[PATCH] powerpc/32s: Don't flush all TLBs when flushing one page
Subsystems:
linux for powerpc (32-bit and 64-bit) , the rest
2397d
8 messages,
2 authors,
2020-02-01 · open the first message on its own page
When flushing a range, the flushing function flushes all TLBs.
When the range is a single page, do a page flush instead.
Signed-off-by: Christophe Leroy <redacted>
---
arch/powerpc/mm/book3s32/tlb.c | 9 ++++++---
1 file changed, 6 insertions(+), 3 deletions(-)
diff --git a/arch/powerpc/mm/book3s32/tlb.c b/arch/powerpc/mm/book3s32/tlb.c
index 2fcd321040ff..8e0089065a7e 100644
--- a/arch/powerpc/mm/book3s32/tlb.c
+++ b/arch/powerpc/mm/book3s32/tlb.c @@ -79,14 +79,17 @@ static void flush_range(struct mm_struct *mm, unsigned long start,
int count ;
unsigned int ctx = mm -> context . id ;
+ start &= PAGE_MASK ; + end = ( end - 1 ) | ~ PAGE_MASK ; if ( ! Hash ) {
- _tlbia (); + if ( end - start == PAGE_SIZE ) + _tlbie ( start ); + else + _tlbia (); return ;
}
- start &= PAGE_MASK ; if ( start >= end )
return ;
- end = ( end - 1 ) | ~ PAGE_MASK ; pmd = pmd_offset ( pud_offset ( pgd_offset ( mm , start ), start ), start );
for (;;) {
pmd_end = (( start + PGDIR_SIZE ) & PGDIR_MASK ) - 1 ; --
2.25.0
On Fri, Jan 31, 2020 at 03:37:34PM +0000, Christophe Leroy wrote: When the range is a single page, do a page flush instead. + start &= PAGE_MASK;
+ end = (end - 1) | ~PAGE_MASK;
if (!Hash) {
- _tlbia();
+ if (end - start == PAGE_SIZE)
+ _tlbie(start);
+ else
+ _tlbia();
return;
}
For just one page, you get end - start == 0 actually?
Segher
Le 31/01/2020 à 16:51, Segher Boessenkool a écrit : On Fri, Jan 31, 2020 at 03:37:34PM +0000, Christophe Leroy wrote: quoted When the range is a single page, do a page flush instead. quoted + start &= PAGE_MASK;
+ end = (end - 1) | ~PAGE_MASK;
if (!Hash) {
- _tlbia();
+ if (end - start == PAGE_SIZE)
+ _tlbie(start);
+ else
+ _tlbia();
return;
}
For just one page, you get end - start == 0 actually?
Oops, good catch.
Indeed you don't get PAGE_SIZE but (PAGE_SIZE - 1) for just one page.
Christophe
On Fri, Jan 31, 2020 at 05:15:20PM +0100, Christophe Leroy wrote: Le 31/01/2020 à 16:51, Segher Boessenkool a écrit : quoted On Fri, Jan 31, 2020 at 03:37:34PM +0000, Christophe Leroy wrote: quoted When the range is a single page, do a page flush instead. quoted + start &= PAGE_MASK;
+ end = (end - 1) | ~PAGE_MASK;
if (!Hash) {
- _tlbia();
+ if (end - start == PAGE_SIZE)
+ _tlbie(start);
+ else
+ _tlbia();
return;
}
For just one page, you get end - start == 0 actually?
Oops, good catch.
Indeed you don't get PAGE_SIZE but (PAGE_SIZE - 1) for just one page.
You have all low bits masked off in both start and end, so you get zero.
You could make the condion read "if (start == end)?
Maybe a nicer way to describe what you do is "if start and end are on the
same memory page, flush that page."
