Hi,
I think we just found a bug in ost_handler.c: ost_brw_write()
(lustre-1.4.9).
Some parts of the function:
rc = obd_commitrw(OBD_BRW_WRITE, req->rq_export, &repbody->oa,
objcount, ioo, npages, local_nb, oti, rc);
if (unlikely(client_cksum != server_cksum && rc == 0)) {
int new_cksum = ost_checksum_bulk(desc);
...
}
ost_nio_pages_put(req, local_nb, npages);
The problem is in obd_commitrw(), for a read it calls
filter_io.c: filter_commitrw_read(), which will call filter_free_dio_pages().
In filter_free_dio_pages() all pages are set to zero, which will cause a
nullpointer dereference in ost_nio_pages_put().
We just commented out the call of filter_free_dio_pages() and on a first and
very brief test, it didn''t cause any problems. But maybe obd_commitrw()
is
supposed to be called after ost_nio_pages_put()?
Thanks,
Bernd
--
Bernd Schubert
Q-Leap Networks GmbH / transtec AG / ttec
is POISON_BULK defined ?>>>>> Bernd Schubert (BS) writes:BS> Hi, BS> I think we just found a bug in ost_handler.c: ost_brw_write() BS> (lustre-1.4.9). BS> Some parts of the function: BS> rc = obd_commitrw(OBD_BRW_WRITE, req->rq_export, &repbody->oa, BS> objcount, ioo, npages, local_nb, oti, rc); BS> if (unlikely(client_cksum != server_cksum && rc == 0)) { BS> int new_cksum = ost_checksum_bulk(desc); BS> ... BS> } BS> ost_nio_pages_put(req, local_nb, npages); BS> The problem is in obd_commitrw(), for a read it calls BS> filter_io.c: filter_commitrw_read(), which will call filter_free_dio_pages(). BS> In filter_free_dio_pages() all pages are set to zero, which will cause a BS> nullpointer dereference in ost_nio_pages_put(). BS> We just commented out the call of filter_free_dio_pages() and on a first and BS> very brief test, it didn''t cause any problems. But maybe obd_commitrw() is BS> supposed to be called after ost_nio_pages_put()? BS> Thanks, BS> Bernd BS> -- BS> Bernd Schubert BS> Q-Leap Networks GmbH / transtec AG / ttec BS> _______________________________________________ BS> Lustre-devel mailing list BS> Lustre-devel@clusterfs.com BS> https://mail.clusterfs.com/mailman/listinfo/lustre-devel
On Friday 16 March 2007 14:18, Alex Tomas wrote:> is POISON_BULK defined ?Thanks, we did oversee this, got defined by accident. However, I still think the code should be fixed to make it work, even when POISON_BULK is defined. Spasibo bol''shoj, Bernd -- Bernd Schubert Q-Leap Networks GmbH / transtec AG / ttec