Today I've repeated the experience described in
http://lkml.org/lkml/2008/3/15/111 and I got these
latencytop numbers(!)
Cause Maximum Percentage
sync_page __lock_page handle_mm_fault do_page_faul103659.0 msec 59.8 %
get_request_wait __make_request generic_make_reque1992.0 msec 10.2 %
get_request_wait __make_request generic_make_reque1620.7 msec 4.6 %
sync_page __lock_page find_lock_page filemap_fault399.3 msec 3.4 %
sync_buffer __wait_on_buffer __bread ext3_get_bran292.9 msec 1.3 %
sync_page __lock_page handle_mm_fault do_page_faul200.5 msec 0.3 %
Scheduler: waiting for cpu 155.9 msec 18.9 %
sync_page __lock_page handle_mm_fault do_page_faul117.5 msec 0.1 %
congestion_wait try_to_free_pages __alloc_pages re103.5 msec 0.2 %
r_code
Process X (2910)
sync_page __lock_page handle_mm_fault do_page_faul103659.0 msec 98.5 %
Scheduler: waiting for cpu 64.6 msec 1.3 %
sync_page __lock_page find_lock_page filemap_fault 51.4 msec 0.2 %lt do_page_fault error_code
sync_buffer __wait_on_buffer __bread ext3_get_bran 17.8 msec 0.0 % ext3_get_block do_mpage_readpage mpage_readpa
sync_page __lock_page handle_mm_fault do_page_faul 9.1 msec 0.0 %ig ault
The experience goes like this:
1) Boot with 2.6.25-rc6-00243-g028011e
2) Log into X
3) Open a 380MB file with xjed
It takes more than 6 minutes to load the file, and meanwhile
I experience very bad desktop interactivity (like 1 minute to
the result of pressing F12 in Window Maker to appear in the
screen).
If, after step 2), I start using firefox, thunderbird, play some
music etc and then close all these apps and go to step 3),
the loading finishes in about 2 minutes (and I have very
good interactivity meanwhile).
So I noticed that when xjed is opening the file it is in
'D' state (reported by ps), and while greping the kernel
source code for 'sync_page' I've found this comment
in mm/filemap.c:
/**
* __lock_page - get a lock on the page, assuming we need to sleep to get it
* @page: the page to lock
*
* Ugly. Running sync_page() in state TASK_UNINTERRUPTIBLE is scary. If some
* random driver's requestfn sets TASK_RUNNING, we could busywait. However
* chances are that on the second loop, the block layer's plug list is empty,
* so sync_page() will then return in state TASK_UNINTERRUPTIBLE.
*/
"Ugly. Running sync_page() in state TASK_UNINTERRUPTIBLE is scary."
It appears that my problem with xjed in 'D' state while loading,
and sync_page() appearing in latencytop with 103 secs of latency
may be related through the "ugliness" described above.
So I was wondering if there is a way to fix this. Note that
this issue does not happen if I load the file after using the
computer for a while, so it is not impossible to have good
interactivity while loading that big file.
I am sorry that I am scattering reports about this issue all
over the place.
--