Segher
Le 31/01/2020 à 20:38, Segher Boessenkool a écrit : On Fri, Jan 31, 2020 at 05:15:20PM +0100, Christophe Leroy wrote: quoted Le 31/01/2020 à 16:51, Segher Boessenkool a écrit : quoted On Fri, Jan 31, 2020 at 03:37:34PM +0000, Christophe Leroy wrote: quoted When the range is a single page, do a page flush instead. quoted + start &= PAGE_MASK;
+ end = (end - 1) | ~PAGE_MASK;
if (!Hash) {
- _tlbia();
+ if (end - start == PAGE_SIZE)
+ _tlbie(start);
+ else
+ _tlbia();
return;
}
For just one page, you get end - start == 0 actually?
Oops, good catch.
Indeed you don't get PAGE_SIZE but (PAGE_SIZE - 1) for just one page.
You have all low bits masked off in both start and end, so you get zero.
You could make the condion read "if (start == end)?
No, in end the low bits are set, that's a BIT OR with ~PAGE_MASK, so it
sets all low bits to 1.
Christophe
On Sat, Feb 01, 2020 at 08:27:03AM +0100, Christophe Leroy wrote: Le 31/01/2020 à 20:38, Segher Boessenkool a écrit : quoted On Fri, Jan 31, 2020 at 05:15:20PM +0100, Christophe Leroy wrote: quoted Le 31/01/2020 à 16:51, Segher Boessenkool a écrit : quoted On Fri, Jan 31, 2020 at 03:37:34PM +0000, Christophe Leroy wrote: quoted When the range is a single page, do a page flush instead. quoted + start &= PAGE_MASK;
+ end = (end - 1) | ~PAGE_MASK;
if (!Hash) {
- _tlbia();
+ if (end - start == PAGE_SIZE)
+ _tlbie(start);
+ else
+ _tlbia();
return;
}
For just one page, you get end - start == 0 actually?
Oops, good catch.
Indeed you don't get PAGE_SIZE but (PAGE_SIZE - 1) for just one page.
You have all low bits masked off in both start and end, so you get zero.
You could make the condion read "if (start == end)?
No, in end the low bits are set, that's a BIT OR with ~PAGE_MASK, so it
sets all low bits to 1.
Oh, wow, yes, I cannot read apparently.
Maybe there are some ROUND_DOWN and ROUND_UP macros you could use?
Segher
Le 01/02/2020 à 15:06, Segher Boessenkool a écrit : On Sat, Feb 01, 2020 at 08:27:03AM +0100, Christophe Leroy wrote: quoted Le 31/01/2020 à 20:38, Segher Boessenkool a écrit : quoted On Fri, Jan 31, 2020 at 05:15:20PM +0100, Christophe Leroy wrote: quoted Le 31/01/2020 à 16:51, Segher Boessenkool a écrit : quoted On Fri, Jan 31, 2020 at 03:37:34PM +0000, Christophe Leroy wrote: quoted When the range is a single page, do a page flush instead. quoted + start &= PAGE_MASK;
+ end = (end - 1) | ~PAGE_MASK;
if (!Hash) {
- _tlbia();
+ if (end - start == PAGE_SIZE)
+ _tlbie(start);
+ else
+ _tlbia();
return;
}
For just one page, you get end - start == 0 actually?
Oops, good catch.
Indeed you don't get PAGE_SIZE but (PAGE_SIZE - 1) for just one page.
You have all low bits masked off in both start and end, so you get zero.
You could make the condion read "if (start == end)?
No, in end the low bits are set, that's a BIT OR with ~PAGE_MASK, so it
sets all low bits to 1.
Oh, wow, yes, I cannot read apparently.
Maybe there are some ROUND_DOWN and ROUND_UP macros you could use?
Yes but my intention was to modify the existing code as less as possible.
What do you think about version v2 of the patch ?
Christophe
On Sat, Feb 01, 2020 at 03:53:12PM +0100, Christophe Leroy wrote: quoted quoted No, in end the low bits are set, that's a BIT OR with ~PAGE_MASK, so it
sets all low bits to 1.
Oh, wow, yes, I cannot read apparently.
Maybe there are some ROUND_DOWN and ROUND_UP macros you could use?
Yes but my intention was to modify the existing code as less as possible.
What do you think about version v2 of the patch ?
It looked fine to me.
Add my
Reviewed-by: Segher Boessenkool <redacted>
if you want.
Segher