Re: [PATCH v2 7/7] swap, put_swap_page: Share more between huge/normal code path
On 07/17/2018 07:56 PM, Huang, Ying wrote: > -.orc_unwind_ip1380 0 > -.orc_unwind 2070 0 > -Total26810 > +.orc_unwind_ip1480 0 > +.orc_unwind 2220 0 > +Total27172 > > The total difference is same: 27172 - 26810 = 362 = 24577 - 24215. > > The text section difference is small: 17927 - 17815 = 112. The > additional size change comes from unwinder information: (1480 + 2220) - > (1380 + 2070) = 250. If the frame pointer unwinder is chosen, this cost > nothing, but if the ORC unwinder is chosen, this is the real difference. > > For 112 text section difference, use 'objdump -t' to get symbol size and > compare, Cool, thanks for doing this! I think what you've done here is great for readability and the binary size increase is well worth the modest size increase.
Re: [PATCH v2 7/7] swap, put_swap_page: Share more between huge/normal code path
On 07/17/2018 07:56 PM, Huang, Ying wrote: > -.orc_unwind_ip1380 0 > -.orc_unwind 2070 0 > -Total26810 > +.orc_unwind_ip1480 0 > +.orc_unwind 2220 0 > +Total27172 > > The total difference is same: 27172 - 26810 = 362 = 24577 - 24215. > > The text section difference is small: 17927 - 17815 = 112. The > additional size change comes from unwinder information: (1480 + 2220) - > (1380 + 2070) = 250. If the frame pointer unwinder is chosen, this cost > nothing, but if the ORC unwinder is chosen, this is the real difference. > > For 112 text section difference, use 'objdump -t' to get symbol size and > compare, Cool, thanks for doing this! I think what you've done here is great for readability and the binary size increase is well worth the modest size increase.
Re: [PATCH v2 7/7] swap, put_swap_page: Share more between huge/normal code path
Dave Hansen writes: > On 07/16/2018 05:55 PM, Huang, Ying wrote: >> text data bss dec hex filename >> base: 24215 2028 340 2658367d7 mm/swapfile.o >> unified: 245772028 340 269456941 mm/swapfile.o > > That's a bit more than I'd expect looking at the rest of the diff. Make > me wonder if we missed an #ifdef somewhere or the compiler is getting > otherwise confused. > > Might be worth a 10-minute look at the disassembly. Dig one step deeper via 'size -A mm/swapfile.o' and diff between base and unified, --- b.s 2018-07-18 09:42:07.872501680 +0800 +++ h.s 2018-07-18 09:50:37.984499168 +0800 @@ -1,6 +1,6 @@ mm/swapfile.o : section size addr -.text17815 0 +.text17927 0 .data 1288 0 .bss 340 0 ___ksymtab_gpl+nr_swap_pages 8 0 @@ -26,8 +26,8 @@ .data.once 1 0 .comment35 0 .note.GNU-stack 0 0 -.orc_unwind_ip1380 0 -.orc_unwind 2070 0 -Total26810 +.orc_unwind_ip1480 0 +.orc_unwind 2220 0 +Total27172 The total difference is same: 27172 - 26810 = 362 = 24577 - 24215. The text section difference is small: 17927 - 17815 = 112. The additional size change comes from unwinder information: (1480 + 2220) - (1380 + 2070) = 250. If the frame pointer unwinder is chosen, this cost nothing, but if the ORC unwinder is chosen, this is the real difference. For 112 text section difference, use 'objdump -t' to get symbol size and compare, --- b.od2018-07-18 10:45:05.768483075 +0800 +++ h.od2018-07-18 10:44:39.556483204 +0800 @@ -30,9 +30,9 @@ 00a3 cluster_list_add_tail 001e __kunmap_atomic.isra.34 018c swap_count_continued -00ac __swap_entry_free 000f put_swap_device.isra.35 00b4 inc_cluster_info_page +006f __swap_entry_free_locked 004a _enable_swap_info 0046 wait_on_page_writeback 002e inode_to_bdi @@ -53,8 +53,8 @@ 0012 __x64_sys_swapon 0011 __ia32_sys_swapon 007a get_swap_device -0032 swap_free -0035 put_swap_page +006e swap_free +0078 put_swap_page 0267 swapcache_free_entries 0058 page_swapcount 003a __swap_count @@ -64,7 +64,7 @@ 011a try_to_free_swap 01fb get_swap_pages 0098 get_swap_page_of_type -01b8 free_swap_and_cache +01e6 free_swap_and_cache 0543 try_to_unuse 000e __x64_sys_swapoff 000d __ia32_sys_swapoff The size of put_swap_page() change is small: 0x78 - 0x35 = 67. But __swap_entry_free() is inlined by compiler, which cause some code dilating. Best Regards, Huang, Ying
