The code used the mmu mutex to protect access to the context's page
tables and invalidation of the MMU cache. Because pgt are per
context, the mmu mutex was a member of the context object.
The problem is that the device has a single MMU invalidation h/w
(per MMU). Therefore, the mmu mutex should not be a property of the
context but a property of the device.
Signed-off-by: Oded Gabbay <ogabbay@kernel.org>
Set the addresses for userspace command buffer dynamically
instead of hard-coded. There is no reason for it to
be hard-coded.
Signed-off-by: Dafna Hirschfeld <dhirschfeld@habana.ai>
Reviewed-by: Oded Gabbay <ogabbay@kernel.org>
Signed-off-by: Oded Gabbay <ogabbay@kernel.org>
Several munmap() calls can be done or a mapped H/W block that has a
larger size than a page size.
Releasing the object should be done only when all mapped range is
unmapped.
Signed-off-by: Tomer Tayar <ttayar@habana.ai>
Reviewed-by: Oded Gabbay <ogabbay@kernel.org>
Signed-off-by: Oded Gabbay <ogabbay@kernel.org>
In hl_hw_block_mmap(), the vma's 'vm_private_data' and 'vm_ops' fields
are assigned before filling the content of the private data.
In between there is a call to the ASIC hw_block_mmap() function, and if
it fails, the vma close function will be called with a bad private data
value.
Fix the order of assignments to avoid this issue.
In hl_hw_block_mmap() the vma's 'vm_private_data and vm_ops are assigned
before setting the
Signed-off-by: Tomer Tayar <ttayar@habana.ai>
Reviewed-by: Oded Gabbay <ogabbay@kernel.org>
Signed-off-by: Oded Gabbay <ogabbay@kernel.org>
map_block() sets the block id handle even if get_hw_block_id() fails,
and in this case it uses block id 0 which might be a valid id.
Modify it to set the handle only if get_hw_block_id() succeeds.
Signed-off-by: Tomer Tayar <ttayar@habana.ai>
Reviewed-by: Oded Gabbay <ogabbay@kernel.org>
Signed-off-by: Oded Gabbay <ogabbay@kernel.org>
Gaudi2 has new MMU units. A PMMU for device->host accesses, and HMMU
for HBM accesses.
The page tables of both MMUs are located in the host's memory (referred
to in the code as host-resident pgt).
Signed-off-by: Moti Haimovski <mhaimovski@habana.ai>
Reviewed-by: Oded Gabbay <ogabbay@kernel.org>
Signed-off-by: Oded Gabbay <ogabbay@kernel.org>
Add the ASIC-specific code for Gaudi2. Supply (almost) all of the
function callbacks that the driver's common code need to initialize,
finalize and submit workloads to the Gaudi2 ASIC.
It also contains the code to initialize the F/W of the Gaudi2 ASIC
and to receive events from the F/W.
It contains new debugfs entry to dump razwi events. razwi is a case
where the device's engines create a transaction that reaches an
invalid destination.
Signed-off-by: Oded Gabbay <ogabbay@kernel.org>
Because in future ASICs the driver will allow the user to set the
page size we need to make sure this data is propagated in all APIs.
In addition, since this is already an ASIC property we no longer need
ASIC function for it.
Signed-off-by: Ohad Sharabi <osharabi@habana.ai>
Reviewed-by: Oded Gabbay <ogabbay@kernel.org>
Signed-off-by: Oded Gabbay <ogabbay@kernel.org>
free_device_memory() ends with if and else, each has a return statement,
followed by another return statement that can never be reached.
Restructure the function and remove this dead code.
Signed-off-by: Tomer Tayar <ttayar@habana.ai>
Reviewed-by: Oded Gabbay <ogabbay@kernel.org>
Signed-off-by: Oded Gabbay <ogabbay@kernel.org>
We dropped support for page sizes that are not power of 2.
Signed-off-by: Ohad Sharabi <osharabi@habana.ai>
Reviewed-by: Oded Gabbay <ogabbay@kernel.org>
Signed-off-by: Oded Gabbay <ogabbay@kernel.org>
There is no need to do memory scrub when unmapping anymore as it is
an overhead as long as we have a single user at any given time.
