From fbd2895e45aebdb3d3ea73a3a796cf3bb9c912da Mon Sep 17 00:00:00 2001 From: Ben Skeggs Date: Wed, 1 Sep 2010 15:24:34 +1000 Subject: [PATCH] drm/nouveau: rework init ordering so nv50_instmem.c can be less bad Reviewed-by: Francisco Jerez Signed-off-by: Ben Skeggs --- drivers/gpu/drm/nouveau/nouveau_bo.c | 5 +- drivers/gpu/drm/nouveau/nouveau_drv.h | 6 +- drivers/gpu/drm/nouveau/nouveau_mem.c | 75 +++-- drivers/gpu/drm/nouveau/nouveau_object.c | 27 +- drivers/gpu/drm/nouveau/nouveau_state.c | 41 +-- drivers/gpu/drm/nouveau/nv04_instmem.c | 51 +--- drivers/gpu/drm/nouveau/nv50_instmem.c | 341 ++++++++++------------- 7 files changed, 243 insertions(+), 303 deletions(-) diff --git a/drivers/gpu/drm/nouveau/nouveau_bo.c b/drivers/gpu/drm/nouveau/nouveau_bo.c index 841c63f2886..8e4a9bce4f3 100644 --- a/drivers/gpu/drm/nouveau/nouveau_bo.c +++ b/drivers/gpu/drm/nouveau/nouveau_bo.c @@ -403,7 +403,10 @@ nouveau_bo_init_mem_type(struct ttm_bo_device *bdev, uint32_t type, man->available_caching = TTM_PL_FLAG_UNCACHED | TTM_PL_FLAG_WC; man->default_caching = TTM_PL_FLAG_WC; - man->gpu_offset = dev_priv->vm_vram_base; + if (dev_priv->card_type == NV_50) + man->gpu_offset = 0x40000000; + else + man->gpu_offset = 0; break; case TTM_PL_TT: switch (dev_priv->gart_info.type) { diff --git a/drivers/gpu/drm/nouveau/nouveau_drv.h b/drivers/gpu/drm/nouveau/nouveau_drv.h index 150fdeea11a..3ba7a649fe5 100644 --- a/drivers/gpu/drm/nouveau/nouveau_drv.h +++ b/drivers/gpu/drm/nouveau/nouveau_drv.h @@ -700,8 +700,10 @@ extern bool nouveau_wait_for_idle(struct drm_device *); extern int nouveau_card_init(struct drm_device *); /* nouveau_mem.c */ -extern int nouveau_mem_detect(struct drm_device *dev); -extern int nouveau_mem_init(struct drm_device *); +extern int nouveau_mem_vram_init(struct drm_device *); +extern void nouveau_mem_vram_fini(struct drm_device *); +extern int nouveau_mem_gart_init(struct drm_device *); +extern void nouveau_mem_gart_fini(struct drm_device *); extern int nouveau_mem_init_agp(struct drm_device *); extern int nouveau_mem_reset_agp(struct drm_device *); extern void nouveau_mem_close(struct drm_device *); diff --git a/drivers/gpu/drm/nouveau/nouveau_mem.c b/drivers/gpu/drm/nouveau/nouveau_mem.c index 6eeaeac5629..fb15a1b0dda 100644 --- a/drivers/gpu/drm/nouveau/nouveau_mem.c +++ b/drivers/gpu/drm/nouveau/nouveau_mem.c @@ -221,7 +221,7 @@ nv50_mem_vm_unbind(struct drm_device *dev, uint64_t virt, uint32_t size) * Cleanup everything */ void -nouveau_mem_close(struct drm_device *dev) +nouveau_mem_vram_fini(struct drm_device *dev) { struct drm_nouveau_private *dev_priv = dev->dev_private; @@ -232,6 +232,19 @@ nouveau_mem_close(struct drm_device *dev) nouveau_ttm_global_release(dev_priv); + if (dev_priv->fb_mtrr >= 0) { + drm_mtrr_del(dev_priv->fb_mtrr, + pci_resource_start(dev->pdev, 1), + pci_resource_len(dev->pdev, 1), DRM_MTRR_WC); + dev_priv->fb_mtrr = -1; + } +} + +void +nouveau_mem_gart_fini(struct drm_device *dev) +{ + nouveau_sgdma_takedown(dev); + if (drm_core_has_AGP(dev) && dev->agp) { struct drm_agp_mem *entry, *tempe; @@ -251,13 +264,6 @@ nouveau_mem_close(struct drm_device *dev) dev->agp->acquired = 0; dev->agp->enabled = 0; } - - if (dev_priv->fb_mtrr) { - drm_mtrr_del(dev_priv->fb_mtrr, - pci_resource_start(dev->pdev, 1), - pci_resource_len(dev->pdev, 1), DRM_MTRR_WC); - dev_priv->fb_mtrr = -1; - } } static uint32_t @@ -363,7 +369,7 @@ nvaa_vram_preinit(struct drm_device *dev) dev_priv->vram_rblock_size = 1; } -int +static int nouveau_mem_detect(struct drm_device *dev) { struct drm_nouveau_private *dev_priv = dev->dev_private; @@ -500,24 +506,27 @@ nouveau_mem_init_agp(struct drm_device *dev) } int -nouveau_mem_init(struct drm_device *dev) +nouveau_mem_vram_init(struct drm_device *dev) { struct drm_nouveau_private *dev_priv = dev->dev_private; struct ttm_bo_device *bdev = &dev_priv->ttm.bdev; - int ret, dma_bits = 32; - - dev_priv->fb_phys = pci_resource_start(dev->pdev, 1); - dev_priv->gart_info.type = NOUVEAU_GART_NONE; + int ret, dma_bits; if (dev_priv->card_type >= NV_50 && pci_dma_supported(dev->pdev, DMA_BIT_MASK(40))) dma_bits = 40; + else + dma_bits = 32; ret = pci_set_dma_mask(dev->pdev, DMA_BIT_MASK(dma_bits)); - if (ret) { - NV_ERROR(dev, "Error setting DMA mask: %d\n", ret); + if (ret) return ret; - } + + ret = nouveau_mem_detect(dev); + if (ret) + return ret; + + dev_priv->fb_phys = pci_resource_start(dev->pdev, 1); ret = nouveau_ttm_global_init(dev_priv); if (ret) @@ -541,7 +550,16 @@ nouveau_mem_init(struct drm_device *dev) pci_resource_len(dev->pdev, 1); dev_priv->fb_mappable_pages >>= PAGE_SHIFT; - /* remove reserved space at end of vram from available amount */ + /* reserve space at end of VRAM for PRAMIN */ + if (dev_priv->chipset == 0x40 || dev_priv->chipset == 0x47 || + dev_priv->chipset == 0x49 || dev_priv->chipset == 0x4b) + dev_priv->ramin_rsvd_vram = (2 * 1024 * 1024); + else + if (dev_priv->card_type >= NV_40) + dev_priv->ramin_rsvd_vram = (1 * 1024 * 1024); + else + dev_priv->ramin_rsvd_vram = (512 * 1024); + dev_priv->fb_available_size -= dev_priv->ramin_rsvd_vram; dev_priv->fb_aper_free = dev_priv->fb_available_size; @@ -562,7 +580,21 @@ nouveau_mem_init(struct drm_device *dev) nouveau_bo_ref(NULL, &dev_priv->vga_ram); } - /* GART */ + dev_priv->fb_mtrr = drm_mtrr_add(pci_resource_start(dev->pdev, 1), + pci_resource_len(dev->pdev, 1), + DRM_MTRR_WC); + return 0; +} + +int +nouveau_mem_gart_init(struct drm_device *dev) +{ + struct drm_nouveau_private *dev_priv = dev->dev_private; + struct ttm_bo_device *bdev = &dev_priv->ttm.bdev; + int ret; + + dev_priv->gart_info.type = NOUVEAU_GART_NONE; + #if !defined(__powerpc__) && !defined(__ia64__) if (drm_device_is_agp(dev) && dev->agp && !nouveau_noagp) { ret = nouveau_mem_init_agp(dev); @@ -590,11 +622,6 @@ nouveau_mem_init(struct drm_device *dev) return ret; } - dev_priv->fb_mtrr = drm_mtrr_add(pci_resource_start(dev->pdev, 1), - pci_resource_len(dev->pdev, 1), - DRM_MTRR_WC); - return 0; } - diff --git a/drivers/gpu/drm/nouveau/nouveau_object.c b/drivers/gpu/drm/nouveau/nouveau_object.c index df445fcb832..b68922f2fe5 100644 --- a/drivers/gpu/drm/nouveau/nouveau_object.c +++ b/drivers/gpu/drm/nouveau/nouveau_object.c @@ -130,7 +130,7 @@ nouveau_gpuobj_new(struct drm_device *dev, struct nouveau_channel *chan, /* if we got a chunk of the aperture, map pages into it */ gpuobj->im_pramin = ramin; - if (!chan && gpuobj->im_pramin) { + if (!chan && gpuobj->im_pramin && dev_priv->ramin_available) { ret = engine->instmem.bind(dev, gpuobj); if (ret) { nouveau_gpuobj_ref(NULL, &gpuobj); @@ -173,7 +173,7 @@ nouveau_gpuobj_new(struct drm_device *dev, struct nouveau_channel *chan, } int -nouveau_gpuobj_early_init(struct drm_device *dev) +nouveau_gpuobj_init(struct drm_device *dev) { struct drm_nouveau_private *dev_priv = dev->dev_private; @@ -186,29 +186,6 @@ nouveau_gpuobj_early_init(struct drm_device *dev) return 0; } -int -nouveau_gpuobj_init(struct drm_device *dev) -{ - struct drm_nouveau_private *dev_priv = dev->dev_private; - struct nouveau_gpuobj *ramht = NULL; - int ret; - - NV_DEBUG(dev, "\n"); - - if (dev_priv->card_type >= NV_50) - return 0; - - ret = nouveau_gpuobj_new_fake(dev, dev_priv->ramht_offset, ~0, - dev_priv->ramht_size, - NVOBJ_FLAG_ZERO_ALLOC, &ramht); - if (ret) - return ret; - - ret = nouveau_ramht_new(dev, ramht, &dev_priv->ramht); - nouveau_gpuobj_ref(NULL, &ramht); - return ret; -} - void nouveau_gpuobj_takedown(struct drm_device *dev) { diff --git a/drivers/gpu/drm/nouveau/nouveau_state.c b/drivers/gpu/drm/nouveau/nouveau_state.c index fec29522298..19eb06dca89 100644 --- a/drivers/gpu/drm/nouveau/nouveau_state.c +++ b/drivers/gpu/drm/nouveau/nouveau_state.c @@ -532,35 +532,26 @@ nouveau_card_init(struct drm_device *dev) if (ret) goto out_display_early; - ret = nouveau_mem_detect(dev); + ret = nouveau_mem_vram_init(dev); if (ret) goto out_bios; - ret = nouveau_gpuobj_early_init(dev); + ret = nouveau_gpuobj_init(dev); if (ret) - goto out_bios; + goto out_vram; - /* Initialise instance memory, must happen before mem_init so we - * know exactly how much VRAM we're able to use for "normal" - * purposes. - */ ret = engine->instmem.init(dev); if (ret) - goto out_gpuobj_early; + goto out_gpuobj; - /* Setup the memory manager */ - ret = nouveau_mem_init(dev); + ret = nouveau_mem_gart_init(dev); if (ret) goto out_instmem; - ret = nouveau_gpuobj_init(dev); - if (ret) - goto out_mem; - /* PMC */ ret = engine->mc.init(dev); if (ret) - goto out_gpuobj; + goto out_gart; /* PGPIO */ ret = engine->gpio.init(dev); @@ -640,15 +631,14 @@ out_gpio: engine->gpio.takedown(dev); out_mc: engine->mc.takedown(dev); -out_gpuobj: - nouveau_gpuobj_takedown(dev); -out_mem: - nouveau_sgdma_takedown(dev); - nouveau_mem_close(dev); +out_gart: + nouveau_mem_gart_fini(dev); out_instmem: engine->instmem.takedown(dev); -out_gpuobj_early: - nouveau_gpuobj_late_takedown(dev); +out_gpuobj: + nouveau_gpuobj_takedown(dev); +out_vram: + nouveau_mem_vram_fini(dev); out_bios: nouveau_bios_takedown(dev); out_display_early: @@ -684,15 +674,14 @@ static void nouveau_card_takedown(struct drm_device *dev) ttm_bo_clean_mm(&dev_priv->ttm.bdev, TTM_PL_VRAM); ttm_bo_clean_mm(&dev_priv->ttm.bdev, TTM_PL_TT); mutex_unlock(&dev->struct_mutex); - nouveau_sgdma_takedown(dev); + nouveau_mem_gart_fini(dev); - nouveau_gpuobj_takedown(dev); - nouveau_mem_close(dev); engine->instmem.takedown(dev); + nouveau_gpuobj_takedown(dev); + nouveau_mem_vram_fini(dev); drm_irq_uninstall(dev); - nouveau_gpuobj_late_takedown(dev); nouveau_bios_takedown(dev); vga_client_register(dev->pdev, NULL, NULL, NULL); diff --git a/drivers/gpu/drm/nouveau/nv04_instmem.c b/drivers/gpu/drm/nouveau/nv04_instmem.c index 3aba7674560..15cd468f4c2 100644 --- a/drivers/gpu/drm/nouveau/nv04_instmem.c +++ b/drivers/gpu/drm/nouveau/nv04_instmem.c @@ -1,6 +1,7 @@ #include "drmP.h" #include "drm.h" #include "nouveau_drv.h" +#include "nouveau_ramht.h" /* returns the size of fifo context */ static int @@ -17,42 +18,6 @@ nouveau_fifo_ctx_size(struct