Re: ACL limit

[Timothy Shimmin <tes@xxxxxxx>]

David Chinner wrote:
> On Mon, Dec 03, 2007 at 10:47:07AM +1100, Timothy Shimmin wrote:
> > David Chinner wrote:
> > > On Fri, Nov 30, 2007 at 07:07:26PM +0100, Jan Engelhardt wrote:
> > > > Hi,
> > > >
> > > >
> > > > is there any way to raise the number of ACLs that can be stored? The 
> > > > current limit of 25 is quite tight, where ext3 allows 124 and jfs 8192.
> > > > Would increasing XFS_ACL_MAX_ENTRIES work (yes, using potentially more 
> > > > memory), i.e. not interfering with the on-disk format?
> > > It would be an on disk format change - older kernels would error out
> > > (-EINVAL) on > 25 ACLs and not check any of them. Hence we'd
> > > probably need a superblock feature bit to indicate that >25 ACEs are
> > > supported in a given ACL.
> > >
> > > But we can work around that (superblock feature bit) and should
> > > be able to extend this out to ~8190 entries. We're doing an ACL
> > > rework ATM, so > 25 entry support should fall out of that....
> > Yeah, it's just an array of entries in an EA value.
> > The EA value is limited to 64K so it's a question of how
> > many entries you can fit into that.
> > (64K - 4)/12 = 5461
>
> Confused - I thought the ACE was 8 bytes:
>
> struct posix_acl_entry {
>         short                   e_tag;
>         unsigned short          e_perm;
>         unsigned int            e_id;
> };
>
> Also, JFS only allows 64k for the xattr as well (jfs_xattr.h):
>
> /* Macros for defining maxiumum number of bytes supported for EAs */
> #define MAXEASIZE       65535
> #define MAXEALISTSIZE   MAXEASIZE
>
> And (64k - 4)/8 = 8191 which is what JFS supports.
>
> Oh:
>
> typedef __uint16_t      xfs_acl_perm_t;
> typedef __int32_t       xfs_acl_type_t;
> typedef __int32_t       xfs_acl_tag_t;
> typedef __int32_t       xfs_acl_id_t;
>
> #define XFS_ACL_MAX_ENTRIES 25
> #define XFS_ACL_NOT_PRESENT (-1)
>
> typedef struct xfs_acl_entry {
>         xfs_acl_tag_t   ae_tag;
>         xfs_acl_id_t    ae_id;
>         xfs_acl_perm_t  ae_perm;
> } xfs_acl_entry_t;
>
> typedef struct xfs_acl {
>         __int32_t       acl_cnt;
>         xfs_acl_entry_t acl_entry[XFS_ACL_MAX_ENTRIES];
> } xfs_acl_t;
>
> An *XFS* ACE is 12 bytes and hence can't be passed directly to the
> generic code. Tim - are we adding a translation layer or storing
> the generic posix acl format on disk?

There is a translation layer and we have preserved the XFS ACL format
on disk since the irix days.
The only exception to that is on IRIX it actually always stores an
array with 25 entries in it (irrespective of the count) which meant
for default inode size the irix acls were always out of line - a bit
silly. Decided not to continue with that on Linux ;-)

So yes, it's the xfs_acl (12 byte aces) we have on disk.

--Tim