kmsg: lseek errors confuse glibc's dprintf
glibc's dprintf implementation does not work correctly with /dev/kmsg file descriptors because glibc treats receiving EBADF and EINVAL from lseek when trying to determine the current file position as errors. See
https://sourceware.org/bugzilla/show_bug.cgi?id=17830
Quoted 1 line
From what I can tell prior to Kay Sievers printk record commit
e11fea92e13fb91c50bacca799a6131c81929986, calling lseek(fd, 0, SEEK_CUR) with such a file descriptor would not return an error.
Prior to Kay's change, Arnd Bergmann's commit 6038f373a3dc1f1c26496e60b6c40b164716f07e seemed to go to some lengths to preserve the successful return code rather than returning (the perhaps more logical) -ESPIPE.
glibc is happy with either a successful return or -ESPIPE.
For maximum compatibility it seems that success should be returned but given Kay's new seek interface perhaps this isn't helpful.
This patch ensures that such a seek succeeds:
kernel/printk/printk.c +6 -1
@@ -693,7 +693,7 @@ static loff_t devkmsg_llseek(struct file *file, loff_t offset, int whence)loff_t ret =0 ;if (!user) -return -EBADF ; +return (whence ==SEEK_CUR ) ?0 : -EBADF ;if (offset)return -ESPIPE ;
@@ -718,6 +718,11 @@ static loff_t devkmsg_llseek(struct file *file, loff_t offset, int whence) user->idx = log_next_idx; user->seq = log_next_seq;break ; +case SEEK_CUR : +/* For compatibility with userspace requesting the +* current file position. */ + ret =0 ; +break ;default : ret = -EINVAL ; }
(although it could be argued that the !user case should return -ESPIPE rather than EBADF since the file descriptor _is_ valid.)
and this patch causes a failure that glibc is prepared to accept:
kernel/printk/printk.c +6 -1
@@ -693,7 +693,7 @@ static loff_t devkmsg_llseek(struct file *file, loff_t offset, int whence)loff_t ret =0 ;if (!user) -return -EBADF ; +return -ESPIPE ;if (offset)return -ESPIPE ;
@@ -718,6 +718,11 @@ static loff_t devkmsg_llseek(struct file *file, loff_t offset, int whence) user->idx = log_next_idx; user->seq = log_next_seq;break ; +case SEEK_CUR : +/* For compatibility with userspace expecting SEEK_CUR +* to not yield EINVAL. */ + ret = -ESPIPE ; +break ;default : ret = -EINVAL ; }
Either makes dprintf work, but is either the right solution?
Thanks.
Mike.
Quoted 26 lines · On Thu, 15 Jan 2015 17:31:32 +0000 Mike Crowe <[email protected]> wrote:
glibc's dprintf implementation does not work correctly with
/dev/kmsgfile descriptors because glibc treats receiving EBADF and EINVAL from lseek when trying to determine the current file position as errors. See
https://sourceware.org/bugzilla/show_bug.cgi?id=17830Quoted 1 line
From what I can tell prior to Kay Sievers printk record commit
e11fea92e13fb91c50bacca799a6131c81929986, callinglseek(fd, 0, SEEK_CUR)with such a file descriptor would not return an error.Prior to Kay's change, Arnd Bergmann's commit
6038f373a3dc1f1c26496e60b6c40b164716f07eseemed to go to some lengths to preserve the successful return code rather than returning (the perhaps more logical)-ESPIPE.glibc is happy with either a successful return or
-ESPIPE.For maximum compatibility it seems that success should be returned but given Kay's new seek interface perhaps this isn't helpful.