drm_device *dev) return 32; } -static void -nv04_instmem_determine_amount(struct drm_device *dev) -{ - struct drm_nouveau_private *dev_priv = dev->dev_private; - int i; - - /* Figure out how much instance memory we need */ - if (dev_priv->card_type >= NV_40) { - /* We'll want more instance memory than this on some NV4x cards. - * There's a 16MB aperture to play with that maps onto the end - * of vram. For now, only reserve a small piece until we know - * more about what each chipset requires. - */ - switch (dev_priv->chipset) { - case 0x40: - case 0x47: - case 0x49: - case 0x4b: - dev_priv->ramin_rsvd_vram = (2 * 1024 * 1024); - break; - default: - dev_priv->ramin_rsvd_vram = (1 * 1024 * 1024); - break; - } - } else { - /*XXX: what *are* the limits on ramin_rsvd_vram = (512 * 1024); - } - NV_DEBUG(dev, "RAMIN size: %dKiB\n", dev_priv->ramin_rsvd_vram >> 10); - - /* Clear all of it, except the BIOS image that's in the first 64KiB */ - for (i = 64 * 1024; i < dev_priv->ramin_rsvd_vram; i += 4) - nv_wi32(dev, i, 0x00000000); -} - static void nv04_instmem_configure_fixed_tables(struct drm_device *dev) { @@ -103,12 +68,24 @@ nv04_instmem_configure_fixed_tables(struct drm_device *dev) int nv04_instmem_init(struct drm_device *dev) { struct drm_nouveau_private *dev_priv = dev->dev_private; + struct nouveau_gpuobj *ramht = NULL; uint32_t offset; int ret; - nv04_instmem_determine_amount(dev); nv04_instmem_configure_fixed_tables(dev); + /* Setup shared RAMHT */ + ret = nouveau_gpuobj_new_fake(dev, dev_priv->ramht_offset, ~0, + dev_priv->ramht_size, + NVOBJ_FLAG_ZERO_ALLOC, &ramht); + if (ret) + return ret; + + ret = nouveau_ramht_new(dev, ramht, &dev_priv->ramht); + nouveau_gpuobj_ref(NULL, &ramht); + if (ret) + return ret; + /* Create a heap to manage RAMIN allocations, we don't allocate * the space that was reserved for RAMHT/FC/RO. */ diff --git a/drivers/gpu/drm/nouveau/nv50_instmem.c b/drivers/gpu/drm/nouveau/nv50_instmem.c index 5c617f807b2..d932594449c 100644 --- a/drivers/gpu/drm/nouveau/nv50_instmem.c +++ b/drivers/gpu/drm/nouveau/nv50_instmem.c @@ -37,27 +37,82 @@ struct nv50_instmem_priv { struct nouveau_gpuobj *fb_bar; }; -#define NV50_INSTMEM_PAGE_SHIFT 12 -#define NV50_INSTMEM_PAGE_SIZE (1 << NV50_INSTMEM_PAGE_SHIFT) -#define NV50_INSTMEM_PT_SIZE(a) (((a) >> 12) << 3) +static void +nv50_channel_del(struct nouveau_channel **pchan) +{ + struct nouveau_channel *chan; -/*NOTE: - Assumes 0x1700 already covers the correct MiB of PRAMIN - */ -#define BAR0_WI32(g, o, v) do { \ - u32 offset = (g)->vinst + (o); \ - nv_wr32(dev, NV_RAMIN + (offset & 0xfffff), (v)); \ -} while (0) + chan = *pchan; + *pchan = NULL; + if (!chan) + return; + + nouveau_gpuobj_ref(NULL, &chan->ramfc); + nouveau_gpuobj_ref(NULL, &chan->vm_pd); + if (chan->ramin_heap.free_stack.next) + drm_mm_takedown(&chan->ramin_heap); + nouveau_gpuobj_ref(NULL, &chan->ramin); + kfree(chan); +} + +static int +nv50_channel_new(struct drm_device *dev, u32 size, + struct nouveau_channel **pchan) +{ + struct drm_nouveau_private *dev_priv = dev->dev_private; + u32 pgd = (dev_priv->chipset == 0x50) ? 0x1400 : 0x0200; + u32 fc = (dev_priv->chipset == 0x50) ? 