Slightly better version, adjusting a few obsoleted comments, adjusting
error message to distinguish write/extend, fixing a thinko in
smgr_cached_nblocks maintenance.
From c786f979b0c38364775e32b9403b79303507d37b Mon Sep 17 00:00:00 2001
From: Thomas Munro <thomas.mu...@gmail.com>
Date: Sat, 9 Mar 2024 16:04:21 +1300
Subject: [PATCH v2 1/2] Provide vectored variant of smgrextend().
Since mdwrite() and mdextend() were basically the same, merge them.
They had different assertions, but those would surely apply to any
implementation of smgr, so move them up to the smgr.c wrapper
level. The other difference was the policy on segment creation, but
that can be captured by having smgrwritev() and smgrextendv() call a
single mdwritev() function with a new "extend" flag.
---
src/backend/storage/smgr/md.c | 106 +++++---------------------------
src/backend/storage/smgr/smgr.c | 32 ++++++----
src/include/storage/md.h | 3 +-
src/include/storage/smgr.h | 12 +++-
4 files changed, 47 insertions(+), 106 deletions(-)
diff --git a/src/backend/storage/smgr/md.c b/src/backend/storage/smgr/md.c
index bf0f3ca76d..9b12584122 100644
--- a/src/backend/storage/smgr/md.c
+++ b/src/backend/storage/smgr/md.c
@@ -447,79 +447,10 @@ mdunlinkfork(RelFileLocatorBackend rlocator, ForkNumber forknum, bool isRedo)
pfree(path);
}
-/*
- * mdextend() -- Add a block to the specified relation.
- *
- * The semantics are nearly the same as mdwrite(): write at the
- * specified position. However, this is to be used for the case of
- * extending a relation (i.e., blocknum is at or beyond the current
- * EOF). Note that we assume writing a block beyond current EOF
- * causes intervening file space to become filled with zeroes.
- */
-void
-mdextend(SMgrRelation reln, ForkNumber forknum, BlockNumber blocknum,
- const void *buffer, bool skipFsync)
-{
- off_t seekpos;
- int nbytes;
- MdfdVec *v;
-
- /* If this build supports direct I/O, the buffer must be I/O aligned. */
- if (PG_O_DIRECT != 0 && PG_IO_ALIGN_SIZE <= BLCKSZ)
- Assert((uintptr_t) buffer == TYPEALIGN(PG_IO_ALIGN_SIZE, buffer));
-
- /* This assert is too expensive to have on normally ... */
-#ifdef CHECK_WRITE_VS_EXTEND
- Assert(blocknum >= mdnblocks(reln, forknum));
-#endif
-
- /*
- * If a relation manages to grow to 2^32-1 blocks, refuse to extend it any
- * more --- we mustn't create a block whose number actually is
- * InvalidBlockNumber. (Note that this failure should be unreachable
- * because of upstream checks in bufmgr.c.)
- */
- if (blocknum == InvalidBlockNumber)
- ereport(ERROR,
- (errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED),
- errmsg("cannot extend file \"%s\" beyond %u blocks",
- relpath(reln->smgr_rlocator, forknum),
- InvalidBlockNumber)));
-
- v = _mdfd_getseg(reln, forknum, blocknum, skipFsync, EXTENSION_CREATE);
-
- seekpos = (off_t) BLCKSZ * (blocknum % ((BlockNumber) RELSEG_SIZE));
-
- Assert(seekpos < (off_t) BLCKSZ * RELSEG_SIZE);
-
- if ((nbytes = FileWrite(v->mdfd_vfd, buffer, BLCKSZ, seekpos, WAIT_EVENT_DATA_FILE_EXTEND)) != BLCKSZ)
- {
- if (nbytes < 0)
- ereport(ERROR,
- (errcode_for_file_access(),
- errmsg("could not extend file \"%s\": %m",
- FilePathName(v->mdfd_vfd)),
- errhint("Check free disk space.")));
- /* short write: complain appropriately */
- ereport(ERROR,
- (errcode(ERRCODE_DISK_FULL),
- errmsg("could not extend file \"%s\": wrote only %d of %d bytes at block %u",
- FilePathName(v->mdfd_vfd),
- nbytes, BLCKSZ, blocknum),
- errhint("Check free disk space.")));
- }
-
- if (!skipFsync && !SmgrIsTemp(reln))
- register_dirty_segment(reln, forknum, v);
-
- Assert(_mdnblocks(reln, forknum, v) <= ((BlockNumber) RELSEG_SIZE));
-}
-
/*
* mdzeroextend() -- Add new zeroed out blocks to the specified relation.