Remove that code and change return value of free_phys_pg_pack to void
Signed-off-by: Dafna Hirschfeld <dhirschfeld@habana.ai>
Reviewed-by: Oded Gabbay <ogabbay@kernel.org>
Signed-off-by: Oded Gabbay <ogabbay@kernel.org>
If hl_mmu_prefetch_cache_range() fails then this code calls
mutex_unlock(&ctx->mmu_lock) when it's no longer holding the mutex.
Fixes: 9e495e2400 ("habanalabs: do MMU prefetch as deferred work")
Signed-off-by: Dan Carpenter <dan.carpenter@oracle.com>
Reviewed-by: Oded Gabbay <ogabbay@kernel.org>
Signed-off-by: Oded Gabbay <ogabbay@kernel.org>
When user requests to prefetch the MMU translations, the driver will
not block the user until prefetch is done.
Instead, the prefetch work will be delegated to a WQ which will do it
in the background.
This way, the prefetch may progress without blocking the user at all.
Signed-off-by: Ohad Sharabi <osharabi@habana.ai>
Reviewed-by: Oded Gabbay <ogabbay@kernel.org>
Signed-off-by: Oded Gabbay <ogabbay@kernel.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Currently, buffers from multiple flows pass through the same infra.
This way, in logs, we are unable to distinguish between buffers that
came from separate flows.
To address this problem, add a "topic" to buffer behavior
descriptor - a string identifier that will be used to identify in logs
the flow this buffer relates to.
Signed-off-by: Yuri Nudelman <ynudelman@habana.ai>
Reviewed-by: Oded Gabbay <ogabbay@kernel.org>
Signed-off-by: Oded Gabbay <ogabbay@kernel.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
The new unified memory manager uses page offset to pass buffer handle
during the mmap operation. One problem with this approach is that it
requires the handle to always be divisible by the page size, else, the
user would not be able to pass it correctly as an argument to the mmap
system call.
Previously, this was achieved by shifting the handle left after alloc
operation, and shifting it right before get operation. This was done in
the user code. This creates code duplication, and, what's worse,
requires some knowledge from the user regarding the handle internal
structure, hurting the encapsulation.
This patch encloses all the page shifts inside memory manager functions.
This way, the user can take the handle as a black box, and simply use
it, without any concert about how it actually works.
Signed-off-by: Yuri Nudelman <ynudelman@habana.ai>
Reviewed-by: Oded Gabbay <ogabbay@kernel.org>
Signed-off-by: Oded Gabbay <ogabbay@kernel.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
This patch let the user decide whether the translations done in the
page tables will be fetched directly to the STLB right after the map.
We want to let the user control whether to perform prefetch upon map
operation.
To do so a memory flag was added, to be used in the MAP ioctl, called
HL_MEM_PREFETCH and if set- the mappings will be fetched directly to
the STLB after map operation.
Signed-off-by: Ohad Sharabi <osharabi@habana.ai>
Reviewed-by: Oded Gabbay <ogabbay@kernel.org>
Signed-off-by: Oded Gabbay <ogabbay@kernel.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Instead of using for_each_sg when iterating sgt that contains dma
entries, use the more proper for_each_sgtable_dma_sg macro.
In addition, both Goya and Gaudi have the exact same implementation
of the asic function that encapsulate the usage of this macro, so
it is better to move that implementation to the common code.
Signed-off-by: Ohad Sharabi <osharabi@habana.ai>
Reviewed-by: Oded Gabbay <ogabbay@kernel.org>
Signed-off-by: Oded Gabbay <ogabbay@kernel.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
drivers/misc/habanalabs/common/memory.c:2137:28: warning: symbol 'hl_ts_behavior' was not declared. Should it be static?
Fixes: 4d530e7d12 ("habanalabs: convert ts to use unified memory manager")
Reported-by: kernel test robot <lkp@intel.com>
Signed-off-by: kernel test robot <lkp@intel.com>
Reviewed-by: Oded Gabbay <ogabbay@kernel.org>
Signed-off-by: Oded Gabbay <ogabbay@kernel.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
The out of memory message is rephrased to more subtle expression as out
of memory may be caused by the user in case of, for example, greedy
allocation.
