xfs: log recovery torn write detection

index c5ecaac..4f880d6 100644
@@ -4245,7 +4245,7 @@ xlog_do_recovery_pass(
 	xfs_daddr_t		blk_no;
 	char			*offset;
 	xfs_buf_t		*hbp, *dbp;
-	int			error = 0, h_size;
+	int			error = 0, h_size, h_len;
 	int			bblks, split_bblks;
 	int			hblks, split_hblks, wrapped_hblks;
 	struct hlist_head	rhash[XLOG_RHASH_SIZE];

@@ -4274,7 +4274,31 @@ xlog_do_recovery_pass(
 		error = xlog_valid_rec_header(log, rhead, tail_blk);
 		if (error)
 			goto bread_err1;
+
+		/*
+		 * xfsprogs has a bug where record length is based on lsunit but
+		 * h_size (iclog size) is hardcoded to 32k. Now that we
+		 * unconditionally CRC verify the unmount record, this means the
+		 * log buffer can be too small for the record and cause an
+		 * overrun.
+		 *
+		 * Detect this condition here. Use lsunit for the buffer size as
+		 * long as this looks like the mkfs case. Otherwise, return an
+		 * error to avoid a buffer overrun.
+		 */
 		h_size = be32_to_cpu(rhead->h_size);
+		h_len = be32_to_cpu(rhead->h_len);
+		if (h_len > h_size) {
+			if (h_len <= log->l_mp->m_logbsize &&
+			    be32_to_cpu(rhead->h_num_logops) == 1) {
+				xfs_warn(log->l_mp,
+		"invalid iclog size (%d bytes), using lsunit (%d bytes)",
+					 h_size, log->l_mp->m_logbsize);
+				h_size = log->l_mp->m_logbsize;
+			} else
+				return -EFSCORRUPTED;
+		}
+
 		if ((be32_to_cpu(rhead->h_version) & XLOG_VERSION_2) &&
 		    (h_size > XLOG_HEADER_CYCLE_SIZE)) {
 			hblks = h_size / XLOG_HEADER_CYCLE_SIZE;

index 4f880d6..236ebaf 100644
@@ -4190,6 +4190,26 @@ xlog_unpack_data(
 	return 0;
 }
 
+/*
+ * Unpack and process a log record.
+ */
+STATIC int
+xlog_recover_process(
+	struct xlog		*log,
+	struct hlist_head	rhash[],
+	struct xlog_rec_header	*rhead,
+	char			*dp,
+	int			pass)
+{
+	int			error;
+
+	error = xlog_unpack_data(rhead, dp, log);
+	if (error)
+		return error;
+
+	return xlog_recover_process_data(log, rhash, rhead, dp, pass);
+}
+
 STATIC int
 xlog_valid_rec_header(
 	struct xlog		*log,

@@ -4432,12 +4452,8 @@ xlog_do_recovery_pass(
 					goto bread_err2;
 			}
 
-			error = xlog_unpack_data(rhead, offset, log);
-			if (error)
-				goto bread_err2;
-
-			error = xlog_recover_process_data(log, rhash,
-							rhead, offset, pass);
+			error = xlog_recover_process(log, rhash, rhead, offset,
+						     pass);
 			if (error)
 				goto bread_err2;
 			blk_no += bblks;

@@ -4465,12 +4481,7 @@ xlog_do_recovery_pass(
 		if (error)
 			goto bread_err2;
 
-		error = xlog_unpack_data(rhead, offset, log);
-		if (error)
-			goto bread_err2;
-
-		error = xlog_recover_process_data(log, rhash,
-						rhead, offset, pass);
+		error = xlog_recover_process(log, rhash, rhead, offset, pass);
 		if (error)
 			goto bread_err2;
 		blk_no += bblks + hblks;

index 236ebaf..9ec4bbd 100644
@@ -4118,46 +4118,6 @@ xlog_recover_process_iunlinks(
 	mp->m_dmevmask = mp_dmevmask;
 }
 
-/*
- * Upack the log buffer data and crc check it. If the check fails, issue a
- * warning if and only if the CRC in the header is non-zero. This makes the
- * check an advisory warning, and the zero CRC check will prevent failure
- * warnings from being emitted when upgrading the kernel from one that does not
- * add CRCs by default.
- *
- * When filesystems are CRC enabled, this CRC mismatch becomes a fatal log
- * corruption failure
- */
-STATIC int
-xlog_unpack_data_crc(
-	struct xlog_rec_header	*rhead,
-	char			*dp,
-	struct xlog		*log)
-{
-	__le32			crc;
-
-	crc = xlog_cksum(log, rhead, dp, be32_to_cpu(rhead->h_len));
-	if (crc != rhead->h_crc) {
-		if (rhead->h_crc || xfs_sb_version_hascrc(&log->l_mp->m_sb)) {
-			xfs_alert(log->l_mp,
-		"log record CRC mismatch: found 0x%x, expected 0x%x.",
-					le32_to_cpu(rhead->h_crc),
-					le32_to_cpu(crc));
-			xfs_hex_dump(dp, 32);
-		}
-
-		/*
-		 * If we've detected a log record corruption, then we can't
-		 * recover past this point. Abort recovery if we are enforcing
-		 * CRC protection by punting an error back up the stack.
-		 */
-		if (xfs_sb_version_hascrc(&log->l_mp->m_sb))
-			return -EFSCORRUPTED;
-	}
-
-	return 0;
-}
-
 STATIC int
 xlog_unpack_data(
 	struct xlog_rec_header	*rhead,

