0 :             if ((behavior & EXTENSION_RETURN_NULL) &&
    1842           0 :                 FILE_POSSIBLY_DELETED(errno))
    1843           0 :                 return NULL;
    1844           0 :             ereport(ERROR,
    1845             :                     (errcode_for_file_access(),
    1846             :                      errmsg("could not open file \"%s\" (target block %u): %m",
    1847             :                             _mdfd_segpath(reln, forknum, nextsegno).str,
    1848             :                             blkno)));
    1849             :         }
    1850             :     }
    1851             : 
    1852      441158 :     return v;
    1853             : }
    1854             : 
    1855             : /*
    1856             :  * Get number of blocks present in a single disk file
    1857             :  */
    1858             : static BlockNumber
    1859     4842524 : _mdnblocks(SMgrRelation reln, ForkNumber forknum, MdfdVec *seg)
    1860             : {
    1861             :     off_t       len;
    1862             : 
    1863     4842524 :     len = FileSize(seg->mdfd_vfd);
    1864     4842524 :     if (len < 0)
    1865           0 :         ereport(ERROR,
    1866             :                 (errcode_for_file_access(),
    1867             :                  errmsg("could not seek to end of file \"%s\": %m",
    1868             :                         FilePathName(seg->mdfd_vfd))));
    1869             :     /* note that this calculation will ignore any partial block at EOF */
    1870     4842524 :     return (BlockNumber) (len / BLCKSZ);
    1871             : }
    1872             : 
    1873             : /*
    1874             :  * Sync a file to disk, given a file tag.  Write the path into an output
    1875             :  * buffer so the caller can use it in error messages.
    1876             :  *
    1877             :  * Return 0 on success, -1 on failure, with errno set.
    1878             :  */
    1879             : int
    1880           0 : mdsyncfiletag(const FileTag *ftag, char *path)
    1881             : {
    1882           0 :     SMgrRelation reln = smgropen(ftag->rlocator, INVALID_PROC_NUMBER);
    1883             :     File        file;
    1884             :     instr_time  io_start;
    1885             :     bool        need_to_close;
    1886             :     int         result,
    1887             :                 save_errno;
    1888             : 
    1889             :     /* See if we already have the file open, or need to open it. */
    1890           0 :     if (ftag->segno < reln->md_num_open_segs[ftag->forknum])
    1891             :     {
    1892           0 :         file = reln->md_seg_fds[ftag->forknum][ftag->segno].mdfd_vfd;
    1893           0 :         strlcpy(path, FilePathName(file), MAXPGPATH);
    1894           0 :         need_to_close = false;
    1895             :     }
    1896             :     else
    1897             :     {
    1898             :         MdPathStr   p;
    1899             : 
    1900           0 :         p = _mdfd_segpath(reln, ftag->forknum, ftag->segno);
    1901           0 :         strlcpy(path, p.str, MD_PATH_STR_MAXLEN);
    1902             : 
    1903           0 :         file = PathNameOpenFile(path, _mdfd_open_flags());
    1904           0 :         if (file < 0)
    1905           0 :             return -1;
    1906           0 :         need_to_close = true;
    1907             :     }
    1908             : 
    1909           0 :     io_start = pgstat_prepare_io_time(track_io_timing);
    1910             : 
    1911             :     /* Sync the file. */
    1912           0 :     result = FileSync(file, WAIT_EVENT_DATA_FILE_SYNC);
    1913           0 :     save_errno = errno;
    1914             : 
    1915           0 :     if (need_to_close)
    1916           0 :         FileClose(file);
    1917             : 
    1918           0 :     pgstat_count_io_op_time(IOOBJECT_RELATION, IOCONTEXT_NORMAL,
    1919             :                             IOOP_FSYNC, io_start, 1, 0);
    1920             : 
    1921           0 :     errno = save_errno;
    1922           0 :     return result;
    1923             : }
    1924             : 
    1925             : /*
    1926             :  * Unlink a file, given a file tag.  Write the path into an output
    1927             :  * buffer so the caller can use it in error messages.
    1928             :  *
    1929             :  * Return 0 on success, -1 on failure, with errno set.
    1930             :  */
    1931             : int
    1932       64820 : mdunlinkfiletag(const FileTag *ftag, char *path)
    1933             : {
    1934             :     RelPathStr  p;
    1935             : 
    1936             :     /* Compute the path. */
    1937       64820 :     p = relpathperm(ftag->rlocator, MAIN_FORKNUM);
    1938       64820 :     strlcpy(path, p.str, MAXPGPATH);
    1939             : 
    1940             :     /* Try to unlink the file. */
    1941       64820 :     return unlink(path);
    1942             : }
    1943             : 
    1944             : /*
    1945             :  * Check if a given candidate request matches a given tag, when processing
    1946             :  * a SYNC_FILTER_REQUEST request.  This will be called for all pending
    1947             :  * requests to find out whether to forget them.
    1948             :  */
    1949             : bool
    1950       14086 : mdfiletagmatches(const FileTag *ftag, const FileTag *candidate)
    1951             : {
    1952             :     /*
    1953             :      * For now we only use filter requests as a way to drop all scheduled
    1954             :      * callbacks relating to a given database, when dropping the database.
