Re: [PATCH 0/2] Report the size of pages backing VMAs in /proc V3

Re: [PATCH 0/2] Report the size of pages backing VMAs in /proc V3

[Albert Cahalan]

Adding " (hpagesize=4096kB)" onto the end of a filename is as vile
as adding " (deleted)" onto the end. If anything is going to change
in this area, it should be the elimination of " (deleted)". These
tags are perfectly legitimate in filenames.

Looping on stat() while chopping off suspected tags is dreadful.
Besides just being gross, it's slow.

gdb will tolerate up to 7 flags, procps will tolerate up to 31 flags,
and both will tolerate anything without a '/' before the filename.

Obviously, every author of a /proc-based tool has been forced to
take a random guess at the ABI. The /proc/*/smaps is so gross and
that I put off writing a parser for years.

What you can probably get away with:

After the "rwxp" stuff you can add 3 more flags. (gdb limit)
You could use 'L' for locked pages, 'R' for swap reservation,
and 'D' for deleted files. It's probably much better to save
space though, since gdb will crash if you add too many flags.
Three characters can be 18 bits if you base-64 encode them,
being careful to avoid the '/' character. (adding 0x30 works)

Right before the filename, you can add anything except a '/'.
You could add a few columns of numbers or a second flags field.

Not that it matters on such a slow-ass file format, but you
can make parsing faster if you encode the page size in one byte.
Simply add 0x30 to the page shift, then print that byte. Note that
this would let you cram the page size into the flags field.

BTW, I'm thinking that the /proc/*/*maps files fail when the
lines exceed 4096 bytes. The pathname may legitimately be that
long, plus it can be backslash escaped, plus there is all the
junk on the beginning.

Re: [PATCH 0/2] Report the size of pages backing VMAs in /proc V3

[Mel Gorman]

On (20/10/08 05:18), Albert Cahalan didst pronounce:
> Adding " (hpagesize=4096kB)" onto the end of a filename is as vile
> as adding " (deleted)" onto the end. If anything is going to change
> in this area, it should be the elimination of " (deleted)". These
> tags are perfectly legitimate in filenames.
> 

I dropped that change altogether in the last series because of concerns like
this. It did have the potential to grow to something weird looking.

> Looping on stat() while chopping off suspected tags is dreadful.
> Besides just being gross, it's slow.
> 

You're probably right. It's a bit weird that it's what you have to do to
figure out if the file in /proc/PID/maps is really there or not.

> gdb will tolerate up to 7 flags, procps will tolerate up to 31 flags,
> and both will tolerate anything without a '/' before the filename.
> 

Understood.

> Obviously, every author of a /proc-based tool has been forced to
> take a random guess at the ABI. The /proc/*/smaps is so gross and
> that I put off writing a parser for years.
> 

I intend to take a stab at it for the purposes of teaching pmap to print
the pagesizes if the smaps change gets picked up.

> What you can probably get away with:
> 
> After the "rwxp" stuff you can add 3 more flags. (gdb limit)
> You could use 'L' for locked pages, 'R' for swap reservation,
> and 'D' for deleted files. It's probably much better to save
> space though, since gdb will crash if you add too many flags.
> Three characters can be 18 bits if you base-64 encode them,
> being careful to avoid the '/' character. (adding 0x30 works)
> 

Ok, noted in case I ever decide to tackle the (deleted) removal. It's
not something I feel strongly about though.

> Right before the filename, you can add anything except a '/'.
> You could add a few columns of numbers or a second flags field.
> 

My fear was about parsers that hard-coded what number field stored the
filename. If a column was added for pagesize for example, then parsers
would think the pagesize was the filename.

> Not that it matters on such a slow-ass file format, but you
> can make parsing faster if you encode the page size in one byte.
> Simply add 0x30 to the page shift, then print that byte. Note that
> this would let you cram the page size into the flags field.
> 

Now, that is an interested idea, albeit it's not one that is easily
human-readable and would need a second parser like pmap but that's ok. If
parsing smaps turns into a total pain in the ass or the performance overhead
of calculating PSS when reading the pagesize becomes a problem, then I'll
try this option. Thanks a lot for that idea.

> BTW, I'm thinking that the /proc/*/*maps files fail when the
> lines exceed 4096 bytes. The pathname may legitimately be that
> long, plus it can be backslash escaped, plus there is all the
> junk on the beginning.
> 

Yes. While it's unlikely to be exceeded, a file could be 4096 bytes long
and the other fields will then cause a problem. It was because of things
like this, I was ok with dropping the idea of adding (attribute[=value])
from the end of the filename.

