Hello Andres,
[...] posix_fadvise().
My current thinking is "maybe yes, maybe no":-), as it may depend on the OS
implementation of posix_fadvise, so it may differ between OS.
As long as fadvise has no 'undirty' option, I don't see how that
problem goes away. You're telling the OS to throw the buffer away, so
unless it ignores it that'll have consequences when you read the page
back in.
Yep, probably.
Note that we are talking about checkpoints, which "write" buffers out
*but* keep them nevertheless. As the buffer is kept, the OS page is a
duplicate, and freeing it should not harm, at least immediatly.
The situation is different if the memory is reused in between, which is
the work of the bgwriter I think, based on LRU/LFU heuristics, but such
writes are not flushed by the current patch.
Now, if a buffer was recently updated it should not be selected by the
bgwriter, if the LRU/LFU heuristics works as expected, which mitigate the
issue somehow...
To sum up, I agree that it is indeed possible that flushing with
posix_fadvise could reduce read OS-memory hits on some systems for some
workloads, although not on Linux, see below.
So the option is best kept as "off" for now, without further data, I'm
fine with that.
[...] I'd say it should then be an os-specific default. No point in
making people work for it needlessly on linux and/or elsewhere.
Ok. Version 9 attached does that, "on" for Linux, "off" for others because
of the potential issues you mentioned.
(Another reason to keep it "off" is that I'm not sure about what
happens with such HD flushing features on virtual servers).
I don't see how that matters? Either the host will entirely ignore
flushing, and thus the sync_file_range and the fsync won't cost much, or
fsync will be honored, in which case the pre-flushing is helpful.
Possibly. I know that I do not know:-) The distance between the database
and real hardware is so great in VM, that I think that it may have any
effect, including good, bad or none:-)
Overall, I'm not pessimistic, because I've seen I/O storms on a FreeBSD host
and it was as bad as Linux (namely the database and even the box was offline
for long minutes...), and if you can avoid that having to read back some
data may be not that bad a down payment.
I don't see how that'd alleviate my fear.
I'm trying to mitigate your fears, not to alleviate them:-)
Sure, the latency for many workloads will be better, but I don't how
that argument says anything about the reads?
It just says that there may be a compromise, better in some case, possibly
not so in others, because posix_fadvise does not really say what the
database would like to say to the OS, this is why I wrote such a large
comment about it in the source file in the first place.
And we'll not just use this in cases it'd be beneficial...
I'm fine if it is off by default for some systems. If people want to avoid
write stalls they can use the option, but it may have adverse effect on
the tps in some cases, that's life? Not using the option also has adverse
effects in some cases, because you have write stalls... and currently you
do not have the choice, so it would be a progress.
The issue is largely mitigated if the data is not removed from
shared_buffers, because the OS buffer is just a copy of already hold data.
What I would do on such systems is to increase shared_buffers and keep
flushing on, that is to count less on the system cache and more on postgres
own cache.
That doesn't work that well for a bunch of reasons. For one it's
completely non-adaptive. With the OS's page cache you can rely on free
memory being used for caching *and* it be available should a query or
another program need lots of memory.
Yep. I was thinking about a dedicated database server, not a shared one.
Overall, I'm not convince that the practice of relying on the OS cache is a
good one, given what it does with it, at least on Linux.
The alternatives aren't super realistic near-term though. Using direct
IO efficiently on the set of operating systems we support is
*hard*. [...]
Sure. This is not necessarily what I had in mind.
Currently pg "write"s stuff to the OS, and then suddenly calls "fsync" out
of the blue, hoping that in between the OS will actually have done a good
job with the underlying hardware. This is pretty naive, the fsync
generates write storms, and the database is offline: trying to improve
these things is the motivation for this patch.
Now if you think of the bgwriter, it does pretty much the same, and
probably may generate plenty of random I/Os, because the underlying
LRU/LFU heuristics used to select buffers does not care about the file
structures.
So I think that to get good performance the database must take some
control over the OS. That does not mean that direct I/O needs to be
involved, although maybe it could, but this patch shows that it is not
needed to improve things.
Now, if someone could provide a dedicated box with posix_fadvise (say
FreeBSD, maybe others...) for testing that would allow to provide data
instead of speculating... and then maybe to decide to change its default
value.
Testing, as an approximation, how it turns out to work on linux would be
a good step.
Do you mean testing with posix_fadvise on Linux?
I did think about it, but the documented behavior of this call on Linux is
disappointing: if the buffer has been written to disk, it is freed by the
OS. If not, nothing is done. Given that the flush is called pretty close
after writes, mostly the buffer will not have been written to disk yet,
and the call would just be a no-op... So I concluded that there is no
point in trying that on Linux because it will have no effect other than
loosing some time, IMO.
Really, a useful test would be FreeBSD, when posix_fadvise does move
things to disk, although the actual offsets & length are ignored, but I do
not think that it would be a problem. I do not know about other systems
and what they do with posix_fadvise.
--
Fabien.
diff --git a/doc/src/sgml/config.sgml b/doc/src/sgml/config.sgml
index e900dcc..1cec243 100644
--- a/doc/src/sgml/config.sgml
+++ b/doc/src/sgml/config.sgml
@@ -2454,6 +2454,28 @@ include_dir 'conf.d'
</listitem>
</varlistentry>
+ <varlistentry id="guc-checkpoint-sort" xreflabel="checkpoint_sort">
+ <term><varname>checkpoint_sort</varname> (<type>bool</type>)
+ <indexterm>
+ <primary><varname>checkpoint_sort</> configuration parameter</primary>
+ </indexterm>
+ </term>
+ <listitem>
+ <para>
+ Whether to sort buffers before writting them out to disk on checkpoint.
+ For a HDD storage, this setting allows to group together
+ neighboring pages written to disk, thus improving performance by
+ reducing random write activity.
+ This sorting should have limited performance effects on SSD backends
+ as such storages have good random write performance, but it may
+ help with wear-leveling so be worth keeping anyway.
+ The default is <literal>on</>.
+ This parameter can only be set in the <filename>postgresql.conf</>
+ file or on the server command line.
+ </para>
+ </listitem>
+ </varlistentry>
+
<varlistentry id="guc-checkpoint-warning" xreflabel="checkpoint_warning">
<term><varname>checkpoint_warning</varname> (<type>integer</type>)
<indexterm>
diff --git a/doc/src/sgml/wal.sgml b/doc/src/sgml/wal.sgml
index e3941c9..f538698 100644
--- a/doc/src/sgml/wal.sgml
+++ b/doc/src/sgml/wal.sgml
@@ -546,6 +546,18 @@
</para>
<para>
+ When hard-disk drives (HDD) are used for terminal data storage
+ <xref linkend="guc-checkpoint-sort"> allows to sort pages
+ so that neighboring pages on disk will be flushed together by
+ chekpoints, reducing the random write load and improving performance.
+ If solid-state drives (SSD) are used, sorting pages induces no benefit
+ as their random write I/O performance is good: this feature could then
+ be disabled by setting <varname>checkpoint_sort</> to <value>off</>.
+ It is possible that sorting may help with SSD wear leveling, so it may
+ be kept on that account.
+ </para>
+
+ <para>
The number of WAL segment files in <filename>pg_xlog</> directory depends on
<varname>min_wal_size</>, <varname>max_wal_size</> and
the amount of WAL generated in previous checkpoint cycles. When old log
diff --git a/src/backend/access/transam/xlog.c b/src/backend/access/transam/xlog.c
index 68e33eb..bee38ab 100644
--- a/src/backend/access/transam/xlog.c
+++ b/src/backend/access/transam/xlog.c
@@ -7995,11 +7995,13 @@ LogCheckpointEnd(bool restartpoint)
sync_secs,
total_secs,
longest_secs,
+ sort_secs,
average_secs;
int write_usecs,
sync_usecs,
total_usecs,
longest_usecs,
+ sort_usecs,
average_usecs;
uint64 average_sync_time;
@@ -8030,6 +8032,10 @@ LogCheckpointEnd(bool restartpoint)
CheckpointStats.ckpt_end_t,
&total_secs, &total_usecs);
+ TimestampDifference(CheckpointStats.ckpt_sort_t,
+ CheckpointStats.ckpt_sort_end_t,
+ &sort_secs, &sort_usecs);
+
/*
* Timing values returned from CheckpointStats are in microseconds.