0x0000 : 0x4200; + struct nouveau_channel *chan; + int ret; + + chan = kzalloc(sizeof(*chan), GFP_KERNEL); + if (!chan) + return -ENOMEM; + chan->dev = dev; + + ret = nouveau_gpuobj_new(dev, NULL, size, 0x1000, 0, &chan->ramin); + if (ret) { + nv50_channel_del(&chan); + return ret; + } + + ret = drm_mm_init(&chan->ramin_heap, 0x6000, chan->ramin->size); + if (ret) { + nv50_channel_del(&chan); + return ret; + } + + ret = nouveau_gpuobj_new_fake(dev, chan->ramin->pinst == ~0 ? ~0 : + chan->ramin->pinst + pgd, + chan->ramin->vinst + pgd, + 0x4000, NVOBJ_FLAG_ZERO_ALLOC, + &chan->vm_pd); + if (ret) { + nv50_channel_del(&chan); + return ret; + } + + ret = nouveau_gpuobj_new_fake(dev, chan->ramin->pinst == ~0 ? ~0 : + chan->ramin->pinst + fc, + chan->ramin->vinst + fc, 0x100, + NVOBJ_FLAG_ZERO_ALLOC, &chan->ramfc); + if (ret) { + nv50_channel_del(&chan); + return ret; + } + + *pchan = chan; + return 0; +} int nv50_instmem_init(struct drm_device *dev) { struct drm_nouveau_private *dev_priv = dev->dev_private; - struct nouveau_channel *chan; - uint32_t c_offset, c_size, c_ramfc, c_vmpd, c_base, pt_size; - uint32_t save_nv001700; - uint64_t v; struct nv50_instmem_priv *priv; + struct nouveau_channel *chan; int ret, i; + u32 tmp; priv = kzalloc(sizeof(*priv), GFP_KERNEL); if (!priv) @@ -68,206 +123,113 @@ nv50_instmem_init(struct drm_device *dev) for (i = 0x1700; i <= 0x1710; i += 4) priv->save1700[(i-0x1700)/4] = nv_rd32(dev, i); - /* Reserve the last MiB of VRAM, we should probably try to avoid - * setting up the below tables over the top of the VBIOS image at - * some point. - */ - dev_priv->ramin_rsvd_vram = 1 << 20; - c_offset = dev_priv->vram_size - dev_priv->ramin_rsvd_vram; - c_size = 128 << 10; - c_vmpd = ((dev_priv->chipset & 0xf0) == 0x50) ? 0x1400 : 0x200; - c_ramfc = ((dev_priv->chipset & 0xf0) == 0x50) ? 0x0 : 0x20; - c_base = c_vmpd + 0x4000; - pt_size = NV50_INSTMEM_PT_SIZE(dev_priv->ramin_size); - - NV_DEBUG(dev, " Rsvd VRAM base: 0x%08x\n", c_offset); - NV_DEBUG(dev, " VBIOS image: 0x%08x\n", - (nv_rd32(dev, 0x619f04) & ~0xff) << 8); - NV_DEBUG(dev, " Aperture size: %d MiB\n", dev_priv->ramin_size >> 20); - NV_DEBUG(dev, " PT size: %d KiB\n", pt_size >> 10); - - /* Determine VM layout, we need to do this first to make sure - * we allocate enough memory for all the page tables. - */ - dev_priv->vm_gart_base = roundup(NV50_VM_BLOCK, NV50_VM_BLOCK); - dev_priv->vm_gart_size = NV50_VM_BLOCK; - - dev_priv->vm_vram_base = dev_priv->vm_gart_base + dev_priv->vm_gart_size; - dev_priv->vm_vram_size = dev_priv->vram_size; - if (dev_priv->vm_vram_size > NV50_VM_MAX_VRAM) - dev_priv->vm_vram_size = NV50_VM_MAX_VRAM; - dev_priv->vm_vram_size = roundup(dev_priv->vm_vram_size, NV50_VM_BLOCK); - dev_priv->vm_vram_pt_nr = dev_priv->vm_vram_size / NV50_VM_BLOCK; - - dev_priv->vm_end = dev_priv->vm_vram_base + dev_priv->vm_vram_size; - - NV_DEBUG(dev, "NV50VM: GART 0x%016llx-0x%016llx\n", - dev_priv->vm_gart_base, - dev_priv->vm_gart_base + dev_priv->vm_gart_size - 1); - NV_DEBUG(dev, "NV50VM: VRAM 0x%016llx-0x%016llx\n", - dev_priv->vm_vram_base, - dev_priv->vm_vram_base + dev_priv->vm_vram_size - 1); - - c_size += dev_priv->vm_vram_pt_nr * (NV50_VM_BLOCK / 65536 * 8); - - /* Map BAR0 PRAMIN aperture over the memory we want to use */ - save_nv001700 = nv_rd32(dev, NV50_PUNK_BAR0_PRAMIN); - nv_wr32(dev, NV50_PUNK_BAR0_PRAMIN, (c_offset >> 16)); - - /* Create a fake channel, and use it as our "dummy" channels 0/127. - * The main reason for creating a channel is so