On 12/15/2017 10:53 AM, Christian König wrote:
>
>> Well this is more or less replicating what you are doing currently
>> but instead of spreading the checks and locking state all over the
>> code, both as local variables and parameters this is keeping it in a
>> single place with correctness checks.
>
> I don't see how that can be correct.
In what sense? It should be doing *exactly* what you're doing now but
use an abstraction. No more no less. But with correctness checks, and
less of passing state around. Specifically what I'm referring to are
things like the locking test this patch is proposing, the "locked"
variable in ttm_mem_evict_first(), and the parameter special code in
ttm_mem_cleanup_refs() that I think are hard to follow and error prone.
Conditional locking like in ttm_mem_clenup_refs() also typically trip
static lock balance checkers.
> As far as I can see it makes TTM responsible about locking again
No it doesn't. It's a helper. The only TTM state in my suggestion was
the "bo::reserved"Â debug variable, which could either be ommited or
count the recursive reservation to aid making sure that all recursive
reservations you do in TTM are undone in TTM.
> and to be honest TTM is a bloody mess regarding this already.
Hmm. IMO strong unspecific wording like this in a design descussions
should be avoided. It tends to take away focus from what's actually
being dicussed by making either part feel bad. And it's typically
unproductive.
>
> My intention in the long term is rather to remove logic from TTM and
> move it back into the drivers. The end result I have in mind is that
> TTM becomes a toolbox instead of the midlayer it is currently.
I'm in favour of that. But I don't think what I proposed is a step away
from that direction. On the contrary. I've attached a POC patch with
the correctness checks stripped, not compile-tested. Much easier to
follow if you ask me, but if you feel so strongly against it, never mind.
Thanks,
Thomas
>
> Regards,
> Christian. qq
>
>>
>>
>> I agree recursive locking is generally frowned upon, but you're
>> already doing it, not by using recursive locks, but by passing
>> locking state around which IMO is worse.
>>
>> Collecting the state in a the operation_ctx will make that
>> usage-pattern more obvious but will help make the code cleaner and
>> less error prone.
>>
>> /Thomas
>>
>>
>>
>>>
>>> Regards,
>>> Christian.
>>>
>>> Am 15.12.2017 um 08:01 schrieb Thomas Hellstrom:
>>>> Roger and Chrisitian,
>>>>
>>>> Correct me if I'm wrong, but It seems to me like a lot of the
>>>> recent changes to ttm_bo.c are to allow recursive reservation
>>>> object locking in the case of shared reservation objects, but only
>>>> in certain functions and with special arguments so it doesn't look
>>>> like recursive locking to the lockdep checker. Wouldn't it be a lot
>>>> cleaner if we were to hide all this in a resurrected
>>>> __ttm_bo_reserve something along the lines of
>>>>
>>>> int __ttm_bo_reserve(struct ttm_bo *bo, struct ttm_operation_ctx
>>>> *ctx) {
>>>> Â Â Â if (ctx && ctx->resv == bo->resv) {
>>>> #ifdef CONFIG_LOCKDEP
>>>> Â Â Â Â Â Â Â WARN_ON(bo->reserved);
>>>> lockdep_assert_held(&ctx->resv);
>>>> Â Â Â Â Â Â ctx->reserve_count++;
>>>> bo->reserved = true;
>>>> #endif
>>>> Â Â Â Â Â Â return0;
>>>> Â Â Â Â } else {
>>>> Â Â Â Â Â Â Â int ret = reservation_object_lock(bo->resv, NULL) ? 0:-EBUSY;
>>>>
>>>> Â Â Â Â Â Â Â if (ret)
>>>> Â Â Â Â Â Â Â Â Â return ret;
>>>> #ifdef CONFIG_LOCKDEP
>>>> Â Â Â Â Â Â WARN_ON(bo->reserved);
>>>> Â Â Â Â Â Â bo->reserved = true;
>>>> #endif
>>>> Â Â Â Â Â Â return 0;
>>>> }
>>>>
>>>> And similar for tryreserve and unreserve? Perhaps with a
>>>> ww_acquire_ctx included somewhere as well...
