On 15/11/2022 17:38, Dave Stevenson wrote:
Hi Dmitry On Tue, 15 Nov 2022 at 14:21, Dmitry Baryshkov <dmitry.baryshkov@xxxxxxxxxx> wrote:On 15/11/2022 17:14, Dave Stevenson wrote:Hi Dmitry On Sun, 13 Nov 2022 at 13:06, Dmitry Baryshkov <dmitry.baryshkov@xxxxxxxxxx> wrote:Hi Dave, On 19/07/2022 16:45, Dave Stevenson wrote:Hi Sam On Mon, 18 Jul 2022 at 21:52, Sam Ravnborg <sam@xxxxxxxxxxxx> wrote:Hi Dave, a long overdue reply on this series. On Fri, Mar 04, 2022 at 03:17:55PM +0000, Dave Stevenson wrote:Hi All Changes from v1: - New patch to refactor drm_bridge_chain_post_disable and drm_bridge_chain_pre_enable to reuse drm_atomic_bridge_chain_post_disable / drm_atomic_bridge_chain_pre_enable but with a NULL state. - New patch that adds a pre_enable_upstream_first to drm_panel. - changed from an OPS flag to a bool "pre_enable_upstream_first" in drm_bridge. - Followed Andrzej's suggestion of using continue in the main loop to avoid needing 2 additional loops (one forward to find the last bridge wanting upstream first, and the second backwards again). - Actioned Laurent's review comments on docs patch. Original cover letter: Hopefully I've cc'ed all those that have bashed this problem around previously, or are otherwise linked to DRM bridges. There have been numerous discussions around how DSI support is currently broken as it doesn't support initialising the PHY to LP-11 and potentially the clock lane to HS prior to configuring the DSI peripheral. There is no op where the interface is initialised but HS video isn't also being sent. Currently you have: - peripheral pre_enable (host not initialised yet) - host pre_enable - encoder enable - host enable - peripheral enable (video already running) vc4 and exynos currently implement the DSI host as an encoder, and split the bridge_chain. This fails if you want to switch to being a bridge and/or use atomic calls as the state of all the elements split off are not added by drm_atomic_add_encoder_bridges.A typically chain looks like this: CRTC => Encoder => Bridge A => Bridge B We have in DRM bridges established what is the "next" bridge - indicated with the direction of the arrows in the drawing. This set of patches introduces the concept of "upstream" bridges. pre_enable_prev_bridge_first would be easier to understand as it uses the current terminology. I get that "upstream" is used in the DSI specification - but we are dealing with bridges that happens to support DSI and more, and mixing the two terminologies is not good. Note: Upstream is also used in a bridge doc section - here it should most likely be updated too.Sure, I have no issues with switching to prev/next from upstream/downstream. To the outsider it can be confusing - in pre_enable and disable, the next bridge to be called is the previous one. At least it is documented.The current approach set a flag that magically makes the core do something else. Have you considered a much more explicit approach? A few helpers like: drm_bridge_pre_enable_prev_bridge() drm_bridge_enable_prev_bridge() drm_bridge_disable_prev_bridge() drm_bridge_post_disable_prev_bridge()No point in drm_bridge_enable_prev_bridge() and drm_bridge_post_disable_prev_bridge() as the call order down the chain will mean that they have already been called. drm_bridge_enable_next_bridge() and drm_bridge_post_disable_next_bridge() possibly.And then update the core so the relevant function is only called once for a bridge. Then the need for DSI lanes in LP-11 can be archived by a call to drm_bridge_pre_enable_prev_bridge()Unfortunately it gets ugly with post_disable. The DSI host controller post_disable will have been called before the DSI peripheral's post_disable, and there are conditions where the peripheral needs to send DSI commands in post_disable (eg panel-asus-z00t-tm5p5-n35596 [1]). Changing all DSI hosts to call drm_bridge_post_disable_next_bridge feels like the wrong thing to do. There are currently hacks in dw-mipi-dsi that do call the next panel/bridge post_disable [2] and it would be nice to get rid of them. Currently the calls aren't tracked for state, so you end up with post_disable being called twice, and panels having to track state (eg jdi_lt070me050000 [3]). [1] tm5p5_nt35596_unprepare() calls tm5p5_nt35596_off() https://elixir.bootlin.com/linux/latest/source/drivers/gpu/drm/panel/panel-asus-z00t-tm5p5-n35596.c#L107 [2] https://elixir.bootlin.com/linux/latest/source/drivers/gpu/drm/bridge/synopsys/dw-mipi-dsi.c#L889 [3] https://elixir.bootlin.com/linux/latest/source/drivers/gpu/drm/panel/panel-jdi-lt070me05000.c#L44This is more explicit than a flag that triggers some magic behaviour. It may even see uses we have not realised yet.Personally it feels like more boilerplate in almost all DSI drivers, and generally I see a push to remove boilerplate.It is late here - so maybe the above is not a good idea tomorrow - but right now I like the simplicity of it. Other than the above I read that a mipi_dsi_host_init() is planned, which is also explicit and simple - good.It's been raised, but the justification for most use cases hasn't been made. The Exynos conversion looks to be doing the wrong thing in checking state, and that's why it is currently needing it. Again it's also more boilerplate. TC358767 is an odd one as it wants the DSI interface enabled very early in order to have a clock for the DP aux channel well before video is running. I had a thought on that, but It looks like I haven't hit send on a reply to Lucas on