On Thu, Jun 06, 2019 at 11:41:07PM +0200, Mark Kettenis wrote:
> As a result of a recent discussion with jsg@, I realized that the
> graphics drivers are (mostly) allocating memory from the dma region.
> Since the the graphics stack can potentially gobble up large amounts
> of memory, this means we can run out of dma memory which makes other
> parts of our kernel quite unhappy. Most of the supported hardware
> actually supports 64-bit DMA just fine, and the drivers already have
> code to handle the exceptions. The diff below makes use of this
> knowledge to (hopefully) safely allocate from "high" memory when
> possible. One big change is that this makes bus_dma(9) 64-bit DMA
> aware in the sense that if the BUS_DMA_64BIT flag is used, we skip the
> "not dma-reachable" panic.
>
> It seems to work fine on my Intel Broadwell laptop. I haven't tested
> this on radeon(4) yet. So further testing, especially on systems with
> 4GB of memory or more is necessary.
>
> Please test.
One of the ways dma32 is set in radeon/amdgpu is along the lines of
dma_bits = rdev->need_dma32 ? 32 : 40;
r = pci_set_dma_mask(rdev->pdev, DMA_BIT_MASK(dma_bits));
if (r) {
rdev->need_dma32 = true;
dma_bits = 32;
which we don't handle. That is for the case where the card supports
64 bit dma but the system doesn't? Not something we should be
concerned about?
Having __GFP_DMA32 set the sign bit on the flags seems like it is asking
for trouble. How about 0x40000000 / bit 30 instead of
0x80000000 / bit 31.
Otherwise looks good.
>
>
> Index: arch/amd64/amd64/bus_dma.c
> ===================================================================
> RCS file: /cvs/src/sys/arch/amd64/amd64/bus_dma.c,v
> retrieving revision 1.50
> diff -u -p -r1.50 bus_dma.c
> --- arch/amd64/amd64/bus_dma.c 14 Oct 2017 04:44:43 -0000 1.50
> +++ arch/amd64/amd64/bus_dma.c 6 Jun 2019 21:19:21 -0000
> @@ -319,7 +319,8 @@ _bus_dmamap_load_raw(bus_dma_tag_t t, bu
> if (plen < sgsize)
> sgsize = plen;
>
> - if (paddr > dma_constraint.ucr_high)
> + if (paddr > dma_constraint.ucr_high &&
> + (map->_dm_flags & BUS_DMA_64BIT) == 0)
> panic("Non dma-reachable buffer at paddr
> %#lx(raw)",
> paddr);
>
> @@ -583,7 +584,8 @@ _bus_dmamap_load_buffer(bus_dma_tag_t t,
> */
> pmap_extract(pmap, vaddr, (paddr_t *)&curaddr);
>
> - if (curaddr > dma_constraint.ucr_high)
> + if (curaddr > dma_constraint.ucr_high &&
> + (map->_dm_flags & BUS_DMA_64BIT) == 0)
> panic("Non dma-reachable buffer at curaddr %#lx(raw)",
> curaddr);
>
> Index: dev/pci/drm/drm_linux.c
> ===================================================================
> RCS file: /cvs/src/sys/dev/pci/drm/drm_linux.c,v
> retrieving revision 1.37
> diff -u -p -r1.37 drm_linux.c
> --- dev/pci/drm/drm_linux.c 4 Jun 2019 12:08:22 -0000 1.37
> +++ dev/pci/drm/drm_linux.c 6 Jun 2019 21:19:21 -0000
> @@ -293,16 +293,19 @@ struct vm_page *
> alloc_pages(unsigned int gfp_mask, unsigned int order)
> {
> int flags = (gfp_mask & M_NOWAIT) ? UVM_PLA_NOWAIT : UVM_PLA_WAITOK;
> + struct uvm_constraint_range *constraint = &no_constraint;
> struct pglist mlist;
>
> if (gfp_mask & M_CANFAIL)
> flags |= UVM_PLA_FAILOK;
> if (gfp_mask & M_ZERO)
> flags |= UVM_PLA_ZERO;
> + if (gfp_mask & __GFP_DMA32)
> + constraint = &dma_constraint;
>
> TAILQ_INIT(&mlist);
> - if (uvm_pglistalloc(PAGE_SIZE << order, dma_constraint.ucr_low,
> - dma_constraint.ucr_high, PAGE_SIZE, 0, &mlist, 1, flags))
> + if (uvm_pglistalloc(PAGE_SIZE << order, constraint->ucr_low,
> + constraint->ucr_high, PAGE_SIZE, 0, &mlist, 1, flags))
> return NULL;
> return TAILQ_FIRST(&mlist);
> }
> Index: dev/pci/drm/include/linux/gfp.h
> ===================================================================
> RCS file: /cvs/src/sys/dev/pci/drm/include/linux/gfp.h,v
> retrieving revision 1.1
> diff -u -p -r1.1 gfp.h
> --- dev/pci/drm/include/linux/gfp.h 14 Apr 2019 10:14:53 -0000 1.1
> +++ dev/pci/drm/include/linux/gfp.h 6 Jun 2019 21:19:21 -0000
> @@ -7,24 +7,25 @@
> #include <sys/malloc.h>
> #include <uvm/uvm_extern.h>
>
> -#define GFP_ATOMIC M_NOWAIT
> -#define GFP_NOWAIT M_NOWAIT
> -#define GFP_KERNEL (M_WAITOK | M_CANFAIL)
> -#define GFP_USER (M_WAITOK | M_CANFAIL)
> -#define GFP_TEMPORARY (M_WAITOK | M_CANFAIL)
> -#define GFP_HIGHUSER 0
> -#define GFP_DMA32 0
> -#define __GFP_NOWARN 0
> -#define __GFP_NORETRY 0
> -#define __GFP_ZERO M_ZERO
> +#define __GFP_ZERO M_ZERO
> +#define __GFP_DMA32 0x80000000
> +#define __GFP_NOWARN 0
> +#define __GFP_NORETRY 0
> #define __GFP_RETRY_MAYFAIL 0
> #define __GFP_MOVABLE 0
> #define __GFP_COMP 0
> -#define GFP_TRANSHUGE_LIGHT 0
> #define __GFP_KSWAPD_RECLAIM 0
> #define __GFP_HIGHMEM 0
> #define __GFP_RECLAIMABLE 0
> -#define __GFP_DMA32 0
> +
> +#define GFP_ATOMIC M_NOWAIT
> +#define GFP_NOWAIT M_NOWAIT
> +#define GFP_KERNEL (M_WAITOK | M_CANFAIL)
> +#define GFP_USER (M_WAITOK | M_CANFAIL)
> +#define GFP_TEMPORARY (M_WAITOK | M_CANFAIL)
> +#define GFP_HIGHUSER 0
> +#define GFP_DMA32 __GFP_DMA32
> +#define GFP_TRANSHUGE_LIGHT 0
>
> static inline bool
> gfpflags_allow_blocking(const unsigned int flags)
> Index: dev/pci/drm/ttm/ttm_tt.c
> ===================================================================
> RCS file: /cvs/src/sys/dev/pci/drm/ttm/ttm_tt.c,v
> retrieving revision 1.8
> diff -u -p -r1.8 ttm_tt.c
> --- dev/pci/drm/ttm/ttm_tt.c 14 Apr 2019 10:14:54 -0000 1.8
> +++ dev/pci/drm/ttm/ttm_tt.c 6 Jun 2019 21:19:22 -0000
> @@ -261,6 +261,7 @@ int ttm_dma_tt_init(struct ttm_dma_tt *t
> uint32_t page_flags)
> {
> struct ttm_tt *ttm = &ttm_dma->ttm;
> + int flags = BUS_DMA_WAITOK;
>
> ttm_tt_init_fields(ttm, bo, page_flags);
>
> @@ -276,8 +277,10 @@ int ttm_dma_tt_init(struct ttm_dma_tt *t
>
> ttm_dma->dmat = bo->bdev->dmat;
>
> + if ((page_flags & TTM_PAGE_FLAG_DMA32) == 0)
> + flags |= BUS_DMA_64BIT;
> if (bus_dmamap_create(ttm_dma->dmat, ttm->num_pages << PAGE_SHIFT,
> - ttm->num_pages, ttm->num_pages << PAGE_SHIFT, 0, BUS_DMA_WAITOK,
> + ttm->num_pages, ttm->num_pages << PAGE_SHIFT, 0, flags,
> &ttm_dma->map)) {
> free(ttm_dma->segs, M_DRM, 0);
> ttm_tt_destroy(ttm);
> @@ -293,6 +296,7 @@ int ttm_sg_tt_init(struct ttm_dma_tt *tt
> uint32_t page_flags)
> {
> struct ttm_tt *ttm = &ttm_dma->ttm;
> + int flags = BUS_DMA_WAITOK;
> int ret;
>
> ttm_tt_init_fields(ttm, bo, page_flags);
> @@ -313,8 +317,10 @@ int ttm_sg_tt_init(struct ttm_dma_tt *tt
>
> ttm_dma->dmat = bo->bdev->dmat;
>
> + if ((page_flags & TTM_PAGE_FLAG_DMA32) == 0)
> + flags |= BUS_DMA_64BIT;
> if (bus_dmamap_create(ttm_dma->dmat, ttm->num_pages << PAGE_SHIFT,
> - ttm->num_pages, ttm->num_pages << PAGE_SHIFT, 0, BUS_DMA_WAITOK,
> + ttm->num_pages, ttm->num_pages << PAGE_SHIFT, 0, flags,
> &ttm_dma->map)) {
> free(ttm_dma->segs, M_DRM, 0);
> ttm_tt_destroy(ttm);
>