This patch ensures that such a seek succeeds:
kernel/printk/printk.c +0
-0
kernel/printk/printk.c static loff_t devkmsg_llseek(struct file *file, loff_t offset, int whence) · line 696
Quoted 5 lines
kernel/printk/printk.c +1
-1@@ -693,7 +693,7 @@ static loff_t devkmsg_llseek(struct file *file, loff_t offset, int whence)loff_t ret =0 ;if (!user) -return -EBADF ; +return (whence ==SEEK_CUR ) ?0 : -EBADF ;
What's actually going on here? What is the significance of file->private_data==NULL and why does this code treat it as an error?
kernel/printk/printk.c static loff_t devkmsg_llseek(struct file *file, loff_t offset, int whence) · line 725
Quoted 12 lines
kernel/printk/printk.c +5
-0if (offset)return -ESPIPE ;@@ -718,6 +718,11 @@ static loff_t devkmsg_llseek(struct file *file, loff_t offset, int whence) user->idx = log_next_idx; user->seq = log_next_seq;break ; +case SEEK_CUR : +/* For compatibility with userspace requesting the +* current file position. */ + ret =0 ; +break ;
Can we actually do something useful here? Return some value which can be fed back into SEEK_SET to restore the file position?
Quoted 3 lines
kernel/printk/printk.c +0
-0default : ret = -EINVAL ; }
Quoted 30 lines · Andrew Morton
Quoted 28 lines · On Thu, 15 Jan 2015 17:31:32 +0000 Mike Crowe <[email protected]> wrote:
glibc's dprintf implementation does not work correctly with
/dev/kmsgfile descriptors because glibc treats receiving EBADF and EINVAL from lseek when trying to determine the current file position as errors. See
https://sourceware.org/bugzilla/show_bug.cgi?id=17830From what I can tell prior to Kay Sievers printk record commit
e11fea92e13fb91c50bacca799a6131c81929986, callinglseek(fd, 0, SEEK_CUR)with such a file descriptor would not return an error.Prior to Kay's change, Arnd Bergmann's commit
6038f373a3dc1f1c26496e60b6c40b164716f07eseemed to go to some lengths to preserve the successful return code rather than returning (the perhaps more logical)-ESPIPE.glibc is happy with either a successful return or
-ESPIPE.For maximum compatibility it seems that success should be returned but given Kay's new seek interface perhaps this isn't helpful.
This patch ensures that such a seek succeeds:
kernel/printk/printk.c +0
-0@@ -693,7 +693,7 @@ static loff_t devkmsg_llseek(struct file *file, loff_t offset, int whence)loff_t ret =0 ;
Quoted 6 lines
kernel/printk/printk.c static loff_t devkmsg_llseek(struct file *file, loff_t offset, int whence) · line 696
Quoted 3 lines
kernel/printk/printk.c +1
-1if (!user) -return -EBADF ; +return (whence ==SEEK_CUR ) ?0 : -EBADF ;What's actually going on here? What is the significance of file->private_data==NULL and why does this code treat it as an error?
Kay is presumably the expert on this but my understanding is that opening /dev/kmsg for writing only is supposed to be as lightweight as possible - there's not even a context structure so file->private_data is NULL (see devmksg_open.) In this mode seeking is not supported. I believe that EBADF is not a particularly helpful error in this case but didn't want to complicate the patch with a separate change.
When /dev/kmsg is opened for reading seeking is supported but the seek behaviour is tailored to (I assume) systemd journal's usage.
Quoted 15 lines
kernel/printk/printk.c static loff_t devkmsg_llseek(struct file *file, loff_t offset, int whence) · line 725
Quoted 12 lines
kernel/printk/printk.c +5
-0if (offset)return -ESPIPE ;@@ -718,6 +718,11 @@ static loff_t devkmsg_llseek(struct file *file, loff_t offset, int whence) user->idx = log_next_idx; user->seq = log_next_seq;break ; +case SEEK_CUR : +/* For compatibility with userspace requesting the +* current file position. */ + ret =0 ; +break ;Can we actually do something useful here? Return some value which can be fed back into
SEEK_SETto restore the file position?
Perhaps that would be useful for someone. It would certainly be more logical than just returning zero.
Quoted 4 lines
kernel/printk/printk.c static loff_t devkmsg_llseek(struct file *file, loff_t offset, int whence) · line 728
Quoted 4 lines
kernel/printk/printk.c +0
-0default : ret = -EINVAL ; }
Mike.
