Hi, Tiberiu, On Fri, Jul 30, 2021 at 04:08:25PM +0000, Tiberiu A Georgescu wrote: > This patch follows up on a previous RFC: > 20210714152426.216217-1-tiberiu.georgescu@xxxxxxxxxxx > > When a page allocated using the MAP_SHARED flag is swapped out, its pagemap > entry is cleared. In many cases, there is no difference between swapped-out > shared pages and newly allocated, non-dirty pages in the pagemap interface. > > Example pagemap-test code (Tested on Kernel Version 5.14-rc3): > #define NPAGES (256) > /* map 1MiB shared memory */ > size_t pagesize = getpagesize(); > char *p = mmap(NULL, pagesize * NPAGES, PROT_READ | PROT_WRITE, > MAP_ANONYMOUS | MAP_SHARED, -1, 0); > /* Dirty new pages. */ > for (i = 0; i < PAGES; i++) > p[i * pagesize] = i; > > Run the above program in a small cgroup, which causes swapping: > /* Initialise cgroup & run a program */ > $ echo 512K > foo/memory.limit_in_bytes > $ echo 60 > foo/memory.swappiness > $ cgexec -g memory:foo ./pagemap-test > > Check the pagemap report. Example of the current expected output: > $ dd if=/proc/$PID/pagemap ibs=8 skip=$(($VADDR / $PAGESIZE)) count=$COUNT | hexdump -C > 00000000 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 |................| > * > 00000710 e1 6b 06 00 00 00 00 a1 9e eb 06 00 00 00 00 a1 |.k..............| > 00000720 6b ee 06 00 00 00 00 a1 a5 a4 05 00 00 00 00 a1 |k...............| > 00000730 5c bf 06 00 00 00 00 a1 90 b6 06 00 00 00 00 a1 |\...............| > > The first pagemap entries are reported as zeroes, indicating the pages have > never been allocated while they have actually been swapped out. > > This patch addresses the behaviour and modifies pte_to_pagemap_entry() to > make use of the XArray associated with the virtual memory area struct > passed as an argument. The XArray contains the location of virtual pages in > the page cache, swap cache or on disk. If they are in either of the caches, > then the original implementation still works. If not, then the missing > information will be retrieved from the XArray. > > Performance > ============ > I measured the performance of the patch on a single socket Xeon E5-2620 > machine, with 128GiB of RAM and 128GiB of swap storage. These were the > steps taken: > > 1. Run example pagemap-test code on a cgroup > a. Set up cgroup with limit_in_bytes=4GiB and swappiness=60; > b. allocate 16GiB (about 4 million pages); > c. dirty 0,50 or 100% of pages; > d. do this for both private and shared memory. > 2. Run `dd if=<PAGEMAP> ibs=8 skip=$(($VADDR / $PAGESIZE)) count=4194304` > for each possible configuration above > a. 3 times for warm up; > b. 10 times to measure performance. > Use `time` or another performance measuring tool. > > Results (averaged over 10 iterations): > +--------+------------+------------+ > | dirty% | pre patch | post patch | > +--------+------------+------------+ > private|anon | 0% | 8.15s | 8.40s | > | 50% | 11.83s | 12.19s | > | 100% | 12.37s | 12.20s | > +--------+------------+------------+ > shared|anon | 0% | 8.17s | 8.18s | > | 50% | (*) 10.43s | 37.43s | > | 100% | (*) 10.20s | 38.59s | > +--------+------------+------------+ > > (*): reminder that pre-patch produces incorrect pagemap entries for swapped > out pages. > > From run to run the above results are stable (mostly <1% stderr). > > The amount of time it takes for a full read of the pagemap depends on the > granularity used by dd to read the pagemap file. Even though the access is > sequential, the script only reads 8 bytes at a time, running pagemap_read() > COUNT times (one time for each page in a 16GiB area). > > To reduce overhead, we can use batching for large amounts of sequential > access. We can make dd read multiple page entries at a time, > allowing the kernel to make optimisations and yield more throughput. > > Performance in real time (seconds) of > `dd if=<PAGEMAP> ibs=8*$BATCH skip=$(($VADDR / $PAGESIZE / $BATCH)) > count=$((4194304 / $BATCH))`: > +---------------------------------+ +---------------------------------+ > | Shared, Anon, 50% dirty | | Shared, Anon, 100% dirty | > +-------+------------+------------+ +-------+------------+------------+ > | Batch | Pre-patch | Post-patch | | Batch | Pre-patch | Post-patch | > +-------+------------+------------+ +-------+------------+------------+ > | 1 | (*) 10.43s | 37.43s | | 1 | (*) 10.20s | 38.59s | > | 2 | (*) 5.25s | 18.77s | | 2 | (*) 5.15s | 19.37s | > | 4 | (*) 2.63s | 9.42s | | 4 | (*) 2.63s | 9.74s | > | 8 | (*) 1.38s | 4.80s | | 8 | (*) 1.35s | 4.94s | > | 16 | (*) 0.73s | 2.46s | | 16 | (*) 0.72s | 2.54s | > | 32 | (*) 0.40s | 1.31s | | 32 | (*) 0.41s | 1.34s | > | 64 | (*) 0.25s | 0.72s | | 64 | (*) 0.24s | 0.74s | > | 128 | (*) 0.16s | 0.43s | | 128 | (*) 0.16s | 0.44s | > | 256 | (*) 0.12s | 0.28s | | 256 | (*) 0.12s | 0.29s | > | 512 | (*) 0.10s | 0.21s | | 512 | (*) 0.10s | 0.22s | > | 1024 | (*) 0.10s | 0.20s | | 1024 | (*) 0.10s | 0.21s | > +-------+------------+------------+ +-------+------------+------------+ > > To conclude, in order to make the most of the underlying mechanisms of > pagemap and xarray, one should be using batching to achieve better > performance. So what I'm still a bit worried is whether it will regress some existing users. Note that existing users can try to read pagemap in their own way; we can't expect all the userspaces to change their behavior due to a kernel change. Meanwhile, from the numbers, it seems to show a 4x speed down due to looking up the page cache no matter the size of ibs=. IOW I don't see a good way to avoid that overhead, so no way to have the userspace run as fast as before. Also note that it's not only affecting the PM_SWAP users; it potentially affects all the /proc/pagemap users as long as there're file-backed memory on the read region of pagemap, which is very sane to happen. That's why I think if we want to persist it, we should still consider starting from the pte marker idea. I do plan to move the pte marker idea forward unless that'll be NACKed upstream for some other reason, because that seems to be the only way for uffd-wp to support file based memories; no matter with a new swp type or with special swap pte. I am even thinking about whether I should propose that with PM_SWAP first because that seems to be a simpler scenario than uffd-wp (which will get the rest uffd-wp patches involved then), then we can have a shared infrastructure. But haven't thought deeper than that. Thanks, -- Peter Xu