@@ -4165,11 +4125,6 @@ xlog_unpack_data(
 	struct xlog		*log)
 {
 	int			i, j, k;
-	int			error;
-
-	error = xlog_unpack_data_crc(rhead, dp, log);
-	if (error)
-		return error;
 
 	for (i = 0; i < BTOBB(be32_to_cpu(rhead->h_len)) &&
 		  i < (XLOG_HEADER_CYCLE_SIZE / BBSIZE); i++) {

@@ -4191,7 +4146,7 @@ xlog_unpack_data(
 }
 
 /*
- * Unpack and process a log record.
+ * CRC check, unpack and process a log record.
  */
 STATIC int
 xlog_recover_process(

@@ -4202,6 +4157,31 @@ xlog_recover_process(
 	int			pass)
 {
 	int			error;
+	__le32			crc;
+
+	/*
+	 * Check the CRC and issue a warning if and only if the CRC in the
+	 * header is non-zero. This is an advisory warning and the zero CRC
+	 * check prevents warnings from being emitted when upgrading the kernel
+	 * from one that does not add CRCs by default.
+	 */
+	crc = xlog_cksum(log, rhead, dp, be32_to_cpu(rhead->h_len));
+	if (crc != le32_to_cpu(rhead->h_crc)) {
+		if (rhead->h_crc || xfs_sb_version_hascrc(&log->l_mp->m_sb)) {
+			xfs_alert(log->l_mp,
+		"log record CRC mismatch: found 0x%x, expected 0x%x.",
+					le32_to_cpu(rhead->h_crc),
+					le32_to_cpu(crc));
+			xfs_hex_dump(dp, 32);
+		}
+
+		/*
+		 * If the filesystem is CRC enabled, this mismatch becomes a
+		 * fatal log corruption failure.
+		 */
+		if (xfs_sb_version_hascrc(&log->l_mp->m_sb))
+			return -EFSCORRUPTED;
+	}
 
 	error = xlog_unpack_data(rhead, dp, log);
 	if (error)

index 9ec4bbd..e0318e8 100644
@@ -4239,10 +4239,12 @@ xlog_do_recovery_pass(
 	struct xlog		*log,
 	xfs_daddr_t		head_blk,
 	xfs_daddr_t		tail_blk,
-	int			pass)
+	int			pass,
+	xfs_daddr_t		*first_bad)	/* out: first bad log rec */
 {
 	xlog_rec_header_t	*rhead;
 	xfs_daddr_t		blk_no;
+	xfs_daddr_t		rhead_blk;
 	char			*offset;
 	xfs_buf_t		*hbp, *dbp;
 	int			error = 0, h_size, h_len;

@@ -4251,6 +4253,7 @@ xlog_do_recovery_pass(
 	struct hlist_head	rhash[XLOG_RHASH_SIZE];
 
 	ASSERT(head_blk != tail_blk);
+	rhead_blk = 0;
 
 	/*
 	 * Read the header of the tail block and get the iclog buffer size from

@@ -4325,7 +4328,7 @@ xlog_do_recovery_pass(
 	}
 
 	memset(rhash, 0, sizeof(rhash));
-	blk_no = tail_blk;
+	blk_no = rhead_blk = tail_blk;
 	if (tail_blk > head_blk) {
 		/*
 		 * Perform recovery around the end of the physical log.

@@ -4436,11 +4439,14 @@ xlog_do_recovery_pass(
 						     pass);
 			if (error)
 				goto bread_err2;
+
 			blk_no += bblks;
+			rhead_blk = blk_no;
 		}
 
 		ASSERT(blk_no >= log->l_logBBsize);
 		blk_no -= log->l_logBBsize;
+		rhead_blk = blk_no;
 	}
 
 	/* read first part of physical log */

@@ -4464,13 +4470,19 @@ xlog_do_recovery_pass(
 		error = xlog_recover_process(log, rhash, rhead, offset, pass);
 		if (error)
 			goto bread_err2;
+
 		blk_no += bblks + hblks;
+		rhead_blk = blk_no;
 	}
 
  bread_err2:
 	xlog_put_bp(dbp);
  bread_err1:
 	xlog_put_bp(hbp);
+
+	if (error && first_bad)
+		*first_bad = rhead_blk;
+
 	return error;
 }
 

