Re: [9fans] kprof

Re: [9fans] kprof

[forsyth]

>>charged 4%.  Chances are against you that this way of accounting will
>>give you the correct answer.

there was a paper explaining how the original version of gprof
gave misleading answers.  i think it was in Software: P & E
but i'm not sure.  the usual net search engines will probably know.

Re: [9fans] kprof

[andrey mirtchovski]

i managed to do something like what i wanted by modifying devkprof.c
like:

allocate 10mins * 60 secs * 1000ms for storing the pc, then at each
call store the pc value passed to _kproftimer() in the array.  (so, it
only works for 10 minutes, which is plenty enough for what we're
looking at)

then the kprof executable prints out the corresponding function name
assiciated with this pc...

here's how the final output looks like (note, i piped it uniq to avoid
crud):

.....
01ced	i8259isr
46a8e	v6tov4
00301	halt
5fac0	runproc
00301	halt
01ced	i8259isr
00301	halt
01ced	i8259isr
00301	halt
5c35b	seg
00301	halt
01ced	i8259isr
00301	halt
5f8d0	runproc
00301	halt
69f21	ilock
00301	halt
01ced	i8259isr
6edd2	poolfree
00301	halt
0a75d	transmit
.....

unfortunately it's not as good as prof -d, but paired with the normal
kprof output some information could be gleaned from it.

if anyone needs the diffs, mail me privately

Re: [9fans] kprof

[rob pike, esq.]

_kproftimer is called from the clock routine.

You're proposing the gprof approach, where you notice 12% of the time
is spent in a routine, which is called in 3 places, so each place gets
charged 4%.  Chances are against you that this way of accounting will
give you the correct answer.

-rob

Re: [9fans] kprof

[andrey mirtchovski]

how about just keeping track of the pc at each call to _kprofinit()
and then translating that into functions the pc is at off-line?  it's
not a general solution, will have gaps and will not be fine-grained
but at least it'll give us an idea (hopefully) where a memmove() is
called from, if the kernel spends 5% of its time there...

how often does _kproftimer() get called?  i'll have to dig deeper...

andrey

Re: [9fans] kprof

[rob pike, esq.]

Prof -d works by rewriting the text of the program to include
statistics-gathering code at each function call/return.  Without
such code at function points in the kernel, kprof -d cannot work.
However, you'll have an awful time managing that statistics-
gathering code in a running kernel with interrupts and such
going on.  Yikes.

-rob

[9fans] kprof

[andrey mirtchovski]

i'm trying to modify kprof to provide a syscall tree the same way
'prof -d' does, but i can't figure out how to handle the data
accumulated in devkprof.c's _kproftimer().

i understand the current way it works -- storing the total time the
kernel has spent in a certain function, but what i can't figure out is
how to store the actual PC values for the entire timespan between a
start and a stop command.

one way is to create a (presumably large!)  array and incrementally
store the current PC value at each call to _kproftimer(), but is that
a decent solution?  am i right in assuming (based on observations :)
that no syscall can be executed in _kproftimer()?

andrey