NVADSP driver tries to allocate shared memory for ADSP
by requesting the DMA APIs to allocate at address hardcoded
in DT. This poses an issue where if kernel has already allocated
that memory to a different driver in same iommu group, this
memory becomes unavailable and ADSP does not boot
Adding support to allocate shared memory for ADSP dynamically
to avoid dependencies on other drivers
Jira EMA-1213
Change-Id: I1bc8d49f17ec8226d34f3c943cccabef97b2afb6
Signed-off-by: Hariharan Sivaraman <hariharans@nvidia.com>
Reviewed-on: https://git-master.nvidia.com/r/2134341
Reviewed-by: Automatic_Commit_Validation_User
Reviewed-by: Uday Gupta <udayg@nvidia.com>
GVS: Gerrit_Virtual_Submit
Tested-by: Niranjan Dighe <ndighe@nvidia.com>
Reviewed-by: Nitin Pai <npai@nvidia.com>
Reviewed-by: mobile promotions <svcmobile_promotions@nvidia.com>
Tested-by: mobile promotions <svcmobile_promotions@nvidia.com>
There is no point in managing iovad that's not binded to
device's IOMMU domain. So this patch just removes iovad
managing code and does a correct iommu_dma_alloc_iova()
to reserve the iova.
A difficulty of using iommu_dma_alloc_iova() is that we
mostly do not use aligned addresses for the target iova,
but IOMMU core aligns the iova address with the size by
power of two due to cache friendly. So the trick here is
to align the start address and to allocate a larger iova
region that would cover the target iova -- the excessive
part will be freed during iommu_dma_free_iova() by IOMMU
core.
Note that the dma_free_coherent() does both iommu_unmap
and iommu_dma_free_iova, so there's no need of an extra
free_iova() in the exit routine. However, for the remap
part, the tmp_iova is allocated and mapped, so it should
be unmapped and freed. So this patch also adds a free()
of tmp_iova in the nvadsp_dma_alloc_at_map() function.
Bug 200444660
Change-Id: I54126c2bd9ff3dd688aaa5c6f9befaad620ab548
Signed-off-by: Nicolin Chen <nicolinc@nvidia.com>
Reviewed-on: https://git-master.nvidia.com/r/2096704
Reviewed-by: Automatic_Commit_Validation_User
GVS: Gerrit_Virtual_Submit
Reviewed-by: Krishna Reddy <vdumpa@nvidia.com>
Reviewed-by: mobile promotions <svcmobile_promotions@nvidia.com>
Tested-by: mobile promotions <svcmobile_promotions@nvidia.com>
The logic of the nvadsp_dma_alloc_and_map_at() function is to
firstly reserve the target iova and secondly allocate a memory
with another dynamical iova.
Since the target iova is reserved, the dynamical iova won't be
possible to be identical. And this lucky check lacks of an exit
routine either.
So this patch just drops it.
Bug 200444660
Change-Id: I5c45e72b1e7315dacd8708fc8facf6bda41c51e5
Signed-off-by: Nicolin Chen <nicolinc@nvidia.com>
Reviewed-on: https://git-master.nvidia.com/r/2098219
Reviewed-by: Automatic_Commit_Validation_User
GVS: Gerrit_Virtual_Submit
Reviewed-by: Krishna Reddy <vdumpa@nvidia.com>
Reviewed-by: mobile promotions <svcmobile_promotions@nvidia.com>
Tested-by: mobile promotions <svcmobile_promotions@nvidia.com>
The dma-mask could be applied to an older kernel as well.
And this is essential for nvadsp driver to use the newer
nvadsp_dma_alloc_and_map_at() function. Though downstream
kernel 4.9 still has dma_alloc_at_coherent() support, it
would be cleaner to have nvadsp code using a unified way
instead of adding older dma_alloc_at_coherent() back with
another KERNEL_VERSION check.
Bug 200444660
Change-Id: I3968155c7439789a398ca499e70598ce1385737a
Signed-off-by: Nicolin Chen <nicolinc@nvidia.com>
Reviewed-on: https://git-master.nvidia.com/r/2096705
Reviewed-by: Automatic_Commit_Validation_User
GVS: Gerrit_Virtual_Submit
Reviewed-by: Krishna Reddy <vdumpa@nvidia.com>
Reviewed-by: mobile promotions <svcmobile_promotions@nvidia.com>
Tested-by: mobile promotions <svcmobile_promotions@nvidia.com>
The parameters of dma_free_coherent() should be in order of:
1) Device pointer 2) Size 3) CPU virtual address 4) Device address
However, the dma_free_coherent() here mixes with the size with the
device address which will trigger kernel bugs. So this patch fixes
the bug by exchanging those two.
