prefcnt is a new tool to do reference counting on all the TAGs, starting
from the list of DW_TAG_subroutine tags and going down thru the return type,
parameter list types, variables and inline expansions in the functions, to
help finding unused stuff, its not so effective because of bugs in gcc
DWARF emitting code for concrete inline instances, i.e. the inline expansions
are not all being emitted, see:
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=29792
Signed-off-by: Arnaldo Carvalho de Melo <acme@mandriva.com>
To represent DW_TAG_variable, for now all the variables in all the lexical
blocks, in addition to the top level function variables are in this list, next
step is to add support for DW_TAG_lexical_block, with support for nesting, and
to associate variables to the right place, be it the function itself (first,
implicit lexical block) or to the lexical blocks they belong too, this will be
useful for calculating stack usage.
So, with what we have now pfunct can do this:
[acme@newtoy guinea_pig-2.6]$ pfunct --variables net/ipv4/built-in.o tcp_v4_remember_stamp
/* net/ipv4/tcp_ipv4.c:1197 */
int tcp_v4_remember_stamp(struct sock * sk);
{
/* variables in tcp_v4_remember_stamp: */
struct inet_sock * inet;
struct tcp_sock * tp;
struct rtable * rt;
struct inet_peer * peer;
int release_it;
}
[acme@newtoy guinea_pig-2.6]$
That is already useful when you don't have the sources, huh? :-)
Signed-off-by: Arnaldo Carvalho de Melo <acme@mandriva.com>
Now we're able to process a kernel built with make allyesconfig on a machine
with 1GB, of course there is still more things to optimize, but I'm lazy and
for now this gives the numbers I wanted to get.
Signed-off-by: Arnaldo Carvalho de Melo <acme@mandriva.com>
Its just totsz / nrexp, i.e. total size of expansions divided by the number of
expansions found, but helps in analising the data.
Signed-off-by: Arnaldo Carvalho de Melo <acme@mandriva.com>
Not safe to directly point to the string parameter, probably its a inline
string, not an indirect one, one more reason to create a string table...
Signed-off-by: Arnaldo Carvalho de Melo <acme@mandriva.com>
[acme@newtoy guinea_pig-2.6]$ pahole -t ../../acme/OUTPUT/qemu/net-2.6/vmlinux | sort -k2 -nr | head -5
list_head 468
__wait_queue_head 466
timespec 466
rw_semaphore 466
plist_head 466
[acme@newtoy guinea_pig-2.6]$
Which leads to another, more non-trivia question, what if a struct
definition is included but there are no references to this function?
Signed-off-by: Arnaldo Carvalho de Melo <acme@mandriva.com>
Now it shows the number that each of the inline functions were expanded in an
object file:
Top 10 inline functions expanded more than once in kernel/sched.o, by total
size of inline expansions:
[acme@newtoy guinea_pig-2.6]$ pfunct --cu_inline_expansions_stats kernel/sched.o | sort -k3 -nr | grep -v ': 1 ' | head -11
kernel/sched.c: 318 10217
get_current: 38 325
finish_task_switch: 2 238
normal_prio: 2 167
__cpus_and: 14 164
find_process_by_pid: 6 152
current_thread_info: 21 149
sched_find_first_bit: 2 148
update_cpu_clock: 2 140
task_rq_unlock: 14 137
variable_test_bit: 14 121
[acme@newtoy guinea_pig-2.6]$
Now we have these options:
[acme@newtoy guinea_pig-2.6]$ pfunct --help
usage: pfunct [options] <file_name> {<function_name>}
where:
-c, --class=<class> functions that have <class> pointer parameters
-g, --goto_labels show number of goto labels
-i, --show_inline_expansions show inline expansions
-C, --cu_inline_expansions_stats show CU inline expansions stats
-s, --sizes show size of functions
-N, --function_name_len show size of functions
-p, --nr_parameters show number or parameters
-S, --variables show number of variables
-V, --verbose be verbose
[acme@newtoy guinea_pig-2.6]$
Signed-off-by: Arnaldo Carvalho de Melo <acme@mandriva.com>
Top five object files (CU, Compilation Unit) per number of inline expansions,
vmlinux being dissected is one built for QEMU, most things as modules, that
are not being taken into account as we're only looking at vmlinux:
[acme@newtoy guinea_pig-2.6]$ pfunct -C ../../acme/OUTPUT/qemu/net-2.6/vmlinux | sort -k2 -nr | head -5 | cut -c40-
net/ipv4/tcp_input.c: 274 20655
fs/buffer.c: 272 4597
kernel/sched.c: 214 3549
kernel/signal.c: 196 2730
fs/ext3/inode.c: 191 7961
[acme@newtoy guinea_pig-2.6]$
Top five object files (CU, Compilation Unit) per total size of inline expansions:
[acme@newtoy guinea_pig-2.6]$ pfunct -C ../../acme/OUTPUT/qemu/net-2.6/vmlinux | sort -k3 -nr | head -5 | cut -c40-
net/ipv4/tcp_input.c: 274 20655
net/xfrm/xfrm_policy.c: 173 11511
kernel/module.c: 95 10826
drivers/char/vt.c: 91 10050
net/xfrm/xfrm_user.c: 150 9682
[acme@newtoy guinea_pig-2.6]$
Signed-off-by: Arnaldo Carvalho de Melo <acme@mandriva.com>
For now we just discard it so that all the variables in a function can be
accounted, but the right thing, as said in the comment added in this cset is to
have a list of lexical blocks, each with a list of its variables so that we can
find the biggest stack usage in functions.