Re: [PATCH v2 7/7] swap, put_swap_page: Share more between huge/normal code path
Dave Hansen writes: > On 07/16/2018 05:55 PM, Huang, Ying wrote: >> text data bss dec hex filename >> base: 24215 2028 340 2658367d7 mm/swapfile.o >> unified: 245772028 340 269456941 mm/swapfile.o > > That's a bit more than I'd expect looking at the rest of the diff. Make > me wonder if we missed an #ifdef somewhere or the compiler is getting > otherwise confused. > > Might be worth a 10-minute look at the disassembly. Dig one step deeper via 'size -A mm/swapfile.o' and diff between base and unified, --- b.s 2018-07-18 09:42:07.872501680 +0800 +++ h.s 2018-07-18 09:50:37.984499168 +0800 @@ -1,6 +1,6 @@ mm/swapfile.o : section size addr -.text17815 0 +.text17927 0 .data 1288 0 .bss 340 0 ___ksymtab_gpl+nr_swap_pages 8 0 @@ -26,8 +26,8 @@ .data.once 1 0 .comment35 0 .note.GNU-stack 0 0 -.orc_unwind_ip1380 0 -.orc_unwind 2070 0 -Total26810 +.orc_unwind_ip1480 0 +.orc_unwind 2220 0 +Total27172 The total difference is same: 27172 - 26810 = 362 = 24577 - 24215. The text section difference is small: 17927 - 17815 = 112. The additional size change comes from unwinder information: (1480 + 2220) - (1380 + 2070) = 250. If the frame pointer unwinder is chosen, this cost nothing, but if the ORC unwinder is chosen, this is the real difference. For 112 text section difference, use 'objdump -t' to get symbol size and compare, --- b.od2018-07-18 10:45:05.768483075 +0800 +++ h.od2018-07-18 10:44:39.556483204 +0800 @@ -30,9 +30,9 @@ 00a3 cluster_list_add_tail 001e __kunmap_atomic.isra.34 018c swap_count_continued -00ac __swap_entry_free 000f put_swap_device.isra.35 00b4 inc_cluster_info_page +006f __swap_entry_free_locked 004a _enable_swap_info 0046 wait_on_page_writeback 002e inode_to_bdi @@ -53,8 +53,8 @@ 0012 __x64_sys_swapon 0011 __ia32_sys_swapon 007a get_swap_device -0032 swap_free -0035 put_swap_page +006e swap_free +0078 put_swap_page 0267 swapcache_free_entries 0058 page_swapcount 003a __swap_count @@ -64,7 +64,7 @@ 011a try_to_free_swap 01fb get_swap_pages 0098 get_swap_page_of_type -01b8 free_swap_and_cache +01e6 free_swap_and_cache 0543 try_to_unuse 000e __x64_sys_swapoff 000d __ia32_sys_swapoff The size of put_swap_page() change is small: 0x78 - 0x35 = 67. But __swap_entry_free() is inlined by compiler, which cause some code dilating. Best Regards, Huang, Ying
Re: [PATCH v2 7/7] swap, put_swap_page: Share more between huge/normal code path
On 07/16/2018 05:55 PM, Huang, Ying wrote: > text data bss dec hex filename > base:24215 2028 340 2658367d7 mm/swapfile.o > unified: 24577 2028 340 269456941 mm/swapfile.o That's a bit more than I'd expect looking at the rest of the diff. Make me wonder if we missed an #ifdef somewhere or the compiler is getting otherwise confused. Might be worth a 10-minute look at the disassembly.
Re: [PATCH v2 7/7] swap, put_swap_page: Share more between huge/normal code path
On 07/16/2018 05:55 PM, Huang, Ying wrote: > text data bss dec hex filename > base:24215 2028 340 2658367d7 mm/swapfile.o > unified: 24577 2028 340 269456941 mm/swapfile.o That's a bit more than I'd expect looking at the rest of the diff. Make me wonder if we missed an #ifdef somewhere or the compiler is getting otherwise confused. Might be worth a 10-minute look at the disassembly.
