On Wed, May 15, 2019 at 09:43:24AM +0200, Paul Kocialkowski wrote: > Hi, > > On Tue, 2019-05-14 at 16:28 +0200, Daniel Vetter wrote: > > On Tue, May 14, 2019 at 4:13 PM Paul Kocialkowski > > <paul.kocialkowski@xxxxxxxxxxx> wrote: > > > Hi, > > > > > > On Tue, 2019-05-14 at 13:09 +0200, Daniel Vetter wrote: > > > > On Mon, May 13, 2019 at 7:14 PM Paul Kocialkowski > > > > <paul.kocialkowski@xxxxxxxxxxx> wrote: > > > > > Hey, > > > > > > > > > > Le lundi 13 mai 2019 à 11:34 +0200, Daniel Vetter a écrit : > > > > > > On Mon, May 13, 2019 at 11:02 AM Paul Kocialkowski > > > > > > <paul.kocialkowski@xxxxxxxxxxx> wrote: > > > > > > > Hi, > > > > > > > > > > > > > > On Fri, 2019-05-10 at 16:54 +0200, Daniel Vetter wrote: > > > > > > > > On Fri, May 10, 2019 at 2:12 PM Paul Kocialkowski > > > > > > > > <paul.kocialkowski@xxxxxxxxxxx> wrote: > > > > > > > > > Hi, > > > > > > > > > > > > > > > > > > On Tue, 2019-05-07 at 21:57 +0530, Ramalingam C wrote: > > > > > > > > > > DRM API for generating uevent for a status changes of connector's > > > > > > > > > > property. > > > > > > > > > > > > > > > > > > > > This uevent will have following details related to the status change: > > > > > > > > > > > > > > > > > > > > HOTPLUG=1, CONNECTOR=<connector_id> and PROPERTY=<property_id> > > > > > > > > > > > > > > > > > > > > Need ACK from this uevent from userspace consumer. > > > > > > > > > > > > > > > > > > So we just had some discussions over on IRC and at about the hotplug > > > > > > > > > issue and came up with similar ideas: > > > > > > > > > https://lists.freedesktop.org/archives/dri-devel/2019-May/217408.html > > > > > > > > > > > > > > > > > > The conclusions of these discussions so far would be to have a more or > > > > > > > > > less fine grain of uevent reporting depending on what happened. The > > > > > > > > > point is that we need to cover different cases: > > > > > > > > > - one or more properties changed; > > > > > > > > > - the connector status changed; > > > > > > > > > - something else about the connector changed (e.g. EDID/modes) > > > > > > > > > > > > > > > > > > For the first case, we can send out: > > > > > > > > > HOTPLUG=1 > > > > > > > > > CONNECTOR=<id> > > > > > > > > > PROPERTY=<id> > > > > > > > > > > > > > > > > > > and no reprobe is required. > > > > > > > > > > > > > > > > > > For the second one, something like: > > > > > > > > > HOTPLUG=1 > > > > > > > > > CONNECTOR=<id> > > > > > > > > > STATUS=Connected/Disconnected > > > > > > > > > > > > > > > > > > and a connector probe is needed for connected, but not for > > > > > > > > > disconnected; > > > > > > > > > > > > > > > > > > For the third one, we can only indicate the connector: > > > > > > > > > HOTPLUG=1 > > > > > > > > > CONNECTOR=<id> > > > > > > > > > > > > > > > > > > and a reprobe of the connector is always needed > > > > > > > > > > > > > > > > There's no material difference between this one and the previous one. > > > > > > > > Plus there's no beenfit in supplying the actual value of the property, > > > > > > > > i.e. we can reuse the same PROPERTY=<id-of-status-property> trick. > > > > > > > > > > > > > > That's the idea, but we need to handle status changes differently than > > > > > > > properties, since as far as I know, connected/unconnected status is not > > > > > > > exposed as a prop for the connector. > > > > > > > > > > > > Oops, totally missed that. "Everything is a property" is kinda > > > > > > new-ish, at least compared to kms. Kinda tempted to just make status > > > > > > into a property. Or another excuse why we should expose the epoch > > > > > > property :-) > > > > > > > > > > Well I think it would make sense anyway, as long as we can make sure it > > > > > stays consistent with the one reported in the connector struct. > > > > > > > > > > > > > Here's why: > > > > > > > > - A side effect of forcing a probe on a connector is that you get to > > > > > > > > read all the properties, so supplying them is kinda pointless. > > > > > > > > > > > > > > Agreed, except for the status case where it's useful to know it's a > > > > > > > disconnect, because we don't need any probe step in that case. > > > > > > > > > > > > > > > - You can read