*
- * Similar to mdextend(), except the relation can be extended by multiple
- * blocks at once and the added blocks will be filled with zeroes.
+ * The added blocks will be filled with zeroes.
*/
void
mdzeroextend(SMgrRelation reln, ForkNumber forknum,
@@ -595,13 +526,7 @@ mdzeroextend(SMgrRelation reln, ForkNumber forknum,
{
int ret;
- /*
- * Even if we don't want to use fallocate, we can still extend a
- * bit more efficiently than writing each 8kB block individually.
- * pg_pwrite_zeros() (via FileZero()) uses pg_pwritev_with_retry()
- * to avoid multiple writes or needing a zeroed buffer for the
- * whole length of the extension.
- */
+ /* Fall back to writing out zeroes. */
ret = FileZero(v->mdfd_vfd,
seekpos, (off_t) BLCKSZ * numblocks,
WAIT_EVENT_DATA_FILE_EXTEND);
@@ -920,19 +845,14 @@ mdreadv(SMgrRelation reln, ForkNumber forknum, BlockNumber blocknum,
/*
* mdwritev() -- Write the supplied blocks at the appropriate location.
*
- * This is to be used only for updating already-existing blocks of a
- * relation (ie, those before the current EOF). To extend a relation,
- * use mdextend().
+ * Note that smgrextendv() and smgrwritev() are different operations, but both
+ * are handled here.
*/
void
mdwritev(SMgrRelation reln, ForkNumber forknum, BlockNumber blocknum,
- const void **buffers, BlockNumber nblocks, bool skipFsync)
+ const void **buffers, BlockNumber nblocks,
+ bool extend, bool skipFsync)
{
- /* This assert is too expensive to have on normally ... */
-#ifdef CHECK_WRITE_VS_EXTEND
- Assert(blocknum < mdnblocks(reln, forknum));
-#endif
-
while (nblocks > 0)
{
struct iovec iov[PG_IOV_MAX];
@@ -945,6 +865,8 @@ mdwritev(SMgrRelation reln, ForkNumber forknum, BlockNumber blocknum,
size_t size_this_segment;
v = _mdfd_getseg(reln, forknum, blocknum, skipFsync,
+ extend ?
+ EXTENSION_CREATE :
EXTENSION_FAIL | EXTENSION_CREATE_RECOVERY);
seekpos = (off_t) BLCKSZ * (blocknum % ((BlockNumber) RELSEG_SIZE));
@@ -992,7 +914,9 @@ mdwritev(SMgrRelation reln, ForkNumber forknum, BlockNumber blocknum,
ereport(ERROR,
(errcode_for_file_access(),
- errmsg("could not write blocks %u..%u in file \"%s\": %m",
+ errmsg(extend ?
