[Intel-gfx] [PATCH v3 1/4] drm/i915: Always do WOPCM partitioning based on real firmware sizes

Thu Apr 19 21:17:44 UTC 2018

On 04/19/2018 08:31 AM, Michal Wajdeczko wrote:
> On Mon, 16 Apr 2018 19:28:04 +0200, Yaodong Li <yaodong.li at intel.com> 
> wrote:
>
>> On 04/13/2018 07:15 PM, Michal Wajdeczko wrote:
>>> On Tue, 10 Apr 2018 02:42:17 +0200, Jackie Li <yaodong.li at intel.com> 
>>> wrote:
>>>
>>>> After enabled the WOPCM write-once registers locking status checking,
>>>> reloading of the i915 module will fail with modparam enable_guc set 
>>>> to 3
>>>> (enable GuC and HuC firmware loading) if the module was originally 
>>>> loaded
>>>> with enable_guc set to 1 (only enable GuC firmware loading).
>>>
>>> Is this frequent and required scenario ? or just for 
>>> debug/development ?
>>>
>> My understanding is this should be a nice to have feature and mainly 
>> for debugging.
>>>> This is
>>>> because WOPCM registers were updated and locked without considering 
>>>> the HuC
>>>> FW size. Since we need both GuC and HuC FW sizes to determine the 
>>>> final
>>>> layout of WOPCM, we should always calculate the WOPCM layout based 
>>>> on the
>>>> actual sizes of the GuC and HuC firmware available for a specific 
>>>> platform
>>>> if we need continue to support enable/disable HuC FW loading 
>>>> dynamically
>>>> with enable_guc modparam.
>>>>
>>>> This patch splits uC firmware fetching into two stages. First stage 
>>>> is to
>>>> fetch the firmware image and verify the firmware header. uC 
>>>> firmware will
>>>> be marked as verified and this will make FW info available for 
>>>> following
>>>> WOPCM layout calculation. The second stage is to create a GEM 
>>>> object and
>>>> copy the FW data into the created GEM object which will only be 
>>>> available
>>>> when GuC/HuC loading is enabled by enable_guc modparam. This will 
>>>> guarantee
>>>> that the WOPCM layout will be always be calculated correctly 
>>>> without making
>>>> any assumptions to the GuC and HuC firmware sizes.
>>>
>>> You are also assuming that on reload exactly the same GuC/HuC firmwares
>>> will bee used as in initial run. This will make this useless for debug/
>>> development scenarios, where custom fw are likely to be specified.
>>>
>> This patch is mainly for providing a real fix to support 
>> enable_guc=1->3->1 use case.
>> It based on the fact that it is inevitable that sometimes we need to 
>> reboot the system
>> if the status of the fw was changed on the file system.
>
> What do you mean by "status of the fw was changed on the file system" ?
> * change of the fw binary/version/size, or
> * change from not-present to present ?
I think it should include all of the presence changes, file updates.
>
>> I am not sure how often we switch between different HuC FW with 
>> different sizes?
>
> Just above you said that you need this "mainly for debugging" so
> I would expect that then different fw sizes are expected.
>
>>> If we want to support enable_guc=1->3->1 scenarios for debug/dev then
>>> maybe more flexible will be other approach that makes allocations from
>>> the other end as proposed in [1]
>>>
>>> [1] https://patchwork.freedesktop.org/patch/212471/
>> Actually, I do think this might be one of the options, and I've also 
>> put some comments on this
>> series. The main concern I have is it still make assumption on the 
>> GuC FW size and may
>
> But in enable_guc=1-->3 scenario, I would assume that the only difference
> will be HuC fw (as with enable=1 we already loaded GuC)
Hmm, my main concern to the current "from the end" solution is it makes 
assumption on
the GuC FW size in order to meet the HW restriction.
>
> If you want just to test different GuC fws, then it is different scenario
> as then enable_guc will always be = 1.
>
what I mean is the "from the end" approach will lead to the same issue 
for different GuC FW sizes - we
may have to reboot the system for GuC FW debugging (different GuC FW 
sizes) even if enable_guc is always
set to 1. However, with the current "from the beginning" way we won't 
run into such problems
for GuC FW debugging (since it's already used the max available space). 
Thus I think we should
define the enable_guc = 1->3->1 as following:
we would support use of enable_guc=1->3->1 correctly without system 
reboot for the present FWs. A system
reboot will be expected (but not necessarily happen if we found current 
partition works for the new FWs)
for any FW changes (including add/remove/update).

if we decide to drop the support for enable_guc=1->3->1 since it's only 
for debugging purpose then we should
expect a system reboot for either "from the end" or "from the beginning" 
solutions since we cannot 100% have
this issue (changing FW without a system boot) solved. Therefore, the 
require of system reboot should not be
a bug when it comes to FW updating.

