make zil max block size tunable

We've observed that on some highly fragmented pools, most metaslab
allocations are small (~2-8KB), but there are some large, 128K
allocations.  The large allocations are for ZIL blocks.  If there is a
lot of fragmentation, the large allocations can be hard to satisfy.

The most common impact of this is that we need to check (and thus load)
lots of metaslabs from the ZIL allocation code path, causing sync writes
to wait for metaslabs to load, which can take a second or more.  In the
worst case, we may not be able to satisfy the allocation, in which case
the ZIL will resort to txg_wait_synced() to ensure the change is on
disk.

To provide a workaround for this, this change adds a tunable that can
reduce the size of ZIL blocks.

External-issue: DLPX-61719
Reviewed-by: George Wilson <george.wilson@delphix.com>
Reviewed-by: Paul Dagnelie <pcd@delphix.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Matthew Ahrens <mahrens@delphix.com>
Closes #8865

module/zfs/zfs_log.c

@@ -20,7 +20,7 @@
  */
 /*
  * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
- * Copyright (c) 2015 by Delphix. All rights reserved.
+ * Copyright (c) 2015, 2018 by Delphix. All rights reserved.
  */
 
 
@@ -541,7 +541,14 @@ zfs_log_write(zilog_t *zilog, dmu_tx_t *tx, int txtype,
 		itx_wr_state_t wr_state = write_state;
 		ssize_t len = resid;
 
-		if (wr_state == WR_COPIED && resid > ZIL_MAX_COPIED_DATA)
+		/*
+		 * A WR_COPIED record must fit entirely in one log block.
+		 * Large writes can use WR_NEED_COPY, which the ZIL will
+		 * split into multiple records across several log blocks
+		 * if necessary.
+		 */
+		if (wr_state == WR_COPIED &&
+		    resid > zil_max_copied_data(zilog))
 			wr_state = WR_NEED_COPY;
 		else if (wr_state == WR_INDIRECT)
 			len = MIN(blocksize - P2PHASE(off, blocksize), resid);

module/zfs/zil.c

@@ -1423,6 +1423,13 @@ uint64_t zil_block_buckets[] = {
     UINT64_MAX
 };
 
+/*
+ * Maximum block size used by the ZIL.  This is picked up when the ZIL is
+ * initialized.  Otherwise this should not be used directly; see
+ * zl_max_block_size instead.
+ */
+int zil_maxblocksize = SPA_OLD_MAXBLOCKSIZE;
+
 /*
  * Start a log block write and advance to the next log block.
  * Calls are serialized.
@@ -1499,9 +1506,7 @@ zil_lwb_write_issue(zilog_t *zilog, lwb_t *lwb)
 	zil_blksz = zilog->zl_cur_used + sizeof (zil_chain_t);
 	for (i = 0; zil_blksz > zil_block_buckets[i]; i++)
 		continue;
-	zil_blksz = zil_block_buckets[i];
-	if (zil_blksz == UINT64_MAX)
-		zil_blksz = SPA_OLD_MAXBLOCKSIZE;
+	zil_blksz = MIN(zil_block_buckets[i], zilog->zl_max_block_size);
 	zilog->zl_prev_blks[zilog->zl_prev_rotor] = zil_blksz;
 	for (i = 0; i < ZIL_PREV_BLKS; i++)
 		zil_blksz = MAX(zil_blksz, zilog->zl_prev_blks[i]);
@@ -1562,13 +1567,47 @@ zil_lwb_write_issue(zilog_t *zilog, lwb_t *lwb)
 	return (nlwb);
 }
 