Hello Mike and all,
Quoted 4 lines · On Thu, 15 Jan 2015 17:31:32 +0000 Mike Crowe <[email protected]> wrote:
glibc's dprintf implementation does not work correctly with
/dev/kmsgfile descriptors because glibc treats receiving EBADF and EINVAL from lseek when trying to determine the current file position as errors. See
https://sourceware.org/bugzilla/show_bug.cgi?id=17830
we need to conclude on this issue. This is a real bug which is ignored for 4 years now. Mike, would you like to re-send a formal patch? I can do it as well, preserving a link to your original patch/report. In case you'd like to post it yourself, I can be a tester and/or provide a reproducer.
Quoted 78 lines
Quoted 1 line
Quoted 1 line
From what I can tell prior to Kay Sievers printk record commit
e11fea92e13fb91c50bacca799a6131c81929986, callinglseek(fd, 0, SEEK_CUR)with such a file descriptor would not return an error.Prior to Kay's change, Arnd Bergmann's commit
6038f373a3dc1f1c26496e60b6c40b164716f07eseemed to go to some lengths to preserve the successful return code rather than returning (the perhaps more logical)-ESPIPE.glibc is happy with either a successful return or
-ESPIPE.For maximum compatibility it seems that success should be returned but given Kay's new seek interface perhaps this isn't helpful.
This patch ensures that such a seek succeeds:
kernel/printk/printk.c +6
-1@@ -693,7 +693,7 @@ static loff_t devkmsg_llseek(struct file *file, loff_t offset, int whence)loff_t ret =0 ;if (!user) -return -EBADF ; +return (whence ==SEEK_CUR ) ?0 : -EBADF ;if (offset)return -ESPIPE ;@@ -718,6 +718,11 @@ static loff_t devkmsg_llseek(struct file *file, loff_t offset, int whence) user->idx = log_next_idx; user->seq = log_next_seq;break ; +case SEEK_CUR : +/* For compatibility with userspace requesting the +* current file position. */ + ret =0 ; +break ;default : ret = -EINVAL ; }(although it could be argued that the !user case should return
-ESPIPErather than EBADF since the file descriptor _is_ valid.)and this patch causes a failure that glibc is prepared to accept:
kernel/printk/printk.c +6
-1@@ -693,7 +693,7 @@ static loff_t devkmsg_llseek(struct file *file, loff_t offset, int whence)loff_t ret =0 ;if (!user) -return -EBADF ; +return -ESPIPE ;if (offset)return -ESPIPE ;@@ -718,6 +718,11 @@ static loff_t devkmsg_llseek(struct file *file, loff_t offset, int whence) user->idx = log_next_idx; user->seq = log_next_seq;break ; +case SEEK_CUR : +/* For compatibility with userspace expecting SEEK_CUR +* to not yield EINVAL. */ + ret = -ESPIPE ; +break ;default : ret = -EINVAL ; }Either makes dprintf work, but is either the right solution?
Thanks.
Mike.
Signature · 1 line
Alexander Sverdlin.
Quoted 11 lines · Alexander Sverdlin
Hello Mike and all,
Quoted 4 lines · On Thu, 15 Jan 2015 17:31:32 +0000 Mike Crowe <[email protected]> wrote:
glibc's dprintf implementation does not work correctly with
/dev/kmsgfile descriptors because glibc treats receiving EBADF and EINVAL from lseek when trying to determine the current file position as errors. See
https://sourceware.org/bugzilla/show_bug.cgi?id=17830
Quoted 5 lines
we need to conclude on this issue. This is a real bug which is ignored for 4 years now. Mike, would you like to re-send a formal patch? I can do it as well, preserving a link to your original patch/report. In case you'd like to post it yourself, I can be a tester and/or provide a reproducer.
The patch needs to be rebased because of the changed file location. I would also suggest adding a "Cc: [email protected]" tag so it will get backported into stable kernels.