@@ -4508,7 +4520,7 @@ xlog_do_log_recovery(
 		INIT_LIST_HEAD(&log->l_buf_cancel_table[i]);
 
 	error = xlog_do_recovery_pass(log, head_blk, tail_blk,
-				      XLOG_RECOVER_PASS1);
+				      XLOG_RECOVER_PASS1, NULL);
 	if (error != 0) {
 		kmem_free(log->l_buf_cancel_table);
 		log->l_buf_cancel_table = NULL;

@@ -4519,7 +4531,7 @@ xlog_do_log_recovery(
 	 * When it is complete free the table of buf cancel items.
 	 */
 	error = xlog_do_recovery_pass(log, head_blk, tail_blk,
-				      XLOG_RECOVER_PASS2);
+				      XLOG_RECOVER_PASS2, NULL);
 #ifdef DEBUG
 	if (!error) {
 		int	i;

index 1c55ccb..8e385f9 100644
@@ -60,6 +60,7 @@ typedef struct xlog_recover {
  */
 #define	XLOG_BC_TABLE_SIZE	64
 
+#define	XLOG_RECOVER_CRCPASS	0
 #define	XLOG_RECOVER_PASS1	1
 #define	XLOG_RECOVER_PASS2	2
 

index e0318e8..1be2590 100644
@@ -4159,13 +4159,27 @@ xlog_recover_process(
 	int			error;
 	__le32			crc;
 
+	crc = xlog_cksum(log, rhead, dp, be32_to_cpu(rhead->h_len));
+
 	/*
-	 * Check the CRC and issue a warning if and only if the CRC in the
-	 * header is non-zero. This is an advisory warning and the zero CRC
-	 * check prevents warnings from being emitted when upgrading the kernel
-	 * from one that does not add CRCs by default.
+	 * Nothing else to do if this is a CRC verification pass. Just return
+	 * if this a record with a non-zero crc. Unfortunately, mkfs always
+	 * sets h_crc to 0 so we must consider this valid even on v5 supers.
+	 * Otherwise, return EFSBADCRC on failure so the callers up the stack
+	 * know precisely what failed.
+	 */
+	if (pass == XLOG_RECOVER_CRCPASS) {
+		if (rhead->h_crc && crc != le32_to_cpu(rhead->h_crc))
+			return -EFSBADCRC;
+		return 0;
+	}
+
+	/*
+	 * We're in the normal recovery path. Issue a warning if and only if the
+	 * CRC in the header is non-zero. This is an advisory warning and the
+	 * zero CRC check prevents warnings from being emitted when upgrading
+	 * the kernel from one that does not add CRCs by default.
 	 */
-	crc = xlog_cksum(log, rhead, dp, be32_to_cpu(rhead->h_len));
 	if (crc != le32_to_cpu(rhead->h_crc)) {
 		if (rhead->h_crc || xfs_sb_version_hascrc(&log->l_mp->m_sb)) {
 			xfs_alert(log->l_mp,

index 1be2590..423c36d 100644
@@ -868,6 +868,79 @@ validate_head:
 }
 
 /*
+ * Seek backwards in the log for log record headers.
+ *
+ * Given a starting log block, walk backwards until we find the provided number
+ * of records or hit the provided tail block. The return value is the number of
+ * records encountered or a negative error code. The log block and buffer
+ * pointer of the last record seen are returned in rblk and rhead respectively.
+ */
+STATIC int
+xlog_rseek_logrec_hdr(
+	struct xlog		*log,
+	xfs_daddr_t		head_blk,
+	xfs_daddr_t		tail_blk,
+	int			count,
+	struct xfs_buf		*bp,
+	xfs_daddr_t		*rblk,
+	struct xlog_rec_header	**rhead,
+	bool			*wrapped)
+{
+	int			i;
+	int			error;
+	int			found = 0;
+	char			*offset = NULL;
+	xfs_daddr_t		end_blk;
+
+	*wrapped = false;
+
+	/*
+	 * Walk backwards from the head block until we hit the tail or the first
+	 * block in the log.
+	 */
+	end_blk = head_blk > tail_blk ? tail_blk : 0;
+	for (i = (int) head_blk - 1; i >= end_blk; i--) {
+		error = xlog_bread(log, i, 1, bp, &offset);
+		if (error)
+			goto out_error;
+
+		if (*(__be32 *) offset == cpu_to_be32(XLOG_HEADER_MAGIC_NUM)) {
+			*rblk = i;
+			*rhead = (struct xlog_rec_header *) offset;
+			if (++found == count)
+				break;
+		}
+	}
+
+	/*
+	 * If we haven't hit the tail block or the log record header count,
+	 * start looking again from the end of the physical log. Note that
+	 * callers can pass head == tail if the tail is not yet known.
+	 */
+	if (tail_blk >= head_blk && found != count) {
+		for (i = log->l_logBBsize - 1; i >= (int) tail_blk; i--) {
+			error = xlog_bread(log, i, 1, bp, &offset);
+			if (error)
+				goto out_error;
+
+			if (*(__be32 *)offset ==
+			    cpu_to_be32(XLOG_HEADER_MAGIC_NUM)) {
+				*wrapped = true;
+				*rblk = i;
+				*rhead = (struct xlog_rec_header *) offset;
+				if (++found == count)
+					break;
+			}
+		}
+	}
+
+	return found;
+
+out_error:
+	return error;
+}
+
+/*
  * Find the sync block number or the tail of the log.
  *
  * This will be the block number of the last record to have its