+ "could not extend blocks %u..%u in file \"%s\": %m" :
+ "could not write blocks %u..%u in file \"%s\": %m",
blocknum,
blocknum + nblocks_this_segment - 1,
FilePathName(v->mdfd_vfd)),
@@ -1638,7 +1562,7 @@ _mdfd_getseg(SMgrRelation reln, ForkNumber forknum, BlockNumber blkno,
{
/*
* Normally we will create new segments only if authorized by the
- * caller (i.e., we are doing mdextend()). But when doing WAL
+ * caller (i.e., we are doing smgrextend()). But when doing WAL
* recovery, create segments anyway; this allows cases such as
* replaying WAL data that has a write into a high-numbered
* segment of a relation that was later deleted. We want to go
@@ -1655,9 +1579,9 @@ _mdfd_getseg(SMgrRelation reln, ForkNumber forknum, BlockNumber blkno,
char *zerobuf = palloc_aligned(BLCKSZ, PG_IO_ALIGN_SIZE,
MCXT_ALLOC_ZERO);
- mdextend(reln, forknum,
- nextsegno * ((BlockNumber) RELSEG_SIZE) - 1,
- zerobuf, skipFsync);
+ smgrextend(reln, forknum,
+ nextsegno * ((BlockNumber) RELSEG_SIZE) - 1,
+ zerobuf, skipFsync);
pfree(zerobuf);
}
flags = O_CREAT;
diff --git a/src/backend/storage/smgr/smgr.c b/src/backend/storage/smgr/smgr.c
index a5b18328b8..ad0f02f205 100644
--- a/src/backend/storage/smgr/smgr.c
+++ b/src/backend/storage/smgr/smgr.c
@@ -82,8 +82,6 @@ typedef struct f_smgr
bool (*smgr_exists) (SMgrRelation reln, ForkNumber forknum);
void (*smgr_unlink) (RelFileLocatorBackend rlocator, ForkNumber forknum,
bool isRedo);
- void (*smgr_extend) (SMgrRelation reln, ForkNumber forknum,
- BlockNumber blocknum, const void *buffer, bool skipFsync);
void (*smgr_zeroextend) (SMgrRelation reln, ForkNumber forknum,
BlockNumber blocknum, int nblocks, bool skipFsync);
bool (*smgr_prefetch) (SMgrRelation reln, ForkNumber forknum,
@@ -94,7 +92,7 @@ typedef struct f_smgr
void (*smgr_writev) (SMgrRelation reln, ForkNumber forknum,
BlockNumber blocknum,
const void **buffers, BlockNumber nblocks,
- bool skipFsync);
+ bool extend, bool skipFsync);
void (*smgr_writeback) (SMgrRelation reln, ForkNumber forknum,
BlockNumber blocknum, BlockNumber nblocks);
BlockNumber (*smgr_nblocks) (SMgrRelation reln, ForkNumber forknum);
@@ -114,7 +112,6 @@ static const f_smgr smgrsw[] = {
.smgr_create = mdcreate,
.smgr_exists = mdexists,
.smgr_unlink = mdunlink,
- .smgr_extend = mdextend,
.smgr_zeroextend = mdzeroextend,
.smgr_prefetch = mdprefetch,
.smgr_readv = mdreadv,
@@ -525,7 +522,7 @@ smgrdounlinkall(SMgrRelation *rels, int nrels, bool isRedo)
/*
- * smgrextend() -- Add a new block to a file.
+ * smgrextendv() -- Add new blocks to a file.
*
* The semantics are nearly the same as smgrwrite(): write at the
* specified position. However, this is to be used for the case of
@@ -534,19 +531,24 @@ smgrdounlinkall(SMgrRelation *rels, int nrels, bool isRedo)
* causes intervening file space to become filled with zeroes.
*/
void
-smgrextend(SMgrRelation reln, ForkNumber forknum, BlockNumber blocknum,
- const void *buffer, bool skipFsync)
+smgrextendv(SMgrRelation reln, ForkNumber forknum, BlockNumber blocknum,
+ const void **buffers, BlockNumber nblocks, bool skipFsync)
{
- smgrsw[reln->smgr_which].smgr_extend(reln, forknum, blocknum,
- buffer, skipFsync);
+ /* This assert is too expensive to have on normally ... */
+#ifdef CHECK_WRITE_VS_EXTEND
+ Assert(blocknum >= smgrnblocks(reln, forknum));
+#endif
+
+ smgrsw[reln->smgr_which].smgr_writev(reln, forknum, blocknum,
+ buffers, nblocks, true, skipFsync);
/*
- * Normally we expect this to increase nblocks by one, but if the cached
+ * Normally we expect this to increase nblocks by N, but if the cached
* value isn't as expected, just invalidate it so the next call asks the
* kernel.
