[openib-general] Re: [PATCH] [RFC] use device max_map_per_fmr in fmr pool

[Or Gerlitz]

One issue i brought here is that the max_map_per_fmr field is not 
implemented by mthca query device, this better be fixed, also the field
name should be changed to max_remaps_per_fmr which is much clearer.

Michael S. Tsirkin wrote:
> Does this actually help performance?
> How about increasing IB_FMR_MAX_REMAPS from 32 to say 128 and checking?

I am not on working on performance optimizations now, the code we wrote 
to the GEN1 driver was using the max possible remaps and i'd like to 
work here this way as well.

> Allocating resources up to a maximum supported by a device is rarely optimal.
> Its easy to imagine a device where FMR number of maps is a shared resource,
> so allocating one FMR with max_map_per_fmr will not let you create any more
> of these.

Why imagine something related to Mellanox proprietary implementation 
over which the number of times you remap an FMR does not consume --any-- 
further resources on the device other (at least for tavor and arbel, 
correct if i am wrong to other devices).