diff options
Diffstat (limited to 'runlist.c')
| -rw-r--r-- | runlist.c | 212 |
1 files changed, 87 insertions, 125 deletions
| @@ -1,172 +1,134 @@ | |||
| 1 | #include <linux/kernel.h> // Kernel types | 1 | /* Copyright 2024 Joshua Bakita |
| 2 | * Helpers for dealing with the runlist and other Host (PFIFO) registers | ||
| 3 | */ | ||
| 4 | #include <linux/printk.h> // For printk() | ||
| 5 | #include <asm/errno.h> // For error defines | ||
| 6 | #include <asm/io.h> // For phys_to_virt() | ||
| 2 | 7 | ||
| 3 | #include "nvdebug.h" | 8 | #include "nvdebug.h" |
| 4 | 9 | ||
| 10 | // Uncomment to, upon BAR2 access failure, return a PRAMIN-based runlist pointer | ||
| 11 | // **If enabled, PRAMIN may not be otherwise used while walking the runlist!** | ||
| 5 | #define FALLBACK_TO_PRAMIN | 12 | #define FALLBACK_TO_PRAMIN |
| 6 | 13 | ||
| 7 | /* Get runlist head and info (incl. length) | 14 | /* Get runlist head and info (incl. length) |
| 8 | @param rl_iter Location at which to store output | 15 | @param rl_id Which runlist to obtain? |
| 9 | @param rl_id Which runlist to obtain? | 16 | @param rl_iter Location at which to store output |
| 17 | @return 0 or -errno on error | ||
| 10 | */ | 18 | */ |
| 11 | int get_runlist_iter(struct nvdebug_state *g, int rl_id, struct runlist_iter *rl_iter) { | 19 | int get_runlist_iter(struct nvdebug_state *g, int rl_id, struct runlist_iter *rl_iter) { |
| 12 | runlist_base_t rl_base; | 20 | uint64_t runlist_iova; |
| 13 | runlist_info_t rl_info; | 21 | enum INST_TARGET runlist_target; |
| 14 | u64 runlist_iova; | 22 | uint16_t runlist_len; |
| 23 | #ifdef FALLBACK_TO_PRAMIN | ||
| 24 | int off; | ||
| 25 | #endif // FALLBACK_TO_PRAMIN | ||
| 26 | // Zero-initialize the runlist iterator | ||
| 15 | *rl_iter = (struct runlist_iter){0}; | 27 | *rl_iter = (struct runlist_iter){0}; |
| 16 | rl_base.raw = nvdebug_readl(g, NV_PFIFO_ENG_RUNLIST_BASE(rl_id)); | 28 | |
| 17 | // Check that reads are working | 29 | // Get runlist location and length using architecture-dependent logic |
| 18 | if (rl_base.raw == -1) | 30 | if (g->chip_id < NV_CHIP_ID_TURING) { |
| 19 | return -EIO; | 31 | eng_runlist_gf100_t rl; |
| 20 | // The address pointed to `regs` + NV_PFIFO_RUNLIST_BASE seems to not be: | 32 | if ((rl.raw = nvdebug_readq(g, NV_PFIFO_ENG_RUNLIST_BASE_GF100(rl_id))) == -1) |
| 21 | // - A GPU address (type is sysmem_coherent) | 33 | return -EIO; |
| 22 | // - A physical address (dereferencing after ioremap crashes) | 34 | runlist_iova = ((uint64_t)rl.ptr) << 12; |
| 23 | // - A kernel virtual address (dereferencing segfaults) | 35 | runlist_target = rl.target; |
| 24 | // So maybe it's some sort of custom thing? This is an address that the GPU | 36 | printk(KERN_INFO "[nvdebug] Runlist %d: %d entries @ %llx in %s (config raw: %#018llx)\n", |
| 25 | // can use, so it would make most sense for it to be a physical address. | 37 | rl_id, rl.len, runlist_iova, target_to_text(rl.target), rl.raw); |
| 26 | // | 38 | runlist_len = rl.len; |
| 27 | // BUT, it can't possibly be a physical address, as it would refer to an | 39 | } else if (g->chip_id < NV_CHIP_ID_AMPERE) { |
| 28 | // address greater than the maximum one on our system (by a lot!). | 40 | runlist_base_tu102_t base; |
