Skip to contentSkip to conversation

xfsprogs: Fix for xfstest 252 hang on ext4

PATCH Allison Henderson 3 participants 8 messages May 23, 2011 – Jun 1, 2011

lore
Share & tools
lore
#0 lore raw

Hi all,

While trying to add more punch hole tests to xfstest, I found that test 252 hangs on ext4 due to a loop in xfsprogs that does not exit. XFS gets out of this loop because there is logic in the loop that looks for the last extent flag and breaks out. But it looks like ext4 does not return a last extent when the file has a hole at the end. I am not sure if this is the correct behavior or not, so I will copy the ext4 folks on this too. Below is a copy of the fix for xfsprogs:

io/fiemap.c +1 -1
index fa990cc..81fc92c 100644
@@ -246,7 +246,7 @@ fiemap_f(
 		       flg_w, _("FLAGS"));
 	}
 
-	while (!last && ((cur_extent + 1) != max_extents)) {
+	while (!last && (cur_extent <= max_extents)) {
 		if (max_extents)
 			num_extents = min(num_extents,
 					  max_extents - (cur_extent + 1));

It looks like the loop enters with last=0, cur_extents=0, and max_extents = 0, and on the first iteration cur_extents get set to 2, so we dont see ((cur_extent + 1) == max_extents for a very long time. I doubt the logic was meant to work that way, so this patch should fix it, but I wanted to make sure that the fiemap for ext4 is working as intended too. Feed back appreciated! Thx all!

Allison Henderson

#1 lore raw
Quoted 3 lines · Allison Henderson

Hi all,

While trying to add more punch hole tests to xfstest, I found that test 252 hangs on ext4 due to a loop in xfsprogs that does not exit.  XFS gets out of this loop because there is logic in the loop that looks for the last extent flag and breaks out.  But it looks like ext4 does not return a last extent when the file has a hole at the end.  I am not sure if this is the correct behavior or not, so I will copy the ext4 folks on this too.  Below is a copy of the fix for xfsprogs:

Hi there,

What's blocksize of the tested ext4? For now, ext4 returns
LAST_EXTENT if the logical offset covered by the extent is greater than file size, so if there is a hole at the end, no last extent is returned. Thx!

Yongqiang.

Quoted 24 lines
io/fiemap.c +1 -1
index fa990cc..81fc92c 100644
@@ -246,7 +246,7 @@ fiemap_f(
                       flg_w, _("FLAGS"));
        }
 
-       while (!last && ((cur_extent + 1) != max_extents)) {
+       while (!last && (cur_extent <= max_extents)) {
                if (max_extents)
                        num_extents = min(num_extents,
                                          max_extents - (cur_extent + 1));

It looks like the loop enters with last=0, cur_extents=0, and max_extents = 0, and on the first iteration cur_extents get set to 2, so we dont see ((cur_extent + 1) == max_extents for a very long time.  I doubt the logic was meant to work that way, so this patch should fix it, but I wanted to make sure that the fiemap for ext4 is working as intended too.  Feed back appreciated!  Thx all!

Allison Henderson
--
To unsubscribe from this list: send the line "unsubscribe linux-ext4" in the body of a message to [email protected]
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Signature · 6 lines
Best Wishes
Yongqiang Yang
--
To unsubscribe from this list: send the line "unsubscribe linux-ext4" in
the body of a message to [email protected]
More majordomo info at  http://vger.kernel.org/majordomo-info.html
#2 lore raw
Quoted 6 lines · Yongqiang Yang
Quoted 3 lines · Allison Henderson

Hi all,

While trying to add more punch hole tests to xfstest, I found that test 252 hangs on ext4 due to a loop in xfsprogs that does not exit. XFS gets out of this loop because there is logic in the loop that looks for the last extent flag and breaks out. But it looks like ext4 does not return a last extent when the file has a hole at the end. I am not sure if this is the correct behavior or not, so I will copy the ext4 folks on this too. Below is a copy of the fix for xfsprogs:

Quoted 8 lines

Hi there,

What's blocksize of the tested ext4? For now, ext4 returns
LAST_EXTENT if the logical offset covered by the extent is greater than file size, so if there is a hole at the end, no last extent is returned. Thx!

Yongqiang.

Hi there,

The block size I've been using is 4096. As long as that behavior is expected, I think the test will be ok with just the xfsprogs fix, though. Thx!

