> > > Could you please try the attached patch which will also show the
> > > user and process that opened these files? It adds three more
fields
> > > when CONFIG_PROC_FILECACHE_EXTRAS is selected.
> > >
> > > Thanks,
> > > Fengguang
> > >
> > > On Thu, Mar 05, 2009 at 01:55:35PM +0200, Markus wrote:
> > > >
> > > > # sort -n -k 3 filecache-2009-03-05 | tail -n 5
> > > > 15886 7112 7112 100 1 d- 00:08
> > > > (tmpfs) /dev/zero\040(deleted)
> > > > 16209 35708 35708 100 1 d- 00:08
> > > > (tmpfs) /dev/zero\040(deleted)
> > > > 16212 82128 82128 100 1 d- 00:08
> > > > (tmpfs) /dev/zero\040(deleted)
> > > > 15887 340024 340024 100 1 d- 00:08
> > > > (tmpfs) /dev/zero\040(deleted)
> > > > 15884 455008 455008 100 1 d- 00:08
> > > > (tmpfs) /dev/zero\040(deleted)
> > > >
> > > > The sum of the third column is 1013 MB.
> > > > To note the biggest ones (or do you want the whole file?)... and
> > thats
> > > > after a sync and a drop_caches! (Can be seen in the commands
given.)
> >
> > I could, but I know where these things belong to. Its from sphinx (a
> > mysql indexer) searchd. It loads parts of the index into memory.
> > The sizes looked well-known and killing the searchd will
reduce "cached"
> > to a normal amount ;)
>
> And it's weird about the file name: /dev/zero. I wonder how it
> managed to create that file, and then delete it, inside a tmpfs!

I dont know exactly. But in the source its just a:
... mmap ( NULL, m_iLength, PROT_READ | PROT_WRITE, MAP_SHARED |
MAP_ANON, -1, 0 );
Perhaps thats the way shared anonymous memory is handled?!

> Just out of curiosity, are they shm objects? Can you show us the
> output of 'df'? In your convenient time.

Thats all:
# df -h
Filesystem Size Used Avail Use% Mounted on
/dev/md6 14G 7.9G 5.9G 58% /
udev 10M 304K 9.8M 3% /dev
cachedir 4.0M 100K 4.0M 3% /lib64/splash/cache
/dev/md4 19G 15G 3.2G 82% /home
/dev/md3 8.3G 4.5G 3.8G 55% /usr/portage
shm 2.0G 0 2.0G 0% /dev/shm

> > I just dont know why its in "cached" (can that be swapped out btw?).
> > But I think thats not a problem of the kernel, but of anonymous
> > mmap-ing.
>
> You know, because the file is created in tmpfs, which is swap-backed.
> By definition the pages here cannot be dropped by third-party.

Hm, ok.

> > I think its resolved, thanks to everybody and Fengguang in
particular!
>
> You are welcome :-)
;)

Have a nice day.
Markus

2009-03-05 14:54:00

by Fengguang Wu

[permalink] [raw]

Subject: Re: drop_caches ...

On Thu, Mar 05, 2009 at 04:43:22PM +0200, Markus wrote:
> > > > Could you please try the attached patch which will also show the
> > > > user and process that opened these files? It adds three more
> fields
> > > > when CONFIG_PROC_FILECACHE_EXTRAS is selected.
> > > >
> > > > Thanks,
> > > > Fengguang
> > > >
> > > > On Thu, Mar 05, 2009 at 01:55:35PM +0200, Markus wrote:
> > > > >
> > > > > # sort -n -k 3 filecache-2009-03-05 | tail -n 5
> > > > > 15886 7112 7112 100 1 d- 00:08
> > > > > (tmpfs) /dev/zero\040(deleted)
> > > > > 16209 35708 35708 100 1 d- 00:08
> > > > > (tmpfs) /dev/zero\040(deleted)
> > > > > 16212 82128 82128 100 1 d- 00:08
> > > > > (tmpfs) /dev/zero\040(deleted)
> > > > > 15887 340024 340024 100 1 d- 00:08
> > > > > (tmpfs) /dev/zero\040(deleted)
> > > > > 15884 455008 455008 100 1 d- 00:08
> > > > > (tmpfs) /dev/zero\040(deleted)
> > > > >
> > > > > The sum of the third column is 1013 MB.
> > > > > To note the biggest ones (or do you want the whole file?)... and
> > > thats
> > > > > after a sync and a drop_caches! (Can be seen in the commands
> given.)
> > >
> > > I could, but I know where these things belong to. Its from sphinx (a
> > > mysql indexer) searchd. It loads parts of the index into memory.
> > > The sizes looked well-known and killing the searchd will
> reduce "cached"
> > > to a normal amount ;)
> >
> > And it's weird about the file name: /dev/zero. I wonder how it
> > managed to create that file, and then delete it, inside a tmpfs!
>
> I dont know exactly. But in the source its just a:
> ... mmap ( NULL, m_iLength, PROT_READ | PROT_WRITE, MAP_SHARED |
> MAP_ANON, -1, 0 );
> Perhaps thats the way shared anonymous memory is handled?!

Good to know this. The corresponding kernel function is:

/**
* shmem_zero_setup - setup a shared anonymous mapping
* @vma: the vma to be mmapped is prepared by do_mmap_pgoff
*/
int shmem_zero_setup(struct vm_area_struct *vma)
{
struct file *file;
loff_t size = vma->vm_end - vma->vm_start;

file = shmem_file_setup("dev/zero", size, vma->vm_flags);

Here goes the /dev/zero ^_^

> > Just out of curiosity, are they shm objects? Can you show us the
> > output of 'df'? In your convenient time.
>
> Thats all:
> # df -h
> Filesystem Size Used Avail Use% Mounted on
> /dev/md6 14G 7.9G 5.9G 58% /
> udev 10M 304K 9.8M 3% /dev
> cachedir 4.0M 100K 4.0M 3% /lib64/splash/cache
> /dev/md4 19G 15G 3.2G 82% /home
> /dev/md3 8.3G 4.5G 3.8G 55% /usr/portage

> shm 2.0G 0 2.0G 0% /dev/shm

shm objects will be accounted here. It's clean.

Thanks,
Fengguang