| 29 | // Maybe I'm reading the runlist base wrong? | 41 | runlist_submit_tu102_t submit; |
| 30 | // Aha, the driver calls it runlist_iova. Sounds like runlist I/O virtual | 42 | if ((base.raw = nvdebug_readq(g, NV_PFIFO_RUNLIST_BASE_TU102(rl_id))) == -1) |
| 31 | // address! So, what's this I/O address space? All I know is that it's what | 43 | return -EIO; |
| 32 | // nvgpu_mem_get_addr() returns. That function returns the result of either: | 44 | if ((submit.raw = nvdebug_readq(g, NV_PFIFO_RUNLIST_SUBMIT_TU102(rl_id))) == -1) |
| 33 | // - gpu_phys_addr which is __nvgpu_sgl_phys on our platform which (?) | 45 | return -EIO; |
| 34 | // converts an IPA to a PA? | 46 | runlist_iova = ((uint64_t)base.ptr) << 12; |
| 35 | // - nvgpu_mem_iommu_translate | 47 | runlist_target = base.target; |
| 36 | // | 48 | runlist_len = submit.len; |
| 37 | // The original memory is allocated with nvgpu_dma_alloc_flags_sys(), which | 49 | } |
| 38 | // returns SYSMEM. | ||
| 39 | // | ||
| 40 | // To convert a physical address to a IOMMU address, we add a bit | ||
| 41 | // | ||
| 42 | // BUT, it turns out that it IS JUST A PHYSICAL ADDRESS! It wasn't working | ||
| 43 | // before because the GPU had simply gone to sleep and invalidated its | ||
| 44 | // register state, so nvgpu_readl() was simply returning garbage. | ||
| 45 | rl_info.raw = nvdebug_readl(g, NV_PFIFO_ENG_RUNLIST(rl_id)); | ||
| 46 | if (rl_info.raw == -1) | ||
| 47 | return -EIO; | ||
| 48 | runlist_iova = ((u64)rl_base.ptr) << 12; | ||
| 49 | printk(KERN_INFO "[nvdebug] Runlist %d @ %llx in %s (config raw: %x)\n", | ||
| 50 | rl_id, runlist_iova, target_to_text(rl_base.target), rl_base.raw); | ||
| 51 | printk(KERN_INFO "[nvdebug] Runlist length %d, ID %d\n", rl_info.len, rl_info.id); | ||
| 52 | // Return early on an empty runlist | 50 | // Return early on an empty runlist |
| 53 | if (!rl_info.len) | 51 | if (!runlist_len) |
| 54 | return 0; | 52 | return 0; |
| 53 | |||
| 55 | // If the runlist is in VID_MEM, search the BAR2/3 page tables for a mapping | 54 | // If the runlist is in VID_MEM, search the BAR2/3 page tables for a mapping |
| 56 | if (rl_base.target == TARGET_VID_MEM) { | 55 | if (runlist_target == TARGET_VID_MEM) { |
| 57 | printk(KERN_WARNING "[nvdebug] Runlist is located in video memory. Access to video memory is experimental."); | 56 | void __iomem *bar2_page_dir; |
| 58 | bar_config_block_t bar1_block, bar2_block; | 57 | bool pdb_is_ver2; |
| 59 | bar1_block.raw = nvdebug_readl(g, NV_PBUS_BAR1_BLOCK); | ||
| 60 | printk(KERN_INFO "[nvdebug] BAR1 inst block @ %llx in %s's %s address space.\n", ((u64)bar1_block.ptr) << 12, target_to_text(bar1_block.target), bar1_block.is_virtual ? "virtual" : "physical"); | ||
| 61 | bar2_block.raw = nvdebug_readl(g, NV_PBUS_BAR2_BLOCK); | ||
| 62 | printk(KERN_INFO "[nvdebug] BAR2 inst block @ %llx in %s's %s address space.\n", ((u64)bar2_block.ptr) << 12, target_to_text(bar2_block.target), bar1_block.is_virtual ? "virtual" : "physical"); | ||
| 63 | uint32_t bar_inst_pramin_offset = vram2PRAMIN(g, (uint64_t)bar2_block.ptr << 12); | ||
| 64 | if (!bar_inst_pramin_offset) { | ||
| 65 | printk(KERN_WARNING "[nvdebug] Unable to find instance block for BAR2/3 in the current NV_PRAMIN window. VRAM inaccessible.\n"); | ||
| 66 | goto attempt_pramin_access; | ||
