Hi Laurent,
On 06/04/18 23:33, Laurent Pinchart wrote:
> Hi Kieran,
>
> Thank you for the patch.
>
> On Thursday, 8 March 2018 02:05:26 EEST Kieran Bingham wrote:
>> Each display list allocates a body to store register values in a dma
>> accessible buffer from a dma_alloc_wc() allocation. Each of these
>> results in an entry in the TLB, and a large number of display list
>
> I'd write it as "IOMMU TLB" to make it clear we're not concerned about CPU MMU
> TLB pressure.
Yes, of course.
>
>> allocations adds pressure to this resource.
>>
>> Reduce TLB pressure on the IPMMUs by allocating multiple display list
>> bodies in a single allocation, and providing these to the display list
>> through a 'body pool'. A pool can be allocated by the display list
>> manager or entities which require their own body allocations.
>>
>> Signed-off-by: Kieran Bingham <kieran.bingham+renesas@xxxxxxxxxxxxxxxx>
>>
>> ---
>> v4:
>> - Provide comment explaining extra allocation on body pool
>> highlighting area for optimisation later.
>>
>> v3:
>> - s/fragment/body/, s/fragments/bodies/
>> - qty -> num_bodies
>> - indentation fix
>> - s/vsp1_dl_body_pool_{alloc,free}/vsp1_dl_body_pool_{create,destroy}/'
>> - Add kerneldoc to non-static functions
>>
>> v2:
>> - assign dlb->dma correctly
>>
>> drivers/media/platform/vsp1/vsp1_dl.c | 163 +++++++++++++++++++++++++++-
>> drivers/media/platform/vsp1/vsp1_dl.h | 8 +-
>> 2 files changed, 171 insertions(+)
>>
>> diff --git a/drivers/media/platform/vsp1/vsp1_dl.c
>> b/drivers/media/platform/vsp1/vsp1_dl.c index 67cc16c1b8e3..0208e72cb356
>> 100644
>> --- a/drivers/media/platform/vsp1/vsp1_dl.c
>> +++ b/drivers/media/platform/vsp1/vsp1_dl.c
>> @@ -45,6 +45,8 @@ struct vsp1_dl_entry {
>> /**
>> * struct vsp1_dl_body - Display list body
>> * @list: entry in the display list list of bodies
>> + * @free: entry in the pool free body list
>
> Could we reuse @list for this purpose ? Unless I'm mistaken, when a body is in
> a pool it doesn't belong to any particular display list, and when it is in a
> display list it isn't in the pool anymore.
I've adapted the @list doc-string to read:
* @list: entry in the display list list of bodies, or body pool free list
Actually, I think I'm tempted to leave this distinct and separate.
If we use the single @list field, then we have calls to vsp1_dl_body_get() which
'might not' add that body to a display list.
Consider the lut_set_table call, if it was called twice without a dl_commit() in
between.
The first call would get a dlb, and populate it with the LUT, setting it as the
active LUT in the entitiy, but not adding it to any list.
The call to vsp1_dl_body_get() *has* to remove it from the free list at that
point. But now the list structure '@list' is poisoned.
Now consider the general case for 'putting' the object back to the pool free
list when it's reference count is zero:
vsp1_dl_body_put() must delete the structure from any existing list and add it
to the pool-free list.
We can use list_move_tail for this:
@@ -286,7 +284,7 @@ void vsp1_dl_body_put(struct vsp1_dl_body *dlb)
dlb->num_entries = 0;
spin_lock_irqsave(&dlb->pool->lock, flags);
- list_add_tail(&dlb->free, &dlb->pool->free);
+ list_move_tail(&dlb->list, &dlb->pool->free);
spin_unlock_irqrestore(&dlb->pool->lock, flags);
However - this fails in the instance above where the dlb did not make it onto a
list, because the call to __list_del_entry() will segfault on the poisoned values.
I don't believe we can expect the callers of vsp1_dl_body_put() to guarantee
that the entity is not on a list either, as that is intrinsic to the refcounting
associated with the object.
I think we have the same issue with the cached body in the vsp1_video object too.
So - leaving this as it is.
>
>> + * @pool: pool to which this body belongs
>> * @vsp1: the VSP1 device
>> * @entries: array of entries
>> * @dma: DMA address of the entries
>> @@ -54,6 +56,9 @@ struct vsp1_dl_entry {
>> */
>> struct vsp1_dl_body {
>> struct list_head list;
>> + struct list_head free;
>> +
>> + struct vsp1_dl_body_pool *pool;
>> struct vsp1_device *vsp1;
>>
>> struct vsp1_dl_entry *entries;
>> @@ -65,6 +70,30 @@ struct vsp1_dl_body {
>> };
>>
>> /**
>> + * struct vsp1_dl_body_pool - display list body pool
>> + * @dma: DMA address of the entries
>> + * @size: size of the full DMA memory pool in bytes
>> + * @mem: CPU memory pointer for the pool
>> + * @bodies: Array of DLB structures for the pool
>> + * @free: List of free DLB entries
>> + * @lock: Protects the pool and free list
>
> The pool and free list ? As far as I can tell the lock only protects the free
> list.