Bug 200444660
Change-Id: Ib0b7855be3868908a4fd59beb938f8bec9e6daee
Signed-off-by: Nicolin Chen <nicolinc@nvidia.com>
Reviewed-on: https://git-master.nvidia.com/r/2098218
Reviewed-by: Automatic_Commit_Validation_User
GVS: Gerrit_Virtual_Submit
Reviewed-by: Krishna Reddy <vdumpa@nvidia.com>
Reviewed-by: mobile promotions <svcmobile_promotions@nvidia.com>
Tested-by: mobile promotions <svcmobile_promotions@nvidia.com>
The function hasn't grabbed the mutex_lock yet at first
condition check. So this patch changes the exit path to
return directly instead of jumping to the mutex_unlock.
Bug 200444660
Change-Id: I122fc9cd145a5de932b4ae55d88c878231926072
Signed-off-by: Nicolin Chen <nicolinc@nvidia.com>
Reviewed-on: https://git-master.nvidia.com/r/2098217
Reviewed-by: Automatic_Commit_Validation_User
GVS: Gerrit_Virtual_Submit
Reviewed-by: Krishna Reddy <vdumpa@nvidia.com>
Reviewed-by: mobile promotions <svcmobile_promotions@nvidia.com>
Tested-by: mobile promotions <svcmobile_promotions@nvidia.com>
The dma_alloc_at_coherent is implemented with a hack in the
iommu core code, which is unlikely able to get upstream. So
this patch removes its callers in the nvadsp code, and then
adds an alternative helper function to do something similar.
The dma_alloc_at_coherent was used to allocate a DMA buffer
and to map it to a specified iova address. And this can be
done by remapping a normal DMA buffer to the iova address.
Bug 200444660
Change-Id: Ic90aec550daa29df69d97c4b230c8f9cc1519166
Signed-off-by: Nicolin Chen <nicolinc@nvidia.com>
Reviewed-on: https://git-master.nvidia.com/r/2083356
Reviewed-by: Krishna Reddy <vdumpa@nvidia.com>
Reviewed-by: Automatic_Commit_Validation_User
Reviewed-by: mobile promotions <svcmobile_promotions@nvidia.com>
Tested-by: mobile promotions <svcmobile_promotions@nvidia.com>
Adsp logger's read incorrectly expected the (remaining)
size of the user-buf to be at least ADSP_TAG - 1 chars.
That has been corrected.
For simplicity there is no state machine introduced to
make sure the ADSP_TAG is always complete in the log.
In the case the tag needs to be added to user-buf without
sufficient size it gets simply truncated.
Bug: 2108805
Change-Id: Ifd137980be521ca3c0932248353f10ea0e92692a
Signed-off-by: Mika Andersson <mikaa@nvidia.com>
Reviewed-on: https://git-master.nvidia.com/r/1770559
Reviewed-by: svc-mobile-coverity <svc-mobile-coverity@nvidia.com>
Reviewed-by: Uday Gupta <udayg@nvidia.com>
Reviewed-by: Viraj Karandikar <vkarandikar@nvidia.com>
Reviewed-by: Ajay Nandakumar M <anandakumarm@nvidia.com>
Reviewed-by: Sandeep Trasi <strasi@nvidia.com>
GVS: Gerrit_Virtual_Submit
Reviewed-by: mobile promotions <svcmobile_promotions@nvidia.com>
Tested-by: mobile promotions <svcmobile_promotions@nvidia.com>
The logger currently blocks in open for finding SOH within the
circular buffer. This blocking is infinite. However, the SOH
is written in buffer for sure when adsp boots.
Hence, added a check to see if in the past if adsp has been
started and then search for SOH else return an error.
Also, the blocking takes place in the read path where the logger
waits in the position where it gets EOT and loops back. This is
not desirable due to recent changes in debugfs and hence now the
read path returns back with an ASCII value of BELL which will not
effect `cat` from writing anything on the console.
Bug 1997011
Change-Id: Ia6e990bcdc1af5c328c24f350af941126beda07a
Signed-off-by: Ajay Nandakumar <anandakumarm@nvidia.com>
Reviewed-on: https://git-master.nvidia.com/r/1662918
Reviewed-by: svc-mobile-coverity <svc-mobile-coverity@nvidia.com>
GVS: Gerrit_Virtual_Submit
Reviewed-by: Sachin Nikam <snikam@nvidia.com>
Reviewed-by: Timo Alho <talho@nvidia.com>
Reviewed-by: mobile promotions <svcmobile_promotions@nvidia.com>
Tested-by: mobile promotions <svcmobile_promotions@nvidia.com>
nvadsp_da_to_va_mappings API provides the ability to get the virtual
address from the physical address. This API does not need to be exposed
as there are no other modules using it and prevent prevent speculative
load related leak.
