Skip to contentSkip to conversation

seq_file: make seq_lseek accept SEEK_END

PATCH In review Alexey Dobriyan 4 participants 4 messages Jun 9, 2008 – Jun 15, 2008 VFS

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 seq_file: make seq_lseek accept SEEK_END

Apologies for delay, such simple thing should have been sent long ago. Joakim, please, confirm.
-------------------------------------------
[PATCH] seq_file: make seq_lseek accept SEEK_END

and pretend seq_files have zero length. This should be enough to fix busybox start-stop-daemon:
http://marc.info/?t=120836691600002&r=1&w=2

It does xlseek(fd, 0, SEEK_END) to estimate amount of memory to malloc but satisfied with 0. Sudden -EINVAL from lseek(2) breaks it.

X-Introduced-By f16278c679aa72e28288435b313ba2d4494d6be5

---

 fs/seq_file.c |    2 ++
 1 file changed, 2 insertions(+)
fs/seq_file.c +2 -0
@@ -254,6 +254,8 @@ loff_t seq_lseek(struct file *file, loff_t offset, int origin)
 	switch (origin) {
 		case 1:
 			offset += file->f_pos;
+		case 2:
+			/* pretend it's zero length */
 		case 0:
 			if (offset < 0)
 				break;
#1 lore raw

busybox has been fixed, at least in trunk, not to do xlseek(fd, 0, SEEK_END) anymore to address this problem. Perhaps there are other apps out there that needs this too, but I don't know of any. Older busyboxes needs this change though.
I have moved on and don't have a setup handy where I can test this. Perhaps someone at the bb list has, CC:ing bb list.

Jocke

Quoted 32 lines · Alexey Dobriyan

Apologies for delay, such simple thing should have been sent long ago. Joakim, please, confirm.
-------------------------------------------
[PATCH] seq_file: make seq_lseek accept SEEK_END

and pretend seq_files have zero length. This should be enough to fix busybox start-stop-daemon:
http://marc.info/?t=120836691600002&r=1&w=2

It does xlseek(fd, 0, SEEK_END) to estimate amount of memory to malloc but satisfied with 0. Sudden -EINVAL from lseek(2) breaks it.

X-Introduced-By f16278c679aa72e28288435b313ba2d4494d6be5

---

 fs/seq_file.c |    2 ++
 1 file changed, 2 insertions(+)
fs/seq_file.c +2 -0
@@ -254,6 +254,8 @@ loff_t seq_lseek(struct file *file, loff_t offset, int origin)
 	switch (origin) {
 		case 1:
 			offset += file->f_pos;
+		case 2:
+			/* pretend it's zero length */
 		case 0:
 			if (offset < 0)
 				break;
#2 lore raw
Quoted 7 lines · On Mon, 09 Jun 2008 13:52:55 +0200 Joakim Tjernlund…

busybox has been fixed, at least in trunk, not to do xlseek(fd, 0, SEEK_END) anymore to address this problem. Perhaps there are other apps out there that needs this too, but I don't know of any. Older busyboxes needs this change though.
I have moved on and don't have a setup handy where I can test this. Perhaps someone at the bb list has, CC:ing bb list.

Nevertheless it'd be nice to fix old versions.

To some extent that depends on when we broke it. See below.

Quoted 21 lines · Alexey Dobriyan
Quoted 19 lines

Apologies for delay, such simple thing should have been sent long ago. Joakim, please, confirm.
-------------------------------------------
[PATCH] seq_file: make seq_lseek accept SEEK_END

and pretend seq_files have zero length. This should be enough to fix busybox start-stop-daemon:
http://marc.info/?t=120836691600002&r=1&w=2

It does xlseek(fd, 0, SEEK_END) to estimate amount of memory to malloc but satisfied with 0. Sudden -EINVAL from lseek(2) breaks it.