@@ -898,8 +971,7 @@ xlog_find_tail(
 	xfs_daddr_t		after_umount_blk;
 	xfs_lsn_t		tail_lsn;
 	int			hblks;
-
-	found = 0;
+	bool			wrapped = false;
 
 	/*
 	 * Find previous log record

@@ -923,37 +995,16 @@ xlog_find_tail(
 	}
 
 	/*
-	 * Search backwards looking for log record header block
+	 * Search backwards through the log looking for the log record header
+	 * block. This wraps all the way back around to the head so something is
+	 * seriously wrong if we can't find it.
 	 */
 	ASSERT(*head_blk < INT_MAX);
-	for (i = (int)(*head_blk) - 1; i >= 0; i--) {
-		error = xlog_bread(log, i, 1, bp, &offset);
-		if (error)
-			goto done;
-
-		if (*(__be32 *)offset == cpu_to_be32(XLOG_HEADER_MAGIC_NUM)) {
-			found = 1;
-			break;
-		}
-	}
-	/*
-	 * If we haven't found the log record header block, start looking
-	 * again from the end of the physical log.  XXXmiken: There should be
-	 * a check here to make sure we didn't search more than N blocks in
-	 * the previous code.
-	 */
-	if (!found) {
-		for (i = log->l_logBBsize - 1; i >= (int)(*head_blk); i--) {
-			error = xlog_bread(log, i, 1, bp, &offset);
-			if (error)
-				goto done;
-
-			if (*(__be32 *)offset ==
-			    cpu_to_be32(XLOG_HEADER_MAGIC_NUM)) {
-				found = 2;
-				break;
-			}
-		}
+	found = xlog_rseek_logrec_hdr(log, *head_blk, *head_blk, 1, bp, &i,
+				      &rhead, &wrapped);
+	if (found < 0) {
+		error = found;
+		goto done;
 	}
 	if (!found) {
 		xfs_warn(log->l_mp, "%s: couldn't find sync record", __func__);

@@ -961,9 +1012,6 @@ xlog_find_tail(
 		ASSERT(0);
 		return -EIO;
 	}
-
-	/* find blk_no of tail of log */
-	rhead = (xlog_rec_header_t *)offset;
 	*tail_blk = BLOCK_LSN(be64_to_cpu(rhead->h_tail_lsn));
 
 	/*

@@ -979,7 +1027,7 @@ xlog_find_tail(
 	log->l_prev_block = i;
 	log->l_curr_block = (int)*head_blk;
 	log->l_curr_cycle = be32_to_cpu(rhead->h_cycle);
-	if (found == 2)
+	if (wrapped)
 		log->l_curr_cycle++;
 	atomic64_set(&log->l_tail_lsn, be64_to_cpu(rhead->h_tail_lsn));
 	atomic64_set(&log->l_last_sync_lsn, be64_to_cpu(rhead->h_lsn));

index 423c36d..5bf8076 100644
@@ -61,6 +61,9 @@ xlog_recover_check_summary(
 #else
 #define	xlog_recover_check_summary(log)
 #endif
+STATIC int
+xlog_do_recovery_pass(
+        struct xlog *, xfs_daddr_t, xfs_daddr_t, int, xfs_daddr_t *);
 
