Currently driver maintain one list for both pending user interrupts
which seeks to wait till CQ reaches it's target value and also the ones
that seeks to get timestamp records when the CQ reaches it's target
value.
This causes delay in handling the waiters which gets higher priority
than the timestamp records.
In order to solve this, let's split the list into two,
one for each case and each one is protected by it's own spinlock.
Waiters will be handled within the interrupt context first,
then the timestamp records will be set.
Freeing the timestamp related memory will be handled in a workqueue.
Signed-off-by: farah kassabri <fkassabri@habana.ai>
Reviewed-by: Tomer Tayar <ttayar@habana.ai>
Signed-off-by: Oded Gabbay <ogabbay@kernel.org>
Currently we use dynamic allocation inside the irq handler
in order to allocate free node to be used for the free jobs.
This operation is expensive, especially when we deal with large
burst of events records that get released at the same time.
The alternative is to have pre allocated pool of free nodes
and just fetch nodes from this pool at irq handling time instead
of allocating them.
In case the pool becomes full, then the driver will fallback to
dynamic allocations.
As part of the optimization also update the unregister flow
upon re-using a timestamp record, by making the operation much
simpler and quicker. We already have the record in the registration
flow and now we just seek to re-use with different interrupt.
Therefore, no need to look for buffer according to the user handle.
Signed-off-by: farah kassabri <fkassabri@habana.ai>
Reviewed-by: Tomer Tayar <ttayar@habana.ai>
Signed-off-by: Oded Gabbay <ogabbay@kernel.org>
There is a potential race between user thread seeking to re-use
a timestamp record with new interrupt id, while this record is still
in the middle of interrupt handling and it is about to be freed.
Imagine the driver set the record in_use to 0 and only then fill the
free_node information. This might lead to unpleasant scenario where
the new registration thread detects the record as free to use, and
change the cq buff address. That will cause the free_node to get
the wrong buffer address to put refcount to.
Signed-off-by: farah kassabri <fkassabri@habana.ai>
Reviewed-by: Oded Gabbay <ogabbay@kernel.org>
Signed-off-by: Oded Gabbay <ogabbay@kernel.org>
alloc_sgt_from_device_pages() includes relatively many parameters, and
in a subsequent change another offset parameter is going to be added.
Using structure fields directly when calling this function, and in
hl_map_dmabuf() it is done twice, makes it a little bit difficult to
understand the meaning of the parameters.
To make it clearer, assign the required values into local variables with
explicit names, and use the variables when calling the function.
Signed-off-by: Tomer Tayar <ttayar@habana.ai>
Reviewed-by: Oded Gabbay <ogabbay@kernel.org>
Signed-off-by: Oded Gabbay <ogabbay@kernel.org>
It is currently allowed for a user to export dma-buf with size and
offset that are not multiples of PAGE_SIZE.
The exported memory is mapped for the importer device, and there it will
be rounded to PAGE_SIZE, leading to actually exporting more than the
user intended to.
To make the user be aware of it, accept only size and offset which are
multiple of PAGE_SIZE.
Signed-off-by: Tomer Tayar <ttayar@habana.ai>
Reviewed-by: Oded Gabbay <ogabbay@kernel.org>
Signed-off-by: Oded Gabbay <ogabbay@kernel.org>
The 'exported_size' member in 'struct hl_vm_phys_pg_pack' is used to
keep the exported dma-buf size, to be later used when the buffer is
mapped.
However it is possible that the same phys_pg_pack will be exported more
than once, and independently of when the mapping takes place.
Remove this member from the phys_pg_pack structure, and simply use the
size in the dma-buf object as the exported size when mapping.
Signed-off-by: Tomer Tayar <ttayar@habana.ai>
Reviewed-by: Oded Gabbay <ogabbay@kernel.org>
Signed-off-by: Oded Gabbay <ogabbay@kernel.org>
For Gaudi1 the exported dma-buf is always composed of a single page, and
therefore the exported size is equal to this page's size.
