reduce lock congestion and improve SMP scalability of the SCSI/ATA stack,
preparing the ground for the coming next GEOM direct dispatch support.
Replace big per-SIM locks with bunch of smaller ones:
- per-LUN locks to protect device and peripheral drivers state;
- per-target locks to protect list of LUNs on target;
- per-bus locks to protect reference counting;
- per-send queue locks to protect queue of CCBs to be sent;
- per-done queue locks to protect queue of completed CCBs;
- remaining per-SIM locks now protect only HBA driver internals.
While holding LUN lock it is allowed (while not recommended for performance
reasons) to take SIM lock. The opposite acquisition order is forbidden.
All the other locks are leaf locks, that can be taken anywhere, but should
not be cascaded. Many functions, such as: xpt_action(), xpt_done(),
xpt_async(), xpt_create_path(), etc. are no longer require (but allow) SIM
lock to be held.
To keep compatibility and solve cases where SIM lock can't be dropped, all
xpt_async() calls in addition to xpt_done() calls are queued to completion
threads for async processing in clean environment without SIM lock held.
Instead of single CAM SWI thread, used for commands completion processing
before, use multiple (depending on number of CPUs) threads. Load balanced
between them using "hash" of the device B:T:L address.
HBA drivers that can drop SIM lock during completion processing and have
sufficient number of completion threads to efficiently scale to multiple
CPUs can use new function xpt_done_direct() to avoid extra context switch.
Make ahci(4) driver to use this mechanism depending on hardware setup.
Sponsored by: iXsystems, Inc.
MFC after: 2 months
r248917, r248918, r248978, r249001, r249014, r249030:
Remove multilevel freezing mechanism, implemented to handle specifics of
the ATA/SATA error recovery, when post-reset recovery commands should be
allocated when queues are already full of payload requests. Instead of
removing frozen CCBs with specified range of priorities from the queue
to provide free openings, use simple hack, allowing explicit CCBs over-
allocation for requests with priority higher (numerically lower) then
CAM_PRIORITY_OOB threshold.
Simplify CCB allocation logic by removing SIM-level allocation queue.
After that SIM-level queue manages only CCBs execution, while allocation
logic is localized within each single device.
Suggested by: gibbs
driver's periphs, acquiring and releaseing periph references while doing it.
Use it to iterate over the lists of ada and da periphs when flushing caches
and putting devices to sleep on shutdown and suspend. Previous code could
panic in theory if some device disappear in the middle of the process.
reporting. It includes:
- removing of error messages controlled by bootverbose, replacing them
with more universal and informative debugging on CAM_DEBUG_INFO level,
that is now built into the kernel by default;
- more close following to the arguments submitted by caller, such as
SF_PRINT_ALWAYS, SF_QUIET_IR and SF_NO_PRINT; consumer knows better which
errors are usual/expected at this point and which are really informative;
- adding two new flags SF_NO_RECOVERY and SF_NO_RETRY to allow caller
specify how much assistance it needs at this point; previously consumers
controlled that by not calling cam_periph_error() at all, but that made
behavior inconsistent and debugging complicated;
- tuning debug messages and taken actions order to make debugging output
more readable and cause-effect relationships visible;
- making camperiphdone() (common device recovery completion handler) to
also use cam_periph_error() in most cases, instead of own dumb code;
- removing manual sense fetching code from cam_periph_error(); I was told
by number of people that it is SIM obligation to fetch sense data, so this
code is useless and only significantly complicates recovery logic;
- making ada, da and pass driver to use cam_periph_error() with new limited
recovery options to handle error recovery and debugging in common way;
as one of results, CAM_REQUEUE_REQ and other retrying statuses are now
working fine with pass driver, that caused many problems before.
- reverting r186891 by raj@ to avoid burning few seconds in tight DELAY()
loops on device probe, while device simply loads media; I think that problem
may already be fixed in other way, and even if it is not, solution must be
different.
Sponsored by: iXsystems, Inc.
MFC after: 2 weeks
in the CAM XPT bus traversal code, and a number of other periph level
issues.
cam_periph.h,
cam_periph.c: Modify cam_periph_acquire() to test the CAM_PERIPH_INVALID
flag prior to allowing a reference count to be gained
on a peripheral. Callers of this function will receive
CAM_REQ_CMP_ERR status in the situation of attempting to
reference an invalidated periph. This guarantees that
a peripheral scheduled for a deferred free will not
be accessed during its wait for destruction.
Panic during attempts to drop a reference count on
a peripheral that already has a zero reference count.
