Thanks for fixing the problem. If pci_map_page() can handle a call with
scp->request_bufflen=0 as a parameter then reverting this line should be ok. Otherwise we should not remove the line but add the initialization of
cmdp->u.raw64.sg_ranz respective cmdp->u.raw.sg_ranz to 0 before (where the initialization of all other elements like "reserved", "mdisc_time" etc. take place).
Thanks,
Achim Leubner
=======================
Achim Leubner
Software Engineer / RAID drivers
ICP vortex?GmbH / Adaptec Inc.
Phone: +49-351-8718291
?
________________________________________
From: Jerome Borsboom [mailto:[email protected]]
Sent: Freitag, 13. Oktober 2006 13:04
To: [email protected]
Cc: Leubner, Achim
Subject: 2.6.18 bug in gdth.c [solved]
Recent changes in the gdth.c driver introduced an 'unable to handle kernel paging request' bug. The offending change seems to be following change in 'gdth_fill_raw_cmd':
@@ -3022,7 +3148,7 @@ #ifdef GDTH_STATISTICS
???????????? }
#endif
-??????? } else {
+??????? } else if (scp->request_bufflen) {
???????????? scp->SCp.Status = GDTH_MAP_SINGLE;
???????????? scp->SCp.Message = PCI_DMA_BIDIRECTIONAL;
???????????? page = virt_to_page(scp->request_buffer);
Reverting this line, make the driver stable again. My hypothesis is that when scp->request_bufflen is 0, then cmdp->u.raw.sg_ranz will not be assigned which makes the subsequent ha->cmd_len calculation misbehave. When you compare gdth_fill_raw_cmd with gdth_fill_cache_cmd, then in the latter function cmdp- >u.cache.sg_canz IS assigned before the conditional 'if (scp- >use_sg)...'
Jerome Borsboom