Skip to contentSkip to conversation

fstests: fixes for 64k pages and dax

PATCH Superseded Jeff Moyer 3 participants 8 messages Feb 5, 2020 – Feb 6, 2020 VFS

v1 of 2 Superseded Latest is v2 →

lore
Share & tools
lore

Jeff Moyer

author
#0 lore raw

cover fstests: fixes for 64k pages and dax

This set of patches fixes a few false positives I encountered when testing DAX on ppc64le (which has a 64k page size).

Patch 1 is actually not specific to non-4k page sizes. Right now we only test for dax incompatibility in the dm flakey target. This means that tests that use dm-thin or the snapshot target will still try to run. Moving the check to _require_dm_target fixes that problem.

Patches 2 and 3 get rid of hard coded block/page sizes in the tests. They run just fine on 64k pages and 64k block sizes.

Even after these patches, there are many more tests that fail in the following configuration:

MKFS_OPTIONS="-b size=65536 -m reflink=0" MOUNT_OPTIONS="-o dax"

One class of failures is tests that create a really small file system size. Some of those tests seem to require the very small size, but others seem like they could live with a slightly bigger size that would then fit the log (the typical failure is a mkfs failure due to not enough blocks for the log). For the former case, I'm tempted to send patches to _notrun those tests, and for the latter, I'd like to bump the file system sizes up. 300MB seems to be large enough to accommodate the log. Would folks be opposed to those approaches?

Another class of failure is tests that either hard-code a block size to trigger a specific error case, or that test a multitude of block sizes. I'd like to send a patch to _notrun those tests if there is a user-specified block size. That will require parsing the MKFS_OPTIONS based on the fs type, of course. Is that something that seems reasonable?

I will follow up with a series of patches to implement those changes if there is consensus on the approach. These first three seemed straight-forward to me, so that's where I'm starting.

Thanks!
Jeff

[PATCH 1/3] dax/dm: disable testing on devices that don't support dax [PATCH 2/3] t_mmap_collision: fix hard-coded page size [PATCH 3/3] xfs/300: modify test to work on any fs block size

author
#1 lore raw

PATCH 1/3 dax/dm: disable testing on devices that don't support dax

Move the hack out of dmflakey and put it into _require_dm_target. This fixes up a lot of missed tests that are failing due to the lack of dax support (such as tests on dm-thin, snapshot, etc).

---
 common/dmflakey |  5 -----
 common/rc       | 11 +++++++++++
 2 files changed, 11 insertions(+), 5 deletions(-)
common/dmflakey +0 -5
index 2af3924d..b4e11ae9 100644
@@ -8,11 +8,6 @@ FLAKEY_ALLOW_WRITES=0
 FLAKEY_DROP_WRITES=1
 FLAKEY_ERROR_WRITES=2
 
-echo $MOUNT_OPTIONS | grep -q dax
-if [ $? -eq 0 ]; then
-	_notrun "Cannot run tests with DAX on dmflakey devices"
-fi
-
 _init_flakey()
 {
 	local BLK_DEV_SIZE=`blockdev --getsz $SCRATCH_DEV`
common/rc +11 -0
index eeac1355..785f34c6 100644
@@ -1874,6 +1874,17 @@ _require_dm_target()
 	_require_sane_bdev_flush $SCRATCH_DEV
 	_require_command "$DMSETUP_PROG" dmsetup
 
+	echo $MOUNT_OPTIONS | grep -q dax
+	if [ $? -eq 0 ]; then
+		case $target in
+		stripe|linear|error)
+			;;
+		*)
+			_notrun "Cannot run tests with DAX on $target devices."
+			;;
+		esac
+	fi
+
 	modprobe dm-$target >/dev/null 2>&1
 
 	$DMSETUP_PROG targets 2>&1 | grep -q ^$target
Signature · 1 line
2.19.1
#2 lore raw

Re: [PATCH 1/3] dax/dm: disable testing on devices that don't support dax

Quoted 15 lines · Jeff Moyer

Move the hack out of dmflakey and put it into _require_dm_target. This fixes up a lot of missed tests that are failing due to the lack of dax support (such as tests on dm-thin, snapshot, etc).

