If device memory scrubbing from hl_device_reset() fails, we return with
an error code but not perform error handling code.
Signed-off-by: Tomer Tayar <ttayar@habana.ai>
Reviewed-by: Oded Gabbay <ogabbay@kernel.org>
Signed-off-by: Oded Gabbay <ogabbay@kernel.org>
This commit enhances the following error messages to also provide the
type of error occurred, this in order to ease debugging of errors
detected during firmware-load.
Signed-off-by: Moti Haimovski <mhaimovski@habana.ai>
Reviewed-by: Oded Gabbay <ogabbay@kernel.org>
Signed-off-by: Oded Gabbay <ogabbay@kernel.org>
Completion timestamp is taken during the actual command submission
release. As the release happens in a work queue, the timestamp taken
is not accurate. Hence, we will take the timestamp in the interrupt
handler itself while propagating it to the release function.
Signed-off-by: Ofir Bitton <obitton@habana.ai>
Reviewed-by: Oded Gabbay <ogabbay@kernel.org>
Signed-off-by: Oded Gabbay <ogabbay@kernel.org>
In order to support more user interrupt types in the future, we
enumerate the user interrupt type instead of using a boolean.
Signed-off-by: Ofir Bitton <obitton@habana.ai>
Reviewed-by: Oded Gabbay <ogabbay@kernel.org>
Signed-off-by: Oded Gabbay <ogabbay@kernel.org>
Handling edma razwi is different than all other engines since edma
uses sft routers. For hbw transactions sft router contain separate
interface for each edma and for lbw there is common interface for
both edma engines of the same dcore.
To handle the razwi correctly we need to:
1. Simplify the calculation of the sft router address.
2. Add razwi handling for edma qm errors, since edma qman doesn't
reports axi error response.
Signed-off-by: Dani Liberman <dliberman@habana.ai>
Reviewed-by: Oded Gabbay <ogabbay@kernel.org>
Signed-off-by: Oded Gabbay <ogabbay@kernel.org>
A device with status malfunction indicates that it can't be used.
In such a case we do not support certain reset types, e.g.,
all kinds of soft-resets (compute reset, inference soft-reset),
and reset upon device release.
A hard-reset is the only way that an unusable device can change its
status. All other reset procedures can't put the device in a reset
procedure, which might ultimately cause the device to change its
status, unintentionally, to become operational again.
Such a scenario has recently occurred, when a user requested
a hard-reset while another heavy user workload was ongoing (reset
request is queued).
Since the workload couldn't finish within reset's timeout limits, the
reset has failed and set a device status malfunction.
Eventually, when the user released the FD, an unsuccessful soft-reset
occurred, hence followed by an additional hard-reset that changed the
ASICs status back to be operational.
Signed-off-by: Koby Elbaz <kelbaz@habana.ai>
Reviewed-by: Oded Gabbay <ogabbay@kernel.org>
Signed-off-by: Oded Gabbay <ogabbay@kernel.org>
AXI transaction id holds information about the initiator which caused
the page fault. In the future it will be translated automatically by
driver to an initiator name.
Signed-off-by: Dani Liberman <dliberman@habana.ai>
Reviewed-by: Oded Gabbay <ogabbay@kernel.org>
Signed-off-by: Oded Gabbay <ogabbay@kernel.org>
Captured addresses of low b/w razwi information contains only the
offset from the cfg base. To make it more user readable, add the cfg
base to it.
Signed-off-by: Dani Liberman <dliberman@habana.ai>
Reviewed-by: Oded Gabbay <ogabbay@kernel.org>
Signed-off-by: Oded Gabbay <ogabbay@kernel.org>
In gaudi2 there night be different routers for low b/w and high b/w
transactions. But in the code that collects razwi information, we used
the same router for high b/w and low b/w.
Fixed it by reading the information also from low b/w routers.
Signed-off-by: Dani Liberman <dliberman@habana.ai>
Reviewed-by: Oded Gabbay <ogabbay@kernel.org>
Signed-off-by: Oded Gabbay <ogabbay@kernel.org>
Protect re-using the same timestamp buffer record before actually
adding it to the to interrupt wait list.
Mark ts buff offset as in use in the spinlock protection area of the
interrupt wait list to avoid getting in the re-use section in
ts_buff_get_kernel_ts_record before adding the node to the list.
this scenario might happen when multiple threads are racing on
same offset and one thread could set data in the ts buff in
ts_buff_get_kernel_ts_record then the other thread takes over
and get to ts_buff_get_kernel_ts_record and we will try
to re-use the same ts buff offset then we will try to
delete a non existing node from the list.
