Merge branch 'PGPROEE9_6_CFS_385' into PGPROEE9_6

Author: vbwagner

src/backend/storage/file/cfs.c

@@ -987,13 +987,16 @@ static bool cfs_gc_file(char* map_path, GC_CALL_KIND background)
 		cfs_state->gc_stat.processedFiles += 1;
 		cfs_gc_processed_segments += 1;
 
+retry:
 		/* temporary lock file for fetching map snapshot */
 		cfs_gc_lock(lock);
 
 		/* Reread variables after locking file */
+		physSize = pg_atomic_read_u32(&map->hdr.physSize);
 		virtSize = pg_atomic_read_u32(&map->hdr.virtSize);
 		n_pages = virtSize / BLCKSZ;
-retry:
+		if (physSize >= CFS_RED_LINE)
+			goto forceWhole;
 		for (i = 0; i < n_pages; i++)
 		{
 			newMap->inodes[i] = map->inodes[i];
@@ -1026,10 +1029,9 @@ static bool cfs_gc_file(char* map_path, GC_CALL_KIND background)
 
 		/* Reread variables after locking file */
 		n_pages1 = n_pages;
+		physSize = pg_atomic_read_u32(&map->hdr.physSize);
 		virtSize = pg_atomic_read_u32(&map->hdr.virtSize);
 		n_pages = virtSize / BLCKSZ;
-		second_pass = 0;
-		second_pass_bytes = 0;
 
 		for (i = 0; i < n_pages; i++)
 		{
@@ -1074,10 +1076,16 @@ static bool cfs_gc_file(char* map_path, GC_CALL_KIND background)
 			memset(newMap->inodes, 0, sizeof(newMap->inodes));
 			elog(LOG, "CFS: retry %d whole gc file %s", second_pass_whole,
 					file_path);
-			if (second_pass_whole == 1)
+			if (second_pass_whole == 1 && physSize < CFS_RETRY_GC_THRESHOLD)
 			{
+				cfs_gc_unlock(lock);
+				/* sleep, cause there is possibly checkpoint is on a way */
+				pg_usleep(CFS_LOCK_MAX_TIMEOUT);
+				second_pass = 0;
+				second_pass_bytes = 0;
 				goto retry;
 			}
+		forceWhole:
 			for (i = 0; i < n_pages; i++)
 			{
 				newMap->inodes[i] = map->inodes[i];
@@ -1275,7 +1283,7 @@ static bool cfs_gc_file(char* map_path, GC_CALL_KIND background)
 		{
 			elog(LOG, "CFS GC worker %d: defragment file %s: old size %u, new size %u, logical size %u, used %u, compression ratio %f, time %ld usec; second pass: pages %u, bytes %u, time %ld"
 					,
-				 MyProcPid, file_path, physSize, newSize, virtSize, usedSize, (double)virtSize/newSize,
+				 MyProcPid, file_path, physSize, newSize, virtSize, newUsed, (double)virtSize/newSize,
 				 secs*USECS_PER_SEC + usecs, second_pass, second_pass_bytes,
 				 secs2*USECS_PER_SEC + usecs2);
 		}

src/backend/storage/smgr/md.c

@@ -451,6 +451,7 @@ mdcreate(SMgrRelation reln, ForkNumber forkNum, bool isRedo)
 void
 mdunlink(RelFileNodeBackend rnode, ForkNumber forkNum, bool isRedo)
 {
+	bool cfs_gc_locked = false;
 	/*
 	 * We have to clean out any pending fsync requests for the doomed
 	 * relation, else the next mdsync() will fail.  There can't be any such
@@ -461,14 +462,33 @@ mdunlink(RelFileNodeBackend rnode, ForkNumber forkNum, bool isRedo)
 	if (!RelFileNodeBackendIsTemp(rnode))
 		ForgetRelationFsyncRequests(rnode.node, forkNum);
 
-	/* Now do the per-fork work */
-	if (forkNum == InvalidForkNumber)
+	if (md_use_compression(rnode, forkNum == InvalidForkNumber ? MAIN_FORKNUM : forkNum))
 	{
-		for (forkNum = 0; forkNum <= MAX_FORKNUM; forkNum++)
+		cfs_gc_locked = true;
+		cfs_control_gc_lock();
+	}
+
+	PG_TRY();
+	{
+		/* Now do the per-fork work */
+		if (forkNum == InvalidForkNumber)
+		{
+			for (forkNum = 0; forkNum <= MAX_FORKNUM; forkNum++)
+				mdunlinkfork(rnode, forkNum, isRedo);
+		}
+		else
 			mdunlinkfork(rnode, forkNum, isRedo);
+		}
 	}
-	else
-		mdunlinkfork(rnode, forkNum, isRedo);
+	PG_CATCH();
+	{
+		if (cfs_gc_locked)
+			cfs_control_gc_unlock();
+		PG_RE_THROW();
+	}
+	PG_END_TRY();
+	if (cfs_gc_locked)
+		cfs_control_gc_unlock();
 }
 
 static void

src/include/storage/cfs.h

@@ -53,6 +53,7 @@ typedef uint64 inode_t;
 #define CFS_INODE_CLEAN_FLAG  ((inode_t)1 << 63)
 
 #define CFS_IMPLICIT_GC_THRESHOLD 0x80000000U /* 2Gb */
+#define CFS_RETRY_GC_THRESHOLD    0x60000000U /* 1.5Gb */
 #define CFS_RED_LINE              0xC0000000U /* 3Gb */
 
 size_t cfs_compress(void* dst, size_t dst_size, void const* src, size_t src_size);