tlb_finish_mmu() bogisity
tlb->need_flush += &__get_cpu_var(quicklist)[0].nr_pages != 0; makes no sense whatsoever. How the hell can you ever get the address of __get_cpu_var(quicklist)[0].nr_pages to be NULL? Postfix operators have higher precedence than prefix ones, so that's &(((__get_cpu_var(quicklist))[0]).nr_pages)
What did you intend here? s/&//, perhaps?
Quoted 8 lines · Al Viro
tlb->need_flush += &__get_cpu_var(quicklist)[0].nr_pages != 0; makes no sense whatsoever. How the hell can you ever get the address of __get_cpu_var(quicklist)[0].nr_pages to be NULL? Postfix operators have higher precedence than prefix ones, so that's &(((__get_cpu_var(quicklist))[0]).nr_pages)What did you intend here? s/&//, perhaps?
I think that thing is bogus in other ways. It plays games with "needs_flush", just because it seems to want the generic code to then call "check_pgt_cache()", not because it actually wants any flushing to take place.
But we already call "check_pgt_cache()" there in tlb_finish_mmu() unconditionally, so I think the whole patch was utter crap.
Or is there any other reason going on here? That quicklist stuff has been totally broken, it's done too many totally invalid things to really be worth even keeping. We should get rid of that unmaintainable hack, and if it has a real performance upside, we should make it part of the *native* mmu-gather infrastructure, instead of maintaining it as some separate and buggy/unmaintainable piece-of-sh*t code.
Linus
Quoted 6 lines · Linus Torvalds
Quoted 1 line
What did you intend here? s/&//, perhaps?
I think that thing is bogus in other ways. It plays games with "
needs_flush", just because it seems to want the generic code to then call "check_pgt_cache()", not because it actually wants any flushing to take place.
It can only free the pages on the quicklist after the flush. So it needs to set need_flush to trigger the flush before check_pgt_cache can trim the quicklist. If need_flush is not set then pages are freed in check_pgt_cache before their TLBs have been flushed.
Quoted 6 lines
Or is there any other reason going on here? That quicklist stuff has been totally broken, it's done too many totally invalid things to really be worth even keeping. We should get rid of that unmaintainable hack, and if it has a real performance upside, we should make it part of the *native* mmu-gather infrastructure, instead of maintaining it as some separate and buggy/unmaintainable piece-of-sh*t code.
Yup as we agreed before it would be best to make the TLB pieces part of the native mmu gather infrastructure. But then they vary quite a lot between arches.
Quicklist: Remove bogus &
include/asm-generic/tlb.h +1 -1
@@ -87,7 +87,7 @@ static inline void tlb_finish_mmu(struct mmu_gather *tlb, unsigned long start, unsigned long end) { #ifdef CONFIG_QUICKLIST - tlb->need_flush += &__get_cpu_var(quicklist)[0].nr_pages != 0; + tlb->need_flush += __get_cpu_var(quicklist)[0].nr_pages != 0;#endif tlb_flush_mmu (tlb, start, end);
Argh. This is indeed bogus. The one reporting the problem states that the patch did not address the issue. The report was regarding pgd freeing which is handled slightly differently from pte frees.
[PATCH] Revert quicklist need->flush fix
Did not fix the reported issue. Apart from other weirdness this causes a bad link between the TLB flushing logic and the quicklists. If there is indeed an issue that an arch needs a tlb flush before free then the arch code needs to set tlb->need_flush before calling quicklist_free.
include/asm-generic/tlb.h +0 -3
@@ -86,9 +86,6 @@ tlb_flush_mmu(struct mmu_gather *tlb, unstatic inline void tlb_finish_mmu (struct mmu_gather *tlb,unsigned long start,unsigned long end) { -#ifdef CONFIG_QUICKLIST - tlb->need_flush += &__get_cpu_var (quicklist)[0 ].nr_pages !=0 ; -#endif tlb_flush_mmu (tlb, start, end);/* keep the page table cache within bounds */