* Convert to the second plus microsecond form that TimestampDifference
@@ -8048,8 +8054,8 @@ LogCheckpointEnd(bool restartpoint)
elog(LOG, "%s complete: wrote %d buffers (%.1f%%); "
"%d transaction log file(s) added, %d removed, %d recycled; "
- "write=%ld.%03d s, sync=%ld.%03d s, total=%ld.%03d s; "
- "sync files=%d, longest=%ld.%03d s, average=%ld.%03d s; "
+ "sort=%ld.%03d s, write=%ld.%03d s, sync=%ld.%03d s, total=%ld.%03d s;"
+ " sync files=%d, longest=%ld.%03d s, average=%ld.%03d s; "
"distance=%d kB, estimate=%d kB",
restartpoint ? "restartpoint" : "checkpoint",
CheckpointStats.ckpt_bufs_written,
@@ -8057,6 +8063,7 @@ LogCheckpointEnd(bool restartpoint)
CheckpointStats.ckpt_segs_added,
CheckpointStats.ckpt_segs_removed,
CheckpointStats.ckpt_segs_recycled,
+ sort_secs, sort_usecs / 1000,
write_secs, write_usecs / 1000,
sync_secs, sync_usecs / 1000,
total_secs, total_usecs / 1000,
diff --git a/src/backend/storage/buffer/buf_init.c b/src/backend/storage/buffer/buf_init.c
index 3ae2848..3bd5eab 100644
--- a/src/backend/storage/buffer/buf_init.c
+++ b/src/backend/storage/buffer/buf_init.c
@@ -65,7 +65,8 @@ void
InitBufferPool(void)
{
bool foundBufs,
- foundDescs;
+ foundDescs,
+ foundCpid;
/* Align descriptors to a cacheline boundary. */
BufferDescriptors = (BufferDescPadded *) CACHELINEALIGN(
@@ -77,10 +78,14 @@ InitBufferPool(void)
ShmemInitStruct("Buffer Blocks",
NBuffers * (Size) BLCKSZ, &foundBufs);
- if (foundDescs || foundBufs)
+ CheckpointBufferIds = (CheckpointSortItem *)
+ ShmemInitStruct("Checkpoint BufferIds",
+ NBuffers * sizeof(CheckpointSortItem), &foundCpid);
+
+ if (foundDescs || foundBufs || foundCpid)
{
- /* both should be present or neither */
- Assert(foundDescs && foundBufs);
+ /* all should be present or neither */
+ Assert(foundDescs && foundBufs && foundCpid);
/* note: this path is only taken in EXEC_BACKEND case */
}
else
@@ -144,5 +149,8 @@ BufferShmemSize(void)
/* size of stuff controlled by freelist.c */
size = add_size(size, StrategyShmemSize());
+ /* size of checkpoint sort array in bufmgr.c */
+ size = add_size(size, mul_size(NBuffers, sizeof(CheckpointSortItem)));
+
return size;
}
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index cd3aaad..ca295f1 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -80,6 +80,7 @@ bool zero_damaged_pages = false;
int bgwriter_lru_maxpages = 100;
double bgwriter_lru_multiplier = 2.0;
bool track_io_timing = false;
+bool checkpoint_sort = true;
/*
* How many buffers PrefetchBuffer callers should try to stay ahead of their
@@ -95,6 +96,9 @@ static bool IsForInput;
/* local state for LockBufferForCleanup */
static volatile BufferDesc *PinCountWaitBuf = NULL;
+/* array of buffer ids & sort criterion of all buffers to checkpoint */
+CheckpointSortItem *CheckpointBufferIds = NULL;
+
/*
* Backend-Private refcount management:
*
@@ -1561,6 +1565,129 @@ UnpinBuffer(volatile BufferDesc *buf, bool fixOwner)
}
}
+/* Compare checkpoint buffers.
+ */
+static int bufcmp(const void * pa, const void * pb)
+{
+ CheckpointSortItem
+ *a = (CheckpointSortItem *) pa,
+ *b = (CheckpointSortItem *) pb;
+
+ /* compare relation */
+ if (a->relNode < b->relNode)
+ return -1;
+ else if (a->relNode > b->relNode)
+ return 1;
+ /* same relation, compare fork */
+ else if (a->forkNum < b->forkNum)
+ return -1;
+ else if (a->forkNum > b->forkNum)
+ return 1;
+ /* same relation/fork, so same segmented "file", compare block number
+ * which are mapped on different segments depending on the number.
+ */
+ else if (a->blockNum < b->blockNum)
+ return -1;
+ else /* should not be the same block anyway... */
+ return 1;
+}
+
+/* Status of buffers to checkpoint for a particular tablespace,
+ * used internally in BufferSync.
+ * - space: oid of the tablespace
+ * - num_to_write: number of checkpoint pages counted for this tablespace
+ * - num_written: number of pages actually written out
+ * - index: scanning position in CheckpointBufferIds for this tablespace
+ */
+typedef struct TableSpaceCheckpointStatus {
+ Oid space;
+ int num_to_write;
+ int num_written;
+ int index;
+} TableSpaceCheckpointStatus;
+
+/* entry structure for table space to count hashtable,
+ * used internally in BufferSync.
+ */
+typedef struct TableSpaceCountEntry {
+ Oid space;
+ int count;
+} TableSpaceCountEntry;
+
+/* return the next buffer to write, or -1.
+ * this function balances buffers over tablespaces.
+ */
+static int
+NextBufferToWrite(
+ TableSpaceCheckpointStatus *spcStatus, int nb_spaces,
+ int *pspace, int num_to_write, int num_written)
+{
+ int space = *pspace, buf_id = -1, index;
+
+ /*
+ * Select a tablespace depending on the current overall progress.
+ *
+ * The progress ratio of each unfinished tablespace is compared to
+ * the overall progress ratio to find one with is not in advance
+ * (i.e. tablespace ratio <= overall ratio).
+ *
+ * Existence: it is bound to exist otherwise the overall progress
+ * ratio would be inconsistent: with positive buffers to write (t1 & t2)
+ * and already written buffers (w1 & w2), we have:
+ *
+ * If w1/t1 > (w1+w2)/(t1+t2) # one table space is in advance
+ * => w1t1+w1t2 > w1t1+w2t1 => w1t2 > w2t1 => w1t2+w2t2 > w2t1+w2t2
+ * => (w1+w2) / (t1+t2) > w2 / t2 # the other one is late
+ *
+ * The round robin ensures that each space is given some attention
+ * till it is over the current ratio, before going to the next.
+ *
+ * Precision: using int32 computations for comparing fractions
+ * (w1 / t1 > w / t <=> w1 t > w t1) seems a bad idea as the values
+ * can overflow 32-bit integers: the limit would be sqrt(2**31) ~
+ * 46340 buffers, i.e. a 362 MB checkpoint. So ensure that 64-bit
+ * integers are used in the comparison.
+ */
+ while (/* compare tablespace vs overall progress ratio:
+ * tablespace written/to_write > overall written/to_write
+ */
+ (int64) spcStatus[space].num_written * num_to_write >
+ (int64) num_written * spcStatus[space].num_to_write)
+ space = (space + 1) % nb_spaces; /* round robin */
+
+ /*
+ * Find a valid buffer in the selected tablespace,
+ * by continuing the tablespace specific buffer scan
+ * where it was left.
+ */
+ index = spcStatus[space].index;
+
+ while (index < num_to_write && buf_id == -1)
+ {
+ volatile BufferDesc *bufHdr;
+
+ buf_id = CheckpointBufferIds[index].buf_id;
+ bufHdr = GetBufferDescriptor(buf_id);
+
+ /* Skip if in another tablespace or not in checkpoint anymore.
