]> git.baikalelectronics.ru Git - kernel.git/commit
parisc: Optimize per-pagetable spinlocks
authorHelge Deller <deller@gmx.de>
Fri, 12 Feb 2021 15:38:52 +0000 (16:38 +0100)
committerHelge Deller <deller@gmx.de>
Fri, 12 Feb 2021 15:39:42 +0000 (16:39 +0100)
commit8e41de55d8d7662ba04e38ba8e99330cf9a1e618
tree5d4431b00e3416bc66f6fa1fc1a21cc530da4568
parent5a3f22cb78fda1faaf1980518828b6c11edc8dee
parisc: Optimize per-pagetable spinlocks

On parisc a spinlock is stored in the next page behind the pgd which
protects against parallel accesses to the pgd. That's why one additional
page (PGD_ALLOC_ORDER) is allocated for the pgd.

Matthew Wilcox suggested that we instead should use a pointer in the
struct page table for this spinlock and noted, that the comments for the
PGD_ORDER and PMD_ORDER defines were wrong.

Both suggestions are addressed with this patch. Instead of having an own
spinlock to protect the pgd, we now switch to use the existing
page_table_lock.  Additionally, beside loading the pgd into cr25 in
switch_mm_irqs_off(), the physical address of this lock is loaded into
cr28 (tr4), so that we can avoid implementing a complicated lookup in
assembly for this lock in the TLB fault handlers.

The existing Hybrid L2/L3 page table scheme (where the pmd is adjacent
to the pgd) has been dropped with this patch.

Remove the locking in set_pte() and the huge-page pte functions too.
They trigger a spinlock recursion on 32bit machines and seem unnecessary.

Suggested-by: Matthew Wilcox <willy@infradead.org>
Fixes: 3bf7047aebbb ("parisc: Use per-pagetable spinlock")
Signed-off-by: John David Anglin <dave.anglin@bell.net>
Signed-off-by: Helge Deller <deller@gmx.de>
arch/parisc/Kconfig
arch/parisc/include/asm/mmu_context.h
arch/parisc/include/asm/page.h
arch/parisc/include/asm/pgalloc.h
arch/parisc/include/asm/pgtable.h
arch/parisc/kernel/asm-offsets.c
arch/parisc/kernel/entry.S
arch/parisc/mm/hugetlbpage.c
arch/parisc/mm/init.c