Bug 200381256
Change-Id: Ic94497ca40402101e4246b914f00e030551b0c4c
Signed-off-by: Ajay Nandakumar <anandakumarm@nvidia.com>
Reviewed-on: https://git-master.nvidia.com/r/1652771
Reviewed-by: mobile promotions <svcmobile_promotions@nvidia.com>
Tested-by: mobile promotions <svcmobile_promotions@nvidia.com>
This reverts commit baa2e9cd86a446e085091494be4247d2a4b64a33
since it causes ADSP tests to fail on T18x and T19x
platforms.
Bug 2057870
Change-Id: Ibfc1ec5dfcd678bf608598edc1f118a452d072e3
Signed-off-by: Bharat Nihalani <bnihalani@nvidia.com>
Reviewed-on: https://git-master.nvidia.com/r/1651935
Reviewed-by: svc-mobile-coverity <svc-mobile-coverity@nvidia.com>
adding ability for ADSP to share data when an exception occurs
on the ADSP. The data is shared through the shared memory across
the host.
adding shared exception data prints on host side for debugging.
bug 1868578
Change-Id: I6645e8ae6ab73e5c9df8374f583dacbed4615536
Signed-off-by: Prateek Patel <prpatel@nvidia.com>
Reviewed-on: https://git-master.nvidia.com/r/1611260
Reviewed-by: Ajay Nandakumar M <anandakumarm@nvidia.com>
Reviewed-by: Automatic_Commit_Validation_User
Reviewed-by: Bharat Nihalani <bnihalani@nvidia.com>
With adma driver as module, following compilation error is seen
"undefined reference to tegra_adma_dump_ch_reg".
In the current patch, adsp code exports a function to set regdump
function call back. This function is used by adsp audio driver to
set the function pointer to tegra_adma_dump_ch_reg. This is done
in tegra210_adsp_init().
Bug 200289390
Reviewed-on: http://git-master/r/1483001
Change-Id: I1ef259cf3b988a41720fb6706d39d9bd6f86aac3
Signed-off-by: Sameer Pujar <spujar@nvidia.com>
(cherry picked from commit 36ccba1262928a6ce8b0f4fff1d4ad2591c4305d)
Reviewed-on: https://git-master.nvidia.com/r/1528158
Signed-off-by: Nitin Kumbhar <nkumbhar@nvidia.com>
Reviewed-on: https://git-master.nvidia.com/r/1537339
Reviewed-by: Automatic_Commit_Validation_User
Reviewed-by: svccoveritychecker <svccoveritychecker@nvidia.com>
GVS: Gerrit_Virtual_Submit
Making the "WFI wait" from interruptible to non-interruptible. This is
done because there is a probalbilty between the thread making a call to
nvadsp_os_suspend / nvadsp_os_start getting signaled or having a
pending signal before ADSP enters WFI and the CPU getting the WFI
interrupt.This results in wait_for_completion_interruptible_timeout()
returning -ERESTARTSYS, due to which nvadsp_os_suspend treats it as
failure although ADSP has entered WFI and the WFI interrupt has been
received on the CCPLEX.
Since this is a race it is not seen often.
Also, reducing the timeouts for WFI and logger wait to 800 and
500 ms respectively. The timeouts have been reduced because during
suspend there are only a few register writes and logger has been
reduced from 1 MB to 16 KB.
Bug 200317530
Bug 1893324
Change-Id: I40b3894f284763c0e4cc45647e1adf2cd074d701
Signed-off-by: Ajay Nandakumar <anandakumarm@nvidia.com>
Reviewed-on: https://git-master.nvidia.com/r/1517838
Signed-off-by: Nitin Kumbhar <nkumbhar@nvidia.com>
Reviewed-on: https://git-master.nvidia.com/r/1537335
Reviewed-by: Automatic_Commit_Validation_User
Reviewed-by: svccoveritychecker <svccoveritychecker@nvidia.com>
GVS: Gerrit_Virtual_Submit
Instead of using mailbox to communicate change in adsp cpu
freq, update it in adsp os args from ccplex-adsp shared
memory. This adsp freq update is needed by EDF scheduler
for scheduling EDF threads. The EDF scheduler gets
current adsp freq from shared memory instead of
scheduler callback.