 /*
  * This structure is used during recovery to record the buf log items which

@@ -941,6 +944,278 @@ out_error:
 }
 
 /*
+ * Seek forward in the log for log record headers.
+ *
+ * Given head and tail blocks, walk forward from the tail block until we find
+ * the provided number of records or hit the head block. The return value is the
+ * number of records encountered or a negative error code. The log block and
+ * buffer pointer of the last record seen are returned in rblk and rhead
+ * respectively.
+ */
+STATIC int
+xlog_seek_logrec_hdr(
+	struct xlog		*log,
+	xfs_daddr_t		head_blk,
+	xfs_daddr_t		tail_blk,
+	int			count,
+	struct xfs_buf		*bp,
+	xfs_daddr_t		*rblk,
+	struct xlog_rec_header	**rhead,
+	bool			*wrapped)
+{
+	int			i;
+	int			error;
+	int			found = 0;
+	char			*offset = NULL;
+	xfs_daddr_t		end_blk;
+
+	*wrapped = false;
+
+	/*
+	 * Walk forward from the tail block until we hit the head or the last
+	 * block in the log.
+	 */
+	end_blk = head_blk > tail_blk ? head_blk : log->l_logBBsize - 1;
+	for (i = (int) tail_blk; i <= end_blk; i++) {
+		error = xlog_bread(log, i, 1, bp, &offset);
+		if (error)
+			goto out_error;
+
+		if (*(__be32 *) offset == cpu_to_be32(XLOG_HEADER_MAGIC_NUM)) {
+			*rblk = i;
+			*rhead = (struct xlog_rec_header *) offset;
+			if (++found == count)
+				break;
+		}
+	}
+
+	/*
+	 * If we haven't hit the head block or the log record header count,
+	 * start looking again from the start of the physical log.
+	 */
+	if (tail_blk > head_blk && found != count) {
+		for (i = 0; i < (int) head_blk; i++) {
+			error = xlog_bread(log, i, 1, bp, &offset);
+			if (error)
+				goto out_error;
+
+			if (*(__be32 *)offset ==
+			    cpu_to_be32(XLOG_HEADER_MAGIC_NUM)) {
+				*wrapped = true;
+				*rblk = i;
+				*rhead = (struct xlog_rec_header *) offset;
+				if (++found == count)
+					break;
+			}
+		}
+	}
+
+	return found;
+
+out_error:
+	return error;
+}
+
+/*
+ * Check the log tail for torn writes. This is required when torn writes are
+ * detected at the head and the head had to be walked back to a previous record.
+ * The tail of the previous record must now be verified to ensure the torn
+ * writes didn't corrupt the previous tail.
+ *
+ * Return an error if CRC verification fails as recovery cannot proceed.
+ */
+STATIC int
+xlog_verify_tail(
+	struct xlog		*log,
+	xfs_daddr_t		head_blk,
+	xfs_daddr_t		tail_blk)
+{
+	struct xlog_rec_header	*thead;
+	struct xfs_buf		*bp;
+	xfs_daddr_t		first_bad;
+	int			count;
+	int			error = 0;
+	bool			wrapped;
+	xfs_daddr_t		tmp_head;
+
+	bp = xlog_get_bp(log, 1);
+	if (!bp)
+		return -ENOMEM;
+
+	/*
+	 * Seek XLOG_MAX_ICLOGS + 1 records past the current tail record to get
+	 * a temporary head block that points after the last possible
+	 * concurrently written record of the tail.
+	 */
+	count = xlog_seek_logrec_hdr(log, head_blk, tail_blk,
+				     XLOG_MAX_ICLOGS + 1, bp, &tmp_head, &thead,
+				     &wrapped);
+	if (count < 0) {
+		error = count;
+		goto out;
+	}
+
+	/*
+	 * If the call above didn't find XLOG_MAX_ICLOGS + 1 records, we ran
+	 * into the actual log head. tmp_head points to the start of the record
+	 * so update it to the actual head block.
+	 */
+	if (count < XLOG_MAX_ICLOGS + 1)
+		tmp_head = head_blk;
+
+	/*
+	 * We now have a tail and temporary head block that covers at least
+	 * XLOG_MAX_ICLOGS records from the tail. We need to verify that these
+	 * records were completely written. Run a CRC verification pass from
+	 * tail to head and return the result.
+	 */
+	error = xlog_do_recovery_pass(log, tmp_head, tail_blk,
+				      XLOG_RECOVER_CRCPASS, &first_bad);
+
+out:
+	xlog_put_bp(bp);
+	return error;
+}
+
+/*
+ * Detect and trim torn writes from the head of the log.
+ *
+ * Storage without sector atomicity guarantees can result in torn writes in the
+ * log in the event of a crash. Our only means to detect this scenario is via
+ * CRC verification. While we can't always be certain that CRC verification
+ * failure is due to a torn write vs. an unrelated corruption, we do know that
+ * only a certain number (XLOG_MAX_ICLOGS) of log records can be written out at
+ * one time. Therefore, CRC verify up to XLOG_MAX_ICLOGS records at the head of
+ * the log and treat failures in this range as torn writes as a matter of
+ * policy. In the event of CRC failure, the head is walked back to the last good
+ * record in the log and the tail is updated from that record and verified.
