Daniel Vetter
2015-Jun-15 07:56 UTC
[Nouveau] [PATCH v2 2/2] drm/nouveau: add GEM_SET_TILING staging ioctl
On Mon, Jun 15, 2015 at 04:09:29PM +0900, Alexandre Courbot wrote:> From: Ari Hirvonen <ahirvonen at nvidia.com> > > Add new NOUVEAU_GEM_SET_TILING ioctl to set correct tiling > mode for imported dma-bufs. This ioctl is staging for now > and enabled with the "staging_tiling" module option. > > Signed-off-by: Ari Hirvonen <ahirvonen at nvidia.com> > [acourbot at nvidia.com: carry upstream, many fixes] > Signed-off-by: Alexandre Courbot <acourbot at nvidia.com> > --- > drm/nouveau/nouveau_bo.c | 18 ++++++++++++ > drm/nouveau/nouveau_bo.h | 2 ++ > drm/nouveau/nouveau_drm.c | 6 ++++ > drm/nouveau/nouveau_gem.c | 58 ++++++++++++++++++++++++++++++++++++++ > drm/nouveau/nouveau_gem.h | 2 ++ > drm/nouveau/nouveau_ttm.c | 13 +-------- > drm/nouveau/uapi/drm/nouveau_drm.h | 8 ++++++ > 7 files changed, 95 insertions(+), 12 deletions(-) > > diff --git a/drm/nouveau/nouveau_bo.c b/drm/nouveau/nouveau_bo.c > index 6edcce1658b7..2a2ebbeb4fc0 100644 > --- a/drm/nouveau/nouveau_bo.c > +++ b/drm/nouveau/nouveau_bo.c > @@ -178,6 +178,24 @@ nouveau_bo_fixup_align(struct nouveau_bo *nvbo, u32 flags, > *size = roundup(*size, PAGE_SIZE); > } > > +void > +nouveau_bo_update_tiling(struct nouveau_drm *drm, struct nouveau_bo *nvbo, > + struct nvkm_mem *mem) > +{ > + switch (drm->device.info.family) { > + case NV_DEVICE_INFO_V0_TESLA: > + if (drm->device.info.chipset != 0x50) > + mem->memtype = (nvbo->tile_flags & 0x7f00) >> 8; > + break; > + case NV_DEVICE_INFO_V0_FERMI: > + case NV_DEVICE_INFO_V0_KEPLER: > + mem->memtype = (nvbo->tile_flags & 0xff00) >> 8; > + break; > + default: > + break; > + } > +} > + > int > nouveau_bo_new(struct drm_device *dev, int size, int align, > uint32_t flags, uint32_t tile_mode, uint32_t tile_flags, > diff --git a/drm/nouveau/nouveau_bo.h b/drm/nouveau/nouveau_bo.h > index e42360983229..87d07e3533eb 100644 > --- a/drm/nouveau/nouveau_bo.h > +++ b/drm/nouveau/nouveau_bo.h > @@ -69,6 +69,8 @@ nouveau_bo_ref(struct nouveau_bo *ref, struct nouveau_bo **pnvbo) > extern struct ttm_bo_driver nouveau_bo_driver; > > void nouveau_bo_move_init(struct nouveau_drm *); > +void nouveau_bo_update_tiling(struct nouveau_drm *, struct nouveau_bo *, > + struct nvkm_mem *); > int nouveau_bo_new(struct drm_device *, int size, int align, u32 flags, > u32 tile_mode, u32 tile_flags, struct sg_table *sg, > struct reservation_object *robj, > diff --git a/drm/nouveau/nouveau_drm.c b/drm/nouveau/nouveau_drm.c > index 28860268cf38..45a2c88ebf8e 100644 > --- a/drm/nouveau/nouveau_drm.c > +++ b/drm/nouveau/nouveau_drm.c > @@ -75,6 +75,10 @@ MODULE_PARM_DESC(runpm, "disable (0), force enable (1), optimus only default (-1 > int nouveau_runtime_pm = -1; > module_param_named(runpm, nouveau_runtime_pm, int, 0400); > > +MODULE_PARM_DESC(staging_tiling, "enable staging GEM_SET_TILING ioctl"); > +int nouveau_staging_tiling = 0; > +module_param_named(staging_tiling, nouveau_staging_tiling, int, 0600);Please use _unsafe here to make sure that setting this option taints the kernel and gives at least a bit of a deterrent. But in the end the policy is still that you can't regress anything if people complain, which means you might end up with a staging ioctl locked down forever. The other part I don't like with this plan is that it looks a bit like it could be easily abused to evade the open source userspace requirement upstream has for new interfaces. Doesn't help that your first staging ioctl doesn't come with links to mesa/hwc/whatever patches attached ;-) Overall I don't think this will help - you need internal branch management anyway, and upstreaming new ABI is somewhat painful for a reason: Screwing things up is really expensive long-term, and the drm community has paid that price a few too many times. Cheers, Daniel> + > static struct drm_driver driver_stub; > static struct drm_driver driver_pci; > static struct drm_driver driver_platform; > @@ -897,6 +901,7 @@ nouveau_ioctls[] = { > DRM_IOCTL_DEF_DRV(NOUVEAU_GEM_CPU_FINI, nouveau_gem_ioctl_cpu_fini, DRM_UNLOCKED|DRM_AUTH|DRM_RENDER_ALLOW), > DRM_IOCTL_DEF_DRV(NOUVEAU_GEM_INFO, nouveau_gem_ioctl_info, DRM_UNLOCKED|DRM_AUTH|DRM_RENDER_ALLOW), > /* Staging ioctls */ > + DRM_IOCTL_DEF_DRV(NOUVEAU_GEM_SET_TILING, nouveau_gem_ioctl_set_tiling, DRM_UNLOCKED|DRM_AUTH|DRM_RENDER_ALLOW), > }; > > long > @@ -1029,6 +1034,7 @@ static void nouveau_display_options(void) > DRM_DEBUG_DRIVER("... runpm : %d\n", nouveau_runtime_pm); > DRM_DEBUG_DRIVER("... vram_pushbuf : %d\n", nouveau_vram_pushbuf); > DRM_DEBUG_DRIVER("... pstate : %d\n", nouveau_pstate); > + DRM_DEBUG_DRIVER("... staging_tiling: %d\n", nouveau_staging_tiling); > } > > static const struct dev_pm_ops nouveau_pm_ops = { > diff --git a/drm/nouveau/nouveau_gem.c b/drm/nouveau/nouveau_gem.c > index 0e690bf19fc9..0e69449798aa 100644 > --- a/drm/nouveau/nouveau_gem.c > +++ b/drm/nouveau/nouveau_gem.c > @@ -172,6 +172,64 @@ nouveau_gem_object_close(struct drm_gem_object *gem, struct drm_file *file_priv) > ttm_bo_unreserve(&nvbo->bo); > } > > +extern int nouveau_staging_tiling; > +int > +nouveau_gem_ioctl_set_tiling(struct drm_device *dev, void *data, > + struct drm_file *file_priv) > +{ > + struct nouveau_drm *drm = nouveau_drm(dev); > + struct nouveau_cli *cli = nouveau_cli(file_priv); > + struct nvkm_fb *pfb = nvxx_fb(&drm->device); > + struct drm_nouveau_gem_set_tiling *req = data; > + struct drm_gem_object *gem; > + struct nouveau_bo *nvbo; > + struct nvkm_vma *vma; > + int ret = 0; > + > + if (!nouveau_staging_tiling) > + return -EINVAL; > + > + if (!pfb->memtype_valid(pfb, req->tile_flags)) { > + NV_PRINTK(error, cli, "bad page flags: 0x%08x\n", req->tile_flags); > + return -EINVAL; > + } > + > + gem = drm_gem_object_lookup(dev, file_priv, req->handle); > + if (!gem) > + return -ENOENT; > + > + nvbo = nouveau_gem_object(gem); > + > + /* We can only change tiling on PRIME-imported buffers */ > + if (nvbo->bo.type != ttm_bo_type_sg) { > + ret = -EINVAL; > + goto out; > + } > + > + if (nvbo->tile_mode != req->tile_mode || > + nvbo->tile_flags != req->tile_flags) { > + ret = ttm_bo_reserve(&nvbo->bo, false, false, false, NULL); > + if (ret) > + goto out; > + > + nvbo->tile_mode = req->tile_mode; > + nvbo->tile_flags = req->tile_flags; > + > + nouveau_bo_update_tiling(drm, nvbo, nvbo->bo.mem.mm_node); > + > + /* remap over existing mapping with new tile parameters */ > + vma = nouveau_bo_vma_find(nvbo, cli->vm); > + if (vma) > + nvkm_vm_map(vma, nvbo->bo.mem.mm_node); > + > + ttm_bo_unreserve(&nvbo->bo); > + } > + > +out: > + drm_gem_object_unreference_unlocked(gem); > + return ret; > +} > + > int > nouveau_gem_new(struct drm_device *dev, int size, int align, uint32_t domain, > uint32_t tile_mode, uint32_t tile_flags, > diff --git a/drm/nouveau/nouveau_gem.h b/drm/nouveau/nouveau_gem.h > index e4049faca780..56e741d98bcd 100644 > --- a/drm/nouveau/nouveau_gem.h > +++ b/drm/nouveau/nouveau_gem.h > @@ -23,6 +23,8 @@ extern void nouveau_gem_object_del(struct drm_gem_object *); > extern int nouveau_gem_object_open(struct drm_gem_object *, struct drm_file *); > extern void nouveau_gem_object_close(struct drm_gem_object *, > struct drm_file *); > +extern int nouveau_gem_ioctl_set_tiling(struct drm_device *, void *, > + struct drm_file *); > extern int nouveau_gem_ioctl_new(struct drm_device *, void *, > struct drm_file *); > extern int nouveau_gem_ioctl_pushbuf(struct drm_device *, void *, > diff --git a/drm/nouveau/nouveau_ttm.c b/drm/nouveau/nouveau_ttm.c > index 18f449715788..d680153db6e5 100644 > --- a/drm/nouveau/nouveau_ttm.c > +++ b/drm/nouveau/nouveau_ttm.c > @@ -174,18 +174,7 @@ nouveau_gart_manager_new(struct ttm_mem_type_manager *man, > > node->page_shift = 12; > > - switch (drm->device.info.family) { > - case NV_DEVICE_INFO_V0_TESLA: > - if (drm->device.info.chipset != 0x50) > - node->memtype = (nvbo->tile_flags & 0x7f00) >> 8; > - break; > - case NV_DEVICE_INFO_V0_FERMI: > - case NV_DEVICE_INFO_V0_KEPLER: > - node->memtype = (nvbo->tile_flags & 0xff00) >> 8; > - break; > - default: > - break; > - } > + nouveau_bo_update_tiling(drm, nvbo, node); > > mem->mm_node = node; > mem->start = 0; > diff --git a/drm/nouveau/uapi/drm/nouveau_drm.h b/drm/nouveau/uapi/drm/nouveau_drm.h > index 4e7e21f41b5c..8f10b16b1473 100644 > --- a/drm/nouveau/uapi/drm/nouveau_drm.h > +++ b/drm/nouveau/uapi/drm/nouveau_drm.h > @@ -64,6 +64,12 @@ struct drm_nouveau_gem_new { > uint32_t align; > }; > > +struct drm_nouveau_gem_set_tiling { > + uint32_t handle; > + uint32_t tile_mode; > + uint32_t tile_flags; > +}; > + > #define NOUVEAU_GEM_MAX_BUFFERS 1024 > struct drm_nouveau_gem_pushbuf_bo_presumed { > uint32_t valid; > @@ -142,6 +148,7 @@ struct drm_nouveau_gem_cpu_fini { > #define DRM_NOUVEAU_GEM_INFO 0x44 > /* range 