On Fri, Mar 25, 2022 at 11:35:49AM +0100, Christian König wrote:
> Am 25.03.22 um 11:13 schrieb Daniel Vetter:
> > On Fri, Mar 11, 2022 at 12:02:44PM +0100, Christian König wrote:
> > > The dma_fence_chain containers can show up in sync_files as well resulting in
> > > warnings that those can't be added to dma_fence_array containers when merging
> > > multiple sync_files together.
> > >
> > > Solve this by using the dma_fence_unwrap iterator to deep dive into the
> > > contained fences and then add those flatten out into a dma_fence_array.
> > >
> > > Signed-off-by: Christian König <christian.koenig@xxxxxxx>
> > I have no idea why we try to keep fences sorted, but oh well it looks like
> > the merging is done correctly.
>
> To be honest I don't fully know either.
>
> Keeping the array sorted by context allows to merge it without adding
> duplicates, but adding duplicates is not an extra overhead to begin with
> because we always allocate memory for the worst case anyway.
>
> Just keeping it around for now.
Hm I guess if we want to keep that we could make that an invariant of dma
fence arrays, i.e. sorted and deduplicated. Usually there should be few
enough fences that de-duping shouldn't be expensive really.
But no idea whether that's worth it.
-Daniel
>
> > Reviewed-by: Daniel Vetter <daniel.vetter@xxxxxxxx>
>
> Thanks,
> Christian.
>
> >
> > > ---
> > > drivers/dma-buf/sync_file.c | 141 +++++++++++++++++++-----------------
> > > 1 file changed, 73 insertions(+), 68 deletions(-)
> > >
> > > diff --git a/drivers/dma-buf/sync_file.c b/drivers/dma-buf/sync_file.c
> > > index 394e6e1e9686..b8dea4ec123b 100644
> > > --- a/drivers/dma-buf/sync_file.c
> > > +++ b/drivers/dma-buf/sync_file.c
> > > @@ -5,6 +5,7 @@
> > > * Copyright (C) 2012 Google, Inc.
> > > */
> > > +#include <linux/dma-fence-unwrap.h>
> > > #include <linux/export.h>
> > > #include <linux/file.h>
> > > #include <linux/fs.h>
> > > @@ -172,20 +173,6 @@ static int sync_file_set_fence(struct sync_file *sync_file,
> > > return 0;
> > > }
> > > -static struct dma_fence **get_fences(struct sync_file *sync_file,
> > > - int *num_fences)
> > > -{
> > > - if (dma_fence_is_array(sync_file->fence)) {
> > > - struct dma_fence_array *array = to_dma_fence_array(sync_file->fence);
> > > -
> > > - *num_fences = array->num_fences;
> > > - return array->fences;
> > > - }
> > > -
> > > - *num_fences = 1;
> > > - return &sync_file->fence;
> > > -}
> > > -
> > > static void add_fence(struct dma_fence **fences,
> > > int *i, struct dma_fence *fence)
> > > {
> > > @@ -210,86 +197,97 @@ static void add_fence(struct dma_fence **fences,
> > > static struct sync_file *sync_file_merge(const char *name, struct sync_file *a,
> > > struct sync_file *b)
> > > {
> > > + struct dma_fence *a_fence, *b_fence, **fences;
> > > + struct dma_fence_unwrap a_iter, b_iter;
> > > + unsigned int index, num_fences;
> > > struct sync_file *sync_file;
> > > - struct dma_fence **fences = NULL, **nfences, **a_fences, **b_fences;
> > > - int i = 0, i_a, i_b, num_fences, a_num_fences, b_num_fences;
> > > sync_file = sync_file_alloc();
> > > if (!sync_file)
> > > return NULL;
> > > - a_fences = get_fences(a, &a_num_fences);
> > > - b_fences = get_fences(b, &b_num_fences);
> > > - if (a_num_fences > INT_MAX - b_num_fences)
> > > - goto err;
> > > + num_fences = 0;
> > > + dma_fence_unwrap_for_each(a_fence, &a_iter, a->fence)
> > > + ++num_fences;
> > > + dma_fence_unwrap_for_each(b_fence, &b_iter, b->fence)
> > > + ++num_fences;
> > > - num_fences = a_num_fences + b_num_fences;
> > > + if (num_fences > INT_MAX)
> > > + goto err_free_sync_file;
> > > fences = kcalloc(num_fences, sizeof(*fences), GFP_KERNEL);
> > > if (!fences)
> > > - goto err;
> > > + goto err_free_sync_file;
> > > /*
> > > - * Assume sync_file a and b are both ordered and have no
> > > - * duplicates with the same context.
> > > + * We can't guarantee that fences in both a and b are ordered, but it is
> > > + * still quite likely.
> > > *
> > > - * If a sync_file can only be created with sync_file_merge
> > > - * and sync_file_create, this is a reasonable assumption.
> > > + * So attempt to order the fences as we pass over them and merge fences
> > > + * with the same context.
