Robert Hancock wrote:
Tejun Heo wrote:
[cc'ing Robert Hancock and NVidia people]
Whole thread can be read from the following URL.
http://thread.gmane.org/gmane.linux.ide/21710
In a nutshell, with ADMA enabled, FLUSH_EXT occasionally times out. I
first suspected faulty disk (reallocation failure on flush) but SMART
reports nothing suspicious and w/ ADMA disabled, the drive works just
fine.
On a side note, on 2.6.22.1, SMART fails from time to time but the
problem went away on 2.6.24-rc6. This was apparently fixed during that
period. I guess we can ignore this for now.
Thanks.
This is kind of a longstanding problem which has been partially worked
around, but it seems not entirely. This is what I had diagnosed some
time ago:
"recently, some issues cropped up with command timeouts when a cache
flush command was immediately followed by an NCQ write. In this case,
sometimes when the NCQ write was issued, the status register changed
from 0x500 (Stopped and Idle) to 0x400 (Stopped) as it normally appears
to, however it seems like the controller would get hung in that state,
and we would time out with no notifiers set, the gen_ctl register not
indicating interrupt status, and the CPB response flags still 0 as we
left them, seemingly indicating the controller hasn't done anything with
it. Then, when the error handler kicks in we clear the GO bit to put it
back into register mode, but the Legacy flag in the status register
doesn't get set (or at least it takes longer than 1 microsecond).
Finally when we do an ADMA channel reset that seems to get it responding
again, until this happens the next time.
From some experimentation, I found that when we are issuing a NCQ
command when the last command was non-NCQ, or vice versa, if I added in
a delay of 20 microseconds between setting up the CPB and writing to the
append register, the problem appeared to go away. Problem is I don't
know if that's because it actually needs this delay, or because it
changes the timing so that it happens to work even though we're doing
something wrong, there's some event we're not waiting for, etc.
I've now verified that no switches between ADMA and register mode occur
near the time of these timeouts. Neither are we reading or writing any
of the ATA shadow registers while we're in ADMA mode."
It seems likely that this is what is happening here (a switch from an
NCQ command to a non-NCQ command, then the non-NCQ times out). It could
be in some cases the 20 microsecond delay is not enough. But it seems
bogus that we should need such an arbitrary delay in the first place.
The question I had for NVIDIA regarding this that I never got answered
was, is there any reason why we would need a delay when switching
between NCQ and non-NCQ commands on ADMA, and if not, is there any known
cause that could cause the controller to get into this seemingly
locked-up state?
Well, I guess I did sort of get an answer, but the only theory was that
the flush and the NCQ commands were being overlapped, which shouldn't be
possible (the libata core guarantees that, and if it didn't work it
would affect all controllers).
I'm kind of wondering if there's something funny going on with the
notifier register stuff, which is supposed to tell us what commands have
completed. We don't really use it at all (we had some problems with
missed completions, etc. when I tried using it, also it doesn't work if
ATAPI is enabled on the other port on the controller, apparently). I
know these controllers will do strange things like not signalling
interrupts for later events if you don't clear the notifiers in just the
right way (that being mostly determined by trial and error).
Or, maybe somehow the flush is getting issued before the controller is
really "ready" for it somehow (it's not finished cleaning up after
preceding NCQ command).
It's pretty hard for me to figure out which of the above might be the
case, especially without access to the detailed controller documentation..
-
To unsubscribe from this list: send the line "unsubscribe linux-ide" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html