0x98..DRM_COMMAND_END (8 entries) is reserved for staging, unstable ioctls */ > #define DRM_NOUVEAU_STAGING_IOCTL 0x58 > +#define DRM_NOUVEAU_GEM_SET_TILING (DRM_NOUVEAU_STAGING_IOCTL + 0x0) > > #define DRM_IOCTL_NOUVEAU_GEM_NEW DRM_IOWR(DRM_COMMAND_BASE + DRM_NOUVEAU_GEM_NEW, struct drm_nouveau_gem_new) > #define DRM_IOCTL_NOUVEAU_GEM_PUSHBUF DRM_IOWR(DRM_COMMAND_BASE + DRM_NOUVEAU_GEM_PUSHBUF, struct drm_nouveau_gem_pushbuf) > @@ -149,5 +156,6 @@ struct drm_nouveau_gem_cpu_fini { > #define DRM_IOCTL_NOUVEAU_GEM_CPU_FINI DRM_IOW (DRM_COMMAND_BASE + DRM_NOUVEAU_GEM_CPU_FINI, struct drm_nouveau_gem_cpu_fini) > #define DRM_IOCTL_NOUVEAU_GEM_INFO DRM_IOWR(DRM_COMMAND_BASE + DRM_NOUVEAU_GEM_INFO, struct drm_nouveau_gem_info) > /* staging ioctls */ > +#define DRM_IOCTL_NOUVEAU_GEM_SET_TILING DRM_IOWR(DRM_COMMAND_BASE + DRM_NOUVEAU_GEM_SET_TILING, struct drm_nouveau_gem_set_tiling) > > #endif /* __NOUVEAU_DRM_H__ */ > -- > 2.4.2 > > _______________________________________________ > Nouveau mailing list > Nouveau at lists.freedesktop.org > http://lists.freedesktop.org/mailman/listinfo/nouveau-- Daniel Vetter Software Engineer, Intel Corporation http://blog.ffwll.ch
Alexandre Courbot
2015-Jun-15 09:08 UTC
[Nouveau] [PATCH v2 2/2] drm/nouveau: add GEM_SET_TILING staging ioctl
On 06/15/2015 04:56 PM, Daniel Vetter wrote:> On Mon, Jun 15, 2015 at 04:09:29PM +0900, Alexandre Courbot wrote: >> From: Ari Hirvonen <ahirvonen at nvidia.com> >> >> Add new NOUVEAU_GEM_SET_TILING ioctl to set correct tiling >> mode for imported dma-bufs. This ioctl is staging for now >> and enabled with the "staging_tiling" module option. >> >> Signed-off-by: Ari Hirvonen <ahirvonen at nvidia.com> >> [acourbot at nvidia.com: carry upstream, many fixes] >> Signed-off-by: Alexandre Courbot <acourbot at nvidia.com> >> --- >> drm/nouveau/nouveau_bo.c | 18 ++++++++++++ >> drm/nouveau/nouveau_bo.h | 2 ++ >> drm/nouveau/nouveau_drm.c | 6 ++++ >> drm/nouveau/nouveau_gem.c | 58 ++++++++++++++++++++++++++++++++++++++ >> drm/nouveau/nouveau_gem.h | 2 ++ >> drm/nouveau/nouveau_ttm.c | 13 +-------- >> drm/nouveau/uapi/drm/nouveau_drm.h | 8 ++++++ >> 7 files changed, 95 insertions(+), 12 deletions(-) >> >> diff --git a/drm/nouveau/nouveau_bo.c b/drm/nouveau/nouveau_bo.c >> index 6edcce1658b7..2a2ebbeb4fc0 100644 >> --- a/drm/nouveau/nouveau_bo.c >> +++ b/drm/nouveau/nouveau_bo.c >> @@ -178,6 +178,24 @@ nouveau_bo_fixup_align(struct nouveau_bo *nvbo, u32 flags, >> *size = roundup(*size, PAGE_SIZE); >> } >> >> +void >> +nouveau_bo_update_tiling(struct nouveau_drm *drm, struct nouveau_bo *nvbo, >> + struct nvkm_mem *mem) >> +{ >> + switch (drm->device.info.family) { >> + case NV_DEVICE_INFO_V0_TESLA: >> + if (drm->device.info.chipset != 0x50) >> + mem->memtype = (nvbo->tile_flags & 0x7f00) >> 8; >> + break; >> + case NV_DEVICE_INFO_V0_FERMI: >> + case NV_DEVICE_INFO_V0_KEPLER: >> + mem->memtype = (nvbo->tile_flags & 0xff00) >> 8; >> + break; >> + default: >> + break; >> + } >> +} >> + >> int >> nouveau_bo_new(struct drm_device *dev, int size, int align, >> uint32_t flags, uint32_t tile_mode, uint32_t tile_flags, >> diff --git a/drm/nouveau/nouveau_bo.h b/drm/nouveau/nouveau_bo.h >> index e42360983229..87d07e3533eb 100644 >> --- a/drm/nouveau/nouveau_bo.h >> +++ b/drm/nouveau/nouveau_bo.h >> @@ -69,6 +69,8 @@ nouveau_bo_ref(struct nouveau_bo *ref, struct nouveau_bo **pnvbo) >> extern struct ttm_bo_driver nouveau_bo_driver; >> >> void nouveau_bo_move_init(struct nouveau_drm *); >> +void nouveau_bo_update_tiling(struct nouveau_drm *, struct nouveau_bo *, >> + struct nvkm_mem *); >> int nouveau_bo_new(struct drm_device *, int size, int align, u32 flags, >> u32 tile_mode, u32 tile_flags, struct sg_table *sg, >> struct reservation_object *robj, >> diff --git a/drm/nouveau/nouveau_drm.c b/drm/nouveau/nouveau_drm.c >> index 28860268cf38..45a2c88ebf8e 100644 >> --- a/drm/nouveau/nouveau_drm.c >> +++ b/drm/nouveau/nouveau_drm.c >> @@ -75,6 +75,10 @@ MODULE_PARM_DESC(runpm, "disable (0), force enable (1), optimus only default (-1 >> int nouveau_runtime_pm = -1; >> module_param_named(runpm, nouveau_runtime_pm, int, 0400); >> >> +MODULE_PARM_DESC(staging_tiling, "enable staging GEM_SET_TILING ioctl"); >> +int nouveau_staging_tiling = 0; >> +module_param_named(staging_tiling, nouveau_staging_tiling, int, 0600); > > Please use _unsafe here to make sure that setting this option taints the > kernel and gives at least a bit of a deterrent. But in the end the policy > is still that you can't regress anything if people complain, which means > you might end up with a staging ioctl locked down forever.That would kind of kill the whole purpose of this patchset. But at the same time the point of having staging ioctls is to say "don't use them in production", so hopefully the message is clear.> The other part I don't like with this plan is that it looks a bit like it > could be easily abused to evade the open source userspace requirement > upstream has for new interfaces. Doesn't help that your first staging > ioctl doesn't come with links to mesa/hwc/whatever patches attached ;-)Well, you could abuse it - no more than 8 times though. ;) The point is not to evade anything though, but rather to have the necessary kernel code land in such a way that we can experiment with Mesa and other user-space.> Overall I don't think this will help - you need internal branch management > anyway, and upstreaming new ABI is somewhat painful for a reason: Screwing > things up is really expensive long-term, and the drm community has paid > that price a few too many times.It seems to me that this staging feature can exactly help with that: allow new ioctls to mature a bit (no longer than a kernel release cycle) and avoid that "ah, I wish we did this differently" moment. But considering the number of ABIs I have driven so far (0), someone more experienced may challenge that belief.