Allison Henderson

Quoted 28 lines
Quoted 25 lines
io/fiemap.c +1 -1
index fa990cc..81fc92c 100644
@@ -246,7 +246,7 @@ fiemap_f(
                       flg_w, _("FLAGS"));
        }
 
-       while (!last&&  ((cur_extent + 1) != max_extents)) {
+       while (!last&&  (cur_extent<= max_extents)) {
                if (max_extents)
                        num_extents = min(num_extents,
                                          max_extents - (cur_extent + 1));

It looks like the loop enters with last=0, cur_extents=0, and max_extents = 0, and on the first iteration cur_extents get set to 2, so we dont see ((cur_extent + 1) == max_extents for a very long time. I doubt the logic was meant to work that way, so this patch should fix it, but I wanted to make sure that the fiemap for ext4 is working as intended too. Feed back appreciated! Thx all!

Allison Henderson
--
To unsubscribe from this list: send the line "unsubscribe linux-ext4" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html

#6 lore raw
In reply to Allison Henderson
Quoted 19 lines · Allison Henderson

Hi all,

While trying to add more punch hole tests to xfstest, I found that test 252 hangs on ext4 due to a loop in xfsprogs that does not exit. XFS gets out of this loop because there is logic in the loop that looks for the last extent flag and breaks out. But it looks like ext4 does not return a last extent when the file has a hole at the end. I am not sure if this is the correct behavior or not, so I will copy the ext4 folks on this too. Below is a copy of the fix for xfsprogs:

io/fiemap.c +1 -1
index fa990cc..81fc92c 100644
@@ -246,7 +246,7 @@ fiemap_f(
  		       flg_w, _("FLAGS"));
  	}
 
-	while (!last&&  ((cur_extent + 1) != max_extents)) {
+	while (!last&&  (cur_extent<= max_extents)) {
  		if (max_extents)
  			num_extents = min(num_extents,
  					  max_extents - (cur_extent + 1));
Quoted 4 lines

It looks like the loop enters with last=0, cur_extents=0, and max_extents = 0, and on the first iteration cur_extents get set to 2, so we dont see ((cur_extent + 1) == max_extents for a very long time. I doubt the logic was meant to work that way, so this patch should fix it, but I wanted to make sure that the fiemap for ext4 is working as intended too. Feed back appreciated! Thx all!

Allison Henderson

Hi all,

I haven't heard much back on this patch, so Im just poking this thread to make sure it doesn't get forgotten. I have some patches out there for punch hole, and I'm currently looking at fixing up some older punch hole tests in the dmapi code, but they wont do much good for ext4 with out this fix. If I could get a quick peek from some one on the xfs list for this patch, that would be much appreciated. Thx all!

Allison Henderson

Quoted 4 lines

_______________________________________________
xfs mailing list
[email protected]
http://oss.sgi.com/mailman/listinfo/xfs

_______________________________________________
xfs mailing list
[email protected]
http://oss.sgi.com/mailman/listinfo/xfs

#7 lore raw
Quoted 39 lines · Allison Henderson
Quoted 34 lines · Allison Henderson

Hi all,

While trying to add more punch hole tests to xfstest, I found that test 252 hangs on ext4 due to a loop in xfsprogs that does not exit. XFS gets out of this loop because there is logic in the loop that looks for the last extent flag and breaks out. But it looks like ext4 does not return a last extent when the file has a hole at the end. I am not sure if this is the correct behavior or not, so I will copy the ext4 folks on this too. Below is a copy of the fix for xfsprogs:

io/fiemap.c +1 -1
index fa990cc..81fc92c 100644
@@ -246,7 +246,7 @@ fiemap_f(
 		       flg_w, _("FLAGS"));
 	}
 
-	while (!last&&  ((cur_extent + 1) != max_extents)) {
+	while (!last&&  (cur_extent<= max_extents)) {
 		if (max_extents)
 			num_extents = min(num_extents,
 					  max_extents - (cur_extent + 1));

It looks like the loop enters with last=0, cur_extents=0, and max_extents = 0, and on the first iteration cur_extents get set to 2, so we dont see ((cur_extent + 1) == max_extents for a very long time. I doubt the logic was meant to work that way, so this patch should fix it, but I wanted to make sure that the fiemap for ext4 is working as intended too. Feed back appreciated! Thx all!