I've removed the reference to the pool for the lock. I don't think much else
needs protecting in the current code, so it should be fine.
>
>> + * @vsp1: the VSP1 device
>> + */
>> +struct vsp1_dl_body_pool {
>> + /* DMA allocation */
>> + dma_addr_t dma;
>> + size_t size;
>> + void *mem;
>> +
>> + /* Body management */
>> + struct vsp1_dl_body *bodies;
>> + struct list_head free;
>> + spinlock_t lock;
>> +
>> + struct vsp1_device *vsp1;
>> +};
>> +
>> +/**
>> * struct vsp1_dl_list - Display list
>> * @list: entry in the display list manager lists
>> * @dlm: the display list manager
>> @@ -105,6 +134,7 @@ enum vsp1_dl_mode {
>> * @active: list currently being processed (loaded) by hardware
>> * @queued: list queued to the hardware (written to the DL registers)
>> * @pending: list waiting to be queued to the hardware
>> + * @pool: body pool for the display list bodies
>> * @gc_work: bodies garbage collector work struct
>> * @gc_bodies: array of display list bodies waiting to be freed
>> */
>> @@ -120,6 +150,8 @@ struct vsp1_dl_manager {
>> struct vsp1_dl_list *queued;
>> struct vsp1_dl_list *pending;
>>
>> + struct vsp1_dl_body_pool *pool;
>> +
>> struct work_struct gc_work;
>> struct list_head gc_bodies;
>> };
>> @@ -128,6 +160,137 @@ struct vsp1_dl_manager {
>> * Display List Body Management
>> */
>>
>> +/**
>> + * vsp1_dl_body_pool_create - Create a pool of bodies from a single
>> allocation
>> + * @vsp1: The VSP1 device
>> + * @num_bodies: The quantity of bodies to allocate
>
> For consistency, s/quantity/number/
Done
>
>> + * @num_entries: The maximum number of entries that the body can contain
>
> Maybe s/the body/a body/ ?
Done
>
>> + * @extra_size: Extra allocation provided for the bodies
>> + *
>> + * Allocate a pool of display list bodies each with enough memory to
>> contain the
>> + * requested number of entries.
>
> How about
>
> the requested number of entries plus the @extra_size.
Done
>
>> + *
>> + * Return a pointer to a pool on success or NULL if memory can't be
>> allocated.
>> + */
>> +struct vsp1_dl_body_pool *
>> +vsp1_dl_body_pool_create(struct vsp1_device *vsp1, unsigned int num_bodies,
>> + unsigned int num_entries, size_t extra_size)
>> +{
>> + struct vsp1_dl_body_pool *pool;
>> + size_t dlb_size;
>> + unsigned int i;
>> +
>> + pool = kzalloc(sizeof(*pool), GFP_KERNEL);
>> + if (!pool)
>> + return NULL;
>> +
>> + pool->vsp1 = vsp1;
>> +
>> + /*
>> + * Todo: 'extra_size' is only used by vsp1_dlm_create(), to allocate
>
> s/Todo/TODO/
>
>> + * extra memory for the display list header. We need only one header per
>> + * display list, not per display list body, thus this allocation is
>> + * extraneous and should be reworked in the future.
>> + */
>
> Any plan to fix this ? :-)
Sort of - The DU interlaced work brings in further need to create pool type
memory - and creates an opportunity to further rework and separate the headers
from the display list body.
(I think something can be built on top of the DU Interlaced patches, they don't
need to be blocked by this TODO:)
But yes, this is in my mind for how to fix it.
Fixed the capitalisation.
