bug reading /proc/sys/kernel/*: only first byte read.
I were debugging a weird problem with busybox, and come across this chunk of strace output:
open("/proc/sys/kernel/osrelease", O_RDONLY) = 3
read(3, "2", 1) = 1 read(3, "", 1) = 0 close(3) = 0
As you can see, after reading one byte from /proc/sys/kernel/osrelease, next read() returns 0, which is treated as end-of-file by an application.
Why busybox does this single-byte reads is another question (many shells does that, in order to be able to stop reading at newline).
But this is definitely a bug in kernel, and should be fixed....
It exists in 2.6.17 and 2.6.18
Thanks.
/mjt
Quoted 11 lines · On Fri, 20 Oct 2006 16:43:39 +0400 Michael Tokarev <[email protected]> wrote:
I were debugging a weird problem with busybox, and come across this chunk of strace output:
open("/proc/sys/kernel/osrelease", O_RDONLY)= 3read(3, "2", 1) = 1 read(3, "", 1) = 0 close(3) = 0As you can see, after reading one byte from
/proc/sys/kernel/osrelease, nextread()returns 0, which is treated as end-of-file by an application.
Quoted 7 lines
Why busybox does this single-byte reads is another question (many shells does that, in order to be able to stop reading at newline).
But this is definitely a bug in kernel, and should be fixed....
It exists in 2.6.17 and 2.6.18
Well this nearly killed me. kernel-side proc handlers are ghastly things.
Could I have this reviewed please? It surely has a hole in it somewhere.
From: Andrew Morton <[email protected]>
If you try to read things like /proc/sys/kernel/osrelease with single-byte reads, you get just one byte and then EOF. This is because _proc_do_string() assumes that the caller is read()ing into a buffer which is large enough to fit the whole string in a single hit.
Fix.
---
kernel/sysctl.c | 37 ++++++++++++++++++++++++++++--------- 1 file changed, 28 insertions(+), 9 deletions(-)
kernel/sysctl.c renamed from kernel/sysctl.c~_proc_do_string-fix-short-reads +28 -9
@@ -1682,8 +1682,7 @@ static int _proc_do_string(void* data, ichar __user *p;char c; -if (!data || !maxlen || !*lenp || - (*ppos && !write)) { +if (!data || !maxlen || !*lenp) { *lenp =0 ;return 0 ; }
@@ -1705,18 +1704,38 @@ static int _proc_do_string(void* data, i ((char *) data)[len] =0 ; *ppos += *lenp; }else { - len =strlen (data); +loff_t pos = *ppos; +const size_t slen =strlen (data); + +/* +* len is the amount of data to copy, and becomes the amount of +* data which was copied +*/ + len = slen; +if (pos > len) { + *lenp =0 ; +return 0 ; + }if (len > maxlen) len = maxlen;if (len > *lenp) len = *lenp; -if (len) -if (copy_to_user (buffer, data, len)) -return -EFAULT ; -if (len < *lenp) { -if (put_user ('\n' , ((char __user *) buffer) + len)) +/* Don't copy past the end of the string */ +if (len > slen - pos) + len = slen - pos; + data += pos; +/* Copy as much of the string as we can */ +if (len) { +if (copy_to_user (buffer, data, len))return -EFAULT ; - len++; + } +/* If we copied the whole string, now write a \n */ +if (len + pos == slen) { +if (len + pos < maxlen) { +if (put_user ('\n' , (char __user *)buffer + len)) +return -EFAULT ; + len++; + } } *lenp = len; *ppos += len;
_
Quoted 30 lines · Andrew Morton
kernel/sysctl.c renamed from kernel/sysctl.c~_proc_do_string-fix-short-reads +15
-6@@ -1705,18 +1704,38 @@ static int _proc_do_string(void* data, i ((char *) data)[len] =0 ; *ppos += *lenp; }else { - len =strlen (data); +loff_t pos = *ppos; +const size_t slen =strlen (data); + +/* +* len is the amount of data to copy, and becomes the amount of +* data which was copied +*/ + len = slen; +if (pos > len) { + *lenp =0 ; +return 0 ; + }if (len > maxlen) len = maxlen;if (len > *lenp) len = *lenp; -if (len) -if (copy_to_user (buffer, data, len)) -return -EFAULT ; -if (len < *lenp) { -if (put_user ('\n' , ((char __user *) buffer) + len)) +/* Don't copy past the end of the string */ +if (len > slen - pos) + len = slen - pos;
kernel/sysctl.c static int _proc_do_string(void* data, i · line 1734
Quoted 10 lines
kernel/sysctl.c renamed from kernel/sysctl.c~_proc_do_string-fix-short-reads +8
-1+ data += pos; +/* Copy as much of the string as we can */ +if (len) { +if (copy_to_user (buffer, data, len))return -EFAULT ; - len++; + } +/* If we copied the whole string, now write a \n */ +if (len + pos == slen) { +if (len + pos < maxlen) {
^^^^^^^^^^^^^^^^^^^^^^^
Shouldn't this be
if (len < *lenp)
?
