[Freedreno] [PATCH] drm/msm: Add syncobj support.

Kristian Høgsberg hoegsberg at gmail.com
Tue Jan 14 17:26:56 UTC 2020


On Tue, Jan 14, 2020 at 8:41 AM Rob Clark <robdclark at chromium.org> wrote:
>
> On Tue, Jan 14, 2020 at 7:58 AM Jordan Crouse <jcrouse at codeaurora.org> wrote:
> >
> > On Tue, Jan 14, 2020 at 01:40:11AM +0100, Bas Nieuwenhuizen wrote:
> > > On Tue, Jan 14, 2020 at 12:41 AM Jordan Crouse <jcrouse at codeaurora.org> wrote:
> > > >
> > > > On Mon, Jan 13, 2020 at 09:25:57PM +0100, Bas Nieuwenhuizen wrote:
> > > > > This
> > > > >
> > > > > 1) Enables core DRM syncobj support.
> > > > > 2) Adds options to the submission ioctl to wait/signal syncobjs.
> > > > >
> > > > > Just like the wait fence fd, this does inline waits. Using the
> > > > > scheduler would be nice but I believe it is out of scope for
> > > > > this work.
> > > > >
> > > > > Support for timeline syncobjs is implemented and the interface
> > > > > is ready for it, but I'm not enabling it yet until there is
> > > > > some code for turnip to use it.
> > > > >
> > > > > The reset is mostly in there because in the presence of waiting
> > > > > and signalling the same semaphores, resetting them after
> > > > > signalling can become very annoying.
> > > > >
> > > > > Signed-off-by: Bas Nieuwenhuizen <bas at basnieuwenhuizen.nl>
> > > > > ---
> > > > >
> > > > > Userspace code in
> > > > >
> > > > > https://gitlab.freedesktop.org/mesa/mesa/merge_requests/2769
> > > > >
> > > > >  drivers/gpu/drm/msm/msm_drv.c        |   6 +-
> > > > >  drivers/gpu/drm/msm/msm_gem_submit.c | 241 ++++++++++++++++++++++++++-
> > > > >  include/uapi/drm/msm_drm.h           |  22 ++-
> > > > >  3 files changed, 265 insertions(+), 4 deletions(-)
> > > > >
> > > > > diff --git a/drivers/gpu/drm/msm/msm_drv.c b/drivers/gpu/drm/msm/msm_drv.c
> > > > > index c84f0a8b3f2c..ca36d6b21d8f 100644
> > > > > --- a/drivers/gpu/drm/msm/msm_drv.c
> > > > > +++ b/drivers/gpu/drm/msm/msm_drv.c
> > > > > @@ -37,9 +37,10 @@
> > > > >   * - 1.4.0 - softpin, MSM_RELOC_BO_DUMP, and GEM_INFO support to set/get
> > > > >   *           GEM object's debug name
> > > > >   * - 1.5.0 - Add SUBMITQUERY_QUERY ioctl
> > > > > + * - 1.6.0 - Syncobj support
> > > > >   */
> > > > >  #define MSM_VERSION_MAJOR    1
> > > > > -#define MSM_VERSION_MINOR    5
> > > > > +#define MSM_VERSION_MINOR    6
> > > > >  #define MSM_VERSION_PATCHLEVEL       0
> > > > >
> > > > >  static const struct drm_mode_config_funcs mode_config_funcs = {
> > > > > @@ -988,7 +989,8 @@ static struct drm_driver msm_driver = {
> > > > >       .driver_features    = DRIVER_GEM |
> > > > >                               DRIVER_RENDER |
> > > > >                               DRIVER_ATOMIC |
> > > > > -                             DRIVER_MODESET,
> > > > > +                             DRIVER_MODESET|
> > > >
> > > > A space before the | would be preferred.
> > >
> > > Done.
> > > >
> > > > > +                             DRIVER_SYNCOBJ,
> > > > >       .open               = msm_open,
> > > > >       .postclose           = msm_postclose,
> > > > >       .lastclose          = drm_fb_helper_lastclose,
> > > > > diff --git a/drivers/gpu/drm/msm/msm_gem_submit.c b/drivers/gpu/drm/msm/msm_gem_submit.c
> > > > > index be5327af16fa..9085229f88e0 100644
> > > > > --- a/drivers/gpu/drm/msm/msm_gem_submit.c
> > > > > +++ b/drivers/gpu/drm/msm/msm_gem_submit.c
> > > > > @@ -8,7 +8,9 @@
> > > > >  #include <linux/sync_file.h>
> > > > >  #include <linux/uaccess.h>
> > > > >
> > > > > +#include <drm/drm_drv.h>
> > > > >  #include <drm/drm_file.h>
> > > > > +#include <drm/drm_syncobj.h>
> > > > >
> > > > >  #include "msm_drv.h"
> > > > >  #include "msm_gpu.h"
> > > > > @@ -394,6 +396,196 @@ static void submit_cleanup(struct msm_gem_submit *submit)
> > > > >       ww_acquire_fini(&submit->ticket);
> > > > >  }
> > > > >
> > > > > +
> > > > > +struct msm_submit_post_dep {
> > > > > +     struct drm_syncobj *syncobj;
> > > > > +     uint64_t point;
> > > > > +     struct dma_fence_chain *chain;
> > > > > +};
> > > > > +
> > > > > +static int msm_wait_deps(struct drm_device *dev,
> > > > > +                         struct drm_file *file,
> > > > > +                         uint64_t in_syncobjs_addr,
> > > > > +                         uint32_t nr_in_syncobjs,
> > > > > +                         struct msm_ringbuffer *ring,
