Some careful testing has revealed yet another cbdata barrier error in
aufs causing similar errors.. this time in aioCancel. Makes the same
damn read bug resurface
How to repeat:
half_closed_clients off (makes detection simpler, not stricly
required)
patch aio to be slow at scheduling operations
run under a malloc debugger such as valgrind ("memory_pools off"
required for valgrind to see this error)
abort a request
Regards
Henrik
On Tuesday 05 November 2002 23.39, Phil Oester wrote:
> After 5 hours of testing with my 'corruption' patch, your patch has
> shown zero corruption (as expected).
>
> I'll load this up on some of our boxes today and see if anything
> breaks ;-)
>
> -Phil
>
> On Tue, Nov 05, 2002 at 12:31:03PM +0100, Henrik Nordstrom wrote:
> > Attached is a preleminary Squid-2.5 patch that inforces the
> > cbdata barrier on aufs reads.
> >
> > Quick summary: The responsibility of copying the read data into
> > the buffer is moved up to the storeAufsRead() layer who can
> > correcly verify if the swapin request has been aborted. It also
> > adds a few cbdata verifications to insure that aborted events is
> > properly ignored and a few on-free checks to avoid memory or
> > filedescriptor leakage.
Received on Sun Nov 10 2002 - 04:30:01 MST
This archive was generated by hypermail pre-2.1.9 : Tue Dec 09 2003 - 16:18:40 MST