>> run into the same issue if the GuC FW failed to meet the requirement.
>> and for debugging purpose it would have the same possible for 
>> different GuC FW debugging.
>>>
>>>>
>>>> v3:
>>>>  - Rebase
>>>>
>>>> Signed-off-by: Jackie Li <yaodong.li at intel.com>
>>>> Cc: Michal Wajdeczko <michal.wajdeczko at intel.com>
>>>> Cc: Sagar Arun Kamble <sagar.a.kamble at intel.com>
>>>> Cc: Michal Winiarski <michal.winiarski at intel.com>
>>>> Cc: John Spotswood <john.a.spotswood at intel.com>
>>>> Cc: Joonas Lahtinen <joonas.lahtinen at linux.intel.com>
>>>> ---
>>>>  drivers/gpu/drm/i915/intel_uc.c    | 14 ++++----------
>>>>  drivers/gpu/drm/i915/intel_uc_fw.c | 31 
>>>> ++++++++++++++++++++-----------
>>>>  drivers/gpu/drm/i915/intel_uc_fw.h |  7 +++++--
>>>>  3 files changed, 29 insertions(+), 23 deletions(-)
>>>>
>>>> diff --git a/drivers/gpu/drm/i915/intel_uc.c 
>>>> b/drivers/gpu/drm/i915/intel_uc.c
>>>> index 1cffaf7..73b8f6c 100644
>>>> --- a/drivers/gpu/drm/i915/intel_uc.c
>>>> +++ b/drivers/gpu/drm/i915/intel_uc.c
>>>> @@ -172,11 +172,8 @@ void intel_uc_init_early(struct 
>>>> drm_i915_private *i915)
>>>>     sanitize_options_early(i915);
>>>> -    if (USES_GUC(i915))
>>>> -        intel_uc_fw_fetch(i915, &guc->fw);
>>>> -
>>>> -    if (USES_HUC(i915))
>>>> -        intel_uc_fw_fetch(i915, &huc->fw);
>>>> +    intel_uc_fw_fetch(i915, &guc->fw, USES_GUC(i915));
>>>> +    intel_uc_fw_fetch(i915, &huc->fw, USES_HUC(i915));
>>>
>>> Hmm, side effect of those unconditional fetches might be unwanted 
>>> warnings
>>> about missing firmwares (on configs with disabled guc) as well as 
>>> extended
>>> driver load time.
>> Hmm, if HAS_GUC is false then fw path would be NULL. The fetch will 
>> return directly.
>
> I was referring to scenario when on platform with HAS_HUC and with
> enable_guc=1 (just submission, no HuC) we will try to fetch HuC fw
> (that may not be present at all) and then drop it as don't need it.
>
I think there are two scenarios here for this specific case - a platform 
with HAS_HUC = 1  and only GuC submission is needed:
0) No HuC FW available - We should expect a system reboot for adding new FW.
1) If HuC FW is present - always get the FW header info in order to 
support possible enable_guc=1->3->1.

IMHO, the problem we have here is that we need to define the use case 
precisely. e.g. whether we shall support
enable_guc=1->3->1 flawlessly? and whether we shall support dynamic HuC 
FW sizes for debugging rather than
supporting dynamic GuC FW sizes for debugging purpose?
>>>
>>> Do we really need to support this corner case enable_guc=1->3 at all 
>>> costs?
>> I think this is the real solution for this issue (with no 
>> assumption). However, we do
>> need to decide whether we should support such a corner case which is 
>> mainly for
>> debugging.
>
> I'm repeating here Joonas' earlier statement:
>
> "Then just require a reboot if after that partitioning,
>  changing the parameter causes the FW not to fit"
>
That's my thought too:)

Regards,
-Jackie