Signed-off-by: farah kassabri <fkassabri@habana.ai>
Reviewed-by: Oded Gabbay <ogabbay@kernel.org>
Signed-off-by: Oded Gabbay <ogabbay@kernel.org>
use argument instead of fixed GFP value for allocation
in Timestamps buffers alloc function.
change data type of size to size_t.
Fixes: 9158bf69e7 ("habanalabs: Timestamps buffers registration")
Signed-off-by: farah kassabri <fkassabri@habana.ai>
Reviewed-by: Oded Gabbay <ogabbay@kernel.org>
Signed-off-by: Oded Gabbay <ogabbay@kernel.org>
Make sure all reserved/pad fields in uapi input structures
are set to 0.
Signed-off-by: farah kassabri <fkassabri@habana.ai>
Reviewed-by: Oded Gabbay <ogabbay@kernel.org>
Signed-off-by: Oded Gabbay <ogabbay@kernel.org>
data is a void * type and does not require a cast.
Signed-off-by: XU pengfei <xupengfei@nfschina.com>
Reviewed-by: Oded Gabbay <ogabbay@kernel.org>
Signed-off-by: Oded Gabbay <ogabbay@kernel.org>
This commit attaches the PCI device address to driver fatal messages
in order to ease debugging in multi-device setups.
Signed-off-by: Moti Haimovski <mhaimovski@habana.ai>
Reviewed-by: Oded Gabbay <ogabbay@kernel.org>
Signed-off-by: Oded Gabbay <ogabbay@kernel.org>
Because f/w does not update razwi info when sending events, remove the
use of it.
The driver is responsible to check if razwi happened and to
collect razwi data.
Signed-off-by: Dani Liberman <dliberman@habana.ai>
Reviewed-by: Oded Gabbay <ogabbay@kernel.org>
Signed-off-by: Oded Gabbay <ogabbay@kernel.org>
Add traces to LBW reads/writes.
This may be handy when debugging configuration failure or events when
tracking configuration flow.
Signed-off-by: Ohad Sharabi <osharabi@habana.ai>
Reviewed-by: Oded Gabbay <ogabbay@kernel.org>
Signed-off-by: Oded Gabbay <ogabbay@kernel.org>
The value in SM_SEI_CAUSE includes the SOB index and not the SOB group
index.
Remove usage of log_mask in sm_sei_cause structure as it was never
used.
Signed-off-by: Carmit Carmel <ccarmel@habana.ai>
Reviewed-by: Oded Gabbay <ogabbay@kernel.org>
Signed-off-by: Oded Gabbay <ogabbay@kernel.org>
This function shall be used whenever components enable/binning masks
should be updated.
Usage is in one of the below cases:
- update user (or default) component masks
- update when getting the masks from FW (either CPUCP or COMMS)
Signed-off-by: Ohad Sharabi <osharabi@habana.ai>
Reviewed-by: Oded Gabbay <ogabbay@kernel.org>
Signed-off-by: Oded Gabbay <ogabbay@kernel.org>
When HL_INFO_USER_MAPPINGS IOCTL is called, we copy_to_user from
a dynamically allocated memory - 'user_mappings'.
Since freeing/allocating it happens in runtime (upon a page fault),
it not unlikely to access it even before being initially allocated
(i.e., accessing a NULL pointer).
The solution is to simply mark the spot when the err info has been
collected, and that way to know whether err info (either page fault
or RAZWI) is available to be read.
Signed-off-by: Koby Elbaz <kelbaz@habana.ai>
Reviewed-by: Oded Gabbay <ogabbay@kernel.org>
Signed-off-by: Oded Gabbay <ogabbay@kernel.org>
From reviewing the code, the line
memset(kdata, 0, usize);
is not needed because kdata is either zeroed by
kdata = kzalloc(asize, GFP_KERNEL);
when allocated at runtime or by
char stack_kdata[128] = {0};
at compile time.
Signed-off-by: Tom Rix <trix@redhat.com>
Reviewed-by: Oded Gabbay <ogabbay@kernel.org>
Signed-off-by: Oded Gabbay <ogabbay@kernel.org>
This refactor makes the code clearer and the new variables' names
better describe their roles.
Signed-off-by: Koby Elbaz <kelbaz@habana.ai>
Reviewed-by: Oded Gabbay <ogabbay@kernel.org>
Signed-off-by: Oded Gabbay <ogabbay@kernel.org>
It appears that, within the sync manager security configuration,
we reconfigure PB registers over and over without any need to do that.
