[PATCH v2 9/9] drm/i915: Use kmap_local_page() in gem/i915_gem_execbuffer.c

[Tvrtko Ursulin]

On 31/03/2023 16:32, Fabio M. De Francesco wrote:
> On venerdì 31 marzo 2023 13:30:20 CEST Tvrtko Ursulin wrote:
>> On 31/03/2023 05:18, Ira Weiny wrote:
>>> Zhao Liu wrote:
>>>> From: Zhao Liu <zhao1.liu at intel.com>
>>>>
>>>> The use of kmap_atomic() is being deprecated in favor of
>>>> kmap_local_page()[1], and this patch converts the calls from
>>>> kmap_atomic() to kmap_local_page().
>>>>
>>>> The main difference between atomic and local mappings is that local
>>>> mappings doesn't disable page faults or preemption (the preemption is
>>>> disabled for !PREEMPT_RT case, otherwise it only disables migration).
>>>>
>>>> With kmap_local_page(), we can avoid the often unwanted side effect of
>>>> unnecessary page faults and preemption disables.
>>>>
>>>> In i915_gem_execbuffer.c, eb->reloc_cache.vaddr is mapped by
>>>> kmap_atomic() in eb_relocate_entry(), and is unmapped by
>>>> kunmap_atomic() in reloc_cache_reset().
>>>
>>> First off thanks for the series and sticking with this.  That said this
>>> patch kind of threw me for a loop because tracing the map/unmap calls did
>>> not make sense to me.  See below.
>>>
>>>> And this mapping/unmapping occurs in two places: one is in
>>>> eb_relocate_vma(), and another is in eb_relocate_vma_slow().
>>>>
>>>> The function eb_relocate_vma() or eb_relocate_vma_slow() doesn't
>>>> need to disable pagefaults and preemption during the above mapping/
>>>> unmapping.
>>>>
>>>> So it can simply use kmap_local_page() / kunmap_local() that can
>>>> instead do the mapping / unmapping regardless of the context.
>>>>
>>>> Convert the calls of kmap_atomic() / kunmap_atomic() to
>>>> kmap_local_page() / kunmap_local().
>>>>
>>>> [1]:
>>>> https://lore.kernel.org/all/20220813220034.806698-1-ira.weiny@intel.com
>>>>
>>>> v2: No code change since v1. Added description of the motivation of
>>>>
>>>>       using kmap_local_page() and "Suggested-by" tag of Fabio.
>>>>
>>>> Suggested-by: Ira Weiny <ira.weiny at intel.com>
>>>> Suggested-by: Fabio M. De Francesco <fmdefrancesco at gmail.com>
>>>> Signed-off-by: Zhao Liu <zhao1.liu at intel.com>
>>>> ---
>>>>
>>>> Suggested by credits:
>>>>     Ira: Referred to his task document, review comments.
>>>>     Fabio: Referred to his boiler plate commit message and his description
>>>>     
>>>>            about why kmap_local_page() should be preferred.
>>>>
>>>> ---
>>>>
>>>>    drivers/gpu/drm/i915/gem/i915_gem_execbuffer.c | 10 +++++-----
>>>>    1 file changed, 5 insertions(+), 5 deletions(-)
>>>>
> 
> [snip]
>   
>> However I am unsure if disabling pagefaulting is needed or not. Thomas,
>> Matt, being the last to touch this area, perhaps you could have a look?
>> Because I notice we have a fallback iomap path which still uses
>> io_mapping_map_atomic_wc. So if kmap_atomic to kmap_local conversion is
>> safe, does the iomap side also needs converting to
>> io_mapping_map_local_wc? Or they have separate requirements?
> 
> AFAIK, the requirements for io_mapping_map_local_wc() are the same as for
> kmap_local_page(): the kernel virtual address is _only_ valid in the caller
> context, and map/unmap nesting must be done in stack-based ordering (LIFO).
> 
> I think a follow up patch could safely switch to io_mapping_map_local_wc() /
> io_mapping_unmap_local_wc since the address is local to context.
> 
> However, not being an expert, reading your note now I suspect that I'm missing
> something. Can I ask why you think that page-faults disabling might be
> necessary?

I am not saying it is, was just unsure and wanted some people who worked on this code most recently to take a look and confirm.

I guess it will work since the copying is done like this anyway:

		/*
		 * This is the fast path and we cannot handle a pagefault
		 * whilst holding the struct mutex lest the user pass in the
		 * relocations contained within a mmaped bo. For in such a case
		 * we, the page fault handler would call i915_gem_fault() and
		 * we would try to acquire the struct mutex again. Obviously
		 * this is bad and so lockdep complains vehemently.
		 */
		pagefault_disable();
		copied = __copy_from_user_inatomic(r, urelocs, count * sizeof(r[0]));
		pagefault_enable();
		if (unlikely(copied)) {
			remain = -EFAULT;
			goto out;
		}

Comment is a bit outdated since we don't use that global "struct mutex" any longer, but in any case, if there is a page fault on the mapping where we need to recurse into i915 again to satisfy if, we seem to have code already to handle it. So kmap_local conversion I *think* can't regress anything.

Patch to convert the io_mapping_map_atomic_wc can indeed come later.

In terms of logistics - if we landed this series to out branch it would be queued only for 6.5. Would that work for you?

Regards,

Tvrtko