[PATCH mdadm v4 0/7] Write Zeroes option for Creating Arrays

[PATCH mdadm v4 0/7] Write Zeroes option for Creating Arrays

[Logan Gunthorpe]

Hi,

This is the next iteration of the patchset that added the discard
option to mdadm. Per feedback from Martin, it's more desirable
to use the write-zeroes functionality than rely on devices to zero
the data on a discard request. This is because standards typically
only require the device to do the best effort to discard data and
may not actually discard (and thus zero) it all in some circumstances.

This version of the patch set adds the --write-zeroes option which
will imply --assume-clean and write zeros to the data region in
each disk before starting the array. This can take some time so
each disk is done in parallel in its own fork. To make the forking
code easier to understand this patch set also starts with some
cleanup of the existing Create code.

We tested write-zeroes requests on a number of modern nvme drives of
various manufacturers and found most are not as optimized as the
discard path. A couple drives that were tested did not support
write-zeroes at all but still performed similarly with the kernel
falling back to writing zero pages. Typically we see it take on the
order of one minute per 100GB of data zeroed.

One reason write-zeroes is slower than discard is that today's NVMe
devices only allow about 2MB to be zeroed in one command where as
the entire drive can typically be discarded in one command. Partly,
this is a limitation of the spec as there are only 16 bits avalaible
in the write-zeros command size but drives still don't max this out.
Hopefully, in the future this will all be optimized a bit more
and this work will be able to take advantage of that.

Logan

--

Changes since v3:
   * Store the pid in a local variable instead of the mdinfo struct
    (per Mariusz and Xiao)

Changes since v2:

   * Use write-zeroes instead of discard to zero the disks (per
     Martin)
   * Due to the time required to zero the disks, each disk is
     now done in parallel with separate forks of the process.
   * In order to add the forking some refactoring was done on the
     Create() function to make it easier to understand
   * Added a pr_info() call so that some prints can be done
     to stdout instead of stdour (per Mariusz)
   * Added KIB_TO_BYTES and SEC_TO_BYTES helpers (per Mariusz)
   * Added a test to the mdadm test suite to test the option
     works.
   * Fixed up how the size and offset are calculated with some
     great information from Xiao.

Changes since v1:

   * Discard the data in the devices later in the create process
     while they are already open. This requires treating the
     s.discard option the same as the s.assume_clean option.
     Per Mariusz.
   * A couple other minor cleanup changes from Mariusz.


*** BLURB HERE ***

Logan Gunthorpe (7):
  Create: goto abort_locked instead of return 1 in error path
  Create: remove safe_mode_delay local variable
  Create: Factor out add_disks() helpers
  mdadm: Introduce pr_info()
  mdadm: Add --write-zeros option for Create
  tests/00raid5-zero: Introduce test to exercise --write-zeros.
  manpage: Add --write-zeroes option to manpage

 Create.c           | 479 ++++++++++++++++++++++++++++-----------------
 ReadMe.c           |   2 +
 mdadm.8.in         |  16 ++
 mdadm.c            |   9 +
 mdadm.h            |   7 +
 tests/00raid5-zero |  12 ++
 6 files changed, 350 insertions(+), 175 deletions(-)
 create mode 100644 tests/00raid5-zero


base-commit: 8b668d4aa3305af5963162b7499b128bd71f8f29
--
2.30.2

[PATCH mdadm v4 1/7] Create: goto abort_locked instead of return 1 in error path

[Logan Gunthorpe]

The return 1 after the fstat_is_blkdev() check should be replaced
with an error return that goes through the error path to unlock
resources locked by this function.

Signed-off-by: Logan Gunthorpe <logang@deltatee.com>
---
 Create.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/Create.c b/Create.c
index 953e73722518..2e8203ecdccd 100644
--- a/Create.c
+++ b/Create.c
@@ -939,7 +939,7 @@ int Create(struct supertype *st, char *mddev,
 						goto abort_locked;
 					}
 					if (!fstat_is_blkdev(fd, dv->devname, &rdev))
-						return 1;
+						goto abort_locked;
 					inf->disk.major = major(rdev);
 					inf->disk.minor = minor(rdev);
 				}
-- 
2.30.2

[PATCH mdadm v4 2/7] Create: remove safe_mode_delay local variable

[Logan Gunthorpe]

All .getinfo_super() call sets the info.safe_mode_delay variables
to a constant value, so no matter what the current state is
that function will always set it to the same value.

Create() calls .getinfo_super() multiple times while creating the array.
The value is stored in a local variable for every disk in the loop
to add disks (so the last disc call takes precedence). The local
variable is then used in the call to sysfs_set_safemode().

This can be simplified by using info.safe_mode_delay directly. The info
variable had .getinfo_super() called on it early in the function so, by the
reasoning above, it will have the same value as the local variable which
can thus be removed.

Doing this allows for factoring out code from Create() in a subsequent
patch.

Signed-off-by: Logan Gunthorpe <logang@deltatee.com>
---
 Create.c | 4 +---
 1 file changed, 1 insertion(+), 3 deletions(-)

diff --git a/Create.c b/Create.c
index 2e8203ecdccd..8ded81dc265d 100644
--- a/Create.c
+++ b/Create.c
@@ -137,7 +137,6 @@ int Create(struct supertype *st, char *mddev,
 	int did_default = 0;
 	int do_default_layout = 0;
 	int do_default_chunk = 0;
-	unsigned long safe_mode_delay = 0;
 	char chosen_name[1024];
 	struct map_ent *map = NULL;
 	unsigned long long newsize;
@@ -952,7 +951,6 @@ int Create(struct supertype *st, char *mddev,
 					goto abort_locked;
 				}
 				st->ss->getinfo_super(st, inf, NULL);
-				safe_mode_delay = inf->safe_mode_delay;
 
 				if (have_container && c->verbose > 0)
 					pr_err("Using %s for device %d\n",
@@ -1065,7 +1063,7 @@ int Create(struct supertype *st, char *mddev,
 						    "readonly");
 				break;
 			}
-			sysfs_set_safemode(&info, safe_mode_delay);
+			sysfs_set_safemode(&info, info.safe_mode_delay);
 			if (err) {
 				pr_err("failed to activate array.\n");
 				ioctl(mdfd, STOP_ARRAY, NULL);
-- 
2.30.2

[PATCH mdadm v4 3/7] Create: Factor out add_disks() helpers

[Logan Gunthorpe]

The Create function is massive with a very large number of variables.
Reading and understanding the function is almost impossible. To help
with this, factor out the two pass loop that adds the disks to the array.

This moves about 160 lines into three new helper functions and removes
a bunch of local variables from the main Create function. The main new
helper function add_disks() does the two pass loop and calls into
add_disk_to_super() and update_metadata(). Factoring out the
latter two helpers also helps to reduce a ton of indentation.

No functional changes intended.

