Re: [RFC V5] Add gup trace points support

From: Shi, Yang
Date: Tue Jan 12 2016 - 15:01:03 EST

Hi Steven,

Any more comments on this series? How should I proceed it?


On 12/9/2015 1:22 PM, Yang Shi wrote:
* Fixed a typo introduced by v4 rebase
* Removed redundant "#define CREATE_TRACE_POINTS" from architecture specifc

* Adopted Steven's suggestion to use "unsigned int" for nr_pages to save
space in ring buffer since it is unlikely to have more than 0xffffffff
pages are touched by gup in one invoke
* Remove unnecessray type cast

* Adopted suggestion from Dave Hansen to move the gup header include to the last
* Adopted comments from Steven:
- Just keep necessary TP_ARGS
* Moved archtichture specific fall-backable fast version trace point after the
do while loop since it may jump to the slow version.
* Not implement recording return value since Steven plans to have it in generic
tracing code

* Adopted commetns from Steven
- remove all reference to tsk->comm since it is unnecessary for non-sched
trace points
- reduce arguments for __get_user_pages trace point and update mm/gup.c
* Added Ralf's acked-by for patch 4/7.

Some background about why I think this might be useful.

When I was profiling some hugetlb related program, I got page-faults event
doubled when hugetlb is enabled. When I looked into the code, I found page-faults
come from two places, do_page_fault and gup. So, I tried to figure out which
play a role (or both) in my use case. But I can't find existing finer tracing
event for sub page-faults in current mainline kernel.

So, I added the gup trace points support to have finer tracing events for
page-faults. The below events are added:


Both __get_user_pages and fixup_user_fault call handle_mm_fault.

Just added trace points to raw version __get_user_pages since all variants
will call it finally to do real work.

Although __get_user_pages_fast doesn't call handle_mm_fault, it might be useful
to have it to distinguish between slow and fast version.