On Fri, May 11, 2018 at 6:57 PM, Felix Kuehling <felix.kuehling at amd.com> wrote: > On 2018-05-11 04:52 AM, Oded Gabbay wrote: >>> --- a/drivers/gpu/drm/amd/amdkfd/kfd_priv.h >>> +++ b/drivers/gpu/drm/amd/amdkfd/kfd_priv.h >>> @@ -41,9 +41,33 @@ >>> >>> #define KFD_SYSFS_FILE_MODE 0444 >>> >>> -#define KFD_MMAP_DOORBELL_MASK 0x8000000000000ull >>> -#define KFD_MMAP_EVENTS_MASK 0x4000000000000ull >>> -#define KFD_MMAP_RESERVED_MEM_MASK 0x2000000000000ull >>> +/* GPU ID hash width in bits */ >>> +#define KFD_GPU_ID_HASH_WIDTH 16 >>> + >>> +/* Use upper bits of mmap offset to store KFD driver specific information. >>> + * BITS[63:62] - Encode MMAP type >>> + * BITS[61:46] - Encode gpu_id. To identify to which GPU the offset belongs to >>> + * BITS[45:0] - MMAP offset value >>> + * >>> + * NOTE: struct vm_area_struct.vm_pgoff uses offset in pages. Hence, these >>> + * defines are w.r.t to PAGE_SIZE >>> + */ >>> +#define KFD_MMAP_TYPE_SHIFT (62 - PAGE_SHIFT) >>> +#define KFD_MMAP_TYPE_MASK (0x3ULL << KFD_MMAP_TYPE_SHIFT) >>> +#define KFD_MMAP_TYPE_DOORBELL (0x3ULL << KFD_MMAP_TYPE_SHIFT) >>> +#define KFD_MMAP_TYPE_EVENTS (0x2ULL << KFD_MMAP_TYPE_SHIFT) >>> +#define KFD_MMAP_TYPE_RESERVED_MEM (0x1ULL << KFD_MMAP_TYPE_SHIFT) >> Isn't this new definition breaks existing user-space library (kfd thunk) ? >> If that is the case we have a problem here. > > No, this does not break user mode, because user mode isn't aware of > these definitions at all. The mmap offset comes from kernel mode and is > opaque to user mode. User mode just passes it back down to kernel mode > in the offset parameter of mmap. So the kernel can change the encoding > of the mmap offset without breaking user mode. > Oops, you are correct of course, I forgot. That's what happens after 4 years of not seeing that particular code :) Oded > Regards, > Felix > >> >> Oded >> >> >