In addition the user is also being notified by an error code.
Signed-off-by: Ohad Sharabi <osharabi@habana.ai>
Reviewed-by: Oded Gabbay <ogabbay@kernel.org>
Signed-off-by: Oded Gabbay <ogabbay@kernel.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
With the introduction of the unified memory manager infrastructure, the
timestamp buffers can be converted to use it.
Signed-off-by: Yuri Nudelman <ynudelman@habana.ai>
Reviewed-by: Oded Gabbay <ogabbay@kernel.org>
Signed-off-by: Oded Gabbay <ogabbay@kernel.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Looking forward we will need to report to the user what is the default
page size used.
This will be done more conveniently by explicitly updating the property
rather than to rely on a "0 meaning default" value.
Signed-off-by: Ohad Sharabi <osharabi@habana.ai>
Reviewed-by: Oded Gabbay <ogabbay@kernel.org>
Signed-off-by: Oded Gabbay <ogabbay@kernel.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
allmodconfig builds on 32-bit architectures fail with the following error.
drivers/misc/habanalabs/common/memory.c: In function 'alloc_device_memory':
drivers/misc/habanalabs/common/memory.c:153:49: error:
cast from pointer to integer of different size
Fix the typecast. While at it, drop other unnecessary typecasts associated
with the same commit.
Fixes: e8458e20e0 ("habanalabs: make sure device mem alloc is page aligned")
Cc: Ohad Sharabi <osharabi@habana.ai>
Signed-off-by: Guenter Roeck <linux@roeck-us.net>
Link: https://lore.kernel.org/r/20220404134859.3278599-1-linux@roeck-us.net
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Working with MMU that supports multiple page sizes requires that mapping
of a page of a certain size will be aligned to the same size (e.g. the
physical address of 32MB page shall be aligned to 32MB).
To achieve this the gen_poll allocation is now using the "align" variant
to comply with the alignment requirements.
Signed-off-by: Ohad Sharabi <osharabi@habana.ai>
Reviewed-by: Oded Gabbay <ogabbay@kernel.org>
Signed-off-by: Oded Gabbay <ogabbay@kernel.org>
In future ASICs the MMU will be able to work with multiple page sizes,
thus a new flag is added to allow the user to set the requested page
size.
This flag is added since the whole DRAM is allocated for the user and
the user also should be familiar with the memory usage use case.
As such, the user may choose to "over allocate" memory in favor of
performance (for instance- large page allocations covers more memory
in less TLB entries).
For example: say available page sizes are of 1MB and 32MB. If user
wants to allocate 40MB the user can either set page size to 1MB and
allocate the exact amount of memory (but will result in 40 TLB entries)
or the user can use 32MB pages, "waste" 8MB of physical memory but
occupy only 2 TLB entries.
Note that this feature will be available only to ASIC that supports
multiple DRAM page sizes.
Signed-off-by: Ohad Sharabi <osharabi@habana.ai>
Reviewed-by: Oded Gabbay <ogabbay@kernel.org>
Signed-off-by: Oded Gabbay <ogabbay@kernel.org>
Use of vfree(), vmalloc_user(), vmalloc() and remap_vmalloc_range()
requires this include in some architectures.
Signed-off-by: Oded Gabbay <ogabbay@kernel.org>
When the code iterates over the free list of physical pages nodes, it
deletes the physical page node which is used as the iterator.
Therefore, we need to use the safe version of the iteration to prevent
use-after-free.
Signed-off-by: Oded Gabbay <ogabbay@kernel.org>
Timestamp registration API allows the user to register
a timestamp record event which will make the driver set
timestamp when CQ counter reaches the target value
and write it to a specific location specified
by the user.
This is a non blocking API, unlike the wait_for_interrupt
which is a blocking one.
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 not something we can do a workaround. It is clearly an error
and we should notify the user that it is an error.