When calling alloc_sgt_from_device_pages(), we pass 0 as the exported
size and internally calculate it as "number of pages * page size".
This makes alloc_sgt_from_device_pages() less clear, because the
exported size parameter is not understood as a restriction on the pages'
size.
Modify to always pass the exported size explicitly.
Signed-off-by: Tomer Tayar <ttayar@habana.ai>
Reviewed-by: Oded Gabbay <ogabbay@kernel.org>
Signed-off-by: Oded Gabbay <ogabbay@kernel.org>
After the heartbeat mechanism is now expanded to be used also
for EQ health check, we shouldn't send heartbeat messages
to FW before driver allow events to be received from FW.
Because if the driver will send two heartbeats before it enables
events to be received from FW, then the EQ health check
will fail and reset the device.
Signed-off-by: farah kassabri <fkassabri@habana.ai>
Reviewed-by: Oded Gabbay <ogabbay@kernel.org>
Signed-off-by: Oded Gabbay <ogabbay@kernel.org>
This is the second patch for applying the eq health check mechanism
which will add support for the interrupt flow for gaudi2 asic.
More info about the interrupt mechanism:
set a dedicated msix for the eq error interrupt, and add
interrupt handler for it.
when FW detects some issue with EQ like EQ_FULL, it'll
raise that interrupt and driver should reset the device.
Driver will inform the FW which msix index to use through
the already existing handshake mechanism which will
send msix info message to fw.
Signed-off-by: farah kassabri <fkassabri@habana.ai>
Reviewed-by: Oded Gabbay <ogabbay@kernel.org>
Signed-off-by: Oded Gabbay <ogabbay@kernel.org>
Add mechanism for fw eq health check. this will be done using two flows:
using the heartbeat mechanism and raising a dedicated interrupt to
indicate an eq failure like EQ full.
This patch will add implementation for the eq heartbeat for gaudi2 asic.
More info about the heartbeat mechanism:
Expand the heartbeat mechanism to monitor a new event that
will be sent from FW upon receiving heartbeat message.
that way driver can know that the eq is working or not.
Signed-off-by: farah kassabri <fkassabri@habana.ai>
Reviewed-by: Oded Gabbay <ogabbay@kernel.org>
Signed-off-by: Oded Gabbay <ogabbay@kernel.org>
Add tsc clock to clock sync info, to enable using this clock for
sampling and sync it with device time.
Signed-off-by: Hen Alon <halon@habana.ai>
Reviewed-by: Oded Gabbay <ogabbay@kernel.org>
Signed-off-by: Oded Gabbay <ogabbay@kernel.org>
Because it is not used and also, for graceful reset to work
those ioctls should run on the compute device.
Signed-off-by: Dafna Hirschfeld <dhirschfeld@habana.ai>
Reviewed-by: Oded Gabbay <ogabbay@kernel.org>
Signed-off-by: Oded Gabbay <ogabbay@kernel.org>
The CPUCP interface is moved to a shared folder outside of accel as
a pre-requisite to upstream the NIC drivers that will also include
this file.
Signed-off-by: David Meriin <dmeriin@habana.ai>
Reviewed-by: Oded Gabbay <ogabbay@kernel.org>
Signed-off-by: Oded Gabbay <ogabbay@kernel.org>
To use drm_ioctl(), move the ioctls to the device specific ioctls
range at [DRM_COMMAND_BASE, DRM_COMMAND_END).
Signed-off-by: Tomer Tayar <ttayar@habana.ai>
Reviewed-by: Oded Gabbay <ogabbay@kernel.org>
Signed-off-by: Oded Gabbay <ogabbay@kernel.org>
Register the compute device as an accel device, and remove the creation
of the habanalabs compute char device.
The IOCTLs in this patch are still handled by the current driver
handler. Moving to DRM IOCTL handling requires moving the IOCTLs
numbers to a specific range, so it will be handled in subsequent
patches.