Daniel Vetter
2015-Jun-16 10:07 UTC
[Nouveau] [PATCH v2 2/2] drm/nouveau: add GEM_SET_TILING staging ioctl
On Mon, Jun 15, 2015 at 06:08:21PM +0900, Alexandre Courbot wrote:> On 06/15/2015 04:56 PM, Daniel Vetter wrote: > >On Mon, Jun 15, 2015 at 04:09:29PM +0900, Alexandre Courbot wrote: > >>From: Ari Hirvonen <ahirvonen at nvidia.com> > >> > >>Add new NOUVEAU_GEM_SET_TILING ioctl to set correct tiling > >>mode for imported dma-bufs. This ioctl is staging for now > >>and enabled with the "staging_tiling" module option. > >> > >>Signed-off-by: Ari Hirvonen <ahirvonen at nvidia.com> > >>[acourbot at nvidia.com: carry upstream, many fixes] > >>Signed-off-by: Alexandre Courbot <acourbot at nvidia.com> > >>--- > >> drm/nouveau/nouveau_bo.c | 18 ++++++++++++ > >> drm/nouveau/nouveau_bo.h | 2 ++ > >> drm/nouveau/nouveau_drm.c | 6 ++++ > >> drm/nouveau/nouveau_gem.c | 58 ++++++++++++++++++++++++++++++++++++++ > >> drm/nouveau/nouveau_gem.h | 2 ++ > >> drm/nouveau/nouveau_ttm.c | 13 +-------- > >> drm/nouveau/uapi/drm/nouveau_drm.h | 8 ++++++ > >> 7 files changed, 95 insertions(+), 12 deletions(-) > >> > >>diff --git a/drm/nouveau/nouveau_bo.c b/drm/nouveau/nouveau_bo.c > >>index 6edcce1658b7..2a2ebbeb4fc0 100644 > >>--- a/drm/nouveau/nouveau_bo.c > >>+++ b/drm/nouveau/nouveau_bo.c > >>@@ -178,6 +178,24 @@ nouveau_bo_fixup_align(struct nouveau_bo *nvbo, u32 flags, > >> *size = roundup(*size, PAGE_SIZE); > >> } > >> > >>+void > >>+nouveau_bo_update_tiling(struct nouveau_drm *drm, struct nouveau_bo *nvbo, > >>+ struct nvkm_mem *mem) > >>+{ > >>+ switch (drm->device.info.family) { > >>+ case NV_DEVICE_INFO_V0_TESLA: > >>+ if (drm->device.info.chipset != 0x50) > >>+ mem->memtype = (nvbo->tile_flags & 0x7f00) >> 8; > >>+ break; > >>+ case NV_DEVICE_INFO_V0_FERMI: > >>+ case NV_DEVICE_INFO_V0_KEPLER: > >>+ mem->memtype = (nvbo->tile_flags & 0xff00) >> 8; > >>+ break; > >>+ default: > >>+ break; > >>+ } > >>+} > >>+ > >> int > >> nouveau_bo_new(struct drm_device *dev, int size, int align, > >> uint32_t flags, uint32_t tile_mode, uint32_t tile_flags, > >>diff --git a/drm/nouveau/nouveau_bo.h b/drm/nouveau/nouveau_bo.h > >>index e42360983229..87d07e3533eb 100644 > >>--- a/drm/nouveau/nouveau_bo.h > >>+++ b/drm/nouveau/nouveau_bo.h > >>@@ -69,6 +69,8 @@ nouveau_bo_ref(struct nouveau_bo *ref, struct nouveau_bo **pnvbo) > >> extern struct ttm_bo_driver nouveau_bo_driver; > >> > >> void nouveau_bo_move_init(struct nouveau_drm *); > >>+void nouveau_bo_update_tiling(struct nouveau_drm *, struct nouveau_bo *, > >>+ struct nvkm_mem *); > >> int nouveau_bo_new(struct drm_device *, int size, int align, u32 flags, > >> u32 tile_mode, u32 tile_flags, struct sg_table *sg, > >> struct reservation_object *robj, > >>diff --git a/drm/nouveau/nouveau_drm.c b/drm/nouveau/nouveau_drm.c > >>index 28860268cf38..45a2c88ebf8e 100644 > >>--- a/drm/nouveau/nouveau_drm.c > >>+++ b/drm/nouveau/nouveau_drm.c > >>@@ -75,6 +75,10 @@ MODULE_PARM_DESC(runpm, "disable (0), force enable (1), optimus only default (-1 > >> int nouveau_runtime_pm = -1; > >> module_param_named(runpm, nouveau_runtime_pm, int, 0400); > >> > >>+MODULE_PARM_DESC(staging_tiling, "enable staging GEM_SET_TILING ioctl"); > >>+int nouveau_staging_tiling = 0; > >>+module_param_named(staging_tiling, nouveau_staging_tiling, int, 0600); > > > >Please