Date:   Fri, 7 Apr 2017 13:56:46 +0800
From:   Wu Hao <hao.wu@...el.com>
To:     Alan Tull <atull@...nel.org>
Cc:     Moritz Fischer <moritz.fischer@...us.com>,
        "linux-fpga@...r.kernel.org" <linux-fpga@...r.kernel.org>,
        linux-kernel <linux-kernel@...r.kernel.org>,
        "Kang, Luwei" <luwei.kang@...el.com>,
        "Zhang, Yi Z" <yi.z.zhang@...el.com>,
        "Whisonant, Tim" <tim.whisonant@...el.com>,
        "Luebbers, Enno" <enno.luebbers@...el.com>,
        "Rao, Shiva" <shiva.rao@...el.com>,
        "Rauer, Christopher" <christopher.rauer@...el.com>,
        "Tull, Alan" <alan.tull@...el.com>,
        Xiao Guangrong <guangrong.xiao@...ux.intel.com>
Subject: Re: [PATCH 11/16] fpga: intel: fme: add partial reconfiguration sub
 feature support

On Thu, Apr 06, 2017 at 02:27:42PM -0500, Alan Tull wrote:
> On Thu, Apr 6, 2017 at 5:57 AM, Wu Hao <hao.wu@...el.com> wrote:
> > On Wed, Apr 05, 2017 at 10:39:05AM -0500, Alan Tull wrote:
> >> On Wed, Apr 5, 2017 at 10:26 AM, Alan Tull <atull@...nel.org> wrote:
> >> > On Wed, Apr 5, 2017 at 6:40 AM, Wu, Hao <hao.wu@...el.com> wrote:
> >> >>> >> The  fpga_image_info struct started life as just image specific info,
> >> >>> >> but I want it to go in the direction of including parameters needed to
> >> >>> >> program it this specific time. Otherwise we are stuck having to keep
> >> >>> >> adding parameters as our use of FPGA develops.  It probably could be
> >> >>> >> documented better as 'information needed to program a FPGA image'
> >> >>> >> rather than strictly 'information about this particular FPGA image'.
> >> >>> >> My patch "fpga-mgr: pass parameters for loading fpga in image info"
> >> >>> >> goes in this direction by having the buf, firmware name, or sg list
> >> >>> >> passed in the info for the added fpga_mgr_load() function.  Actually I
> >> >>> >> should probably simplify the API and get rid of fpga_mgr_buf_load,
> >> >>> >> fpga_mgr_buf_load_sg, and fpga_mgr_firmware_load and require people to
> >> >>> >> use fpga_mgr_load (passing all parameters in fpga_image_info).
> >> >>> >>
> >> >>> >
> >> >>> > Make sense.
> >> >>> >
> >> >>> >> > It may be a
> >> >>> >> > little confusing. One rough idea is that keep this info under fpga region
> >> >>> >> > (maybe its private data), and pass the fpga-region to fpga_mgr_buf_load,
> >> >>> >>
> >> >>> >> Yes, keep this info in fpga-region.  When the region wants to program
> >> >>> >> using fpga-mgr, add the region id to fpga_image_info.  I propose
> >> >>> >> calling it region_id.
> >> >>> >
> >> >>> > Hm.. Do we need a function which moves info from region to image info?
> >> >>>
> >> >>> No, just code that sets that variable in the struct before calling the
> >> >>> fpga_region_program_fpga function.
> >> >>>
> >> >>> >
> >> >>> > Another idea is, add a priv to fpga_image_info, and use a common function
> >> >>> > to pass the fpga_region's priv to fpga_image_info's priv before PR.
> >> >>> > fpga-mgr then knows fpga_region priv info from the fpga_image_info.
> >> >>> >
> >> >>>
> >> >>> Adding priv would make the interface for fpga-mgr non-uniform.  The point
> >> >>> of having a fpga-mgr framework is that there
> >> >>> is the potential of the upper layers working for different FPGA devices.
> >> >>> If the interface for each FPGA device were different, that would then
> >> >>> be broken.
> >> >>>
> >> >>
> >> >> I mean drivers can register their own fpga-mgr ops, and handle priv of
> >> >> fpga_image_info in driver specific way for pr (e.g write_init function).
> >> >> We don't need to change the any upper layer interfaces.
> >> >
> >> > I'm trying to avoid driver specific ways of doing things.  Think of this
> >> > all as a set of blocks and we want to be able to switch out individual
> >> > blocks in the future.  It's future-proofing and also making code more
> >> > generally usable.
> >> >