Signed-off-by: Tomer Tayar <ttayar@habana.ai>
Reviewed-by: Oded Gabbay <ogabbay@kernel.org>
Signed-off-by: Oded Gabbay <ogabbay@kernel.org>
User gets notification for every engine error report, but he still
lacks the exact engine information. Hence, we allow user to query
for the exact engine reported an error.
Signed-off-by: Ofir Bitton <obitton@habana.ai>
Reviewed-by: Oded Gabbay <ogabbay@kernel.org>
Signed-off-by: Oded Gabbay <ogabbay@kernel.org>
After being notified about certain errors, user is expected to finish
his post-errors actions and to release the device within some timeout,
after which is deice is being reset.
The default timeout value is 5 sec, which in some case is not enough for
a user application to collect debug data.
Increase the default value to 30 sec.
Signed-off-by: Tomer Tayar <ttayar@habana.ai>
Reviewed-by: Oded Gabbay <ogabbay@kernel.org>
Signed-off-by: Oded Gabbay <ogabbay@kernel.org>
It is possible for FW to request reserved space in dram.
If the device supports this option, it will retrieve the size from the
f/w and will reserve it.
Currently we add the common code infrastructure to support it.
Signed-off-by: Dani Liberman <dliberman@habana.ai>
Reviewed-by: Oded Gabbay <ogabbay@kernel.org>
Signed-off-by: Oded Gabbay <ogabbay@kernel.org>
Add dump of an error reported from f/w during boot time.
This error indicates a failure with setting temperature threshold.
Signed-off-by: Ofir Bitton <obitton@habana.ai>
Reviewed-by: Oded Gabbay <ogabbay@kernel.org>
Signed-off-by: Oded Gabbay <ogabbay@kernel.org>
If scrubbing memory after user released device has failed it means
the device is in a bad state and should be reset.
Signed-off-by: Oded Gabbay <ogabbay@kernel.org>
Reviewed-by: Ofir Bitton <obitton@habana.ai>
Our simulator supports idle check so no need anymore to check if pdev
exists.
Signed-off-by: Oded Gabbay <ogabbay@kernel.org>
Reviewed-by: Ofir Bitton <obitton@habana.ai>
When the driver needs to abort waiters for interrupts, for cases
such as critical events that occur and driver need to do hard reset,
in such scenario the driver will complete the fence to wake up the
waiting thread, and will set the fence error indication.
The return value of the completion API will be greater than 0
since it will return the timeout, but as this indicates successful
completion, the driver should mark it as aborted.
Signed-off-by: farah kassabri <fkassabri@habana.ai>
Reviewed-by: Oded Gabbay <ogabbay@kernel.org>
Signed-off-by: Oded Gabbay <ogabbay@kernel.org>
Protect against concurrency of user requesting to register a timestamp
offset (where the driver fills the timestamp when the command submission
has finished executing) to a specific user interrupt ID. The
protection is basically to allow only one timestamp registration
request to be handled at a time.
This is needed because the user can decide to re-use a timestamp
offset (register an already registered offset, to a different
interrupt ID). This means the request will cause the timestamp node to
move from one interrupt list to another interrupt list. In such
scenario, without proper protection, we could end up adding the same
node twice to the interrupts wait lists.
Signed-off-by: farah kassabri <fkassabri@habana.ai>
Reviewed-by: Oded Gabbay <ogabbay@kernel.org>
Signed-off-by: Oded Gabbay <ogabbay@kernel.org>
Every time an FD is returned to the user, the driver adds
a corresponding private structure to the list.
Yet, it's still a list of private structures rather than of FDs.
Remove, as well, an unnecessary comment.
Signed-off-by: Koby Elbaz <kelbaz@habana.ai>
Reviewed-by: Oded Gabbay <ogabbay@kernel.org>
Signed-off-by: Oded Gabbay <ogabbay@kernel.org>
Because we might still be using related resources, decrementing PID's
reference count should be done at later stages of the device release.
