[PATCH v5] bus: mhi: host: don't free bhie tables during suspend/hibernation

[Muhammad Usama Anjum]

On 5/16/25 7:59 PM, Jeff Hugo wrote:
> On 5/14/2025 2:14 AM, Muhammad Usama Anjum wrote:
>> Fix dma_direct_alloc() failure at resume time during bhie_table
>> allocation because of memory pressure. There is a report where at
>> resume time, the memory from the dma doesn't get allocated and MHI
>> fails to re-initialize.
>>
>> To fix it, don't free the memory at power down during suspend /
>> hibernation. Instead, use the same allocated memory again after every
>> resume / hibernation. This patch has been tested with resume and
>> hibernation both.
>>
>> There are two allocations of bhie; rddm and fbc. Optimize both of those
> 
> There are 3, but you touch 2.  I just commented on this in v4.  Only
> touching two is fine (the device for the 3rd one doesn't need this), but
> the documentation must be accurate.
Yeah, for 3rd one in mhi_load_image_bhie(), this optimization isn't
needed. Let me rephrase and resent v6.

> 
>> allocations. The rddm is of constant size for a given hardware. While
>> the fbc_image size depends on the firmware. If the firmware changes,
>> we'll free and allocate new memory for it. This patch is moticated from
> 
> moticated?  Motivated maybe?
I'm using checkpatch --strict --codespell and hopping that it'll catch
typos. But it misses some of them. Do you use any such tool?

> 
>> the ath12k [1] and ath11k [2] patches. They don't free the memory and
>> reuse the same memory if new size is same. The firmware caching hasn't
>> been implemented for the drivers other than the nouveau. (The changing
>> of firmware isn't tested/supported for wireless drivers. But let's
>> follow the example patches here.)
>>
>> [1] https://lore.kernel.org/all/20240419034034.2842-1-
>> quic_bqiang at quicinc.com/
>> [2] https://lore.kernel.org/all/20220506141448.10340-1-
>> quic_akolli at quicinc.com/
>>
>> Tested-on: WCN6855 WLAN.HSP.1.1-03926.13-
>> QCAHSPSWPL_V2_SILICONZ_CE-2.52297.6
>> Tested-on: WCN7850 hw2.0 WLAN.HMT.1.1.c5-00284-
>> QCAHMTSWPL_V1.0_V2.0_SILICONZ-3
>>
>> Acked-by: Jeff Johnson <jeff.johnson at oss.qualcomm.com>
>> Tested-by: Baochen Qiang <quic_bqiang at quicinc.com>
>> Signed-off-by: Muhammad Usama Anjum <usama.anjum at collabora.com>
>> ---
>> Changes since v1:
>> - Don't free bhie tables during suspend/hibernation only
>> - Handle fbc_image changed size correctly
>> - Remove fbc_image getting set to NULL in *free_bhie_table()
>>
>> Changes since v2:
>> - Remove the new mhi_partial_unprepare_after_power_down() and instead
>>    update mhi_power_down_keep_dev() to use
>>    mhi_power_down_unprepare_keep_dev() as suggested by Mani
>> - Update all users of this API such as ath12k (previously only ath11k
>>    was updated)
>> - Define prev_fw_sz in docs
>> - Do better alignment of comments
>>
>> Changes since v3:
>> - Fix state machine of ath12k by setting ATH12K_MHI_DEINIT with
>>    ATH12K_MHI_POWER_OFF_KEEP_DEV state (Thanks Sebastian for testing and
>>    finding the problem)
>> - Use static with mhi_power_down_unprepare_keep_dev()
>> - Remove crash log as it was showing that kworker wasn't able to
>>    allocate memory.
>>
>> Changes since v4:
>> - Update desctiption
> 
> While I'm commenting on spelling, "description"
> 
> 
> Code looks ok to me. I think we just need to hash out some of the
> documentation and this will be good to go.
Thanks

> 


-- 
Regards,
Usama