Signed-off-by: Logan Gunthorpe <logang@deltatee.com>
---
 Create.c | 382 +++++++++++++++++++++++++++++++------------------------
 1 file changed, 213 insertions(+), 169 deletions(-)

diff --git a/Create.c b/Create.c
index 8ded81dc265d..6a0446644e04 100644
--- a/Create.c
+++ b/Create.c
@@ -91,6 +91,214 @@ int default_layout(struct supertype *st, int level, int verbose)
 	return layout;
 }
 
+static int add_disk_to_super(int mdfd, struct shape *s, struct context *c,
+		struct supertype *st, struct mddev_dev *dv,
+		struct mdinfo *info, int have_container, int major_num)
+{
+	dev_t rdev;
+	int fd;
+
+	if (dv->disposition == 'j') {
+		info->disk.raid_disk = MD_DISK_ROLE_JOURNAL;
+		info->disk.state = (1<<MD_DISK_JOURNAL);
+	} else if (info->disk.raid_disk < s->raiddisks) {
+		info->disk.state = (1<<MD_DISK_ACTIVE) |
+			(1<<MD_DISK_SYNC);
+	} else {
+		info->disk.state = 0;
+	}
+
+	if (dv->writemostly == FlagSet) {
+		if (major_num == BITMAP_MAJOR_CLUSTERED) {
+			pr_err("Can not set %s --write-mostly with a clustered bitmap\n",dv->devname);
+			return 1;
+		} else {
+			info->disk.state |= (1<<MD_DISK_WRITEMOSTLY);
+		}
+
+	}
+
+	if (dv->failfast == FlagSet)
+		info->disk.state |= (1<<MD_DISK_FAILFAST);
+
+	if (have_container) {
+		fd = -1;
+	} else {
+		if (st->ss->external && st->container_devnm[0])
+			fd = open(dv->devname, O_RDWR);
+		else
+			fd = open(dv->devname, O_RDWR|O_EXCL);
+
+		if (fd < 0) {
+			pr_err("failed to open %s after earlier success - aborting\n",
+			       dv->devname);
+			return 1;
+		}
+		if (!fstat_is_blkdev(fd, dv->devname, &rdev))
+			return 1;
+		info->disk.major = major(rdev);
+		info->disk.minor = minor(rdev);
+	}
+	if (fd >= 0)
+		remove_partitions(fd);
+	if (st->ss->add_to_super(st, &info->disk, fd, dv->devname,
+				 dv->data_offset)) {
+		ioctl(mdfd, STOP_ARRAY, NULL);
+		return 1;
+	}
+	st->ss->getinfo_super(st, info, NULL);
+
+	if (have_container && c->verbose > 0)
+		pr_err("Using %s for device %d\n",
+		       map_dev(info->disk.major, info->disk.minor, 0),
+		       info->disk.number);
+
+	if (!have_container) {
+		/* getinfo_super might have lost these ... */
+		info->disk.major = major(rdev);
+		info->disk.minor = minor(rdev);
+	}
+
+	return 0;
+}
+
+static int update_metadata(int mdfd, struct shape *s, struct supertype *st,
+			   struct map_ent **map, struct mdinfo *info,
+			   char *chosen_name)
+{
+	struct mdinfo info_new;
+	struct map_ent *me = NULL;
+
+	/* check to see if the uuid has changed due to these
+	 * metadata changes, and if so update the member array
+	 * and container uuid.  Note ->write_init_super clears
+	 * the subarray cursor such that ->getinfo_super once
+	 * again returns container info.
+	 */
+	st->ss->getinfo_super(st, &info_new, NULL);
+	if (st->ss->external && is_container(s->level) &&
+	    !same_uuid(info_new.uuid, info->uuid, 0)) {
+		map_update(map, fd2devnm(mdfd),
+			   info_new.text_version,
+			   info_new.uuid, chosen_name);
+		me = map_by_devnm(map, st->container_devnm);
+	}
+
+	if (st->ss->write_init_super(st)) {
+		st->ss->free_super(st);
+		return 1;
+	}
+
+	/*
+	 * Before activating the array, perform extra steps
+	 * required to configure the internal write-intent
+	 * bitmap.
+	 */
+	if (info_new.consistency_policy == CONSISTENCY_POLICY_BITMAP &&
+	    st->ss->set_bitmap && st->ss->set_bitmap(st, info)) {
+		st->ss->free_super(st);
+		return 1;
+	}
+
+	/* update parent container uuid */
+	if (me) {
+		char *path = xstrdup(me->path);
+
+		st->ss->getinfo_super(st, &info_new, NULL);
+		map_update(map, st->container_devnm, info_new.text_version,
+			   info_new.uuid, path);
+		free(path);
+	}
+
+	flush_metadata_updates(st);
+	st->ss->free_super(st);
+
+	return 0;
+}
+
+static int add_disks(int mdfd, struct mdinfo *info, struct shape *s,
+		     struct context *c, struct supertype *st,
+		     struct map_ent **map, struct mddev_dev *devlist,
+		     int total_slots, int have_container, int insert_point,
+		     int major_num, char *chosen_name)
+{
+	struct mddev_dev *moved_disk = NULL;
+	int pass, raid_disk_num, dnum;
+	struct mddev_dev *dv;
+	struct mdinfo *infos;
+	int ret = 0;
+
+	infos = xmalloc(sizeof(*infos) * total_slots);
+	enable_fds(total_slots);
+	for (pass = 1; pass <= 2; pass++) {
+		for (dnum = 0, raid_disk_num = 0, dv = devlist; dv;
+		     dv = (dv->next) ? (dv->next) : moved_disk, dnum++) {
+			if (dnum >= total_slots)
+				abort();
+			if (dnum == insert_point) {
+				raid_disk_num += 1;
+				moved_disk = dv;
+				continue;
+			}
+			if (strcasecmp(dv->devname, "missing") == 0) {
+				raid_disk_num += 1;
+				continue;
+			}
+			if (have_container)
+				moved_disk = NULL;
+			if (have_container && dnum < total_slots - 1)
+				/* repeatedly use the container */
+				moved_disk = dv;
+
+			switch(pass) {
+			case 1:
+				infos[dnum] = *info;
+				infos[dnum].disk.number = dnum;
+				infos[dnum].disk.raid_disk = raid_disk_num++;
+
+				if (dv->disposition == 'j')
+					raid_disk_num--;
+
+				ret = add_disk_to_super(mdfd, s, c, st, dv,
+						&infos[dnum], have_container,
+						major_num);
+				if (ret)
+					goto out;
+
+				break;
+			case 2:
+				infos[dnum].errors = 0;
+
+				ret = add_disk(mdfd, st, info, &infos[dnum]);
+				if (ret) {
+					pr_err("ADD_NEW_DISK for %s failed: %s\n",
+					       dv->devname, strerror(errno));
+					if (errno == EINVAL &&
+					    info->array.level == 0) {
+						pr_err("Possibly your kernel doesn't support RAID0 layouts.\n");
+						pr_err("Either upgrade, or use --layout=dangerous\n");
+					}
+					goto out;
+				}
+				break;
+			}
+			if (!have_container &&
+			    dv == moved_disk && dnum != insert_point) break;
+		}
+
+		if (pass == 1) {
+			ret = update_metadata(mdfd, s, st, map, info,
+					      chosen_name);
+			if (ret)
+				goto out;
+		}
+	}
+
+out:
+	free(infos);
+	return ret;
+}
+
 int Create(struct supertype *st, char *mddev,
 	   char *name, int *uuid,
 	   int subdevs, struct mddev_dev *devlist,
@@ -117,7 +325,7 @@ int Create(struct supertype *st, char *mddev,
 	unsigned long long minsize = 0, maxsize = 0;
 	char *mindisc = NULL;
 	char *maxdisc = NULL;
-	int dnum, raid_disk_num;
+	int dnum;
 	struct mddev_dev *dv;
 	dev_t rdev;
 	int fail = 0, warn = 0;
@@ -126,14 +334,13 @@ int Create(struct supertype *st, char *mddev,
 	int missing_disks = 0;
 	int insert_point = subdevs * 2; /* where to insert a missing drive */
 	int total_slots;
-	int pass;
 	int rv;
 	int bitmap_fd;
 	int have_container = 0;
 	int container_fd = -1;
 	int need_mdmon = 0;
 	unsigned long long bitmapsize;
-	struct mdinfo info, *infos;
+	struct mdinfo info;
 	int did_default = 0;
 	int do_default_layout = 0;
 	int do_default_chunk = 0;
@@ -869,174 +1076,11 @@ int Create(struct supertype *st, char *mddev,
 		}
 	}
 