A good place is right after the representing private structure is
removed from LKD's list.
Signed-off-by: Koby Elbaz <kelbaz@habana.ai>
Reviewed-by: Oded Gabbay <ogabbay@kernel.org>
Signed-off-by: Oded Gabbay <ogabbay@kernel.org>
As part of driver teardown, we attempt to kill all user processes.
It shouldn't fail, but if it does we want to print the error code that
the kapi returned to us.
Signed-off-by: Koby Elbaz <kelbaz@habana.ai>
Reviewed-by: Oded Gabbay <ogabbay@kernel.org>
Signed-off-by: Oded Gabbay <ogabbay@kernel.org>
The preboot used to statically allocate memory for the comms descriptor
on the device memory when driver requested the descriptor information.
Now preboot moved to dynamic memory allocation where it wants to check
the size the driver expects vs. what the f/w expects.
Note there are no backward compatibility issues as older f/w versions
simply ignore this value.
Signed-off-by: farah kassabri <fkassabri@habana.ai>
Reviewed-by: Oded Gabbay <ogabbay@kernel.org>
Signed-off-by: Oded Gabbay <ogabbay@kernel.org>
hl_device_status() returns the status of an acquired device.
If a device is going down (following an rmmod cmd),
it should be marked as an unusable/malfunctioning device, and
hence should not be acquired.
However, since this was not the case so far (i.e., a device going
down would inaccurately return 'in reset' status allowing the user
to acquire the device) it introduced a bug where as part of a reset
flow, the driver could not kill processes that have not run yet, and
since those processes aren't blocked from reacquiring a device,
we get eventually a new flow of a driver attempting to kill all
processes in a list that can't be ever really empty.
Signed-off-by: Koby Elbaz <kelbaz@habana.ai>
Reviewed-by: Oded Gabbay <ogabbay@kernel.org>
Signed-off-by: Oded Gabbay <ogabbay@kernel.org>
It is useful for debug to know which user process have acquired the
device.
Add this info to the relevant debug print, in addition to the already
printed user context's ASID.
Signed-off-by: Tomer Tayar <ttayar@habana.ai>
Reviewed-by: Oded Gabbay <ogabbay@kernel.org>
Signed-off-by: Oded Gabbay <ogabbay@kernel.org>
When an ioctl fails, it is useful to know what is the task command name
and the full ioctl request code, in addition to the task pid and the
ioctl number.
Add the additional information to the relevant debug error prints.
Signed-off-by: Tomer Tayar <ttayar@habana.ai>
Reviewed-by: Oded Gabbay <ogabbay@kernel.org>
Signed-off-by: Oded Gabbay <ogabbay@kernel.org>
If hl_device_cond_reset() is called while a reset is already pending but
hasn't started, the reset request will be dropped.
If the flags of the new request are more severe, e.g. a hard reset while
the pending reset is a compute reset, the eventual reset won't be
suitable for the device status.
To prevent such cases, update the pending reset flags with the new
requests flags before the requests are dropped.
Signed-off-by: Tomer Tayar <ttayar@habana.ai>
Reviewed-by: Oded Gabbay <ogabbay@kernel.org>
Signed-off-by: Oded Gabbay <ogabbay@kernel.org>
When a H/W event is received while a user is registered to events, no
immediate hard reset will happen, and instead the user will be notified
and will have some time to handle it and eventually release the
device, after which the reset will be done.
If a user, as part of the handling and as part of the cleanup steps
towards releasing the device, unregisters from receiving those events,
and at that time an adjacent H/W event is received, it will be assumed
that the user is not registered to events and thus an immediate hard
reset is required.
To prevent such an unwanted immediate reset, modify the driver to
perform it if the user is not registered to events AND we don't already
have a pending reset for a previous H/W event.