+ */
+STATIC int
+xlog_verify_head(
+	struct xlog		*log,
+	xfs_daddr_t		*head_blk,	/* in/out: unverified head */
+	xfs_daddr_t		*tail_blk,	/* out: tail block */
+	struct xfs_buf		*bp,
+	xfs_daddr_t		*rhead_blk,	/* start blk of last record */
+	struct xlog_rec_header	**rhead,	/* ptr to last record */
+	bool			*wrapped)	/* last rec. wraps phys. log */
+{
+	struct xlog_rec_header	*tmp_rhead;
+	struct xfs_buf		*tmp_bp;
+	xfs_daddr_t		first_bad;
+	xfs_daddr_t		tmp_rhead_blk;
+	int			found;
+	int			error;
+	bool			tmp_wrapped;
+
+	/*
+	 * Search backwards through the log looking for the log record header
+	 * block. This wraps all the way back around to the head so something is
+	 * seriously wrong if we can't find it.
+	 */
+	found = xlog_rseek_logrec_hdr(log, *head_blk, *head_blk, 1, bp, rhead_blk,
+				      rhead, wrapped);
+	if (found < 0)
+		return found;
+	if (!found) {
+		xfs_warn(log->l_mp, "%s: couldn't find sync record", __func__);
+		return -EIO;
+	}
+
+	*tail_blk = BLOCK_LSN(be64_to_cpu((*rhead)->h_tail_lsn));
+
+	/*
+	 * Now that we have a tail block, check the head of the log for torn
+	 * writes. Search again until we hit the tail or the maximum number of
+	 * log record I/Os that could have been in flight at one time. Use a
+	 * temporary buffer so we don't trash the rhead/bp pointer from the
+	 * call above.
+	 */
+	tmp_bp = xlog_get_bp(log, 1);
+	if (!tmp_bp)
+		return -ENOMEM;
+	error = xlog_rseek_logrec_hdr(log, *head_blk, *tail_blk,
+				      XLOG_MAX_ICLOGS, tmp_bp, &tmp_rhead_blk,
+				      &tmp_rhead, &tmp_wrapped);
+	xlog_put_bp(tmp_bp);
+	if (error < 0)
+		return error;
+
+	/*
+	 * Now run a CRC verification pass over the records starting at the
+	 * block found above to the current head. If a CRC failure occurs, the
+	 * log block of the first bad record is saved in first_bad.
+	 */
+	error = xlog_do_recovery_pass(log, *head_blk, tmp_rhead_blk,
+				      XLOG_RECOVER_CRCPASS, &first_bad);
+	if (error == -EFSBADCRC) {
+		/*
+		 * We've hit a potential torn write. Reset the error and warn
+		 * about it.
+		 */
+		error = 0;
+		xfs_warn(log->l_mp,
+"Torn write (CRC failure) detected at log block 0x%llx. Truncating head block from 0x%llx.",
+			 first_bad, *head_blk);
+
+		/*
+		 * Get the header block and buffer pointer for the last good
+		 * record before the bad record.
+		 *
+		 * Note that xlog_find_tail() clears the blocks at the new head
+		 * (i.e., the records with invalid CRC) if the cycle number
+		 * matches the the current cycle.
+		 */
+		found = xlog_rseek_logrec_hdr(log, first_bad, *tail_blk, 1, bp,
+					      rhead_blk, rhead, wrapped);
+		if (found < 0)
+			return found;
+		if (found == 0)		/* XXX: right thing to do here? */
+			return -EIO;
+
+		/*
+		 * Reset the head block to the starting block of the first bad
+		 * log record and set the tail block based on the last good
+		 * record.
+		 *
+		 * Bail out if the updated head/tail match as this indicates
+		 * possible corruption outside of the acceptable
+		 * (XLOG_MAX_ICLOGS) range. This is a job for xfs_repair...
+		 */
+		*head_blk = first_bad;
+		*tail_blk = BLOCK_LSN(be64_to_cpu((*rhead)->h_tail_lsn));
+		if (*head_blk == *tail_blk) {
+			ASSERT(0);
+			return 0;
+		}
+
+		/*
+		 * Now verify the tail based on the updated head. This is
+		 * required because the torn writes trimmed from the head could
+		 * have been written over the tail of a previous record. Return
+		 * any errors since recovery cannot proceed if the tail is
+		 * corrupt.
+		 *
+		 * XXX: This leaves a gap in truly robust protection from torn
+		 * writes in the log. If the head is behind the tail, the tail
+		 * pushes forward to create some space and then a crash occurs
+		 * causing the writes into the previous record's tail region to
+		 * tear, log recovery isn't able to recover.
+		 *
+		 * How likely is this to occur? If possible, can we do something
+		 * more intelligent here? Is it safe to push the tail forward if
+		 * we can determine that the tail is within the range of the
+		 * torn write (e.g., the kernel can only overwrite the tail if
+		 * it has actually been pushed forward)? Alternatively, could we
+		 * somehow prevent this condition at runtime?
+		 */
+		error = xlog_verify_tail(log, *head_blk, *tail_blk);
+	}
+
+	return error;
+}
+
+/*
  * Find the sync block number or the tail of the log.
  *
  * This will be the block number of the last record to have its

