On Thu, Apr 28, 2022 at 03:28:49PM +0300, Mathias Nyman wrote:
> On 28.4.2022 6.03, Jung Daehwan wrote:
> > On Wed, Apr 27, 2022 at 07:25:21PM +0300, Mathias Nyman wrote:
> >> On 26.4.2022 12.18, Daehwan Jung wrote:
> >>> This driver is for Samsung Exynos xhci host conroller. It uses xhci-plat
> >>> driver mainly and extends some functions by xhci hooks and overrides.
> >>>
> >>> It supports USB Audio offload with Co-processor. It only cares DCBAA,
> >>> Device Context, Transfer Ring, Event Ring, and ERST. They are allocated
> >>> on specific address with xhci hooks. Co-processor could use them directly
> >>> without xhci driver after then.
> >>>
> >>> Signed-off-by: Daehwan Jung <dh10.jung@xxxxxxxxxxx>
> >>
> >> I have to agree with Krzysztof's comments, this is an odd driver stub.
> >>
> >> Perhaps open up a bit how the Exynos offloading works so we can figure out
> >> in more detail what the hardware needs from software.
> >>
> >> (...)
> >
> >>> +static int xhci_alloc_segments_for_ring_uram(struct xhci_hcd *xhci,
> >>> + struct xhci_segment **first, struct xhci_segment **last,
> >>> + unsigned int num_segs, unsigned int cycle_state,
> >>> + enum xhci_ring_type type, unsigned int max_packet, gfp_t flags,
> >>> + u32 endpoint_type)
> >>> +{
> >>> + struct xhci_segment *prev;
> >>> + bool chain_links = false;
> >>> +
> >>> + while (num_segs > 0) {
> >>> + struct xhci_segment *next = NULL;
> >>> +
> >>> + if (!next) {
> >>> + prev = *first;
> >>> + while (prev) {
> >>> + next = prev->next;
> >>> + xhci_segment_free(xhci, prev);
> >>> + prev = next;
> >>> + }
> >>> + return -ENOMEM;
> >>
> >> This always return -ENOMEM
> >
> > Yes. it's right to return error here.
> >
>
> Still don't think that is the case.
>
> So if the num_segs value passed to a function named
> xhci_alloc_segments_for_ring_uram() is anything else than 0, it will
> automatically return -ENOMEM?
>
> >>
> >> Also this whole function never allocates or remaps any memory.
> >
> > This fuctions is for link segments. Right below function(xhci_ring_alloc_uram)
> > allocates.
>
> Still doesn't allocate any ring segments.
> Below function only allocates memory for the
> ring structure that contains pointers to segments.
>
When I re-check it, it has a problem as you said.
I will modify it on next submission. Thanks.
Best Regards,
Jung Daehwan
> >
> >>
> >>> + }
> >>> + xhci_link_segments(prev, next, type, chain_links);
> >>> +
> >>> + prev = next;
> >>> + num_segs--;
> >>> + }
> >>> + xhci_link_segments(prev, *first, type, chain_links);
> >>> + *last = prev;
> >>> +
> >>> + return 0;
> >>> +}
> >>> +
> >>> +static struct xhci_ring *xhci_ring_alloc_uram(struct xhci_hcd *xhci,
> >>> + unsigned int num_segs, unsigned int cycle_state,
> >>> + enum xhci_ring_type type, unsigned int max_packet, gfp_t flags,
> >>> + u32 endpoint_type)
> >>> +{
> >>> + struct xhci_ring *ring;
> >>> + int ret;
> >>> + struct device *dev = xhci_to_hcd(xhci)->self.sysdev;
> >>> +
> >>> + ring = kzalloc_node(sizeof(*ring), flags, dev_to_node(dev));
> >>> + if (!ring)
> >>> + return NULL;
> >>> +
> >>> + ring->num_segs = num_segs;
> >>> + ring->bounce_buf_len = max_packet;
> >>> + INIT_LIST_HEAD(&ring->td_list);
> >>> + ring->type = type;
> >>> + if (num_segs == 0)
> >>> + return ring;
> >>> +
> >>> + ret = xhci_alloc_segments_for_ring_uram(xhci, &ring->first_seg,
> >>> + &ring->last_seg, num_segs, cycle_state, type,
> >>> + max_packet, flags, endpoint_type);
> >>> + if (ret)
> >>> + goto fail;
> >>> +
> >>> + /* Only event ring does not use link TRB */
> >>> + if (type != TYPE_EVENT) {
> >>> + /* See section 4.9.2.1 and 6.4.4.1 */
> >>> + ring->last_seg->trbs[TRBS_PER_SEGMENT - 1].link.control |=
> >>> + cpu_to_le32(LINK_TOGGLE);
> >>
> >> No memory was allocated for trbs
> >
> > Allcation function for trbs are missed. It's done by ioremap.
> > I will add it on next submission. Thanks for the comment.
> >
> >>
> >> A lot of this code seems to exists just to avoid xhci driver from allocating
> >> dma capable memory, we can refactor the existing xhci_mem_init() and move
> >> dcbaa and event ring allocation and other code to their own overridable
> >> functions.
> >>
> >> This way we can probably get rid of a lot of the code in this series.
> >
> > Yes right. I think it's proper. Do you agree with it or have better way
> > to do it?
>
> Could be, but I don't have a good picture of how this Exynos audio offloading
> works, so it's hard to guess.
>
> Thanks
> -Mathias
>