+ * No lock is acquired, see comments below.
+ */
+ if (spcStatus[space].space != bufHdr->tag.rnode.spcNode ||
+ ! (bufHdr->flags & BM_CHECKPOINT_NEEDED))
+ {
+ index ++;
+ buf_id = -1;
+ }
+ }
+
+ /* Update tablespace writing status, will start over at next index */
+ spcStatus[space].index = index+1;
+
+ *pspace = space;
+
+ return buf_id;
+}
+
/*
* BufferSync -- Write out all dirty buffers in the pool.
*
@@ -1574,11 +1701,13 @@ UnpinBuffer(volatile BufferDesc *buf, bool fixOwner)
static void
BufferSync(int flags)
{
- int buf_id;
- int num_to_scan;
+ int buf_id = -1;
int num_to_write;
int num_written;
int mask = BM_DIRTY;
+ HTAB *spcBuffers;
+ TableSpaceCheckpointStatus *spcStatus = NULL;
+ int nb_spaces, space;
/* Make sure we can handle the pin inside SyncOneBuffer */
ResourceOwnerEnlargeBuffers(CurrentResourceOwner);
@@ -1609,6 +1738,18 @@ BufferSync(int flags)
* certainly need to be written for the next checkpoint attempt, too.
*/
num_to_write = 0;
+
+ /* initialize oid -> int buffer count hash table */
+ {
+ HASHCTL ctl;
+
+ MemSet(&ctl, 0, sizeof(HASHCTL));
+ ctl.keysize = sizeof(Oid);
+ ctl.entrysize = sizeof(TableSpaceCountEntry);
+ spcBuffers = hash_create("Number of buffers to write per tablespace",
+ 16, &ctl, HASH_ELEM | HASH_BLOBS);
+ }
+
for (buf_id = 0; buf_id < NBuffers; buf_id++)
{
volatile BufferDesc *bufHdr = GetBufferDescriptor(buf_id);
@@ -1621,32 +1762,111 @@ BufferSync(int flags)
if ((bufHdr->flags & mask) == mask)
{
+ Oid spc;
+ TableSpaceCountEntry * entry;
+ bool found;
+
bufHdr->flags |= BM_CHECKPOINT_NEEDED;
+ CheckpointBufferIds[num_to_write].buf_id = buf_id;
+ CheckpointBufferIds[num_to_write].relNode = bufHdr->tag.rnode.relNode;
+ CheckpointBufferIds[num_to_write].forkNum = bufHdr->tag.forkNum;
+ CheckpointBufferIds[num_to_write].blockNum = bufHdr->tag.blockNum;
num_to_write++;
+
+ /* keep track of per tablespace buffers */
+ spc = bufHdr->tag.rnode.spcNode;
+ entry = (TableSpaceCountEntry *)
+ hash_search(spcBuffers, (void *) &spc, HASH_ENTER, &found);
+
+ if (found) entry->count++;
+ else entry->count = 1;
}
UnlockBufHdr(bufHdr);
}
if (num_to_write == 0)
+ {
+ hash_destroy(spcBuffers);
return; /* nothing to do */
+ }
TRACE_POSTGRESQL_BUFFER_SYNC_START(NBuffers, num_to_write);
+ /* Build checkpoint tablespace buffer status */
+ nb_spaces = hash_get_num_entries(spcBuffers);
+ spcStatus = (TableSpaceCheckpointStatus *)
+ palloc(sizeof(TableSpaceCheckpointStatus) * nb_spaces);
+
+ {
+ int index = 0;
+ HASH_SEQ_STATUS hseq;
+ TableSpaceCountEntry * entry;
+
+ hash_seq_init(&hseq, spcBuffers);
+ while ((entry = (TableSpaceCountEntry *) hash_seq_search(&hseq)))
+ {
+ Assert(index < nb_spaces);
+ spcStatus[index].space = entry->space;
+ spcStatus[index].num_to_write = entry->count;
+ spcStatus[index].num_written = 0;
+ /* should it be randomized? chosen with some criterion? */
+ spcStatus[index].index = 0;
+
+ index ++;
+ }
+ }
+
+ hash_destroy(spcBuffers);
+ spcBuffers = NULL;
+
/*
- * Loop over all buffers again, and write the ones (still) marked with
- * BM_CHECKPOINT_NEEDED. In this loop, we start at the clock sweep point
- * since we might as well dump soon-to-be-recycled buffers first.
+ * Sort buffer ids to help find sequential writes.
*
- * Note that we don't read the buffer alloc count here --- that should be
- * left untouched till the next BgBufferSync() call.
+ * Note: Buffers are not locked in any way during sorting, but that's ok:
+ * A change in the buffer header is only relevant when it changes the
+ * buffer's identity. If the identity has changed it'll have been
+ * written out by BufferAlloc(), so there's no need for checkpointer to
+ * write it out anymore. The buffer might also get written out by a
+ * backend or bgwriter, but that's equally harmless.
+ *
+ * Marked buffers must not be move during the checkpoint.
+ * Also, qsort implementation should be resilient to occasional
+ * contradictions (cmp(a,b) != -cmp(b,a)) because of possible
+ * concurrent changes.
*/
- buf_id = StrategySyncStart(NULL, NULL);
- num_to_scan = NBuffers;
+ CheckpointStats.ckpt_sort_t = GetCurrentTimestamp();
+
+ if (checkpoint_sort)
+ {
+ qsort(CheckpointBufferIds, num_to_write, sizeof(CheckpointSortItem),
+ bufcmp);
+ }
+
+ CheckpointStats.ckpt_sort_end_t = GetCurrentTimestamp();
+
+ /*
+ * Loop over buffers to write through CheckpointBufferIds,
+ * and write the ones (still) marked with BM_CHECKPOINT_NEEDED,
+ * with some round robin over table spaces so as to balance writes,
+ * so that buffer writes move forward roughly proportionally for each
+ * tablespace.
+ *
+ * Termination: if a tablespace is selected by the inner while loop
+ * (see argument there), its index is incremented and will eventually
+ * reach num_to_write, mark this table space scanning as done and
+ * decrement the number of (active) spaces, which will thus reach 0.
+ */
+ space = 0;
num_written = 0;
- while (num_to_scan-- > 0)
+
+ while (nb_spaces != 0)
{
- volatile BufferDesc *bufHdr = GetBufferDescriptor(buf_id);
+ volatile BufferDesc *bufHdr = NULL;
+ buf_id = NextBufferToWrite(spcStatus, nb_spaces, &space,
+ num_to_write, num_written);
+ if (buf_id != -1)
+ bufHdr = GetBufferDescriptor(buf_id);
/*
* We don't need to acquire the lock here, because we're only looking
@@ -1660,39 +1880,45 @@ BufferSync(int flags)
* write the buffer though we didn't need to. It doesn't seem worth
* guarding against this, though.
*/
- if (bufHdr->flags & BM_CHECKPOINT_NEEDED)
+ if (bufHdr != NULL && bufHdr->flags & BM_CHECKPOINT_NEEDED)
{
if (SyncOneBuffer(buf_id, false) & BUF_WRITTEN)
{
TRACE_POSTGRESQL_BUFFER_SYNC_WRITTEN(buf_id);
BgWriterStats.m_buf_written_checkpoints++;
+ spcStatus[space].num_written++;
num_written++;
/*
- * We know there are at most num_to_write buffers with
- * BM_CHECKPOINT_NEEDED set; so we can stop scanning if
- * num_written reaches num_to_write.
- *
- * Note that num_written doesn't include buffers written by
- * other backends, or by the bgwriter cleaning scan. That
- * means that the estimate of how much progress we've made is
- * conservative, and also that this test will often fail to
- * trigger. But it seems worth making anyway.
- */
- if (num_written >= num_to_write)
- break;
-
- /*
* Sleep to throttle our I/O rate.