    1955             :      * We'll return true for all candidates that have the same database OID as
    1956             :      * the ftag from the SYNC_FILTER_REQUEST request, so they're forgotten.
    1957             :      */
    1958       14086 :     return ftag->rlocator.dbOid == candidate->rlocator.dbOid;
    1959             : }
    1960             : 
    1961             : /*
    1962             :  * AIO completion callback for mdstartreadv().
    1963             :  */
    1964             : static PgAioResult
    1965     2255710 : md_readv_complete(PgAioHandle *ioh, PgAioResult prior_result, uint8 cb_data)
    1966             : {
    1967     2255710 :     PgAioTargetData *td = pgaio_io_get_target_data(ioh);
    1968     2255710 :     PgAioResult result = prior_result;
    1969             : 
    1970     2255710 :     if (prior_result.result < 0)
    1971             :     {
    1972          14 :         result.status = PGAIO_RS_ERROR;
    1973          14 :         result.id = PGAIO_HCB_MD_READV;
    1974             :         /* For "hard" errors, track the error number in error_data */
    1975          14 :         result.error_data = -prior_result.result;
    1976          14 :         result.result = 0;
    1977             : 
    1978             :         /*
    1979             :          * Immediately log a message about the IO error, but only to the
    1980             :          * server log. The reason to do so immediately is that the originator
    1981             :          * might not process the query result immediately (because it is busy
    1982             :          * doing another part of query processing) or at all (e.g. if it was
    1983             :          * cancelled or errored out due to another IO also failing).  The
    1984             :          * definer of the IO will emit an ERROR when processing the IO's
    1985             :          * results
    1986             :          */
    1987          14 :         pgaio_result_report(result, td, LOG_SERVER_ONLY);
    1988             : 
    1989          14 :         return result;
    1990             :     }
    1991             : 
    1992             :     /*
    1993             :      * As explained above smgrstartreadv(), the smgr API operates on the level
    1994             :      * of blocks, rather than bytes. Convert.
    1995             :      */
    1996     2255696 :     result.result /= BLCKSZ;
    1997             : 
    1998             :     Assert(result.result <= td->smgr.nblocks);
    1999             : 
    2000     2255696 :     if (result.result == 0)
    2001             :     {
    2002             :         /* consider 0 blocks read a failure */
    2003           4 :         result.status = PGAIO_RS_ERROR;
    2004           4 :         result.id = PGAIO_HCB_MD_READV;
    2005           4 :         result.error_data = 0;
    2006             : 
    2007             :         /* see comment above the "hard error" case */
    2008           4 :         pgaio_result_report(result, td, LOG_SERVER_ONLY);
    2009             : 
    2010           4 :         return result;
    2011             :     }
    2012             : 
    2013     2255692 :     if (result.status != PGAIO_RS_ERROR &&
    2014     2255692 :         result.result < td->smgr.nblocks)
    2015             :     {
    2016             :         /* partial reads should be retried at upper level */
    2017          24 :         result.status = PGAIO_RS_PARTIAL;
    2018          24 :         result.id = PGAIO_HCB_MD_READV;
    2019             :     }
    2020             : 
    2021     2255692 :     return result;
    2022             : }
    2023             : 
    2024             : /*
    2025             :  * AIO error reporting callback for mdstartreadv().
    2026             :  *
    2027             :  * Errors are encoded as follows:
    2028             :  * - PgAioResult.error_data != 0 encodes IO that failed with that errno
    2029             :  * - PgAioResult.error_data == 0 encodes IO that didn't read all data
    2030             :  */
    2031             : static void
    2032          56 : md_readv_report(PgAioResult result, const PgAioTargetData *td, int elevel)
    2033             : {
    2034             :     RelPathStr  path;
    2035             : 
    2036          56 :     path = relpathbackend(td->smgr.rlocator,
    2037             :                           td->smgr.is_temp ? MyProcNumber : INVALID_PROC_NUMBER,
    2038             :                           td->smgr.forkNum);
    2039             : 
    2040          56 :     if (result.error_data != 0)
    2041             :     {
    2042             :         /* for errcode_for_file_access() and %m */
    2043          28 :         errno = result.error_data;
    2044             : 
    2045          28 :         ereport(elevel,
    2046             :                 errcode_for_file_access(),
    2047             :                 errmsg("could not read blocks %u..%u in file \"%s\": %m",
    2048             :                        td->smgr.blockNum,
    2049             :                        td->smgr.blockNum + td->smgr.nblocks - 1,
    2050             :                        path.str));
    2051             :     }
    2052             :     else
    2053             :     {
    2054             :         /*
    2055             :          * NB: This will typically only be output in debug messages, while
    2056             :          * retrying a partial IO.
    2057             :          */
    2058          28 :         ereport(elevel,
    2059             :                 errcode(ERRCODE_DATA_CORRUPTED),
    2060             :                 errmsg("could not read blocks %u..%u in file \"%s\": read only %zu of %zu bytes",
    2061             :                        td->smgr.blockNum,
    2062             :                        td->smgr.blockNum + td->smgr.nblocks - 1,
    2063             :                        path.str,
    2064             :                        result.result * (size_t) BLCKSZ,
    2065             :                        td->smgr.nblocks * (size_t) BLCKSZ));
    2066             :     }
    2067          38 : }

Generated by: LCOV version 1.14