Prevent excess SimpleLruTruncate() deletion.

Every core SLRU wraps around.  With the exception of pg_notify, the wrap
point can fall in the middle of a page.  Account for this in the
PagePrecedes callback specification and in SimpleLruTruncate()'s use of
said callback.  Update each callback implementation to fit the new
specification.  This changes SerialPagePrecedesLogically() from the
style of asyncQueuePagePrecedes() to the style of CLOGPagePrecedes().
(Whereas pg_clog and pg_serial share a key space, pg_serial is nothing
like pg_notify.)  The bug fixed here has the same symptoms and user
followup steps as 592a589.  Back-patch
to 9.5 (all supported versions).

Reviewed by Andrey Borodin and (in earlier versions) by Tom Lane.

Discussion: https://postgr.es/m/20190202083822.GC32531@gust.leadboat.com

Author: nmisch

src/backend/access/transam/clog.c

@@ -694,6 +694,7 @@ CLOGShmemInit(void)
 	SimpleLruInit(XactCtl, "Xact", CLOGShmemBuffers(), CLOG_LSNS_PER_PAGE,
 				  XactSLRULock, "pg_xact", LWTRANCHE_XACT_BUFFER,
 				  SYNC_HANDLER_CLOG);
+	SlruPagePrecedesUnitTests(XactCtl, CLOG_XACTS_PER_PAGE);
 }
 
 /*
@@ -912,13 +913,22 @@ TruncateCLOG(TransactionId oldestXact, Oid oldestxid_datoid)
 
 
 /*
- * Decide which of two CLOG page numbers is "older" for truncation purposes.
+ * Decide whether a CLOG page number is "older" for truncation purposes.
  *
  * We need to use comparison of TransactionIds here in order to do the right
- * thing with wraparound XID arithmetic.  However, if we are asked about
- * page number zero, we don't want to hand InvalidTransactionId to
- * TransactionIdPrecedes: it'll get weird about permanent xact IDs.  So,
- * offset both xids by FirstNormalTransactionId to avoid that.
+ * thing with wraparound XID arithmetic.  However, TransactionIdPrecedes()
+ * would get weird about permanent xact IDs.  So, offset both such that xid1,
+ * xid2, and xid2 + CLOG_XACTS_PER_PAGE - 1 are all normal XIDs; this offset
+ * is relevant to page 0 and to the page preceding page 0.
+ *
+ * The page containing oldestXact-2^31 is the important edge case.  The
+ * portion of that page equaling or following oldestXact-2^31 is expendable,
+ * but the portion preceding oldestXact-2^31 is not.  When oldestXact-2^31 is
+ * the first XID of a page and segment, the entire page and segment is
+ * expendable, and we could truncate the segment.  Recognizing that case would
+ * require making oldestXact, not just the page containing oldestXact,
+ * available to this callback.  The benefit would be rare and small, so we
+ * don't optimize that edge case.
  */
 static bool
 CLOGPagePrecedes(int page1, int page2)
@@ -927,11 +937,12 @@ CLOGPagePrecedes(int page1, int page2)
 	TransactionId xid2;
 
 	xid1 = ((TransactionId) page1) * CLOG_XACTS_PER_PAGE;
-	xid1 += FirstNormalTransactionId;
+	xid1 += FirstNormalTransactionId + 1;
 	xid2 = ((TransactionId) page2) * CLOG_XACTS_PER_PAGE;
-	xid2 += FirstNormalTransactionId;
+	xid2 += FirstNormalTransactionId + 1;
 
-	return TransactionIdPrecedes(xid1, xid2);
+	return (TransactionIdPrecedes(xid1, xid2) &&
+			TransactionIdPrecedes(xid1, xid2 + CLOG_XACTS_PER_PAGE - 1));
 }
 
 

src/backend/access/transam/commit_ts.c

@@ -557,6 +557,7 @@ CommitTsShmemInit(void)
 				  CommitTsSLRULock, "pg_commit_ts",
 				  LWTRANCHE_COMMITTS_BUFFER,
 				  SYNC_HANDLER_COMMIT_TS);
+	SlruPagePrecedesUnitTests(CommitTsCtl, COMMIT_TS_XACTS_PER_PAGE);
 
 	commitTsShared = ShmemInitStruct("CommitTs shared",
 									 sizeof(CommitTimestampShared),
@@ -927,14 +928,27 @@ AdvanceOldestCommitTsXid(TransactionId oldestXact)
 