Signed-off-by: Oded Gabbay <ogabbay@kernel.org>
Freeing phys_pg_pack includes calling to scrubbing functions of the
device's memory, taking locks and possibly even calling reset.
This is not something that should be done while holding a device-wide
spinlock.
Therefore, save the relevant objects on a local linked-list and after
releasing the spinlock, traverse that list and free the phys_pg_pack
objects.
Signed-off-by: Oded Gabbay <ogabbay@kernel.org>
Unify variables related to device reset, which will help us to
add some new reset functionality in future patches.
Signed-off-by: Ofir Bitton <obitton@habana.ai>
Reviewed-by: Oded Gabbay <ogabbay@kernel.org>
Signed-off-by: Oded Gabbay <ogabbay@kernel.org>
-ENOTTY is returned in case of error in the ioctl arguments themselves,
such as function that doesn't exists.
In all other cases, where the error is in the arguments of the custom
data structures that we define that are passed in the various ioctls,
we need to return -EINVAL.
Signed-off-by: Oded Gabbay <ogabbay@kernel.org>
Currently there is a deadlock in driver in scenarios where MMU
cache invalidation fails. The issue is basically device reset
being performed without releasing the MMU mutex.
The solution is to skip device reset as it is not necessary.
In addition we introduce a slight code refactor that prints the
invalidation error from a single location.
Signed-off-by: Ofir Bitton <obitton@habana.ai>
Reviewed-by: Oded Gabbay <ogabbay@kernel.org>
Signed-off-by: Oded Gabbay <ogabbay@kernel.org>
The PCI MMU cache is two layered. The upper layer, memcache, uses cache
lines, the bottom layer doesn't.
Hence, after PMMU map operation we have to invalidate memcache, to avoid
the situation where the new entry is already in the cache due to its
cache line being fully in the cache.
However, we do not have to invalidate the lower cache, and here we can
optimize, since cache invalidation is time consuming.
Signed-off-by: Yuri Nudelman <ynudelman@habana.ai>
Reviewed-by: Oded Gabbay <ogabbay@kernel.org>
Signed-off-by: Oded Gabbay <ogabbay@kernel.org>
The enum vm_type was abused, used once as a value (indication
memory type for map) and once as a flag (for cache invalidation).
This makes it hard to add new and still keep it meaningful, hence it
is better to split into one enum for values and one for flags.
Signed-off-by: Yuri Nudelman <ynudelman@habana.ai>
Reviewed-by: Oded Gabbay <ogabbay@kernel.org>
Signed-off-by: Oded Gabbay <ogabbay@kernel.org>
VA blocks are currently stored in an inconsistent way. Sometimes block
end is inclusive, sometimes exclusive. This leads to wrong size
calculations in certain cases, plus could lead to a segmentation fault
in case mapping process fails in the middle and we try to roll it back.
Need to make this consistent - start inclusive till end inclusive.
For example, the regions table may now look like this:
0x0000 - 0x1fff : allocated
0x2000 - 0x2fff : free
0x3000 - 0x3fff : allocated
Signed-off-by: Yuri Nudelman <ynudelman@habana.ai>
Reviewed-by: Oded Gabbay <ogabbay@kernel.org>
Signed-off-by: Oded Gabbay <ogabbay@kernel.org>
In order to better track where in the kernel the dma-buf code is used,
put the symbols in the namespace DMA_BUF and modify all users of the
symbols to properly import the namespace to not break the build at the
same time.
Now the output of modinfo shows the use of these symbols, making it
easier to watch for users over time:
$ modinfo drivers/misc/fastrpc.ko | grep import
import_ns: DMA_BUF
Cc: "Pan, Xinhui" <Xinhui.Pan@amd.com>
Cc: David Airlie <airlied@linux.ie>
Cc: Maarten Lankhorst <maarten.lankhorst@linux.intel.com>
Cc: Maxime Ripard <mripard@kernel.org>
Cc: Thomas Zimmermann <tzimmermann@suse.de>
Cc: Mauro Carvalho Chehab <mchehab@kernel.org>
Cc: dri-devel@lists.freedesktop.org
Acked-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Acked-by: Christian König <christian.koenig@amd.com>
Acked-by: Arnd Bergmann <arnd@arndb.de>
Acked-by: Sumit Semwal <sumit.semwal@linaro.org>
Acked-by: Alex Deucher <alexander.deucher@amd.com>
Link: https://lore.kernel.org/r/20211010124628.17691-1-gregkh@linuxfoundation.org
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Implement the calls to the dma-buf kernel api to create a dma-buf
object backed by FD.