-	infos = xmalloc(sizeof(*infos) * total_slots);
-	enable_fds(total_slots);
-	for (pass = 1; pass <= 2; pass++) {
-		struct mddev_dev *moved_disk = NULL; /* the disk that was moved out of the insert point */
-
-		for (dnum = 0, raid_disk_num = 0, dv = devlist; dv;
-		     dv = (dv->next) ? (dv->next) : moved_disk, dnum++) {
-			int fd;
-			struct mdinfo *inf = &infos[dnum];
-
-			if (dnum >= total_slots)
-				abort();
-			if (dnum == insert_point) {
-				raid_disk_num += 1;
-				moved_disk = dv;
-				continue;
-			}
-			if (strcasecmp(dv->devname, "missing") == 0) {
-				raid_disk_num += 1;
-				continue;
-			}
-			if (have_container)
-				moved_disk = NULL;
-			if (have_container && dnum < info.array.raid_disks - 1)
-				/* repeatedly use the container */
-				moved_disk = dv;
-
-			switch(pass) {
-			case 1:
-				*inf = info;
-
-				inf->disk.number = dnum;
-				inf->disk.raid_disk = raid_disk_num++;
-
-				if (dv->disposition == 'j') {
-					inf->disk.raid_disk = MD_DISK_ROLE_JOURNAL;
-					inf->disk.state = (1<<MD_DISK_JOURNAL);
-					raid_disk_num--;
-				} else if (inf->disk.raid_disk < s->raiddisks)
-					inf->disk.state = (1<<MD_DISK_ACTIVE) |
-						(1<<MD_DISK_SYNC);
-				else
-					inf->disk.state = 0;
-
-				if (dv->writemostly == FlagSet) {
-					if (major_num == BITMAP_MAJOR_CLUSTERED) {
-						pr_err("Can not set %s --write-mostly with a clustered bitmap\n",dv->devname);
-						goto abort_locked;
-					} else
-						inf->disk.state |= (1<<MD_DISK_WRITEMOSTLY);
-				}
-				if (dv->failfast == FlagSet)
-					inf->disk.state |= (1<<MD_DISK_FAILFAST);
-
-				if (have_container)
-					fd = -1;
-				else {
-					if (st->ss->external &&
-					    st->container_devnm[0])
-						fd = open(dv->devname, O_RDWR);
-					else
-						fd = open(dv->devname, O_RDWR|O_EXCL);
-
-					if (fd < 0) {
-						pr_err("failed to open %s after earlier success - aborting\n",
-							dv->devname);
-						goto abort_locked;
-					}
-					if (!fstat_is_blkdev(fd, dv->devname, &rdev))
-						goto abort_locked;
-					inf->disk.major = major(rdev);
-					inf->disk.minor = minor(rdev);
-				}
-				if (fd >= 0)
-					remove_partitions(fd);
-				if (st->ss->add_to_super(st, &inf->disk,
-							 fd, dv->devname,
-							 dv->data_offset)) {
-					ioctl(mdfd, STOP_ARRAY, NULL);
-					goto abort_locked;
-				}
-				st->ss->getinfo_super(st, inf, NULL);
-
-				if (have_container && c->verbose > 0)
-					pr_err("Using %s for device %d\n",
-						map_dev(inf->disk.major,
-							inf->disk.minor,
-							0), dnum);
-
-				if (!have_container) {
-					/* getinfo_super might have lost these ... */
-					inf->disk.major = major(rdev);
-					inf->disk.minor = minor(rdev);
-				}
-				break;
-			case 2:
-				inf->errors = 0;
-
-				rv = add_disk(mdfd, st, &info, inf);
-
-				if (rv) {
-					pr_err("ADD_NEW_DISK for %s failed: %s\n",
-					       dv->devname, strerror(errno));
-					if (errno == EINVAL &&
-					    info.array.level == 0) {
-						pr_err("Possibly your kernel doesn't support RAID0 layouts.\n");
-						pr_err("Either upgrade, or use --layout=dangerous\n");
-					}
-					goto abort_locked;
-				}
-				break;
-			}
-			if (!have_container &&
-			    dv == moved_disk && dnum != insert_point) break;
-		}
-		if (pass == 1) {
-			struct mdinfo info_new;
-			struct map_ent *me = NULL;
-
-			/* check to see if the uuid has changed due to these
-			 * metadata changes, and if so update the member array
-			 * and container uuid.  Note ->write_init_super clears
-			 * the subarray cursor such that ->getinfo_super once
-			 * again returns container info.
-			 */
-			st->ss->getinfo_super(st, &info_new, NULL);
-			if (st->ss->external && !is_container(s->level) &&
-			    !same_uuid(info_new.uuid, info.uuid, 0)) {
-				map_update(&map, fd2devnm(mdfd),
-					   info_new.text_version,
-					   info_new.uuid, chosen_name);
-				me = map_by_devnm(&map, st->container_devnm);
-			}
-
-			if (st->ss->write_init_super(st)) {
-				st->ss->free_super(st);
-				goto abort_locked;
-			}
-			/*
-			 * Before activating the array, perform extra steps
-			 * required to configure the internal write-intent
-			 * bitmap.
-			 */
-			if (info_new.consistency_policy ==
-				    CONSISTENCY_POLICY_BITMAP &&
-			    st->ss->set_bitmap &&
-			    st->ss->set_bitmap(st, &info)) {
-				st->ss->free_super(st);
-				goto abort_locked;
-			}
-
-			/* update parent container uuid */
-			if (me) {
-				char *path = xstrdup(me->path);
-
-				st->ss->getinfo_super(st, &info_new, NULL);
-				map_update(&map, st->container_devnm,
-					   info_new.text_version,
-					   info_new.uuid, path);
-				free(path);
-			}
+	if (add_disks(mdfd, &info, s, c, st, &map, devlist, total_slots,
+		      have_container, insert_point, major_num, chosen_name))
+		goto abort_locked;
 
-			flush_metadata_updates(st);
-			st->ss->free_super(st);
-		}
-	}
 	map_unlock(&map);
-	free(infos);
 
 	if (is_container(s->level)) {
 		/* No need to start.  But we should signal udev to
-- 
2.30.2

[PATCH mdadm v4 4/7] mdadm: Introduce pr_info()

[Logan Gunthorpe]

Feedback was given to avoid informational pr_err() calls that print
to stderr, even though that's done all through out the code.

Using printf() directly doesn't maintain the same format (an "mdadm"
prefix on every line.

So introduce pr_info() which prints to stdout with the same format
and use it for a couple informational pr_err() calls in Create().

Future work can make this call used in more cases.

Signed-off-by: Logan Gunthorpe <logang@deltatee.com>
---
 Create.c | 7 ++++---
 mdadm.h  | 2 ++
 2 files changed, 6 insertions(+), 3 deletions(-)

diff --git a/Create.c b/Create.c
index 6a0446644e04..4acda30c5256 100644
--- a/Create.c
+++ b/Create.c
@@ -984,11 +984,12 @@ int Create(struct supertype *st, char *mddev,
 
 			mdi = sysfs_read(-1, devnm, GET_VERSION);
 
-			pr_err("Creating array inside %s container %s\n",
+			pr_info("Creating array inside %s container %s\n",
 				mdi?mdi->text_version:"managed", devnm);
 			sysfs_free(mdi);
 		} else
-			pr_err("Defaulting to version %s metadata\n", info.text_version);
+			pr_info("Defaulting to version %s metadata\n",
+				info.text_version);
 	}
 
 	map_update(&map, fd2devnm(mdfd), info.text_version,
@@ -1145,7 +1146,7 @@ int Create(struct supertype *st, char *mddev,
 			ioctl(mdfd, RESTART_ARRAY_RW, NULL);
 		}
 		if (c->verbose >= 0)
-			pr_err("array %s started.\n", mddev);
+			pr_info("array %s started.\n", mddev);
 		if (st->ss->external && st->container_devnm[0]) {
 			if (need_mdmon)
 				start_mdmon(st->container_devnm);
diff --git a/mdadm.h b/mdadm.h
index 3673494e560b..18c24915e94c 100644
--- a/mdadm.h
+++ b/mdadm.h
@@ -1798,6 +1798,8 @@ static inline int xasprintf(char **strp, const char *fmt, ...) {
 #endif
 #define cont_err(fmt ...) fprintf(stderr, "       " fmt)
 
+#define pr_info(fmt, args...) printf("%s: "fmt, Name, ##args)
+
 void *xmalloc(size_t len);
 void *xrealloc(void *ptr, size_t len);
 void *xcalloc(size_t num, size_t size);
-- 
2.30.2

[PATCH mdadm v4 5/7] mdadm: Add --write-zeros option for Create

[Logan Gunthorpe]

Add the --write-zeros option for Create which will send a write zeros
request to all the disks before assembling the array. After zeroing
the array, the disks will be in a known clean state and the initial
sync may be skipped.

Writing zeroes is best used when there is a hardware offload method
to zero the data. But even still, zeroing can take several minutes on
a large device. Because of this, all disks are zeroed in parallel using
their own forked process and a message is printed to the user. The main
process will proceed only after all the zeroing processes have completed
successfully.

Signed-off-by: Logan Gunthorpe <logang@deltatee.com>

fixup! mdadm: Add --write-zeros option for Create
---
 Create.c | 90 ++++++++++++++++++++++++++++++++++++++++++++++++++++++--
 ReadMe.c |  2 ++
 mdadm.c  |  9 ++++++
 mdadm.h  |  5 ++++
 4 files changed, 104 insertions(+), 2 deletions(-)

diff --git a/Create.c b/Create.c
index 4acda30c5256..36d63244f9ad 100644
--- a/Create.c
+++ b/Create.c
@@ -26,6 +26,8 @@
 #include	"md_u.h"
 #include	"md_p.h"
 #include	<ctype.h>
+#include	<fcntl.h>
+#include	<sys/wait.h>
 
 static int round_size_and_verify(unsigned long long *size, int chunk)
 {
@@ -91,9 +93,77 @@ int default_layout(struct supertype *st, int level, int verbose)
 	return layout;
 }
 
+static pid_t write_zeroes_fork(int fd, struct shape *s, struct supertype *st,
+			       struct mddev_dev *dv)
+
+{
+	unsigned long long offset_bytes, size_bytes;
+	int ret = 0;
+	pid_t pid;
+
+	size_bytes = KIB_TO_BYTES(s->size);
+
+	/*
+	 * If size_bytes is zero, this is a zoned raid array where
+	 * each disk is of a different size and uses its full
+	 * disk. Thus zero the entire disk.
+	 */
+	if (!size_bytes && !get_dev_size(fd, dv->devname, &size_bytes))
+		return -1;
+
+	if (dv->data_offset != INVALID_SECTORS)
+		offset_bytes = SEC_TO_BYTES(dv->data_offset);
+	else
+		offset_bytes = SEC_TO_BYTES(st->data_offset);
+
+	pr_info("zeroing data from %lld to %lld on: %s\n",
+		offset_bytes, size_bytes, dv->devname);
+
+	pid = fork();
+	if (pid < 0) {
+		pr_err("Could not fork to zero disks: %m\n");
+		return pid;
+	} else if (pid != 0) {
+		return pid;
+	}
+
+	if (fallocate(fd, FALLOC_FL_ZERO_RANGE | FALLOC_FL_KEEP_SIZE,
+		      offset_bytes, size_bytes)) {
+		pr_err("zeroing %s failed: %m\n", dv->devname);
+		ret = 1;
+	}
+
+	exit(ret);
+}
+
+static int wait_for_zero_forks(int *zero_pids, int count)
+{
+	int wstatus, ret = 0, i;
+	bool waited = false;
+
+	for (i = 0; i < count; i++) {
+		if (!zero_pids[i])
+			continue;
+
+		waited = true;
+		waitpid(zero_pids[i], &wstatus, 0);
+
+		if (!WIFEXITED(wstatus) || WEXITSTATUS(wstatus))
+			ret = 1;
+	}
+
+	if (ret)
+		pr_err("zeroing failed!\n");
+	else if (waited)
+		pr_info("zeroing finished\n");
+
+	return ret;
+}
+
 static int add_disk_to_super(int mdfd, struct shape *s, struct context *c,
 		struct supertype *st, struct mddev_dev *dv,
-		struct mdinfo *info, int have_container, int major_num)
+		struct mdinfo *info, int have_container, int major_num,
+		int *zero_pid)
 {
 	dev_t rdev;
 	int fd;
@@ -148,6 +218,14 @@ static int add_disk_to_super(int mdfd, struct shape *s, struct context *c,
 	}
 	st->ss->getinfo_super(st, info, NULL);
 
