Re: [Xen-devel] [PATCH] tools/libxc: send page-in requests in batches in linux_privcmd_map_foreign_bulk

["Andres Lagar-Cavilla" <andres@xxxxxxxxxxxxxxxx>]

> On Thu, Jan 26, Andres Lagar-Cavilla wrote:
>
>> Olaf,
>> if I get a pattern of errors in the err array of the form:
>>
>> {ENOENT, EWHATEVER, ENOENT, EWHATEVER}
>> with EWHATEVER != ENOENT
>>
>> this patch does not optimize anything.
>
> It will poke all gfns instead of a single one before going to sleep.

All gfns that are contiguous in the arr array and have all failed with
ENOENT.

>
>> Why not alloc a separate arr2 array, move the gfn's that failed with
>> ENOENT there, and retry the whole new arr2 block? You only need to
>> allocate arr2 once, to size num. That will batch always, everything.
>
> The errors for gfns are most likely fragmented. Merging the gfns means
> they will get a different addr. So the overall layout of arr and err
> needs to remain the same.

Ummh yeah, addr is the problem. But as you say, the error for the gfns are
fragmented, so that seriously limits your batching abilities.

You could pass the whole arr sub-segment encompassing the first and last
gfn that failed with ENOENT. Successful maps within that array will be
re-done by the hypervisor, at no correctness cost. I would imagine that
the extra work is offset by the gains, but that remains to be seen.

Andres
>
> Olaf
>



_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxxxxxxxx
http://lists.xensource.com/xen-devel