| Age | Commit message (Collapse) | Author |
|---|
|
is causing problems when trying to boot sparc64 from an isp(4).
Verified to fix the sparc64/isp(4) regression by krw@; ok deraadt@
|
|
adapter_target to if their adapter isn't addressable on the bus.
ok dlg@, krw@
|
|
These will be used to replace scsi_link's adapter_buswidth and luns
fields, but for now we stay compatible with existing SCSI adapter
driver conventions while I update them to set the scsibus_attach_args
fields directly.
ok dlg@
|
|
needs either of them.
ok krw@
|
|
scsi_init_inquiry(). cut the compiled INQUIRY code over to it.
ok and tweaks from krw@
ok matthew@
|
|
it up by using the usb devices iSerial thing.
ok deraadt@
|
|
the big change is how paths between mpath capable devices and the
kernel are managed.
originally the midlayer would steal the links to the devices and
hide them behind mpath. all the changes an adapter made to a link
(eg activate or detach), the midlayer had to test if it was an mpath
link and then call special mpath code to handle it.
the original code also assumed that all paths behaved the same, but
the reality is that different devices have different command sets
and behaviours. figuring out which behaviour to pick and prioritising
them is basically the same job autoconf does with match and attach.
rather than special casing mpath in the midlayer and reimplimenting
autoconf, this turns paths into actual device drivers with match
and attach routines. after they figure out if the path is active,
they then give it to mpath(4) to use as a backend.
i have written drivers for symmetric access devices (sym(4)) where
all paths to the same logical unit are as good as each other,
lsi/engenio arrays (rdac(4), and emc arrays (emc(4)).
the rdac and emc drivers only detect active paths at attach time,
the do not cope if the controller changes state unless you unplug
the path and plug it in again to retest the active state. they also
do not have support for directing array failover.
operating and hoplugging has been tested with mpii(4), fc and sas
mpi(4), and iscsi via vscsi (claudio did this too).
ok krw@ deraadt@
|
|
way st(4) does. Have both decline to open read-only devices for
anything but read-only access.
Suggestion to fail opens rather than individual i/o's from deraadt@.
Problem USB device found and donated by chefren, who also tested
diffs. Thanks!
ok dlg@ marco@
|
|
than 1 i/o active at once. This reduces the chances that concurrent
i/o's for such devices will confuse the device or the adapter code.
It also eliminates a reason for adapter code to maintain its own
queues.
Tweak all drivers that fake INQUIRY results to set the SID_CmdQue
flag, thus continuing to claim to be able to do tagged i/o.
Positive feedback from matthew@ and marco@ for an earlier version.
ok dlg@
|
|
src/sys/dev/ata/atascsi.c r1.92 solves the problem that my scsi_link
change tickled.
tested by josh elsasser who reported the problem in pr6470
deraadt@ is letting it in again so it can get widespread testing
|
|
disks in atascsi.
as reported in pr6470
|
|
scsi device can have in flight. instead of counting users of openings on
the bus by taking away from the openings value, count the number of
pending commands on the bus in a new pending variable.
this lets us know how many outstanding commands there are. we can then use
that to make sure that all commands a device has generated get completed
before detaching the device. this helps avoid resource leaks and use after
frees.
tested by me on pci ehci/umass, fc mpi, and sas mpi.
tested by jakemsr on cardbus ohci umass.
it found issues in sas mpi which were fixed as a result of this diff.
ok krw@
|
|
all from deraadt@
tested by me with hotplugged disks on mpi(4)
|
|
device are going away, this will walk the pool and link queues and
wake up processes that are sleeping while waiting for an io or xs.
they will return NULL to the scsi_{xs,io}_get callers, which should
then check if they device is still alive. all other handlers that
are registered on the queues should be removed by their owners
before the destroy/shutdown funcs are called.
lots of help and discussion with matthew@
ok matthew@
|
|
and update atascsi(4) to make use of it. (Other HBAs will be updated
post-release.) Should allow for use of SATA drives with >2^32 LBAs.
ok deraadt@, dlg@, krw@
|
|
of a separate struct which the ioh struct includes for no good reason
anymore. just put the vars directly in the ioh.
this removes this useless abstraction.
ok krw@ matthew@
|
|
initialization strategy, rather than pretending to do user reference
counting. Previously, we would re-initialize the SCSI pool(9)s, which
had the fun consequence of causing sysctl(kern.pool.npools) to
infinite loop at IPL_VM.
ok krw@
|
|
last bastardized sync and buf call through scsi_scsi_cmd(). Flatten
code to call scsi_xs_sync() directly for all commands.