In cam_periph_list(), use a local sbuf with SBUF_FIXEDLEN
set so that mallocs do not occur while the xpt topology
lock is held, regardless of the allocation policy of the
passed in sbuf.
Add a new routine, cam_periph_release_locked_buses(),
that can be called when the caller already holds
the CAM topology lock.
Add some extra debugging for duplicate peripheral
allocations in cam_periph_alloc().
Treat CAM_DEV_NOT_THERE much the same as a selection
timeout (AC_LOST_DEVICE is emitted), but forgo retries.
cam_xpt.c: Revamp the way the EDT traversal code does locking
and reference counting. This was broken, since it
assumed that the EDT would not change during
traversal, but that assumption is no longer valid.
So, to prevent devices from going away while we
traverse the EDT, make sure we properly lock
everything and hold references on devices that
we are using.
The two peripheral driver traversal routines should
be examined. xptpdperiphtraverse() holds the
topology lock for the entire time it runs.
xptperiphtraverse() is now locked properly, but
only holds the topology lock while it is traversing
the list, and not while the traversal function is
running.
The bus locking code in xptbustraverse() should
also be revisited at a later time, since it is
complex and should probably be simplified.
scsi_da.c: Pay attention to the return value from cam_periph_acquire().
Return 0 always from daclose() even if the disk is now gone.
Add some rudimentary error injection support.
scsi_sg.c: Fix reference counting in the sg(4) driver.
The sg driver was calling cam_periph_release() on close,
but never called cam_periph_acquire() (which increments
the reference count) on open.
The periph code correctly complained that the sg(4)
driver was trying to decrement the refcount when it
was already 0.
Sponsored by: Spectra Logic
MFC after: 2 weeks
other device attributes stored in the CAM Existing Device Table (EDT).
This includes some infrastructure requried by the enclosure services
driver to export physical path information.
Make the CAM device advanced info interface accept store requests.
sys/cam/scsi/scsi_all.c:
sys/cam/scsi/scsi_all.h:
- Replace scsi_get_sas_addr() with a scsi_get_devid() which takes
a callback that decides whether to accept a particular descriptor.
Provide callbacks for NAA IEEE Registered addresses and for SAS
addresses, replacing the old function. This is needed because
the old function doesn't work for an enclosure address for a SAS
device, which is not flagged as a SAS address, but is NAA IEEE
Registered. It may be worthwhile merging this interface with the
devid match interface.
- Add a few more defines for some device ID fields.
sbin/camcontrol/camcontrol.c:
- Update for the CCB_DEV_ADVINFO interface change.
cam/cam_xpt_internal.h:
- Add the new fields for the physical path string to the CAM EDT.
cam/cam_ccb.h:
- Rename CCB_GDEV_ADVINFO to simply CCB_DEV_ADVINFO, and the ccb
structure to ccb_dev_advinfo.
- Add a flag that changes this CCB's action to store, rather than
the default, retrieve.
- Add a new buffer type, CDAI_TYPE_PHYS_PATH, for the new CAM EDT
physpath field.
- Remove the never-implemented transport & proto flags.
cam/cam_xpt.c:
cam/cam_xpt.h:
- Add xpt_getattr(), which provides a wrapper for fetching a device's
attribute using the GEOM strings as key. This method currently
supports "GEOM::ident" and "GEOM::physpath".
Submitted by: will
Reviewed by : gibbs
Extend the XPT_DEV_MATCH api to allow a device search by device ID.
As far as the API is concerned, device ID is a binary blob to be
interpreted by the transport layer. The SCSI implementation assumes
it is an array of VPD device ID descriptors.
sys/cam/cam_ccb.h:
Create a new structure, device_id_match_pattern, and
update the XPT_DEV_MATCH datastructures and flags so
that this pattern type can be used.
sys/cam/cam_xpt.c:
- A single pattern matching on both inquiry data and device
ID is invalid. Report any violators.
- Pass device ID match requests through to the new routine
scsi_devid_match(). The direct call of a SCSI routine is
a layering violation, but no worse than the one a few
lines up that checks inquiry data. Defer cleaning this
up until our future, larger, rototilling of CAM.
- Zero out cam_ed and cam_et nodes on allocation. Prior to
this change, device_id_len and device_id were not inialized,
preventing proper detection of the presence of this
information.
sys/cam/scsi/scsi_all.c:
sys/cam/scsi/scsi_all.h:
Add the scsi_match_devid() routine.