Signed-off-by: Koby Elbaz <kelbaz@habana.ai>
Reviewed-by: Oded Gabbay <ogabbay@kernel.org>
Signed-off-by: Oded Gabbay <ogabbay@kernel.org>
During device acquire, the driver is using a QMAN for clearing some
registers. In order to avoid internal races, the driver verifies
the device is idle before submitting the register clear job.
This check introduces an issue, as debug mode will cause the device
to be non-idle which will lead to device acquire failure.
In order to overcome this issue we can entirely remove the idle
check as the driver is using the QMAN only when there is no active
context.
Signed-off-by: Ofir Bitton <obitton@habana.ai>
Reviewed-by: Oded Gabbay <ogabbay@kernel.org>
Signed-off-by: Oded Gabbay <ogabbay@kernel.org>
When entering an IOCTL, the driver prints a message in case device is
not operational. This message should be printed in debug level as
it can spam the kernel log and it is not an error.
Signed-off-by: Oded Gabbay <ogabbay@kernel.org>
Remove the distinction between user CB and kernel CB, and verify for
both that they are not destroyed more than once.
As kernel CB might be taken from the pre-allocated CB pool, so we need
to clear the handle destroyed indication when returning a CB to the
pool.
Signed-off-by: Tomer Tayar <ttayar@habana.ai>
Reviewed-by: Oded Gabbay <ogabbay@kernel.org>
Signed-off-by: Oded Gabbay <ogabbay@kernel.org>
When doing p2p with a NIC device, the NIC needs to make sure all the
writes to the HBM (through the PCI bar of the Gaudi device) were
flushed.
It can be done by either the NIC or the host reading through the PCI
bar.
To support the host side, we supply a simple uapi to perform this flush
through the driver, because the user can't create such a transaction
by itself (the PCI bar isn't exposed to normal users).
Signed-off-by: Ohad Sharabi <osharabi@habana.ai>
Reviewed-by: Oded Gabbay <ogabbay@kernel.org>
Signed-off-by: Oded Gabbay <ogabbay@kernel.org>
Now that we have a subsystem for compute accelerators, move the
habanalabs driver to it.
This patch only moves the files and fixes the Makefiles. Future
patches will change the existing code to register to the accel
subsystem and expose the accel device char files instead of the
habanalabs device char files.
Update the MAINTAINERS file to reflect this change.
Signed-off-by: Oded Gabbay <ogabbay@kernel.org>
Each of the user contexts has two command queues, one for compute engine
and one for the copy engine. Command queues are allocated and registered
in the device when the first job (command buffer) is submitted from
the user space to the VPU device. The userspace provides a list of
GEM buffer object handles to submit to the VPU, the driver resolves
buffer handles, pins physical memory if needed, increments ref count
for each buffer and stores pointers to buffer objects in
the ivpu_job objects that track jobs submitted to the device.
The VPU signals job completion with an asynchronous message that
contains the job id passed to firmware when the job was submitted.
Currently, the driver supports simple scheduling logic
where jobs submitted from user space are immediately pushed
to the VPU device command queues. In the future, it will be
extended to use hardware base scheduling and/or drm_sched.
Co-developed-by: Andrzej Kacprowski <andrzej.kacprowski@linux.intel.com>
Signed-off-by: Andrzej Kacprowski <andrzej.kacprowski@linux.intel.com>
Signed-off-by: Jacek Lawrynowicz <jacek.lawrynowicz@linux.intel.com>
Reviewed-by: Oded Gabbay <ogabbay@kernel.org>
Reviewed-by: Jeffrey Hugo <quic_jhugo@quicinc.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Link: https://patchwork.freedesktop.org/patch/msgid/20230117092723.60441-7-jacek.lawrynowicz@linux.intel.com
The IPC driver is used to send and receive messages to/from firmware
running on the VPU.
The only supported IPC message format is Job Submission Model (JSM)
defined in vpu_jsm_api.h header.
Co-developed-by: Andrzej Kacprowski <andrzej.kacprowski@linux.intel.com>
Signed-off-by: Andrzej Kacprowski <andrzej.kacprowski@linux.intel.com>
Co-developed-by: Krystian Pradzynski <krystian.pradzynski@linux.intel.com>
Signed-off-by: Krystian Pradzynski <krystian.pradzynski@linux.intel.com>
Signed-off-by: Jacek Lawrynowicz <jacek.lawrynowicz@linux.intel.com>
Reviewed-by: Oded Gabbay <ogabbay@kernel.org>
Reviewed-by: Jeffrey Hugo <quic_jhugo@quicinc.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Link: https://patchwork.freedesktop.org/patch/msgid/20230117092723.60441-5-jacek.lawrynowicz@linux.intel.com
Adds four types of GEM-based BOs for the VPU:
- shmem
- internal
- prime
All types are implemented as struct ivpu_bo, based on
struct drm_gem_object. VPU address is allocated when buffer is created
except for imported prime buffers that allocate it in BO_INFO IOCTL due
to missing file_priv arg in gem_prime_import callback.
