Bug 6282

Summary:	wrong madvise(MADV_DONTNEED) semantic
Product:	Memory Management	Reporter:	Samuel Thibault (samuel.thibault)
Component:	Other	Assignee:	Natalie Protasevich (protasnb)
Status:	ASSIGNED ---
Severity:	normal	CC:	protasnb, xerofoify
Priority:	P2
Hardware:	i386
OS:	Linux
Kernel Version:	any since madvise implementation (somewhere in 2.4)	Subsystem:
Regression:	No	Bisected commit-id:
Attachments:	initialization dmesg (in case it can be useful) additionnal dmesg on modprobe atyfb testcase

Description Samuel Thibault 2006-03-24 14:42:44 UTC

Hi,

That

Comment 1 Andrew Morton 2006-03-24 14:47:32 UTC

Please take it to linux-kernel.  Include the testcase if poss, thanks.

Comment 2 Samuel Thibault 2006-03-24 14:52:27 UTC

URL of thread:
http://marc.theaimsgroup.com/?t=111996860600002&r=1&w=2&n=13

Comment 3 Samuel Thibault 2006-03-25 06:37:26 UTC

Created attachment 7667 [details]
initialization dmesg (in case it can be useful)

Comment 4 Samuel Thibault 2006-03-25 06:38:10 UTC

Created attachment 7668 [details]
additionnal dmesg on modprobe atyfb

Comment 5 Samuel Thibault 2006-03-25 06:39:14 UTC

Comment on attachment 7667 [details]
initialization dmesg (in case it can be useful)

Oops, wrong bug report

Comment 6 Samuel Thibault 2006-03-25 06:39:37 UTC

Comment on attachment 7668 [details]
additionnal dmesg on modprobe atyfb

Oops, wrong bug report

Comment 7 Samuel Thibault 2007-03-01 10:05:13 UTC

See glibc bug http://sourceware.org/bugzilla/show_bug.cgi?id=3458
Drepper "fixed" the issue for now by ignoring MADV_DONTNEED advice.

Comment 8 Samuel Thibault 2007-03-01 10:05:55 UTC

Note: that of course does not mean that we shouldn't implement something that really is what POSIX expects from MADV_DONTNEED

Comment 9 Natalie Protasevich 2008-03-20 00:40:08 UTC

Samuel, have you tried advice in #1 about posting this to LKML?
Thanks.

Comment 10 Samuel Thibault 2008-03-20 03:42:45 UTC

See  http://marc.info/?l=linux-kernel&m=114324188901936&w=2  but no luck.  There are related threads:  http://marc.info/?l=linux-kernel&m=117561753326551&w=2 http://marc.info/?l=linux-kernel&m=117626557716611&w=2 http://marc.info/?l=linux-kernel&m=117679424209965&w=2 http://marc.info/?l=linux-kernel&m=117798097213454&w=2  And they indeed agree that our MADV_DONTNEED is not POSIX and that we should add a POSIX_MADV_DONTNEED, but don't do anything about it. So it looks like yes, the issue was raised and people know about it, but nobody really cares and glibc is still wrapping it with a nop...

Comment 11 xerofoify 2014-06-25 18:58:15 UTC

This bug is very old. Please try on newer kernel.
Cheers Nick

Comment 12 Samuel Thibault 2014-06-29 19:10:24 UTC

Created attachment 141411 [details]
testcase

Hello,

Yes, this is still a problem on Linux 3.15.  I have attached again the
testcase, which is still emitting "FAILED".

Samuel

Comment 13 xerofoify 2014-08-08 19:29:54 UTC

I have read your test case and would like to known on what lines it's failing or does it just fail all togethor.
Nick

Comment 14 Samuel Thibault 2014-08-11 20:40:09 UTC

Here is the output I get:

c = 00204a16      file msync:  NO madvise:  NO ok
c = 05542a27      file msync: YES madvise:  NO ok
c = 169c39e2      file msync:  NO madvise: YES FAILED
c = 2eab4956      file msync: YES madvise: YES FAILED
c = 0bb0b8b4 anonymous msync:  NO madvise:  NO ok
c = 0bd068c9 anonymous msync: YES madvise:  NO ok
c = 3e5d224d anonymous msync:  NO madvise: YES FAILED
c = 436c8d9c anonymous msync: YES madvise: YES FAILED

Whenever madvise(MADV_DONTNEED) is called, the value of the cookie gets
lost.