*/
CheckpointWriteDelay(flags, (double) num_written / num_to_write);
}
}
- if (++buf_id >= NBuffers)
- buf_id = 0;
+ /*
+ * Detect checkpoint end for a tablespace: either the scan is done
+ * or all tablespace buffers have been written out. If so, the
+ * another active tablespace status is moved in place of the current
+ * one and the next round will start on this one, or maybe round about.
+ * Note: maybe an exchange could be made instead in order to keep
+ * informations about the closed table space, but this is currently
+ * not used afterwards.
+ */
+ if (spcStatus[space].index >= num_to_write ||
+ spcStatus[space].num_written >= spcStatus[space].num_to_write)
+ {
+ nb_spaces--;
+ if (space != nb_spaces)
+ spcStatus[space] = spcStatus[nb_spaces];
+ else
+ space = 0;
+ }
}
+ pfree(spcStatus);
+ spcStatus = NULL;
+
/*
* Update checkpoint statistics. As noted above, this doesn't include
* buffers written by other backends or bgwriter scan.
diff --git a/src/backend/utils/misc/guc.c b/src/backend/utils/misc/guc.c
index b3dac51..cf1e505 100644
--- a/src/backend/utils/misc/guc.c
+++ b/src/backend/utils/misc/guc.c
@@ -1013,6 +1013,17 @@ static struct config_bool ConfigureNamesBool[] =
false,
NULL, NULL, NULL
},
+
+ {
+ {"checkpoint_sort", PGC_SIGHUP, WAL_CHECKPOINTS,
+ gettext_noop("Whether disk-page buffers are sorted on checkpoints."),
+ NULL
+ },
+ &checkpoint_sort,
+ true,
+ NULL, NULL, NULL
+ },
+
{
{"log_connections", PGC_SU_BACKEND, LOGGING_WHAT,
gettext_noop("Logs each successful connection."),
diff --git a/src/backend/utils/misc/postgresql.conf.sample b/src/backend/utils/misc/postgresql.conf.sample
index e5d275d..e84f380 100644
--- a/src/backend/utils/misc/postgresql.conf.sample
+++ b/src/backend/utils/misc/postgresql.conf.sample
@@ -201,6 +201,7 @@
#max_wal_size = 1GB
#min_wal_size = 80MB
#checkpoint_completion_target = 0.5 # checkpoint target duration, 0.0 - 1.0
+#checkpoint_sort = on # sort buffers on checkpoint
#checkpoint_warning = 30s # 0 disables
# - Archiving -
diff --git a/src/include/access/xlog.h b/src/include/access/xlog.h
index 6dacee2..dbd4757 100644
--- a/src/include/access/xlog.h
+++ b/src/include/access/xlog.h
@@ -186,6 +186,8 @@ extern bool XLOG_DEBUG;
typedef struct CheckpointStatsData
{
TimestampTz ckpt_start_t; /* start of checkpoint */
+ TimestampTz ckpt_sort_t; /* start buffer sorting */
+ TimestampTz ckpt_sort_end_t; /* end of sorting */
TimestampTz ckpt_write_t; /* start of flushing buffers */
TimestampTz ckpt_sync_t; /* start of fsyncs */
TimestampTz ckpt_sync_end_t; /* end of fsyncs */
diff --git a/src/include/storage/buf_internals.h b/src/include/storage/buf_internals.h
index 521ee1c..7fde0dc 100644
--- a/src/include/storage/buf_internals.h
+++ b/src/include/storage/buf_internals.h
@@ -210,6 +210,22 @@ extern PGDLLIMPORT BufferDescPadded *BufferDescriptors;
/* in localbuf.c */
extern BufferDesc *LocalBufferDescriptors;
+/* in bufmgr.c */
+
+/*
+ * Structure to sort buffers per file on checkpoints.
+ *
+ * Maybe the sort criterion could be compacted to reduce memory requirement
+ * and for faster comparison?
+ */
+typedef struct CheckpointSortItem {
+ int buf_id;
+ Oid relNode;
+ ForkNumber forkNum; /* only 4 values */
+ BlockNumber blockNum;
+} CheckpointSortItem;
+
+extern CheckpointSortItem *CheckpointBufferIds;
/*
* Internal routines: only called by bufmgr
diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index ec0a254..c228f39 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -54,6 +54,7 @@ extern int bgwriter_lru_maxpages;
extern double bgwriter_lru_multiplier;
extern bool track_io_timing;
extern int target_prefetch_pages;
+extern bool checkpoint_sort;
/* in buf_init.c */
extern PGDLLIMPORT char *BufferBlocks;
diff --git a/doc/src/sgml/config.sgml b/doc/src/sgml/config.sgml
index 1cec243..917b2fb 100644
--- a/doc/src/sgml/config.sgml
+++ b/doc/src/sgml/config.sgml
@@ -2497,6 +2497,24 @@ include_dir 'conf.d'
</listitem>
</varlistentry>
+ <varlistentry id="guc-checkpoint-flush-to-disk" xreflabel="checkpoint_flush_to_disk">
+ <term><varname>checkpoint_flush_to_disk</varname> (<type>bool</type>)
+ <indexterm>
+ <primary><varname>checkpoint_flush_to_disk</> configuration parameter</primary>
+ </indexterm>
+ </term>
+ <listitem>
+ <para>
+ When writing data for a checkpoint, hint the underlying OS that the
+ data must be sent to disk as soon as possible. This may help smoothing
+ disk I/O writes and avoid a stall when fsync is issued at the end of
+ the checkpoint, but it may also reduce average performance.
+ This setting may have no effect on some platforms.
+ The default is <literal>on</> on Linux, <literal>off</> otherwise.
+ </para>
+ </listitem>
+ </varlistentry>
+
<varlistentry id="guc-min-wal-size" xreflabel="min_wal_size">
<term><varname>min_wal_size</varname> (<type>integer</type>)
<indexterm>
diff --git a/doc/src/sgml/wal.sgml b/doc/src/sgml/wal.sgml
index f538698..eea6668 100644
--- a/doc/src/sgml/wal.sgml
+++ b/doc/src/sgml/wal.sgml
@@ -558,6 +558,17 @@
</para>
<para>
+ On Linux and POSIX platforms, <xref linkend="guc-checkpoint-flush-to-disk">
+ allows to hint the OS that pages written on checkpoints must be flushed
+ to disk quickly. Otherwise, these pages may be kept in cache for some time,
+ inducing a stall later when <literal>fsync</> is called to actually
+ complete the checkpoint. This setting helps to reduce transaction latency,
+ but it may also have a small adverse effect on the average transaction rate
+ at maximum throughput. This feature probably brings no benefit on SSD,
+ as the I/O write latency is small on such hardware, thus it may be disabled.
+ </para>
+
+ <para>
The number of WAL segment files in <filename>pg_xlog</> directory depends on
<varname>min_wal_size</>, <varname>max_wal_size</> and
the amount of WAL generated in previous checkpoint cycles. When old log
diff --git a/src/backend/access/heap/rewriteheap.c b/src/backend/access/heap/rewriteheap.c
index bcce3e3..f565dc4 100644
--- a/src/backend/access/heap/rewriteheap.c
+++ b/src/backend/access/heap/rewriteheap.c
@@ -918,7 +918,7 @@ logical_heap_rewrite_flush_mappings(RewriteState state)
* Note that we deviate from the usual WAL coding practices here,
* check the above "Logical rewrite support" comment for reasoning.