Internal buffers are pinned on creation, the rest of buffers types
can be pinned on demand (in SUBMIT IOCTL).
Buffer VPU address, allocated pages and mappings are released when the
buffer is destroyed.
Eviction mechanism is planned for future versions.
Add two new IOCTLs: BO_CREATE, BO_INFO
Signed-off-by: Jacek Lawrynowicz <jacek.lawrynowicz@linux.intel.com>
Reviewed-by: Oded Gabbay <ogabbay@kernel.org>
Reviewed-by: Jeffrey Hugo <quic_jhugo@quicinc.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Link: https://patchwork.freedesktop.org/patch/msgid/20230117092723.60441-4-jacek.lawrynowicz@linux.intel.com
VPU Memory Management Unit is based on ARM MMU-600.
It allows the creation of multiple virtual address spaces for
the device and map noncontinuous host memory (there is no dedicated
memory on the VPU).
Address space is implemented as a struct ivpu_mmu_context, it has an ID,
drm_mm allocator for VPU addresses and struct ivpu_mmu_pgtable that
holds actual 3-level, 4KB page table.
Context with ID 0 (global context) is created upon driver initialization
and it's mainly used for mapping memory required to execute
the firmware.
Contexts with non-zero IDs are user contexts allocated each time
the devices is open()-ed and they map command buffers and other
workload-related memory.
Workloads executing in a given contexts have access only
to the memory mapped in this context.
This patch is has two main files:
- ivpu_mmu_context.c handles MMU page tables and memory mapping
- ivpu_mmu.c implements a driver that programs the MMU device
Co-developed-by: Karol Wachowski <karol.wachowski@linux.intel.com>
Signed-off-by: Karol Wachowski <karol.wachowski@linux.intel.com>
Co-developed-by: Krystian Pradzynski <krystian.pradzynski@linux.intel.com>
Signed-off-by: Krystian Pradzynski <krystian.pradzynski@linux.intel.com>
Signed-off-by: Jacek Lawrynowicz <jacek.lawrynowicz@linux.intel.com>
Reviewed-by: Oded Gabbay <ogabbay@kernel.org>
Reviewed-by: Jeffrey Hugo <quic_jhugo@quicinc.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Link: https://patchwork.freedesktop.org/patch/msgid/20230117092723.60441-3-jacek.lawrynowicz@linux.intel.com
VPU stands for Versatile Processing Unit and it's a CPU-integrated
inference accelerator for Computer Vision and Deep Learning
applications.
The VPU device consist of following components:
- Buttress - provides CPU to VPU integration, interrupt, frequency and
power management.
- Memory Management Unit (based on ARM MMU-600) - translates VPU to
host DMA addresses, isolates user workloads.
- RISC based microcontroller - executes firmware that provides job
execution API for the kernel-mode driver
- Neural Compute Subsystem (NCS) - does the actual work, provides
Compute and Copy engines.
- Network on Chip (NoC) - network fabric connecting all the components
This driver supports VPU IP v2.7 integrated into Intel Meteor Lake
client CPUs (14th generation).
Module sources are at drivers/accel/ivpu and module name is
"intel_vpu.ko".
This patch includes only very besic functionality:
- module, PCI device and IRQ initialization
- register definitions and low level register manipulation functions
- SET/GET_PARAM ioctls
- power up without firmware
Co-developed-by: Krystian Pradzynski <krystian.pradzynski@linux.intel.com>
Signed-off-by: Krystian Pradzynski <krystian.pradzynski@linux.intel.com>
Signed-off-by: Jacek Lawrynowicz <jacek.lawrynowicz@linux.intel.com>
Reviewed-by: Oded Gabbay <ogabbay@kernel.org>
Reviewed-by: Jeffrey Hugo <quic_jhugo@quicinc.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Link: https://patchwork.freedesktop.org/patch/msgid/20230117092723.60441-2-jacek.lawrynowicz@linux.intel.com
This was a mismerge of a semantic conflict in my merge of the driver
core updates, where commit ff62b8e658 ("driver core: make struct
class.devnode() take a const *") changed the devnode function pointer
type. In the meantime, the drm tree I merged earlier had introduced a
new use of that in commit 8bf4889762 ("drivers/accel: define kconfig
and register a new major").