+	if (fd >= 0 && s->write_zeroes) {
+		*zero_pid = write_zeroes_fork(fd, s, st, dv);
+		if (*zero_pid <= 0) {
+			ioctl(mdfd, STOP_ARRAY, NULL);
+			return 1;
+		}
+	}
+
 	if (have_container && c->verbose > 0)
 		pr_err("Using %s for device %d\n",
 		       map_dev(info->disk.major, info->disk.minor, 0),
@@ -224,10 +302,12 @@ static int add_disks(int mdfd, struct mdinfo *info, struct shape *s,
 {
 	struct mddev_dev *moved_disk = NULL;
 	int pass, raid_disk_num, dnum;
+	int zero_pids[total_slots];
 	struct mddev_dev *dv;
 	struct mdinfo *infos;
 	int ret = 0;
 
+	memset(zero_pids, 0, sizeof(zero_pids));
 	infos = xmalloc(sizeof(*infos) * total_slots);
 	enable_fds(total_slots);
 	for (pass = 1; pass <= 2; pass++) {
@@ -261,7 +341,7 @@ static int add_disks(int mdfd, struct mdinfo *info, struct shape *s,
 
 				ret = add_disk_to_super(mdfd, s, c, st, dv,
 						&infos[dnum], have_container,
-						major_num);
+						major_num, &zero_pids[dnum]);
 				if (ret)
 					goto out;
 
@@ -287,6 +367,10 @@ static int add_disks(int mdfd, struct mdinfo *info, struct shape *s,
 		}
 
 		if (pass == 1) {
+			ret = wait_for_zero_forks(zero_pids, total_slots);
+			if (ret)
+				goto out;
+
 			ret = update_metadata(mdfd, s, st, map, info,
 					      chosen_name);
 			if (ret)
@@ -295,6 +379,8 @@ static int add_disks(int mdfd, struct mdinfo *info, struct shape *s,
 	}
 
 out:
+	if (ret)
+		wait_for_zero_forks(zero_pids, total_slots);
 	free(infos);
 	return ret;
 }
diff --git a/ReadMe.c b/ReadMe.c
index 50a5e36d05fc..9424bfc3eeca 100644
--- a/ReadMe.c
+++ b/ReadMe.c
@@ -138,6 +138,7 @@ struct option long_options[] = {
     {"size",	  1, 0, 'z'},
     {"auto",	  1, 0, Auto}, /* also for --assemble */
     {"assume-clean",0,0, AssumeClean },
+    {"write-zeroes",0,0, WriteZeroes },
     {"metadata",  1, 0, 'e'}, /* superblock format */
     {"bitmap",	  1, 0, Bitmap},
     {"bitmap-chunk", 1, 0, BitmapChunk},
@@ -390,6 +391,7 @@ char Help_create[] =
 "  --write-journal=      : Specify journal device for RAID-4/5/6 array\n"
 "  --consistency-policy= : Specify the policy that determines how the array\n"
 "                     -k : maintains consistency in case of unexpected shutdown.\n"
+"  --write-zeroes        : Write zeroes to the disks before creating. This will bypass initial sync.\n"
 "\n"
 ;
 
diff --git a/mdadm.c b/mdadm.c
index 972adb524dfb..141838bd394f 100644
--- a/mdadm.c
+++ b/mdadm.c
@@ -602,6 +602,10 @@ int main(int argc, char *argv[])
 			s.assume_clean = 1;
 			continue;
 
+		case O(CREATE, WriteZeroes):
+			s.write_zeroes = 1;
+			continue;
+
 		case O(GROW,'n'):
 		case O(CREATE,'n'):
 		case O(BUILD,'n'): /* number of raid disks */
@@ -1306,6 +1310,11 @@ int main(int argc, char *argv[])
 		}
 	}
 
+	if (s.write_zeroes && !s.assume_clean) {
+		pr_info("Disk zeroing requested, setting --assume-clean to skip resync\n");
+		s.assume_clean = 1;
+	}
+
 	if (!mode && devs_found) {
 		mode = MISC;
 		devmode = 'Q';
diff --git a/mdadm.h b/mdadm.h
index 18c24915e94c..82e920fb523a 100644
--- a/mdadm.h
+++ b/mdadm.h
@@ -273,6 +273,9 @@ static inline void __put_unaligned32(__u32 val, void *p)
 
 #define ARRAY_SIZE(x) (sizeof(x)/sizeof(x[0]))
 
+#define KIB_TO_BYTES(x)	((x) << 10)
+#define SEC_TO_BYTES(x)	((x) << 9)
+
 extern const char Name[];
 
 struct md_bb_entry {
@@ -433,6 +436,7 @@ extern char Version[], Usage[], Help[], OptionHelp[],
  */
 enum special_options {
 	AssumeClean = 300,
+	WriteZeroes,
 	BitmapChunk,
 	WriteBehind,
 	ReAdd,
@@ -593,6 +597,7 @@ struct shape {
 	int	bitmap_chunk;
 	char	*bitmap_file;
 	int	assume_clean;
+	bool	write_zeroes;
 	int	write_behind;
 	unsigned long long size;
 	unsigned long long data_offset;
-- 
2.30.2

[PATCH mdadm v4 6/7] tests/00raid5-zero: Introduce test to exercise --write-zeros.

[Logan Gunthorpe]

Attempt to create a raid5 array with --write-zeros. If it is successful
check the array to ensure it is in sync.

If it is unsuccessful and an unsupported error is printed, skip the
test.