*/
- written = FileWrite(src->vfd, waldata_start, len);
+ written = FileWrite(src->vfd, waldata_start, len, false, NULL);
if (written != len)
ereport(ERROR,
(errcode_for_file_access(),
diff --git a/src/backend/access/nbtree/nbtree.c b/src/backend/access/nbtree/nbtree.c
index cf4a6dc..4b5e9cd 100644
--- a/src/backend/access/nbtree/nbtree.c
+++ b/src/backend/access/nbtree/nbtree.c
@@ -203,7 +203,7 @@ btbuildempty(PG_FUNCTION_ARGS)
/* Write the page. If archiving/streaming, XLOG it. */
PageSetChecksumInplace(metapage, BTREE_METAPAGE);
smgrwrite(index->rd_smgr, INIT_FORKNUM, BTREE_METAPAGE,
- (char *) metapage, true);
+ (char *) metapage, true, false, NULL);
if (XLogIsNeeded())
log_newpage(&index->rd_smgr->smgr_rnode.node, INIT_FORKNUM,
BTREE_METAPAGE, metapage, false);
diff --git a/src/backend/access/nbtree/nbtsort.c b/src/backend/access/nbtree/nbtsort.c
index f95f67a..ea7a45d 100644
--- a/src/backend/access/nbtree/nbtsort.c
+++ b/src/backend/access/nbtree/nbtsort.c
@@ -315,7 +315,7 @@ _bt_blwritepage(BTWriteState *wstate, Page page, BlockNumber blkno)
{
/* overwriting a block we zero-filled before */
smgrwrite(wstate->index->rd_smgr, MAIN_FORKNUM, blkno,
- (char *) page, true);
+ (char *) page, true, false, NULL);
}
pfree(page);
diff --git a/src/backend/access/spgist/spginsert.c b/src/backend/access/spgist/spginsert.c
index bceee8d..b700efb 100644
--- a/src/backend/access/spgist/spginsert.c
+++ b/src/backend/access/spgist/spginsert.c
@@ -170,7 +170,7 @@ spgbuildempty(PG_FUNCTION_ARGS)
/* Write the page. If archiving/streaming, XLOG it. */
PageSetChecksumInplace(page, SPGIST_METAPAGE_BLKNO);
smgrwrite(index->rd_smgr, INIT_FORKNUM, SPGIST_METAPAGE_BLKNO,
- (char *) page, true);
+ (char *) page, true, false, NULL);
if (XLogIsNeeded())
log_newpage(&index->rd_smgr->smgr_rnode.node, INIT_FORKNUM,
SPGIST_METAPAGE_BLKNO, page, false);
@@ -180,7 +180,7 @@ spgbuildempty(PG_FUNCTION_ARGS)
PageSetChecksumInplace(page, SPGIST_ROOT_BLKNO);
smgrwrite(index->rd_smgr, INIT_FORKNUM, SPGIST_ROOT_BLKNO,
- (char *) page, true);
+ (char *) page, true, false, NULL);
if (XLogIsNeeded())
log_newpage(&index->rd_smgr->smgr_rnode.node, INIT_FORKNUM,
SPGIST_ROOT_BLKNO, page, true);
@@ -190,7 +190,7 @@ spgbuildempty(PG_FUNCTION_ARGS)
PageSetChecksumInplace(page, SPGIST_NULL_BLKNO);
smgrwrite(index->rd_smgr, INIT_FORKNUM, SPGIST_NULL_BLKNO,
- (char *) page, true);
+ (char *) page, true, false, NULL);
if (XLogIsNeeded())
log_newpage(&index->rd_smgr->smgr_rnode.node, INIT_FORKNUM,
SPGIST_NULL_BLKNO, page, true);
diff --git a/src/backend/postmaster/checkpointer.c b/src/backend/postmaster/checkpointer.c
index 3b3a09e..e361907 100644
--- a/src/backend/postmaster/checkpointer.c
+++ b/src/backend/postmaster/checkpointer.c
@@ -665,7 +665,8 @@ ImmediateCheckpointRequested(void)
* fraction between 0.0 meaning none, and 1.0 meaning all done.
*/
void
-CheckpointWriteDelay(int flags, double progress)
+CheckpointWriteDelay(int flags, double progress,
+ FileFlushContext * context, int ctx_size)
{
static int absorb_counter = WRITES_PER_ABSORB;
@@ -700,6 +701,26 @@ CheckpointWriteDelay(int flags, double progress)
*/
pgstat_send_bgwriter();
+#if defined(HAVE_SYNC_FILE_RANGE) || defined(HAVE_POSIX_FADVISE)
+
+ /*
+ * Before sleeping, flush written blocks for each tablespace.
+ */
+ if (checkpoint_flush_to_disk)
+ {
+ int i;
+
+ for (i = 0; i < ctx_size; i++)
+ {
+ if (context[i].ncalls != 0)
+ {
+ PerformFileFlush(&context[i]);
+ ResetFileFlushContext(&context[i]);
+ }
+ }
+ }
+#endif /* HAVE_SYNC_FILE_RANGE || HAVE_POSIX_FADVISE */
+
/*
* This sleep used to be connected to bgwriter_delay, typically 200ms.
* That resulted in more frequent wakeups if not much work to do.
diff --git a/src/backend/storage/buffer/bufmgr.c b/src/backend/storage/buffer/bufmgr.c
index ca295f1..3bd2043 100644
--- a/src/backend/storage/buffer/bufmgr.c
+++ b/src/backend/storage/buffer/bufmgr.c
@@ -80,6 +80,8 @@ bool zero_damaged_pages = false;
int bgwriter_lru_maxpages = 100;
double bgwriter_lru_multiplier = 2.0;
bool track_io_timing = false;
+/* hint to move writes to high priority */
+bool checkpoint_flush_to_disk = DEFAULT_CHECKPOINT_FLUSH_TO_DISK;
bool checkpoint_sort = true;
/*
@@ -400,7 +402,8 @@ static bool PinBuffer(volatile BufferDesc *buf, BufferAccessStrategy strategy);
static void PinBuffer_Locked(volatile BufferDesc *buf);
static void UnpinBuffer(volatile BufferDesc *buf, bool fixOwner);
static void BufferSync(int flags);
-static int SyncOneBuffer(int buf_id, bool skip_recently_used);
+static int SyncOneBuffer(int buf_id, bool skip_recently_used,
+ bool flush_to_disk, FileFlushContext *context);
static void WaitIO(volatile BufferDesc *buf);
static bool StartBufferIO(volatile BufferDesc *buf, bool forInput);
static void TerminateBufferIO(volatile BufferDesc *buf, bool clear_dirty,
@@ -413,7 +416,8 @@ static volatile BufferDesc *BufferAlloc(SMgrRelation smgr,
BlockNumber blockNum,
BufferAccessStrategy strategy,
bool *foundPtr);
-static void FlushBuffer(volatile BufferDesc *buf, SMgrRelation reln);
+static void FlushBuffer(volatile BufferDesc *buf, SMgrRelation reln,
+ bool flush_to_disk, FileFlushContext *context);
static void AtProcExit_Buffers(int code, Datum arg);
static void CheckForBufferLeaks(void);
static int rnode_comparator(const void *p1, const void *p2);
@@ -1022,7 +1026,7 @@ BufferAlloc(SMgrRelation smgr, char relpersistence, ForkNumber forkNum,
smgr->smgr_rnode.node.dbNode,
smgr->smgr_rnode.node.relNode);
- FlushBuffer(buf, NULL);
+ FlushBuffer(buf, NULL, false, NULL);
LWLockRelease(buf->content_lock);
TRACE_POSTGRESQL_BUFFER_WRITE_DIRTY_DONE(forkNum, blockNum,
@@ -1708,6 +1712,7 @@ BufferSync(int flags)
HTAB *spcBuffers;
TableSpaceCheckpointStatus *spcStatus = NULL;
int nb_spaces, space;
+ FileFlushContext * spcContext = NULL;
/* Make sure we can handle the pin inside SyncOneBuffer */
ResourceOwnerEnlargeBuffers(CurrentResourceOwner);
@@ -1793,10 +1798,12 @@ BufferSync(int flags)
TRACE_POSTGRESQL_BUFFER_SYNC_START(NBuffers, num_to_write);
- /* Build checkpoint tablespace buffer status */
+ /* Build checkpoint tablespace buffer status & flush context arrays */
nb_spaces = hash_get_num_entries(spcBuffers);
spcStatus = (TableSpaceCheckpointStatus *)
palloc(sizeof(TableSpaceCheckpointStatus) * nb_spaces);
+ spcContext = (FileFlushContext *)
+ palloc(sizeof(FileFlushContext) * nb_spaces);
{
int index = 0;
@@ -1813,6 +1820,12 @@ BufferSync(int flags)
/* should it be randomized? chosen with some criterion? */
spcStatus[index].index = 0;
+#if defined(HAVE_SYNC_FILE_RANGE) || defined(HAVE_POSIX_FADVISE)
+
+ ResetFileFlushContext(&spcContext[index]);
+
+#endif /* HAVE_SYNC_FILE_RANGE || HAVE_POSIX_FADVISE */
+
index ++;
}
}
@@ -1882,7 +1895,8 @@ BufferSync(int flags)
*/
if (bufHdr != NULL && bufHdr->flags & BM_CHECKPOINT_NEEDED)
{
- if (SyncOneBuffer(buf_id, false) & BUF_WRITTEN)
+ if (SyncOneBuffer(buf_id, false, checkpoint_flush_to_disk,
+ &spcContext[space]) & BUF_WRITTEN)
{
TRACE_POSTGRESQL_BUFFER_SYNC_WRITTEN(buf_id);
BgWriterStats.m_buf_written_checkpoints++;
@@ -1892,7 +1906,8 @@ BufferSync(int flags)
/*
* Sleep to throttle our I/O rate.
