aboutsummaryrefslogtreecommitdiffstats
diff options
context:
space:
mode:
Diffstat
-rw-r--r--bus.c50
-rw-r--r--mmu.c206
-rw-r--r--nvdebug.h158
-rw-r--r--nvdebug_entry.c2
-rw-r--r--runlist.c33
5 files changed, 268 insertions, 181 deletions
diff --git a/bus.c b/bus.c
index 802b6df..951ac77 100644
--- a/bus.c
+++ b/bus.c
@@ -57,35 +57,25 @@ relocate:
57 window.base = (u32)(addr >> 16); // Safe, due to above range check 57 window.base = (u32)(addr >> 16); // Safe, due to above range check
58 window.target = target; 58 window.target = target;
59 nvdebug_writel(g, NV_PBUS_BAR0_WINDOW, window.raw); 59 nvdebug_writel(g, NV_PBUS_BAR0_WINDOW, window.raw);
60 // Wait for the window to move by re-reading (as done in nvgpu driver)
61 (void) nvdebug_readl(g, NV_PBUS_BAR0_WINDOW);
60 return (int)(addr & 0xffffull); 62 return (int)(addr & 0xffffull);
61} 63}
62 64
63 65/* Get a copy of the BAR2 page directory configuration (base and aperture)
64/* Get a persistent pointer to the page directory base 66 @param pd Pointer at which to store the configuration, including a pointer
65 @param pdb Dereferencable pointer to the zeroeth entry of top-level page 67 and aperture for the zeroth entry of the top-level page directory
66 directory (PD3) for the BAR2 register region. 68 (PD3 for V2 page tables). This pointer **may not** be directly
67 Note: The returned pointer will be into the PRAMIN space. If the PRAMIN 69 dereferencable, and the caller may need to shift the BAR2 window.
68 window is moved to a region that does not cover the BAR2 page table, 70 @return 0 on success, -errno on error.
69 this ***will move the window***. 71 Note: This may move the PRAMIN window.
70 Note: Even if the page table is located in SYS_MEM, we route reads/writes via
71 PRAMIN. This ensures that we always see what the GPU sees, and that
72 includes any passes through I/O MMUs or IOVA spaces.
73*/ 72*/
74int get_bar2_pdb(struct nvdebug_state *g, void **pdb, bool *is_v2_pdb) { 73int get_bar2_pdb(struct nvdebug_state *g, page_dir_config_t* pd) {
75 static void* cached_pdb = NULL;
76 static bool cached_is_v2_pdb = false;
77 static long pd_hash = 0;
78 int ret; 74 int ret;
79 bar_config_block_t bar2_block; 75 bar_config_block_t bar2_block;
80 page_dir_config_t pd_config;
81 uint64_t pdb_vram;
82 76
83 // Use cached base as long as it's still pointing to the same thing 77 if (!pd)
84 if (cached_pdb && readl(cached_pdb) == pd_hash) { 78 return -EINVAL;
85 *pdb = cached_pdb;
86 *is_v2_pdb = cached_is_v2_pdb;
87 return 0;
88 }
89 79
90 if (!g->bar2) 80 if (!g->bar2)
91 return -ENXIO; 81 return -ENXIO;
@@ -107,24 +97,10 @@ int get_bar2_pdb(struct nvdebug_state *g, void **pdb, bool *is_v2_pdb) {
107 } 97 }
108 printk(KERN_INFO "[nvdebug] BAR2 inst block at off %x in PRAMIN\n", ret); 98 printk(KERN_INFO "[nvdebug] BAR2 inst block at off %x in PRAMIN\n", ret);
109 // Pull the page directory base configuration from the instance block 99 // Pull the page directory base configuration from the instance block
110 if ((pd_config.raw = nvdebug_readq(g, NV_PRAMIN + ret + NV_PRAMIN_PDB_CONFIG_OFF)) == -1) { 100 if ((pd->raw = nvdebug_readq(g, NV_PRAMIN + ret + NV_PRAMIN_PDB_CONFIG_OFF)) == -1) {
111 printk(KERN_ERR "[nvdebug] Unable to read BAR2/3 PDB configuration! BAR2/3 inaccessible.\n"); 101 printk(KERN_ERR "[nvdebug] Unable to read BAR2/3 PDB configuration! BAR2/3 inaccessible.\n");
112 return -ENOTSUPP; 102 return -ENOTSUPP;
113 } 103 }
114 pdb_vram = pd_config.page_dir_hi;
115 pdb_vram <<= 20;
116 pdb_vram |= pd_config.page_dir_lo;
117 pdb_vram <<= 12;
118 printk(KERN_INFO "[nvdebug] BAR2 PDB @ %llx (config raw: %llx)\n", pdb_vram, pd_config.raw);
119 // Setup PRAMIN to point at the page directory
120 if ((ret = addr_to_pramin_mut(g, pdb_vram, pd_config.target)) < 0) {
121 printk(KERN_ERR "[nvdebug] Invalid BAR2/3 PDB configuration! BAR2/3 inaccessible.\n");
122 return ret;
123 }
124
125 *pdb = cached_pdb = g->regs + NV_PRAMIN + ret;
126 pd_hash = readl(cached_pdb);
127 *is_v2_pdb = cached_is_v2_pdb = pd_config.is_ver2;
128 104
129 return 0; 105 return 0;
130} 106}
diff --git a/mmu.c b/mmu.c
index e420864..70c00f9 100644
--- a/mmu.c
+++ b/mmu.c
@@ -1,117 +1,129 @@
1// Helpers to deal with NVIDIA's MMU and associated page tables 1/* Copyright 2024 Joshua Bakita
2 * Helpers to deal with NVIDIA's MMU and associated page tables
3 */
4#include <linux/err.h> // ERR_PTR() etc.
