ZIL: "crash" the ZIL if the pool suspends during fallback

cmd/zilstat.in

@@ -47,6 +47,7 @@ cols = {
 	"cec":       [5,         1000,       "zil_commit_error_count"],
 	"csc":       [5,         1000,       "zil_commit_stall_count"],
 	"cSc":       [5,         1000,       "zil_commit_suspend_count"],
+	"cCc":       [5,         1000,       "zil_commit_crash_count"],
 	"ic":        [5,         1000,       "zil_itx_count"],
 	"iic":       [5,         1000,       "zil_itx_indirect_count"],
 	"iib":       [5,         1024,       "zil_itx_indirect_bytes"],

cmd/ztest.c

@@ -1995,7 +1995,7 @@ ztest_log_write(ztest_ds_t *zd, dmu_tx_t *tx, lr_write_t *lr)
 	    dmu_read(zd->zd_os, lr->lr_foid, lr->lr_offset, lr->lr_length,
 	    ((lr_write_t *)&itx->itx_lr) + 1, DMU_READ_NO_PREFETCH |
 	    DMU_KEEP_CACHING) != 0) {
-		zil_itx_destroy(itx);
+		zil_itx_destroy(itx, 0);
 		itx = zil_itx_create(TX_WRITE, sizeof (*lr));
 		write_state = WR_NEED_COPY;
 	}
@@ -2965,7 +2965,7 @@ ztest_zil_commit(ztest_ds_t *zd, uint64_t id)
 
 	(void) pthread_rwlock_rdlock(&zd->zd_zilog_lock);
 
-	zil_commit(zilog, ztest_random(ZTEST_OBJECTS));
+	VERIFY0(zil_commit(zilog, ztest_random(ZTEST_OBJECTS)));
 
 	/*
 	 * Remember the committed values in zd, which is in parent/child
@@ -7933,7 +7933,7 @@ ztest_freeze(void)
 	 */
 	while (BP_IS_HOLE(&zd->zd_zilog->zl_header->zh_log)) {
 		ztest_dmu_object_alloc_free(zd, 0);
-		zil_commit(zd->zd_zilog, 0);
+		VERIFY0(zil_commit(zd->zd_zilog, 0));
 	}
 
 	txg_wait_synced(spa_get_dsl(spa), 0);
@@ -7975,7 +7975,7 @@ ztest_freeze(void)
 	/*
 	 * Commit all of the changes we just generated.
 	 */
-	zil_commit(zd->zd_zilog, 0);
+	VERIFY0(zil_commit(zd->zd_zilog, 0));
 	txg_wait_synced(spa_get_dsl(spa), 0);
 
 	/*

config/kernel-sb-wb-err.m4

@@ -0,0 +1,27 @@
+# dnl
+# dnl 5.8 (735e4ae5ba28) introduced a superblock scoped errseq_t to use to
+# dnl record writeback errors for syncfs() to return. Up until 5.17, when
+# dnl sync_fs errors were returned directly, this is the only way for us to
+# dnl report an error from syncfs().
+# dnl
+AC_DEFUN([ZFS_AC_KERNEL_SRC_SUPER_BLOCK_S_WB_ERR], [
+	ZFS_LINUX_TEST_SRC([super_block_s_wb_err], [
+		#include <linux/fs.h>
+
+		static const struct super_block
+		    sb __attribute__ ((unused)) = {
+			.s_wb_err = 0,
+		};
+	],[])
+])
+
+AC_DEFUN([ZFS_AC_KERNEL_SUPER_BLOCK_S_WB_ERR], [
+	AC_MSG_CHECKING([whether super_block has s_wb_err])
+	ZFS_LINUX_TEST_RESULT([super_block_s_wb_err], [
+		AC_MSG_RESULT(yes)
+		AC_DEFINE(HAVE_SUPER_BLOCK_S_WB_ERR, 1,
+			[have super_block s_wb_err])
+	],[
+		AC_MSG_RESULT(no)
+	])
+])

config/kernel.m4

@@ -131,6 +131,7 @@ AC_DEFUN([ZFS_AC_KERNEL_TEST_SRC], [
 	ZFS_AC_KERNEL_SRC_FILE
 	ZFS_AC_KERNEL_SRC_PIN_USER_PAGES
 	ZFS_AC_KERNEL_SRC_TIMER
+	ZFS_AC_KERNEL_SRC_SUPER_BLOCK_S_WB_ERR
 	case "$host_cpu" in
 		powerpc*)
 			ZFS_AC_KERNEL_SRC_CPU_HAS_FEATURE
@@ -246,6 +247,7 @@ AC_DEFUN([ZFS_AC_KERNEL_TEST_RESULT], [
 	ZFS_AC_KERNEL_FILE
 	ZFS_AC_KERNEL_PIN_USER_PAGES
 	ZFS_AC_KERNEL_TIMER
+	ZFS_AC_KERNEL_SUPER_BLOCK_S_WB_ERR
 	case "$host_cpu" in
 		powerpc*)
 			ZFS_AC_KERNEL_CPU_HAS_FEATURE

include/os/freebsd/spl/sys/debug.h

@@ -69,6 +69,10 @@
 #define	__maybe_unused __attribute__((unused))
 #endif
 
+#ifndef __must_check
+#define	__must_check __attribute__((__warn_unused_result__))
+#endif
+
 /*
  * Without this, we see warnings from objtool during normal Linux builds when
  * the kernel is built with CONFIG_STACK_VALIDATION=y:

include/os/freebsd/spl/sys/vm.h

@@ -35,6 +35,8 @@
 extern const int zfs_vm_pagerret_bad;
 extern const int zfs_vm_pagerret_error;
 extern const int zfs_vm_pagerret_ok;
+extern const int zfs_vm_pagerret_pend;
+extern const int zfs_vm_pagerret_again;
 extern const int zfs_vm_pagerput_sync;
 extern const int zfs_vm_pagerput_inval;
 

include/os/linux/spl/sys/debug.h

@@ -69,6 +69,10 @@
 #define	__maybe_unused __attribute__((unused))
 #endif
 
+#ifndef __must_check
+#define	__must_check __attribute__((__warn_unused_result__))
+#endif
+
 /*
  * Without this, we see warnings from objtool during normal Linux builds when
  * the kernel is built with CONFIG_STACK_VALIDATION=y:

include/sys/zil.h

@@ -456,7 +456,7 @@ typedef enum {
 	WR_NUM_STATES	/* number of states */
 } itx_wr_state_t;
 
-typedef void (*zil_callback_t)(void *data);
+typedef void (*zil_callback_t)(void *data, int err);
 
 typedef struct itx {
 	list_node_t	itx_node;	/* linkage on zl_itx_list */
@@ -498,10 +498,13 @@ typedef struct zil_stats {
 	 *            (see zil_commit_writer_stall())
 	 * - suspend: ZIL suspended
 	 *            (see zil_commit(), zil_get_commit_list())
+	 * -   crash: ZIL crashed
+	 *            (see zil_crash(), zil_commit(), ...)
 	 */
 	kstat_named_t zil_commit_error_count;
 	kstat_named_t zil_commit_stall_count;
 	kstat_named_t zil_commit_suspend_count;
+	kstat_named_t zil_commit_crash_count;
 
 	/*
 	 * Number of transactions (reads, writes, renames, etc.)
@@ -549,6 +552,7 @@ typedef struct zil_sums {
 	wmsum_t zil_commit_error_count;
 	wmsum_t zil_commit_stall_count;
 	wmsum_t zil_commit_suspend_count;
+	wmsum_t zil_commit_crash_count;
 	wmsum_t zil_itx_count;
 	wmsum_t zil_itx_indirect_count;
 	wmsum_t zil_itx_indirect_bytes;
@@ -606,12 +610,11 @@ extern boolean_t zil_destroy(zilog_t *zilog, boolean_t keep_first);
 extern void	zil_destroy_sync(zilog_t *zilog, dmu_tx_t *tx);
 
 extern itx_t	*zil_itx_create(uint64_t txtype, size_t lrsize);
-extern void	zil_itx_destroy(itx_t *itx);
+extern void	zil_itx_destroy(itx_t *itx, int err);
 extern void	zil_itx_assign(zilog_t *zilog, itx_t *itx, dmu_tx_t *tx);
 
 extern void	zil_async_to_sync(zilog_t *zilog, uint64_t oid);
-extern void	zil_commit(zilog_t *zilog, uint64_t oid);
-extern void	zil_commit_impl(zilog_t *zilog, uint64_t oid);
+extern int __must_check	zil_commit(zilog_t *zilog, uint64_t oid);
 extern void	zil_remove_async(zilog_t *zilog, uint64_t oid);
 
 extern int	zil_reset(const char *osname, void *txarg);

include/sys/zil_impl.h

@@ -221,6 +221,7 @@ struct zilog {
 	uint64_t	zl_cur_left;	/* current burst remaining size */
 	uint64_t	zl_cur_max;	/* biggest record in current burst */
 	list_t		zl_lwb_list;	/* in-flight log write list */
+	list_t		zl_lwb_crash_list; /* log writes in-flight at crash */
 	avl_tree_t	zl_bp_tree;	/* track bps during log parse */
 	clock_t		zl_replay_time;	/* lbolt of when replay started */
 	uint64_t	zl_replay_blks;	/* number of log blocks replayed */
@@ -245,6 +246,9 @@ struct zilog {
 	 */
 	uint64_t	zl_max_block_size;
 
+	/* After crash, txg to restart zil */
+	uint64_t	zl_restart_txg;
+
 	/* Pointer for per dataset zil sums */
 	zil_sums_t *zl_sums;
 };

lib/libspl/include/sys/debug.h

@@ -38,4 +38,8 @@
 #define	__maybe_unused __attribute__((unused))
 #endif
 
+#ifndef __must_check
+#define	__must_check __attribute__((warn_unused_result))
+#endif
+
 #endif

module/os/freebsd/spl/spl_vm.c

@@ -43,6 +43,8 @@
 const int zfs_vm_pagerret_bad = VM_PAGER_BAD;
 const int zfs_vm_pagerret_error = VM_PAGER_ERROR;
 const int zfs_vm_pagerret_ok = VM_PAGER_OK;
+const int zfs_vm_pagerret_pend = VM_PAGER_PEND;
+const int zfs_vm_pagerret_again = VM_PAGER_AGAIN;
 const int zfs_vm_pagerput_sync = VM_PAGER_PUT_SYNC;
 const int zfs_vm_pagerput_inval = VM_PAGER_PUT_INVAL;
 

module/os/freebsd/zfs/zfs_vfsops.c

@@ -455,8 +455,13 @@ zfs_sync(vfs_t *vfsp, int waitfor)
 			return (0);
 		}
 
-		if (zfsvfs->z_log != NULL)
-			zil_commit(zfsvfs->z_log, 0);
+		if (zfsvfs->z_log != NULL) {
+			error = zil_commit(zfsvfs->z_log, 0);
+			if (error != 0) {
+				zfs_exit(zfsvfs, FTAG);
+				return (error);
+			}
+		}
 
 		zfs_exit(zfsvfs, FTAG);
 	} else {