 
 /*
- * Decide which of two commitTS page numbers is "older" for truncation
- * purposes.
+ * Decide whether a commitTS page number is "older" for truncation purposes.
+ * Analogous to CLOGPagePrecedes().
  *
- * We need to use comparison of TransactionIds here in order to do the right
- * thing with wraparound XID arithmetic.  However, if we are asked about
- * page number zero, we don't want to hand InvalidTransactionId to
- * TransactionIdPrecedes: it'll get weird about permanent xact IDs.  So,
- * offset both xids by FirstNormalTransactionId to avoid that.
+ * At default BLCKSZ, (1 << 31) % COMMIT_TS_XACTS_PER_PAGE == 128.  This
+ * introduces differences compared to CLOG and the other SLRUs having (1 <<
+ * 31) % per_page == 0.  This function never tests exactly
+ * TransactionIdPrecedes(x-2^31, x).  When the system reaches xidStopLimit,
+ * there are two possible counts of page boundaries between oldestXact and the
+ * latest XID assigned, depending on whether oldestXact is within the first
+ * 128 entries of its page.  Since this function doesn't know the location of
+ * oldestXact within page2, it returns false for one page that actually is
+ * expendable.  This is a wider (yet still negligible) version of the
+ * truncation opportunity that CLOGPagePrecedes() cannot recognize.
+ *
+ * For the sake of a worked example, number entries with decimal values such
+ * that page1==1 entries range from 1.0 to 1.999.  Let N+0.15 be the number of
+ * pages that 2^31 entries will span (N is an integer).  If oldestXact=N+2.1,
+ * then the final safe XID assignment leaves newestXact=1.95.  We keep page 2,
+ * because entry=2.85 is the border that toggles whether entries precede the
+ * last entry of the oldestXact page.  While page 2 is expendable at
+ * oldestXact=N+2.1, it would be precious at oldestXact=N+2.9.
  */
 static bool
 CommitTsPagePrecedes(int page1, int page2)
@@ -943,11 +957,12 @@ CommitTsPagePrecedes(int page1, int page2)
 	TransactionId xid2;
 
 	xid1 = ((TransactionId) page1) * COMMIT_TS_XACTS_PER_PAGE;
-	xid1 += FirstNormalTransactionId;
+	xid1 += FirstNormalTransactionId + 1;
 	xid2 = ((TransactionId) page2) * COMMIT_TS_XACTS_PER_PAGE;
-	xid2 += FirstNormalTransactionId;
+	xid2 += FirstNormalTransactionId + 1;
 
-	return TransactionIdPrecedes(xid1, xid2);
+	return (TransactionIdPrecedes(xid1, xid2) &&
+			TransactionIdPrecedes(xid1, xid2 + COMMIT_TS_XACTS_PER_PAGE - 1));
 }
 
 

src/backend/access/transam/multixact.c

@@ -1852,11 +1852,13 @@ MultiXactShmemInit(void)
 				  MultiXactOffsetSLRULock, "pg_multixact/offsets",
 				  LWTRANCHE_MULTIXACTOFFSET_BUFFER,
 				  SYNC_HANDLER_MULTIXACT_OFFSET);
+	SlruPagePrecedesUnitTests(MultiXactOffsetCtl, MULTIXACT_OFFSETS_PER_PAGE);
 	SimpleLruInit(MultiXactMemberCtl,
 				  "MultiXactMember", NUM_MULTIXACTMEMBER_BUFFERS, 0,
 				  MultiXactMemberSLRULock, "pg_multixact/members",
 				  LWTRANCHE_MULTIXACTMEMBER_BUFFER,
 				  SYNC_HANDLER_MULTIXACT_MEMBER);
+	/* doesn't call SimpleLruTruncate() or meet criteria for unit tests */
 