> > > */
> > > - for (i_a = i_b = 0; i_a < a_num_fences && i_b < b_num_fences; ) {
> > > - struct dma_fence *pt_a = a_fences[i_a];
> > > - struct dma_fence *pt_b = b_fences[i_b];
> > > - if (pt_a->context < pt_b->context) {
> > > - add_fence(fences, &i, pt_a);
> > > + index = 0;
> > > + for (a_fence = dma_fence_unwrap_first(a->fence, &a_iter),
> > > + b_fence = dma_fence_unwrap_first(b->fence, &b_iter);
> > > + a_fence || b_fence; ) {
> > > +
> > > + if (!b_fence) {
> > > + add_fence(fences, &index, a_fence);
> > > + a_fence = dma_fence_unwrap_next(&a_iter);
> > > +
> > > + } else if (!a_fence) {
> > > + add_fence(fences, &index, b_fence);
> > > + b_fence = dma_fence_unwrap_next(&b_iter);
> > > +
> > > + } else if (a_fence->context < b_fence->context) {
> > > + add_fence(fences, &index, a_fence);
> > > + a_fence = dma_fence_unwrap_next(&a_iter);
> > > - i_a++;
> > > - } else if (pt_a->context > pt_b->context) {
> > > - add_fence(fences, &i, pt_b);
> > > + } else if (b_fence->context < a_fence->context) {
> > > + add_fence(fences, &index, b_fence);
> > > + b_fence = dma_fence_unwrap_next(&b_iter);
> > > +
> > > + } else if (__dma_fence_is_later(a_fence->seqno, b_fence->seqno,
> > > + a_fence->ops)) {
> > > + add_fence(fences, &index, a_fence);
> > > + a_fence = dma_fence_unwrap_next(&a_iter);
> > > + b_fence = dma_fence_unwrap_next(&b_iter);
> > > - i_b++;
> > > } else {
> > > - if (__dma_fence_is_later(pt_a->seqno, pt_b->seqno,
> > > - pt_a->ops))
> > > - add_fence(fences, &i, pt_a);
> > > - else
> > > - add_fence(fences, &i, pt_b);
> > > -
> > > - i_a++;
> > > - i_b++;
> > > + add_fence(fences, &index, b_fence);
> > > + a_fence = dma_fence_unwrap_next(&a_iter);
> > > + b_fence = dma_fence_unwrap_next(&b_iter);
> > > }
> > > }
> > > - for (; i_a < a_num_fences; i_a++)
> > > - add_fence(fences, &i, a_fences[i_a]);
> > > -
> > > - for (; i_b < b_num_fences; i_b++)
> > > - add_fence(fences, &i, b_fences[i_b]);
> > > -
> > > - if (i == 0)
> > > - fences[i++] = dma_fence_get(a_fences[0]);
> > > + if (index == 0)
> > > + add_fence(fences, &index, dma_fence_get_stub());
> > > - if (num_fences > i) {
> > > - nfences = krealloc_array(fences, i, sizeof(*fences), GFP_KERNEL);
> > > - if (!nfences)
> > > - goto err;
> > > + if (num_fences > index) {
> > > + struct dma_fence **tmp;
> > > - fences = nfences;
> > > + /* Keep going even when reducing the size failed */
> > > + tmp = krealloc_array(fences, index, sizeof(*fences),
> > > + GFP_KERNEL);
> > > + if (tmp)
> > > + fences = tmp;
> > > }
> > > - if (sync_file_set_fence(sync_file, fences, i) < 0)
> > > - goto err;
> > > + if (sync_file_set_fence(sync_file, fences, index) < 0)
> > > + goto err_put_fences;
> > > strlcpy(sync_file->user_name, name, sizeof(sync_file->user_name));
> > > return sync_file;
> > > -err:
> > > - while (i)
> > > - dma_fence_put(fences[--i]);
> > > +err_put_fences:
> > > + while (index)
> > > + dma_fence_put(fences[--index]);
> > > kfree(fences);
> > > +
> > > +err_free_sync_file:
> > > fput(sync_file->file);
> > > return NULL;
> > > -
> > > }
> > > static int sync_file_release(struct inode *inode, struct file *file)
> > > @@ -398,11 +396,13 @@ static int sync_fill_fence_info(struct dma_fence *fence,
> > > static long sync_file_ioctl_fence_info(struct sync_file *sync_file,
> > > unsigned long arg)
> > > {
> > > - struct sync_file_info info;
> > > struct sync_fence_info *fence_info = NULL;
> > > - struct dma_fence **fences;
> > > + struct dma_fence_unwrap iter;
> > > + struct sync_file_info info;
> > > + unsigned int num_fences;
> > > + struct dma_fence *fence;
> > > + int ret;
> > > __u32 size;
> > > - int num_fences, ret, i;
> > > if (copy_from_user(&info, (void __user *)arg, sizeof(info)))
> > > return -EFAULT;
> > > @@ -410,7 +410,9 @@ static long sync_file_ioctl_fence_info(struct sync_file *sync_file,
> > > if (info.flags || info.pad)
> > > return -EINVAL;
> > > - fences = get_fences(sync_file, &num_fences);
> > > + num_fences = 0;
> > > + dma_fence_unwrap_for_each(fence, &iter, sync_file->fence)
> > > + ++num_fences;
> > > /*
> > > * Passing num_fences = 0 means that userspace doesn't want to
> > > @@ -433,8 +435,11 @@ static long sync_file_ioctl_fence_info(struct sync_file *sync_file,
> > > if (!fence_info)
> > > return -ENOMEM;
> > > - for (i = 0; i < num_fences; i++) {
> > > - int status = sync_fill_fence_info(fences[i], &fence_info[i]);
> > > + num_fences = 0;
> > > + dma_fence_unwrap_for_each(fence, &iter, sync_file->fence) {
> > > + int status;
> > > +
> > > + status = sync_fill_fence_info(fence, &fence_info[num_fences++]);
> > > info.status = info.status <= 0 ? info.status : status;
> > > }
> > > --
> > > 2.25.1
> > >
>
--
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch
