- Use the existing gasket_interrupt_msix_cleanup at suspend time to
remove our IRQs from the system. Without doing this, we leave 13 IRQs
per device registered, which can cause issues during suspend if many
devices are loaded in the system.
- Restore them at resume time, using the gasket_interrupt_reinit
functionality.
- This allows successful suspend without any errors taking the CPUs
down, and passing the multi-tpu stress test after resume.
Change-Id: Ied1aca8605c0cb3b64ba591d05312d10cf45343f
After host suspend/resume operation device is power cycled and
loses all custom register settings.
All settings are restored on first device open except thermal
monitoring related settings.
Signed-off-by: Leonid Lobachev <leonidl@google.com>
Change-Id: I3a7c69efd919da0992c585874d7fe906a7768ee7
The variables "mapping" and "tmp" are unused in
gasket_page_table_unmap_dmabuf().
Signed-off-by: Waqar Hameed <waqarh@axis.com>
Change-Id: I0da67e5e91fb8c8f34b13e7227afb8410bcbc138
Add a space before the open parentheses in an if-statement, to follow
Linux kernel coding style.
Signed-off-by: Waqar Hameed <waqarh@axis.com>
Change-Id: Ic6658aecdde508762117364669872d69dcb0e8e6
Buffers to be mapped as DMA_TO_DEVICE must be writable due to the
'write' argument to get_user_pages_fast being hard coded to 1.
This prevents userspace from passing in pointers returned from
mmap(fd, PROT_READ), instead mmap(fd, PROT_READ | PROT_WRITE)
must be used even if the buffer is supposed to be read only.
Instead set writable to non zero iff direction != DMA_TO_DEVICE,
ie only buffers with DMA_FROM_DEVICE or DMA_BIDIRECTIONAL are
required to be writable.
Change-Id: I21c97b2bb855e11ecaa6ae9e81cf8b463b4aedaa
Signed-off-by: Jonas Larsson <ljonas@google.com>
This allows userspace to pass in dma-buf file descriptors for
mapping in addition to userspace virtual memory pointers,
leaving cache maintenance to the dma-buf exporter and userspace
application.
Bug: 149764192
Change-Id: Iee8b581db0a461fe13ee8712d47d66f8fb16e7c9
Signed-off-by: Jonas Larsson <ljonas@google.com>
This is needed to avoid the need for manual dkms driver installation
with --force on systems where upstream staging kernel module is present
already.
Change-Id: Ifb3dc930757835888ae3e13d1607d5fd7bc710ed
This allows to force dma buffers to be located in 32bit low address
space to workaround issues with some pcie controllers.
Change-Id: I7f77081882b6dffb91f034f8491ad8ffa1375fe7
Signed-off-by: Leonid Lobachev <leonidl@google.com>
Coherent buffer mmap code assumes DMA addresses are physical addresses,
which is patently wrong when a downstream IOMMU allocates and
translates IOVAs for DMA mappings. Use dma_mmap_coherent().
Bug: 140076240
Bug: 111709861
Change-Id: Ieaa052378844aaf126ea3bc9dba544e7edb37bc7
Signed-off-by: Todd Poynor <toddpoynor@google.com>
The extended address bit is at position 63 (c.f. apex_driver.c:60). The
device address will thus not fit in a ulong on 32-bit ARM systems.
Use u64 instead.
[toddpoynor@google.com: more fixups, fix debug printk param order]
Bug: 137378503
Signed-off-by: Waqar Hameed <waqarh@axis.com>
Change-Id: If801b2b97d88c1329f3a359507e265837150571c
In at least the case of an mfd_cell multi-function device child, the
child platform device does not inherit the DMA / IOMMU configuration
of the parent. Copying that configuration is non-straightforward or
even likely impossible when the parent device is PCI and holds
pointers to IOMMU-related state only made available for PCI devices.
Add a gasket call to allow platform chip drivers to register their
PCI parent as the device to use for DMA API calls, such that the
chip drivers can workaround this.
