[ofa-general][PATCH] mlx4_core: Multi Protocol support

[Yevgeny Petrilin]

Thank you for the comrehensive review.
I will split the patches by topics and send them separately,
along with the fixes to other remarks you made

Yevgeny


-----Original Message-----
From: Roland Dreier [mailto:rdreier at cisco.com] 
Sent: Wednesday, April 16, 2008 6:34 PM
To: Yevgeny Petrilin
Cc: general at lists.openfabrics.org
Subject: Re: [ofa-general][PATCH] mlx4_core: Multi Protocol support

Your email has

 > Content-Type: text/plain; charset=ISO-8859-1; format=flowed

and the format=flowed means that the patch gets corrupted and won't
apply.  So when you resend, please fix.

I don't think we can really apply this as one patch -- it does too many
things at once and needs to be split up... I think pretty much each of
these items is independent and could be a separate patch:

 >         1.  Mlx4 device now holds the actual protocol for each port.
 >             The port types are determined through module parameters
of through sysfs
 >             interface. The requested types are verified with firmware
capabilities
 >             in order to determine the actual port protocol.
 >         2.  The driver now manages Mac and Vlan tables used by
customers of the low
 >             level driver. Corresponding commands were added.
 >         3.  Completion eq's are created per cpu. Created cq's are
attached to an eq by
 >             "Round Robin" algorithm, unless a specific eq was
requested.
 >         4.  Creation of a collapsed cq support was added.
 >         5.  Additional reserved qp ranges were added. There is a
range for the customers
 >             of the low level driver (IB, Ethernet, FCoE).
 >         6.  Qp allocation process changed.
 >             First a qp range should be reserved, then qps can be
allocated from that
 >             range. This is to support the ability to allocate
consecutive qps.
 >             Appropriate changes were made in the allocation
mechanism.
 >         7.  Common actions to all HW resource management (Doorbell
allocation,
 >             Buffer allocation, Mtt write) were moved to the low level
driver.

Also, on the other hand, the current two patches are too split up: if I
apply this patch then mlx4_ib won't compile until the second patch goes
in too.  Which means someone trying to bisect an mlx4 bug gets into
trouble.  So please make sure that everything still compiles and works
after each patch is applied.

By the way, the multiple EQ stuff is a pretty major change in
behavior... are we really ready for this?  Round robin seems like it
could easily lead to worst-case behavior for some plausible workloads.

Finally, checkpatch.pl shows a few minor whitespace problems... please
fix when you resend.

 - R.