And of course this happens when I'm traveling with my laptop, and thus
didn't do a full allmodconfig build between every pull and before
pushing my work out.
So I only noticed later as I was doing my full build.
Brown-paper-bag-by: Linus Torvalds <torvalds@linux-foundation.org>
Fixes: 71a7507afb Merge tag 'driver-core-6.2-rc1' of git://git.kernel.org/pub/scm/linux/kernel/git/gregkh/driver-core
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
The accelerator devices are exposed to user-space using a dedicated
major. In addition, they are represented in /dev with new, dedicated
device char names: /dev/accel/accel*. This is done to make sure any
user-space software that tries to open a graphic card won't open
the accelerator device by mistake.
The above implies that the minor numbering should be separated from
the rest of the DRM devices. However, to avoid code duplication, we
want the drm_minor structure to be able to represent the accelerator
device.
To achieve this, we add a new drm_minor* to drm_device that represents
the accelerator device. This pointer is initialized for drivers that
declare they handle compute accelerator, using a new driver feature
flag called DRIVER_COMPUTE_ACCEL. It is important to note that this
driver feature is mutually exclusive with DRIVER_RENDER. Devices that
want to expose both graphics and compute device char files should be
handled by two drivers that are connected using the auxiliary bus
framework.
In addition, we define a different IDR to handle the accelerators
minors. This is done to make the minor's index be identical to the
device index in /dev/. Any access to the IDR is done solely
by functions in accel_drv.c, as the IDR is define as static. The
DRM core functions call those functions in case they detect the minor's
type is DRM_MINOR_ACCEL.
We define a separate accel_open function (from drm_open) that the
accel drivers should set as their open callback function. Both these
functions eventually call the same drm_open_helper(), which had to be
changed to be non-static so it can be called from accel_drv.c.
accel_open() only partially duplicates drm_open as I removed some code
from it that handles legacy devices.
To help new drivers, I defined DEFINE_DRM_ACCEL_FOPS macro to easily
set the required function operations pointers structure.
Signed-off-by: Oded Gabbay <ogabbay@kernel.org>
Reviewed-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Reviewed-by: Jeffrey Hugo <quic_jhugo@quicinc.com>
Reviewed-by: Dave Airlie <airlied@redhat.com>
Acked-by: Thomas Zimmermann <tzimmermann@suse.de>
Acked-by: Jacek Lawrynowicz <jacek.lawrynowicz@linux.intel.com>
Tested-by: Jacek Lawrynowicz <jacek.lawrynowicz@linux.intel.com>
Reviewed-by: Melissa Wen <mwen@igalia.com>
Add a new Kconfig for the accel subsystem. The Kconfig currently
contains only the basic CONFIG_DRM_ACCEL option that will be used to
decide whether to compile the accel registration code. Therefore, the
kconfig option is defined as bool.
The accel code will be compiled as part of drm.ko and will be called
directly from the DRM core code. The reason we compile it as part of
drm.ko and not as a separate module is because of cyclic dependency
between drm.ko and the separate module (if it would have existed).
This is due to the fact that DRM core code calls accel functions and
vice-versa.
The accelerator devices will be exposed to the user space with a new,
dedicated major number - 261.
The accel init function registers the new major number as a char device
and create corresponding sysfs and debugfs root entries, similar to
what is done in DRM init function.
I added a new header called drm_accel.h to include/drm/, that will hold
the prototypes of the drm_accel.c functions. In case CONFIG_DRM_ACCEL
is set to 'N', that header will contain empty inline implementations of
those functions, to allow DRM core code to compile successfully
without dependency on CONFIG_DRM_ACCEL.
I Updated the MAINTAINERS file accordingly with the newly added folder
and I have taken the liberty to appropriate the dri-devel mailing list
and the dri-devel IRC channel for the accel subsystem.
Signed-off-by: Oded Gabbay <ogabbay@kernel.org>
Reviewed-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Reviewed-by: Jeffrey Hugo <quic_jhugo@quicinc.com>
Reviewed-by: Dave Airlie <airlied@redhat.com>
Acked-by: Thomas Zimmermann <tzimmermann@suse.de>
Acked-by: Jacek Lawrynowicz <jacek.lawrynowicz@linux.intel.com>
Tested-by: Jacek Lawrynowicz <jacek.lawrynowicz@linux.intel.com>
Reviewed-by: Melissa Wen <mwen@igalia.com>
Tomer Tayar