---
 common/dmflakey |  5 -----
 common/rc       | 11 +++++++++++
 2 files changed, 11 insertions(+), 5 deletions(-)
common/dmflakey +0 -0
index 2af3924d..b4e11ae9 100644

common/dmflakey FLAKEY_ALLOW_WRITES=0 · line 14

Quoted 7 lines
common/dmflakey +0 -4
@@ -8,11 +8,6 @@ FLAKEY_ALLOW_WRITES=0
 FLAKEY_DROP_WRITES=1
 FLAKEY_ERROR_WRITES=2
 
-echo $MOUNT_OPTIONS | grep -q dax
-if [ $? -eq 0 ]; then
-	_notrun "Cannot run tests with DAX on dmflakey devices"
-fi

If we need to remove this for common/dmflakey, why not do the same thing in common/dmthin and common/dmdelay etc ?

Quoted 29 lines
common/dmflakey +0 -1
-
 _init_flakey()
 {
 	local BLK_DEV_SIZE=`blockdev --getsz $SCRATCH_DEV`
common/rc +11 -0
index eeac1355..785f34c6 100644
@@ -1874,6 +1874,17 @@ _require_dm_target()
 	_require_sane_bdev_flush $SCRATCH_DEV
 	_require_command "$DMSETUP_PROG" dmsetup
 
+	echo $MOUNT_OPTIONS | grep -q dax
+	if [ $? -eq 0 ]; then
+		case $target in
+		stripe|linear|error)
+			;;
+		*)
+			_notrun "Cannot run tests with DAX on $target devices."
+			;;
+		esac
+	fi
+
 	modprobe dm-$target >/dev/null 2>&1
 
 	$DMSETUP_PROG targets 2>&1 | grep -q ^$target
Signature · 1 line
2.19.1
author
#3 lore raw

Re: [PATCH 1/3] dax/dm: disable testing on devices that don't support dax

Quoted 19 lines · Zorro Lang
Quoted 18 lines · Jeff Moyer

Move the hack out of dmflakey and put it into _require_dm_target. This fixes up a lot of missed tests that are failing due to the lack of dax support (such as tests on dm-thin, snapshot, etc).

---
 common/dmflakey |  5 -----
 common/rc       | 11 +++++++++++
 2 files changed, 11 insertions(+), 5 deletions(-)
common/dmflakey +0 -0
index 2af3924d..b4e11ae9 100644
@@ -8,11 +8,6 @@ FLAKEY_ALLOW_WRITES=0
 FLAKEY_DROP_WRITES=1
 FLAKEY_ERROR_WRITES=2
 
Quoted 7 lines

common/dmflakey FLAKEY_ALLOW_WRITES=0 · line 14

Quoted 4 lines
common/dmflakey +0 -4
-echo $MOUNT_OPTIONS | grep -q dax
-if [ $? -eq 0 ]; then
-	_notrun "Cannot run tests with DAX on dmflakey devices"
-fi

If we need to remove this for common/dmflakey, why not do the same thing in common/dmthin and common/dmdelay etc ?

I didn't realize they had this same code. I'll make that change, thanks!