Quoted 7 lines
kernel/sysctl.c renamed from kernel/sysctl.c~_proc_do_string-fix-short-reads +4
-0+if (put_user ('\n' , (char __user *)buffer + len)) +return -EFAULT ; + len++; + } } *lenp = len; *ppos += len;
Quoted 9 lines · Oleg Nesterov
Review
Quoted 2 lines · Andrew Morton
(quoted diff) +1
-0+ if (len + pos < maxlen) {^^^^^^^^^^^^^^^^^^^^^^^Shouldn't this be
if (len < *lenp)?
On the other hand. If we may assume that original code was correct, we can make a simpler patch?
kernel/sysctl.c renamed from kernel/sysctl.c~ +12 -4
@@ -1683,13 +1683,12 @@ static int _proc_do_string(void* data, isize_t len;char __user *p;char c; - -if (!data || !maxlen || !*lenp || - (*ppos && !write)) { + +if (!data || !maxlen || !*lenp) { *lenp =0 ;return 0 ; } - +if (write) { len =0 ; p = buffer;
@@ -1710,6 +1709,15 @@ static int _proc_do_string(void* data, i len =strlen (data);if (len > maxlen) len = maxlen; + +if (*ppos > len) { + *lenp =0 ; +return 0 ; + } + + data += *ppos; + len -= *ppos; +if (len > *lenp) len = *lenp;if (len)
Quoted 9 lines · Oleg Nesterov
Quoted 8 lines · Oleg Nesterov
Review
Quoted 2 lines · Andrew Morton
(quoted diff) +1
-0+ if (len + pos < maxlen) {^^^^^^^^^^^^^^^^^^^^^^^Shouldn't this be
if (len < *lenp)
Quoted 4 lines
Quoted 1 line
?
On the other hand. If we may assume that original code was correct, we can make a simpler patch?
I think there is an issue worth fixing her.
However as far as I can tell the code has this limitation deliberately for simplicity.
Getting the string side of this fixed even by itself is worthwhile, although it might be worth teach people about sys_uname and /bin/uname. It seems is the biggest thing people look at /proc/sys/ for...
For the non-string data the only way we can do this is to generate it into a temporary buffer and the read out the part of the buffer the user is requesting. Isn't there something in seq_file that will do this?
On the other side I'm fairly certain we can't support short writes to magic sysctl files on the write side of things.
I do have to say I like Oleg short version of this patch.
Eric
Eric W. Biederman wrote:
[]
Quoted 7 lines
However as far as I can tell the code has this limitation deliberately for simplicity.
Getting the string side of this fixed even by itself is worthwhile, although it might be worth teach people about
sys_unameand/bin/uname. It seems is the biggest thing people look at/proc/sys/for...
I come across this very issue when I discovered another bug, using un-initialized pipefs structures during early hotplug calls. I know about /bin/uname, but it requires a pipe, so wont work. Obvious alternative is /proc/sys/version, which is alot faster too (no fork+exec overhead).
/mjt