-- 
Mel Gorman
Part-time Phd Student                          Linux Technology Center
University of Limerick                         IBM Dublin Software Lab

Re: [PATCH 0/2] Report the size of pages backing VMAs in /proc V3

[Albert Cahalan]

On Mon, Oct 20, 2008 at 6:06 AM, Mel Gorman <mel@csn.ul.ie> wrote:
> On (20/10/08 05:18), Albert Cahalan didst pronounce:

>> Looping on stat() while chopping off suspected tags is dreadful.
>> Besides just being gross, it's slow.
>
> You're probably right. It's a bit weird that it's what you have to do to
> figure out if the file in /proc/PID/maps is really there or not.

Actually you can't do this, because of directory permissions.

>> Obviously, every author of a /proc-based tool has been forced to
>> take a random guess at the ABI. The /proc/*/smaps is so gross and
>> that I put off writing a parser for years.
>
> I intend to take a stab at it for the purposes of teaching pmap to print
> the pagesizes if the smaps change gets picked up.

FYI, "KernelPageSize" is at least unique under the perfect
hash function I'm using to parse the damn smaps file.

hash = ( ( (s[8]&15) + (s[1]&15) ) ^ (s[0]&3) ) & 31;

I have to wonder if we'll be getting mixed page sizes
within a single mapping, making such info unusable.

>> Right before the filename, you can add anything except a '/'.
>> You could add a few columns of numbers or a second flags field.
>
> My fear was about parsers that hard-coded what number field stored the
> filename. If a column was added for pagesize for example, then parsers
> would think the pagesize was the filename.

It's possible. Every parser I've examined does strchr()
or similar to find that '/' character.

Maybe try some dummy patches in a linux-next kernel?
Give each one a month. You could do "xyz" concatenated
to the flags, a second "rwx" concatenated to the flags,
a single column of "0" before the filename, and several
columns of "parsertest" before the filename.

> Now, that is an interested idea, albeit it's not one that is easily
> human-readable and would need a second parser like pmap but that's ok. If
> parsing smaps turns into a total pain in the ass

I assure you that parsing smaps is a total pain in the ass,
especially if you want tolerable performance. Something
like "top" is not viable if it performs like a Python script.

>> BTW, I'm thinking that the /proc/*/*maps files fail when the
>> lines exceed 4096 bytes. The pathname may legitimately be that
>> long, plus it can be backslash escaped, plus there is all the
>> junk on the beginning.
>
> Yes. While it's unlikely to be exceeded, a file could be 4096 bytes long
> and the other fields will then cause a problem. It was because of things
> like this, I was ok with dropping the idea of adding (attribute[=value])
> from the end of the filename.

"unlikely" is not something one should trust. I think you
can even get a name longer than 4096 bytes if you make
directories relative to the current directory and keep
changing directories as you make the directories.
Then double that with backslashes becoming \\ or
newlines becoming \n (must be escaped) in the output.

I think /proc/*/maps has been broken ever since it was
converted to seq_file, and maybe ever since it got filenames.
Prior to the filenames, lines were fixed-width records.

Re: [PATCH 0/2] Report the size of pages backing VMAs in /proc V3

[Mel Gorman]

On Mon, Oct 20, 2008 at 02:07:28PM -0400, Albert Cahalan wrote:
> On Mon, Oct 20, 2008 at 6:06 AM, Mel Gorman <mel@csn.ul.ie> wrote:
> > On (20/10/08 05:18), Albert Cahalan didst pronounce:
> 
> >> Looping on stat() while chopping off suspected tags is dreadful.
> >> Besides just being gross, it's slow.
> >
> > You're probably right. It's a bit weird that it's what you have to do to
> > figure out if the file in /proc/PID/maps is really there or not.
> 
> Actually you can't do this, because of directory permissions.
> 

Good point.

> >> Obviously, every author of a /proc-based tool has been forced to
> >> take a random guess at the ABI. The /proc/*/smaps is so gross and
> >> that I put off writing a parser for years.
> >
> > I intend to take a stab at it for the purposes of teaching pmap to print
> > the pagesizes if the smaps change gets picked up.
> 
> FYI, "KernelPageSize" is at least unique under the perfect
> hash function I'm using to parse the damn smaps file.
> 
> hash = ( ( (s[8]&15) + (s[1]&15) ) ^ (s[0]&3) ) & 31;
> 

Good to know.

> I have to wonder if we'll be getting mixed page sizes
> within a single mapping, making such info unusable.
> 

It's not planned right now, but even if it is, KernelPageSize would
remain as the intended page size. VMAs would either split around each
mixed page size in which case there will be separate VMAs or an
additional field will be added that indicates what number of each
pagesize makes up the mapping.

> >> Right before the filename, you can add anything except a '/'.
> >> You could add a few columns of numbers or a second flags field.
> >
> > My fear was about parsers that hard-coded what number field stored the
> > filename. If a column was added for pagesize for example, then parsers
> > would think the pagesize was the filename.
> 
> It's possible. Every parser I've examined does strchr()
> or similar to find that '/' character.
> 

I might be the only criminal. A mucky shell script used awk to display
field X and everything past it to find the filename. A more rational
person would have used strchr or found the first / with cut or similar.

