On Fri, Nov 24 2000, Boszormenyi Zoltan wrote:
> Hi!
>
> I tried 2.4.0-test11 (plain, +ac1/2) with and without
> Jens' blk-11 patch. This indeed performs (much) better
> when there is only high disk activity but cdrecord
> starts up _very_ slowly if the kernel was compiled with
> blk-11. It does not happen if blk-11 is not applied.
>
> I stopped cdrecord before it started writing because of
> this suspicious slowness and I did not want to create a bad CD.
>
> Other data points:
> The CD-writer is a Yamaha-6416 (SCSI version).
> The SCSI card is a Diamond Fireport-40 (Symbios 53c875j)
> I tested both the in-kernel 1.6b and 1.7.2 versions of the
> sym53c8xx driver.
>
> The slowdown was experienced in every case where
> the kernel contained blk-11.
You might want to send messages such as this one to me
as well, so I don't miss them :-)
The problem is due to sg assuming that scsi_do_req will
fire the request queue immediately to let the command
inject complete. This was never really the case, even
in the stock kernel. Here's a quick-and-dirty patch
against test11+blk-11 attached, untested but it should
fix the delays.
--
* Jens Axboe <[EMAIL PROTECTED]>
* SuSE Labs
--- /opt/kernel/linux-2.4.0-test11/drivers/scsi/sg.c Tue Oct 24 22:58:20 2000
+++ drivers/scsi/sg.c Fri Nov 24 21:57:50 2000
@@ -689,6 +689,7 @@
(void *)SRpnt->sr_buffer, hp->dxfer_len,
sg_cmd_done_bh, timeout, SG_DEFAULT_RETRIES);
/* dxfer_len overwrites SRpnt->sr_bufflen, hence need for b_malloc_len */
+ generic_unplug_device(&SRpnt->sr_device->request_queue);
return 0;
}