we can use the gpuobj - * code. However, it's probably worth noting that NVIDIA also setup - * their channels 0/127 with the same values they configure here. - * So, there may be some other reason for doing this. - * - * Have to create the entire channel manually, as the real channel - * creation code assumes we have PRAMIN access, and we don't until - * we're done here. - */ - chan = kzalloc(sizeof(*chan), GFP_KERNEL); - if (!chan) + /* Global PRAMIN heap */ + ret = drm_mm_init(&dev_priv->ramin_heap, 0, dev_priv->ramin_size); + if (ret) { + NV_ERROR(dev, "Failed to init RAMIN heap\n"); return -ENOMEM; - chan->id = 0; - chan->dev = dev; - chan->file_priv = (struct drm_file *)-2; - dev_priv->fifos[0] = dev_priv->fifos[127] = chan; - - INIT_LIST_HEAD(&chan->ramht_refs); + } - /* Channel's PRAMIN object + heap */ - ret = nouveau_gpuobj_new_fake(dev, 0, c_offset, c_size, 0, &chan->ramin); + /* we need a channel to plug into the hw to control the BARs */ + ret = nv50_channel_new(dev, 128*1024, &dev_priv->fifos[0]); if (ret) return ret; + chan = dev_priv->fifos[127] = dev_priv->fifos[0]; - if (drm_mm_init(&chan->ramin_heap, c_base, c_size - c_base)) - return -ENOMEM; - - /* RAMFC + zero channel's PRAMIN up to start of VM pagedir */ - ret = nouveau_gpuobj_new_fake(dev, c_ramfc, c_offset + c_ramfc, - 0x4000, 0, &chan->ramfc); + /* allocate page table for PRAMIN BAR */ + ret = nouveau_gpuobj_new(dev, chan, (dev_priv->ramin_size >> 12) * 8, + 0x1000, NVOBJ_FLAG_ZERO_ALLOC, + &priv->pramin_pt); if (ret) return ret; - for (i = 0; i < c_vmpd; i += 4) - BAR0_WI32(chan->ramin, i, 0); + nv_wo32(chan->vm_pd, 0x0000, priv->pramin_pt->vinst | 0x63); + nv_wo32(chan->vm_pd, 0x0004, 0); - /* VM page directory */ - ret = nouveau_gpuobj_new_fake(dev, c_vmpd, c_offset + c_vmpd, - 0x4000, 0, &chan->vm_pd); + /* DMA object for PRAMIN BAR */ + ret = nouveau_gpuobj_new(dev, chan, 6*4, 16, 0, &priv->pramin_bar); if (ret) return ret; - for (i = 0; i < 0x4000; i += 8) { - BAR0_WI32(chan->vm_pd, i + 0x00, 0x00000000); - BAR0_WI32(chan->vm_pd, i + 0x04, 0x00000000); - } - - /* PRAMIN page table, cheat and map into VM at 0x0000000000. - * We map the entire fake channel into the start of the PRAMIN BAR - */ - ret = nouveau_gpuobj_new(dev, chan, pt_size, 0x1000, 0, - &priv->pramin_pt); + nv_wo32(priv->pramin_bar, 0x00, 0x7fc00000); + nv_wo32(priv->pramin_bar, 0x04, dev_priv->ramin_size - 1); + nv_wo32(priv->pramin_bar, 0x08, 0x00000000); + nv_wo32(priv->pramin_bar, 0x0c, 0x00000000); + nv_wo32(priv->pramin_bar, 0x10, 0x00000000); + nv_wo32(priv->pramin_bar, 0x14, 0x00000000); + + /* map channel into PRAMIN, gpuobj didn't do it for us */ + ret = nv50_instmem_bind(dev, chan->ramin); if (ret) return ret; - v = c_offset | 1; - if (dev_priv->vram_sys_base) { - v += dev_priv->vram_sys_base; - v |= 0x30; - } + /* poke regs... */ + nv_wr32(dev, 0x001704, 0x00000000 | (chan->ramin->vinst >> 12)); + nv_wr32(dev, 0x001704, 0x40000000 | (chan->ramin->vinst >> 12)); + nv_wr32(dev, 0x00170c, 0x80000000 | (priv->pramin_bar->cinst >> 4)); - i = 0; - while (v < dev_priv->vram_sys_base + c_offset + c_size) { - BAR0_WI32(priv->pramin_pt, i + 0, lower_32_bits(v)); - BAR0_WI32(priv->pramin_pt, i + 4, upper_32_bits(v)); - v += 0x1000; - i += 8; + tmp = nv_ri32(dev, 0); + nv_wi32(dev, 0, ~tmp); + if (nv_ri32(dev, 0) != ~tmp) { + NV_ERROR(dev, "PRAMIN