Skip to contentSkip to conversation

mkfs: set rtstart from user-specified dblocks

PATCH In review Darrick J. Wong 2 participants 3 messages Jan 6 – Jan 7

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 mkfs: set rtstart from user-specified dblocks

From: Darrick J. Wong <[email protected]>

generic/211 fails to format the disk on a system with an internal zoned device. Poking through the shell scripts, it's apparently doing this:

# mkfs.xfs -d size=629145600 -r size=629145600 -b size=4096 -m metadir=1,autofsck=1,uquota,gquota,pquota, -r zoned=1 -d rtinherit=1 /dev/sdd
size 629145600 specified for data subvolume is too large, maximum is 131072 blocks

Strange -- we asked for 629M data and rt sections, the device is 20GB in size, but it claims insufficient space in the data subvolume.

Further analysis shows that open_devices is setting rtstart to 1% of the size of the data volume (or no less than 300M) and rounding that up to the nearest power of two (512M). Hence the 131072 number.

But wait, we said that we wanted a 629M data section. Let's set rtstart to the same value if the user didn't already provide one, instead of using the default value.

Cc <[email protected]> # v6.15.0
Fixes 2e5a737a61d34e ("xfs_mkfs: support creating file system with zoned RT devices")
---
 mkfs/xfs_mkfs.c |   31 +++++++++++++++++++++----------
 1 file changed, 21 insertions(+), 10 deletions(-)
mkfs/xfs_mkfs.c +21 -10
index b34407725f76df..ab3d74790bbcb8 100644
@@ -3720,17 +3720,28 @@ open_devices(
 		zt->rt.zone_capacity = zt->data.zone_capacity;
 		zt->rt.nr_zones = zt->data.nr_zones - zt->data.nr_conv_zones;
 	} else if (cfg->sb_feat.zoned && !cfg->rtstart && !xi->rt.dev) {
-		/*
-		 * By default reserve at 1% of the total capacity (rounded up to
-		 * the next power of two) for metadata, but match the minimum we
-		 * enforce elsewhere. This matches what SMR HDDs provide.
-		 */
-		uint64_t rt_target_size = max((xi->data.size + 99) / 100,
-					      BTOBB(300 * 1024 * 1024));
+		if (cfg->dblocks) {
+			/*
+			 * If the user specified the size of the data device
+			 * but not the start of the internal rt device, set
+			 * the internal rt volume to start at the end of the
+			 * data device.
+			 */
+			cfg->rtstart = cfg->dblocks << (cfg->blocklog - BBSHIFT);
+		} else {
+			/*
+			 * By default reserve at 1% of the total capacity
+			 * (rounded up to the next power of two) for metadata,
+			 * but match the minimum we enforce elsewhere. This
+			 * matches what SMR HDDs provide.
+			 */
+			uint64_t rt_target_size = max((xi->data.size + 99) / 100,
+						      BTOBB(300 * 1024 * 1024));
 
-		cfg->rtstart = 1;
-		while (cfg->rtstart < rt_target_size)
-			cfg->rtstart <<= 1;
+			cfg->rtstart = 1;
+			while (cfg->rtstart < rt_target_size)
+				cfg->rtstart <<= 1;
+		}
 	}
 
 	if (cfg->rtstart) {
#1 lore raw
Quoted 4 lines · Darrick J. Wong

From: Darrick J. Wong <[email protected]>

generic/211 fails to format the disk on a system with an internal zoned device. Poking through the shell scripts, it's apparently doing this:

On some anyway. We've had this on our CI machines as well, and it's been on my todo list why it doesn't happen everywhere, so thanks for digging into this and getting it off my todo list :)

Quoted 34 lines
# mkfs.xfs -d size=629145600 -r size=629145600 -b size=4096 -m metadir=1,autofsck=1,uquota,gquota,pquota, -r zoned=1 -d rtinherit=1 /dev/sdd
size 629145600 specified for data subvolume is too large, maximum is 131072 blocks

Strange -- we asked for 629M data and rt sections, the device is 20GB in size, but it claims insufficient space in the data subvolume.

Further analysis shows that open_devices is setting rtstart to 1% of the size of the data volume (or no less than 300M) and rounding that up to the nearest power of two (512M). Hence the 131072 number.

But wait, we said that we wanted a 629M data section. Let's set rtstart to the same value if the user didn't already provide one, instead of using the default value.

Cc <[email protected]> # v6.15.0
Fixes 2e5a737a61d34e ("xfs_mkfs: support creating file system with zoned RT devices")
---
 mkfs/xfs_mkfs.c |   31 +++++++++++++++++++++----------
 1 file changed, 21 insertions(+), 10 deletions(-)
mkfs/xfs_mkfs.c +0 -5
index b34407725f76df..ab3d74790bbcb8 100644
@@ -3720,17 +3720,28 @@ open_devices(
 		zt->rt.zone_capacity = zt->data.zone_capacity;
 		zt->rt.nr_zones = zt->data.nr_zones - zt->data.nr_conv_zones;
 	} else if (cfg->sb_feat.zoned && !cfg->rtstart && !xi->rt.dev) {
-		/*
-		 * By default reserve at 1% of the total capacity (rounded up to
-		 * the next power of two) for metadata, but match the minimum we
-		 * enforce elsewhere. This matches what SMR HDDs provide.
-		 */

mkfs/xfs_mkfs.c open_devices( · line 3730

Quoted 10 lines
mkfs/xfs_mkfs.c +8 -2
-		uint64_t rt_target_size = max((xi->data.size + 99) / 100,
-					      BTOBB(300 * 1024 * 1024));
+		if (cfg->dblocks) {
+			/*
+			 * If the user specified the size of the data device
+			 * but not the start of the internal rt device, set
+			 * the internal rt volume to start at the end of the
+			 * data device.
+			 */
+			cfg->rtstart = cfg->dblocks << (cfg->blocklog - BBSHIFT);

