vmcore: prevent PT_NOTE p_memsz overflow during header update
Subsystems (2)
PATCH v2 1/1 vmcore: prevent PT_NOTE p_memsz overflow during header update
Currently, update_note_header_size_elf64() and update_note_header_size_elf32() will add the size of a PT_NOTE entry to real_sz even if that causes real_sz to exceeds max_sz. This patch corrects the while loop logic in those routines to ensure that does not happen and prints a warning if a PT_NOTE entry is dropped.
One possible negative side effect of exceeding the max_sz limit is an allocation failure in merge_note_headers_elf64() or merge_note_headers_elf32() which would produce console output such as the following while booting the crash kernel.
vmalloc: allocation failure: 14076997632 bytes swapper/0: page allocation failure: order:0, mode:0x80d2
CPU: 0 PID: 1 Comm: swapper/0 Not tainted 3.10.0-gbp1 #7 ffffffff817dcc30 ffff88003025fc28 ffffffff815bdb0b ffff88003025fcb0 ffffffff8113b3d0 ffffffff817dcc30 ffff88003025fc48 ffffc90000000018 ffff88003025fcc0 ffff88003025fc60 ffff88003025fc80 ffff88002b5df980 Call Trace: [<ffffffff815bdb0b>] dump_stack+0x19/0x1b [<ffffffff8113b3d0>] warn_alloc_failed+0xf0/0x160 [<ffffffff81a1267d>] ? merge_note_headers_elf64.constprop.9+0x116/0x24a [<ffffffff8116d34e>] __vmalloc_node_range+0x19e/0x250 [<ffffffff81210454>] ? read_from_oldmem.part.0+0xa4/0xe0 [<ffffffff8116d4ec>] vmalloc_user+0x4c/0x70 [<ffffffff81a1267d>] ? merge_note_headers_elf64.constprop.9+0x116/0x24a [<ffffffff81a1267d>] merge_note_headers_elf64.constprop.9+0x116/0x24a [<ffffffff81a12cce>] vmcore_init+0x2d4/0x76c [<ffffffff8120f9f0>] ? kcore_update_ram+0x1f0/0x1f0 [<ffffffff81063b92>] ? walk_system_raange+0x112/0x130 [<ffffffff81a129fa>] ? merge_note_headers_elf32.constprop.8+0x249/0x249 [<ffffffff810020e2>] do_one_initcall+0xe2/0x190 [<ffffffff819e20c4>] kernel_init_freeable+0x17c/0x207 [<ffffffff819e18d0>] ? do_early_param+0x88/0x88 [<ffffffff815a0d20>] ? rest_init+0x80/0x80 [<ffffffff815a0d2e>] kernel_init+0xe/0x180 [<ffffffff815cd8ac>] ret_from_fork+0x7c/0xb0 [<ffffffff815a0d20>] ? rest_init+0x80/0x80
Kdump: vmcore not initialized
kdump: dump target is /dev/sda4
kdump: saving to /sysroot//var/crash/127.0.0.1-2014.01.28-13:58:52/
kdump: saving vmcore-dmesg.txt
Cannot open /proc/vmcore: No such file or directory
kdump: saving vmcore-dmesg.txt failed
kdump: saving vmcore
kdump: saving vmcore failed
This type of failure has been seen on a four socket prototype system with certain memory configurations. Most PT_NOTE sections have a single entry similar to:
n_namesz = 0x5
n_descsz = 0x150
n_type = 0x1
Occasionally, a second entry is encountered with very large n_namesz and n_descsz sizes:
n_namesz = 0x80000008
n_descsz = 0x510ae163
n_type = 0x80000008
Not yet sure of the source of these extra entries, they seem bogus, but they shouldn't cause crash dump to fail.