*/
- CheckpointWriteDelay(flags, (double) num_written / num_to_write);
+ CheckpointWriteDelay(flags, (double) num_written / num_to_write,
+ spcContext, nb_spaces);
}
}
@@ -1908,6 +1923,13 @@ BufferSync(int flags)
if (spcStatus[space].index >= num_to_write ||
spcStatus[space].num_written >= spcStatus[space].num_to_write)
{
+#if defined(HAVE_SYNC_FILE_RANGE) || defined(HAVE_POSIX_FADVISE)
+
+ PerformFileFlush(&spcContext[space]);
+ ResetFileFlushContext(&spcContext[space]);
+
+#endif /* HAVE_SYNC_FILE_RANGE || HAVE_POSIX_FADVISE */
+
nb_spaces--;
if (space != nb_spaces)
spcStatus[space] = spcStatus[nb_spaces];
@@ -1918,6 +1940,8 @@ BufferSync(int flags)
pfree(spcStatus);
spcStatus = NULL;
+ pfree(spcContext);
+ spcContext = NULL;
/*
* Update checkpoint statistics. As noted above, this doesn't include
@@ -2165,7 +2189,8 @@ BgBufferSync(void)
/* Execute the LRU scan */
while (num_to_scan > 0 && reusable_buffers < upcoming_alloc_est)
{
- int buffer_state = SyncOneBuffer(next_to_clean, true);
+ int buffer_state =
+ SyncOneBuffer(next_to_clean, true, false, NULL);
if (++next_to_clean >= NBuffers)
{
@@ -2242,7 +2267,8 @@ BgBufferSync(void)
* Note: caller must have done ResourceOwnerEnlargeBuffers.
*/
static int
-SyncOneBuffer(int buf_id, bool skip_recently_used)
+SyncOneBuffer(int buf_id, bool skip_recently_used, bool flush_to_disk,
+ FileFlushContext * context)
{
volatile BufferDesc *bufHdr = GetBufferDescriptor(buf_id);
int result = 0;
@@ -2283,7 +2309,7 @@ SyncOneBuffer(int buf_id, bool skip_recently_used)
PinBuffer_Locked(bufHdr);
LWLockAcquire(bufHdr->content_lock, LW_SHARED);
- FlushBuffer(bufHdr, NULL);
+ FlushBuffer(bufHdr, NULL, flush_to_disk, context);
LWLockRelease(bufHdr->content_lock);
UnpinBuffer(bufHdr, true);
@@ -2545,9 +2571,16 @@ BufferGetTag(Buffer buffer, RelFileNode *rnode, ForkNumber *forknum,
*
* If the caller has an smgr reference for the buffer's relation, pass it
* as the second parameter. If not, pass NULL.
+ *
+ * The third parameter tries to hint the OS that a high priority write is meant,
+ * possibly because io-throttling is already managed elsewhere.
+ * The last parameter holds the current flush context that accumulates flush
+ * requests to be performed in one call, instead of being performed on a buffer
+ * per buffer basis.
*/
static void
-FlushBuffer(volatile BufferDesc *buf, SMgrRelation reln)
+FlushBuffer(volatile BufferDesc *buf, SMgrRelation reln, bool flush_to_disk,
+ FileFlushContext * context)
{
XLogRecPtr recptr;
ErrorContextCallback errcallback;
@@ -2636,7 +2669,9 @@ FlushBuffer(volatile BufferDesc *buf, SMgrRelation reln)
buf->tag.forkNum,
buf->tag.blockNum,
bufToWrite,
- false);
+ false,
+ flush_to_disk,
+ context);
if (track_io_timing)
{
@@ -3058,7 +3093,9 @@ FlushRelationBuffers(Relation rel)
bufHdr->tag.forkNum,
bufHdr->tag.blockNum,
localpage,
- false);
+ false,
+ false,
+ NULL);
bufHdr->flags &= ~(BM_DIRTY | BM_JUST_DIRTIED);
@@ -3092,7 +3129,7 @@ FlushRelationBuffers(Relation rel)
{
PinBuffer_Locked(bufHdr);
LWLockAcquire(bufHdr->content_lock, LW_SHARED);
- FlushBuffer(bufHdr, rel->rd_smgr);
+ FlushBuffer(bufHdr, rel->rd_smgr, false, NULL);
LWLockRelease(bufHdr->content_lock);
UnpinBuffer(bufHdr, true);
}
@@ -3144,7 +3181,7 @@ FlushDatabaseBuffers(Oid dbid)
{
PinBuffer_Locked(bufHdr);
LWLockAcquire(bufHdr->content_lock, LW_SHARED);
- FlushBuffer(bufHdr, NULL);
+ FlushBuffer(bufHdr, NULL, false, NULL);
LWLockRelease(bufHdr->content_lock);
UnpinBuffer(bufHdr, true);
}
diff --git a/src/backend/storage/buffer/localbuf.c b/src/backend/storage/buffer/localbuf.c
index 3144afe..114a0a6 100644
--- a/src/backend/storage/buffer/localbuf.c
+++ b/src/backend/storage/buffer/localbuf.c
@@ -208,7 +208,9 @@ LocalBufferAlloc(SMgrRelation smgr, ForkNumber forkNum, BlockNumber blockNum,
bufHdr->tag.forkNum,
bufHdr->tag.blockNum,
localpage,
- false);
+ false,
+ false,
+ NULL);
/* Mark not-dirty now in case we error out below */
bufHdr->flags &= ~BM_DIRTY;
diff --git a/src/backend/storage/file/buffile.c b/src/backend/storage/file/buffile.c
index ea4d689..fb3b383 100644
--- a/src/backend/storage/file/buffile.c
+++ b/src/backend/storage/file/buffile.c
@@ -317,7 +317,7 @@ BufFileDumpBuffer(BufFile *file)
return; /* seek failed, give up */
file->offsets[file->curFile] = file->curOffset;
}
- bytestowrite = FileWrite(thisfile, file->buffer + wpos, bytestowrite);
+ bytestowrite = FileWrite(thisfile, file->buffer + wpos, bytestowrite, false, NULL);
if (bytestowrite <= 0)
return; /* failed to write */
file->offsets[file->curFile] += bytestowrite;
diff --git a/src/backend/storage/file/fd.c b/src/backend/storage/file/fd.c
index 1ba4946..daf03e4 100644
--- a/src/backend/storage/file/fd.c
+++ b/src/backend/storage/file/fd.c
@@ -1344,8 +1344,97 @@ retry:
return returnCode;
}
+#if defined(HAVE_SYNC_FILE_RANGE) || defined(HAVE_POSIX_FADVISE)
+
+void
+ResetFileFlushContext(FileFlushContext * context)
+{
+ context->fd = 0;
+ context->ncalls = 0;
+ context->offset = 0;
+ context->nbytes = 0;
+ context->filename = NULL;
+}
+
+void
+PerformFileFlush(FileFlushContext * context)
+{
+ if (context->ncalls != 0)
+ {
+ int rc;
+
+#if defined(HAVE_SYNC_FILE_RANGE)
+
+ /* Linux: tell the memory manager to move these blocks to io so
+ * that they are considered for being actually written to disk.