*/
if (reln->smgr_cached_nblocks[forknum] == blocknum)
- reln->smgr_cached_nblocks[forknum] = blocknum + 1;
+ reln->smgr_cached_nblocks[forknum] = blocknum + nblocks;
else
reln->smgr_cached_nblocks[forknum] = InvalidBlockNumber;
}
@@ -633,8 +635,14 @@ void
smgrwritev(SMgrRelation reln, ForkNumber forknum, BlockNumber blocknum,
const void **buffers, BlockNumber nblocks, bool skipFsync)
{
+ /* This assert is too expensive to have on normally ... */
+#ifdef CHECK_WRITE_VS_EXTEND
+ Assert(blocknum + nblocks <= smgrnblocks(reln, forknum));
+#endif
+
smgrsw[reln->smgr_which].smgr_writev(reln, forknum, blocknum,
- buffers, nblocks, skipFsync);
+ buffers, nblocks, false,
+ skipFsync);
}
/*
diff --git a/src/include/storage/md.h b/src/include/storage/md.h
index 620f10abde..8746a48a5f 100644
--- a/src/include/storage/md.h
+++ b/src/include/storage/md.h
@@ -36,7 +36,8 @@ extern void mdreadv(SMgrRelation reln, ForkNumber forknum, BlockNumber blocknum,
void **buffers, BlockNumber nblocks);
extern void mdwritev(SMgrRelation reln, ForkNumber forknum,
BlockNumber blocknum,
- const void **buffers, BlockNumber nblocks, bool skipFsync);
+ const void **buffers, BlockNumber nblocks,
+ bool extend, bool skipFsync);
extern void mdwriteback(SMgrRelation reln, ForkNumber forknum,
BlockNumber blocknum, BlockNumber nblocks);
extern BlockNumber mdnblocks(SMgrRelation reln, ForkNumber forknum);
diff --git a/src/include/storage/smgr.h b/src/include/storage/smgr.h
index fc5f883ce1..7ffea4332d 100644
--- a/src/include/storage/smgr.h
+++ b/src/include/storage/smgr.h
@@ -86,8 +86,9 @@ extern void smgrreleaserellocator(RelFileLocatorBackend rlocator);
extern void smgrcreate(SMgrRelation reln, ForkNumber forknum, bool isRedo);
extern void smgrdosyncall(SMgrRelation *rels, int nrels);
extern void smgrdounlinkall(SMgrRelation *rels, int nrels, bool isRedo);
-extern void smgrextend(SMgrRelation reln, ForkNumber forknum,
- BlockNumber blocknum, const void *buffer, bool skipFsync);
+extern void smgrextendv(SMgrRelation reln, ForkNumber forknum,
+ BlockNumber blocknum,
+ const void **buffers, BlockNumber nblocks, bool skipFsync);
extern void smgrzeroextend(SMgrRelation reln, ForkNumber forknum,
BlockNumber blocknum, int nblocks, bool skipFsync);
extern bool smgrprefetch(SMgrRelation reln, ForkNumber forknum,
@@ -124,4 +125,11 @@ smgrwrite(SMgrRelation reln, ForkNumber forknum, BlockNumber blocknum,
smgrwritev(reln, forknum, blocknum, &buffer, 1, skipFsync);
}
+static inline void
+smgrextend(SMgrRelation reln, ForkNumber forknum, BlockNumber blocknum,
+ const void *buffer, bool skipFsync)
+{
+ smgrextendv(reln, forknum, blocknum, &buffer, 1, skipFsync);
+}
+
#endif /* SMGR_H */
--
2.43.2
From bf8f807a6c0d34a3447bb8ef6c7cc014be5c3603 Mon Sep 17 00:00:00 2001
From: Thomas Munro <thomas.mu...@gmail.com>
Date: Sat, 9 Mar 2024 16:54:56 +1300
Subject: [PATCH v2 2/2] Use vectored I/O for bulk writes.
Zero-extend using smgrzeroextend(). Extend using smgrextendv(). Write
using smgrwritev().
---
src/backend/storage/smgr/bulk_write.c | 83 +++++++++++++++++++--------
1 file changed, 58 insertions(+), 25 deletions(-)
diff --git a/src/backend/storage/smgr/bulk_write.c b/src/backend/storage/smgr/bulk_write.c
index 4a10ece4c3..2a566a80f6 100644
--- a/src/backend/storage/smgr/bulk_write.c
+++ b/src/backend/storage/smgr/bulk_write.c
@@ -8,7 +8,7 @@
* the regular buffer manager and the bulk loading interface!