Signed-off-by: Logan Gunthorpe <logang@deltatee.com>
---
 tests/00raid5-zero | 12 ++++++++++++
 1 file changed, 12 insertions(+)
 create mode 100644 tests/00raid5-zero

diff --git a/tests/00raid5-zero b/tests/00raid5-zero
new file mode 100644
index 000000000000..7d0f05a12539
--- /dev/null
+++ b/tests/00raid5-zero
@@ -0,0 +1,12 @@
+
+if mdadm -CfR $md0 -l 5 -n3 $dev0 $dev1 $dev2 --write-zeroes ; then
+  check nosync
+  echo check > /sys/block/md0/md/sync_action;
+  check wait
+elif grep "zeroing [^ ]* failed: Operation not supported" \
+     $targetdir/stderr; then
+  echo "write-zeros not supported, skipping"
+else
+  echo >&2 "ERROR: mdadm return failure without not supported message"
+  exit 1
+fi
-- 
2.30.2

[PATCH mdadm v4 7/7] manpage: Add --write-zeroes option to manpage

[Logan Gunthorpe]

Document the new --write-zeroes option in the manpage.

Signed-off-by: Logan Gunthorpe <logang@deltatee.com>
---
 mdadm.8.in | 16 ++++++++++++++++
 1 file changed, 16 insertions(+)

diff --git a/mdadm.8.in b/mdadm.8.in
index 70c79d1e6e76..3beb475fd955 100644
--- a/mdadm.8.in
+++ b/mdadm.8.in
@@ -837,6 +837,22 @@ array is resynced at creation.  From Linux version 3.0,
 .B \-\-assume\-clean
 can be used with that command to avoid the automatic resync.
 
+.TP
+.BR \-\-write-zeroes
+When creating an array, send write zeroes requests to all the block
+devices.  This should zero the data area on all disks such that the
+initial sync is not necessary and, if successfull, will behave
+as if
+.B \-\-assume\-clean
+was specified.
+.IP
+This is intended for use with devices that have hardware offload for
+zeroing, but despit this zeroing can still take several minutes for
+large disks.  Thus a message is printed before and after zeroing and
+each disk is zeroed in parallel with the others.
+.IP
+This is only meaningful with --create.
+
 .TP
 .BR \-\-backup\-file=
 This is needed when
-- 
2.30.2

Re: [PATCH mdadm v4 0/7] Write Zeroes option for Creating Arrays

[Xiao Ni]

Hi Logan

I did a test with the patchset. There is a problem like this:

mdadm -CR /dev/md0 -l5 -n3 /dev/nvme1n1 /dev/nvme2n1 /dev/nvme0n1 --write-zero
mdadm: zeroing data from 135266304 to 960061505536 on: /dev/nvme1n1
mdadm: zeroing data from 135266304 to 960061505536 on: /dev/nvme2n1
mdadm: zeroing data from 135266304 to 960061505536 on: /dev/nvme0n1

I ran ctrl+c when waiting, then the raid can't be created anymore. Because the
processes that write zero to nvmes are stuck.

ps auxf | grep mdadm
root       68764  0.0  0.0   9216  1104 pts/0    S+   21:09   0:00
         \_ grep --color=auto mdadm
root       68633  0.1  0.0  27808   336 pts/0    D    21:04   0:00
mdadm -CR /dev/md0 -l5 -n3 /dev/nvme1n1 /dev/nvme2n1 /dev/nvme0n1
--write-zero
root       68634  0.2  0.0  27808   336 pts/0    D    21:04   0:00
mdadm -CR /dev/md0 -l5 -n3 /dev/nvme1n1 /dev/nvme2n1 /dev/nvme0n1
--write-zero
root       68635  0.0  0.0  27808   336 pts/0    D    21:04   0:00
mdadm -CR /dev/md0 -l5 -n3 /dev/nvme1n1 /dev/nvme2n1 /dev/nvme0n1
--write-zero

Regards
Xiao

On Sat, Oct 8, 2022 at 4:10 AM Logan Gunthorpe <logang@deltatee.com> wrote:
>
> Hi,
>
> This is the next iteration of the patchset that added the discard
> option to mdadm. Per feedback from Martin, it's more desirable
> to use the write-zeroes functionality than rely on devices to zero
> the data on a discard request. This is because standards typically
> only require the device to do the best effort to discard data and
> may not actually discard (and thus zero) it all in some circumstances.
>
> This version of the patch set adds the --write-zeroes option which
> will imply --assume-clean and write zeros to the data region in
> each disk before starting the array. This can take some time so
> each disk is done in parallel in its own fork. To make the forking
> code easier to understand this patch set also starts with some
> cleanup of the existing Create code.
>
> We tested write-zeroes requests on a number of modern nvme drives of
> various manufacturers and found most are not as optimized as the
> discard path. A couple drives that were tested did not support
> write-zeroes at all but still performed similarly with the kernel
> falling back to writing zero pages. Typically we see it take on the
> order of one minute per 100GB of data zeroed.
>
> One reason write-zeroes is slower than discard is that today's NVMe
> devices only allow about 2MB to be zeroed in one command where as
> the entire drive can typically be discarded in one command. Partly,
> this is a limitation of the spec as there are only 16 bits avalaible
> in the write-zeros command size but drives still don't max this out.
> Hopefully, in the future this will all be optimized a bit more
> and this work will be able to take advantage of that.
>
> Logan
>
> --
>
> Changes since v3:
>    * Store the pid in a local variable instead of the mdinfo struct
>     (per Mariusz and Xiao)
>
> Changes since v2:
>
>    * Use write-zeroes instead of discard to zero the disks (per
>      Martin)
>    * Due to the time required to zero the disks, each disk is
>      now done in parallel with separate forks of the process.
>    * In order to add the forking some refactoring was done on the
>      Create() function to make it easier to understand
>    * Added a pr_info() call so that some prints can be done
>      to stdout instead of stdour (per Mariusz)
>    * Added KIB_TO_BYTES and SEC_TO_BYTES helpers (per Mariusz)
>    * Added a test to the mdadm test suite to test the option
>      works.
>    * Fixed up how the size and offset are calculated with some
>      great information from Xiao.
>
> Changes since v1:
>
>    * Discard the data in the devices later in the create process
>      while they are already open. This requires treating the
>      s.discard option the same as the s.assume_clean option.
>      Per Mariusz.
>    * A couple other minor cleanup changes from Mariusz.
>
>
> *** BLURB HERE ***
>
> Logan Gunthorpe (7):
>   Create: goto abort_locked instead of return 1 in error path
>   Create: remove safe_mode_delay local variable
>   Create: Factor out add_disks() helpers
>   mdadm: Introduce pr_info()
>   mdadm: Add --write-zeros option for Create
>   tests/00raid5-zero: Introduce test to exercise --write-zeros.
>   manpage: Add --write-zeroes option to manpage
>
>  Create.c           | 479 ++++++++++++++++++++++++++++-----------------
>  ReadMe.c           |   2 +
>  mdadm.8.in         |  16 ++
>  mdadm.c            |   9 +
>  mdadm.h            |   7 +
>  tests/00raid5-zero |  12 ++
>  6 files changed, 350 insertions(+), 175 deletions(-)
>  create mode 100644 tests/00raid5-zero
>
>
> base-commit: 8b668d4aa3305af5963162b7499b128bd71f8f29
> --
> 2.30.2
>