+/*
+ * Maximum amount of write data that can be put into single log block.
+ */
+uint64_t
+zil_max_log_data(zilog_t *zilog)
+{
+	return (zilog->zl_max_block_size -
+	    sizeof (zil_chain_t) - sizeof (lr_write_t));
+}
+
+/*
+ * Maximum amount of log space we agree to waste to reduce number of
+ * WR_NEED_COPY chunks to reduce zl_get_data() overhead (~12%).
+ */
+static inline uint64_t
+zil_max_waste_space(zilog_t *zilog)
+{
+	return (zil_max_log_data(zilog) / 8);
+}
+
+/*
+ * Maximum amount of write data for WR_COPIED.  For correctness, consumers
+ * must fall back to WR_NEED_COPY if we can't fit the entire record into one
+ * maximum sized log block, because each WR_COPIED record must fit in a
+ * single log block.  For space efficiency, we want to fit two records into a
+ * max-sized log block.
+ */
+uint64_t
+zil_max_copied_data(zilog_t *zilog)
+{
+	return ((zilog->zl_max_block_size - sizeof (zil_chain_t)) / 2 -
+	    sizeof (lr_write_t));
+}
+
 static lwb_t *
 zil_lwb_commit(zilog_t *zilog, itx_t *itx, lwb_t *lwb)
 {
 	lr_t *lrcb, *lrc;
 	lr_write_t *lrwb, *lrw;
 	char *lr_buf;
-	uint64_t dlen, dnow, lwb_sp, reclen, txg;
+	uint64_t dlen, dnow, lwb_sp, reclen, txg, max_log_data;
 
 	ASSERT(MUTEX_HELD(&zilog->zl_issuer_lock));
 	ASSERT3P(lwb, !=, NULL);
@@ -1617,15 +1656,27 @@ cont:
 	 * For WR_NEED_COPY optimize layout for minimal number of chunks.
 	 */
 	lwb_sp = lwb->lwb_sz - lwb->lwb_nused;
+	max_log_data = zil_max_log_data(zilog);
 	if (reclen > lwb_sp || (reclen + dlen > lwb_sp &&
-	    lwb_sp < ZIL_MAX_WASTE_SPACE && (dlen % ZIL_MAX_LOG_DATA == 0 ||
-	    lwb_sp < reclen + dlen % ZIL_MAX_LOG_DATA))) {
+	    lwb_sp < zil_max_waste_space(zilog) &&
+	    (dlen % max_log_data == 0 ||
+	    lwb_sp < reclen + dlen % max_log_data))) {
 		lwb = zil_lwb_write_issue(zilog, lwb);
 		if (lwb == NULL)
 			return (NULL);
 		zil_lwb_write_open(zilog, lwb);
 		ASSERT(LWB_EMPTY(lwb));
 		lwb_sp = lwb->lwb_sz - lwb->lwb_nused;
+
+		/*
+		 * There must be enough space in the new, empty log block to
+		 * hold reclen.  For WR_COPIED, we need to fit the whole
+		 * record in one block, and reclen is the header size + the
+		 * data size. For WR_NEED_COPY, we can create multiple
+		 * records, splitting the data into multiple blocks, so we
+		 * only need to fit one word of data per block; in this case
+		 * reclen is just the header size (no data).
+		 */
 		ASSERT3U(reclen + MIN(dlen, sizeof (uint64_t)), <=, lwb_sp);
 	}
 
@@ -3114,6 +3165,7 @@ zil_alloc(objset_t *os, zil_header_t *zh_phys)
 	zilog->zl_dirty_max_txg = 0;
 	zilog->zl_last_lwb_opened = NULL;
 	zilog->zl_last_lwb_latency = 0;
+	zilog->zl_max_block_size = zil_maxblocksize;
 
 	mutex_init(&zilog->zl_lock, NULL, MUTEX_DEFAULT, NULL);
 	mutex_init(&zilog->zl_issuer_lock, NULL, MUTEX_DEFAULT, NULL);
@@ -3627,5 +3679,8 @@ MODULE_PARM_DESC(zil_nocacheflush, "Disable ZIL cache flushes");
 
 module_param(zil_slog_bulk, ulong, 0644);
 MODULE_PARM_DESC(zil_slog_bulk, "Limit in bytes slog sync writes per commit");
+
+module_param(zil_maxblocksize, int, 0644);
+MODULE_PARM_DESC(zil_maxblocksize, "Limit in bytes of ZIL log block size");
 /* END CSTYLED */
 #endif

module/zfs/zvol.c

@@ -684,7 +684,7 @@ zvol_log_write(zvol_state_t *zv, dmu_tx_t *tx, uint64_t offset,
 		itx_wr_state_t wr_state = write_state;
 		ssize_t len = size;
 
-		if (wr_state == WR_COPIED && size > ZIL_MAX_COPIED_DATA)
+		if (wr_state == WR_COPIED && size > zil_max_copied_data(zilog))
 			wr_state = WR_NEED_COPY;
 		else if (wr_state == WR_INDIRECT)
 			len = MIN(blocksize - P2PHASE(offset, blocksize), size);