Can prune_icache safely discard inodes which have only clean pages? (2.4.18)

Can prune_icache safely discard inodes which have only clean pages? (2.4.18)

[Bill Davenport]

I've got a system which has a fairly large amount of physical memory (2GB)
that experiences
performance problems after a large number of files have been accessed.

Looking into the slab_info I discovered that a very large number of inodes
are currently
present in the system (along with many buffer headers). Digging deeper I was
able to determine
that most of the inodes were on the inode_unused chain, but were being
skipped over during
the prunce_icache processing because they have a non-zero number of pages
(i_data.nrpages).
Looking a bit deeper I discovered that most of the inodes had only pages
that are on the
clean_pages list, with these pages also accounting for many of the buffer
heads.

The system wasn't attempting to free these pages (presumably since it still
had a fair
amount of physical memory available, so it didn't need to do this).

Is there any danger in changing prune_icache to also pick an inode for
pruning if it has
a non-zero page count where the dirty_list and locked_list are empty?

In particular, the existing code in fs/inode.c looks somewhat like:

 #define CAN_UNUSE(inode) \
  ((((inode)->i_state | (inode)->i_data.nrpages) == 0)  && \
   !inode_has_buffers(inode))
 void prune_icache(int goal)
 {
  ...
  while (entry != &inode_unused)
  {
   ...
   if (inode->i_state & (I_FREEING|I_CLEAR|I_LOCK))
    continue;
   if (!CAN_UNUSE(inode))
    continue;

   Remove inode from i_hash and add to freeable
  }
  ...
  dispose_list(freeable);
  ...
 }

and I'd like to change it to:

 void prune_icache(int goal)
 {
  ...
  while (entry != &inode_unused)
  {
   ...
   if (inode->i_state & (I_FREEING|I_CLEAR|I_LOCK))
    continue;
   if ((inode->i_state != 0) || inode_has_buffers(inode))
    continue;
   if (inode->i_data.nrpages != 0) {
    if ((!list_empty(&inode->i_data.dirty_pages)) ||
        (!list_empty(&inode->i_data.locked_pages))) {
     /* skip if any dirty or locked pages */
     continue;
    }
   }

   Remove inode from i_hash and add to freeable
  }
  ...
  dispose_list(freeable);
  ...
 }

Re: Can prune_icache safely discard inodes which have only clean pages? (2.4.18)

[Andrew Morton]

Bill Davenport wrote:
> 
> I've got a system which has a fairly large amount of physical memory (2GB)
> that experiences
> performance problems after a large number of files have been accessed.
> 
> ...

Your analysis is 100% correct.  It's a problem.

There's a fix for this in Andrea's kernel.
 
> ...
> 
> and I'd like to change it to:
> 
>  void prune_icache(int goal)
>  {
>   ...
>   while (entry != &inode_unused)
>   {
>    ...
>    if (inode->i_state & (I_FREEING|I_CLEAR|I_LOCK))
>     continue;
>    if ((inode->i_state != 0) || inode_has_buffers(inode))
>     continue;
>    if (inode->i_data.nrpages != 0) {
>     if ((!list_empty(&inode->i_data.dirty_pages)) ||
>         (!list_empty(&inode->i_data.locked_pages))) {
>      /* skip if any dirty or locked pages */
>      continue;
>     }
>    }

locked_pages tends to hold clean, unlocked pages, alas.  Testing
->dirty_pages makes sense.

If there are no dirty pages then you can run invalidate_inode_pages();
chances are, that will bring ->nrpages to zero, and all is well.