+ */
+ rc = sync_file_range(context->fd, context->offset, context->nbytes,
+ SYNC_FILE_RANGE_WRITE);
+
+#elif defined(HAVE_POSIX_FADVISE)
+
+ /* Others: say that data should not be kept in memory...
+ * This is not exactly what we want to say, because we want to write
+ * the data for durability but we may need it later nevertheless.
+ * It seems that Linux would free the memory *if* the data has
+ * already been written do disk, else the "dontneed" call is ignored.
+ * For FreeBSD this may have the desired effect of moving the
+ * data to the io layer, although the system does not seem to
+ * take into account the provided offset & size, so it is rather
+ * rough...
+ */
+ rc = posix_fadvise(context->fd, context->offset, context->nbytes,
+ POSIX_FADV_DONTNEED);
+
+#endif
+
+ if (rc < 0)
+ ereport(ERROR,
+ (errcode_for_file_access(),
+ errmsg("could not flush block " INT64_FORMAT
+ " on " INT64_FORMAT " blocks in file \"%s\": %m",
+ context->offset / BLCKSZ,
+ context->nbytes / BLCKSZ,
+ context->filename)));
+ }
+}
+
+void
+FileAsynchronousFlush(FileFlushContext * context,
+ int fd, off_t offset, off_t nbytes, char * filename)
+{
+ if (context->ncalls != 0 && context->fd == fd)
+ {
+ /* Same file: merge current flush with previous ones */
+ off_t new_offset = offset < context->offset? offset: context->offset;
+
+ context->nbytes =
+ (context->offset + context->nbytes > offset + nbytes ?
+ context->offset + context->nbytes : offset + nbytes) -
+ new_offset;
+ context->offset = new_offset;
+ context->ncalls ++;
+ }
+ else
+ {
+ /* file has changed; actually flush previous file before restarting
+ * to accumulate flushes
+ */
+ PerformFileFlush(context);
+
+ context->fd = fd;
+ context->ncalls = 1;
+ context->offset = offset;
+ context->nbytes = nbytes;
+ context->filename = filename;
+ }
+}
+
+#endif /* HAVE_SYNC_FILE_RANGE || HAVE_POSIX_FADVISE */
+
int
-FileWrite(File file, char *buffer, int amount)
+FileWrite(File file, char *buffer, int amount, bool flush_to_disk,
+ FileFlushContext * context)
{
int returnCode;
@@ -1395,6 +1484,28 @@ retry:
if (returnCode >= 0)
{
+
+#if defined(HAVE_SYNC_FILE_RANGE) || defined(HAVE_POSIX_FADVISE)
+
+ /*
+ * Calling "write" tells the OS that pg wants to write some page to disk,
+ * however when it is really done is chosen by the OS.
+ * Depending on other disk activities this may be delayed significantly,
+ * maybe up to an "fsync" call, which could induce an IO write surge.
+ * When checkpointing pg is doing its own throttling and the result
+ * should really be written to disk with high priority, so as to meet
+ * the completion target.
+ * This call hints that such write have a higher priority.
+ */
+ if (flush_to_disk && returnCode == amount && errno == 0)
+ {
+ FileAsynchronousFlush(context,
+ VfdCache[file].fd, VfdCache[file].seekPos,
+ amount, VfdCache[file].fileName);
+ }
+
+#endif /* HAVE_SYNC_FILE_RANGE || HAVE_POSIX_FADVISE */
+
VfdCache[file].seekPos += returnCode;
/* maintain fileSize and temporary_files_size if it's a temp file */
diff --git a/src/backend/storage/smgr/md.c b/src/backend/storage/smgr/md.c
index 42a43bb..dbf057f 100644
--- a/src/backend/storage/smgr/md.c
+++ b/src/backend/storage/smgr/md.c
@@ -531,7 +531,7 @@ mdextend(SMgrRelation reln, ForkNumber forknum, BlockNumber blocknum,
errmsg("could not seek to block %u in file \"%s\": %m",
blocknum, FilePathName(v->mdfd_vfd))));
- if ((nbytes = FileWrite(v->mdfd_vfd, buffer, BLCKSZ)) != BLCKSZ)
+ if ((nbytes = FileWrite(v->mdfd_vfd, buffer, BLCKSZ, false, NULL)) != BLCKSZ)
{
if (nbytes < 0)
ereport(ERROR,
@@ -738,7 +738,8 @@ mdread(SMgrRelation reln, ForkNumber forknum, BlockNumber blocknum,
*/
void
mdwrite(SMgrRelation reln, ForkNumber forknum, BlockNumber blocknum,
- char *buffer, bool skipFsync)
+ char *buffer, bool skipFsync, bool flush_to_disk,
+ FileFlushContext * context)
{
off_t seekpos;
int nbytes;
@@ -767,7 +768,7 @@ mdwrite(SMgrRelation reln, ForkNumber forknum, BlockNumber blocknum,
errmsg("could not seek to block %u in file \"%s\": %m",
blocknum, FilePathName(v->mdfd_vfd))));
- nbytes = FileWrite(v->mdfd_vfd, buffer, BLCKSZ);
+ nbytes = FileWrite(v->mdfd_vfd, buffer, BLCKSZ, flush_to_disk, context);
TRACE_POSTGRESQL_SMGR_MD_WRITE_DONE(forknum, blocknum,
reln->smgr_rnode.node.spcNode,
diff --git a/src/backend/storage/smgr/smgr.c b/src/backend/storage/smgr/smgr.c
index 244b4ea..2db3cd3 100644
--- a/src/backend/storage/smgr/smgr.c
+++ b/src/backend/storage/smgr/smgr.c
@@ -52,7 +52,8 @@ typedef struct f_smgr
void (*smgr_read) (SMgrRelation reln, ForkNumber forknum,
BlockNumber blocknum, char *buffer);
void (*smgr_write) (SMgrRelation reln, ForkNumber forknum,
- BlockNumber blocknum, char *buffer, bool skipFsync);
+ BlockNumber blocknum, char *buffer, bool skipFsync,
+ bool flush_to_disk, FileFlushContext *context);
BlockNumber (*smgr_nblocks) (SMgrRelation reln, ForkNumber forknum);
void (*smgr_truncate) (SMgrRelation reln, ForkNumber forknum,
BlockNumber nblocks);
@@ -643,10 +644,11 @@ smgrread(SMgrRelation reln, ForkNumber forknum, BlockNumber blocknum,
*/
void
smgrwrite(SMgrRelation reln, ForkNumber forknum, BlockNumber blocknum,
- char *buffer, bool skipFsync)
+ char *buffer, bool skipFsync, bool flush_to_disk,
+ FileFlushContext * context)
{
(*(smgrsw[reln->smgr_which].smgr_write)) (reln, forknum, blocknum,
- buffer, skipFsync);
+ buffer, skipFsync, flush_to_disk, context);
}
/*
diff --git a/src/backend/utils/misc/guc.c b/src/backend/utils/misc/guc.c
index cf1e505..94b0d5b 100644
--- a/src/backend/utils/misc/guc.c
+++ b/src/backend/utils/misc/guc.c
@@ -158,6 +158,7 @@ static bool check_bonjour(bool *newval, void **extra, GucSource source);
static bool check_ssl(bool *newval, void **extra, GucSource source);
static bool check_stage_log_stats(bool *newval, void **extra, GucSource source);
static bool check_log_stats(bool *newval, void **extra, GucSource source);
+static bool check_flush_to_disk(bool *newval, void **extra, GucSource source);
static bool check_canonical_path(char **newval, void **extra, GucSource source);
static bool check_timezone_abbreviations(char **newval, void **extra, GucSource source);
static void assign_timezone_abbreviations(const char *newval, void *extra);