Change-Id: I4a4b6de67f5d10197f8ea02d06522090b150d659
Signed-off-by: Todd Poynor <toddpoynor@google.com>
Make sure to use dma_sync_single_for_device whenever updating the
contents of an extended page table. This performs the cache flush using
the correct function from the DMA-API. Also move this call outside of
the inner loop to improve performance.
Tested: Ran run_tests with inception, testfullyconnected with parameter
caching, and rnntransducerdecoderpie107m on AT Enterprise.
Change-Id: Ie3c63113dde8cafe448d515e8970810afc9564ac
Signed-off-by: Nick Ewalt <nicholasewalt@google.com>
Add that number of page table entries and extended address bit offset
are configurable. Update example virtual address format to be more
consistent with typical usage.
Change-Id: I73f9674df74b6d8ee5ed6e5d95ab12dcc56e29cc
Signed-off-by: Nick Ewalt <nicholasewalt@google.com>
Signed-off-by: Todd Poynor <toddpoynor@google.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Add a brief description and URL for more information on the Apex device,
an Edge TPU (Tensorflow Processing Unit) machine learning accelerator.
Change-Id: Ia7b40414e9f0745feb8ced2cb3aa7e9c6baa0c3d
Signed-off-by: Todd Poynor <toddpoynor@google.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Add support for platform devices in the gasket framework.
Change-Id: I1e3625f9222b37925a2a672c9ea6eed31aefa35f
Signed-off-by: Todd Poynor <toddpoynor@google.com>
Add interrupt type DEVICE_MANAGED, indicating that the device driver
manages interrupt setup and handling. Future non-PCI wire interrupts
will use this type, calling gasket_handle_interrupt(), which is now made
non-static, to call into the gasket framework in order to update sysfs
files for interrupt counts and other framework interfaces.
Change-Id: Ie0be3d950ed2706f7ada848c19ddf7017e9623b2
Signed-off-by: Todd Poynor <toddpoynor@google.com>
Refactor gasket_components_to_dev_address to be faster and easier to
understand. The old implementation was unnecessarily complex and masked
the page_index for simple addresses but not extended ones. It makes the
most sense for this function to perform no such masking.
Change-Id: If5f68cb6f1bf3000a5d5b77aab89cb8e391c3cf4
Signed-off-by: Nick Ewalt <nicholasewalt@google.com>
Signed-off-by: Todd Poynor <toddpoynor@google.com>
The maximum number of entries in the page table is configurable at
initialization time and should be used in gasket_extended_lvl0_page_idx.
Change-Id: I39f2dba48083a3eef3951ecd9659ee9f4901275e
Signed-off-by: Nick Ewalt <nicholasewalt@google.com>
Signed-off-by: Todd Poynor <toddpoynor@google.com>
Previously pages would have never been unmapped in this case.
Change-Id: I71b6447267933865ab23a3ba3830e7ce3d3e0e57
Signed-off-by: Nick Ewalt <nicholasewalt@google.com>
Signed-off-by: Todd Poynor <toddpoynor@google.com>
Correctly clean up data structure state in gasket_alloc_coherent_memory
error path, to ensure no double free on the stale pointer value.
Change-Id: Ifbc426e857d6bd05a13a725f2c2fdb07dedc098c
Signed-off-by: Nick Ewalt <nicholasewalt@google.com>
Signed-off-by: Todd Poynor <toddpoynor@google.com>
Only call dma_unmap_page if there was an associated dma_map_page.
Change-Id: I6234719f8ab6b1e5e069c21e2f639eee4f22b039
Signed-off-by: Nick Ewalt <nicholasewalt@google.com>
Signed-off-by: Todd Poynor <toddpoynor@google.com>
This allows for more precise dma_direction in the dma_map_page requests.
Also leaves room for adding more flags later.
Change-Id: I46feecc2638b7ad0d34548cc10837549a3928011
Signed-off-by: Nick Ewalt <nicholasewalt@google.com>
Signed-off-by: Todd Poynor <toddpoynor@google.com>
Remove unnecessary ptes[i].status update in gasket_perform_unmapping.
The vlaue will be cleared in the following memset.
Change-Id: I5aeafa1085203b5dc24b3303037ece1f37acd6c6
Signed-off-by: Nick Ewalt <nicholasewalt@google.com>
Signed-off-by: Todd Poynor <toddpoynor@google.com>
Extended page tables can be mapped as DMA_TO_DEVICE.
Calls to dma_unmap_page should match corresponding dma_map_page.
Change-Id: I2d09fd5d66c364e190b073fabe089ce5eaf426f6
Signed-off-by: Nick Ewalt <nicholasewalt@google.com>
Signed-off-by: Todd Poynor <toddpoynor@google.com>
gasket_alloc_coherent_memory() has some unnecessary code related to out
of memory checking that will never hit the condition checked, remove.
Change-Id: Ifb09e7ff31d8ae41a486639a2abc6ce26c4d85ef
Signed-off-by: Todd Poynor <toddpoynor@google.com>
Return -EINVAL on mapping failures, instead of -1, which triggers a
checkpatch error.
Change-Id: I91a62a72caa5ca09e07bd4897e72f4919495c040
Signed-off-by: Todd Poynor <toddpoynor@google.com>
Add debug logs for device enable/disable events, remove logs for
callbacks (the called functions can generate their own logs if needed).
Change-Id: I861077c3210452d9ac6606e973bac5c14eeef859
Signed-off-by: Todd Poynor <toddpoynor@google.com>
Devices not using MSIX don't use the msix_initialized field, don't
require it to be set in the interrupt system status check. The general
check for interrupts configured that follows can cover both MSIX and
device-managed interrupts.
Change-Id: I83a86cb4db7cea30c30377dd82b033ef372d69c0
Signed-off-by: Todd Poynor <toddpoynor@google.com>
Flags should be specified for dma_alloc_coherent() call. Use
GFP_KERNEL, it's fine to sleep here.
Change-Id: Iec4ad9550e8829b67e99b22246e40e0b21600c54
Signed-off-by: Todd Poynor <toddpoynor@google.com>
Actually free gasket_coherent_page_entries and just call
gasket_free_coherent_memory_all from gasket_free_coherent_memory.
Change-Id: I1d016abd5f66e0620a37f6851e43fcfd6c96868a
Signed-off-by: Nick Ewalt <nicholasewalt@google.com>
Add module parameters to control default values for HW temperature
monitoring module.
Signed-off-by: Leonid Lobachev <leonidl@google.com>
Change-Id: I69f7cd9c3c0f55a3a36efba35f48a5da37b3b89e
Perform periodic polling of apex temperature and scale down
its frequency if temperature is above one of defined trip points.
Added sysfs nodes to control related settings on per device basis.
Added module parameters to set allow setting default value for all
devices in system.
Signed-off-by: Leonid Lobachev <leonidl@google.com>
Change-Id: I5065282cba9f0888abd4086601bda89174f24705
Port of change by Rob Springer to fix leak of eventfd context objects in
gasket framework interrupt code.
Bug: 131928433
Bug: 80495595
Reported-by: Jann Horn <jannh@google.com>
Reported-by: Neel Natu <neelnatu@google.com>
Test: Resnet50PB, PortraitSegmentation
Change-Id: I6bdea5ab4d3a7435e7f51c96817f0542d185b22c
Signed-off-by: Todd Poynor <toddpoynor@google.com>
Add node to read current temperature and nodes
to set and enable/disable thermal warnings.
Change-Id: I8a12f6e33679573107686fe32be2e4a8bb0f9cc7
Signed-off-by: Leonid Lobachev <leonidl@google.com>
For driver build as module pci_fixup_header approach doesn't work.
This causes apex pci class to remain 0xff and BARs remain
unclaimed. Adding a retry with explicit fixup function call
workarounds the problem.
Bug: 129287752
Change-Id: Id5ff85ef427d3740706a214d3beb1c9bbf42ab5a
Signed-off-by: Leonid Lobachev <leonidl@google.com>
