Re: [PATCH v4] bus: mhi: host: don't free bhie tables during suspend/hibernation

[Jeff Hugo <jeff.hugo@oss.qualcomm.com>]

On 5/13/2025 12:44 AM, Muhammad Usama Anjum wrote:
> On 5/12/25 11:46 PM, Jeff Hugo wrote:
>> On 5/6/2025 8:49 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.
>>>
>>> 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.
>>
>> Why is it valid to load new firmware as a result of suspend?  I don't
>> users would expect that.
> I'm not sure its valid or not. Like other users, I also don't expect
> that firmware would get changed. It doesn't seem to be tested and hence
> supported case.
> 
> But other drivers have code which have implementation like this. I'd
> mentioned previously that this patch was motivated from the ath12k [1]
> and ath11k [2] patches. They don't free the memory and reuse the same
> memory if new size is same.

It feels like this justification needs to be detailed in the commit 
text. I suspect at some point we'll have another MHI device where the FW 
will need to be cached.

>>> diff --git a/drivers/bus/mhi/host/boot.c b/drivers/bus/mhi/host/boot.c
>>> index efa3b6dddf4d2..bc8459798bbee 100644
>>> --- a/drivers/bus/mhi/host/boot.c
>>> +++ b/drivers/bus/mhi/host/boot.c
>>> @@ -584,10 +584,17 @@ void mhi_fw_load_handler(struct mhi_controller
>>> *mhi_cntrl)
>>>         * device transitioning into MHI READY state
>>>         */
>>>        if (fw_load_type == MHI_FW_LOAD_FBC) {
>>
>> Why is this FBC specific?
> It seems we allocate fbc_image only when firmware load type is
> FW_LOAD_FBC. I'm just optimizing the buffer allocation here.

We alloc bhie tables in non FBC usecases. Is this somehow an FBC 
specific issue? Perhaps you could clarify the limits of this solution in 
the commit text?

> 
>>
>>> -        ret = mhi_alloc_bhie_table(mhi_cntrl, &mhi_cntrl->fbc_image,
>>> fw_sz);
>>> -        if (ret) {
>>> -            release_firmware(firmware);
>>> -            goto error_fw_load;
>>> +        if (mhi_cntrl->fbc_image && fw_sz != mhi_cntrl->prev_fw_sz) {
>>> +            mhi_free_bhie_table(mhi_cntrl, mhi_cntrl->fbc_image);
>>> +            mhi_cntrl->fbc_image = NULL;
>>> +        }
>>> +        if (!mhi_cntrl->fbc_image) {
>>> +            ret = mhi_alloc_bhie_table(mhi_cntrl, &mhi_cntrl-
>>>> fbc_image, fw_sz);
>>> +            if (ret) {
>>> +                release_firmware(firmware);
>>> +                goto error_fw_load;
>>> +            }
>>> +            mhi_cntrl->prev_fw_sz = fw_sz;
>>>            }
>>>              /* Load the firmware into BHIE vec table */
>>> diff --git a/drivers/bus/mhi/host/pm.c b/drivers/bus/mhi/host/pm.c
>>> index e6c3ff62bab1d..107d71b4cc51a 100644
>>> --- a/drivers/bus/mhi/host/pm.c
>>> +++ b/drivers/bus/mhi/host/pm.c
>>> @@ -1259,10 +1259,19 @@ void mhi_power_down(struct mhi_controller
>>> *mhi_cntrl, bool graceful)
>>>    }
>>>    EXPORT_SYMBOL_GPL(mhi_power_down);
>>>    +static void __mhi_power_down_unprepare_keep_dev(struct
>>> mhi_controller *mhi_cntrl)
>>> +{
>>> +    mhi_cntrl->bhi = NULL;
>>> +    mhi_cntrl->bhie = NULL;
>>
>> Why?
> This function is shorter version of mhi_unprepare_after_power_down(). As
> we need different code path in case of suspend/hibernation case, I was
> adding a new API which Mani asked me remove and consolidate into
> mhi_power_down_keep_dev() instead. So this static function has been
> added. [3]

I don't understand the need to zero these out.  Also, if you are copying 
part of the functionality of mhi_unprepare_after_power_down(), shouldn't 
that functionality be moved into your new API to eliminate duplication?