On Fri, Sep 02, 2011 at 12:08:53PM +0200, Eric Dumazet wrote: > Le vendredi 02 septembre 2011 à 11:39 +0200, Eric Dumazet a écrit : > > Le jeudi 01 septembre 2011 à 16:50 -0500, Jon Mason a écrit : > > > > > I believe modifying the MRRS values is what is causing the issues. > > > Can you try the attached patch and verify that it also resolves the > > > issue? > > > > I tested this patch and can confirm this solves the corruption problem. > > > > But my disk is _much_ slower than before > > > > # hdparm -t /dev/sda1 > > > > Before : > > > > Timing buffered disk reads: 254 MB in 3.02 seconds = 84.16 MB/sec > > > > After : > > > > Timing buffered disk reads: 120 MB in 3.04 seconds = 39.42 MB/sec > > Hmm, this speed regression is probably old : the 84MB/s was with the > standard debian 6.0.2 kernel (2.6.32-5-amd64) > This regression might be due to these two patches: d0be5ec8693944c2e2fc0de70fda9dbc1b93bd7d [SCSI] hpsa: do readl after writel in main i/o path to ensure commands don't get lost. Apparently we've been doin it rong for a decade, but only lately do we run into problems. and fec62c368b9c8b05d5124ca6c3b8336b537f26f3 [SCSI] hpsa: do not attempt to read from a write-only register Most smartarrays tolerate it, but a few new ones don't. Without this change some newer Smart Arrays will lock up and i/o will grind to a halt. with the second patch being a correction to the first. It seems like the readl after the writel should not be needed, and wasn't needed for a very long time, but there is some very hard to trigger and not yet well understood problem in which very occasionally a command would get lost and the driver thinks a command is out, but controller firmware thinks all commands are completed -- a circumstance which tends to make things grind to a halt. Those two patches avoid that problem. -- steve -- To unsubscribe from this list: send the line "unsubscribe linux-pci" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