readback failed\n"); + return -EIO; } + nv_wi32(dev, 0, tmp); - while (i < pt_size) { - BAR0_WI32(priv->pramin_pt, i + 0, 0x00000000); - BAR0_WI32(priv->pramin_pt, i + 4, 0x00000000); - i += 8; - } + dev_priv->ramin_available = true; + + /* Determine VM layout */ + dev_priv->vm_gart_base = roundup(NV50_VM_BLOCK, NV50_VM_BLOCK); + dev_priv->vm_gart_size = NV50_VM_BLOCK; + + dev_priv->vm_vram_base = dev_priv->vm_gart_base + dev_priv->vm_gart_size; + dev_priv->vm_vram_size = dev_priv->vram_size; + if (dev_priv->vm_vram_size > NV50_VM_MAX_VRAM) + dev_priv->vm_vram_size = NV50_VM_MAX_VRAM; + dev_priv->vm_vram_size = roundup(dev_priv->vm_vram_size, NV50_VM_BLOCK); + dev_priv->vm_vram_pt_nr = dev_priv->vm_vram_size / NV50_VM_BLOCK; - BAR0_WI32(chan->vm_pd, 0x00, priv->pramin_pt->vinst | 0x63); - BAR0_WI32(chan->vm_pd, 0x04, 0x00000000); + dev_priv->vm_end = dev_priv->vm_vram_base + dev_priv->vm_vram_size; + + NV_DEBUG(dev, "NV50VM: GART 0x%016llx-0x%016llx\n", + dev_priv->vm_gart_base, + dev_priv->vm_gart_base + dev_priv->vm_gart_size - 1); + NV_DEBUG(dev, "NV50VM: VRAM 0x%016llx-0x%016llx\n", + dev_priv->vm_vram_base, + dev_priv->vm_vram_base + dev_priv->vm_vram_size - 1); /* VRAM page table(s), mapped into VM at +1GiB */ for (i = 0; i < dev_priv->vm_vram_pt_nr; i++) { - ret = nouveau_gpuobj_new(dev, chan, NV50_VM_BLOCK / 0x10000 * 8, - 0, 0, &chan->vm_vram_pt[i]); + ret = nouveau_gpuobj_new(dev, NULL, NV50_VM_BLOCK / 0x10000 * 8, + 0, NVOBJ_FLAG_ZERO_ALLOC, + &chan->vm_vram_pt[i]); if (ret) { - NV_ERROR(dev, "Error creating VRAM page tables: %d\n", - ret); + NV_ERROR(dev, "Error creating VRAM PGT: %d\n", ret); dev_priv->vm_vram_pt_nr = i; return ret; } - /*XXX: double-check this is ok */ dev_priv->vm_vram_pt[i] = chan->vm_vram_pt[i]; - for (v = 0; v < dev_priv->vm_vram_pt[i]->size; v += 4) - BAR0_WI32(dev_priv->vm_vram_pt[i], v, 0); - - BAR0_WI32(chan->vm_pd, 0x10 + (i*8), - chan->vm_vram_pt[i]->vinst | 0x61); - BAR0_WI32(chan->vm_pd, 0x14 + (i*8), 0); + nv_wo32(chan->vm_pd, 0x10 + (i*8), + chan->vm_vram_pt[i]->vinst | 0x61); + nv_wo32(chan->vm_pd, 0x14 + (i*8), 0); } - /* DMA object for PRAMIN BAR */ - ret = nouveau_gpuobj_new(dev, chan, 6*4, 16, 0, &priv->pramin_bar); - if (ret) - return ret; - BAR0_WI32(priv->pramin_bar, 0x00, 0x7fc00000); - BAR0_WI32(priv->pramin_bar, 0x04, dev_priv->ramin_size - 1); - BAR0_WI32(priv->pramin_bar, 0x08, 0x00000000); - BAR0_WI32(priv->pramin_bar, 0x0c, 0x00000000); - BAR0_WI32(priv->pramin_bar, 0x10, 0x00000000); - BAR0_WI32(priv->pramin_bar, 0x14, 0x00000000); - /* DMA object for FB BAR */ ret = nouveau_gpuobj_new(dev, chan, 6*4, 16, 0, &priv->fb_bar); if (ret) return ret; - BAR0_WI32(priv->fb_bar, 0x00, 0x7fc00000); - BAR0_WI32(priv->fb_bar, 0x04, 0x40000000 + - pci_resource_len(dev->pdev, 1) - 1); - BAR0_WI32(priv->fb_bar, 0x08, 0x40000000); - BAR0_WI32(priv->fb_bar, 0x0c, 0x00000000); - BAR0_WI32(priv->fb_bar, 0x10, 0x00000000); - BAR0_WI32(priv->fb_bar, 0x14, 0x00000000); - - /* Poke the relevant regs, and pray it works :) */ - nv_wr32(dev, NV50_PUNK_BAR_CFG_BASE, (chan->ramin->vinst >> 12)); - nv_wr32(dev, NV50_PUNK_UNK1710, 0); - nv_wr32(dev, NV50_PUNK_BAR_CFG_BASE, (chan->ramin->vinst >> 12) | - NV50_PUNK_BAR_CFG_BASE_VALID); - nv_wr32(dev, NV50_PUNK_BAR1_CTXDMA, (priv->fb_bar->cinst >> 