>
>> + dlb_size = num_entries * sizeof(struct vsp1_dl_entry) + extra_size;
>> + pool->size = dlb_size * num_bodies;
>> +
>> + pool->bodies = kcalloc(num_bodies, sizeof(*pool->bodies), GFP_KERNEL);
>> + if (!pool->bodies) {
>> + kfree(pool);
>> + return NULL;
>> + }
>> +
>> + pool->mem = dma_alloc_wc(vsp1->bus_master, pool->size, &pool->dma,
>> + GFP_KERNEL);
>> + if (!pool->mem) {
>> + kfree(pool->bodies);
>> + kfree(pool);
>> + return NULL;
>> + }
>> +
>> + spin_lock_init(&pool->lock);
>> + INIT_LIST_HEAD(&pool->free);
>> +
>> + for (i = 0; i < num_bodies; ++i) {
>> + struct vsp1_dl_body *dlb = &pool->bodies[i];
>> +
>> + dlb->pool = pool;
>> + dlb->max_entries = num_entries;
>> +
>> + dlb->dma = pool->dma + i * dlb_size;
>> + dlb->entries = pool->mem + i * dlb_size;
>> +
>> + list_add_tail(&dlb->free, &pool->free);
>> + }
>> +
>> + return pool;
>> +}
>> +
>> +/**
>> + * vsp1_dl_body_pool_destroy - Release a body pool
>> + * @pool: The body pool
>> + *
>> + * Release all components of a pool allocation.
>> + */
>> +void vsp1_dl_body_pool_destroy(struct vsp1_dl_body_pool *pool)
>> +{
>> + if (!pool)
>> + return;
>> +
>> + if (pool->mem)
>> + dma_free_wc(pool->vsp1->bus_master, pool->size, pool->mem,
>> + pool->dma);
>> +
>> + kfree(pool->bodies);
>> + kfree(pool);
>> +}
>> +
>> +/**
>> + * vsp1_dl_body_get - Obtain a body from a pool
>> + * @pool: The body pool
>> + *
>> + * Obtain a body from the pool allocation without blocking.
>
> "the pool allocation" ? Did you mean just "the pool" ?
That sounds better :D Done.
>
>> + *
>> + * Returns a display list body or NULL if there are none available.
>> + */
>> +struct vsp1_dl_body *vsp1_dl_body_get(struct vsp1_dl_body_pool *pool)
>> +{
>> + struct vsp1_dl_body *dlb = NULL;
>> + unsigned long flags;
>> +
>> + spin_lock_irqsave(&pool->lock, flags);
>> +
>> + if (!list_empty(&pool->free)) {
>> + dlb = list_first_entry(&pool->free, struct vsp1_dl_body, free);
>> + list_del(&dlb->free);
>> + }
>> +
>> + spin_unlock_irqrestore(&pool->lock, flags);
>> +
>> + return dlb;
>> +}
>> +
>> +/**
>> + * vsp1_dl_body_put - Return a body back to its pool
>> + * @dlb: The display list body
>> + *
>> + * Return a body back to the pool, and reset the num_entries to clear the
>> list.
>> + */
>> +void vsp1_dl_body_put(struct vsp1_dl_body *dlb)
>> +{
>> + unsigned long flags;
>> +
>> + if (!dlb)
>> + return;
>> +
>> + dlb->num_entries = 0;
>> +
>> + spin_lock_irqsave(&dlb->pool->lock, flags);
>> + list_add_tail(&dlb->free, &dlb->pool->free);
>> + spin_unlock_irqrestore(&dlb->pool->lock, flags);
>> +}
>> +
>> /*
>> * Initialize a display list body object and allocate DMA memory for the
>> body * data. The display list body object is expected to have been
>> initialized to diff --git a/drivers/media/platform/vsp1/vsp1_dl.h
>> b/drivers/media/platform/vsp1/vsp1_dl.h index cf57f986b69a..031032e304d2
>> 100644
>> --- a/drivers/media/platform/vsp1/vsp1_dl.h
>> +++ b/drivers/media/platform/vsp1/vsp1_dl.h
>> @@ -17,6 +17,7 @@
>>
>> struct vsp1_device;
>> struct vsp1_dl_body;
>> +struct vsp1_dl_body_pool;
>> struct vsp1_dl_list;
>> struct vsp1_dl_manager;
>>
>> @@ -34,6 +35,13 @@ void vsp1_dl_list_put(struct vsp1_dl_list *dl);
>> void vsp1_dl_list_write(struct vsp1_dl_list *dl, u32 reg, u32 data);
>> void vsp1_dl_list_commit(struct vsp1_dl_list *dl);
>>
>> +struct vsp1_dl_body_pool *
>> +vsp1_dl_body_pool_create(struct vsp1_device *vsp1, unsigned int num_bodies,
>> + unsigned int num_entries, size_t extra_size);
>> +void vsp1_dl_body_pool_destroy(struct vsp1_dl_body_pool *pool);
>> +struct vsp1_dl_body *vsp1_dl_body_get(struct vsp1_dl_body_pool *pool);
>> +void vsp1_dl_body_put(struct vsp1_dl_body *dlb);
>> +
>> struct vsp1_dl_body *vsp1_dl_body_alloc(struct vsp1_device *vsp1,
>> unsigned int num_entries);
>> void vsp1_dl_body_free(struct vsp1_dl_body *dlb);
>
>
