On 6 August 2015 at 18:54, Rainer Jung <rainer.j...@kippdata.de> wrote:
> Am 06.08.2015 um 16:36 schrieb Rainer Jung:
>>
>> Position of crash:
>>
>> #0 0xff29f194 in vparse_tuple (pool=pool@entry=0x35f10,
>> fmt=fmt@entry=0xffbff264, ap=ap@entry=0xffbff20c,
>> items=<error reading variable: Unhandled dwarf expression opcode
>> 0xfa>, items=<error reading variable: Unhandled dwarf expression opcode
>> 0xfa>)
>> at subversion/libsvn_ra_svn/marshal.c:1310
>> 1310 *va_arg(*ap, apr_uint64_t *) = elt->u.number;
>>
>> elt is:
>>
>> (gdb) print *elt
>> $2 = {kind = SVN_RA_SVN_NUMBER, u = {number = 2, string = 0x0, word =
>> 0x0, list = 0x0}}
>>
>> The addresses are:
>>
>> elt address is: 0x7012c
>> elt->u and elt->u.number addresses are both: 0x70134
>>
>> and the crash happens when elt->u.number is being accessed as an
>> apr_uint64_t under this address which is only 4 byte aligned.
>>
>> I haven't tracked down, where elt is actually being allocated. That
>> would be the place to make sure, it is 8 byte aligned. It should be
>> automatic if allocated using its type svn_ra_svn_item_t, but maybe it is
>> allocated in a more generic way with a type the compiler can not align
>> correctly for the later use as svn_ra_svn_item_t.
>
>
> I think the root cause is here (file subversion/libsvn_ra_svn/marshal.c):
>
> 1082 /* Allocate an APR array with room for (initially) 4 items.
> 1083 * We do this manually because lists are the most frequent
> protocol
> 1084 * element, often used to frame a single, optional value. We
> save
> 1085 * about 20% of total protocol handling time. */
> 1086 char *buffer = apr_palloc(pool, sizeof(apr_array_header_t)
> 1087 + 4 *
> sizeof(svn_ra_svn_item_t));
> 1088 svn_ra_svn_item_t *data
> 1089 = (svn_ra_svn_item_t *)(buffer +
> sizeof(apr_array_header_t));
> 1090
> 1091 item->kind = SVN_RA_SVN_LIST;
> 1092 item->u.list = (apr_array_header_t *)buffer;
>
> "buffer" is not specifically aligned, the array members in "item->u.list =
> (apr_array_header_t *)buffer" could be misaligned.
>
> The following (ugly) workaround fixes it for me:
>
> --- subversion/libsvn_ra_svn/marshal.c.kpdt_orig Fri Feb 13 12:17:40
> 2015
> +++ subversion/libsvn_ra_svn/marshal.c Thu Aug 6 17:46:58 2015
> @@ -1083,10 +1083,16 @@
> * We do this manually because lists are the most frequent protocol
> * element, often used to frame a single, optional value. We save
> * about 20% of total protocol handling time. */
> - char *buffer = apr_palloc(pool, sizeof(apr_array_header_t)
> +
> + /* Make sure the data part of the buffer has appropriate alignment
> + by prefixing it with a size that fits the needed
> apr_array_header_t
> + but is itself highly aligned. */
> + size_t offset = sizeof(apr_array_header_t) / 8 * 8;
> +
> + char *buffer = apr_palloc(pool, offset
> + 4 * sizeof(svn_ra_svn_item_t));
> svn_ra_svn_item_t *data
> - = (svn_ra_svn_item_t *)(buffer + sizeof(apr_array_header_t));
> + = (svn_ra_svn_item_t *)(buffer + offset);
>
> item->kind = SVN_RA_SVN_LIST;
> item->u.list = (apr_array_header_t *)buffer;
>
> But of course its a bit rough, because it would apply on all platforms, even
> if not needed. Also on some (future?) platforms, the alignment for 8 bytes
> might not always be correct.
>
> It's a bit tragic that this code part is prefixed with:
>
> * We do this manually because lists are the most frequent protocol
> * element, often used to frame a single, optional value. We save
> * about 20% of total protocol handling time. */
>
> and the trap is that doing it manually often is harder than expected.
> Switching to apr_array_make() would have not introduced this bug, but of
> course you did it for a reason.
>
I really doubt that saving one apr_palloc() call even for such hot
space could give 20% improvements: apr_palloc() is pretty optimized.
May be all changes in r1485851 commit gives 20% protocol handling time
improvement, but not this particural change.
--
Ivan Zhakov