Signed-off-by: Arnaldo Carvalho de Melo <acme@mandriva.com>
Reducing the memory footprint, but more has to be done, such as to take
advantage of the strings table when handling indirect strings.
Signed-off-by: Arnaldo Carvalho de Melo <acme@mandriva.com>
With some of the options in pfunct, such as:
[acme@newtoy guinea_pig-2.6]$ pahole --sizes kernel/sched.o | sort -k2 -nr | head -5
pglist_data: 3456
task_struct: 2704
rq: 2480
mmu_gather: 2040
zone: 1664
[acme@newtoy guinea_pig-2.6]$
Signed-off-by: Arnaldo Carvalho de Melo <acme@mandriva.com>
So that we can find all the cus for some specific class
(cus__find_class_by_name), or traverse all the CUs (cus__for_each_cu),
etc.
Now we don't look at just the first CU in multi-CU files (vmlinux, etc).
Signed-off-by: Arnaldo Carvalho de Melo <acme@mandriva.com>
types for each CU, for now when working on multi-CU files (vmlinux, any binary
with more than one object file linked) we look only at the first CU when
looking for a specific class or function name, this will be fixed in the
upcoming csets, but doesn't affect the case when we don't specify a class or
function name, where all the CU's are traversed.
Signed-off-by: Arnaldo Carvalho de Melo <acme@mandriva.com>
except for --classes, that asks that all non-inlined functions that receive as
one of its parameter a pointer to the specified class be printed, just the name
if --verbose is not used, or its complete prototype and where it is in the
source code:
[acme@newtoy net-2.6]$ pfunct --verbose --class=inode fs/ext3/built-in.o
/* /pub/scm/linux/kernel/git/acme/net-2.6/fs/ext3/balloc.c 606 */
void ext3_free_blocks(handle_t * handle, struct inode * inode, ext3_fsblk_t block, long unsigned int count);
/* /pub/scm/linux/kernel/git/acme/net-2.6/fs/ext3/balloc.c 1404 */
ext3_fsblk_t ext3_new_blocks(handle_t * handle, struct inode * inode, ext3_fsblk_t goal, long unsigned int * count, int * errp);
/* /pub/scm/linux/kernel/git/acme/net-2.6/fs/ext3/balloc.c 1668 */
ext3_fsblk_t ext3_new_block(handle_t * handle, struct inode * inode, ext3_fsblk_t goal, int * errp);
/* /pub/scm/linux/kernel/git/acme/net-2.6/fs/ext3/balloc.c 346 */
void ext3_init_block_alloc_info(struct inode * inode);
/* /pub/scm/linux/kernel/git/acme/net-2.6/fs/ext3/balloc.c 388 */
void ext3_discard_reservation(struct inode * inode);
Signed-off-by: Arnaldo Carvalho de Melo <acme@mandriva.com>
so that one can traverse all the classes loaded by classes__load.
Also export classes__find_by_id().
Signed-off-by: Arnaldo Carvalho de Melo <acme@mandriva.com>
see that the function was indeed inlined:
[acme@newtoy net-2.6]$ pfunct kernel/sched.o task_running
/* /pub/scm/linux/kernel/git/acme/net-2.6/kernel/sched.c 304 */
inline int task_running(struct rq * rq, struct task_struct * p);
[acme@newtoy net-2.6]$
Signed-off-by: Arnaldo Carvalho de Melo <acme@mandriva.com>
print just the structs with holes, that indeed is what we do now if no class
name is passed.
Signed-off-by: Arnaldo Carvalho de Melo <acme@mandriva.com>