X-Introduced-By f16278c679aa72e28288435b313ba2d4494d6be5

---

 fs/seq_file.c |    2 ++
 1 file changed, 2 insertions(+)
Quoted 13 lines

fs/seq_file.c loff_t seq_lseek(struct file *file, loff_t offset, int origin) · line 261

Quoted 13 lines
fs/seq_file.c +2 -0
@@ -254,6 +254,8 @@ loff_t seq_lseek(struct file *file, loff_t offset, int origin)
 	switch (origin) {
 		case 1:
 			offset += file->f_pos;
+		case 2:
+			/* pretend it's zero length */
 		case 0:
 			if (offset < 0)
 				break;

Is returning zero logical, given what we do with SEEK_SET?

afait we _could_ just do

	offset = <largenum>
	/* fall through */

and let the SEEK_SET code do its thing. But I might have misread it. But still, it looks like it'd be possible to return a true(ish) value from SEEK_END?

Quoted 1 line
Quoted 1 line

X-Introduced-By: f16278c679aa72e28288435b313ba2d4494d6be5

Please don't do that. Instead, quote the commit ID _and_ the first few lines of the commit itself, like:

commit f16278c679aa72e28288435b313ba2d4494d6be5

Author: Hans Rosenfeld <[email protected]>
Date:   Fri Mar 21 18:46:59 2008 -0500

Change pagemap output format to allow for future reporting of huge pages

And here we see one reason: that commit cannot be the one which introduced this problem.

#3 lore raw
Quoted 36 lines · Andrew Morton
Quoted 7 lines · On Mon, 09 Jun 2008 13:52:55 +0200 Joakim Tjernlund…

busybox has been fixed, at least in trunk, not to do xlseek(fd, 0, SEEK_END) anymore to address this problem. Perhaps there are other apps out there that needs this too, but I don't know of any. Older busyboxes needs this change though.
I have moved on and don't have a setup handy where I can test this. Perhaps someone at the bb list has, CC:ing bb list.

Nevertheless it'd be nice to fix old versions.

To some extent that depends on when we broke it. See below.

Quoted 21 lines · Alexey Dobriyan
Quoted 19 lines

Apologies for delay, such simple thing should have been sent long ago. Joakim, please, confirm.
-------------------------------------------
[PATCH] seq_file: make seq_lseek accept SEEK_END

and pretend seq_files have zero length. This should be enough to fix busybox start-stop-daemon:
http://marc.info/?t=120836691600002&r=1&w=2

It does xlseek(fd, 0, SEEK_END) to estimate amount of memory to malloc but satisfied with 0. Sudden -EINVAL from lseek(2) breaks it.

X-Introduced-By f16278c679aa72e28288435b313ba2d4494d6be5

---

 fs/seq_file.c |    2 ++
 1 file changed, 2 insertions(+)
Quoted 15 lines
Quoted 5 lines
Quoted 5 lines
fs/seq_file.c +0 -0
@@ -254,6 +254,8 @@ loff_t seq_lseek(struct file *file, loff_t offset, int origin)
 	switch (origin) {
 		case 1:
Quoted 8 lines

fs/seq_file.c loff_t seq_lseek(struct file *file, loff_t offset, int origin) · line 261

Quoted 8 lines
fs/seq_file.c +2 -0
 			offset += file->f_pos;
+		case 2:
+			/* pretend it's zero length */
 		case 0:
 			if (offset < 0)
 				break;

Is returning zero logical, given what we do with SEEK_SET?

Definitely illogical. :-)

Quoted 6 lines

afait we _could_ just do

	offset = <largenum>
	/* fall through */

and let the SEEK_SET code do its thing. But I might have misread it.

If it's largenum, not enough output will be generated during each iteration until kmalloc() will refuse to give more memory.

Quoted 2 lines

But still, it looks like it'd be possible to return a true(ish) value from SEEK_END?

I think this is doable.

On the other side is
a) ugliness (remember how much output each item generates, remember ->index, allocate memory to do that),
b) SEEK_END reported and reports zero-length for other proc files (and sysctls), and nobody really cared.