STATUS without forcing a reprobe, if you want to avoid > > > > > > > > the reprobe for disconnected. I'd kinda not recommend that though, > > > > > > > > feels a bit like overoptimizing. And for reasonable connectors (i.e. > > > > > > > > dp) reprobing a disconnected output is fast. HDMI is ... less > > > > > > > > reasonable unfortunately, but oh well. > > > > > > > > > > > > > > How would that be retreived then? From the looks of it, that's a > > > > > > > MODE_GETCONNECTOR ioctl and I was under the impression this is what > > > > > > > does the full reprobe. > > > > > > > > > > > > drmGetConnector vs drmGetConnectorCurrent. > > > > > > > > > > Ah right, forgot about that one, thanks. > > > > > > > > > > > > Not sure what issues could arise in case of disconnect without reprobe > > > > > > > -- at least I don't see why userspace should have to do anything in > > > > > > > particular except no longer using the connector, even in complex DP MST > > > > > > > cases. > > > > > > > > > > > > connector->status might be a lie without a full reprobe, and wrongly > > > > > > indicate that the connector is disconnected while there's still > > > > > > something plugged in. I'm not sure we've fixed those bugs by now > > > > > > (usually it's around "hpd indicates disconnected" vs. "i2c indicates > > > > > > connected, and we can't break this because every intel platform ever > > > > > > shipped has a few devices where this is somehow broken, irrespective > > > > > > of the sink). > > > > > > > > > > Mhh either way, I think it's up to the driver to report that and make > > > > > it consistent. I think we have poll helpers to make up for cases where > > > > > hotplug is not available too. So I'm not sure why a full reprobe would > > > > > be needed: drivers just need to do the right thing. > > > > > > > > > > > > > - There's no way to only reprobe status, you can only ever reprobe > > > > > > > > everything with the current ioctl and implementations. Having an > > > > > > > > option to reprobe only parts of it doesn't seem useful to me (we need > > > > > > > > to read the EDID anyway, and that's the expensive part of reprobing in > > > > > > > > almost all cases). > > > > > > > > > > > > > > Agreed. > > > > > > > > > > > > > > > In a way PROPERTY=<status-prop-id> simply tells userspace that it > > > > > > > > needs to reprobe this connector. > > > > > > > > > > > > > > I thought we could access the props alone, which avoids doing a reprobe > > > > > > > when the kernel knows that only a prop or a set of props changed and do > > > > > > > not require a full reprobe. That's the first case I was mentionning. > > > > > > > > > > > > > > > At that point we need to figure out whether this is a good uapi or > > > > > > > > not, and that's where the epoch comes in. There's two reasons for an > > > > > > > > epoch: > > > > > > > > - We need it internally because I'm not goinig to wire a new return > > > > > > > > value through hundreds of connector probe functions. It's much easier > > > > > > > > to have an epoch counter which we set from e.g. drm_set_edid and > > > > > > > > similar functions that update probe state. > > > > > > > > > > > > > > I don't think I'm following what issue this is trying to solve > > > > > > > internally. > > > > > > > > > > > > So I'm assuming that if we handle a hotplug, we only want to generate > > > > > > one uevent for that, not one for every little thing that changed. > > > > > > There's two ways to implement this logic: > > > > > > - With some epoch counter and a helper function you can call everytime > > > > > > something changed (e.g. status, or edid, or anything else we care > > > > > > about e.g. from dp aux). This won't need much (if any) driver changes, > > > > > > because we can just put these into the relevant helper/core functions > > > > > > (like edid update, or dp aux reading or whatever). > > > > > > - Wiring a new return value through the entire stack (and _all_ the > > > > > > kms drivers) so that the probe helpers could aggregate this like they > > > > > > currently do. > > > > > > > > > > > > One of these is a lot less typing. > > > > > > > > > > Oh I had missed this issue. Yeah of course if we start