Re: [PATCH 6/9] XArray: internal node is a xa_node when it is bigger than XA_ZERO_ENTRY

[Matthew Wilcox]

On Mon, Mar 30, 2020 at 10:20:13PM +0000, Wei Yang wrote:
> On Mon, Mar 30, 2020 at 07:27:08AM -0700, Matthew Wilcox wrote:
> >On Mon, Mar 30, 2020 at 02:13:50PM +0000, Wei Yang wrote:
> >> On Mon, Mar 30, 2020 at 06:49:03AM -0700, Matthew Wilcox wrote:
> >> >On Mon, Mar 30, 2020 at 01:45:19PM +0000, Wei Yang wrote:
> >> >> On Mon, Mar 30, 2020 at 05:50:06AM -0700, Matthew Wilcox wrote:
> >> >> >On Mon, Mar 30, 2020 at 12:36:40PM +0000, Wei Yang wrote:
> >> >> >> As the comment mentioned, we reserved several ranges of internal node
> >> >> >> for tree maintenance, 0-62, 256, 257. This means a node bigger than
> >> >> >> XA_ZERO_ENTRY is a normal node.
> >> >> >> 
> >> >> >> The checked on XA_ZERO_ENTRY seems to be more meaningful.
> >> >> >
> >> >> >257-1023 are also reserved, they just aren't used yet.  XA_ZERO_ENTRY
> >> >> >is not guaranteed to be the largest reserved entry.
> >> >> 
> >> >> Then why we choose 4096?
> >> >
> >> >Because 4096 is the smallest page size supported by Linux, so we're
> >> >guaranteed that anything less than 4096 is not a valid pointer.
> >> 
> 
> So you want to say, the 4096 makes sure XArray will not store an address in
> first page? If this is the case, I have two suggestions:
> 
>   * use PAGE_SIZE would be more verbose?

But also incorrect, because it'll be different on different architectures.
It's 4096.  That's all.

>   * a node is an internal entry, do we need to compare with xa_mk_internal()
>     instead?

No.  4096 is better because it's a number which is easily expressible in
many CPU instruction sets.  4094 is much less likely to be an easy number
to encode.

> >(it is slightly out of date; the XArray actually supports storing unaligned
> >pointers now, but that's not relevant to this discussion)
> 
> Ok, maybe this document need to update.

Did you want to send a patch?