lkml.org 
[lkml]   [2022]   [Feb]   [8]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
SubjectRe: [PATCH] fsi: occ: Improve response status checking
From

On 2/7/22 03:56, Joel Stanley wrote:
> On Mon, 31 Jan 2022 at 15:29, Eddie James <eajames@linux.ibm.com> wrote:
>>
>> On 1/30/22 23:56, Joel Stanley wrote:
>>> On Mon, 10 Jan 2022 at 15:58, Eddie James <eajames@linux.ibm.com> wrote:
>>>> If the driver sequence number coincidentally equals the previous
>>>> command response sequence number, the driver may proceed with
>>>> fetching the entire buffer before the OCC has processed the current
>>>> command. To be sure the correct response is obtained, check the
>>>> command type and also retry if any of the response parameters have
>>>> changed when the rest of the buffer is fetched.
>>>>
>>>> Signed-off-by: Eddie James <eajames@linux.ibm.com>
>>>> ---
>>>> drivers/fsi/fsi-occ.c | 63 ++++++++++++++++++++++++++++++-------------
>>>> 1 file changed, 44 insertions(+), 19 deletions(-)
>>>>
>>>> diff --git a/drivers/fsi/fsi-occ.c b/drivers/fsi/fsi-occ.c
>>>> index 7eaab1be0aa4..67569282dd69 100644
>>>> --- a/drivers/fsi/fsi-occ.c
>>>> +++ b/drivers/fsi/fsi-occ.c
>>>> @@ -451,6 +451,15 @@ static int occ_trigger_attn(struct occ *occ)
>>>> return rc;
>>>> }
>>>>
>>>> +static void fsi_occ_print_timeout(struct occ *occ, struct occ_response *resp,
>>>> + u8 seq_no, u8 cmd_type)
>>>> +{
>>>> + dev_err(occ->dev,
>>>> + "resp timeout status=%02x seq=%d cmd=%d, our seq=%d cmd=%d\n",
>>>> + resp->return_status, resp->seq_no, resp->cmd_type, seq_no,
>>>> + cmd_type);
>>>> +}
>>>> +
>>>> int fsi_occ_submit(struct device *dev, const void *request, size_t req_len,
>>>> void *response, size_t *resp_len)
>>>> {
>>>> @@ -461,12 +470,14 @@ int fsi_occ_submit(struct device *dev, const void *request, size_t req_len,
>>>> struct occ_response *resp = response;
>>>> size_t user_resp_len = *resp_len;
>>>> u8 seq_no;
>>>> + u8 cmd_type;
>>>> u16 checksum = 0;
>>>> u16 resp_data_length;
>>>> const u8 *byte_request = (const u8 *)request;
>>>> - unsigned long start;
>>>> + unsigned long end;
>>>> int rc;
>>>> size_t i;
>>>> + bool retried = false;
>>>>
>>>> *resp_len = 0;
>>>>
>>>> @@ -478,6 +489,8 @@ int fsi_occ_submit(struct device *dev, const void *request, size_t req_len,
>>>> return -EINVAL;
>>>> }
>>>>
>>>> + cmd_type = byte_request[1];
>>>> +
>>>> /* Checksum the request, ignoring first byte (sequence number). */
>>>> for (i = 1; i < req_len - 2; ++i)
>>>> checksum += byte_request[i];
>>>> @@ -509,30 +522,30 @@ int fsi_occ_submit(struct device *dev, const void *request, size_t req_len,
>>>> if (rc)
>>>> goto done;
>>>>
>>>> - /* Read occ response header */
>>>> - start = jiffies;
>>>> +retry:
>>>> + end = jiffies + timeout;
>>>> do {
>>>> + /* Read occ response header */
>>>> rc = occ_getsram(occ, 0, resp, 8);
>>>> if (rc)
>>>> goto done;
>>>>
>>>> if (resp->return_status == OCC_RESP_CMD_IN_PRG ||
>>>> resp->return_status == OCC_RESP_CRIT_INIT ||
>>>> - resp->seq_no != seq_no) {
>>>> - rc = -ETIMEDOUT;
>>>> -
>>>> - if (time_after(jiffies, start + timeout)) {
>>>> - dev_err(occ->dev, "resp timeout status=%02x "
>>>> - "resp seq_no=%d our seq_no=%d\n",
>>>> - resp->return_status, resp->seq_no,
>>>> - seq_no);
>>>> + resp->seq_no != seq_no || resp->cmd_type != cmd_type) {
>>> You're testing for two different types of conditions. The first is
>>> when the SBE is busy doing something else:
>>>
>>> if (resp->return_status == OCC_RESP_CMD_IN_PRG ||
>>> resp->return_status == OCC_RESP_CRIT_INIT ||
>>>
>>> And the others are when the message is not for the current user:
>>>
>>> resp->seq_no != seq_no || resp->cmd_type != cmd_type) {
>>>
>>> Should we be separating them out? It makes sense that the first means
>>> we should keep trying. For the second case should we bail straight
>>> away, instead of waiting for the timeout?
>>
>> They're really the same thing actually. If the sequence number or
>> command type is incorrect, it means the OCC is processing a different
>> command, and we need to wait for it to get to our command.
> And we sometimes see one but not the other (ie, the return_status
> doesn't cover all cases)?