-Jeff

Quoted 30 lines
Quoted 30 lines
common/dmflakey +0 -1
-
 _init_flakey()
 {
 	local BLK_DEV_SIZE=`blockdev --getsz $SCRATCH_DEV`
common/rc +11 -0
index eeac1355..785f34c6 100644
@@ -1874,6 +1874,17 @@ _require_dm_target()
 	_require_sane_bdev_flush $SCRATCH_DEV
 	_require_command "$DMSETUP_PROG" dmsetup
 
+	echo $MOUNT_OPTIONS | grep -q dax
+	if [ $? -eq 0 ]; then
+		case $target in
+		stripe|linear|error)
+			;;
+		*)
+			_notrun "Cannot run tests with DAX on $target devices."
+			;;
+		esac
+	fi
+
 	modprobe dm-$target >/dev/null 2>&1
 
 	$DMSETUP_PROG targets 2>&1 | grep -q ^$target
Signature · 1 line
2.19.1
author
#4 lore raw

PATCH 2/3 t_mmap_collision: fix hard-coded page size

In reply to Jeff Moyer

Fix the test to run on non-4k page size systems.

---
 src/t_mmap_collision.c | 40 +++++++++++++++++++++-------------------
 1 file changed, 21 insertions(+), 19 deletions(-)
src/t_mmap_collision.c +21 -19
index d547bc05..c872f4e2 100644
@@ -25,13 +25,12 @@
 #include <sys/types.h>
 #include <unistd.h>
 
-#define PAGE(a) ((a)*0x1000)
-#define FILE_SIZE PAGE(4)
-
 void *dax_data;
 int nodax_fd;
 int dax_fd;
 bool done;
+static int pagesize;
+static int file_size;
 
 #define err_exit(op)                                                          \
 {                                                                             \
@@ -49,18 +48,18 @@ void punch_hole_fn(void *ptr)
 		read = 0;
 
 		do {
-			rc = pread(nodax_fd, dax_data + read, FILE_SIZE - read,
+			rc = pread(nodax_fd, dax_data + read, file_size - read,
 					read);
 			if (rc > 0)
 				read += rc;
 		} while (rc > 0);
 
-		if (read != FILE_SIZE || rc != 0)
+		if (read != file_size || rc != 0)
 			err_exit("pread");
 
 		rc = fallocate(dax_fd,
 				FALLOC_FL_PUNCH_HOLE | FALLOC_FL_KEEP_SIZE,
-				0, FILE_SIZE);
+				0, file_size);
 		if (rc < 0)
 			err_exit("fallocate");
 
@@ -81,18 +80,18 @@ void zero_range_fn(void *ptr)
 		read = 0;
 
 		do {
-			rc = pread(nodax_fd, dax_data + read, FILE_SIZE - read,
+			rc = pread(nodax_fd, dax_data + read, file_size - read,
 					read);
 			if (rc > 0)
 				read += rc;
 		} while (rc > 0);
 
-		if (read != FILE_SIZE || rc != 0)
+		if (read != file_size || rc != 0)
 			err_exit("pread");
 
 		rc = fallocate(dax_fd,
 				FALLOC_FL_ZERO_RANGE | FALLOC_FL_KEEP_SIZE,
-				0, FILE_SIZE);
+				0, file_size);
 		if (rc < 0)
 			err_exit("fallocate");
 
@@ -113,11 +112,11 @@ void truncate_down_fn(void *ptr)
 
 		if (ftruncate(dax_fd, 0) < 0)
 			err_exit("ftruncate");
-		if (fallocate(dax_fd, 0, 0, FILE_SIZE) < 0)
+		if (fallocate(dax_fd, 0, 0, file_size) < 0)
 			err_exit("fallocate");
 
 		do {
-			rc = pread(nodax_fd, dax_data + read, FILE_SIZE - read,
+			rc = pread(nodax_fd, dax_data + read, file_size - read,
 					read);
 			if (rc > 0)
 				read += rc;
@@ -142,15 +141,15 @@ void collapse_range_fn(void *ptr)
 	while (!done) {
 		read = 0;
 