that one - too many distractions.Have we seen a new revision of some of these? Chances are high that I have missed it then.No, still on V2. Other than Dmitry's comment over updating parade-ps8640 and dropping drm_bridge_chain_*, no real comments had been made.It's been a while now. Do you still plan to pursue this patchset?If there was anything that could actually be worked on, then I'm happy to respin it, but if the approach is generally being rejected then I don't want to waste the effort. I'm not totally clear who the maintainers are that the final arbiters and need to sign off on this. drm_bridge.c falls to Maarten, Maxime, and Thomas for "DRM DRIVERS AND MISC GPU PATCHES" drm_panel.c falls to Thierry and Sam for "DRM PANEL DRIVERS", and then Maarten, Maxime, and Thomas. Only Sam has responded publicly. I have had discussions with Maxime, but it's not directly his area of knowledge. Looking at the patch series: Patch 1: Your comment "update parade-ps8640 to use drm_atomic_bridge_chain_". It looks like patchset [1] by Sam does this, but the patchset went wrong and is missing patches 8-11 and therefore hasn't been merged. Patch 2: Comment from Jagan that it's like an old patch. It has similarities, but isn't the same. Patch 3: R-b by you (thank you), but concerns from Jagan which I still don't understand. Without clarification on the issue and whether my suggested alternative place for the hook solves the issue, IMHO it's not worth respinning. Patch 4: R-b Laurent. This cover note got totally subverted with Exynos issues. Sam did request use of prev / next instead of upstream / downstream, which can be done and perhaps warrants a respin now.[personal notice: I'd prefer something less strange, e.g. an explicit calls to mipi_dsi_host, but as this patchset seems to fix the issues, I'm fine with it].That can fix the power up sequence, but how do you propose telling the DSI controller NOT to power down in post_disable before the DSI peripheral post_disable has potentially sent DCS commands - i.e. the case you were discussing on Friday in [2].I thought that the same 'call the parent beforehand' switch applied to the deinit paths, didn't it?My proposed flag does indeed swap the order of post_disable as well. Perhaps I was misunderstanding your personal preference. I was taking it as an explicit mipi_dsi_host call to initialise the DSI link, which then also needs an explicit mipi_dsi_host call to tear it down as well. In that situation there is a need to rework the bridge chain post_disable to allow for the panel post_disable to send DCS commands before the DSI host is disabled.If a panel/bridge driver doesn't call mipi_dsi_host_init then the expectation must be that it will be called by the DSI controller's pre_enable, and deinit from post_disable. Likewise init & deinit would be called if host_transfer is used when the host isn't initialised. If the panel/bridge driver explicitly calls mipi_dsi_host_init, then does that mandate that it must also call mipi_dsi_host_deinint. The controller post_disable is then effectively a no-op. This can be covered in documentation, but also leaves the potential for strange behaviour if the requirement is not followed, and I can't think of a nice place to drop a WARN to flag the issue in the driver. TBH The lack of interest in solving the issues almost makes me want to just document the total brokenness of it and throw in the towel. Seeing as we as Raspberry Pi run a vendor kernel, we can run with downstream patches until those who care finally make a decision for mainline. I'd prefer to solve it properly, but it requires some engagement from the community.I see. I can probably try spinning a patchset doing explicit mipi_dsi calls. Let's see if it gains more attention.Is it better to have 2 competing patchsets floating around, or try and get responses on one first? I'll try and get an updated set out today.
I'm a bit in a tough position here. I can't say that I like this approach, but it seems to fix all the issues that we have with DSI hosts, so it's better than the current state.
Whilst the main use case now is DSI devices that want the bus powered up, is it only restricted to DSI or are there potentially other links that have a similar requirement?
eDP has even more strange requirements: several panels have strict requirement to power it off completely before being able to power it up again (see all the patches from Douglas Anderson to get AUX bus to work properly). And powering up a bus requires waiting for the HPD pin.
In the eDP case powering up the bus was solved using autosuspend on the panel and on the host sides. However I don't remember whether eDP has anything comparable to DSI's LP state
It seems something is broken with respect to reviewing of core drm patches touching strange areas. My patchset improving drm_bridge_connector HPD also didn't gain a lot of responses.I do appreciate that folks are busy, but there does seem to be a tendency of "it might not be the optimum solution, however I haven't time to think it through so let's just ignore the problem". Particularly when it is purely an in-kernel API with no impact on userspace, I do find that quite frustrating.
Fully agree.
DaveI'll do a respin now second guessing Jagan's comment, and then give it a month before giving up Cheers Dave [1] https://patchwork.freedesktop.org/series/106422/#rev5 [2] https://lists.freedesktop.org/archives/dri-devel/2022-November/379693.html
-- With best wishes Dmitry
