On martedì 20 giugno 2023 20:01:48 CEST Sumitra Sharma wrote:
> Remove unnecessary calls to kmap{,_atomic}() when acquiring
> pages using GFP_DMA32.
>
> The GFP_DMA32 uses the DMA32 zone to satisfy the allocation
> requests. Therefore, pages allocated with GFP_DMA32 cannot
> come from Highmem.
>
> Avoid using calls to kmap_local_page() / kunmap_local() and
> kmap() / kunmap() in the psb_mmu_alloc_pd function. Instead,
> utilize page_address().
>
> Remove the usage of kmap_atomic() / kunmap_atomic() in the
> psb_mmu_alloc_pt function. Use page_address() instead.
>
> Substitute kmap_atomic(pt->p) / kunmap_atomic(pt->v) calls
> in the psb_mmu_pt_alloc_map_lock() and psb_mmu_pt_unmap_unlock()
> functions with page_address(pt->p). This is possible as
> pt = psb_mmu_alloc_pt(pd) allocates a page using
> pt->p = alloc_page(GFP_DMA32).
Sumitra,
I'm sorry because this patch cannot acked with this commit message.
This commit message is missing two _really_ important information. Therefore,
it is not acked. Please check again what I write below and either add the
missing information or change the code accordingly...
You should assure everybody that the code between the old kmap_atomic() /
kunmap_atomic() doesn't depend either on implicit pagefault_disable() or
preempt_disable() calls or both.
Please read again the section of the Highmem's documentation regarding
kmap_atomic() at https://docs.kernel.org/mm/highmem.html
In particular take care to read and understand what "[] the code between calls
to kmap_atomic() and kunmap_atomic() may implicitly depend on the side effects
of atomic mappings, i.e. disabling page faults or preemption, or both. In that
case, explicit calls to pagefault_disable() or preempt_disable() or both must
be made in conjunction with the use of kmap_local_page().".
Please study carefully also the following patch from Zhao, suggested by Ira
and reviewed by Ira and I: "[PATCH v2 3/9] drm/i915: Use kmap_local_page() in
gem/i915_gem_shmem.c". It's not yet reached upstream so you need to find it in
lore.kernel.org at https://lore.kernel.org/lkml/20230329073220.3982460-4-zhao1.liu@xxxxxxxxxxxxxxx/
Please note that, in turn, that patch also contains a link to a patch from Ira
who too had to disable faults (https://lore.kernel.org/all/
20220813220034.806698-1-ira.weiny@xxxxxxxxx)
I haven't yet looked at your code. If any sections do depend on those implicit
disables, you should act accordingly and add one or both of the above-
mentioned calls, even in cases where you get rid of local mappings.
Instead if the sections don't depend on the mentioned side effects, you should
write something like what I wrote in "[PATCH] NFS: Convert kmap_atomic() to
kmap_local_folio()" that you can find at https://lore.kernel.org/lkml/
20230503172411.3356-1-fmdefrancesco@xxxxxxxxx/ or, by by using "git show
4b71e2416ec4".
Thanks for working on this,
Fabio
>
> Suggested-by: Ira Weiny <ira.weiny@xxxxxxxxx>
> Signed-off-by: Sumitra Sharma <sumitraartsy@xxxxxxxxx>
> ---
> drivers/gpu/drm/gma500/mmu.c | 17 +++++------------
> 1 file changed, 5 insertions(+), 12 deletions(-)
>
> diff --git a/drivers/gpu/drm/gma500/mmu.c b/drivers/gpu/drm/gma500/mmu.c
> index a70b01ccdf70..59aa5661e56a 100644
> --- a/drivers/gpu/drm/gma500/mmu.c
> +++ b/drivers/gpu/drm/gma500/mmu.c
> @@ -184,20 +184,15 @@ struct psb_mmu_pd *psb_mmu_alloc_pd(struct
> psb_mmu_driver *driver, pd->invalid_pte = 0;
> }
>
> - v = kmap_local_page(pd->dummy_pt);
> + v = page_address(pd->dummy_pt);
> for (i = 0; i < (PAGE_SIZE / sizeof(uint32_t)); ++i)
> v[i] = pd->invalid_pte;
>
> - kunmap_local(v);
> -
> - v = kmap_local_page(pd->p);
> + v = page_address(pd->p);
> for (i = 0; i < (PAGE_SIZE / sizeof(uint32_t)); ++i)
> v[i] = pd->invalid_pde;
>
> - kunmap_local(v);
> -
> - clear_page(kmap(pd->dummy_page));
> - kunmap(pd->dummy_page);
> + clear_page(page_address(pd->dummy_page));
>
> pd->tables = vmalloc_user(sizeof(struct psb_mmu_pt *) * 1024);
> if (!pd->tables)
> @@ -279,7 +274,7 @@ static struct psb_mmu_pt *psb_mmu_alloc_pt(struct
> psb_mmu_pd *pd)
>
> spin_lock(lock);
>
> - v = kmap_atomic(pt->p);
> + v = page_address(pt->p);
> clf = (uint8_t *) v;
> ptes = (uint32_t *) v;
> for (i = 0; i < (PAGE_SIZE / sizeof(uint32_t)); ++i)
> @@ -293,7 +288,6 @@ static struct psb_mmu_pt *psb_mmu_alloc_pt(struct
> psb_mmu_pd *pd) }
> mb();
> }
> - kunmap_atomic(v);
> spin_unlock(lock);
>
> pt->count = 0;
> @@ -339,7 +333,7 @@ static struct psb_mmu_pt
*psb_mmu_pt_alloc_map_lock(struct
> psb_mmu_pd *pd, atomic_set(&pd->driver->needs_tlbflush, 1);
> }
> }
> - pt->v = kmap_atomic(pt->p);
> + pt->v = page_address(pt->p);
> return pt;
> }
>
> @@ -365,7 +359,6 @@ static void psb_mmu_pt_unmap_unlock(struct psb_mmu_pt
*pt)
> struct psb_mmu_pd *pd = pt->pd;
> uint32_t *v;
>
> - kunmap_atomic(pt->v);
> if (pt->count == 0) {
> v = kmap_atomic(pd->p);
> v[pt->index] = pd->invalid_pde;
> --
> 2.25.1