Re: [PATCH mdadm v4 0/7] Write Zeroes option for Creating Arrays

[Logan Gunthorpe]

@ccing Martin hoping he has an opinion on the write zeroes interface

On 2022-10-11 19:09, Xiao Ni wrote:
> Hi Logan
> 
> I did a test with the patchset. There is a problem like this:
> 
> mdadm -CR /dev/md0 -l5 -n3 /dev/nvme1n1 /dev/nvme2n1 /dev/nvme0n1 --write-zero
> mdadm: zeroing data from 135266304 to 960061505536 on: /dev/nvme1n1
> mdadm: zeroing data from 135266304 to 960061505536 on: /dev/nvme2n1
> mdadm: zeroing data from 135266304 to 960061505536 on: /dev/nvme0n1
> 
> I ran ctrl+c when waiting, then the raid can't be created anymore. Because the
> processes that write zero to nvmes are stuck.
> 
> ps auxf | grep mdadm
> root       68764  0.0  0.0   9216  1104 pts/0    S+   21:09   0:00
>          \_ grep --color=auto mdadm
> root       68633  0.1  0.0  27808   336 pts/0    D    21:04   0:00
> mdadm -CR /dev/md0 -l5 -n3 /dev/nvme1n1 /dev/nvme2n1 /dev/nvme0n1
> --write-zero
> root       68634  0.2  0.0  27808   336 pts/0    D    21:04   0:00
> mdadm -CR /dev/md0 -l5 -n3 /dev/nvme1n1 /dev/nvme2n1 /dev/nvme0n1
> --write-zero
> root       68635  0.0  0.0  27808   336 pts/0    D    21:04   0:00
> mdadm -CR /dev/md0 -l5 -n3 /dev/nvme1n1 /dev/nvme2n1 /dev/nvme0n1
> --write-zero

Yes, this is because the fallocate() call that the child processes use
to write_zeros will submit a large number of bios in the kernel and then
wait with submit_bio_wait() which is non-interruptible. So when the
child processes get the SIGINT, they will not stop until after the
fallocate() call completes which will pretty much be after the entire
disk is zeroed. So if you are zeroing a very large disk, those processes
will be stuck around for several minutes after the parent process
terminates; though they do go away eventually.

There aren't many great solutions for this:

1) We could install as signal handler in the parent so it sticks around
until the zeroing is complete. This would mean mdadm will not be able to
be terminated while the zeroing is occurring and the user has to wait.

2) We could split up the fallocate call into multiple calls to zero the
entire disk. This would allow a quicker ctrl-c to occur, however it's
not clear what the best size would be to split it into. Even zeroing 1GB
can take a few seconds, but the smaller we go, the less efficient it
will be if the block layer and devices ever get write-zeroes optimized
in the same way discard has been optimized (with NVMe, discard only
requires a single command to handle the entire disk where as
write-zeroes requires a minimum of one command per 2MB of data to zero).
I was hoping write-zeroes could be made faster in the future, at least
for NVMe.

Thoughts?

Logan

Re: [PATCH mdadm v4 0/7] Write Zeroes option for Creating Arrays

[Martin K. Petersen]

Logan,

> 2) We could split up the fallocate call into multiple calls to zero
> the entire disk. This would allow a quicker ctrl-c to occur, however
> it's not clear what the best size would be to split it into. Even
> zeroing 1GB can take a few seconds,

FWIW, we default to 32MB per request in SCSI unless the device
explicitly advertises wanting something larger.

> (with NVMe, discard only requires a single command to handle the
> entire disk

In NVMe there's a limit of 64K blocks per range and 256 ranges per
request. So 8GB or 64GB per request for discard depending on the block
size. So presumably it will take several operations to deallocate an
entire drive.

> where as write-zeroes requires a minimum of one command per 2MB of
> data to zero).

32MB for 512-byte blocks and 256MB for 4096-byte blocks. Which matches
how it currently works for SCSI devices.

> I was hoping write-zeroes could be made faster in the future, at least
> for NVMe.

Deallocate had a bit of a head start and vendors are still catching up
in the zeroing department. Some drives do support using Deallocate for
zeroing and we quirk those in the driver so they should perform OK with
your change.

-- 
Martin K. Petersen	Oracle Linux Engineering

Re: [PATCH mdadm v4 0/7] Write Zeroes option for Creating Arrays

[Xiao Ni]

On Thu, Oct 13, 2022 at 9:34 AM Martin K. Petersen
<martin.petersen@oracle.com> wrote:
>
>
> Logan,
>
> > 2) We could split up the fallocate call into multiple calls to zero
> > the entire disk. This would allow a quicker ctrl-c to occur, however
> > it's not clear what the best size would be to split it into. Even
> > zeroing 1GB can take a few seconds,
>
> FWIW, we default to 32MB per request in SCSI unless the device
> explicitly advertises wanting something larger.

Hi Martin

If the upper layer submit 1GB request, SCSI will split them and handle 32MB
requests in default. If the upper layer wants SCSI to handle 1GB one time, it
needs to pass some information to SCSI, right?

>
> > (with NVMe, discard only requires a single command to handle the
> > entire disk
>
> In NVMe there's a limit of 64K blocks per range and 256 ranges per
> request. So 8GB or 64GB per request for discard depending on the block
> size. So presumably it will take several operations to deallocate an
> entire drive.

Could you tell the command how to check the block size.
blockdev --getsz tells the sector size of the device. It should not be the
block size you mentioned here.

It looks like we can decide the best size once the blocksize is found.

Regards
Xiao

>
> > where as write-zeroes requires a minimum of one command per 2MB of
> > data to zero).
>
> 32MB for 512-byte blocks and 256MB for 4096-byte blocks. Which matches
> how it currently works for SCSI devices.
>
> > I was hoping write-zeroes could be made faster in the future, at least
> > for NVMe.
>
> Deallocate had a bit of a head start and vendors are still catching up
> in the zeroing department. Some drives do support using Deallocate for
> zeroing and we quirk those in the driver so they should perform OK with
> your change.
>
> --
> Martin K. Petersen      Oracle Linux Engineering
>

Re: [PATCH mdadm v4 0/7] Write Zeroes option for Creating Arrays

[Martin K. Petersen]

Xiao,

> If the upper layer submit 1GB request, SCSI will split them and handle
> 32MB requests in default. If the upper layer wants SCSI to handle 1GB
> one time, it needs to pass some information to SCSI, right?

It is the device that defines its capabilities, not Linux. If the device
states it wants 32MB per request, then that's what we'll issue.

>> In NVMe there's a limit of 64K blocks per range and 256 ranges per
>> request. So 8GB or 64GB per request for discard depending on the block
>> size. So presumably it will take several operations to deallocate an
>> entire drive.
>
> Could you tell the command how to check the block size.
> blockdev --getsz tells the sector size of the device. It should not be the
> block size you mentioned here.

blockdev --getss will tell you the device's logical block size.