Quoted 2 lines

Allison Henderson

Hi all,

Quoted 7 lines

I haven't heard much back on this patch, so Im just poking this thread to make sure it doesn't get forgotten. I have some patches out there for punch hole, and I'm currently looking at fixing up some older punch hole tests in the dmapi code, but they wont do much good for ext4 with out this fix. If I could get a quick peek from some one on the xfs list for this patch, that would be much appreciated. Thx all!

If ext4 is not setting the last extent flag on the last extent then that's an ext4 bug that the test has detected, right? And so you should be fixing ext4 rather than modifying the test to hide the different behaviour?

Cheer

Signature · 2 lines
Dave Chinner
[email protected]
#3 lore raw
In reply to Allison Henderson
Quoted 76 lines · Allison Henderson
Quoted 12 lines · Yongqiang Yang
Quoted 9 lines · Allison Henderson

Hi all,

While trying to add more punch hole tests to xfstest, I found that test 252 hangs on ext4 due to a loop in xfsprogs that does not exit. XFS gets out of this loop because there is logic in the loop that looks for the last extent flag and breaks out. But it looks like ext4 does not return a last extent when the file has a hole at the end. I am not sure if this is the correct behavior or not, so I will copy the ext4 folks on this too. Below is a copy of the fix for xfsprogs:

Quoted 8 lines

Hi there,

What's blocksize of the tested ext4? For now, ext4 returns LAST_EXTENT if the logical offset covered by the extent is greater than file size, so if there is a hole at the end, no last extent is returned. Thx!

Yongqiang.

Hi there,

The block size I've been using is 4096. As long as that behavior is expected, I think the test will be ok with just the xfsprogs fix, though. Thx!

Allison Henderson

Quoted 38 lines
Quoted 24 lines
io/fiemap.c +1 -1
index fa990cc..81fc92c 100644
@@ -246,7 +246,7 @@ fiemap_f(
 flg_w, _("FLAGS"));
 }
 
- while (!last&& ((cur_extent + 1) != max_extents)) {
+ while (!last&& (cur_extent<= max_extents)) {
 if (max_extents)
 num_extents = min(num_extents,
 max_extents - (cur_extent + 1));

It looks like the loop enters with last=0, cur_extents=0, and max_extents = 0, and on the first iteration cur_extents get set to 2, so we dont see ((cur_extent + 1) == max_extents for a very long time. I doubt the logic was meant to work that way, so this patch should fix it, but I wanted to make sure that the fiemap for ext4 is working as intended too. Feed back appreciated! Thx all!

Quoted 6 lines

Allison Henderson
--
To unsubscribe from this list: send the line "unsubscribe linux-ext4" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html

Hi all,

Quoted 7 lines

I haven't heard much back on this patch, so Im just poking this thread to make sure it doesn't get forgotten. I have some patches out there for punch hole, and I'm currently looking at fixing up some older punch hole tests in the dmapi code, but they wont do much good for ext4 with out this fix. If I could get a quick peek from some one on the xfs list for this patch, that would be much appreciated. Thx all!

Quoted 8 lines

If ext4 is not setting the last extent flag on the last extent then that's an ext4 bug that the test has detected, right? And so you should be fixing ext4 rather than modifying the test to hide the different behaviour?

Cheer

-- Dave Chinner [email protected]

Hi All,

Sorry, I should have poked the thread with Yongqiang's response, so I will move the dialog into this thread. At the moment, it sounds like the fiemap for ext4 is working as intended. Yongqiang, do you agree that the fiemap for ext4 should be changed? I think you are more familiar with this part of the code than I am, and I just want to make sure we find a solution that everyone is happy with. Thx!

Allison Henderson

Quoted 4 lines

_______________________________________________
xfs mailing list
[email protected]
http://oss.sgi.com/mailman/listinfo/xfs

_______________________________________________
xfs mailing list
[email protected]
http://oss.sgi.com/mailman/listinfo/xfs

#4 lore raw
Quoted 87 lines · Allison Henderson
Quoted 86 lines · Allison Henderson
Quoted 14 lines · Yongqiang Yang
Quoted 10 lines · Allison Henderson

Hi all,

While trying to add more punch hole tests to xfstest, I found that test 252 hangs on ext4 due to a loop in xfsprogs that does not exit. XFS gets out of this loop because there is logic in the loop that looks for the last extent flag and breaks out. But it looks like ext4 does not return a last extent when the file has a hole at the end. I am not sure if this is the correct behavior or not, so I will copy the ext4 folks on this too. Below is a copy of the fix for xfsprogs:

Quoted 8 lines

Hi there,

What's blocksize of the tested ext4? For now, ext4 returns LAST_EXTENT if the logical offset covered by the extent is greater than file size, so if there is a hole at the end, no last extent is returned. Thx!

