Re: mm: Question about the use of 'accessed' flags and pte_young() helper
On Tue, Oct 20, 2020 at 05:52:07PM +0200, Vlastimil Babka wrote: > On 10/8/20 11:49 AM, Christophe Leroy wrote: > > In a 10 years old commit > > (https://github.com/linuxppc/linux/commit/d069cb4373fe0d451357c4d3769623a7564dfa9f), > > powerpc 8xx has > > made the handling of PTE accessed bit conditional to CONFIG_SWAP. > > Since then, this has been extended to some other powerpc variants. > > > > That commit means that when CONFIG_SWAP is not selected, the accessed bit > > is not set by SW TLB miss > > handlers, leading to pte_young() returning garbage, or should I say > > possibly returning false > > allthough a page has been accessed since its access flag was reset. > > > > Looking at various mm/ places, pte_young() is used independent of > > CONFIG_SWAP > > > > Is it still valid the not manage accessed flags when CONFIG_SWAP is not > > selected ? > > AFAIK it's wrong, reclaim needs it to detect accessed pages on inactive > list, via page_referenced(), including file pages (page cache) where > CONFIG_SWAP plays no role. Maybe it was different 10 years ago. Yes, we require this bit for properly aging mmapped file pages. The underlying assumption in the referenced commit is incorrect. > > If yes, should pte_young() always return true in that case ? > > It should best work as intended. If not possible, true is maybe better, as > false will lead to inactive file list thrashing. An unconditional true will cause mmapped file pages to be permanently mlocked / unevictable. Either way will break some workloads. The only good answer is the truth :-)
Re: mm: Question about the use of 'accessed' flags and pte_young() helper
On 10/8/20 11:49 AM, Christophe Leroy wrote: In a 10 years old commit (https://github.com/linuxppc/linux/commit/d069cb4373fe0d451357c4d3769623a7564dfa9f), powerpc 8xx has made the handling of PTE accessed bit conditional to CONFIG_SWAP. Since then, this has been extended to some other powerpc variants. That commit means that when CONFIG_SWAP is not selected, the accessed bit is not set by SW TLB miss handlers, leading to pte_young() returning garbage, or should I say possibly returning false allthough a page has been accessed since its access flag was reset. Looking at various mm/ places, pte_young() is used independent of CONFIG_SWAP Is it still valid the not manage accessed flags when CONFIG_SWAP is not selected ? AFAIK it's wrong, reclaim needs it to detect accessed pages on inactive list, via page_referenced(), including file pages (page cache) where CONFIG_SWAP plays no role. Maybe it was different 10 years ago. If yes, should pte_young() always return true in that case ? It should best work as intended. If not possible, true is maybe better, as false will lead to inactive file list thrashing. While we are at it, I'm wondering whether powerpc should redefine arch_faults_on_old_pte() On some variants of powerpc, accessed flag is managed by HW. On others, it is managed by SW TLB miss handlers via page fault handling. Thanks Christophe
mm: Question about the use of 'accessed' flags and pte_young() helper
In a 10 years old commit (https://github.com/linuxppc/linux/commit/d069cb4373fe0d451357c4d3769623a7564dfa9f), powerpc 8xx has made the handling of PTE accessed bit conditional to CONFIG_SWAP. Since then, this has been extended to some other powerpc variants. That commit means that when CONFIG_SWAP is not selected, the accessed bit is not set by SW TLB miss handlers, leading to pte_young() returning garbage, or should I say possibly returning false allthough a page has been accessed since its access flag was reset. Looking at various mm/ places, pte_young() is used independent of CONFIG_SWAP Is it still valid the not manage accessed flags when CONFIG_SWAP is not selected ? If yes, should pte_young() always return true in that case ? While we are at it, I'm wondering whether powerpc should redefine arch_faults_on_old_pte() On some variants of powerpc, accessed flag is managed by HW. On others, it is managed by SW TLB miss handlers via page fault handling. Thanks Christophe