| 67 | } | ||
| 68 | /* TODO: Support BAR1? | ||
| 69 | bar_inst_pramin_offset = vram2PRAMIN(g, bar1_block.ptr << 12); | ||
| 70 | if (!bar_inst_pramin_offset) { | ||
| 71 | printk(KERN_WARNING "[nvdebug] Unable to find instance block for BAR1 in the current NV_PRAMIN window. VRAM inaccessible.\n"); | ||
| 72 | return -EOPNOTSUPP; | ||
| 73 | }*/ | ||
| 74 | // Instance blocks (size == 1kb) contain many things, but we only care about | ||
| 75 | // the section which describes the location of the page directory (page table) | ||
| 76 | uint32_t bar_pdb_config_pramin_offset = bar_inst_pramin_offset + NV_PRAMIN_PDB_CONFIG_OFF; | ||
| 77 | page_dir_config_t pd_config; | ||
| 78 | pd_config.raw = nvdebug_readq(g, bar_pdb_config_pramin_offset + NV_PRAMIN); | ||
| 79 | uint64_t bar_pdb_vram_addr = pd_config.page_dir_hi; | ||
| 80 | bar_pdb_vram_addr <<= 20; | ||
| 81 | bar_pdb_vram_addr |= pd_config.page_dir_lo; | ||
| 82 | bar_pdb_vram_addr <<= 12; | ||
| 83 | printk(KERN_INFO "[nvdebug] BAR2 PDB @ %llx in %s of version %s (config raw: %llx)\n", bar_pdb_vram_addr, target_to_text(pd_config.target), pd_config.is_ver2 ? "2" : "1", pd_config.raw); | ||
| 84 | // TODO: SYSMEM support for page table location | ||
| 85 | if (pd_config.target != TARGET_VID_MEM) { | ||
| 86 | printk(KERN_WARNING "[nvdebug] BAR2 PDB is in an unsupported location.\n"); | ||
| 87 | goto attempt_pramin_access; | ||
| 88 | } | ||
| 89 | uint32_t bar_pdb_pramin_offset = vram2PRAMIN(g, bar_pdb_vram_addr); | ||
| 90 | if (!bar_pdb_pramin_offset) { | ||
| 91 | printk(KERN_WARNING "[nvdebug] Unable to find page directory BAR2/3 in the current NV_PRAMIN window. VRAM inaccessible.\n"); | ||
| 92 | goto attempt_pramin_access; | ||
| 93 | } | ||
| 94 | uint64_t runlist_bar_vaddr; | 58 | uint64_t runlist_bar_vaddr; |
| 95 | if (pd_config.is_ver2) | 59 | |
| 96 | runlist_bar_vaddr = search_page_directory(g, g->regs + NV_PRAMIN + bar_pdb_pramin_offset, phy2PRAMIN, runlist_iova); | 60 | if (get_bar2_pdb(g, &bar2_page_dir, &pdb_is_ver2) < 0) |
| 61 | return -EIO; | ||
| 62 | |||
| 63 | if (pdb_is_ver2) | ||
| 64 | runlist_bar_vaddr = search_page_directory(g, bar2_page_dir, phy2PRAMIN, runlist_iova); | ||
| 97 | else | 65 | else |
| 98 | runlist_bar_vaddr = search_v1_page_directory(g, g->regs + NV_PRAMIN + bar_pdb_pramin_offset, phy2PRAMIN, runlist_iova); | 66 | runlist_bar_vaddr = search_v1_page_directory(g, bar2_page_dir, phy2PRAMIN, runlist_iova); |
| 99 | if (!runlist_bar_vaddr) { | 67 | if (!runlist_bar_vaddr) { |
| 100 | printk(KERN_WARNING "[nvdebug] Unable to find runlist mapping in BAR2/3 page tables.\n"); | 68 | printk(KERN_WARNING "[nvdebug] Unable to find runlist mapping in BAR2/3 page tables.\n"); |
| 101 | goto attempt_pramin_access; | 69 | goto attempt_pramin_access; |
| 102 | } | 70 | } |
| 103 | printk(KERN_INFO "[nvdebug] Runlist @ %llx in BAR2 virtual address space.\n", runlist_bar_vaddr); | 71 | printk(KERN_INFO "[nvdebug] Runlist @ %llx in BAR2 virtual address space.\n", runlist_bar_vaddr); |
| 104 | /* XXX: Old test code | 72 | if (!g->bar2) { |
| 105 | uint32_t bar2_pd_pramin_offset = vram_to_pramin_off(bar2_pd); | ||
| 106 | //walk_pd_subtree(bar2_pd_pramin_offset); | ||
| 107 | uint64_t runlist_bar2_vaddr = search_pd_subtree(bar2_pd_pramin_offset, runlist_iova); | ||