reporting > > > > > property changes, a hotplug will be lots of such changes. > > > > > > > > > > So an epoch counter property would indeed also solve the reprobing > > > > > problem. But I think it would be nice to keep the ability to be > > > > > notified of what changed precisely via uevent. I'm not really buying > > > > > the "missing uevent" thing so much and I think we can reasonably expect > > > > > that it will be useful. Events could be aggregated (which the epoch > > > > > counter would probably also allow) and sent out altogether when the > > > > > connector status changes (along with the status information). I think > > > > > we're under-using uevent currently, and feel like this should be fixed > > > > > regardless of the full reprobe issue. > > > > > > > > I see two options: > > > > > > > > 1) kernel aggregates uevents a bit, and sends out one (per connector > > > > that changed) indicating that sink related state changed. Userspace > > > > listens to that, and does a drmGetConnector as a result to update > > > > itself. Most of this code needs to exist anyway. > > > > > > > > 2) kernel sends out updates as we go, userspace reassembles everything > > > > again (and not really an idea when all the updates are in, see Lyude's > > > > question. Plus userspace still needs to have the drmGetConnector path, > > > > at least for initial setup. > > > > > > > > I don't see why 2) is any better than 1), and it has some clear downsides. > > > > > > My main point is that when the kernel knows some individual property > > > changed, it should update its internal state and report what prop > > > changed *in cases that don't require a full reprobe* and report only > > > which connector changed when a connector reprobe is needed anyway. > > > > > > Maybe we could even have quirks for drivers to indicate that common > > > actions that shouldn't require a reprobe actually do for this driver in > > > particular, because of whatever instability we want to be confident > > > about handling. I believe this should be abstracted away in DRM and not > > > reflected on the userspace API. > > > > > > So it's not just about providing per-property updates, the updates > > > reported this way have to be "standalone" and never imply that anything > > > else about the connector's state could have change. Driver's guarantee. > > > > I'm not sure what that guarantee buys us. You can never assume that if > > you get this event that only that property changed, since other stuff > > might have changed meanwhile. > > Then you get a second event notification and act accordingly. And the > kernel will group individual property change events into logical events > that only get reported once, too, so it will be rare in practice. I > don't see what issue we have at hand. > > > And especially for a sink change we know that a lot of properties > > (including stuff like connector->status here that's not yet a > > property) will have changed. And event with e.g. PROPERTY=epoch would > > simply indicate that all the sink related things have been updated, > > and you can get them with drmGetConnectorCurrent. > > That's what I had in mind with reporting CONNECTOR= without any > PROPERTY. Reporting that epoch change feels weird to me, epoch is more > of a meta-information about the connector that userspace should > probably query on its own. > > > We also need some way to indicate that an entire batch of updates have > > completed if we send out events for every part of this. Sending out > > bigger events neatly sidesteps that issue. > > As we discussed, grouping property changes into logical changes seems > agreeable to everyone. Yeah no one's disagreeing on that I guess, it's just that there's ways to get there that involve more or less typing :-) > > > > > With that, I agree that a global epoch counter from the connector would > > > > > be a good quick way for userspace to tell if a connector changed or not > > > > > since the last time it checked. > > > > > > > > > > > > > - If userspace misses an event and there's no epoch, we're forcing > > > > > > > > userspace to reprobe everything. Use case would be if a compositor is > > > > > > > > switched away we probably don't