However, for APE1 (i.e. chips = T21x), timer prescalar
must still be updated through mailbox each time adsp cpu
freq is changed.
Bug 200322504
Change-Id: I26f14f5c4f5c2369d67dfcd73dc6b9664df2ca26
Signed-off-by: Nitin Kumbhar <nkumbhar@nvidia.com>
Reviewed-on: https://git-master/r/1512069
(cherry picked from commit 76e8a4c19de2fa5f0d91ebefac68aa0b864cb0eb)
Reviewed-on: https://git-master/r/1512864
(cherry picked from commit 3100911879f9ea9930a515f6d55257caefcef869)
Reviewed-on: https://git-master.nvidia.com/r/1517749
Reviewed-on: https://git-master.nvidia.com/r/1537334
Reviewed-by: Automatic_Commit_Validation_User
Reviewed-by: svccoveritychecker <svccoveritychecker@nvidia.com>
GVS: Gerrit_Virtual_Submit
Dump name (first 4 chars) of current thread and irq
number of last interrupt handled from hwmbox reg 4 and 7
respectively.
Bug 200295526
Change-Id: I4b53521cbfece2177e3cf6ed80b559b58e1d1c0a
Signed-off-by: Nitin Kumbhar <nkumbhar@nvidia.com>
Reviewed-on: http://git-master/r/1475487
(cherry picked from commit 64a799ae7d8e8f8783212fb4c28209cf760b38a7)
Reviewed-on: http://git-master/r/1478825
(cherry picked from commit cab60f6977fc8070e4206db01e2a2d444860c90c)
Reviewed-on: https://git-master/r/1513585
Reviewed-on: https://git-master.nvidia.com/r/1537329
Reviewed-by: Automatic_Commit_Validation_User
Reviewed-by: svccoveritychecker <svccoveritychecker@nvidia.com>
GVS: Gerrit_Virtual_Submit
The state of ADSP is logged by writing to a mailbox
register. Adding support to kernel driver to dump ADSP
state when crash happens.
Jira EMA-404
Bug 1901511
Bug 1893324
Bug 200239577
Change-Id: Ia2beb84a7d5bc871f7062bd6a661333af0b0e24d
Signed-off-by: Hariharan Sivaraman <hariharans@nvidia.com>
Reviewed-on: http://git-master/r/1229864
(cherry picked from commit 0027e7a51f6134b4a7b332c17f7c7c26ac727ebc)
Signed-off-by: Nitin Kumbhar <nkumbhar@nvidia.com>
Reviewed-on: https://git-master.nvidia.com/r/1537327
Reviewed-by: svccoveritychecker <svccoveritychecker@nvidia.com>
GVS: Gerrit_Virtual_Submit
The request irqs are moved just before starting ADSP and freed when
adsp is suspended/stopped.
This is since the new agic driver is based on a device-driver model
and requires all interrupts to be freed before it could suspend.
Bug 200270956
Change-Id: I8ecd05ebe52020f11be79b9a1da37a85fed432ac
Signed-off-by: Ajay Nandakumar <anandakumarm@nvidia.com>
Reviewed-on: http://git-master/r/1478838
Signed-off-by: Nitin Kumbhar <nkumbhar@nvidia.com>
Reviewed-on: https://git-master.nvidia.com/r/1537326
Reviewed-by: Automatic_Commit_Validation_User
Reviewed-by: svccoveritychecker <svccoveritychecker@nvidia.com>
GVS: Gerrit_Virtual_Submit
When ADSP OS is loaded from adsp.elf, it loads two sections : adsp os
and the vector table. The vector table is copied to a local buffer
before writing to the EVP registers.
There is a probability that, if the adsp.elf is hacked / modified by
an attacker, it can write to other APE register spaces. Hence, the
destination address obtained from the elf is checked with evp base
before copying.
Also, this fixes out of array-bound writing to evp buffer.
Bug 1684844
Change-Id: I8981dc5a5db8e0c0653ff46a1576df2c82e420be
Signed-off-by: Ajay Nandakumar <anandakumarm@nvidia.com>
Reviewed-on: http://git-master/r/1118936
(cherry picked from commit 9f7120e03e66b5f6e2bf67f09063da20945be238)
Reviewed-on: http://git-master/r/1458894
Signed-off-by: Nitin Kumbhar <nkumbhar@nvidia.com>
Reviewed-on: https://git-master.nvidia.com/r/1537317
Reviewed-by: svccoveritychecker <svccoveritychecker@nvidia.com>
GVS: Gerrit_Virtual_Submit