@@ -966,9 +1241,10 @@ xlog_find_tail(
 	xlog_op_header_t	*op_head;
 	char			*offset = NULL;
 	xfs_buf_t		*bp;
-	int			error, i, found;
+	int			error;
 	xfs_daddr_t		umount_data_blk;
 	xfs_daddr_t		after_umount_blk;
+	xfs_daddr_t		rhead_blk;
 	xfs_lsn_t		tail_lsn;
 	int			hblks;
 	bool			wrapped = false;

@@ -995,24 +1271,16 @@ xlog_find_tail(
 	}
 
 	/*
-	 * Search backwards through the log looking for the log record header
-	 * block. This wraps all the way back around to the head so something is
-	 * seriously wrong if we can't find it.
+	 * Trim the head block back to skip over torn records. We can have
+	 * multiple log I/Os in flight at any time, so we assume CRC failures
+	 * back through the previous several records are torn writes and skip
+	 * them.
 	 */
 	ASSERT(*head_blk < INT_MAX);
-	found = xlog_rseek_logrec_hdr(log, *head_blk, *head_blk, 1, bp, &i,
-				      &rhead, &wrapped);
-	if (found < 0) {
-		error = found;
+	error = xlog_verify_head(log, head_blk, tail_blk, bp, &rhead_blk,
+				 &rhead, &wrapped);
+	if (error)
 		goto done;
-	}
-	if (!found) {
-		xfs_warn(log->l_mp, "%s: couldn't find sync record", __func__);
-		xlog_put_bp(bp);
-		ASSERT(0);
-		return -EIO;
-	}
-	*tail_blk = BLOCK_LSN(be64_to_cpu(rhead->h_tail_lsn));
 
 	/*
 	 * Reset log values according to the state of the log when we

@@ -1024,7 +1292,7 @@ xlog_find_tail(
 	 * written was complete and ended exactly on the end boundary
 	 * of the physical log.
 	 */
-	log->l_prev_block = i;
+	log->l_prev_block = rhead_blk;
 	log->l_curr_block = (int)*head_blk;
 	log->l_curr_cycle = be32_to_cpu(rhead->h_cycle);
 	if (wrapped)

@@ -1062,12 +1330,12 @@ xlog_find_tail(
 	} else {
 		hblks = 1;
 	}
-	after_umount_blk = (i + hblks + (int)
+	after_umount_blk = (rhead_blk + hblks + (int)
 		BTOBB(be32_to_cpu(rhead->h_len))) % log->l_logBBsize;
 	tail_lsn = atomic64_read(&log->l_tail_lsn);
 	if (*head_blk == after_umount_blk &&
 	    be32_to_cpu(rhead->h_num_logops) == 1) {
-		umount_data_blk = (i + hblks) % log->l_logBBsize;
+		umount_data_blk = (rhead_blk + hblks) % log->l_logBBsize;
 		error = xlog_bread(log, umount_data_blk, 1, bp, &offset);
 		if (error)
 			goto done;

index f52c72a..887c443 100644
@@ -1188,10 +1188,16 @@ xlog_iodone(xfs_buf_t *bp)
 	int			aborted = 0;
 
 	/*
-	 * Race to shutdown the filesystem if we see an error.
+	 * Race to shutdown the filesystem if we see an error or the iclog is in
+	 * IOABORT state. The IOABORT state is only set in DEBUG mode to inject
+	 * CRC errors into log recovery.
 	 */
-	if (XFS_TEST_ERROR(bp->b_error, l->l_mp,
-			XFS_ERRTAG_IODONE_IOERR, XFS_RANDOM_IODONE_IOERR)) {
+	if (XFS_TEST_ERROR(bp->b_error, l->l_mp, XFS_ERRTAG_IODONE_IOERR,
+			   XFS_RANDOM_IODONE_IOERR) ||
+	    iclog->ic_state & XLOG_STATE_IOABORT) {
+		if (iclog->ic_state & XLOG_STATE_IOABORT)
+			iclog->ic_state &= ~XLOG_STATE_IOABORT;
+
 		xfs_buf_ioerror_alert(bp, __func__);
 		xfs_buf_stale(bp);
 		xfs_force_shutdown(l->l_mp, SHUTDOWN_LOG_IO_ERROR);