Add a helper function for extracting peripherial driver names
sys/cam/cam_periph.c:
sys/cam/cam_periph.h:
Add the cam_periph_list() method which fills an sbuf
with a comma delimited list of the peripheral instances
associated with a given CAM path.
Add a helper functions for SCSI commands used by the SES driver.
sys/cam/scsi/scsi_all.c:
sys/cam/scsi/scsi_all.h:
Add structure definitions and csio filling functions for
the receive diagnostic results and send diagnostic commands.
Misc CAM XPT cleanups.
sys/cam/cam_xpt.c:
Broadcast AC_FOUND_DEVICE and AC_PATH_REGISTERED
events at the time async event handlers are attached
even when registering just for events on a partitular
SIM. Previously, you had to register for these
events on all SIMs in the system in order to get
the initial broadcast even though subsequent device
and path arrivals would be delivered.
sys/cam/cam_xpt.c:
Remove SIM mutex held asserts from path accessors.
CAM paths are reference counted and it is this
reference count, not the sim mutex, that garantees
they are stable.
Sponsored by: Spectra Logic Corporation
- Unify bus reset/probe sequence. Whenever bus attached at boot or later,
CAM will automatically reset and scan it. It allows to remove duplicate
code from many drivers.
- Any bus, attached before CAM completed it's boot-time initialization,
will equally join to the process, delaying boot if needed.
- New kern.cam.boot_delay loader tunable should help controllers that
are still unable to register their buses in time (such as slow USB/
PCCard/ CardBus devices), by adding one more event to wait on boot.
- To allow synchronization between different CAM levels, concept of
requests priorities was extended. Priorities now split between several
"run levels". Device can be freezed at specified level, allowing higher
priority requests to pass. For example, no payload requests allowed,
until PMP driver enable port. ATA XPT negotiate transfer parameters,
periph driver configure caching and so on.
- Frozen requests are no more counted by request allocation scheduler.
It fixes deadlocks, when frozen low priority payload requests occupying
slots, required by higher levels to manage theit execution.
- Two last changes were holding proper ATA reinitialization and error
recovery implementation. Now it is done: SATA controllers and Port
Multipliers now implement automatic hot-plug and should correctly
recover from timeouts and bus resets.
- Improve SCSI error recovery for devices on buses without automatic sense
reporting, such as ATAPI or USB. For example, it allows CAM to wait, while
CD drive loads disk, instead of immediately return error status.
- Decapitalize diagnostic messages and make them more readable and sensible.
- Teach PMP driver to limit maximum speed on fan-out ports.
- Make boot wait for PMP scan completes, and make rescan more reliable.
- Fix pass driver, to return CCB to user level in case of error.
- Increase number of retries in cd driver, as device may return several UAs.
- Remove CAM_PERIPH_POLLED flag. It is broken by design. Polling can't be
periph flag. May be SIM, may be CCB, but now it works fine just without it.
- Remove check unused for at least five years. If we will ever have non-BIO
devices in CAM, this check is smallest of what we will need.
- If several controllers complete requests same time, call swi_sched()
only once.
- Reduce code duplication in ATA XPT and PMP driver.
- Move PIO size setting from ada driver to ATA XPT. It is XPT business
to negotiate transfer details. ada driver is now stateless.
- Report PIO size to SIM. It is required for correct PATA SIM operation.
- Tune PMP scan timings. It workarounds some problems with SiI.
- If reset hapens during PMP initialization - restart it.
- Introduce early-initialized periph drivers, which are used during initial
scan process. Use it for xpt, probe, aprobe and pmp. It gives pmp chance
to finish scan before mountroot and numerate devices in right order.
int. All of its callers pass in cmd as a u_long, so this has
always been a dangerous type demotion. It was spooted by clang/llvm
trying to do a type promotion and sign extension within
cam_periph_ioctl.
Submitted by: rdivacky
use to synchornize and protect all data objects that are used for that
SIM. Drivers that are not yet MPSAFE register Giant and operate as
usual. RIght now, no drivers are MPSAFE, though a few will be changed
in the coming week as this work settles down.
The driver API has changed, so all CAM drivers will need to be recompiled.
The userland API has not changed, so tools like camcontrol do not need to
be recompiled.
for unknown events.
A number of modules return EINVAL in this instance, and I have left
those alone for now and instead taught MOD_QUIESCE to accept this
as "didn't do anything".
in geom_disk.c.