use _unsafe here to make sure that setting this option taints the > >kernel and gives at least a bit of a deterrent. But in the end the policy > >is still that you can't regress anything if people complain, which means > >you might end up with a staging ioctl locked down forever. > > That would kind of kill the whole purpose of this patchset. But at the same > time the point of having staging ioctls is to say "don't use them in > production", so hopefully the message is clear. > > >The other part I don't like with this plan is that it looks a bit like it > >could be easily abused to evade the open source userspace requirement > >upstream has for new interfaces. Doesn't help that your first staging > >ioctl doesn't come with links to mesa/hwc/whatever patches attached ;-) > > Well, you could abuse it - no more than 8 times though. ;) > > The point is not to evade anything though, but rather to have the necessary > kernel code land in such a way that we can experiment with Mesa and other > user-space. > > >Overall I don't think this will help - you need internal branch management > >anyway, and upstreaming new ABI is somewhat painful for a reason: Screwing > >things up is really expensive long-term, and the drm community has paid > >that price a few too many times. > > It seems to me that this staging feature can exactly help with that: allow > new ioctls to mature a bit (no longer than a kernel release cycle) and avoid > that "ah, I wish we did this differently" moment. But considering the number > of ABIs I have driven so far (0), someone more experienced may challenge > that belief.Maybe some follow up from irc discussions is in order: It's really a judgment call whether it makes sense. Imo the problem is that marking something as staging is way too tempting and you get sloppy and end up with a baked-in abi that you don't really want. Since in the end it doesn't matter what you declare as staging but whether there are people complaining about regressions when you break it (famously demonstrated years ago with nouveau in staging and giant flamours when nouveau broke their abi, resulting in Linus ulitmately demanding that nouveau gets destaged asap). Given that it's imo better to just roll with a final abi. I think in the past we've overdone things a bit with experimental abis in the drm core, only the atomic stuff looks big enough in retrospective to really justify staging it. Everything else was just chickening out from committing to things right away. In i915 the only thing we stage nowadays is new hw enabling, simply because we start so early nowadays that with the first cut we simply don't know yet whether the code will work perfectly on final silicon. And we've been bitten in the past by regression reports where people upgraded from old kernels to versions with experimental support and couldn't boot any more at all. But that's the only exception. You also mention the issue with having a common assembly ground for internal patches. Staging ioctls isn't going to solve that, you need internal trees anyway. We have lots of those for i915, you just can't see them ;-) Summary of my stance: In almost all cases staging ioctls is just an illusion and doesn't give you real benefits. Cheers, Daniel -- Daniel Vetter Software Engineer, Intel Corporation http://blog.ffwll.ch