Quoted 42 lines
Quoted 42 lines
Quoted 1 line
Quoted 1 line
From what I can tell prior to Kay Sievers printk record commit
e11fea92e13fb91c50bacca799a6131c81929986, callinglseek(fd, 0, SEEK_CUR)with such a file descriptor would not return an error.Prior to Kay's change, Arnd Bergmann's commit
6038f373a3dc1f1c26496e60b6c40b164716f07eseemed to go to some lengths to preserve the successful return code rather than returning (the perhaps more logical)-ESPIPE.glibc is happy with either a successful return or
-ESPIPE.For maximum compatibility it seems that success should be returned but given Kay's new seek interface perhaps this isn't helpful.
This patch ensures that such a seek succeeds:
kernel/printk/printk.c +6
-1@@ -693,7 +693,7 @@ static loff_t devkmsg_llseek(struct file *file, loff_t offset, int whence)loff_t ret =0 ;if (!user) -return -EBADF ; +return (whence ==SEEK_CUR ) ?0 : -EBADF ;if (offset)return -ESPIPE ;@@ -718,6 +718,11 @@ static loff_t devkmsg_llseek(struct file *file, loff_t offset, int whence) user->idx = log_next_idx; user->seq = log_next_seq;break ; +case SEEK_CUR : +/* For compatibility with userspace requesting the +* current file position. */ + ret =0 ; +break ;default : ret = -EINVAL ; }
Quoted 2 lines
Quoted 2 lines
(although it could be argued that the !user case should return
-ESPIPErather than EBADF since the file descriptor _is_ valid.)
I don't think the !user case can ever be hit, I would just leave that to return -BADF and not touch it.
Quoted 13 lines
Quoted 13 lines
kernel/printk/printk.c +5
-0@@ -718,6 +718,11 @@ static loff_t devkmsg_llseek(struct file *file, loff_t offset, int whence) user->idx = log_next_idx; user->seq = log_next_seq;break ; +case SEEK_CUR : +/* For compatibility with userspace expecting SEEK_CUR +* to not yield EINVAL. */ + ret = -ESPIPE ; +break ;default : ret = -EINVAL ; }
Quoted 1 line
Quoted 1 line
Either makes dprintf work, but is either the right solution?
I'd vote for -ESPIPE, for consistency with the offset!=0 case, but etiher one is fine with me here.
Arnd
Quoted 65 lines · Arnd Bergmann
Quoted 11 lines · Alexander Sverdlin
Hello Mike and all,
Quoted 4 lines · On Thu, 15 Jan 2015 17:31:32 +0000 Mike Crowe <[email protected]> wrote:
glibc's dprintf implementation does not work correctly with
/dev/kmsgfile descriptors because glibc treats receiving EBADF and EINVAL from lseek when trying to determine the current file position as errors. See
https://sourceware.org/bugzilla/show_bug.cgi?id=17830Quoted 5 lines
we need to conclude on this issue. This is a real bug which is ignored for 4 years now. Mike, would you like to re-send a formal patch? I can do it as well, preserving a link to your original patch/report. In case you'd like to post it yourself, I can be a tester and/or provide a reproducer.
The patch needs to be rebased because of the changed file location. I would also suggest adding a "Cc: [email protected]" tag so it will get backported into stable kernels.
Quoted 42 lines
Quoted 42 lines
Quoted 1 line
Quoted 1 line
From what I can tell prior to Kay Sievers printk record commit
e11fea92e13fb91c50bacca799a6131c81929986, callinglseek(fd, 0, SEEK_CUR)with such a file descriptor would not return an error.Prior to Kay's change, Arnd Bergmann's commit
6038f373a3dc1f1c26496e60b6c40b164716f07eseemed to go to some lengths to preserve the successful return code rather than returning (the perhaps more logical)-ESPIPE.glibc is happy with either a successful return or
-ESPIPE.For maximum compatibility it seems that success should be returned but given Kay's new seek interface perhaps this isn't helpful.