>>>>
>>>> /Thomas
>>>>
>>>>
>>>>
>>>>
>>>> On 12/14/2017 09:10 AM, Roger He wrote:
>>>>> Change-Id: I0c6ece0decd18d30ccc94e5c7ca106d351941c62
>>>>> Signed-off-by: Roger He <Hongbo.He at amd.com>
>>>>> ---
>>>>> Â drivers/gpu/drm/ttm/ttm_bo.c | 11 +++++------
>>>>> Â 1 file changed, 5 insertions(+), 6 deletions(-)
>>>>>
>>>>> diff --git a/drivers/gpu/drm/ttm/ttm_bo.c
>>>>> b/drivers/gpu/drm/ttm/ttm_bo.c
>>>>> index 098b22e..ba5b486 100644
>>>>> --- a/drivers/gpu/drm/ttm/ttm_bo.c
>>>>> +++ b/drivers/gpu/drm/ttm/ttm_bo.c
>>>>> @@ -707,7 +707,6 @@ bool ttm_bo_eviction_valuable(struct
>>>>> ttm_buffer_object *bo,
>>>>> Â EXPORT_SYMBOL(ttm_bo_eviction_valuable);
>>>>> Â Â static int ttm_mem_evict_first(struct ttm_bo_device *bdev,
>>>>> -Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â struct reservation_object *resv,
>>>>> Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â uint32_t mem_type,
>>>>> Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â const struct ttm_place *place,
>>>>> Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â struct ttm_operation_ctx *ctx)
>>>>> @@ -722,8 +721,9 @@ static int ttm_mem_evict_first(struct
>>>>> ttm_bo_device *bdev,
>>>>> Â Â Â Â Â spin_lock(&glob->lru_lock);
>>>>> Â Â Â Â Â for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i) {
>>>>> Â Â Â Â Â Â Â Â Â list_for_each_entry(bo, &man->lru[i], lru) {
>>>>> -Â Â Â Â Â Â Â Â Â Â Â if (bo->resv == resv) {
>>>>> -Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â if (list_empty(&bo->ddestroy))
>>>>> +Â Â Â Â Â Â Â Â Â Â Â if (bo->resv == ctx->resv) {
>>>>> +Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â if (!ctx->allow_reserved_eviction &&
>>>>> +Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â list_empty(&bo->ddestroy))
>>>>> Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â continue;
>>>>> Â Â Â Â Â Â Â Â Â Â Â Â Â } else {
>>>>> Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â locked = reservation_object_trylock(bo->resv);
>>>>> @@ -835,7 +835,7 @@ static int ttm_bo_mem_force_space(struct
>>>>> ttm_buffer_object *bo,
>>>>> Â Â Â Â Â Â Â Â Â Â Â Â Â return ret;
>>>>> Â Â Â Â Â Â Â Â Â if (mem->mm_node)
>>>>> Â Â Â Â Â Â Â Â Â Â Â Â Â break;
>>>>> -Â Â Â Â Â Â Â ret = ttm_mem_evict_first(bdev, bo->resv, mem_type,
>>>>> place, ctx);
>>>>> +Â Â Â Â Â Â Â ret = ttm_mem_evict_first(bdev, mem_type, place, ctx);
>>>>> Â Â Â Â Â Â Â Â Â if (unlikely(ret != 0))
>>>>> Â Â Â Â Â Â Â Â Â Â Â Â Â return ret;
>>>>> Â Â Â Â Â } while (1);
>>>>> @@ -1332,8 +1332,7 @@ static int ttm_bo_force_list_clean(struct
>>>>> ttm_bo_device *bdev,
>>>>> Â Â Â Â Â for (i = 0; i < TTM_MAX_BO_PRIORITY; ++i) {
>>>>> Â Â Â Â Â Â Â Â Â while (!list_empty(&man->lru[i])) {
>>>>> Â Â Â Â Â Â Â Â Â Â Â Â Â spin_unlock(&glob->lru_lock);
>>>>> -Â Â Â Â Â Â Â Â Â Â Â ret = ttm_mem_evict_first(bdev, NULL, mem_type,
>>>>> -Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â NULL, &ctx);
>>>>> +Â Â Â Â Â Â Â Â Â Â Â ret = ttm_mem_evict_first(bdev, mem_type, NULL, &ctx);
>>>>> Â Â Â Â Â Â Â Â Â Â Â Â Â if (ret)
>>>>> Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â return ret;
>>>>> Â Â Â Â Â Â Â Â Â Â Â Â Â spin_lock(&glob->lru_lock);
>>>>
>>>>
>>
>> _______________________________________________
>> amd-gfx mailing list
>> amd-gfx at lists.freedesktop.org
>> https://lists.freedesktop.org/mailman/listinfo/amd-gfx
-------------- next part --------------
A non-text attachment was scrubbed...
Name: 0001-drm-ttm-tryreserve_shared-POC.patch
Type: text/x-patch
Size: 5550 bytes
Desc: not available
URL: <https://lists.freedesktop.org/archives/amd-gfx/attachments/20171215/c93f03f2/attachment-0001.bin>
