in the queue -- do not attempt to requeue it. We only poll in two cases:
1) We have a non-interrupting controller. In this case, execution of the
previous command should have left the queue empty. (Perhaps there should be
a KASSERT() to this effect.)
2) We're in the shutdown path, either doing a cache sync or a dump. In these
cases, the retry behavior is useless, because we will no longer get
interrupts to notify us that the earlier commands completed. Instead we
just spin for a few seconds and fail anyway. (XXX We should probably clear
the queue explicitly so that the shutdown/dump will succeed.)
calling scsipi_make_xs() and scsipi_execute_xs(), instead push these into
scsipi_command. Make bustype_cmd and PHOLD/PRELE be called from
scsipi_execute_xs(). This allows us to create a xfer structure -- possibly on
the stack -- and call scsipi_execute_xs() directly.
-convert submatch() style functions (passed to config_search() or
config_found_sm()) to the locator passing variants
-pass interface attributes in some cases
-make submatch() functions look uniformly as far as possible
-avoid macros which just hide cfdata members, and reduce dependencies
on "locators.h"
scsipi_done() from their scsipi_request().
For this, add a struct scsipi_xfer * argument to scsipi_command().
If not NULL scsipi_command() will use this to enqueue this xfer, otherwise
it'll try to allocate a new one. This scsipi_xfer has to be allocated
and initialised by scsipi_make_xs() or equivalent.
In xxstart(), allocate a scsipi_xfer using scsipi_make_xs(), and if not NULL,
dequeue the buffer before calling scsipi_command(). This makes sure that
scsipi_command() will not fail, and also makes sure that xxstart() won't
be called again between the BUFQ_PEEK() and BUFQ_GET().
Fix "dequeued wrong buf" panics reported by Juergen Hannken-Illjes in
private mail and Andreas Wrede on current-users@.
Thanks to Jason Thorpe and Chuck Silver for review, and Andreas Wrede for
testing the patch.
sd3(mpt0:0:1:0): unable to allocate scsipi_xfer
sd3: not queued, error 12
Havard Eidnes's analysis of this problem is that the scsipi_xfer pool is
competing for resources with other pools, including the the inode and vnode
pools which can grow quite large.
*_scsipi_cmd(): don't biodone the buffer if scsipi_make_xs() fails, let the
caller deal with the problem
start function of block devices drivers: dequeue the buffer after the
scsipi_command() call. If scsipi_command() fails with ENOMEM don't dequeue
the buffer, and schedule a callout to call the start function after
some delay.
scsipi_init(): prime the scsipi_xfer_pool with one page. This ensure that
there is always some scsipi_xfer to play with. If scsipi_command() fails
because of pool_get(), we're sure there will be resources available later,
when the pending commands have completed.
Reviewed by Jason Thorpe and Havard Eidnes.
Todo: remove the "unable to allocate scsipi_xfer" and "not queued, error %d"
printfs, but I choose to keep them for now, to help make sure the code does
what it should.
wdc_regs structure, and array of which (indexed per channel) is pointed
to by struct wdc_softc.
- Move the resulting wdc_channel structure to atavar.h and rename it to
ata_channel. Rename the corresponding flags.
- Add a "ch_ndrive" member to struct ata_channel, which indicates the
maximum number of drives that can be present on the channel. For now,
this is always 2. Add an ATA_MAXDRIVES constant that places an upper
limit on this value, also currently 2.
