On Thu, 2021-10-28 at 12:39 +0100, Mauro Carvalho Chehab wrote:
> Em Thu, 28 Oct 2021 13:12:45 +0900
> Tsuchiya Yuto <kitakar@xxxxxxxxx> escreveu:
>
> > <Adding back people/list to Cc>
> >
> > On Tue, 2021-10-26 at 09:26 +0100, Mauro Carvalho Chehab wrote:
> > > Em Mon, 18 Oct 2021 01:19:44 +0900
> > > Tsuchiya Yuto <kitakar@xxxxxxxxx> escreveu:
> > >
> > > > Currently, the `port >= N_CSI_PORTS || err` checks for ISP2400 are always
> > > > evaluated as true because the err variable is set to `-EINVAL` on
> > > > declaration but the variable is never used until the evaluation.
> > > >
> > > > Looking at the diff of commit 3c0538fbad9f ("media: atomisp: get rid of
> > > > most checks for ISP2401 version"), the `port >= N_CSI_PORTS` check is
> > > > for ISP2400 and the err variable check is for ISP2401. Fix this issue
> > > > by adding ISP version test there accordingly.
> > > >
> > > > Yes, there are other better ways to fix this issue, like adding support
> > > > for ISP2400 to ia_css_mipi_is_source_port_valid(). In this way, we can
> > > > unify the following test:
> > > >
> > > > if (!IS_ISP2401)
> > > > port = (unsigned int)pipe->stream->config.source.port.port;
> > > > else
> > > > err = ia_css_mipi_is_source_port_valid(pipe, &port);
> > > >
> > > > However, the IS_ISP2401 test here (formerly `ifdef ISP2401`) is not
> > > > a result of real hardware difference, but just a result of the following
> > > > two different versions of driver merged by tools [1]:
> > > >
> > > > - ISP2400: irci_stable_candrpv_0415_20150521_0458
> > > > - ISP2401: irci_ecr-master_20150911_0724
> > >
> > > No.
> > >
> > > While I don't have any internal information from the hardware manufacturer,
> > > I guess you misinterpreted things here. 2400 and 2401 are different
> > > hardware versions. See atomisp_pci_probe() logic.
> > >
> > > Basically, Cherrytail and Anniedale comes with 2401. Older Atom CPUs
> > > (Merrifield and Baytrail) comes with 2400.
> >
> > Yes, indeed, 2400 and 2401 are different hardware. When they (I mean who
> > originally wrote atomisp driver non-upstream) needed to distinguish
> > between ISP2400 and ISP2401, they used the ifdefs like the following:
> >
> > - USE_INPUT_SYSTEM_VERSION_2 (for both ISP2400/ISP2401)
> > - USE_INPUT_SYSTEM_VERSION_2401 (for ISP2401)
> > ...
> >
> > I think this is a sign that the atomisp driver supports both
> > ISP2400/ISP2401 in a single version.
>
> Actually, supporting both on a single version is part of Alan's work.
>
> It seems he used the generation tool to produce a version for 2400, and
> then re-used it to generate for 2401. It then used some scripting tool
> to convert the differences on #ifdef ISP2401. See:
>
> a49d25364dfb ("staging/atomisp: Add support for the Intel IPU v2")
>
> There are things there like:
>
> +#ifdef ISP2401
> +
> +#endif
>
> I did a large cleanup work to get rid of those ifdefs, replacing them
> by runtime logic.
>
> The end goal is to have a single compile-time driver that works for
> both 2400 and 2401.
>
> This is not possible yet, as there are some registers that are mapped
> on different addresses, depending on the hardware version, and making
> it generic requires a lot of work and tests. So, for now, we need to
> have a compile-time option to select between both.
>
> > Indeed, the upstreamed atomisp uses irci_stable_candrpv_0415_20150521_0458
> > for ISP2400 and IIUC it was working on Bay Trail. On the other hand,
> > intel-aero is a kernel for Cherry Trail and uses the same version
> > irci_stable_candrpv_0415_20150521_0458.
> >
> > So, both ISP version ISP2400/ISP2401 can be supported by a single
> > driver version.
>
> I See. OK!
>
> > > > We should eventually remove (not unify) such tests caused by just a
> > > > driver version difference and use just one version of driver. So, for
> > > > now, let's avoid further unification.
