[ Added the perf tracepoint maintainers ]
On Tue, 11 Feb 2014 22:59:58 -0800
Linus Torvalds <torvalds@xxxxxxxxxxxxxxxxxxxx> wrote:
> On Tue, Feb 11, 2014 at 10:31 PM, Dave Chinner <david@xxxxxxxxxxxxx> wrote:
> > FYI, just creating lots of files with open(O_CREAT):
> > [ 348.718357] fs_mark (4828) used greatest stack depth: 2968 bytes left
> > [ 348.769846] fs_mark (4814) used greatest stack depth: 2312 bytes left
> > [ 349.777717] fs_mark (4826) used greatest stack depth: 2280 bytes left
> > [ 418.139415] fs_mark (4928) used greatest stack depth: 1936 bytes left
> > [ 460.492282] fs_mark (4993) used greatest stack depth: 1336 bytes left
> > [ 544.825418] fs_mark (5104) used greatest stack depth: 1112 bytes left
> > [ 689.503970] fs_mark (5265) used greatest stack depth: 1000 bytes left
> > We've got absolutely no spare stack space anymore in the IO path.
> > And the IO path can't get much simpler than filesystem -> virtio
> > block device.
> Ugh, that's bad. A thousand bytes of stack space is much too close to
> any limits.
> Do you have the stack traces for these things so that we can look at
> worst offenders?
> If the new block-mq code is to blame, it needs to be fixed.
> __virtblk_add_req() has a 300-byte stack frame, it seems. Looking
> elsewhere, blkdev_issue_discard() has 350 bytes of stack frame, but is
> hopefully not in any normal path - online discard is moronic, and I'm
> assuming XFS doesn't do that.
> There's a lot of 200+ byte stack frames in block/blk-core.s, and they
> all seem to be of the type perf_trace_block_buffer() - things created
> with DECLARE_EVENT_CLASS(), afaik. Why they all have 200+ bytes of
> frame, I have no idea. That sounds like a potential disaster too,
> although hopefully it's mostly leaf functions - but leaf functions
> *deep* in the callchain. Tejun? Steven, why _do_ they end up with such
> huge frames?
The perf_trace_##event is defined in include/trace/ftrace.h.
There we have this:
perf_trace_##call(void *__data, proto) \
struct ftrace_event_call *event_call = __data; \
struct ftrace_data_offsets_##call __maybe_unused __data_offsets;\
struct ftrace_raw_##call *entry; \
struct pt_regs __regs; \
u64 __addr = 0, __count = 1; \
struct task_struct *__task = NULL; \
struct hlist_head *head; \
int __entry_size; \
int __data_size; \
int rctx; \
Mostly pointers except for two structures. The __data_offests, is
dynamically defined, and only consists of values from the tracepoint
entry_structure that defines dynamic arrays. But the other structure on
the stack looks a bit harrier. The pt_regs structure.
That's what? 21 unsigned longs? 21 * 8 = 168. I think that's the
Peter and Frederic, is there a way not to store that on the stack?
> And if the stack use comes from the VFS layer, we can probably work on
> that too. But I don't think that has really changed much lately..