 	/* Initialize our shared state struct */
 	MultiXactState = ShmemInitStruct("Shared MultiXact State",
@@ -2982,6 +2984,14 @@ TruncateMultiXact(MultiXactId newOldestMulti, Oid newOldestMultiDB)
 	 * truncate the members SLRU.  So we first scan the directory to determine
 	 * the earliest offsets page number that we can read without error.
 	 *
+	 * When nextMXact is less than one segment away from multiWrapLimit,
+	 * SlruScanDirCbFindEarliest can find some early segment other than the
+	 * actual earliest.  (MultiXactOffsetPagePrecedes(EARLIEST, LATEST)
+	 * returns false, because not all pairs of entries have the same answer.)
+	 * That can also arise when an earlier truncation attempt failed unlink()
+	 * or returned early from this function.  The only consequence is
+	 * returning early, which wastes space that we could have liberated.
+	 *
 	 * NB: It's also possible that the page that oldestMulti is on has already
 	 * been truncated away, and we crashed before updating oldestMulti.
 	 */
@@ -3096,15 +3106,11 @@ TruncateMultiXact(MultiXactId newOldestMulti, Oid newOldestMultiDB)
 }
 
 /*
- * Decide which of two MultiXactOffset page numbers is "older" for truncation
- * purposes.
+ * Decide whether a MultiXactOffset page number is "older" for truncation
+ * purposes.  Analogous to CLOGPagePrecedes().
  *
- * We need to use comparison of MultiXactId here in order to do the right
- * thing with wraparound.  However, if we are asked about page number zero, we
- * don't want to hand InvalidMultiXactId to MultiXactIdPrecedes: it'll get
- * weird.  So, offset both multis by FirstMultiXactId to avoid that.
- * (Actually, the current implementation doesn't do anything weird with
- * InvalidMultiXactId, but there's no harm in leaving this code like this.)
+ * Offsetting the values is optional, because MultiXactIdPrecedes() has
+ * translational symmetry.
  */
 static bool
 MultiXactOffsetPagePrecedes(int page1, int page2)
@@ -3113,15 +3119,17 @@ MultiXactOffsetPagePrecedes(int page1, int page2)
 	MultiXactId multi2;
 
 	multi1 = ((MultiXactId) page1) * MULTIXACT_OFFSETS_PER_PAGE;
-	multi1 += FirstMultiXactId;
+	multi1 += FirstMultiXactId + 1;
 	multi2 = ((MultiXactId) page2) * MULTIXACT_OFFSETS_PER_PAGE;
-	multi2 += FirstMultiXactId;
+	multi2 += FirstMultiXactId + 1;
 
-	return MultiXactIdPrecedes(multi1, multi2);
+	return (MultiXactIdPrecedes(multi1, multi2) &&
+			MultiXactIdPrecedes(multi1,
+								multi2 + MULTIXACT_OFFSETS_PER_PAGE - 1));
 }
 
 /*
- * Decide which of two MultiXactMember page numbers is "older" for truncation
+ * Decide whether a MultiXactMember page number is "older" for truncation
  * purposes.  There is no "invalid offset number" so use the numbers verbatim.
  */
 static bool
@@ -3133,7 +3141,9 @@ MultiXactMemberPagePrecedes(int page1, int page2)
 	offset1 = ((MultiXactOffset) page1) * MULTIXACT_MEMBERS_PER_PAGE;
 	offset2 = ((MultiXactOffset) page2) * MULTIXACT_MEMBERS_PER_PAGE;
 
-	return MultiXactOffsetPrecedes(offset1, offset2);
+	return (MultiXactOffsetPrecedes(offset1, offset2) &&
+			MultiXactOffsetPrecedes(offset1,
+									offset2 + MULTIXACT_MEMBERS_PER_PAGE - 1));
 }
 
 /*

src/backend/access/transam/slru.c

@@ -1230,11 +1230,6 @@ SimpleLruTruncate(SlruCtl ctl, int cutoffPage)
 	/* update the stats counter of truncates */
 	pgstat_count_slru_truncate(shared->slru_stats_idx);
 
-	/*
-	 * The cutoff point is the start of the segment containing cutoffPage.
-	 */
-	cutoffPage -= cutoffPage % SLRU_PAGES_PER_SEGMENT;
-
 	/*
 	 * Scan shared memory and remove any pages preceding the cutoff page, to
 	 * ensure we won't rewrite them later.  (Since this is normally called in
@@ -1247,9 +1242,7 @@ restart:;
 