[PATCH v2 7/7] swap, put_swap_page: Share more between huge/normal code path
In this patch, locking related code is shared between huge/normal code path in put_swap_page() to reduce code duplication. And `free_entries == 0` case is merged into more general `free_entries != SWAPFILE_CLUSTER` case, because the new locking method makes it easy. The added lines is same as the removed lines. But the code size is increased when CONFIG_TRANSPARENT_HUGEPAGE=n. text data bss dec hex filename base: 24215 2028 340 2658367d7 mm/swapfile.o unified: 24577 2028 340 269456941 mm/swapfile.o Signed-off-by: "Huang, Ying" Cc: Dave Hansen Cc: Michal Hocko Cc: Johannes Weiner Cc: Shaohua Li Cc: Hugh Dickins Cc: Minchan Kim Cc: Rik van Riel Cc: Daniel Jordan Cc: Dan Williams --- mm/swapfile.c | 20 ++-- 1 file changed, 10 insertions(+), 10 deletions(-) diff --git a/mm/swapfile.c b/mm/swapfile.c index fec28f6c05b0..cd75f449896b 100644 --- a/mm/swapfile.c +++ b/mm/swapfile.c @@ -1280,8 +1280,8 @@ void put_swap_page(struct page *page, swp_entry_t entry) if (!si) return; + ci = lock_cluster_or_swap_info(si, offset); if (nr == SWAPFILE_CLUSTER) { - ci = lock_cluster(si, offset); VM_BUG_ON(!cluster_is_huge(ci)); map = si->swap_map + offset; for (i = 0; i < SWAPFILE_CLUSTER; i++) { @@ -1290,13 +1290,9 @@ void put_swap_page(struct page *page, swp_entry_t entry) if (val == SWAP_HAS_CACHE) free_entries++; } - if (!free_entries) { - for (i = 0; i < SWAPFILE_CLUSTER; i++) - map[i] &= ~SWAP_HAS_CACHE; - } cluster_clear_huge(ci); - unlock_cluster(ci); if (free_entries == SWAPFILE_CLUSTER) { + unlock_cluster_or_swap_info(si, ci); spin_lock(>lock); ci = lock_cluster(si, offset); memset(map, 0, SWAPFILE_CLUSTER); @@ -1307,12 +1303,16 @@ void put_swap_page(struct page *page, swp_entry_t entry) return; } } - if (nr == 1 || free_entries) { - for (i = 0; i < nr; i++, entry.val++) { - if (!__swap_entry_free(si, entry, SWAP_HAS_CACHE)) - free_swap_slot(entry); + for (i = 0; i < nr; i++, entry.val++) { + if (!__swap_entry_free_locked(si, offset + i, SWAP_HAS_CACHE)) { + unlock_cluster_or_swap_info(si, ci); + free_swap_slot(entry); + if (i == nr - 1) + return; + lock_cluster_or_swap_info(si, offset); } } + unlock_cluster_or_swap_info(si, ci); } #ifdef CONFIG_THP_SWAP -- 2.16.4
[PATCH v2 7/7] swap, put_swap_page: Share more between huge/normal code path
In this patch, locking related code is shared between huge/normal code path in put_swap_page() to reduce code duplication. And `free_entries == 0` case is merged into more general `free_entries != SWAPFILE_CLUSTER` case, because the new locking method makes it easy. The added lines is same as the removed lines. But the code size is increased when CONFIG_TRANSPARENT_HUGEPAGE=n. text data bss dec hex filename base: 24215 2028 340 2658367d7 mm/swapfile.o unified: 24577 2028 340 269456941 mm/swapfile.o Signed-off-by: "Huang, Ying" Cc: Dave Hansen Cc: Michal Hocko Cc: Johannes Weiner Cc: Shaohua Li Cc: Hugh Dickins Cc: Minchan Kim Cc: Rik van Riel Cc: Daniel Jordan Cc: Dan Williams --- mm/swapfile.c | 20 ++-- 1 file changed, 10 insertions(+), 10 deletions(-) diff --git a/mm/swapfile.c b/mm/swapfile.c index fec28f6c05b0..cd75f449896b 100644 --- a/mm/swapfile.c +++ b/mm/swapfile.c @@ -1280,8 +1280,8 @@ void put_swap_page(struct page *page, swp_entry_t entry) if (!si) return; + ci = lock_cluster_or_swap_info(si, offset); if (nr == SWAPFILE_CLUSTER) { - ci = lock_cluster(si, offset); VM_BUG_ON(!cluster_is_huge(ci)); map = si->swap_map + offset; for (i = 0; i < SWAPFILE_CLUSTER; i++) { @@ -1290,13 +1290,9 @@ void put_swap_page(struct page *page, swp_entry_t entry) if (val == SWAP_HAS_CACHE) free_entries++; } - if (!free_entries) { - for (i = 0; i < SWAPFILE_CLUSTER; i++) - map[i] &= ~SWAP_HAS_CACHE; - } cluster_clear_huge(ci); - unlock_cluster(ci); if (free_entries == SWAPFILE_CLUSTER) { + unlock_cluster_or_swap_info(si, ci); spin_lock(>lock); ci = lock_cluster(si, offset); memset(map, 0, SWAPFILE_CLUSTER); @@ -1307,12 +1303,16 @@ void put_swap_page(struct page *page, swp_entry_t entry) return; } } - if (nr == 1 || free_entries) { - for (i = 0; i < nr; i++, entry.val++) { - if (!__swap_entry_free(si, entry, SWAP_HAS_CACHE)) - free_swap_slot(entry); + for (i = 0; i < nr; i++, entry.val++) { + if (!__swap_entry_free_locked(si, offset + i, SWAP_HAS_CACHE)) { + unlock_cluster_or_swap_info(si, ci); + free_swap_slot(entry); + if (i == nr - 1) + return; + lock_cluster_or_swap_info(si, offset); } } + unlock_cluster_or_swap_info(si, ci); } #ifdef CONFIG_THP_SWAP -- 2.16.4