|Summary:||[patch] swap_pager.c to release multiple swap blocks at a time|
|Component:||kern||Assignee:||freebsd-bugs mailing list <bugs>|
|Severity:||Affects Only Me||CC:||alc, dougm, emaste, kib, markj|
Description ota 2017-11-23 21:41:46 UTC
Created attachment 188225 [details] release multiple blocks at a time This allows to release multiple swap blocks via swp_pager_freeswapspace(). For each call, swp_pager_freeswapspace() locks. Releasing multiple blocks helps to reduce number of locks. If a swblk contains continuous blocks, it reduces up to SWAP_META_PAGES-1 locks. This patch covers 2 case: 1. when a process exits and all of its swap blocks are released and 2. when a process is paged in from swap space.
Comment 1 ota 2017-12-02 03:22:42 UTC
Created attachment 188460 [details] swapoff to release multiple blocks at a time https://svnweb.freebsd.org/base?view=revision&revision=326329 change works better for the first patchset and already committed to the head. This one is to release multiple blocks from "swapoff" command. The patch allows to take SWB_NPAGES blocks at once from backing store. On i386, it is 32 blocks maximum compare to 1 block at a time.
Comment 2 Konstantin Belousov 2017-12-02 13:43:55 UTC
The patch is not applicable to HEAD, please update it. For the swap_pager_swapoff() chunk, it seems that the new variable offset can be used before initialized. That said, why optimize swapoff at all ?
Comment 3 ota 2017-12-03 04:59:39 UTC
(In reply to Konstantin Belousov from comment #2) Indeed, 1st patch chunk failed and I overlooked it. Also, I missed offset initialization. To answer your question, I started with 2 reasons. #1 - I started looking at optimizing page-in cases and process exit cases where they had released 1 block at a time although swp_pager_freeswapspace() takes npages to free. This was the 3rd one. #2 - swp_pager_force_pagein() comment says "it doesn't work" and I wondered why. The reason I found is we need to call vm_page_grab for the number of pages we want to page in. Summing up, after I figured vm_page part, the code started working okay and based on my observation of "top", multi-block page-in was considerably faster - reading 1 block 32 times compare to 32 blocks a time. It looksvm_object_t is locked during swapoff operation (while paging-in), too. During research, I found swap_pager_getpages() calls swap_pager_haspage() anyway and swap_pager_haspage() checks how many continuous blocks are behind. So, I started researching from swap_pager_haspage() side instead of swblk side. I had few questions like when to lock and how and when to increment vm_object_pip. I posted here for some feed backs. While checking patch failure against HEAD, I realized swp_pager_force_pagein() in HEAD had a change for (m->valid == VM_PAGE_BITS_ALL) case. This change cleared some of my questions. I will update my changes and post a new one for 11-STABLE and HEAD.
Comment 4 ota 2017-12-07 03:47:43 UTC
Created attachment 188607 [details] swapoff to release multiple blocks at a time for 11-stable Initialize offset. Split vm operations into functions. It looks I need to call vm_page_grab and cannot use rbehind to call swap_pager_getpages(). vm_page_alloc() is called by swap_pager_getpages() but I got a crash.
Comment 5 ota 2017-12-07 03:48:36 UTC
Created attachment 188608 [details] swapoff to release multiple blocks at a time for 12-head
Comment 6 Konstantin Belousov 2017-12-09 11:55:22 UTC
(In reply to ota from comment #5) Please: 1. Generate the patch for HEAD using svn diff -x -U 999999 2. Publish the patch on https://reviews.freebsd.org, assigning me as the reviewer.
Comment 7 ota 2017-12-14 06:27:53 UTC
(In reply to Konstantin Belousov from comment #6) I created https://reviews.freebsd.org/D13484 for this case. I have another change-set for a improvement and also created https://reviews.freebsd.org/D13485
Comment 8 Ed Maste 2019-05-21 19:33:39 UTC
Triaged during 2019 Waterloo Hackathon, Doug would you be interested in shepherding this change in?