 	/*
 	 * While we are holding the lock, make an important safety check: the
-	 * planned cutoff point must be <= the current endpoint page. Otherwise we
-	 * have already wrapped around, and proceeding with the truncation would
-	 * risk removing the current segment.
+	 * current endpoint page must not be eligible for removal.
 	 */
 	if (ctl->PagePrecedes(shared->latest_page_number, cutoffPage))
 	{
@@ -1281,8 +1274,11 @@ restart:;
 		 * Hmm, we have (or may have) I/O operations acting on the page, so
 		 * we've got to wait for them to finish and then start again. This is
 		 * the same logic as in SlruSelectLRUPage.  (XXX if page is dirty,
-		 * wouldn't it be OK to just discard it without writing it?  For now,
-		 * keep the logic the same as it was.)
+		 * wouldn't it be OK to just discard it without writing it?
+		 * SlruMayDeleteSegment() uses a stricter qualification, so we might
+		 * not delete this page in the end; even if we don't delete it, we
+		 * won't have cause to read its data again.  For now, keep the logic
+		 * the same as it was.)
 		 */
 		if (shared->page_status[slotno] == SLRU_PAGE_VALID)
 			SlruInternalWritePage(ctl, slotno, NULL);
@@ -1377,19 +1373,134 @@ SlruDeleteSegment(SlruCtl ctl, int segno)
 	LWLockRelease(shared->ControlLock);
 }
 