Airplane typos shaken out by various.
ok dlg@ matthew@ deraadt@
|
|
|
|
to interpret sense errors. This is initialized to the basic
interpretation routine, and specific scsi drivers (sd/st/cd) can
replace this with their own. While here kill EJUSTRETURN dance and
make more specialized interpretation routines directly call the
basic routine if desired.
Fixes by matthew@ to my first diff. Most original work by dlg@.
ok matthew@ marco@ dlg@
|
|
sparsely populated buses.
ok dlg@, krw@
|
|
in any meaningful way, so just get rid of it.
ok krw@, dlg@
|
|
ioctl requests, and dont pass the proc pointers around for any ioctl
requests in scsi land at all. neither were used, so trim the fat.
ok krw@ marco@
|
|
replaced by BUFQ's.
OK krw@, dlg@
|
|
patch from matthew dempsky
|
|
runqueue. less is more sometimes.
|
|
|
|
an xshandler gets put on a series of lists as it allocates different
resources, and uses the same tailq entry on each of these lists as
its only supposed to be on one of them at a time. however, it was
possible for the xshandler to be added to both at the same time,
therefore corrupting the lists and leading to a panic.
this diff moves from using separate flags for each queue an xshandler
could be on to having a single state variable that shows which one
it is on (or not on). this prevents an xshandler on the io runqueue
from being added to the openings runqueue, which in turn prevents
the list corruption.
some operations have been reordered to avoid races and complexity
in this little state machine.
|
|
instead of optimistically trying to use a resource by executing an
xs and then failing when there's no room for it, this puts things
that want to use the hardware on a runqueue. as resources become
available on the bus then consumers on the runqueue are popped off
and guaranteed access to the resource.
the resources are generally "ccbs" in adapter drivers, so this
abstracts a way for the midlayer to get access to them into something
called iopools.
it also provides a callback api for consumers of resources to use:
the scsi_ioh api for things that want direct access to the ccbs,
and the scsi_xsh api for things that want to issue a scsi_xfer on
the bus. these apis have been modelled on the timeout api.
scsi_xs_get and therefore scsi_scs_cmd have been cut over to using these
apis internally, so if they are allowed to sleep then can wait on the
runqueue for a resource to become available and therefore guarantee that
when executed on an adapter providing an iopool that they will succeed.
ok krw@ beck@ marco@
tested by many including krw@ beck@ mk@ okan@ todd@
|
|
void. Use XS_NO_CCB error in the scsi command (xs) to report the
NO_CCB condition. Eliminates all SUCCESSFULLY_QUEUED and COMPLETE
confusion and untangles the midlayer from the adapter a bit more.
Eyes and some fixes by miod@
There may be some compile issues on little used (i.e. I don't have
any) drivers but the change is mechanical and thus easy to remedy.
ok dlg@
|
|
one place for easier debugging and maintenance. No intended functional
changes.
ok dlg@
|
|
scsi_xs_show.
No functional change.
|
|
and do the appropriate sleeps, retries, error processing, and finally
returns an errno style value to the caller.
this cuts scsi_scsi_cmd, the ioctl code, sd_flush, and scsi_inquiry over
to scsi_xs_sync.
ok krw@
|
|
runtime out into a separate state variable. only operate on the state bits
with atomic ops. introduce the DYING state so things that sleep can figure
out if they should keep going or not.
|
|
was to try to play dangerous games with tagged queuing.
ok marco@
|
|
the midlayer.
|
|
requested by krw@ after spending a week munging through my code.
|
|
the two issues affecting it last time are gone. the first, mishandling of
TRY_AGAIN_LATER is not relevant now that krw got rid of TRY_AGAIN_LATER.
the second, the misbehaving IBM disk was found to be a problem with siop
using ordered tags on most ops combined with the speed of the new code.
putting this in so we can move forward.