> Maybe try some dummy patches in a linux-next kernel?
> Give each one a month. You could do "xyz" concatenated
> to the flags, a second "rwx" concatenated to the flags,
> a single column of "0" before the filename, and several
> columns of "parsertest" before the filename.
> 

That sounds reasonable.

> > Now, that is an interested idea, albeit it's not one that is easily
> > human-readable and would need a second parser like pmap but that's ok. If
> > parsing smaps turns into a total pain in the ass
> 
> I assure you that parsing smaps is a total pain in the ass,
> especially if you want tolerable performance. Something
> like "top" is not viable if it performs like a Python script.
> 

I had assumed that smaps + performance were mutually exclusive because
of the PSS calculation and any active monitoring from something like top
would blow bigtime. That's why I tried modifying maps as well.

> >> BTW, I'm thinking that the /proc/*/*maps files fail when the
> >> lines exceed 4096 bytes. The pathname may legitimately be that
> >> long, plus it can be backslash escaped, plus there is all the
> >> junk on the beginning.
> >
> > Yes. While it's unlikely to be exceeded, a file could be 4096 bytes long
> > and the other fields will then cause a problem. It was because of things
> > like this, I was ok with dropping the idea of adding (attribute[=value])
> > from the end of the filename.
> 
> "unlikely" is not something one should trust. I think you
> can even get a name longer than 4096 bytes if you make
> directories relative to the current directory and keep
> changing directories as you make the directories.
> Then double that with backslashes becoming \\ or
> newlines becoming \n (must be escaped) in the output.
> 
> I think /proc/*/maps has been broken ever since it was
> converted to seq_file, and maybe ever since it got filenames.
> Prior to the filenames, lines were fixed-width records.
> 

You could be right. Only one way to find out for sure really.

-- 
Mel Gorman
Part-time Phd Student                          Linux Technology Center
University of Limerick                         IBM Dublin Software Lab

[PATCH 0/2] Report the size of pages backing VMAs in /proc V3

[Mel Gorman]

The following two patches add support for printing the size of pages used by
the kernel and the MMU to back VMAs. This can be used by a user to verify
that a hugepage-aware application is using the expected page sizes.

The first patch prints the size of page used by the kernel when allocating
pages for a VMA in /proc/pid/smaps. The second patch reports on
the size of page used by the MMU as it can differ - for example on POWER
using 64K as a base pagesize on older processors.

Changelog since V2
  o Drop changes to /proc/pid/maps - could not get agreement and it affects
    procps. Patch to procps was posted but fell into silence. Dropping
    patch as smaps gives the necessary information, just with a bit more
    legwork by the user
  o Drop redundant VM_BUG_ON (Alexey)

Changelog since V1
  o Fix build failure on !CONFIG_HUGETLB_PAGE
  o Uninline helper functions
  o Distinguish between base pagesize and MMU pagesize

 arch/powerpc/include/asm/hugetlb.h |    6 ++++++
 arch/powerpc/mm/hugetlbpage.c      |    7 +++++++
 fs/proc/task_mmu.c                 |    8 ++++++--
 include/linux/hugetlb.h            |    6 ++++++
 mm/hugetlb.c                       |   29 +++++++++++++++++++++++++++++
 5 files changed, 54 insertions(+), 2 deletions(-)

[PATCH 0/2] Report the size of pages backing VMAs in /proc V3

[Mel Gorman]

The following two patches add support for printing the size of pages used
by the kernel to back VMAs in maps and smaps. This can be used by a user
to verify that a hugepage-aware application is using the expected page sizes.
In one case the pagesize used by the MMU differs from the size used by the
kernel. This is on PPC64 using 64K as a base page size running on a processor
that does not support 64K in the MMU. In this case, the kernel uses 64K pages
but the MMU is still using 4K.

The first patch prints the size of page used by the kernel when allocating
pages for a VMA in /proc/pid/smaps and should not be considered too
contentious as it is highly unlikely to break any parsers.  The second patch
reports the size of page used by hugetlbfs regions in /proc/pid/maps. There is
a possibility that the final patch will break parsers but they are arguably
already broken. More details are in the patches themselves.

Thanks to KOSAKI Motohiro for rebasing the patches onto mmotm, reviewing
and testing.

Changelog since V2
  o Drop printing of MMUPageSize (mel)
  o Rebase onto mmotm (KOSAKI Motohiro)

Changelog since V1
  o Fix build failure on !CONFIG_HUGETLB_PAGE
  o Uninline helper functions
  o Distinguish between base pagesize and MMU pagesize

 fs/proc/task_mmu.c      |   27 ++++++++++++++++++---------
 include/linux/hugetlb.h |    3 +++
 mm/hugetlb.c            |   17 +++++++++++++++++
 3 files changed, 38 insertions(+), 9 deletions(-)