Re: Terminally slow (75 seconds) on some steps

[Daniel Jacobowitz <drow at mvista dot com>]

On Sat, Aug 16, 2003 at 12:36:52PM -0400, Andrew Cagney wrote:
> 
> >See SYMBOL_SET_NAMES - I fixed that fairly recently.  Also,
> >add_psymbol_to_list is how most partial symbols are added; that uses
> >the same name cache.  If you're talking about the call to bcache in
> >add_psymbol_with_dem_name_to_list, it's only used by hpread.c.  Fixing
> >this did in fact improve the hit rate.
> 
> Ah.  So the only bcache call is to add the partial symbol to the psymbol 
> cache.  That narrows things down a bit.  Wonder if setting 
> CHAIN_LENGTH_THRESHOLD lower would help.
> 
> >>>c011f6b8 663       2.3393     vmlinux                  do_anonymous_page
> >>>00000000 622       2.1946     XFree86                  (no symbols)
> >>>08134b74 574       2.0253     gdb                      read_partial_die
> >
> >>
> >>Ok, so its spending a lot of time reading partial symbols.
> >
> >
> >I've got some plans in this direction... just haven't had time to do
> >anything about it yet :(
> 
> What about the kernel?  3.5 seconds copying data to the user space is 
> suprising.  Can linux, given page sized read requests do it using page 
> swapping?
> 
> The other question is where in GDB (?) the requests are comming from.  I 
> suspect target read, but it could be symbol file related.

First of all, we should be using mmap to access symbol files, at least
when there are no relocations.  We probably should be using it even
when there _are_ relocations; however, that will require
re-implementing some simple ELF relocation functions in GDB.  I've long
thought this is necessary.

Second of all, the copy_to_user is likely to come more from ptrace than
from reading.  But that's just a guess really.

-- 
Daniel Jacobowitz
MontaVista Software                         Debian GNU/Linux Developer