Re: [RFC 7/7] mm: madvise support MADV_ANONYMOUS_FILTER and MADV_FILE_FILTER

[Minchan Kim]

On Mon, May 27, 2019 at 02:44:11PM +0200, Michal Hocko wrote:
> On Mon 27-05-19 16:58:11, Minchan Kim wrote:
> > On Tue, May 21, 2019 at 08:26:28AM +0200, Michal Hocko wrote:
> > > On Tue 21-05-19 11:55:33, Minchan Kim wrote:
> > > > On Mon, May 20, 2019 at 11:28:01AM +0200, Michal Hocko wrote:
> > > > > [cc linux-api]
> > > > > 
> > > > > On Mon 20-05-19 12:52:54, Minchan Kim wrote:
> > > > > > System could have much faster swap device like zRAM. In that case, swapping
> > > > > > is extremely cheaper than file-IO on the low-end storage.
> > > > > > In this configuration, userspace could handle different strategy for each
> > > > > > kinds of vma. IOW, they want to reclaim anonymous pages by MADV_COLD
> > > > > > while it keeps file-backed pages in inactive LRU by MADV_COOL because
> > > > > > file IO is more expensive in this case so want to keep them in memory
> > > > > > until memory pressure happens.
> > > > > > 
> > > > > > To support such strategy easier, this patch introduces
> > > > > > MADV_ANONYMOUS_FILTER and MADV_FILE_FILTER options in madvise(2) like
> > > > > > that /proc/<pid>/clear_refs already has supported same filters.
> > > > > > They are filters could be Ored with other existing hints using top two bits
> > > > > > of (int behavior).
> > > > > 
> > > > > madvise operates on top of ranges and it is quite trivial to do the
> > > > > filtering from the userspace so why do we need any additional filtering?
> > > > > 
> > > > > > Once either of them is set, the hint could affect only the interested vma
> > > > > > either anonymous or file-backed.
> > > > > > 
> > > > > > With that, user could call a process_madvise syscall simply with a entire
> > > > > > range(0x0 - 0xFFFFFFFFFFFFFFFF) but either of MADV_ANONYMOUS_FILTER and
> > > > > > MADV_FILE_FILTER so there is no need to call the syscall range by range.
> > > > > 
> > > > > OK, so here is the reason you want that. The immediate question is why
> > > > > cannot the monitor do the filtering from the userspace. Slightly more
> > > > > work, all right, but less of an API to expose and that itself is a
> > > > > strong argument against.
> > > > 
> > > > What I should do if we don't have such filter option is to enumerate all of
> > > > vma via /proc/<pid>/maps and then parse every ranges and inode from string,
> > > > which would be painful for 2000+ vmas.
> > > 
> > > Painful is not an argument to add a new user API. If the existing API
> > > suits the purpose then it should be used. If it is not usable, we can
> > > think of a different way.
> > 
> > I measured 1568 vma parsing overhead of /proc/<pid>/maps in ARM64 modern
> > mobile CPU. It takes 60ms and 185ms on big cores depending on cpu governor.
> > It's never trivial.
> 
> This is not the only option. Have you tried to simply use
> /proc/<pid>/map_files interface? This will provide you with all the file
> backed mappings.

I compared maps vs. map_files with 3036 file-backed vma.
Test scenario is to dump all of vmas of the process and parse address
ranges.
For map_files, it's easy to parse each address range because directory name
itself is range. However, in case of maps, I need to parse each range
line by line so need to scan all of lines.

(maps cover additional non-file-backed vmas so nr_vma is a little bigger)

performance mode:
map_files: nr_vma 3036 usec 13387
maps     : nr_vma 3078 usec 12923

powersave mode:

map_files: nr_vma 3036 usec 52614
maps     : nr_vma 3078 usec 41089

map_files is slower than maps if we dump all of vmas. I guess directory
operation needs much more jobs(e.g., dentry lookup, instantiation)
compared to maps.