[RFC v4 06/21] ath10k: sdio support

[Erik Stromdahl]

On 2017-03-10 13:43, Valo, Kalle wrote:
> "Valo, Kalle" <kvalo at qca.qualcomm.com> writes:
>
>> Erik Stromdahl <erik.stromdahl at gmail.com> writes:
>>
>>> sdio/mailbox HIF implementation.
>>>
>>> Signed-off-by: Erik Stromdahl <erik.stromdahl at gmail.com>
>>
>> I'm looking at this more carefully now and noticed this:
>>
>>> +static int ath10k_sdio_bmi_credits(struct ath10k *ar)
>>> +{
>>> +	int ret;
>>> +	u32 addr, *cmd_credits;
>>> +	unsigned long timeout;
>>> +
>>> +	cmd_credits = kzalloc(sizeof(*cmd_credits), GFP_KERNEL);
>>> +	if (!cmd_credits) {
>>> +		ret = -ENOMEM;
>>> +		goto err;
>>> +	}
>>> +
>>> +	/* Read the counter register to get the command credits */
>>> +	addr = MBOX_COUNT_DEC_ADDRESS + ATH10K_HIF_MBOX_NUM_MAX * 4;
>>> +
>>> +	timeout = jiffies + BMI_COMMUNICATION_TIMEOUT_HZ;
>>> +	while (time_before(jiffies, timeout) && !*cmd_credits) {
>>> +		/* Hit the credit counter with a 4-byte access, the first byte
>>> +		 * read will hit the counter and cause a decrement, while the
>>> +		 * remaining 3 bytes has no effect. The rationale behind this
>>> +		 * is to make all HIF accesses 4-byte aligned.
>>> +		 */
>>> +		ret = ath10k_sdio_read_write_sync(ar, addr,
>>> +						  (u8 *)cmd_credits,
>>> +						  sizeof(*cmd_credits),
>>> +						  HIF_RD_SYNC_BYTE_INC);
>>> +		if (ret) {
>>> +			ath10k_warn(ar,
>>> + "Unable to decrement the command credit count register: %d\n",
>>> +				    ret);
>>> +			goto err_free;
>>> +		}
>>> +
>>> +		/* The counter is only 8 bits.
>>> +		 * Ignore anything in the upper 3 bytes
>>> +		 */
>>> +		*cmd_credits &= 0xFF;
>>> +	}
>>> +
>>> +	if (!*cmd_credits) {
>>> +		ath10k_warn(ar, "bmi communication timeout\n");
>>> +		ret = -ETIMEDOUT;
>>> +		goto err_free;
>>> +	}
>>> +
>>> +	return 0;
>>> +err_free:
>>> +	kfree(cmd_credits);
>>> +err:
>>> +	return ret;
>>> +}
>>
>> AFAICS we are leaking cmd_credits if there's no error. Or is the buffer
>> freed somewhere within the mmc stack or something? The reason why I ask
>> is that I saw the same pattern in multiple functions so I'm curious.
>
> Also I'm worried about endianness. We are reading from the device
> directly to an u32 variable and not converting the bytes. Is the MMC
> subsystem doing the conversion, I guess not?
>
You are right, there is definitely a memory leak (and there are similar problems
in a couple of other functions as well as you have pointed out).

This was introduced in version 3 of the
RFC when I removed the bounce buffer from ath6kl. Instead I introduced a bunch of
local "bounce" buffers in order to make sure that the buffers passed to the sdio
subsystem is dma-able (malloc'ed buffer instead of stack allocated).

Regarding endianess: That particular code construct is an artifact from ath6kl.
I am not sure it makes any sense to use a u32 in that particular case.
A u8 array is most likely more convenient.

It is really nice you have found some time to review the patches!

--
Erik