--- fs/proc/vmcore.c | 16 ++++++++++------ 1 file changed, 10 insertions(+), 6 deletions(-)
fs/proc/vmcore.c +10 -6
@@ -468,12 +468,14 @@ static int __init update_note_header_size_elf64(const Elf64_Ehdr *ehdr_ptr)return rc; } nhdr_ptr = notes_section; -while (real_sz < max_sz) { -if (nhdr_ptr->n_namesz ==0 ) -break ; +while (nhdr_ptr->n_namesz !=0 ) { sz =sizeof (Elf64_Nhdr) + ((nhdr_ptr->n_namesz +3 ) & ~3 ) + ((nhdr_ptr->n_descsz +3 ) & ~3 ); +if ((real_sz + sz) > max_sz) { +pr_warn ("Warning: dropping PT_NOTE entry\n" ); +break ; + } real_sz += sz; nhdr_ptr = (Elf64_Nhdr *)((char *)nhdr_ptr + sz); }
@@ -648,12 +650,14 @@ static int __init update_note_header_size_elf32(const Elf32_Ehdr *ehdr_ptr)return rc; } nhdr_ptr = notes_section; -while (real_sz < max_sz) { -if (nhdr_ptr->n_namesz ==0 ) -break ; +while (nhdr_ptr->n_namesz !=0 ) { sz =sizeof (Elf32_Nhdr) + ((nhdr_ptr->n_namesz +3 ) & ~3 ) + ((nhdr_ptr->n_descsz +3 ) & ~3 ); +if ((real_sz + sz) > max_sz) { +pr_warn ("Warning: dropping PT_NOTE entry\n" ); +break ; + } real_sz += sz; nhdr_ptr = (Elf32_Nhdr *)((char *)nhdr_ptr + sz); }
Signature · 1 line
1.8.3.2
On Mon, Feb 03, 2014 at 01:18:38PM -0700, Greg Pearson wrote:
[..]
Quoted 9 lines
fs/proc/vmcore.c +0
-1@@ -468,12 +468,14 @@ static int __init update_note_header_size_elf64(const Elf64_Ehdr *ehdr_ptr)return rc; } nhdr_ptr = notes_section; -while (real_sz < max_sz) {
fs/proc/vmcore.c static int __init update_note_header_size_elf64(const Elf64_Ehdr *ehdr_ptr) · line 478
Quoted 10 lines
fs/proc/vmcore.c +5
-2-if (nhdr_ptr->n_namesz ==0 ) -break ; +while (nhdr_ptr->n_namesz !=0 ) { sz =sizeof (Elf64_Nhdr) + ((nhdr_ptr->n_namesz +3 ) & ~3 ) + ((nhdr_ptr->n_descsz +3 ) & ~3 ); +if ((real_sz + sz) > max_sz) { +pr_warn ("Warning: dropping PT_NOTE entry\n" ); +break ; + }
Hi Greg,
Couple of minor nits.
I think it is a good idea to give more data in warning which tells why are we dropping a note entry. May be something like.
"Warning: Total note entry size exceeded PT_NOTE memsz. Dropping PT_NOTE entry, n_namesz=<> n_descsz=<>".
Secondly, if there is only on note entry in a PT_NOTE header and we drop it, then that PT_NOTE header is empty and needs to be cleaned up.
I think you will have to modify get_note_number_and_size_elf64() and other relevant functions which are not expecting ->p_memsz=0.
Thanks
Vivek
Quoted 24 lines · Vivek Goyal
On Mon, Feb 03, 2014 at 01:18:38PM -0700, Greg Pearson wrote:
[..]
Quoted 9 lines
fs/proc/vmcore.c +0
-1@@ -468,12 +468,14 @@ static int __init update_note_header_size_elf64(const Elf64_Ehdr *ehdr_ptr)return rc; } nhdr_ptr = notes_section; -while (real_sz < max_sz) {fs/proc/vmcore.c static int __init update_note_header_size_elf64(const Elf64_Ehdr *ehdr_ptr) · line 478
Quoted 10 lines
fs/proc/vmcore.c +5
-2-if (nhdr_ptr->n_namesz ==0 ) -break ; +while (nhdr_ptr->n_namesz !=0 ) { sz =sizeof (Elf64_Nhdr) + ((nhdr_ptr->n_namesz +3 ) & ~3 ) + ((nhdr_ptr->n_descsz +3 ) & ~3 ); +if ((real_sz + sz) > max_sz) { +pr_warn ("Warning: dropping PT_NOTE entry\n" ); +break ; + }Hi Greg,
Quoted 6 lines
Couple of minor nits.