@@ -1838,6 +1844,23 @@ xlog_sync(
 	/* calculcate the checksum */
 	iclog->ic_header.h_crc = xlog_cksum(log, &iclog->ic_header,
 					    iclog->ic_datap, size);
+#ifdef DEBUG
+	/*
+	 * Intentionally corrupt the log record CRC based on the error injection
+	 * frequency, if defined. This facilitates testing log recovery in the
+	 * event of torn writes. Hence, set the IOABORT state to abort the log
+	 * write on I/O completion and shutdown the fs. The subsequent mount
+	 * detects the bad CRC and attempts to recover.
+	 */
+	if (log->l_badcrc_factor &&
+	    (prandom_u32() % log->l_badcrc_factor == 0)) {
+		iclog->ic_header.h_crc &= 0xAAAAAAAA;
+		iclog->ic_state |= XLOG_STATE_IOABORT;
+		xfs_warn(log->l_mp,
+	"Intentionally corrupted log record at LSN 0x%llx. Shutdown imminent.",
+			 be64_to_cpu(iclog->ic_header.h_lsn));
+	}
+#endif
 
 	bp->b_io_length = BTOBB(count);
 	bp->b_fspriv = iclog;

@@ -2791,11 +2814,19 @@ xlog_state_do_callback(
 		}
 	} while (!ioerrors && loopdidcallbacks);
 
+#ifdef DEBUG
 	/*
-	 * make one last gasp attempt to see if iclogs are being left in
-	 * limbo..
+	 * Make one last gasp attempt to see if iclogs are being left in limbo.
+	 * If the above loop finds an iclog earlier than the current iclog and
+	 * in one of the syncing states, the current iclog is put into
+	 * DO_CALLBACK and the callbacks are deferred to the completion of the
+	 * earlier iclog. Walk the iclogs in order and make sure that no iclog
+	 * is in DO_CALLBACK unless an earlier iclog is in one of the syncing
+	 * states.
+	 *
+	 * Note that SYNCING|IOABORT is a valid state so we cannot just check
+	 * for ic_state == SYNCING.
 	 */
-#ifdef DEBUG
 	if (funcdidcallbacks) {
 		first_iclog = iclog = log->l_iclog;
 		do {

@@ -2810,7 +2841,7 @@ xlog_state_do_callback(
 			 * IOERROR - give up hope all ye who enter here
 			 */
 			if (iclog->ic_state == XLOG_STATE_WANT_SYNC ||
-			    iclog->ic_state == XLOG_STATE_SYNCING ||
+			    iclog->ic_state & XLOG_STATE_SYNCING ||
 			    iclog->ic_state == XLOG_STATE_DONE_SYNC ||
 			    iclog->ic_state == XLOG_STATE_IOERROR )
 				break;

index 8daba74..ed88963 100644
@@ -62,6 +62,7 @@ static inline uint xlog_get_client_id(__be32 i)
 #define XLOG_STATE_CALLBACK  0x0020 /* Callback functions now */
 #define XLOG_STATE_DIRTY     0x0040 /* Dirty IC log, not ready for ACTIVE status*/
 #define XLOG_STATE_IOERROR   0x0080 /* IO error happened in sync'ing log */
+#define XLOG_STATE_IOABORT   0x0100 /* force abort on I/O completion (debug) */
 #define XLOG_STATE_ALL	     0x7FFF /* All possible valid flags */
 #define XLOG_STATE_NOTUSED   0x8000 /* This IC log not being used */
 

@@ -410,6 +411,8 @@ struct xlog {
 	/* The following field are used for debugging; need to hold icloglock */
 #ifdef DEBUG
 	void			*l_iclog_bak[XLOG_MAX_ICLOGS];
+	/* log record crc error injection factor */
+	uint32_t		l_badcrc_factor;
 #endif
 
 };

index ee70f5d..641d625 100644
@@ -255,11 +255,47 @@ write_grant_head_show(
 }
 XFS_SYSFS_ATTR_RO(write_grant_head);
 
+#ifdef DEBUG
+STATIC ssize_t
+log_badcrc_factor_store(
+	struct kobject	*kobject,
+	const char	*buf,
+	size_t		count)
+{
+	struct xlog	*log = to_xlog(kobject);
+	int		ret;
+	uint32_t	val;
+
+	ret = kstrtouint(buf, 0, &val);
+	if (ret)
+		return ret;
+
+	log->l_badcrc_factor = val;
+
+	return count;
+}
+
+STATIC ssize_t
+log_badcrc_factor_show(
+	struct kobject	*kobject,
+	char		*buf)
+{
+	struct xlog	*log = to_xlog(kobject);
+
+	return snprintf(buf, PAGE_SIZE, "%d\n", log->l_badcrc_factor);
+}
+
+XFS_SYSFS_ATTR_RW(log_badcrc_factor);
+#endif	/* DEBUG */
+
 static struct attribute *xfs_log_attrs[] = {
 	ATTR_LIST(log_head_lsn),
 	ATTR_LIST(log_tail_lsn),
 	ATTR_LIST(reserve_grant_head),
 	ATTR_LIST(write_grant_head),
+#ifdef DEBUG
+	ATTR_LIST(log_badcrc_factor),
+#endif
 	NULL,
 };