Bug in 2.4.26 in mm/filemap.c when using RLIMIT_RSS
Subsystems (2)
Use of rlim[RLIMIT_RSS] in mm/filemap.c is wrong.
It is passed down to kernel as a number of bytes but is being used as a number of pages.
There is also a misinformative comment in fs/proc/array.c in proc_pid_stat where it says
mm ? mm->rss : 0, /* you might want to shift this left 3 */ the number 3 should probably be PAGE_SHIFT-10.
Signature · 3 lines
Ake Sandgren, HPC2N, Umea University, S-90187 Umea, Sweden Internet: [email protected] Phone: +46 90 7866134 Fax: +46 90 7866126 Mobile: +46 70 7716134 WWW: http://www.hpc2n.umu.se
Quoted 8 lines · Ake
Use of rlim[RLIMIT_RSS] in
mm/filemap.cis wrong.
It is passed down to kernel as a number of bytes but is being used as a number of pages.There is also a misinformative comment in
fs/proc/array.cinproc_pid_statwhere it says
mm ?mm->rss: 0, /* you might want to shift this left 3 */ the number 3 should probably be PAGE_SHIFT-10.
Amazing that this has never been noticed before - I bet not many people use RSS limits with madvise().
This transform the rlimit in pages before the comparison, can you please test it.
mm/filemap.c renamed from mm/filemap.c.orig +3 -0
@@ -2609,6 +2609,9 @@ error = -EIO ; rlim_rss = current->rlim ? current->rlim[RLIMIT_RSS ].rlim_cur :LONG_MAX ;/* default: see resource.h */ + + rlim_rss = (rlim_rss &PAGE_MASK ) >>PAGE_SHIFT ; +if ((vma->vm_mm->rss + (end - start)) > rlim_rss)return error;
Quoted 4 lines · Marcelo Tosatti
Quoted 4 lines
There is also a misinformative comment in
fs/proc/array.cinproc_pid_statwhere it says
mm ?mm->rss: 0, /* you might want to shift this left 3 */ the number 3 should probably be PAGE_SHIFT-10.
Don't forget the comment (it's still there in 2.6) :-)
Quoted 2 lines
Amazing that this has never been noticed before - I bet not many people use RSS limits with
madvise().
Neither do we. :-)
I just found it when trying to figure out if the kernel actually used any of the rlimits and if so how.
(Trying to make PBS work correctly)
Quoted 6 lines
This transform the rlimit in pages before the comparison, can you please test it.
mm/filemap.c line 2616
Quoted 8 lines
mm/filemap.c renamed from mm/filemap.c.orig +3
-0@@ -2609,6 +2609,9 @@ error = -EIO ; rlim_rss = current->rlim ? current->rlim[RLIMIT_RSS ].rlim_cur :LONG_MAX ;/* default: see resource.h */ + + rlim_rss = (rlim_rss &PAGE_MASK ) >>PAGE_SHIFT ; +if ((vma->vm_mm->rss + (end - start)) > rlim_rss)return error;
Since we don't use it i can't really test it, but a visual inspection is good enough in this simple case. And since this is the only place in 2.4 that RLIMIT_RSS is used, the problem is solved.
BTW do you know if there is any plans for 2.6++ to actually use RLIMIT_RSS? I saw a hint in that direction in mm/thrash.c grab_swap_token but it is commented out and only skeleton code...
I'm asking since the above code piece has been removed.
Signature · 3 lines
Ake Sandgren, HPC2N, Umea University, S-90187 Umea, Sweden Internet: [email protected] Phone: +46 90 7866134 Fax: +46 90 7866126 Mobile: +46 70 7716134 WWW: http://www.hpc2n.umu.se
Quoted 36 lines · Ake
Quoted 4 lines · Marcelo Tosatti
Quoted 1 line
There is also a misinformative comment in
fs/proc/array.cin
proc_pid_statwhere it saysQuoted 2 lines
mm ?
mm->rss: 0, /* you might want to shift this left 3 */ the number 3 should probably be PAGE_SHIFT-10.Don't forget the comment (it's still there in 2.6) :-)
Quoted 2 lines
Amazing that this has never been noticed before - I bet not many people use RSS limits with
madvise().Neither do we. :-)
I just found it when trying to figure out if the kernel actually used any of the rlimits and if so how.
(Trying to make PBS work correctly)Quoted 6 lines
This transform the rlimit in pages before the comparison, can you please test it.
mm/filemap.c line 2616
Quoted 8 lines
mm/filemap.c renamed from mm/filemap.c.orig +3
-0@@ -2609,6 +2609,9 @@ error = -EIO ; rlim_rss = current->rlim ? current->rlim[RLIMIT_RSS ].rlim_cur :LONG_MAX ;/* default: see resource.h */ + + rlim_rss = (rlim_rss &PAGE_MASK ) >>PAGE_SHIFT ; +if ((vma->vm_mm->rss + (end - start)) > rlim_rss)return error;Since we don't use it i can't really test it, but a visual inspection is good enough in this simple case. And since this is the only place in 2.4 that
RLIMIT_RSSis used, the problem is solved.
Quoted 3 lines
BTW do you know if there is any plans for 2.6++ to actually use
RLIMIT_RSS? I saw a hint in that direction inmm/thrash.cgrab_swap_tokenbut it is commented out and only skeleton code...