I think it is a good idea to give more data in warning which tells why are we dropping a note entry. May be something like.
"Warning: Total note entry size exceeded
PT_NOTEmemsz. DroppingPT_NOTEentry, n_namesz=<> n_descsz=<>".
Sounds good I'll add more information to the pr_warn().
Quoted 6 lines
Secondly, if there is only on note entry in a
PT_NOTEheader and we drop it, then thatPT_NOTEheader is empty and needs to be cleaned up.I think you will have to modify
get_note_number_and_size_elf64()and other relevant functions which are not expecting ->p_memsz=0.
What about treating this as an error condition and adding a check to the update_note_header_size_elf32()/update_note_header_size_elf64() routines that would return a failure, something like the following at the end of the routine:
if (real_sz == 0)
return -EINVALI could also add a pr_warn() with a message indicating no PT_NOTE entries were found.
This seems like a lower risk change to handle the case I'm currently seeing as opposed to changing the code to handle a p_memsz==0.
Thoughts?
--
Greg
Quoted 49 lines · Greg
Quoted 24 lines · Vivek Goyal
On Mon, Feb 03, 2014 at 01:18:38PM -0700, Greg Pearson wrote:
[..]
Quoted 9 lines
fs/proc/vmcore.c +0
-1@@ -468,12 +468,14 @@ static int __init update_note_header_size_elf64(const Elf64_Ehdr *ehdr_ptr)return rc; } nhdr_ptr = notes_section; -while (real_sz < max_sz) {fs/proc/vmcore.c static int __init update_note_header_size_elf64(const Elf64_Ehdr *ehdr_ptr) · line 478
Quoted 10 lines
fs/proc/vmcore.c +5
-2-if (nhdr_ptr->n_namesz ==0 ) -break ; +while (nhdr_ptr->n_namesz !=0 ) { sz =sizeof (Elf64_Nhdr) + ((nhdr_ptr->n_namesz +3 ) & ~3 ) + ((nhdr_ptr->n_descsz +3 ) & ~3 ); +if ((real_sz + sz) > max_sz) { +pr_warn ("Warning: dropping PT_NOTE entry\n" ); +break ; + }Hi Greg,
Quoted 6 lines
Couple of minor nits.
I think it is a good idea to give more data in warning which tells why are we dropping a note entry. May be something like.
"Warning: Total note entry size exceeded
PT_NOTEmemsz. DroppingPT_NOTEentry, n_namesz=<> n_descsz=<>".Sounds good I'll add more information to the
pr_warn().Quoted 6 lines
Secondly, if there is only on note entry in a
PT_NOTEheader and we drop it, then thatPT_NOTEheader is empty and needs to be cleaned up.I think you will have to modify
get_note_number_and_size_elf64()and other relevant functions which are not expecting ->p_memsz=0.What about treating this as an error condition and adding a check to the update_note_header_size_elf32()/update_note_header_size_elf64() routines that would return a failure, something like the following at the end of the routine:
if (real_sz == 0) return -EINVAL
Quoted 7 lines
I could also add a
pr_warn()with a message indicating noPT_NOTEentries were found.This seems like a lower risk change to handle the case I'm currently seeing as opposed to changing the code to handle a p_memsz==0.
Thoughts?
If you retrun with -EINVAL, that means /proc/vmcore will not be created. This probably is too harsh an action. If a PT_NOTE is corrupt, we can just ignore it and continue with construction of /proc/vmcore. We will just loose cpu registers of one cpu. May be that cpu was idling and it did not matter for the purpose of crash analysis.
IMO, we can just ignore empty PT_NOTE (real_sz), give a warning and move on.
Thanks
Vivek