1479577 - Firefox Quantum worker processes seem unswappable

Status: RESOLVED INVALID

(Core :: Memory Allocator, defect)

Description

[nemesis]

User Agent: Mozilla/5.0 (X11; Linux x86_64; rv:61.0) Gecko/20100101 Firefox/61.0
Build ID: 20180704003137

Steps to reproduce:

Observed in atop (SWAPSZ is 0 for all content procs) and via pmap -XX that 'Web Content' worker processes have a large amount of locked memory.

$ ps ax | grep firefox.*contentproc | grep -v grep | awk '{print $1}' | while read PID; do ps ax | grep "^\s*${PID}" | grep -v grep; sudo pmap -x /proc/${PID} | grep 'anon' | sort -g -k3 | tail -10 | cut -f1 -d\ | sed 's/^0*//' | xargs -I{} sh -c "sudo pmap -XX /proc/${PID} | sed 's/\s\+/    /g' | grep {} | cut -f2,25 -d\  "; done

 4302 ?        Sl   397:21 /usr/lib/firefox/firefox -contentproc -childID 5 -isForBrowser -prefsLen 88354 -schedulerPrefs 0001,2 -greomni /usr/lib/firefox/omni.ja -appomni /usr/lib/firefox/browser/omni.ja -appdir /usr/lib/firefox/browser 3207 true tab
7f88d4a00000    8412
7f888f700000    8568
7f88e7800000    9056
7f88ae600000    9228
7f88a8300000    9488
7f88e8700000    9824
7f88dbd00000    10316
7f88d7700000    10604
7f88d5500000    16856
7f88dd100000    25196
11696 ?        Sl    14:04 /usr/lib/firefox/firefox -contentproc -childID 50 -isForBrowser -prefsLen 88348 -schedulerPrefs 0001,2 -greomni /usr/lib/firefox/omni.ja -appomni /usr/lib/firefox/browser/omni.ja -appdir /usr/lib/firefox/browser 3207 true tab
7ffbaf000000    6116
7ffb94a00000    6144
7ffbab000000    7116
7ffba4900000    10132
7ffb97aff000    23464
7ffba5500000    33620
7ffb90800000    36104
7ffba7900000    38756
7ffbac000000    40840
7ffb997ff000    115016
28583 ?        Sl     0:01 /usr/lib/firefox/firefox -contentproc -childID 51 -isForBrowser -prefsLen 88348 -schedulerPrefs 0001,2 -greomni /usr/lib/firefox/omni.ja -appomni /usr/lib/firefox/browser/omni.ja -appdir /usr/lib/firefox/browser 3207 true tab
7fe20c52c000    24
ebdb164d000     128
7fe2004c9000    180
7fe1f1600000    280
7fe20a700000    908
7fe1f5700000    992
7fe1efb00000    2608
7fe1f2400000    3864
7fe1f0900000    4040
7fe1edc00000    4104



Actual results:

It seems pages belonging to the heap of worker processes cannot be swapped out, creating global memory pressure and swapstorms on low memory (8GB) systems.


Expected results:

Is this done for an articulable reason?  Is my analysis wrong?

Comment 1

[Mike Hommey [:glandium]]

I don't know what Locked means in /proc/$pid/smaps when VmFlags doesn't contain `lo`, and I see no such regions in my running Firefox.
I'll just note that a simple program that does `mmap()` followed by `memset()` ends up in the same state: Locked with a large value, but no `lo` in VmFlags.
So it would seem like this is not actually a problem.

Comment 2

[nemesis]

Thanks for investigating.  Could you speculate as to why the worker processes refuse to swap even under pressure?  This is very frustrating on a system with 8GB or less.