Yongqiang.

Hi there,

The block size I've been using is 4096. As long as that behavior is expected, I think the test will be ok with just the xfsprogs fix, though. Thx!

Allison Henderson

Quoted 38 lines
Quoted 24 lines
io/fiemap.c +1 -1
index fa990cc..81fc92c 100644
@@ -246,7 +246,7 @@ fiemap_f(
 flg_w, _("FLAGS"));
 }
 
- while (!last&& ((cur_extent + 1) != max_extents)) {
+ while (!last&& (cur_extent<= max_extents)) {
 if (max_extents)
 num_extents = min(num_extents,
 max_extents - (cur_extent + 1));

It looks like the loop enters with last=0, cur_extents=0, and max_extents = 0, and on the first iteration cur_extents get set to 2, so we dont see ((cur_extent + 1) == max_extents for a very long time. I doubt the logic was meant to work that way, so this patch should fix it, but I wanted to make sure that the fiemap for ext4 is working as intended too. Feed back appreciated! Thx all!

Quoted 6 lines

Allison Henderson
--
To unsubscribe from this list: send the line "unsubscribe linux-ext4" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html

Hi all,

Quoted 7 lines

I haven't heard much back on this patch, so Im just poking this thread to make sure it doesn't get forgotten. I have some patches out there for punch hole, and I'm currently looking at fixing up some older punch hole tests in the dmapi code, but they wont do much good for ext4 with out this fix.  If I could get a quick peek from some one on the xfs list for this patch, that would be much appreciated.  Thx all!

If ext4 is not setting the last extent flag on the last extent then that's an ext4 bug that the test has detected, right? And so you should be fixing ext4 rather than modifying the test to hide the different behaviour?

Cheer

Quoted 8 lines
Quoted 1 line

-- Dave Chinner [email protected]

Hi All,

Sorry, I should have poked the thread with Yongqiang's response, so I will move the dialog into this thread.  At the moment, it sounds like the fiemap for ext4 is working as intended.  Yongqiang, do you agree that the fiemap for ext4 should be changed?  I think you are more familiar with this part of

Yes, I agree. ext4 should return LAST extent. I am thinking if we can find a new solution collecting extents.

Maybe we can insert delayed extents into extent tree. This way fiemap will be simpler and much more efficient.

I would like to throw out the proposal inserting delayed-extents into extent tree. What will it bring? AFAIK it will bring:
1. We need to down i_data_sem in delayed write-path to insert delayed-extents into the tree without journaling it.

2. When we come to block allocation, we can convert delayed-extents to normal extents.

There is a problem that the solution can be only used in ordered mode. So what are your opinions?

Yongqiang.

Quoted 15 lines

the code than I am, and I just want to make sure we find a solution that everyone is happy with.  Thx!

Allison Henderson

Quoted 4 lines

_______________________________________________
xfs mailing list
[email protected]
http://oss.sgi.com/mailman/listinfo/xfs

Signature · 6 lines
Best Wishes
Yongqiang Yang
--
To unsubscribe from this list: send the line "unsubscribe linux-ext4" in
the body of a message to [email protected]
More majordomo info at  http://vger.kernel.org/majordomo-info.html
#5 lore raw
Quoted 108 lines · Yongqiang Yang
Quoted 87 lines · Allison Henderson
Quoted 86 lines · Allison Henderson
Quoted 14 lines · Yongqiang Yang
Quoted 10 lines · Allison Henderson

Hi all,

While trying to add more punch hole tests to xfstest, I found that test 252 hangs on ext4 due to a loop in xfsprogs that does not exit. XFS gets out of this loop because there is logic in the loop that looks for the last extent flag and breaks out. But it looks like ext4 does not return a last extent when the file has a hole at the end. I am not sure if this is the correct behavior or not, so I will copy the ext4 folks on this too. Below is a copy of the fix for xfsprogs:

Quoted 8 lines

Hi there,

What's blocksize of the tested ext4? For now, ext4 returns LAST_EXTENT if the logical offset covered by the extent is greater than file size, so if there is a hole at the end, no last extent is returned. Thx!

