Re: systemd memory usage
On Fri, 2011-06-10 at 15:55 -0400, Adam Jackson wrote:
On 6/10/11 9:07 AM, Denys Vlasenko wrote:
> At the very least, I would like to see its memory consumption
> to go down substantially.
Let's try to turn this into something constructive. I'll start with
David Malcolm's rather nifty, if unpolished, 'heap' plugin for gdb:
https://fedorahosted.org/gdb-heap/wiki
Woot! Fame at last! :)
Various comments inline below...
Simply running the 'heap' command tells me that systemd is
using 21M of
malloc'd data, and that about 11M of it is entirely within allocations
that are all 2064 bytes apiece. Everything else, in comparison, is
pretty insignificant:
Domain Kind Detail Count Allocated size
------------- ----------- ------------ ------- --------------
uncategorized 2064 bytes 5,370 11,083,680
[...snip...]
"uncategorized" means that gdb-heap doesn't know how to deal with bytes
in question. If it's from a library, and you figure out what the above
is, we could add a categorizer for it.
C string data 12,964 630,672
...
TOTAL 127,119 21,934,576
That's a pretty unusual size, 2064 bytes. That works out to 2048
+ 16,
though, which are much more natural-sounding numbers. A quick
experiment with a demo program (allocate a 32-byte struct and then call
pause()) shows that the 16 is actually the overhead from malloc itself:
Domain Kind Detail Count Allocated size
------------- ---- -------- ----- --------------
uncategorized 48 bytes 1 48
TOTAL 1 48
Yes: IIRC, the "size" is actually the offset to the next chunk. glibc's
malloc allocates two size_t fields at the top of each chunk.
So now the problem is simply to find a 2048-sized allocation within
systemd, or one of its libraries. Anyone interested in a homework problem?
gdb-heap has a search/querying feature:
If you run :
(gdb) heap select size == 2048
it will try to show addresses and partial hexdumps of the relevant
buffers. (I _think_ so, but you might need 2064 instead).
The addresses it reports are those of where the allocation begins (as
seen by the program) i.e. it does the offset for you.
Given an address, there's also a
(gdb) hexdump ADDR
command, which you can use to scroll through memory.
If you stare at the hexdump you can sometimes figure out what the data
is.
If you think that an allocation is of some given type, you can also try
casting the data, e.g.:
(gdb) print *(some_struct*)ADDR
and see if looks meaningful.
Hope this is helpful.
Dave