[PATCH 06/83] drm/radeon: Add kfd-->kgd interfaces of memory allocation/mapping

[Jerome Glisse]

On Fri, Jul 11, 2014 at 12:50:06AM +0300, Oded Gabbay wrote:
> This patch adds new interfaces to kfd2kgd_calls structure.
> 
> The new interfaces allow the kfd driver to :
> 
> 1. Allocated video memory through the radeon driver
> 2. Map and unmap video memory with GPUVM through the radeon driver
> 3. Map and unmap system memory with GPUVM through the radeon driver
> 
> Signed-off-by: Oded Gabbay <oded.gabbay at amd.com>
> ---
>  drivers/gpu/drm/radeon/radeon_kfd.c | 129 ++++++++++++++++++++++++++++++++++++
>  include/linux/radeon_kfd.h          |  23 +++++++
>  2 files changed, 152 insertions(+)
> 
> diff --git a/drivers/gpu/drm/radeon/radeon_kfd.c b/drivers/gpu/drm/radeon/radeon_kfd.c
> index 1b859b5..66ee36b 100644
> --- a/drivers/gpu/drm/radeon/radeon_kfd.c
> +++ b/drivers/gpu/drm/radeon/radeon_kfd.c
> @@ -25,9 +25,31 @@
>  #include <drm/drmP.h>
>  #include "radeon.h"
>  
> +struct kgd_mem {
> +	struct radeon_bo *bo;
> +	u32 domain;
> +};
> +
> +static int allocate_mem(struct kgd_dev *kgd, size_t size, size_t alignment,
> +		enum kgd_memory_pool pool, struct kgd_mem **memory_handle);
> +
> +static void free_mem(struct kgd_dev *kgd, struct kgd_mem *memory_handle);
> +
> +static int gpumap_mem(struct kgd_dev *kgd, struct kgd_mem *mem, uint64_t *vmid0_address);
> +static void ungpumap_mem(struct kgd_dev *kgd, struct kgd_mem *mem);
> +
> +static int kmap_mem(struct kgd_dev *kgd, struct kgd_mem *mem, void **ptr);
> +static void unkmap_mem(struct kgd_dev *kgd, struct kgd_mem *mem);
> +
>  static uint64_t get_vmem_size(struct kgd_dev *kgd);
>  
>  static const struct kfd2kgd_calls kfd2kgd = {
> +	.allocate_mem = allocate_mem,
> +	.free_mem = free_mem,
> +	.gpumap_mem = gpumap_mem,
> +	.ungpumap_mem = ungpumap_mem,
> +	.kmap_mem = kmap_mem,
> +	.unkmap_mem = unkmap_mem,
>  	.get_vmem_size = get_vmem_size,
>  };
>  
> @@ -96,6 +118,113 @@ void radeon_kfd_device_fini(struct radeon_device *rdev)
>  	}
>  }
>  
> +static u32 pool_to_domain(enum kgd_memory_pool p)
> +{
> +	switch (p) {
> +	case KGD_POOL_FRAMEBUFFER: return RADEON_GEM_DOMAIN_VRAM;
> +	default: return RADEON_GEM_DOMAIN_GTT;
> +	}
> +}
> +
> +static int allocate_mem(struct kgd_dev *kgd, size_t size, size_t alignment,
> +		enum kgd_memory_pool pool, struct kgd_mem **memory_handle)
> +{
> +	struct radeon_device *rdev = (struct radeon_device *)kgd;
> +	struct kgd_mem *mem;
> +	int r;
> +
> +	mem = kzalloc(sizeof(struct kgd_mem), GFP_KERNEL);
> +	if (!mem)
> +		return -ENOMEM;
> +
> +	mem->domain = pool_to_domain(pool);
> +
> +	r = radeon_bo_create(rdev, size, alignment, true, mem->domain, NULL, &mem->bo);
> +	if (r) {
> +		kfree(mem);
> +		return r;
> +	}
> +
> +	*memory_handle = mem;
> +	return 0;
> +}
> +
> +static void free_mem(struct kgd_dev *kgd, struct kgd_mem *mem)
> +{
> +	/* Assume that KFD will never free gpumapped or kmapped memory. This is not quite settled. */
> +	radeon_bo_unref(&mem->bo);
> +	kfree(mem);
> +}
> +
> +static int gpumap_mem(struct kgd_dev *kgd, struct kgd_mem *mem, uint64_t *vmid0_address)
> +{
> +	int r;
> +
> +	r = radeon_bo_reserve(mem->bo, true);
> +
> +	/*
> +	 * ttm_bo_reserve can only fail if the buffer reservation lock
> +	 * is held in circumstances that would deadlock
> +	 */
> +	BUG_ON(r != 0);
> +	r = radeon_bo_pin(mem->bo, mem->domain, vmid0_address);
> +	radeon_bo_unreserve(mem->bo);
> +
> +	return r;
> +}

What is lifetime of such object ? Are they limited in size and number ? How
can the GFX side ie radeon force unmap them ?