This patch ensures that such a seek succeeds:
kernel/printk/printk.c +6
-1@@ -693,7 +693,7 @@ static loff_t devkmsg_llseek(struct file *file, loff_t offset, int whence)loff_t ret =0 ;if (!user) -return -EBADF ; +return (whence ==SEEK_CUR ) ?0 : -EBADF ;if (offset)return -ESPIPE ;@@ -718,6 +718,11 @@ static loff_t devkmsg_llseek(struct file *file, loff_t offset, int whence) user->idx = log_next_idx; user->seq = log_next_seq;break ; +case SEEK_CUR : +/* For compatibility with userspace requesting the +* current file position. */ + ret =0 ; +break ;default : ret = -EINVAL ; }
Quoted 5 lines
Quoted 2 lines
Quoted 2 lines
(although it could be argued that the !user case should return
-ESPIPErather than EBADF since the file descriptor _is_ valid.)I don't think the !user case can ever be hit, I would just leave that to return -BADF and not touch it.
I originally wrote the patch for linux-3.8(!) and back then a write-only kmsg instance had no private data. It looks like this changed with 750afe7babd117daabebf4855da18e4418ea845e in v4.8, so user should now always be valid.
Quoted 13 lines
Quoted 13 lines
Quoted 13 lines
kernel/printk/printk.c +5
-0@@ -718,6 +718,11 @@ static loff_t devkmsg_llseek(struct file *file, loff_t offset, int whence) user->idx = log_next_idx; user->seq = log_next_seq;break ; +case SEEK_CUR : +/* For compatibility with userspace expecting SEEK_CUR +* to not yield EINVAL. */ + ret = -ESPIPE ; +break ;default : ret = -EINVAL ; }
Quoted 4 lines
Quoted 1 line
Quoted 1 line
Either makes dprintf work, but is either the right solution?
I'd vote for
-ESPIPE, for consistency with the offset!=0 case, but etiher one is fine with me here.
It seems that this is the version we've successfully been running in our kernel since then, so that's good.
Thanks.
Mike.
Quoted 10 lines · Alexander Sverdlin
Hello Mike and all,
Quoted 4 lines · On Thu, 15 Jan 2015 17:31:32 +0000 Mike Crowe <[email protected]> wrote:
glibc's dprintf implementation does not work correctly with
/dev/kmsgfile descriptors because glibc treats receiving EBADF and EINVAL from lseek when trying to determine the current file position as errors. See
https://sourceware.org/bugzilla/show_bug.cgi?id=17830
Quoted 5 lines
we need to conclude on this issue. This is a real bug which is ignored for 4 years now. Mike, would you like to re-send a formal patch? I can do it as well, preserving a link to your original patch/report. In case you'd like to post it yourself, I can be a tester and/or provide a reproducer.
Hi Alexander,
I'm sorry for dropping the ball on this. I think I was travelling at the time. :( Thank you for picking it up.
I don't mind rebasing my patch once I know which of my two solutions is preferred. It seems that the -ESPIPE one is the current favourite.
Thanks.
Quoted 73 lines
Quoted 73 lines
Quoted 1 line
Quoted 1 line
From what I can tell prior to Kay Sievers printk record commit
e11fea92e13fb91c50bacca799a6131c81929986, callinglseek(fd, 0, SEEK_CUR)with such a file descriptor would not return an error.Prior to Kay's change, Arnd Bergmann's commit
6038f373a3dc1f1c26496e60b6c40b164716f07eseemed to go to some lengths to preserve the successful return code rather than returning (the perhaps more logical)-ESPIPE.glibc is happy with either a successful return or
-ESPIPE.For maximum compatibility it seems that success should be returned but given Kay's new seek interface perhaps this isn't helpful.