-		if (fallocate(dax_fd, 0, 0, FILE_SIZE) < 0)
+		if (fallocate(dax_fd, 0, 0, file_size) < 0)
 			err_exit("fallocate 1");
-		if (fallocate(dax_fd, FALLOC_FL_COLLAPSE_RANGE, 0, PAGE(1)) < 0)
+		if (fallocate(dax_fd, FALLOC_FL_COLLAPSE_RANGE, 0, pagesize) < 0)
 			err_exit("fallocate 2");
-		if (fallocate(dax_fd, 0, 0, FILE_SIZE) < 0)
+		if (fallocate(dax_fd, 0, 0, file_size) < 0)
 			err_exit("fallocate 3");
 
 		do {
-			rc = pread(nodax_fd, dax_data + read, FILE_SIZE - read,
+			rc = pread(nodax_fd, dax_data + read, file_size - read,
 					read);
 			if (rc > 0)
 				read += rc;
@@ -192,6 +191,9 @@ int main(int argc, char *argv[])
 		exit(0);
 	}
 
+	pagesize = getpagesize();
+	file_size = 4 * pagesize;
+
 	dax_fd = open(argv[1], O_RDWR|O_CREAT, S_IRUSR|S_IWUSR);
 	if (dax_fd < 0)
 		err_exit("dax_fd open");
@@ -202,15 +204,15 @@ int main(int argc, char *argv[])
 
 	if (ftruncate(dax_fd, 0) < 0)
 		err_exit("dax_fd ftruncate");
-	if (fallocate(dax_fd, 0, 0, FILE_SIZE) < 0)
+	if (fallocate(dax_fd, 0, 0, file_size) < 0)
 		err_exit("dax_fd fallocate");
 
 	if (ftruncate(nodax_fd, 0) < 0)
 		err_exit("nodax_fd ftruncate");
-	if (fallocate(nodax_fd, 0, 0, FILE_SIZE) < 0)
+	if (fallocate(nodax_fd, 0, 0, file_size) < 0)
 		err_exit("nodax_fd fallocate");
 
-	dax_data = mmap(NULL, FILE_SIZE, PROT_READ|PROT_WRITE, MAP_SHARED,
+	dax_data = mmap(NULL, file_size, PROT_READ|PROT_WRITE, MAP_SHARED,
 			dax_fd, 0);
 	if (dax_data == MAP_FAILED)
 		err_exit("mmap");
@@ -220,7 +222,7 @@ int main(int argc, char *argv[])
 	run_test(&truncate_down_fn);
 	run_test(&collapse_range_fn);
 
-	if (munmap(dax_data, FILE_SIZE) != 0)
+	if (munmap(dax_data, file_size) != 0)
 		err_exit("munmap");
 
 	err = close(dax_fd);
Signature · 1 line
2.19.1
author
#5 lore raw

PATCH 3/3 xfs/300: modify test to work on any fs block size

In reply to Jeff Moyer

The test currently assumes a file system block size of 4k. It will work just fine on any user-specified block size, though.

---
 tests/xfs/300 | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)
tests/xfs/300 +2 -1
index 28608b81..4f1c927a 100755
@@ -50,8 +50,9 @@ $XFS_IO_PROG -f -c "pwrite -S 0x63 0 4096" $SCRATCH_MNT/attrvals >> $seqres.full
 cat $SCRATCH_MNT/attrvals | attr -s name $SCRATCH_MNT/$seq.test >> $seqres.full 2>&1
 
 # Fragment the file by writing backwards
+bs=$(_get_file_block_size $SCRATCH_MNT)
 for I in `seq 6 -1 0`; do
-	dd if=/dev/zero of=$SCRATCH_MNT/$seq.test seek=$I bs=4k \
+	dd if=/dev/zero of=$SCRATCH_MNT/$seq.test seek=$I bs=${bs} \
 	   oflag=direct count=1 conv=notrunc >> $seqres.full 2>&1
 done
 
Signature · 1 line
2.19.1
#6 lore raw
In reply to Jeff Moyer

[cc [email protected]]

Quoted 36 lines · Jeff Moyer

This set of patches fixes a few false positives I encountered when testing DAX on ppc64le (which has a 64k page size).

