[PATCH] powerpc/mm: Don't kmap_atomic() in pte_offset_map() on PPC32

On PPC32, pte_offset_map() does a kmap_atomic() in order to support
page tables allocated in high memory, just like ARM and x86/32.

But since at least 2008 and commit 8054a3428fbe ("powerpc: Remove dead
CONFIG_HIGHPTE"), page tables are never allocated in high memory.

When the page is in low mem, kmap_atomic() just returns the page
address but still disable preemption and pagefault. And it is
not an inlined function, so we suffer function call for no reason.

Make pte_offset_map() the same as pte_offset_kernel() and make
pte_unmap() void, in the same way as PPC64 which doesn't have HIGHMEM.

Signed-off-by: Christophe Leroy <redacted>
---
 arch/powerpc/include/asm/book3s/32/pgtable.h | 6 ++----
 arch/powerpc/include/asm/nohash/32/pgtable.h | 6 ++----
 2 files changed, 4 insertions(+), 8 deletions(-)

diff --git a/arch/powerpc/include/asm/book3s/32/pgtable.h b/arch/powerpc/include/asm/book3s/32/pgtable.h
index 8bf8cdd92ea5..e490c1012c27 100644
--- a/arch/powerpc/include/asm/book3s/32/pgtable.h
+++ b/arch/powerpc/include/asm/book3s/32/pgtable.h

@@ -367,10 +367,8 @@ static inline void __ptep_set_access_flags(struct vm_area_struct *vma,
 	(((address) >> PAGE_SHIFT) & (PTRS_PER_PTE - 1))
 #define pte_offset_kernel(dir, addr)	\
 	((pte_t *) pmd_page_vaddr(*(dir)) + pte_index(addr))
-#define pte_offset_map(dir, addr)		\
-	((pte_t *)(kmap_atomic(pmd_page(*(dir))) + \
-		   (pmd_page_vaddr(*(dir)) & ~PAGE_MASK)) + pte_index(addr))
-#define pte_unmap(pte)		kunmap_atomic(pte)
+#define pte_offset_map(dir, addr)	pte_offset_kernel((dir), (addr))
+static inline void pte_unmap(pte_t *pte) { }
 
 /*
  * Encode and decode a swap entry.

diff --git a/arch/powerpc/include/asm/nohash/32/pgtable.h b/arch/powerpc/include/asm/nohash/32/pgtable.h
index 5ad7e439dcf4..fcd22737eb5b 100644
--- a/arch/powerpc/include/asm/nohash/32/pgtable.h
+++ b/arch/powerpc/include/asm/nohash/32/pgtable.h

@@ -373,10 +373,8 @@ static inline int pte_young(pte_t pte)
 #define pte_offset_kernel(dir, addr)	\
 	(pmd_bad(*(dir)) ? NULL : (pte_t *)pmd_page_vaddr(*(dir)) + \
 				  pte_index(addr))
-#define pte_offset_map(dir, addr)		\
-	((pte_t *)(kmap_atomic(pmd_page(*(dir))) + \
-		   (pmd_page_vaddr(*(dir)) & ~PAGE_MASK)) + pte_index(addr))
-#define pte_unmap(pte)		kunmap_atomic(pte)
+#define pte_offset_map(dir, addr)	pte_offset_kernel((dir), (addr))
+static inline void pte_unmap(pte_t *pte) { }
 
 /*
  * Encode and decode a swap entry.

-- 
2.25.0

On Mon, 2020-02-17 at 09:41:35 UTC, Christophe Leroy wrote:

On PPC32, pte_offset_map() does a kmap_atomic() in order to support
page tables allocated in high memory, just like ARM and x86/32.

But since at least 2008 and commit 8054a3428fbe ("powerpc: Remove dead
CONFIG_HIGHPTE"), page tables are never allocated in high memory.

When the page is in low mem, kmap_atomic() just returns the page
address but still disable preemption and pagefault. And it is
not an inlined function, so we suffer function call for no reason.

Make pte_offset_map() the same as pte_offset_kernel() and make
pte_unmap() void, in the same way as PPC64 which doesn't have HIGHMEM.

Signed-off-by: Christophe Leroy <redacted>

Applied to powerpc next, thanks.

https://git.kernel.org/powerpc/c/6453f9ed9d4e4b4cdf201bf34bf460c436bf50ea

cheers