We block the option to mmap the DMA-BUF object because we don't support
DIRECT_IO and implicit P2P. We only implement support for explicit P2P
through importing the FD of the DMA-BUF.
In the export phase, we provide to the DMA-BUF object an array of pages
that represent the device's memory area. During the map callback,
we convert the array of pages into an SGT. We split/merge the pages
according to the dma max segment size of the importer.
To get the DMA address of the PCI bar, we use the dma_map_resources()
kernel API, because our device memory is not backed by page struct
and this API doesn't need page struct to map the physical address to
a DMA address.
We set the orig_nents member of the SGT to be 0, to indicate to other
drivers that we don't support CPU mappings.
Note that in Habanalabs's ASICs, the device memory is pinned and
immutable. Therefore, there is no need for dynamic mappings and pinning
callbacks.
Also note that in GAUDI we don't have an MMU towards the device memory
and the user works on physical addresses. Therefore, the user doesn't
pass through the kernel driver to allocate memory there. As a result,
only for GAUDI we receive from the user a device memory physical address
(instead of a handle) and a size.
We check the p2p distance using pci_p2pdma_distance_many() and refusing
to map dmabuf in case the distance doesn't allow p2p.
Signed-off-by: Tomer Tayar <ttayar@habana.ai>
Reviewed-by: Oded Gabbay <ogabbay@kernel.org>
Reviewed-by: Gal Pressman <galpress@amazon.com>
Reviewed-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Acked-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Signed-off-by: Oded Gabbay <ogabbay@kernel.org>
Fix 2 areas in the code where it's possible the code will
go to sleep while holding a spinlock.
Reported-by: Dan Carpenter <dan.carpenter@oracle.com>
Signed-off-by: farah kassabri <fkassabri@habana.ai>
Reviewed-by: Oded Gabbay <ogabbay@kernel.org>
Signed-off-by: Oded Gabbay <ogabbay@kernel.org>
It is useful to have the ability to see which user address was pinned
to which physical address during the initial mapping. We already have
all that info stored, but no means to search this data (which may be
quite large).
Signed-off-by: Yuri Nudelman <ynudelman@habana.ai>
Reviewed-by: Oded Gabbay <ogabbay@kernel.org>
Signed-off-by: Oded Gabbay <ogabbay@kernel.org>
Currently userptr endpoint in debugfs prints out virtual addresses
in the user process memory space, without specifying their owner process
ID. User space virtual address is meaningless without knowing the owner
process.
Signed-off-by: Yuri Nudelman <ynudelman@habana.ai>
Reviewed-by: Oded Gabbay <ogabbay@kernel.org>
Signed-off-by: Oded Gabbay <ogabbay@kernel.org>
The driver quietly handles memory mappings that were not freed so no
need to print a warning about that when user closes the FD.
Accordingly, revise the text that is printed in case the device is
still in use after the user process closed the FD.
Signed-off-by: Oded Gabbay <ogabbay@kernel.org>
Sometimes we may need to disable optimization of using huge pages
in our memory management code. Add such a flag to the function that
creates the list of physical pages that would be programmed into the
device MMU.
Signed-off-by: Oded Gabbay <ogabbay@kernel.org>
A new user flag is required to make memory map hint mandatory, in
contrast to the current situation where it is best effort.
This is due to the requirement to map certain data to specific
pre-determined device virtual address ranges.
Signed-off-by: Yuri Nudelman <ynudelman@habana.ai>
Reviewed-by: Oded Gabbay <ogabbay@kernel.org>
Signed-off-by: Oded Gabbay <ogabbay@kernel.org>