Overly long line here.

mkfs/xfs_mkfs.c open_devices( · line 3739

Quoted 9 lines
mkfs/xfs_mkfs.c +9 -0
+		} else {
+			/*
+			 * By default reserve at 1% of the total capacity
+			 * (rounded up to the next power of two) for metadata,
+			 * but match the minimum we enforce elsewhere. This
+			 * matches what SMR HDDs provide.
+			 */
+			uint64_t rt_target_size = max((xi->data.size + 99) / 100,
+						      BTOBB(300 * 1024 * 1024));

And here as well. Maybe split the sizse calculation into a helper to make these a bit more readable?

Otherwise this looks good.

#2 lore raw
Quoted 57 lines · Christoph Hellwig
Quoted 4 lines · Darrick J. Wong

From: Darrick J. Wong <[email protected]>

generic/211 fails to format the disk on a system with an internal zoned device. Poking through the shell scripts, it's apparently doing this:

On some anyway. We've had this on our CI machines as well, and it's been on my todo list why it doesn't happen everywhere, so thanks for digging into this and getting it off my todo list :)

Quoted 34 lines
# mkfs.xfs -d size=629145600 -r size=629145600 -b size=4096 -m metadir=1,autofsck=1,uquota,gquota,pquota, -r zoned=1 -d rtinherit=1 /dev/sdd
size 629145600 specified for data subvolume is too large, maximum is 131072 blocks

Strange -- we asked for 629M data and rt sections, the device is 20GB in size, but it claims insufficient space in the data subvolume.

Further analysis shows that open_devices is setting rtstart to 1% of the size of the data volume (or no less than 300M) and rounding that up to the nearest power of two (512M). Hence the 131072 number.

But wait, we said that we wanted a 629M data section. Let's set rtstart to the same value if the user didn't already provide one, instead of using the default value.

Cc <[email protected]> # v6.15.0
Fixes 2e5a737a61d34e ("xfs_mkfs: support creating file system with zoned RT devices")
---
 mkfs/xfs_mkfs.c |   31 +++++++++++++++++++++----------
 1 file changed, 21 insertions(+), 10 deletions(-)
mkfs/xfs_mkfs.c +0 -5
index b34407725f76df..ab3d74790bbcb8 100644
@@ -3720,17 +3720,28 @@ open_devices(
 		zt->rt.zone_capacity = zt->data.zone_capacity;
 		zt->rt.nr_zones = zt->data.nr_zones - zt->data.nr_conv_zones;
 	} else if (cfg->sb_feat.zoned && !cfg->rtstart && !xi->rt.dev) {
-		/*
-		 * By default reserve at 1% of the total capacity (rounded up to
-		 * the next power of two) for metadata, but match the minimum we
-		 * enforce elsewhere. This matches what SMR HDDs provide.
-		 */

mkfs/xfs_mkfs.c open_devices( · line 3730

Quoted 10 lines
mkfs/xfs_mkfs.c +8 -2
-		uint64_t rt_target_size = max((xi->data.size + 99) / 100,
-					      BTOBB(300 * 1024 * 1024));
+		if (cfg->dblocks) {
+			/*
+			 * If the user specified the size of the data device
+			 * but not the start of the internal rt device, set
+			 * the internal rt volume to start at the end of the
+			 * data device.
+			 */
+			cfg->rtstart = cfg->dblocks << (cfg->blocklog - BBSHIFT);

Overly long line here.

Quoted 12 lines

mkfs/xfs_mkfs.c open_devices( · line 3739

Quoted 9 lines
mkfs/xfs_mkfs.c +9 -0
+		} else {
+			/*
+			 * By default reserve at 1% of the total capacity
+			 * (rounded up to the next power of two) for metadata,
+			 * but match the minimum we enforce elsewhere. This
+			 * matches what SMR HDDs provide.
+			 */
+			uint64_t rt_target_size = max((xi->data.size + 99) / 100,
+						      BTOBB(300 * 1024 * 1024));

And here as well. Maybe split the sizse calculation into a helper to make these a bit more readable?

Fixed:

static uint64_t
calc_rtstart(
const struct mkfs_params *cfg,
const struct libxfs_init *xi)

{
	uint64_t			rt_target_size;
	uint64_t			rtstart = 1;

	if (cfg->dblocks) {
		/*
		 * If the user specified the size of the data device but not
		 * the start of the internal rt device, set the internal rt
		 * volume to start at the end of the data device.
		 */
		return cfg->dblocks << (cfg->blocklog - BBSHIFT);
	}

	/*
	 * By default reserve at 1% of the total capacity (rounded up to the
	 * next power of two) for metadata, but match the minimum we enforce
	 * elsewhere. This matches what SMR HDDs provide.
	 */
	rt_target_size = max((xi->data.size + 99) / 100,
			     BTOBB(300 * 1024 * 1024));

	while (rtstart < rt_target_size)
		rtstart <<= 1;
	return rtstart;
}

and then the callsite turns into cfg->rtstart = calc_rtstart(...).

Quoted 1 line

Otherwise this looks good.

Thanks!

--D