Patch 1 is actually not specific to non-4k page sizes. Right now we only test for dax incompatibility in the dm flakey target. This means that tests that use dm-thin or the snapshot target will still try to run. Moving the check to _require_dm_target fixes that problem.

Patches 2 and 3 get rid of hard coded block/page sizes in the tests. They run just fine on 64k pages and 64k block sizes.

Even after these patches, there are many more tests that fail in the following configuration:

MKFS_OPTIONS="-b size=65536 -m reflink=0" MOUNT_OPTIONS="-o dax"

One class of failures is tests that create a really small file system size. Some of those tests seem to require the very small size, but others seem like they could live with a slightly bigger size that would then fit the log (the typical failure is a mkfs failure due to not enough blocks for the log). For the former case, I'm tempted to send patches to _notrun those tests, and for the latter, I'd like to bump the file system sizes up. 300MB seems to be large enough to accommodate the log. Would folks be opposed to those approaches?

Another class of failure is tests that either hard-code a block size to trigger a specific error case, or that test a multitude of block sizes. I'd like to send a patch to _notrun those tests if there is a user-specified block size. That will require parsing the MKFS_OPTIONS based on the fs type, of course. Is that something that seems reasonable?

I will follow up with a series of patches to implement those changes if there is consensus on the approach. These first three seemed straight-forward to me, so that's where I'm starting.

Quoted 6 lines

Thanks!
Jeff

[PATCH 1/3] dax/dm: disable testing on devices that don't support dax [PATCH 2/3] t_mmap_collision: fix hard-coded page size [PATCH 3/3] xfs/300: modify test to work on any fs block size

Hi Jeff,

You probably should be sending fstests patches to [email protected], otherwise they probably won't get noticed by the fstests maintainer...

Cheers,

Dave.

Signature · 2 lines
Dave Chinner
[email protected]
author
#7 lore raw
Quoted 46 lines · Dave Chinner

[cc [email protected]]

Quoted 36 lines · Jeff Moyer

This set of patches fixes a few false positives I encountered when testing DAX on ppc64le (which has a 64k page size).

Patch 1 is actually not specific to non-4k page sizes. Right now we only test for dax incompatibility in the dm flakey target. This means that tests that use dm-thin or the snapshot target will still try to run. Moving the check to _require_dm_target fixes that problem.

Patches 2 and 3 get rid of hard coded block/page sizes in the tests. They run just fine on 64k pages and 64k block sizes.

Even after these patches, there are many more tests that fail in the following configuration:

MKFS_OPTIONS="-b size=65536 -m reflink=0" MOUNT_OPTIONS="-o dax"

One class of failures is tests that create a really small file system size. Some of those tests seem to require the very small size, but others seem like they could live with a slightly bigger size that would then fit the log (the typical failure is a mkfs failure due to not enough blocks for the log). For the former case, I'm tempted to send patches to _notrun those tests, and for the latter, I'd like to bump the file system sizes up. 300MB seems to be large enough to accommodate the log. Would folks be opposed to those approaches?

Another class of failure is tests that either hard-code a block size to trigger a specific error case, or that test a multitude of block sizes. I'd like to send a patch to _notrun those tests if there is a user-specified block size. That will require parsing the MKFS_OPTIONS based on the fs type, of course. Is that something that seems reasonable?

I will follow up with a series of patches to implement those changes if there is consensus on the approach. These first three seemed straight-forward to me, so that's where I'm starting.

Quoted 6 lines

Thanks!
Jeff

[PATCH 1/3] dax/dm: disable testing on devices that don't support dax [PATCH 2/3] t_mmap_collision: fix hard-coded page size [PATCH 3/3] xfs/300: modify test to work on any fs block size

Quoted 5 lines

Hi Jeff,

You probably should be sending fstests patches to [email protected], otherwise they probably won't get noticed by the fstests maintainer...

Hm, somehow I didn't know about that list. I'll send v2 there, thanks!

-Jeff