@@ -1025,6 +1026,17 @@ static struct config_bool ConfigureNamesBool[] =
},
{
+ {"checkpoint_flush_to_disk", PGC_SIGHUP, WAL_CHECKPOINTS,
+ gettext_noop("Hint that checkpoint's writes are high priority."),
+ NULL
+ },
+ &checkpoint_flush_to_disk,
+ /* see bufmgr.h: true on Linux, false otherwise */
+ DEFAULT_CHECKPOINT_FLUSH_TO_DISK,
+ check_flush_to_disk, NULL, NULL
+ },
+
+ {
{"log_connections", PGC_SU_BACKEND, LOGGING_WHAT,
gettext_noop("Logs each successful connection."),
NULL
@@ -9806,6 +9818,21 @@ check_log_stats(bool *newval, void **extra, GucSource source)
}
static bool
+check_flush_to_disk(bool *newval, void **extra, GucSource source)
+{
+/* This test must be consistent with the one in FileWrite (storage/file/fd.c)
+ */
+#if ! (defined(HAVE_SYNC_FILE_RANGE) || defined(HAVE_POSIX_FADVISE))
+ /* just warn if it has no effect */
+ ereport(WARNING,
+ (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
+ errmsg("Setting \"checkpoint_flush_to_disk\" has no effect "
+ "on this platform.")));
+#endif /* HAVE_SYNC_FILE_RANGE || HAVE_POSIX_FADVISE */
+ return true;
+}
+
+static bool
check_canonical_path(char **newval, void **extra, GucSource source)
{
/*
diff --git a/src/backend/utils/misc/postgresql.conf.sample b/src/backend/utils/misc/postgresql.conf.sample
index e84f380..a5495da 100644
--- a/src/backend/utils/misc/postgresql.conf.sample
+++ b/src/backend/utils/misc/postgresql.conf.sample
@@ -202,6 +202,8 @@
#min_wal_size = 80MB
#checkpoint_completion_target = 0.5 # checkpoint target duration, 0.0 - 1.0
#checkpoint_sort = on # sort buffers on checkpoint
+#checkpoint_flush_to_disk = ? # send buffers to disk on checkpoint
+ # default is on if Linux, off otherwise
#checkpoint_warning = 30s # 0 disables
# - Archiving -
diff --git a/src/include/postmaster/bgwriter.h b/src/include/postmaster/bgwriter.h
index a49c208..f9c8ca1 100644
--- a/src/include/postmaster/bgwriter.h
+++ b/src/include/postmaster/bgwriter.h
@@ -16,6 +16,7 @@
#define _BGWRITER_H
#include "storage/block.h"
+#include "storage/fd.h"
#include "storage/relfilenode.h"
@@ -29,7 +30,8 @@ extern void BackgroundWriterMain(void) pg_attribute_noreturn();
extern void CheckpointerMain(void) pg_attribute_noreturn();
extern void RequestCheckpoint(int flags);
-extern void CheckpointWriteDelay(int flags, double progress);
+extern void CheckpointWriteDelay(int flags, double progress,
+ FileFlushContext * context, int ctx_size);
extern bool ForwardFsyncRequest(RelFileNode rnode, ForkNumber forknum,
BlockNumber segno);
diff --git a/src/include/storage/bufmgr.h b/src/include/storage/bufmgr.h
index c228f39..4fd3ff5 100644
--- a/src/include/storage/bufmgr.h
+++ b/src/include/storage/bufmgr.h
@@ -54,6 +54,14 @@ extern int bgwriter_lru_maxpages;
extern double bgwriter_lru_multiplier;
extern bool track_io_timing;
extern int target_prefetch_pages;
+
+#ifdef HAVE_SYNC_FILE_RANGE
+#define DEFAULT_CHECKPOINT_FLUSH_TO_DISK true
+#else
+#define DEFAULT_CHECKPOINT_FLUSH_TO_DISK false
+#endif /* HAVE_SYNC_FILE_RANGE */
+
+extern bool checkpoint_flush_to_disk;
extern bool checkpoint_sort;
/* in buf_init.c */
diff --git a/src/include/storage/fd.h b/src/include/storage/fd.h
index 7eabe09..c740ee7 100644
--- a/src/include/storage/fd.h
+++ b/src/include/storage/fd.h
@@ -59,6 +59,22 @@ extern int max_files_per_process;
*/
extern int max_safe_fds;
+/* FileFlushContext:
+ * This structure is used to accumulate several flush requests on a file
+ * into a larger flush request.
+ * - fd: file descriptor of the file
+ * - ncalls: number of flushes merged together
+ * - offset: starting offset (minimum of all offset)
+ * - nbytes: size (minimum extent to cover all flushed data)
+ * - filename: filename of fd for error messages
+ */
+typedef struct FileFlushContext{
+ int fd;
+ int ncalls;
+ off_t offset;
+ off_t nbytes;
+ char * filename;
+} FileFlushContext;
/*
* prototypes for functions in fd.c
@@ -70,7 +86,12 @@ extern File OpenTemporaryFile(bool interXact);
extern void FileClose(File file);
extern int FilePrefetch(File file, off_t offset, int amount);
extern int FileRead(File file, char *buffer, int amount);
-extern int FileWrite(File file, char *buffer, int amount);
+extern void ResetFileFlushContext(FileFlushContext * context);
+extern void PerformFileFlush(FileFlushContext * context);
+extern void FileAsynchronousFlush(FileFlushContext * context,
+ int fd, off_t offset, off_t nbytes, char * filename);
+extern int FileWrite(File file, char *buffer, int amount, bool flush_to_disk,
+ FileFlushContext * context);
extern int FileSync(File file);
extern off_t FileSeek(File file, off_t offset, int whence);
extern int FileTruncate(File file, off_t offset);
diff --git a/src/include/storage/smgr.h b/src/include/storage/smgr.h
index 69a624f..a46a70c 100644
--- a/src/include/storage/smgr.h
+++ b/src/include/storage/smgr.h
@@ -16,6 +16,7 @@
#include "fmgr.h"
#include "storage/block.h"
+#include "storage/fd.h"
#include "storage/relfilenode.h"
@@ -95,7 +96,8 @@ extern void smgrprefetch(SMgrRelation reln, ForkNumber forknum,
extern void smgrread(SMgrRelation reln, ForkNumber forknum,
BlockNumber blocknum, char *buffer);
extern void smgrwrite(SMgrRelation reln, ForkNumber forknum,
- BlockNumber blocknum, char *buffer, bool skipFsync);
+ BlockNumber blocknum, char *buffer, bool skipFsync,
+ bool flush_to_disk, FileFlushContext * context);
extern BlockNumber smgrnblocks(SMgrRelation reln, ForkNumber forknum);
extern void smgrtruncate(SMgrRelation reln, ForkNumber forknum,
BlockNumber nblocks);
@@ -120,8 +122,9 @@ extern void mdprefetch(SMgrRelation reln, ForkNumber forknum,
BlockNumber blocknum);
extern void mdread(SMgrRelation reln, ForkNumber forknum, BlockNumber blocknum,
char *buffer);
-extern void mdwrite(SMgrRelation reln, ForkNumber forknum,
- BlockNumber blocknum, char *buffer, bool skipFsync);
+extern void mdwrite(SMgrRelation reln, ForkNumber forknum, BlockNumber blocknum,
+ char *buffer, bool skipFsync, bool flush_to_disk,
+ FileFlushContext * context);
extern BlockNumber mdnblocks(SMgrRelation reln, ForkNumber forknum);
extern void mdtruncate(SMgrRelation reln, ForkNumber forknum,
BlockNumber nblocks);
--
Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)
To make changes to your subscription:
http://www.postgresql.org/mailpref/pgsql-hackers