ok krw@ "commit please" marco@
|
|
wasn't supported. Nuke #define of ESCAPE_NOT_SUPPORTED at the same
time.
ok miod@
|
|
ok marco@
|
|
causing a weird problems on an alpha and also appears responsible for
isp(4) weirdness i havent had a chance to examine yet.
sigh, this makes me sad.
|
|
marco@ discovered my rewrite retried commands rejected by an adapter
from a timeout, which was trivially starved by normal io going to
disks. this diff allows an xs to be marked as XS_NO_CCB, which will
cause it to be returned to the device driver to be retried as part
of the normal io queue.
tested by krw@ marco@ johan@
ok krw@ marco@
|
|
the fix for the NO_CCB breakage will follow shortly.
tested by krw@ marco@ johan@
ok krw@ marco@
|
|
TRY_AGAIN_LATER. NO_CCB is a timer based mechanism that can trivially
be made to fail by running IO to two or more disks simultaneously. The
TRY_AGAIN_LATER thing is more subtle because it now is a permanent
failure instead of transient however this is much harder to hit because
something must have gone wrong before it hits.
ok deraadt krw miod
|
|
many luns, or the entire bus), dont report ENXIO as an error to the
caller. this broke autoconf when it tried to forcefully remove a
bus such as umass and it thought there was a failure.
this introduces a way for scsi hbas to call activate/deactivate on
a device based on its target/lun address via a call to scsi_activate().
they can then schedule the actual detach/attach in a thread later via
scsi_req_probe/detach.
the mpi changes tweak the sas event handling code to use these apis
to properly handle attaches and detaches of disks. event handling
is still disabled till i can make it less chatty.
umass breakage reported by form@
|
|
scsibusdetach wasnt doign it properly, so we would be leaking on detach in
some cases.
now, with the introduction of mpath, the scsi_link structures can
represent a path to a mpath node as well as normal devices. this
intercepts the device activate entrypoints and sends them to mpath
if it it in use rather than assuming a device is always there. the
scsibusdetach change ensures that detach always ends up handling
the mpath node case too.
hotplug bus functionality (eg, usb) tested by form@
|
|
uses of it in our tree.
ok krw@ deraadt@
|
|
previously a devid was a structure containing its type, length, and
a pointer to the actual devid value. this has been changed so a
devid is a header followed immediately by the memory making up the
id value. this allows the header and its value to be allocated
together.
devids are now reference counted, so multiple things (eg, the mpath
node handlers and the various scsi_link structures) can share the
same allocation safely. this also frees devids when scsi_links go
away, which was previously not done.
if mpath is enabled, then print the devids out as part of the devices
attach line.
|
|
with it which became extremely annoying with what mpath wants to
do.
the major change is a new interface for submitting scsi commands.
previously the only way for drivers like sd, cd, st, etc to push
commands onto the hardware was via scsi_scsi_cmd(). the problem
with scsi_scsi_cmd is that it doesnt tell the caller if the command
failed, was queued, or completed unless you shoved a buf down with
it. this is important for mpath which wants to know what the physical
path to the device did so it can report it back to the midlayer
which called it.
this provides a new api which lets drivers like cd/sd/st/mpath etc
allocate an xs, fill it in, and provide a completion routine which
the midlayer will call with the state of the command when it is
finished with it. the caller is then responsible for freeing the
xs.
from the hba side of thing, the return code from the scsi_cmd
entrypoint is largely ignored now, and it is now always the
responsibility of the hba driver to call scsi_done when it has
completed the io, rather than returning COMPLETE and expecting the
midlayer to do it for you.
i have emulated scsi_scsi_cmd on top of this new api so existing
users of it will continue to work. sd(4) has been reworked to use
the new api directly to both demonstrate its use and test that the
new api actually does work.
this diff was mostly written in a day at f2k9. thanks to miod for poking
through hba drivers to help mitigate against fallout from the change to
the COMPLETE semantic. this has been reviewed by krw who didnt spot
anything wrong.
thanks to dave del debbio for testing.
ok deraadt@
|
|
midlayer that a scsi device has appeared or dissapeared. the midlayer will
queue an event and run it in the system workq (which has process context)
to handle the attach or detach.
|