want to piss of the current compositor > > > > > > > > by blocking it's own probe kernel calls by doing our own (probe is > > > > > > > > single-threaded in the kernel through the dev->mode_config.mutex). If > > > > > > > > it can read the epoch property (which it can do without forcing a > > > > > > > > reprobe) userspace would know which connectors it needs to check and > > > > > > > > reprobe. > > > > > > > > > > > > > > > > Hence why epoch, it's a bit more robust userspace api. Ofc you could > > > > > > > > also require that userspace needs to keep parsing all uevents and make > > > > > > > > a list of all connectors it needs to reprobe when it's back to being > > > > > > > > the active compositor. But just comparing a current epoch with the one > > > > > > > > you cached from the last full probe is much easier. > > > > > > > > > > > > > > Fair enough, I think it's a fine idea for robustness yes, but I think > > > > > > > we could also provide extra info in the uevent when relevant and not > > > > > > > rely on that entirely. > > > > > > > > > > > > See above, with drmGetConnectorCurrent there's no need to provide more > > > > > > than what's needed in the uevent, since userspace can get everything > > > > > > else at the cost of one ioctl, without reprobing. With a bit of > > > > > > engineering work we could even avoid taking the expensive > > > > > > dev->mode_config.mutex lock for this fastpath. > > > > > > > > > > > > > > Another thing: None of this we can for connectors with unreliable hdp. > > > > > > > > Or at least you'll piss of users if you cache always. The sad thing is > > > > > > > > that HDMI is unreliable, at least on some machines/screen combos (you > > > > > > > > never get a hpd irq if you plug in/unplug). So real compositors still > > > > > > > > need to reprobe when the user asks for it. igt can probably get away > > > > > > > > without reprobing. > > > > > > > > > > > > > > I wonder how that is handled currently and how a user action can solve > > > > > > > the issue without any notification from the kernel. Maybe a need a > > > > > > > better understanding of that case to have a clearer idea. > > > > > > > > > > > > User opens the screen configuration tool -> usually at that point the > > > > > > tool/compositor force a full reprobe, which then often triggers the > > > > > > automatic reconfiguring. E.g. on one laptop I have here when I plug in > > > > > > random shit projectors at conferences nothing happens, until I run > > > > > > xrandr, which triggers the full reprobe, which then makes the kernel > > > > > > realize something change, sending and uevent, which starts the > > > > > > automatic reconfigure machinery. > > > > > > > > > > Oh right hehe, I definitely do that blind alt+f2 gnome-control-center > > > > > to get me out of an off-panel situation much too often. > > > > > > > > > > > There's also the issue that there's machines with hpd storms (even on > > > > > > DP, where you really need hpd to work to be compliant), and we have to > > > > > > turn of the hpd irq to keep the machine useable. > > > > > > > > > > I was under the impression that we switch to polling when a hpd storm > > > > > is detected in i915 (but that's a vague memory from my summer > > > > > internship at Intel 2 years ago). > > > > > > > > We do poll, but the issue is still that polling doesn't do the same > > > > thing as full reprobe. One just calls ->detect, the other ->detect + > > > > ->get_modes. This is a bit a silliness of the helpers and source of > > > > lots of confusion. A possible fix might be to always call both. > > > > > > I think that would make a lot of sense. I don't really get why we have > > > to wait for userspace to do a reprobe to get the new modes when the > > > kernel can be absolutely sure that the modes changed and need to be > > > refreshed. > > > > Don't let the names trick you, it's an entirely arbitrary split, > > historically grown :-) > > Seems like it would be sane to fix that up at this point. Sure, the problem is fixing up everyone. > > > > Plus even when that mess is sorted there's still the issue of broken > > > > hw, and we have no idea how much/where/which exactly. Except that > > > > every