Yes, we definitely can see each one without the others, so we have to
check for them all.


>
>>
>>> How often do we see one vs the other?
>>>
>>>> + if (time_after(jiffies, end)) {
>>>> + fsi_occ_print_timeout(occ, resp, seq_no,
>>>> + cmd_type);
>>>> + rc = -ETIMEDOUT;
>>>> goto done;
>>>> }
>>>>
>>>> set_current_state(TASK_UNINTERRUPTIBLE);
>>>> schedule_timeout(wait_time);
>>>> + } else {
>>>> + break;
>>>> }
>>>> - } while (rc);
>>>> + } while (true);
>>> Use while (true) instead of do { } while (true) to make it clearer
>>> what's going on. Or refactor it to put the time_after in the while(),
>>> as this is what the loop is waiting on.
>>
>> OK, I went in circles (pun intended) working on this loop, but I'll try
>> and make it look better.
>>
>>
>>>> /* Extract size of response data */
>>>> resp_data_length = get_unaligned_be16(&resp->data_length);
>>>> @@ -543,17 +556,29 @@ int fsi_occ_submit(struct device *dev, const void *request, size_t req_len,
>>>> goto done;
>>>> }
>>>>
>>>> - dev_dbg(dev, "resp_status=%02x resp_data_len=%d\n",
>>>> - resp->return_status, resp_data_length);
>>>> -
>>>> - /* Grab the rest */
>>>> + /* Now get the entire response; get header again in case it changed */
>>>> if (resp_data_length > 1) {
>>>> - /* already got 3 bytes resp, also need 2 bytes checksum */
>>>> - rc = occ_getsram(occ, 8, &resp->data[3], resp_data_length - 1);
>>>> + rc = occ_getsram(occ, 0, resp, resp_data_length + 7);
>>>> if (rc)
>>>> goto done;
>>>> +
>>>> + if (resp->return_status == OCC_RESP_CMD_IN_PRG ||
>>>> + resp->return_status == OCC_RESP_CRIT_INIT ||
>>>> + resp->seq_no != seq_no || resp->cmd_type != cmd_type) {
>>>> + if (!retried) {
>>>> + retried = true;
>>>> + goto retry;
>>> Not sure about this.
>>>
>>> How often did this situation come up?
>>>
>>> Did you consider instead returning an error here?
>>
>> Well I can't say it's frequent, but hitting this condition was what
>> drove making this change in the first place. It needs to be handled.
> I was concerned about the pattern of using goto back up the function.
>
> Would it make more sense the have the caller retry, instead of adding
> retries in the sbefifo driver?


I've refactored it in v2. I don't think making the caller retry makes
sense, since it would be a lot of wasted work, since we only need to
re-read the response at that point.


>
>> Previously if this occurrred, we got a checksum error, so it effectively
>> already returned an error.
>>
>>
>>>> + }
>>>> +
>>>> + fsi_occ_print_timeout(occ, resp, seq_no, cmd_type);
>>>> + rc = -ETIMEDOUT;
>>>> + goto done;
>>>> + }
>>>> }
>>>>
>>>> + dev_dbg(dev, "resp_status=%02x resp_data_len=%d\n",
>>>> + resp->return_status, resp_data_length);
>>>> +
>>>> occ->client_response_size = resp_data_length + 7;
>>>> rc = occ_verify_checksum(occ, resp, resp_data_length);
>>>>
>>>> @@ -598,7 +623,7 @@ static int occ_probe(struct platform_device *pdev)
>>>> occ->version = (uintptr_t)of_device_get_match_data(dev);
>>>> occ->dev = dev;
>>>> occ->sbefifo = dev->parent;
>>>> - occ->sequence_number = 1;
>>>> + occ->sequence_number = (u8)((jiffies % 0xff) + 1);
>>> This is interesting. You didn't mention this in the commit message;
>>> you're trying to get a random number for the sequence number?
>>
>> Yea, this reduces the chances of hitting that retry above. If it's
>> always 1, then every time the driver is bound it tries the first command
>> with the same sequence number. This is a problem when FSI scanning with
>> the host already running, as the driver gets unbound/rebound several
>> times in a row, and we easily hit the checksum problem, since we proceed
>> to get the full response even though it's not for the latest command,
>> and then the buffer is updated at the same time. So using a non-zero
>> random number is very helpful.
> Makes sense. Perhaps do something like this instead of hand rolling it?
>
> get_random_bytes(occ->sequence_number, 1);


I thought about this, but I ended up just adding a comment and sticking
with jiffies. get_random_bytes seems a little heavy-handed, especially
since you're supposed to call wait_for_random_bytes first.


Thanks,

Eddie


>
> If you could add some of your explanations to the commit message, I'd
> like to get this fix merged soon.
>
> Cheers,
>
> Joel
>
>
>
>>
>> Thanks,
>>
>> Eddie
>>
>>
>>>> mutex_init(&occ->occ_lock);
>>>>
>>>> if (dev->of_node) {
>>>> --
>>>> 2.27.0
>>>>

\
 
 \ /
  Last update: 2022-02-08 16:27    [W:0.068 / U:2.192 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site