-- 
Martin K. Petersen	Oracle Linux Engineering

Re: [PATCH mdadm v4 0/7] Write Zeroes option for Creating Arrays

[Xiao Ni]

On Wed, Oct 26, 2022 at 10:42 AM Martin K. Petersen
<martin.petersen@oracle.com> wrote:
>
>
> Xiao,
>
> > If the upper layer submit 1GB request, SCSI will split them and handle
> > 32MB requests in default. If the upper layer wants SCSI to handle 1GB
> > one time, it needs to pass some information to SCSI, right?
>
> It is the device that defines its capabilities, not Linux. If the device
> states it wants 32MB per request, then that's what we'll issue.
>
> >> In NVMe there's a limit of 64K blocks per range and 256 ranges per
> >> request. So 8GB or 64GB per request for discard depending on the block
> >> size. So presumably it will take several operations to deallocate an
> >> entire drive.
> >
> > Could you tell the command how to check the block size.
> > blockdev --getsz tells the sector size of the device. It should not be the
> > block size you mentioned here.
>
> blockdev --getss will tell you the device's logical block size.
>
> --
> Martin K. Petersen      Oracle Linux Engineering
>

Hi Martin

Thanks for all the explanations :)

Regards
Xiao

Re: [PATCH mdadm v4 0/7] Write Zeroes option for Creating Arrays

[Logan Gunthorpe]

Sorry a little late responding to this.

On 2022-10-12 19:33, Martin K. Petersen wrote:
> 
> Logan,
> 
>> 2) We could split up the fallocate call into multiple calls to zero
>> the entire disk. This would allow a quicker ctrl-c to occur, however
>> it's not clear what the best size would be to split it into. Even
>> zeroing 1GB can take a few seconds,
> 
> FWIW, we default to 32MB per request in SCSI unless the device
> explicitly advertises wanting something larger.
> 
>> (with NVMe, discard only requires a single command to handle the
>> entire disk
> 
> In NVMe there's a limit of 64K blocks per range and 256 ranges per
> request. So 8GB or 64GB per request for discard depending on the block
> size. So presumably it will take several operations to deallocate an
> entire drive.
> 
>> where as write-zeroes requires a minimum of one command per 2MB of
>> data to zero).
> 
> 32MB for 512-byte blocks and 256MB for 4096-byte blocks. Which matches
> how it currently works for SCSI devices.

The 2MB I was referring to was the typical maximum we see on real
devices. We tested a number of NVMe drives from a number of different
vendors and found most to be a maximum of 2MB, some devices had 512KB.
Which is unfortunate.

>> I was hoping write-zeroes could be made faster in the future, at least
>> for NVMe.
> 
> Deallocate had a bit of a head start and vendors are still catching up
> in the zeroing department. Some drives do support using Deallocate for
> zeroing and we quirk those in the driver so they should perform OK with
> your change.

Yeah, my hope is that larger zeroing requests can be supported which
will be handled performantly by deallocating the device. So I don't want
mdadm to slow this down by splitting the request to the kernel into a
number of smaller requests. But this seems to be the only way forward
because the request is uninterruptible and we don't want to hang the
user for several minutes.

Logan

Re: [PATCH mdadm v4 0/7] Write Zeroes option for Creating Arrays

[Kinga Tanska]

On Fri,  7 Oct 2022 14:10:30 -0600
Logan Gunthorpe <logang@deltatee.com> wrote:

> Hi,
> 
> This is the next iteration of the patchset that added the discard
> option to mdadm. Per feedback from Martin, it's more desirable
> to use the write-zeroes functionality than rely on devices to zero
> the data on a discard request. This is because standards typically
> only require the device to do the best effort to discard data and
> may not actually discard (and thus zero) it all in some circumstances.
> 
> This version of the patch set adds the --write-zeroes option which
> will imply --assume-clean and write zeros to the data region in
> each disk before starting the array. This can take some time so
> each disk is done in parallel in its own fork. To make the forking
> code easier to understand this patch set also starts with some
> cleanup of the existing Create code.
> 
> We tested write-zeroes requests on a number of modern nvme drives of
> various manufacturers and found most are not as optimized as the
> discard path. A couple drives that were tested did not support
> write-zeroes at all but still performed similarly with the kernel
> falling back to writing zero pages. Typically we see it take on the
> order of one minute per 100GB of data zeroed.
> 
> One reason write-zeroes is slower than discard is that today's NVMe
> devices only allow about 2MB to be zeroed in one command where as
> the entire drive can typically be discarded in one command. Partly,
> this is a limitation of the spec as there are only 16 bits avalaible
> in the write-zeros command size but drives still don't max this out.
> Hopefully, in the future this will all be optimized a bit more
> and this work will be able to take advantage of that.
> 
> Logan
> 
> --
> 
> Changes since v3:
>    * Store the pid in a local variable instead of the mdinfo struct
>     (per Mariusz and Xiao)
> 
> Changes since v2:
> 
>    * Use write-zeroes instead of discard to zero the disks (per
>      Martin)
>    * Due to the time required to zero the disks, each disk is
>      now done in parallel with separate forks of the process.
>    * In order to add the forking some refactoring was done on the
>      Create() function to make it easier to understand
>    * Added a pr_info() call so that some prints can be done
>      to stdout instead of stdour (per Mariusz)
>    * Added KIB_TO_BYTES and SEC_TO_BYTES helpers (per Mariusz)
>    * Added a test to the mdadm test suite to test the option
>      works.
>    * Fixed up how the size and offset are calculated with some
>      great information from Xiao.
> 
> Changes since v1:
> 
>    * Discard the data in the devices later in the create process
>      while they are already open. This requires treating the
>      s.discard option the same as the s.assume_clean option.
>      Per Mariusz.
>    * A couple other minor cleanup changes from Mariusz.
> 
> 
> *** BLURB HERE ***
> 
> Logan Gunthorpe (7):
>   Create: goto abort_locked instead of return 1 in error path
>   Create: remove safe_mode_delay local variable
>   Create: Factor out add_disks() helpers
>   mdadm: Introduce pr_info()
>   mdadm: Add --write-zeros option for Create
>   tests/00raid5-zero: Introduce test to exercise --write-zeros.
>   manpage: Add --write-zeroes option to manpage
> 
>  Create.c           | 479
> ++++++++++++++++++++++++++++----------------- ReadMe.c           |
> 2 + mdadm.8.in         |  16 ++
>  mdadm.c            |   9 +
>  mdadm.h            |   7 +
>  tests/00raid5-zero |  12 ++
>  6 files changed, 350 insertions(+), 175 deletions(-)
>  create mode 100644 tests/00raid5-zero
> 
> 
> base-commit: 8b668d4aa3305af5963162b7499b128bd71f8f29
> --
> 2.30.2

Acked-by: Kinga Tanska <kinga.tanska@linux.intel.com>