> >> > fpga_mgr_info is part of the interface for calls to fpga-mgr to do
> >> > reprogramming.  My patchset will push it further in that direction
> >> > as pointers to the image are added to fpga_mgr_info.
> >> >
> >> > Adding 'priv' to fpga_mgr_info makes that interface specific to a this driver.
> >> > It's better to add a region_id variable to fpga_mgr_info that may not be used by
> >> > all fpga-mgr drivers.  The current model is that a fpga-mgr driver checks
> >> > fpga_mgr_info flags to see if its correct.  The fpga-mgr driver can check
> >> > any other needed fpga_mgr_info variables and return error if the params
> >> > look invalid.  And ignore any it doesn't need.
> >> >
> >> > I don't think priv belongs in fpga_image_info.  priv tends to be information
> >> > for a specific instance of a driver that can have several instances.  priv
> >> > usually stores a driver's memory mappings, interrupts, etc for that instance.
> >> > It's called private info as it is info that other blocks don't need to know
> >> > and don't get to look at.  It's private.  So priv as in interface strikes me as
> >> > not private any more and is sort of a red flag.
> >
> > Thanks a lot for the details. It's more clear to me now. :)
> >
> >> >
> >> > Alan
> >>
> >> Besides that, I see that region_id is needed both for the fme_pr and
> >> for determining which port to disable/enable.  So adding it to the
> >> fpga_image_info would allow the fpga-region to figure out which
> >> bridge to control as well as pass it to the mgr.
> >>
> >
> > Do we need to add an 'id' to fpga-bridge too?
> 
> You don't have to, but you could.  See below.
> 
> >
> > Per my understanding this FME driver should create fpga-region for each
> > accelerator, and make sure each fpga-region has correct region_id. When
> > this fpga-region wants to be programmed by fpga-mgr, then it add this
> > info to fpga_image_info, and fpga-mgr knows which region to program via
> > this fpga_image_info.
> 
> The fpga_image_info is passed both to fpga-mgr and fpga-bridge.  So
> both will get the region_id through the info.
> 
> >
> > For each fpga-region, FME driver needs to create one fpga-bridge and link
> > it to region's bridge_list. When fpga_region_program_fpga is invoked for
> > PR. This fpga-bridge could be disabled before PR and re-enabled after PR
> > automatically. If fpga-bridge contains this 'id' information, then driver
> > knows which port to enable/disable to implement of the enable_set function
> > under this fpga-bridge.
> 
> Could implement it either way: the bridge could either use the
> bridge_id that has been passed to it in the info or code that creates
> the bridge could create private data and pass it to
> fpga_bridge_register().  But the bridge will have the id because it
> will be passed in the info, so it's not really needed.

Thanks for the suggestion. :)

The reason I was considering adding id to fpga_bridge is, in the case
driver creates all bridges (and regions/manager) in the initialization
code, but at that time, nobody knows the actual fpga_image_info.

In order to pass fpga_image_info to the bridges of give region during PR
(e.g in fpga_region_program_fpga function), then driver must implement
region->get_bridges function, but I feel we may not need get_bridges
function for this case, as all bridges have already been linked to
fpga-region during initialization.

So I prefer the second method that keep this id in the private data. :)

Thanks
Hao

> 
> Alan
> 
> >
> > Thanks
> > Hao
> >
> >> >
> >> >>
> >> >> If you prefer the region_id for fpga_image_info, we can go with region_id
> >> >> for sure. : )
> >> >>
> >> >> Thanks
> >> >> Hao
> --
> To unsubscribe from this list: send the line "unsubscribe linux-fpga" in
> the body of a message to majordomo@...r.kernel.org
> More majordomo info at  http://vger.kernel.org/majordomo-info.html

Powered by blists - more mailing lists