*
* We bypass the buffer manager to avoid the locking overhead, and call
- * smgrextend() directly. A downside is that the pages will need to be
+ * smgrextendv() directly. A downside is that the pages will need to be
* re-read into shared buffers on first use after the build finishes. That's
* usually a good tradeoff for large relations, and for small relations, the
* overhead isn't very significant compared to creating the relation in the
@@ -45,8 +45,6 @@
#define MAX_PENDING_WRITES XLR_MAX_BLOCK_ID
-static const PGIOAlignedBlock zero_buffer = {{0}}; /* worth BLCKSZ */
-
typedef struct PendingWrite
{
BulkWriteBuffer buf;
@@ -225,35 +223,70 @@ smgr_bulk_flush(BulkWriteState *bulkstate)
for (int i = 0; i < npending; i++)
{
- BlockNumber blkno = pending_writes[i].blkno;
- Page page = pending_writes[i].buf->data;
-
+ Page page;
+ const void *pages[16];
+ BlockNumber blkno;
+ int nblocks;
+ int max_nblocks;
+
+ /* Prepare to write the first block. */
+ blkno = pending_writes[i].blkno;
+ page = pending_writes[i].buf->data;
PageSetChecksumInplace(page, blkno);
+ pages[0] = page;
+ nblocks = 1;
- if (blkno >= bulkstate->pages_written)
+ /* Zero-extend any missing space before the first block. */
+ if (blkno > bulkstate->pages_written)
+ {
+ int nzeroblocks;
+
+ nzeroblocks = blkno - bulkstate->pages_written;
+ smgrzeroextend(bulkstate->smgr, bulkstate->forknum,
+ bulkstate->pages_written, nzeroblocks, true);
+ bulkstate->pages_written += nzeroblocks;
+ }
+
+ if (blkno < bulkstate->pages_written)
{
/*
- * If we have to write pages nonsequentially, fill in the space
- * with zeroes until we come back and overwrite. This is not
- * logically necessary on standard Unix filesystems (unwritten
- * space will read as zeroes anyway), but it should help to avoid
- * fragmentation. The dummy pages aren't WAL-logged though.
+ * We're overwriting. Clamp at the existing size, because we
+ * can't mix writing and extending in a single operation.
*/
- while (blkno > bulkstate->pages_written)
- {
- /* don't set checksum for all-zero page */
- smgrextend(bulkstate->smgr, bulkstate->forknum,
- bulkstate->pages_written++,
- &zero_buffer,
- true);
- }
-
- smgrextend(bulkstate->smgr, bulkstate->forknum, blkno, page, true);
- bulkstate->pages_written = pending_writes[i].blkno + 1;
+ max_nblocks = Min(lengthof(pages),
+ bulkstate->pages_written - blkno);
+ }
+ else
+ {
+ /* We're extending. */
+ Assert(blkno == bulkstate->pages_written);
+ max_nblocks = lengthof(pages);
+ }
+
+ /* Find as many consecutive blocks as we can. */
+ while (i + 1 < npending &&
+ pending_writes[i + 1].blkno == blkno + nblocks &&
+ nblocks < max_nblocks)
+ {
+ page = pending_writes[++i].buf->data;
+ PageSetChecksumInplace(page, pending_writes[i].blkno);
+ pages[nblocks++] = page;
+ }
+
+ /* Extend or overwrite. */
+ if (blkno == bulkstate->pages_written)
+ {
+ smgrextendv(bulkstate->smgr, bulkstate->forknum, blkno, pages, nblocks, true);
+ bulkstate->pages_written += nblocks;
}
else
- smgrwrite(bulkstate->smgr, bulkstate->forknum, blkno, page, true);
- pfree(page);
+ {
+ Assert(blkno + nblocks <= bulkstate->pages_written);
+ smgrwritev(bulkstate->smgr, bulkstate->forknum, blkno, pages, nblocks, true);
+ }
+
+ for (int j = 0; j < nblocks; ++j)
+ pfree(pending_writes[i - j].buf->data);
}
bulkstate->npending = 0;
--
2.43.2
