For some time, we see various unaligned exceptions running apt-get on hppa:
http(13559): unaligned access to 0x00000000fa703d49 at ip=0x00000000f9f0a9bb
handle_unaligned: 37 callbacks suppressed
http(13810): unaligned access to 0x00000000fa703d49 at ip=0x00000000f9f0a9bb
http(13810): unaligned access to 0x00000000fa703d49 at ip=0x00000000f9f0a9c3
http(13810): unaligned access to 0x00000000fa703d49 at ip=0x00000000f9f0cdf3
http(13810): unaligned access to 0x00000000fa703d49 at ip=0x00000000f9f0cecf
http(13810): unaligned access to 0x00000000fa703d4d at ip=0x00000000f9f0c69b
Helge and myself tracked the first of these exceptions to the following line
hp->rcode = NOERROR; /* default */
The argument answer has the type u_char *. Thus, the function __libc_res_nquery
should nominally be prepared to access the HEADER struct on a byte boundary.
However, the struct HEADER is not defined with the packed attribute, so accesses
to the bit fields in the struct are done with word rather than byte accesses.
This causes the above faults.
This is very inefficient on strict alignment targets such as hppa and ia64, and
slow on x86, etc.
Adding "__attribute__((packed))" to the HEADER typedef appears to eliminate
the unaligned accesses from http.
New patch here:
New patch is here:
Did this ever get landed? https://sourceware.org/legacy-ml/libc-alpha/2016-06/msg01020.html
On 2022-01-30 10:29 p.m., sam at gentoo dot org wrote:
> Did this ever get landed?
No but Debian has been carrying it for some time.
(In reply to dave.anglin from comment #5)
> On 2022-01-30 10:29 p.m., sam at gentoo dot org wrote:
> > Did this ever get landed?
> No but Debian has been carrying it for some time.
I spoke to azanella and we agreed the patch seems reasonable:
[17:44:45] <azanella> sam_, this has triggered an issue with or1k recently https://email@example.com/ and I think the latest patch from david should be ok (although I am not very found of the packet usage)
[17:52:21] <sam_> oh I'd missed that one
[18:03:55] <sam_> are you able to apply it or do I need to ask david to do it?
[18:09:35] <azanella> sam_, could you please ask sam to repost it, so we can proper review it?
[18:39:27] <sam_> azanella: ofc, will do
Could you resend it to libc-alpha and then we can get this in?
On 2022-03-14 2:40 p.m., sam at gentoo dot org wrote:
> Could you resend it to libc-alpha and then we can get this in?
Will do after recheck on master.
Patch v4 here:
commit 05dec22d7be722987ff07aebf9690f6078b3c4e9 (HEAD -> master, origin/master, origin/HEAD)
Author: John David Anglin <firstname.lastname@example.org>
Date: Tue Mar 22 17:35:54 2022 +0000
resolv: Fix unaligned accesses to fields in HEADER struct
The structure HEADER is normally aligned to a word boundary but
sometimes it needs to be accessed when aligned on a byte boundary.
This change defines a new typedef, UHEADER, with alignment 1.
It is used to ensure the fields are accessed with byte loads and
stores when necessary.
V4: Change to res_mkquery.c deleted. Small whitespace fix.
V5: Move UHEADER typedef to resolv/resolv-internal.h. Replace all
HEADER usage with UHEADER in resolv/res_send.c.
Signed-off-by: John David Anglin <email@example.com>
Reviewed-by: Adhemerval Zanella <firstname.lastname@example.org>