This patch ensures that such a seek succeeds:
kernel/printk/printk.c +6
-1@@ -693,7 +693,7 @@ static loff_t devkmsg_llseek(struct file *file, loff_t offset, int whence)loff_t ret =0 ;if (!user) -return -EBADF ; +return (whence ==SEEK_CUR ) ?0 : -EBADF ;if (offset)return -ESPIPE ;@@ -718,6 +718,11 @@ static loff_t devkmsg_llseek(struct file *file, loff_t offset, int whence) user->idx = log_next_idx; user->seq = log_next_seq;break ; +case SEEK_CUR : +/* For compatibility with userspace requesting the +* current file position. */ + ret =0 ; +break ;default : ret = -EINVAL ; }(although it could be argued that the !user case should return
-ESPIPErather than EBADF since the file descriptor _is_ valid.)and this patch causes a failure that glibc is prepared to accept:
kernel/printk/printk.c +6
-1@@ -693,7 +693,7 @@ static loff_t devkmsg_llseek(struct file *file, loff_t offset, int whence)loff_t ret =0 ;if (!user) -return -EBADF ; +return -ESPIPE ;if (offset)return -ESPIPE ;@@ -718,6 +718,11 @@ static loff_t devkmsg_llseek(struct file *file, loff_t offset, int whence) user->idx = log_next_idx; user->seq = log_next_seq;break ; +case SEEK_CUR : +/* For compatibility with userspace expecting SEEK_CUR +* to not yield EINVAL. */ + ret = -ESPIPE ; +break ;default : ret = -EINVAL ; }
Quoted 1 line
Quoted 1 line
Either makes dprintf work, but is either the right solution?
Mike.
Hi!
On 21/03/2019 13:38, Mike Crowe wrote:
[...]
Quoted 2 lines
I don't mind rebasing my patch once I know which of my two solutions is preferred. It seems that the
-ESPIPEone is the current favourite.
I think:
1. Since Linux v4.8 (!user) case in not possible any more if the device file was opened in write-only mode, so you are probably safe to omit this chunk at all.
2. Nothing comes to my mind, what one could return as "current position" from the lseek() call on /dev/kmsg. So without good idea what to return, supporting offset == 0 (the only thing we ever can support on SEEK_CUR) makes no sense to me.
Documentation/ABI/testing/dev-kmsg makes no promise on SEEK_CUR either, therefore I'd say, -ESPIPE unconditionally is the right choice.
Quoted 77 lines
Quoted 73 lines
Quoted 73 lines
Quoted 1 line
Quoted 1 line
From what I can tell prior to Kay Sievers printk record commit
e11fea92e13fb91c50bacca799a6131c81929986, callinglseek(fd, 0, SEEK_CUR)with such a file descriptor would not return an error.Prior to Kay's change, Arnd Bergmann's commit
6038f373a3dc1f1c26496e60b6c40b164716f07eseemed to go to some lengths to preserve the successful return code rather than returning (the perhaps more logical)-ESPIPE.glibc is happy with either a successful return or
-ESPIPE.For maximum compatibility it seems that success should be returned but given Kay's new seek interface perhaps this isn't helpful.
This patch ensures that such a seek succeeds:
kernel/printk/printk.c +6
-1@@ -693,7 +693,7 @@ static loff_t devkmsg_llseek(struct file *file, loff_t offset, int whence)loff_t ret =0 ;if (!user) -return -EBADF ; +return (whence ==SEEK_CUR ) ?0 : -EBADF ;if (offset)return -ESPIPE ;@@ -718,6 +718,11 @@ static loff_t devkmsg_llseek(struct file *file, loff_t offset, int whence) user->idx = log_next_idx; user->seq = log_next_seq;break ; +case SEEK_CUR : +/* For compatibility with userspace requesting the +* current file position. */ + ret =0 ; +break ;default : ret = -EINVAL ; }(although it could be argued that the !user case should return
-ESPIPErather than EBADF since the file descriptor _is_ valid.)and this patch causes a failure that glibc is prepared to accept:
kernel/printk/printk.c +6
-1@@ -693,7 +693,7 @@ static loff_t devkmsg_llseek(struct file *file, loff_t offset, int whence)loff_t ret =0 ;if (!user) -return -EBADF ; +return -ESPIPE ;if (offset)return -ESPIPE ;@@ -718,6 +718,11 @@ static loff_t devkmsg_llseek(struct file *file, loff_t offset, int whence) user->idx = log_next_idx; user->seq = log_next_seq;break ; +case SEEK_CUR : +/* For compatibility with userspace expecting SEEK_CUR +* to not yield EINVAL. */ + ret = -ESPIPE ; +break ;default : ret = -EINVAL ; }Quoted 1 line
Quoted 1 line
Either makes dprintf work, but is either the right solution?
Mike.
Signature · 2 lines
Best regards, Alexander Sverdlin.