+/*
+ * Determine whether a segment is okay to delete.
+ *
+ * segpage is the first page of the segment, and cutoffPage is the oldest (in
+ * PagePrecedes order) page in the SLRU containing still-useful data.  Since
+ * every core PagePrecedes callback implements "wrap around", check the
+ * segment's first and last pages:
+ *
+ * first<cutoff  && last<cutoff:  yes
+ * first<cutoff  && last>=cutoff: no; cutoff falls inside this segment
+ * first>=cutoff && last<cutoff:  no; wrap point falls inside this segment
+ * first>=cutoff && last>=cutoff: no; every page of this segment is too young
+ */
+static bool
+SlruMayDeleteSegment(SlruCtl ctl, int segpage, int cutoffPage)
+{
+	int			seg_last_page = segpage + SLRU_PAGES_PER_SEGMENT - 1;
+
+	Assert(segpage % SLRU_PAGES_PER_SEGMENT == 0);
+
+	return (ctl->PagePrecedes(segpage, cutoffPage) &&
+			ctl->PagePrecedes(seg_last_page, cutoffPage));
+}
+
+#ifdef USE_ASSERT_CHECKING
+static void
+SlruPagePrecedesTestOffset(SlruCtl ctl, int per_page, uint32 offset)
+{
+	TransactionId lhs,
+				rhs;
+	int			newestPage,
+				oldestPage;
+	TransactionId newestXact,
+				oldestXact;
+
+	/*
+	 * Compare an XID pair having undefined order (see RFC 1982), a pair at
+	 * "opposite ends" of the XID space.  TransactionIdPrecedes() treats each
+	 * as preceding the other.  If RHS is oldestXact, LHS is the first XID we
+	 * must not assign.
+	 */
+	lhs = per_page + offset;	/* skip first page to avoid non-normal XIDs */
+	rhs = lhs + (1U << 31);
+	Assert(TransactionIdPrecedes(lhs, rhs));
+	Assert(TransactionIdPrecedes(rhs, lhs));
+	Assert(!TransactionIdPrecedes(lhs - 1, rhs));
+	Assert(TransactionIdPrecedes(rhs, lhs - 1));
+	Assert(TransactionIdPrecedes(lhs + 1, rhs));
+	Assert(!TransactionIdPrecedes(rhs, lhs + 1));
+	Assert(!TransactionIdFollowsOrEquals(lhs, rhs));
+	Assert(!TransactionIdFollowsOrEquals(rhs, lhs));
+	Assert(!ctl->PagePrecedes(lhs / per_page, lhs / per_page));
+	Assert(!ctl->PagePrecedes(lhs / per_page, rhs / per_page));
+	Assert(!ctl->PagePrecedes(rhs / per_page, lhs / per_page));
+	Assert(!ctl->PagePrecedes((lhs - per_page) / per_page, rhs / per_page));
+	Assert(ctl->PagePrecedes(rhs / per_page, (lhs - 3 * per_page) / per_page));
+	Assert(ctl->PagePrecedes(rhs / per_page, (lhs - 2 * per_page) / per_page));
+	Assert(ctl->PagePrecedes(rhs / per_page, (lhs - 1 * per_page) / per_page)
+		   || (1U << 31) % per_page != 0);	/* See CommitTsPagePrecedes() */
+	Assert(ctl->PagePrecedes((lhs + 1 * per_page) / per_page, rhs / per_page)
+		   || (1U << 31) % per_page != 0);
+	Assert(ctl->PagePrecedes((lhs + 2 * per_page) / per_page, rhs / per_page));
+	Assert(ctl->PagePrecedes((lhs + 3 * per_page) / per_page, rhs / per_page));
+	Assert(!ctl->PagePrecedes(rhs / per_page, (lhs + per_page) / per_page));
+
+	/*
+	 * GetNewTransactionId() has assigned the last XID it can safely use, and
+	 * that XID is in the *LAST* page of the second segment.  We must not
+	 * delete that segment.
+	 */
+	newestPage = 2 * SLRU_PAGES_PER_SEGMENT - 1;
+	newestXact = newestPage * per_page + offset;
+	Assert(newestXact / per_page == newestPage);
+	oldestXact = newestXact + 1;
+	oldestXact -= 1U << 31;
+	oldestPage = oldestXact / per_page;
+	Assert(!SlruMayDeleteSegment(ctl,
+								 (newestPage -
+								  newestPage % SLRU_PAGES_PER_SEGMENT),
+								 oldestPage));
+
+	/*
+	 * GetNewTransactionId() has assigned the last XID it can safely use, and
+	 * that XID is in the *FIRST* page of the second segment.  We must not
+	 * delete that segment.
+	 */
+	newestPage = SLRU_PAGES_PER_SEGMENT;
+	newestXact = newestPage * per_page + offset;
+	Assert(newestXact / per_page == newestPage);
+	oldestXact = newestXact + 1;
+	oldestXact -= 1U << 31;
+	oldestPage = oldestXact / per_page;
+	Assert(!SlruMayDeleteSegment(ctl,
+								 (newestPage -
+								  newestPage % SLRU_PAGES_PER_SEGMENT),
+								 oldestPage));
+}
+
+/*
+ * Unit-test a PagePrecedes function.
+ *
+ * This assumes every uint32 >= FirstNormalTransactionId is a valid key.  It
+ * assumes each value occupies a contiguous, fixed-size region of SLRU bytes.
+ * (MultiXactMemberCtl separates flags from XIDs.  AsyncCtl has
+ * variable-length entries, no keys, and no random access.  These unit tests
+ * do not apply to them.)
+ */
+void
+SlruPagePrecedesUnitTests(SlruCtl ctl, int per_page)
+{
+	/* Test first, middle and last entries of a page. */
+	SlruPagePrecedesTestOffset(ctl, per_page, 0);
+	SlruPagePrecedesTestOffset(ctl, per_page, per_page / 2);
+	SlruPagePrecedesTestOffset(ctl, per_page, per_page - 1);
+}
+#endif
+
 /*
  * SlruScanDirectory callback
- *		This callback reports true if there's any segment prior to the one
- *		containing the page passed as "data".
+ *		This callback reports true if there's any segment wholly prior to the
+ *		one containing the page passed as "data".
  */
 bool
 SlruScanDirCbReportPresence(SlruCtl ctl, char *filename, int segpage, void *data)
 {
 	int			cutoffPage = *(int *) data;
 
-	cutoffPage -= cutoffPage % SLRU_PAGES_PER_SEGMENT;
-
-	if (ctl->PagePrecedes(segpage, cutoffPage))
+	if (SlruMayDeleteSegment(ctl, segpage, cutoffPage))
 		return true;			/* found one; don't iterate any more */
 
 	return false;				/* keep going */
@@ -1404,7 +1515,7 @@ SlruScanDirCbDeleteCutoff(SlruCtl ctl, char *filename, int segpage, void *data)
 {
 	int			cutoffPage = *(int *) data;
 
-	if (ctl->PagePrecedes(segpage, cutoffPage))
+	if (SlruMayDeleteSegment(ctl, segpage, cutoffPage))
 		SlruInternalDeleteSegment(ctl, segpage / SLRU_PAGES_PER_SEGMENT);
 
 	return false;				/* keep going */