Re: [PATCH] Revised fix for download corruption

On Monday 04 November 2002 23.01, Robert Collins wrote:

> I'm not 100% of Henriks either at this point. We should be seeing
> *much* more dangling pointer problems, and SIGSEGV's if we are
> writing randomly on aborted writes.

The overwrites is not very "random". There is a almost 100% certainity
the reads will hit another 4K data buffer as these are mempooled, and
in most cases before the buffer have been reused.

Also this only occurs on "half_closed_clients off" and the client
aborts the request on a cache hit while data is currently beeing
swapped in or in certain conditions on pipelined requests where the
next reply is a cache hit.

So we have:
  * Overwrites which are almost constrained to overwrite the same type
of data for another request
  * Only on specific request patterns which is not very often seen in
normal browsing.
  * Very timing dependent, with a relatively small window.

Regards
Henrik
Received on Mon Nov 04 2002 - 16:50:12 MST