time we relied on hpd status, we got regression reports and had > > > > to revert. > > > > > > I think we could configure the notification behavior per-driver and > > > some events that wouldn't require a full probe on some drivers could do > > > so on other drivers. Userspace should find out whether to reprobe or > > > not dynamically based on what's in the uevent anyway, so it's not even > > > an inconsistency in the interface that 2 events be reported differently > > > in 2 drivers. I think that'd be better than the current situation we're > > > in anyway. If some drivers want be paranoid and always ask for > > > a reprobe of the connectors because of unreliable hw, their choice. > > > > The problem isn't typing the code, it's making sure we don't break the > > world with this. Even if you limit this to i915 alone. That's why I'm > > suggesting we untangle things as much as possible, makes it easier to > > do the (pretty much inevitable) revert. > > I really strongly believe that cases of broken hardware should not be a > show-stopper for improving the situation for every other non-broken > case ou there. Sure, there's the policy of no regressions, but it has > limits too. Wanting to support broken hardware is a noble goal, but I > think we should frankly ask ourselves at what cost we want it, and how > much we want to let that degrade the situation for other non-broken > hardware. The policy of "no regression" has pretty much no limits. Either you fix up everything, or it's not going to happen. Or there's some reasonable forward-compatible option. -Daniel > And in practice, I still don't see why there have to be regressions. If > a driver aims to support broken setups, it should deal with that itself > and not expect the core to. With what I'm proposing, that could mean > reporting that a full reprobe is needed on basically any event like we > do today. But please, let other drivers do better than that if they > can. > > Cheers, > > Paul > > > -Daniel > > > > > Cheers, > > > > > > Paul > > > > > > > -Daniel > > > > > > > > > Cheers, > > > > > > > > > > Paul > > > > > > > > > > > Cheers, Daniel > > > > > > > > > > > > > Cheers, > > > > > > > > > > > > > > Paul > > > > > > > > > > > > > > > -Daniel > > > > > > > > > > > > > > > > > Then we still have the legacy case: > > > > > > > > > HOTPLUG=1 > > > > > > > > > > > > > > > > > > where userspace is expected to reprobe all the connectors. > > > > > > > > > > > > > > > > > > I think this would deserve to be a separate series on its own. So I am > > > > > > > > > proposing to take this one off your plate and come up with another > > > > > > > > > seres implementing this proposal. What do you think? > > > > > > > > > > > > > > > > > > Cheers, > > > > > > > > > > > > > > > > > > Paul > > > > > > > > > > > > > > > > > > > v2: > > > > > > > > > > Minor fixes at KDoc comments [Daniel] > > > > > > > > > > v3: > > > > > > > > > > Check the property is really attached with connector [Daniel] > > > > > > > > > > > > > > > > > > > > Signed-off-by: Ramalingam C <ramalingam.c@xxxxxxxxx> > > > > > > > > > > Reviewed-by: Daniel Vetter <daniel.vetter@xxxxxxxx> > > > > > > > > > > --- > > > > > > > > > > drivers/gpu/drm/drm_sysfs.c | 35 +++++++++++++++++++++++++++++++++++ > > > > > > > > > > include/drm/drm_sysfs.h | 5 ++++- > > > > > > > > > > 2 files changed, 39 insertions(+), 1 deletion(-) > > > > > > > > > > > > > > > > > > > > diff --git a/drivers/gpu/drm/drm_sysfs.c b/drivers/gpu/drm/drm_sysfs.c > > > > > > > > > > index 18b1ac442997..63fa951a20db 100644 > > > > > > > > > > --- a/drivers/gpu/drm/drm_sysfs.c > > > > > > > > > > +++ b/drivers/gpu/drm/drm_sysfs.c > > > > > > > > > > @@ -21,6 +21,7 @@ > > > > > > > > > > #include <drm/drm_sysfs.h> > > > > > > > > > > #include <drm/drmP.h> > > > > > > > > > > #include "drm_internal.h" > > > > > > > > > > +#include "drm_crtc_internal.h" > > > > > > > > > > > > > > > > > > > > #define to_drm_minor(d) dev_get_drvdata(d) > > > > > > > > > > #define to_drm_connector(d) dev_get_drvdata(d) > > > > > > > > > > @@ -320,6 +321,9 @@ void drm_sysfs_lease_event(struct drm_device *dev) > > > > > > > > > > * Send a uevent for the DRM device specified by @dev. Currently we only > > > > > > > > > > * set HOTPLUG=1 in the uevent environment, but this could be expanded to > > > > > > > > > > * deal with other types of events. > > > > > > > > > > + * > > > > > > > > > > + * Any new uapi should be using the drm_sysfs_connector_status_event() > > > > > > > > > > + * for uevents on connector status change. > > > > > > > > > > */ > > > > > > > > > > void drm_sysfs_hotplug_event(struct drm_device *dev) > > > > > > > > > > { > > > > > > > > > > @@ -332,6 +336,37 @@ void drm_sysfs_hotplug_event(struct drm_device *dev) > > > > > > > > > > } > > > > > > > > > > EXPORT_SYMBOL(drm_sysfs_hotplug_event); > > > > > > > > > > > > > > > > > > > > +/** > > > > > > > > > > + * drm_sysfs_connector_status_event - generate a DRM uevent for connector > > > > > > > > > > + * property status change > > > > > > > > > > + * @connector: connector on which property status changed > > > > > > > > > > + * @property: connector property whoes status changed. > > > > > > > > > > + * > > > > > > > > > > + * Send a uevent for the DRM device specified by @dev. Currently we > > > > > > > > > > + * set HOTPLUG=1 and connector id along with the attached property id > > > > > > > > > > + * related to the status change. > > > > > > > > > > + */ > > > > > > > > > > +void drm_sysfs_connector_status_event(struct drm_connector *connector, > > > > > > > > > > + struct drm_property *property) > > > > > > > > > > +{ > > > > > > > > > > + struct drm_device *dev = connector->dev; > > > > > > > > > > + char hotplug_str[] = "HOTPLUG=1", conn_id[30], prop_id[30]; > > > > > > > > > > + char *envp[4] = { hotplug_str, conn_id, prop_id, NULL }; > > > > > > > > > > + > > > > > > > > > > + WARN_ON(!drm_mode_obj_find_prop_id(&connector->base, > > > > > > > > > > + property->base.id)); > > > > > > > > > > + > > > > > > > > > > + snprintf(conn_id, ARRAY_SIZE(conn_id), > > > > > > > > > > + "CONNECTOR=%u", connector->base.id); > > > > > > > > > > + snprintf(prop_id, ARRAY_SIZE(prop_id), > > > > > > > > > > + "PROPERTY=%u", property->base.id); > > > > > > > > > > + > > > > > > > > > > + DRM_DEBUG("generating connector status event\n"); > > > > > > > > > > + > > > > > > > > > > + kobject_uevent_env(&dev->primary->kdev->kobj, KOBJ_CHANGE, envp); > > > > > > > > > > +} > > > > > > > > > > +EXPORT_SYMBOL(drm_sysfs_connector_status_event); > > > > > > > > > > + > > > > > > > > > > static void drm_sysfs_release(struct device *dev) > > > > > > > > > > { > > > > > > > > > > kfree(dev); > > > > > > > > > > diff --git a/include/drm/drm_sysfs.h b/include/drm/drm_sysfs.h > > > > > > > > > > index 4f311e836cdc..d454ef617b2c 100644 > > > > > > > > > > --- a/include/drm/drm_sysfs.h > > > > > > > > > > +++ b/include/drm/drm_sysfs.h > > > > > > > > > > @@ -4,10 +4,13 @@ > > > > > > > > > > > > > > > > > > > > struct drm_device; > > > > > > > > > > struct device; > > > > > > > > > > +struct drm_connector; > > > > > > > > > > +struct drm_property; > > > > > > > > > > > > > > > > > > > > int drm_class_device_register(struct device *dev); > > > > > > > > > > void drm_class_device_unregister(struct device *dev); > > > > > > > > > > > > > > > > > > > > void drm_sysfs_hotplug_event(struct drm_device *dev); > > > > > > > > > > - > > > > > > > > > > +void drm_sysfs_connector_status_event(struct drm_connector *connector, > > > > > > > > > > + struct drm_property *property); > > > > > > > > > > #endif > > > > > > > > > -- > > > > > > > > > Paul Kocialkowski, Bootlin > > > > > > > > > Embedded Linux and kernel engineering > > > > > > > > > https://bootlin.com > > > > > > > > > > > > > > > > -- > > > > > > > Paul Kocialkowski, Bootlin > > > > > > > Embedded Linux and kernel engineering > > > > > > > https://bootlin.com > > > > > > > > > > -- > > > Paul Kocialkowski, Bootlin > > > Embedded Linux and kernel engineering > > > https://bootlin.com > > > > > > > > -- > Paul Kocialkowski, Bootlin > Embedded Linux and kernel engineering > https://bootlin.com > -- Daniel Vetter Software Engineer, Intel Corporation http://blog.ffwll.ch _______________________________________________ dri-devel mailing list dri-devel@xxxxxxxxxxxxxxxxxxxxx https://lists.freedesktop.org/mailman/listinfo/dri-devel