Signed-off-by: Tomer Tayar <ttayar@habana.ai>
Reviewed-by: Oded Gabbay <ogabbay@kernel.org>
Signed-off-by: Oded Gabbay <ogabbay@kernel.org>
Two functions got added with normal prototypes for debugfs, but not
alternative when building without it:
drivers/accel/habanalabs/common/device.c: In function 'hl_device_init':
drivers/accel/habanalabs/common/device.c:2177:14: error: implicit declaration of function 'hl_debugfs_device_init'; did you mean 'hl_debugfs_init'? [-Werror=implicit-function-declaration]
drivers/accel/habanalabs/common/device.c:2305:9: error: implicit declaration of function 'hl_debugfs_device_fini'; did you mean 'hl_debugfs_remove_file'? [-Werror=implicit-function-declaration]
Add stubs for these as well.
Fixes: 3b9abb4fa6 ("accel/habanalabs: expose debugfs files later")
Signed-off-by: Arnd Bergmann <arnd@arndb.de>
Reviewed-by: Tomer Tayar <ttayar@habana.ai>
Acked-by: Oded Gabbay <ogabbay@kernel.org>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Link: https://patchwork.freedesktop.org/patch/msgid/20230609120636.3969045-1-arnd@kernel.org
Moved error info reset code to single function for future use from
other places in the driver.
Signed-off-by: Dani Liberman <dliberman@habana.ai>
Reviewed-by: Oded Gabbay <ogabbay@kernel.org>
Signed-off-by: Oded Gabbay <ogabbay@kernel.org>
In order to increase reliability of the event queue interface,
we apply to Gaudi2 the same mechanism we have in Gaudi1.
The extra validation is basically checking that the received
event index matches the expected index.
Signed-off-by: Ofir Bitton <obitton@habana.ai>
Reviewed-by: Oded Gabbay <ogabbay@kernel.org>
Signed-off-by: Oded Gabbay <ogabbay@kernel.org>
We don't want to allow users to spam the kernel log and sending
ioctls with bad opcodes is a sure way to do it.
Signed-off-by: Oded Gabbay <ogabbay@kernel.org>
Reviewed-by: Ofir Bitton <obitton@habana.ai>
There were a few places where simulator only code got into the upstream.
Remove those places that can confuse other developers.
Fixes: 2a0a839b6a ("habanalabs: extend fatal messages to contain PCI info")
Cc: Moti Haimovski <mhaimovski@habana.ai>
Cc: Dan Carpenter <dan.carpenter@linaro.org>
Signed-off-by: Oded Gabbay <ogabbay@kernel.org>
Any FW component we load must be followed by a corresponding state
update. However, it seems that so far we skipped doing so for the
bootfit case, so fix that.
Signed-off-by: Koby Elbaz <kelbaz@habana.ai>
Reviewed-by: Oded Gabbay <ogabbay@kernel.org>
Signed-off-by: Oded Gabbay <ogabbay@kernel.org>
As mmu disable mode is only used for bring-up stages, let's remove this
option and all code related to it.
Signed-off-by: Ofir Bitton <obitton@habana.ai>
Reviewed-by: Oded Gabbay <ogabbay@kernel.org>
Signed-off-by: Oded Gabbay <ogabbay@kernel.org>
If a workload got stuck, we print an error to the kernel log about it.
Add to that print the configured max timeout value, as that value is
not fixed between ASICs and in addition it can be configured using
a kernel module parameter.
Signed-off-by: Oded Gabbay <ogabbay@kernel.org>
Reviewed-by: Ofir Bitton <obitton@habana.ai>
This commit fixes a memory leak caused when clearing the user_mappings
info when a new context is opened immediately after user_mapping is
captured and a hard reset is performed.
Signed-off-by: Moti Haimovski <mhaimovski@habana.ai>
Reviewed-by: Dani Liberman <dliberman@habana.ai>
Reviewed-by: Oded Gabbay <ogabbay@kernel.org>
Signed-off-by: Oded Gabbay <ogabbay@kernel.org>
Make the argument names specify the registers array represent
registers that should be unsecured so the user can access them.