Yongqiang.

Hi there,

The block size I've been using is 4096. As long as that behavior is expected, I think the test will be ok with just the xfsprogs fix, though. Thx!

Allison Henderson

Quoted 38 lines
Quoted 24 lines
io/fiemap.c +1 -1
index fa990cc..81fc92c 100644
@@ -246,7 +246,7 @@ fiemap_f(
 flg_w, _("FLAGS"));
 }
 
- while (!last&&  ((cur_extent + 1) != max_extents)) {
+ while (!last&&  (cur_extent<= max_extents)) {
 if (max_extents)
 num_extents = min(num_extents,
 max_extents - (cur_extent + 1));

It looks like the loop enters with last=0, cur_extents=0, and max_extents = 0, and on the first iteration cur_extents get set to 2, so we dont see ((cur_extent + 1) == max_extents for a very long time. I doubt the logic was meant to work that way, so this patch should fix it, but I wanted to make sure that the fiemap for ext4 is working as intended too. Feed back appreciated! Thx all!

Quoted 6 lines

Allison Henderson
--
To unsubscribe from this list: send the line "unsubscribe linux-ext4" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html

Hi all,

Quoted 7 lines

I haven't heard much back on this patch, so Im just poking this thread to make sure it doesn't get forgotten. I have some patches out there for punch hole, and I'm currently looking at fixing up some older punch hole tests in the dmapi code, but they wont do much good for ext4 with out this fix. If I could get a quick peek from some one on the xfs list for this patch, that would be much appreciated. Thx all!

If ext4 is not setting the last extent flag on the last extent then that's an ext4 bug that the test has detected, right? And so you should be fixing ext4 rather than modifying the test to hide the different behaviour?

Cheer

Quoted 8 lines
Quoted 1 line

-- Dave Chinner [email protected]

Hi All,

Sorry, I should have poked the thread with Yongqiang's response, so I will move the dialog into this thread. At the moment, it sounds like the fiemap for ext4 is working as intended. Yongqiang, do you agree that the fiemap for ext4 should be changed? I think you are more familiar with this part of

Yes, I agree. ext4 should return LAST extent. I am thinking if we can find a new solution collecting extents.

Maybe we can insert delayed extents into extent tree. This way fiemap will be simpler and much more efficient.

I would like to throw out the proposal inserting delayed-extents into extent tree. What will it bring? AFAIK it will bring:
1. We need to down i_data_sem in delayed write-path to insert delayed-extents into the tree without journaling it.

Quoted 7 lines

2. When we come to block allocation, we can convert delayed-extents to normal extents.

There is a problem that the solution can be only used in ordered mode. So what are your opinions?

Yongqiang.

Hi All,

Well, I am not yet familiar with how the code for ordered mode works, so I may not be much help here, but I do think that if we could get it to work, it would help simplify a lot of things including fiemap and punch hole. I know that the earlier versions of punch hole were a little complicated because of the different mechanisms needed to identify mapped extents, delayed extents and holes. Eventually what we did was to flush out the data in the region to be punched out in order to avoid race conditions, and also to simplify the logic. This has introduced some problems in some of the existing xfstests punch hole tests, because the tests want to see a hole in unwritten extents instead of written extents. I was working on some optimization patches to avoid this, but if this proposal is put in place, I think I could optimize that fix a lot. And I know the fiemap routines would get simpler sense they would only have to deal with the extent tree. If we decide to do what you propose, I will wait on doing any punch hole or fiemap patches to take advantage of that. I'll poke around with the modes too to see if there's any thing we can do to get around that problem. Thx!

Allison Henderson

Quoted 10 lines
Quoted 10 lines

the code than I am, and I just want to make sure we find a solution that everyone is happy with. Thx!

Allison Henderson

Quoted 9 lines
Quoted 6 lines
Quoted 4 lines

_______________________________________________
xfs mailing list
[email protected]
http://oss.sgi.com/mailman/listinfo/xfs

--
To unsubscribe from this list: send the line "unsubscribe linux-ext4" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html