4) | - NV50_PUNK_BAR1_CTXDMA_VALID); - nv_wr32(dev, NV50_PUNK_BAR3_CTXDMA, (priv->pramin_bar->cinst >> 4) | - NV50_PUNK_BAR3_CTXDMA_VALID); - + nv_wo32(priv->fb_bar, 0x00, 0x7fc00000); + nv_wo32(priv->fb_bar, 0x04, 0x40000000 + + pci_resource_len(dev->pdev, 1) - 1); + nv_wo32(priv->fb_bar, 0x08, 0x40000000); + nv_wo32(priv->fb_bar, 0x0c, 0x00000000); + nv_wo32(priv->fb_bar, 0x10, 0x00000000); + nv_wo32(priv->fb_bar, 0x14, 0x00000000); + + nv_wr32(dev, 0x001708, 0x80000000 | (priv->fb_bar->cinst >> 4)); for (i = 0; i < 8; i++) nv_wr32(dev, 0x1900 + (i*4), 0); - dev_priv->ramin_available = true; - - /* Assume that praying isn't enough, check that we can re-read the - * entire fake channel back from the PRAMIN BAR */ - for (i = 0; i < c_size; i += 4) { - if (nv_rd32(dev, NV_RAMIN + i) != nv_ri32(dev, i)) { - NV_ERROR(dev, "Error reading back PRAMIN at 0x%08x\n", - i); - return -EINVAL; - } - } - - nv_wr32(dev, NV50_PUNK_BAR0_PRAMIN, save_nv001700); - - /* Global PRAMIN heap */ - if (drm_mm_init(&dev_priv->ramin_heap, c_size, dev_priv->ramin_size - c_size)) { - NV_ERROR(dev, "Failed to init RAMIN heap\n"); - } - /*XXX: incorrect, but needed to make hash func "work" */ dev_priv->ramht_offset = 0x10000; dev_priv->ramht_bits = 9; @@ -288,6 +250,8 @@ nv50_instmem_takedown(struct drm_device *dev) if (!priv) return; + dev_priv->ramin_available = false; + /* Restore state from before init */ for (i = 0x1700; i <= 0x1710; i += 4) nv_wr32(dev, i, priv->save1700[(i - 0x1700) / 4]); @@ -302,13 +266,8 @@ nv50_instmem_takedown(struct drm_device *dev) nouveau_gpuobj_ref(NULL, &chan->vm_vram_pt[i]); dev_priv->vm_vram_pt_nr = 0; - nouveau_gpuobj_ref(NULL, &chan->vm_pd); - nouveau_gpuobj_ref(NULL, &chan->ramfc); - nouveau_gpuobj_ref(NULL, &chan->ramin); - drm_mm_takedown(&chan->ramin_heap); - - dev_priv->fifos[0] = dev_priv->fifos[127] = NULL; - kfree(chan); + nv50_channel_del(&dev_priv->fifos[0]); + dev_priv->fifos[127] = NULL; } dev_priv->engine.instmem.priv = NULL; @@ -341,9 +300,11 @@ nv50_instmem_resume(struct drm_device *dev) struct nouveau_gpuobj *ramin = chan->ramin; int i; - nv_wr32(dev, NV50_PUNK_BAR0_PRAMIN, (ramin->vinst >> 16)); + dev_priv->ramin_available = false; + dev_priv->ramin_base = ~0; for (i = 0; i < ramin->size; i += 4) - BAR0_WI32(ramin, i, ramin->im_backing_suspend[i/4]); + nv_wo32(ramin, i, ramin->im_backing_suspend[i/4]); + dev_priv->ramin_available = true; vfree(ramin->im_backing_suspend); ramin->im_backing_suspend = NULL; @@ -370,7 +331,7 @@ nv50_instmem_populate(struct drm_device *dev, struct nouveau_gpuobj *gpuobj, if (gpuobj->im_backing) return -EINVAL; - *sz = ALIGN(*sz, NV50_INSTMEM_PAGE_SIZE); + *sz = ALIGN(*sz, 4096); if (*sz == 0) return -EINVAL; @@ -438,7 +399,7 @@ nv50_instmem_bind(struct drm_device *dev, struct nouveau_gpuobj *gpuobj) while (pte < pte_end) { nv_wo32(pramin_pt, (pte * 4) + 0, lower_32_bits(vram)); nv_wo32(pramin_pt, (pte * 4) + 4, upper_32_bits(vram)); - vram += NV50_INSTMEM_PAGE_SIZE; + vram += 0x1000; pte += 2; } dev_priv->engine.instmem.flush(dev); @@ -460,6 +421,10 @@ nv50_instmem_unbind(struct drm_device *dev, struct nouveau_gpuobj *gpuobj) if (gpuobj->im_bound == 0) return -EINVAL; + /* can happen during late takedown */ + if (unlikely(!dev_priv->ramin_available)) + return 0; + pte = (gpuobj->im_pramin->start >> 12) << 1; pte_end = ((gpuobj->im_pramin->size >> 12) << 1) + pte; -- 2.46.0