Signed-off-by: Koby Elbaz <kelbaz@habana.ai>
Reviewed-by: Oded Gabbay <ogabbay@kernel.org>
Signed-off-by: Oded Gabbay <ogabbay@kernel.org>
Make the description of @regs_range_array and @regs_range_array_size
to @user_regs_range_array and @user_regs_range_array_size to silence
the warnings:
drivers/accel/habanalabs/common/security.c:506: warning: Function parameter or member 'user_regs_range_array' not described in 'hl_init_pb_ranges_single_dcore'
drivers/accel/habanalabs/common/security.c:506: warning: Function parameter or member 'user_regs_range_array_size' not described in 'hl_init_pb_ranges_single_dcore'
drivers/accel/habanalabs/common/security.c:506: warning: Excess function parameter 'regs_range_array' description in 'hl_init_pb_ranges_single_dcore'
drivers/accel/habanalabs/common/security.c:506: warning: Excess function parameter 'regs_range_array_size' description in 'hl_init_pb_ranges_single_dcore'
Reported-by: Abaci Robot <abaci@linux.alibaba.com>
Closes: https://bugzilla.openanolis.cn/show_bug.cgi?id=4940
Signed-off-by: Yang Li <yang.lee@linux.alibaba.com>
Reviewed-by: Oded Gabbay <ogabbay@kernel.org>
Signed-off-by: Oded Gabbay <ogabbay@kernel.org>
Currently, we rely on COMMS protocol's ack to verify that WFE command
has been acknowledged by the FW. However, this does not guarantee that
the device status has been updated.
Although unlikely, this could trigger a race since the driver expects
the device to be halted at that stage, but it might not be.
Therefore, we increase WFE's robustness by polling on the status
register that will be updated once the device is actually halted.
Signed-off-by: Koby Elbaz <kelbaz@habana.ai>
Reviewed-by: Oded Gabbay <ogabbay@kernel.org>
Signed-off-by: Oded Gabbay <ogabbay@kernel.org>
Currently the debugfs root folder and files for a device are created at
an early step, before the device initialization and before the char
device and sysfs files are exposed to user.
As there is no real reason not to do it together with the device
creation, postpone it to be done right afterwards.
The initialization of the debugfs entry structure is left in its
current position because it is used before creating the files.
Signed-off-by: Tomer Tayar <ttayar@habana.ai>
Reviewed-by: Oded Gabbay <ogabbay@kernel.org>
Signed-off-by: Oded Gabbay <ogabbay@kernel.org>
Currently upon a heartbeat failure, we don't know if the failure
is due to firmware hang or due to a bad PCI link. Hence, we
are reading a PCI config space register with a known value (vendor ID)
so we will know which of the two possibilities caused the heartbeat
failure.
Signed-off-by: Ofir Bitton <obitton@habana.ai>
Reviewed-by: Oded Gabbay <ogabbay@kernel.org>
Signed-off-by: Oded Gabbay <ogabbay@kernel.org>
Aborting CS completions should be in command_submission.c but aborting
waiting for user interrupts should be in device.c.
This separation is also for adding more abort operations in the future.
Signed-off-by: Koby Elbaz <kelbaz@habana.ai>
Reviewed-by: Oded Gabbay <ogabbay@kernel.org>
Signed-off-by: Oded Gabbay <ogabbay@kernel.org>
Sync Stream Encapsulated Signal Handlers can be managed from different
contexts, and as such they are protected via a spin_lock.
However, spin_lock was unnecessarily protecting a larger code section
than really needed, covering a sleepable code section as well.
Since spin_lock disables preemption, it could lead to sleeping in
atomic context.
Signed-off-by: Koby Elbaz <kelbaz@habana.ai>
Reviewed-by: Oded Gabbay <ogabbay@kernel.org>
Signed-off-by: Oded Gabbay <ogabbay@kernel.org>
