Re: [PATCH 04/17] media: atomisp: pci: do not use err var when checking port validity for ISP2400

From: Mauro Carvalho Chehab
Date: Mon Nov 01 2021 - 10:11:16 EST


Em Mon, 01 Nov 2021 22:38:55 +0900
Tsuchiya Yuto <kitakar@xxxxxxxxx> escreveu:

> 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.

No. What I said is that the if (ISP2401) and the remaining ifdefs are because
of BYT x CHT. The worse part of them are related to those files
(See Makefile):

obj-byt = \
pci/css_2400_system/hive/ia_css_isp_configs.o \
pci/css_2400_system/hive/ia_css_isp_params.o \
pci/css_2400_system/hive/ia_css_isp_states.o \

obj-cht = \
pci/css_2401_system/hive/ia_css_isp_configs.o \
pci/css_2401_system/hive/ia_css_isp_params.o \
pci/css_2401_system/hive/ia_css_isp_states.o \
pci/css_2401_system/host/csi_rx.o \
pci/css_2401_system/host/ibuf_ctrl.o \
pci/css_2401_system/host/isys_dma.o \
pci/css_2401_system/host/isys_irq.o \
pci/css_2401_system/host/isys_stream2mmio.o

Those define regmaps for 2400 and 2401. I was able to remove a lot
of things from the old css_2400/css_2401 directories, but the ones
there at pci/*/css*/ia_css_isp_*.c are a lot more complex, and would
require some mapping functions to allow the same driver to work with
both BYT and CHT.

The better would be to test the driver first at BYT, fix issues (if any) and
then write the mapping code.

> 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).

Maybe Hans could help us on that. I guess he has an Asus T100 device,
which is BYT based.

Hans, if you're willing to do the tests, you could either use nvt
or v4l2grab to test it.

It seems that BYT has an additional issue, though: the ISP seems to
require 12 non-visible lines/columns (in addition to 16 ones required
by CHT?) for it to work.

So, you may need to tweak the resolution a bit, as otherwise the
driver will return an -EINVAL.

See:

https://git.linuxtv.org/media_stage.git/commit/?id=dcbc4f570495dbd490975979471687cbe2246f99

For the workaround I had to add for CHT to properly report the
visible resolution.

Regards,
Mauro

Regards,
Mauro