Because pining is a big NO we only pin a handfull of buffer and the only
thing we allow userspace to pin are front buffer associated with crtc
which means there is a limited number of such buffer and there is a
legitimate use for pining them.

To me this looks like anyone can pin vram and thus starve the GFX side,
read DDOS.

Cheers,
Jérôme

> +
> +static void ungpumap_mem(struct kgd_dev *kgd, struct kgd_mem *mem)
> +{
> +	int r;
> +
> +	r = radeon_bo_reserve(mem->bo, true);
> +
> +	/*
> +	 * ttm_bo_reserve can only fail if the buffer reservation lock
> +	 * is held in circumstances that would deadlock
> +	 */
> +	BUG_ON(r != 0);
> +	r = radeon_bo_unpin(mem->bo);
> +
> +	/*
> +	 * This unpin only removed NO_EVICT placement flags
> +	 * and should never fail
> +	 */
> +	BUG_ON(r != 0);
> +	radeon_bo_unreserve(mem->bo);
> +}
> +
> +static int kmap_mem(struct kgd_dev *kgd, struct kgd_mem *mem, void **ptr)
> +{
> +	int r;
> +
> +	r = radeon_bo_reserve(mem->bo, true);
> +
> +	/*
> +	 * ttm_bo_reserve can only fail if the buffer reservation lock
> +	 * is held in circumstances that would deadlock
> +	 */
> +	BUG_ON(r != 0);
> +	r = radeon_bo_kmap(mem->bo, ptr);
> +	radeon_bo_unreserve(mem->bo);
> +
> +	return r;
> +}
> +
> +static void unkmap_mem(struct kgd_dev *kgd, struct kgd_mem *mem)
> +{
> +	int r;
> +
> +	r = radeon_bo_reserve(mem->bo, true);
> +	/*
> +	 * ttm_bo_reserve can only fail if the buffer reservation lock
> +	 * is held in circumstances that would deadlock
> +	 */
> +	BUG_ON(r != 0);
> +	radeon_bo_kunmap(mem->bo);
> +	radeon_bo_unreserve(mem->bo);
> +}
> +
>  static uint64_t get_vmem_size(struct kgd_dev *kgd)
>  {
>  	struct radeon_device *rdev = (struct radeon_device *)kgd;
> diff --git a/include/linux/radeon_kfd.h b/include/linux/radeon_kfd.h
> index 28cddf5..c7997d4 100644
> --- a/include/linux/radeon_kfd.h
> +++ b/include/linux/radeon_kfd.h
> @@ -36,6 +36,14 @@ struct pci_dev;
>  struct kfd_dev;
>  struct kgd_dev;
>  
> +struct kgd_mem;
> +
> +enum kgd_memory_pool {
> +	KGD_POOL_SYSTEM_CACHEABLE = 1,
> +	KGD_POOL_SYSTEM_WRITECOMBINE = 2,
> +	KGD_POOL_FRAMEBUFFER = 3,
> +};
> +
>  struct kgd2kfd_shared_resources {
>  	void __iomem *mmio_registers; /* Mapped pointer to GFX MMIO registers. */
>  
> @@ -57,6 +65,21 @@ struct kgd2kfd_calls {
>  };
>  
>  struct kfd2kgd_calls {
> +	/* Memory management. */
> +	int (*allocate_mem)(struct kgd_dev *kgd,
> +				size_t size,
> +				size_t alignment,
> +				enum kgd_memory_pool pool,
> +				struct kgd_mem **memory_handle);
> +
> +	void (*free_mem)(struct kgd_dev *kgd, struct kgd_mem *memory_handle);
> +
> +	int (*gpumap_mem)(struct kgd_dev *kgd, struct kgd_mem *mem, uint64_t *vmid0_address);
> +	void (*ungpumap_mem)(struct kgd_dev *kgd, struct kgd_mem *mem);
> +
> +	int (*kmap_mem)(struct kgd_dev *kgd, struct kgd_mem *mem, void **ptr);
> +	void (*unkmap_mem)(struct kgd_dev *kgd, struct kgd_mem *mem);
> +
>  	uint64_t (*get_vmem_size)(struct kgd_dev *kgd);
>  };
>  
> -- 
> 1.9.1
>