| 108 | page_dir_entry_t pde_0; | ||
| 109 | pde_0.raw = nvdebug_readl(g, NV_PRAMIN + bar2_pd_pramin_offset); | ||
| 110 | uint32_t pde_1 = nvdebug_readl(g, NV_PRAMIN + vram_to_pramin_off(((u64)pde_0.addr) << 12)); | ||
| 111 | uint64_t pde_bar2_vaddr = search_pd_subtree(bar2_pd_pramin_offset, ((u64)pde_0.addr) << 12); | ||
| 112 | uint32_t pde_2 = readl(g->bar3 + pde_bar2_vaddr); | ||
| 113 | printk(KERN_INFO "[nvdebug] PDE0 via PRAMIN: %x, via BAR3: %x\n", pde_1, pde_2); | ||
| 114 | */ | ||
| 115 | if (!g->bar3) { | ||
| 116 | printk(KERN_WARNING "[nvdebug] BAR2/3 not mapped.\n"); | 73 | printk(KERN_WARNING "[nvdebug] BAR2/3 not mapped.\n"); |
| 117 | return -ENODEV; | 74 | return -ENODEV; |
| 118 | } | 75 | } |
| 119 | rl_iter->curr_entry = g->bar2 + runlist_bar_vaddr; | 76 | rl_iter->curr_entry = g->bar2 + runlist_bar_vaddr; |
| 120 | } else { | 77 | } else { |
| 121 | // Directly access the runlist if stored in SYS_MEM (physically addressed) | 78 | // Directly access the runlist if stored in SYS_MEM (physically addressed) |
| 122 | rl_iter->curr_entry = phys_to_virt(runlist_iova); | 79 | // XXX: SYS_MEM is an IOMMU address on some platforms, causing this to crash |
| 80 | rl_iter->curr_entry = (void*)phys_to_virt(runlist_iova); | ||
| 123 | } | 81 | } |
| 124 | rl_iter->rl_info = rl_info; | 82 | rl_iter->len = runlist_len; |
| 125 | return 0; | 83 | return 0; |
| 84 | |||
| 126 | attempt_pramin_access: | 85 | attempt_pramin_access: |
| 127 | #ifdef FALLBACK_TO_PRAMIN | 86 | #ifdef FALLBACK_TO_PRAMIN |
| 128 | printk(KERN_INFO "[nvdebug] Attempting to move PRAMIN window to runlist as BAR2/3-based access failed [DANGEROUS SIDE EFFECTS]!\n"); | 87 | printk(KERN_INFO "[nvdebug] Attempting to move PRAMIN window to runlist as BAR2/3-based access failed [DANGEROUS SIDE EFFECTS]!\n"); |
| 129 | bar0_window_t win; | 88 | if ((off = addr_to_pramin_mut(g, runlist_iova, runlist_target)) == -1) |
| 130 | win.base = (runlist_iova >> 16); | 89 | return off; |
| 131 | win.target = TARGET_VID_MEM; | ||
| 132 | // Shift PRAMIN window. This will cause problems if it races with driver code | ||
| 133 | // that tries to do the same, or expects the window not to move. | ||
| 134 | nvdebug_writel(g, NV_PBUS_BAR0_WINDOW, win.raw); | ||
| 135 | uint32_t off = vram2PRAMIN(g, runlist_iova); | ||
| 136 | // Workaround bug for if `off` should be zero (vram2PRAMIN normally returns | ||
| 137 | // this on error) | ||
| 138 | if (!off && (runlist_iova & 0xffffull != runlist_iova)) { | ||
| 139 | printk(KERN_INFO "[nvdebug] Unable to shift PRAMIN to runlist. Aborting...\n"); | ||
| 140 | return -EOPNOTSUPP; | ||
| 141 | } | ||
| 142 | rl_iter->curr_entry = g->regs + NV_PRAMIN + off; | 90 | rl_iter->curr_entry = g->regs + NV_PRAMIN + off; |
| 143 | rl_iter->rl_info = rl_info; | 91 | rl_iter->len = runlist_len; |
| 144 | return 0; | 92 | return 0; |
| 145 | #else | 93 | #else |
| 146 | return -EOPNOTSUPP; | 94 | return -EOPNOTSUPP; |
| 147 | #endif // FALLBACK_TO_PRAMIN | 95 | #endif // FALLBACK_TO_PRAMIN |
| 148 | } | 96 | } |
| 149 | 97 | ||
| 98 | /* Trigger a preempt of the specified TSG | ||
| 99 | @param tsg_id ID of TSG to preempt. | ||
| 100 | @return 0 or -errno on error | ||
| 101 | |||
| 102 | Note: If no other TSGs exist in the associ | ||