> > > > > +                         struct drm_syncobj ***syncobjs)
> > > > > +{
> > > > > +     struct drm_msm_gem_submit_syncobj *syncobj_descs;
> > > > > +     int ret = 0;
> > > > > +     uint32_t i, j;
> > > > > +
> > > > > +     syncobj_descs = kmalloc_array(nr_in_syncobjs, sizeof(*syncobj_descs),
> > > > > +                                   GFP_KERNEL);
> > > > > +     if (!syncobj_descs)
> > > > > +             return -ENOMEM;
> > > > > +
> > > > We would want to watch out here for fuzzers and malicious actors that try to
> > > > force an enormous memory allocation. It seems like we should be able to
> > > > iteratively read each fences in the loop and skip this memory allocation.
> > > >
> > > > > +     *syncobjs = kcalloc(nr_in_syncobjs, sizeof(**syncobjs), GFP_KERNEL);
> > > > > +     if (!syncobjs) {
> > > > > +             ret = -ENOMEM;
> > > > > +             goto out_syncobjs;
> > > > > +     }
> > > >
> > > > Alas no good way to skip this one. But it seems that syncobjs should only
> > > > contain descriptors with MSM_SUBMIT_SYNCOBJ_RESET set. I'm not very familiar
> > > > with how fences work so I'm not sure how common this path is. Would the same
> > > > fuzzer or malicious actor be able to double the destruction by forcing a large
> > > > allocation that doesn't even end up getting used?
> > >
> > > In real usecases I expect MSM_SUBMIT_SYNCOBJ_RESET to be set for 50%+
> > > of the entries and the number of entries to be < 10.
> > >
> > > I can certainly start doing a copy_from_user per entry and save one of
> > > the array. (I was under the impression that copy_from_user was
> > > expensive but if it is not, okay)
> >
> > I guess with recent exploit mitigations it is more expensive, but it shouldn't
> > be too bad if your nominal use cases are somewhere in the area of 10. That
> > said...
> >
> > > Overall though, there is a real issue of wanting to delay all write
> > > actions until we are sure the ioctl will succeed. That will mean we
> > > need to have arrays that are on the order of a UINT32_MAX elements if
> > > we assume full range on the inputs. How much is it worth trying to
> > > squeeze the syncobjs_to_reset, given that a malicious caller could
> > > just set all the reset flags? Especially since a malicious actor would
> > > instead just cause large allocations in the post_deps instead which we
> > > always need to allocate.
> > >
> > > What is the thread model here and what significant improvements can be
> > > made to avoid issues?
> >
> > I'm mostly worried about dealing with fuzzers who will throw you the full u32
> > range and I'm always worried about providing easy ways for non-trusted users to
> > exert memory pressure.
> >
> > > The only thing I could think of is that by doing krealloc we require
> > > the user to commit to using similar amount of memory in userspace.
> > > However, that comes at the significant complexity cost of handling
> > > reallocing and handling the failures of that.
> >
> > If there needs to be a 1:1 relationship between the user and the kernel then
> > I agree krealloc isn't a great idea.
> >
> > > Thoughts?
> >
> > Should we just stick with the classics and restrict the maximum number of fences
> > to a fixed number? 50?  128? You would want the synobjs allocation to fit within
> > a page anyway so 4096 / sizeof(struct drm_syncobj) might be a good start.
> > Assuming we don't run up against any angry tests that try to allocate hundreds
> > of fences this should do and you don't have to worry about the copy_to_user cost
> > you mention above.
> >
>
> I suppose in the end it isn't *too* much different from the existing
> bos/cmds tables on the submit ioctl, is it?  Maybe we should be caring
> about those too, but we don't currently.
>
> Is there a way to allocate memory on the kernel side which abides by
> userspace process limits?  That kinda sounds like what we need here.

Alternatively we can use the fixed-size-bulk copy pattern where we
allocate a page, copy as many fences as will fit, process them,
iterate. We still get a bulk copy from userspace, we have a fixed size
kernel allocation but we don't restrict the number of fences.

> BR,
> -R
> _______________________________________________
> Freedreno mailing list
> Freedreno at lists.freedesktop.org
> https://lists.freedesktop.org/mailman/listinfo/freedreno


More information about the Freedreno mailing list