> > > >
> > > > [1] The function ia_css_mipi_is_source_port_valid() and its usage is
> > > > added on updating css version to irci_master_20150701_0213
> > > > https://raw.githubusercontent.com/intel/ProductionKernelQuilts/cht-m1stable-2016_ww31/uefi/cht-m1stable/patches/cam-0439-atomisp2-css2401-and-2401_legacy-irci_master_2015070.patch
> > > > ("atomisp2: css2401 and 2401_legacy-irci_master_20150701_0213")
> > >
> > > What happens is that there is a 2401 and a 2401 "legacy". It sounds
> > > that this due to some different software stacks that are reflected both
> > > at the firmware and at the driver.
> >
> > Yeah, I'm not sure what the "legacy" is. It might be a reference of
> > `ISP2401_NEW_INPUT_SYSTEM` (css_2401_csi2p_system) and
> > non-`ISP2401_NEW_INPUT_SYSTEM` (css_2401_system).
> >
> > > -
> > >
> > > On other words, this patch requires some rework, as otherwise it will break
> > > support for Baytrail.
> >
> > You mean "this patch"? then, I intended this patch is rather a fix for
> > ISP2400 case! The err variable for ISP2400 case is always true because
> > it is not used before the error check:
> >
> > int
> > allocate_mipi_frames(struct ia_css_pipe *pipe,
> > struct ia_css_stream_info *info)
> > {
> > int err = -EINVAL;
> > [...]
> > if (!IS_ISP2401)
> > port = (unsigned int)pipe->stream->config.source.port.port;
> > else
> > err = ia_css_mipi_is_source_port_valid(pipe, &port);
> >
> > assert(port < N_CSI_PORTS);
> >
> > if (port >= N_CSI_PORTS || err) {
> > ia_css_debug_dtrace(IA_CSS_DEBUG_TRACE_PRIVATE,
> > "allocate_mipi_frames(%p) exit: error: port is not correct (port=%d).\n",
> > pipe, port);
> > return -EINVAL;
> > }
> >
> > The first usage of err variable is ia_css_mipi_is_source_port_valid()
> > for IS_ISP2401 case, but it's not used for ISP2400 case. This causes
> > the evaluation `port >= N_CSI_PORTS || err` always true for ISP2400 case,
> > meaning it will be always treated as a error.
>
> Ok. Had you test the driver with Baytrail?
Unfortunately, no. I don't have a Baytrail device to test (yet?). I
noticed this issue anyway when I tried removing `#ifdef ISP2401`.
This is not directly related to this series, but how we should reduce
the ifdef usage in the future? Here are my two ideas:
1. (my initial idea) remove `#ifdef ISP2401` part and make ISP2401
part completely irci_stable_candrpv_0415_20150521_0458
this way does not require (relatively) much human work I think.
But as Mauro says, the `#ifdef ISP2401` part (irci_ecr-master_20150911_0724)
is basically an improved version. So, we may also:
2. continue unifying `#ifdef ISP2401` and `#ifndef ISP2401` parts
but this way needs more human work I think though. And if we go this
way, I also need to rewrite this patch as mentioned in the commit
message.
> >
> > > Also, patch 13 should be dropped, as the firmware versions for 2400 are
> > > different
> >
> > The firmware version for 2400 on the upstreamed atomisp is
> > irci_stable_candrpv_0415_20150521_0458 :-)
> >
> > static const char *isp2400_release_version = STR(irci_stable_candrpv_0415_20150521_0458);
> > static const char *isp2401_release_version = STR(irci_ecr - master_20150911_0724);
> >
> > The intention of that patch is rather, it clarifies ISP2401 is now using
> > the same driver (css) version as ISP2400.
>
> Ok.
>
> > > - and maybe patches 8 to 12 may need more work in order to not
> > > touch 2400.
> >
> > Those patches do not break ISP2400, because what they do for ISP2400
> > is that, they remove members from `struct`s which were initially inside
> > of `ifdef ISP2401`. And because these removed members were initially
> > inside of the ifdefs, the usage was also inside the ifdefs.
>
> Did you test on Baytrail (ISP2400), and with the compile-time option
> enabled/disabled?
Sorry, I should have clarified on the cover later. For ISP2400, I did
compile test only (CONFIG_VIDEO_ATOMISP_ISP2401 enabled/disabled).
> Regards,
> Mauro