Nope, RLIMIT_RSS does not seem to be used at all in v2.6:
linux-2.6.11-rc1 # findgrep RLIMIT_RSS./fs/proc/array.c./include/asm-arm26/resource.h./include/asm-ia64/resource.h./include/asm-frv/resource.h./include/asm-i386/resource.h./include/asm-alpha/resource.h./include/asm-ppc64/resource.h./include/asm-x86_64/resource.h./include/asm-s390/resource.h./include/asm-mips/resource.h./include/asm-sparc64/resource.h./include/asm-cris/resource.h./include/asm-v850/resource.h./include/asm-m68k/resource.h./include/asm-sparc/resource.h./include/asm-parisc/resource.h./include/asm-arm/resource.h./include/asm-sh/resource.h./include/asm-m32r/resource.h./include/asm-h8300/resource.h./include/asm-ppc/resource.h
Quoted 1 line
I'm asking since the above code piece has been removed.
Its there for compatibility reasons, support for it might be added in the future?
Quoted 14 lines · Marcelo Tosatti
Quoted 13 lines · Ake
mm/filemap.c line 2616
Quoted 12 lines · Marcelo Tosatti
mm/filemap.c renamed from mm/filemap.c.orig +3
-0@@ -2609,6 +2609,9 @@ error = -EIO ; rlim_rss = current->rlim ? current->rlim[RLIMIT_RSS ].rlim_cur :LONG_MAX ;/* default: see resource.h */ + + rlim_rss = (rlim_rss &PAGE_MASK ) >>PAGE_SHIFT ; +if ((vma->vm_mm->rss + (end - start)) > rlim_rss)return error;
Isn't the right fix just to remove rlim_rss and this RLIMIT_RSS code from here? It's pretty silly for madvise alone to be taking notice of it. Presumably the code crept in by mistake from some tree which was using an RLIMIT_RSS patch on 2.4.
Quoted 4 lines
Quoted 4 lines
Since we don't use it i can't really test it, but a visual inspection is good enough in this simple case. And since this is the only place in 2.4 that
RLIMIT_RSSis used, the problem is solved.
Quoted 8 lines
Quoted 3 lines
BTW do you know if there is any plans for 2.6++ to actually use
RLIMIT_RSS? I saw a hint in that direction inmm/thrash.cgrab_swap_tokenbut it is commented out and only skeleton code...Nope,
RLIMIT_RSSdoes not seem to be used at all in v2.6:Its there for compatibility reasons, support for it might be added in the future?
Rik had a patch implementing RLIMIT_RSS in 2.6-mm for a while. But I think there were a couple of problems with it, and no great demand for the feature, so Andrew dropped it until someone makes a better case for it.
Hugh
Quoted 4 lines · Hugh Dickins
Quoted 3 lines · Marcelo Tosatti
Quoted 2 lines · Ake
Quoted 1 line · Marcelo Tosatti
Quoted 16 lines
Quoted 11 lines
Quoted 11 lines
mm/filemap.c line 2616
Quoted 11 lines
mm/filemap.c renamed from mm/filemap.c.orig +3
-0@@ -2609,6 +2609,9 @@ error = -EIO ; rlim_rss = current->rlim ? current->rlim[RLIMIT_RSS ].rlim_cur :LONG_MAX ;/* default: see resource.h */ + + rlim_rss = (rlim_rss &PAGE_MASK ) >>PAGE_SHIFT ; +if ((vma->vm_mm->rss + (end - start)) > rlim_rss)return error;Isn't the right fix just to remove
rlim_rssand thisRLIMIT_RSScode from here? It's pretty silly for madvise alone to be taking notice of it. Presumably the code crept in by mistake from some tree which was using anRLIMIT_RSSpatch on 2.4.
True. Will do it.
Quoted 19 lines
Quoted 4 lines
Quoted 4 lines
Since we don't use it i can't really test it, but a visual inspection is good enough in this simple case. And since this is the only place in 2.4 that
RLIMIT_RSSis used, the problem is solved.Quoted 8 lines
Quoted 3 lines
BTW do you know if there is any plans for 2.6++ to actually use
RLIMIT_RSS? I saw a hint in that direction inmm/thrash.cgrab_swap_tokenbut it is commented out and only skeleton code...Nope,
RLIMIT_RSSdoes not seem to be used at all in v2.6:Its there for compatibility reasons, support for it might be added in the future?
Rik had a patch implementing
RLIMIT_RSSin 2.6-mm for a while. But I think there were a couple of problems with it, and no great demand for the feature, so Andrew dropped it until someone makes a better case for it.Hugh
Quoted 6 lines · Hugh Dickins
Quoted 6 lines
Quoted 3 lines
BTW do you know if there is any plans for 2.6++ to actually use
RLIMIT_RSS? I saw a hint in that direction inmm/thrash.cgrab_swap_tokenbut it is commented out and only skeleton code...Nope,
RLIMIT_RSSdoes not seem to be used at all in v2.6:
Quoted 7 lines
Quoted 2 lines
Its there for compatibility reasons, support for it might be added in the future?
Rik had a patch implementing
RLIMIT_RSSin 2.6-mm for a while. But I think there were a couple of problems with it, and no great demand for the feature, so Andrew dropped it until someone makes a better case for it.
Well, the use for it is for compute clusters where you really would like to be able to limit this. Esp on smp boxes where there is multiple compute jobs running simultaneously. Be it mpi or separate jobs you really want to limit their RSS use so they don't steal memory from each other.
Signature · 3 lines
Ake Sandgren, HPC2N, Umea University, S-90187 Umea, Sweden Internet: [email protected] Phone: +46 90 7866134 Fax: +46 90 7866126 Mobile: +46 70 7716134 WWW: http://www.hpc2n.umu.se