As a side effect this makes a lot of #include <sys/devicestat.h>
lines not needed and some biofinish() calls can be reduced to
biodone() again.
kernel. Justin agress that there is no other reasonable alternative to
do automatic rescans on connect.
The problem is that when a new device attaches to a SIM (SCSI host
controller) we need to send a XPT_SCAN_BUS command to the SIM using
xpt_action. This requires however that there is a peripheral available
to take the command (otherwise xpt_done and later bomb). The RESCAN
ioctl uses the same periph.
This enables a USB mass storage drive to do an automatic rescan on
connection of the drive.
The automatic dropping of a CAM entry on disconnection was already
working (asynchronous event).
The next thing to do is find someone to commit a change to vpo to do the
same thing. Just port umass_cam_rescan and friends across to that
driver.
Approved by: gibbs
is an application space macro and the applications are supposed to be free
to use it as they please (but cannot). This is consistant with the other
BSD's who made this change quite some time ago. More commits to come.
types allow the reporting of error counts and other statistics. Currently
we provide information on the last BDR or bus reset as well as active
transaction inforamtion, but this will be expanded as more information is
added to aid in error recovery.
Use the 'last reset' information to better handle bus settle delays.
Peripheral drivers now control whether a bus settle delay occurs and
for how long. This allows target mode peripheral drivers to avoid
having their device queue frozen by the XPT for what shoudl only be
initiator type behavior.
Don't perform a bus reset if the target device is incapable of performing
transfer negotiation (e.g. Fiber Channel).
If we don't perform a bus reset but the controller is capable of transfer
negotiations, force negotiations on the first transaction to go to the
device. This ensures that we aren't tripped up by a left over negotiation
from the prom, BIOS, loader, etc.
Add a default async handler funstion to cam_periph.c to remove duplicated
code in all initiator type peripheral drivers.
Allow mapping of XPT_CONT_TARGET_IO ccbs from userland. They are
itentical to XPT_SCSI_IO ccbs as far as data mapping is concerned.
to a device failed.
In theory, the same steps that happen when we get an AC_LOST_DEVICE async
notification should have been taken when a driver fails to attach. In
practice, that wasn't the case.
This only affected the da, cd and ch drivers, but the fix affects all
peripheral drivers.
There were several possible problems:
- In the da driver, we didn't remove the peripheral's softc from the da
driver's linked list of softcs. Once the peripheral and softc got
removed, we'd get a kernel panic the next time the timeout routine
called dasendorderedtag().
- In the da, cd and possibly ch drivers, we didn't remove the
peripheral's devstat structure from the devstat queue. Once the
peripheral and softc were removed, this could cause a panic if anyone
tried to access device statistics. (one component of the linked list
wouldn't exist anymore)
- In the cd driver, we didn't take the peripheral off the changer run
queue if it was scheduled to run. In practice, it's highly unlikely,
and maybe impossible that the peripheral would have been on the
changer run queue at that stage of the probe process.
The fix is:
- Add a new peripheral callback function (the "oninvalidate" function)
that is called the first time cam_periph_invalidate() is called for a
peripheral.
- Create new foooninvalidate() routines for each peripheral driver. This
routine is always called at splsoftcam(), and contains all the stuff
that used to be in the AC_LOST_DEVICE case of the async callback
handler.
- Move the devstat cleanup call to the destructor/cleanup routines, since
some of the drivers do I/O in their close routines.
- Make sure that when we're flushing the buffer queue, we traverse it at
splbio().
- Add a check for the invalid flag in the pt driver's open routine.
Reviewed by: gibbs
one error recovery action oustanding for a given peripheral.
This is bad for several reasons. The first problem is that the error
recovery actions would likely be to fix the same problem. (e.g., we
queue 5 CCBs to a disk, and the first one comes back with 0x04,0x02. We
start error recovery, and the second one comes back with the same status.
Then the third one comes back, and so on. Each one causes the drive to get
nailed with a start unit, when we really only need one.)
The other problem is that we only have space to store one CCB while we're
doing error recovery. The subsequent error recovery actions that got
started were over-writing the CCBs from previous error recovery actions,
but we still tried to call the done routine N times for N error recovery
actions. Each call to dadone() was done with the same CCB, though. So on
the second one, we got a "biodone: buffer not busy" panic, since the buffer
in question had already been through biodone().
In any case, this fixes things so that any any given time, there's only one
error recovery action outstanding for any given peripheral driver.
Reviewed by: gibbs
Reported by: Philippe Regnauld <regnauld@deepo.prosa.dk>
[ Philippe wins the "bug finder of the week" award ]