5#include <linux/iommu.h> // iommu_get_domain_for_dev() and iommu_iova_to_phys()
2#include <linux/kernel.h> // Kernel types 6#include <linux/kernel.h> // Kernel types
3 7
4#include "nvdebug.h" 8#include "nvdebug.h"
5 9
6/* One of the oldest ways to access video memory on NVIDIA GPUs is by using 10// Uncomment to print every PDE and PTE walked for debugging
7 a configurable 1MB window into VRAM which is mapped into BAR0 (register) 11//#define DEBUG
8 space starting at offset NV_PRAMIN. This is still supported on NVIDIA GPUs 12#ifdef DEBUG
9 and appear to be used today to bootstrap page table configuration. 13#define printk_debug printk
14#else
15#define printk_debug(...)
16#endif
10 17
11 Why is it mapped at a location called NVIDIA Private RAM Instance? Because 18/* Convert a page directory (PD) pointer and aperture to be kernel-accessible
12 this used to point to the entirety of intance RAM, which was seperate from
13 VRAM on older NVIDIA GPUs.
14*/
15 19
16/* Convert a physical VRAM address to an offset in the PRAMIN window 20 I/O MMU handling inspired by amdgpu_iomem_read() in amdgpu_ttm.c of the
17 @param addr VRAM address to convert 21 AMDGPU driver.
18 @return -errno on error, PRAMIN offset on success
19 22
20 Note: Use off2PRAMIN() instead if you want a dereferenceable address 23 @param addr Pointer from page directory entry (PDE)
21 Note: PRAMIN window is only 1MB, so returning an int is safe 24 @param pd_ap PD-type aperture (target address space) for `addr`
22*/ 25 @return A dereferencable kernel address, or an ERR_PTR-wrapped error
23static int vram2PRAMIN(struct nvdebug_state *g, uint64_t addr) { 26 */
24 uint64_t pramin_base_va; 27void __iomem *pd_deref(struct nvdebug_state *g, uintptr_t addr, enum PD_TARGET pd_ap) {
25 bar0_window_t window; 28 struct iommu_domain *dom;
26 window.raw = nvdebug_readl(g, NV_PBUS_BAR0_WINDOW); 29 phys_addr_t phys;
27 // Check if the address is valid (49 bits are addressable on-GPU) 30
28 if (addr & ~0x0001ffffffffffff) { 31 // Validate arguments
29 printk(KERN_ERR "[nvdebug] Invalid address %llx passed to %s!\n", 32 if (unlikely(!IS_PD_TARGET(pd_ap) || pd_ap == PD_AND_TARGET_INVALID || !addr))
30 addr, __func__); 33 return ERR_PTR(-EINVAL);
31 return -EINVAL; 34
35 // VID_MEM accesses are the simple common-case
36 if (pd_ap == PD_AND_TARGET_VID_MEM) {
37 // Using BAR2 requires a page-table traversal. As this function is part
38 // of the page-table traversal process, it must instead use PRAMIN.
39 int off = addr_to_pramin_mut(g, addr, TARGET_VID_MEM);
40 if (off < 0)
41 return ERR_PTR(off);
42 return g->regs + NV_PRAMIN + off;
32 } 43 }
33 // For unclear (debugging?) reasons, PRAMIN can point to SYSMEM 44 /* SYS_MEM accesses are rare. Only nvgpu (Jetson driver), nouveau, and this
34 if (window.target != TARGET_VID_MEM) 45 * driver are known to create page directory entries in SYS_MEM.
35 return -EFAULT; 46 *
36 pramin_base_va = ((uint64_t)window.base) << 16; 47 * On systems using an I/O MMU, or some other I/O virtual address space,
37 // Protect against out-of-bounds accesses 48 * these are **not** physical addresses, and must first be translated
38 if (addr < pramin_base_va || addr > pramin_base_va + NV_PRAMIN_LEN) 49 * through the I/O MMU before use.
39 return -ERANGE; 50 * Example default meaning of a SYS_MEM address for a few CPUs:
40 return addr - pramin_base_va; 51 * - Jetson Xavier : physical address
41} 52 * - AMD 3950X : I/O MMU address
53 * - Phenom II x4 : physical address
54 */
55 // Check for, and translate through, the I/O MMU (if any)
56 if ((dom = iommu_get_domain_for_dev(g->dev))) {
57 phys = iommu_iova_to_phys(dom, addr);
58 printk(KERN_ERR "[nvdebug] I/O MMU translated SYS_MEM I/O VA %#lx to physical address %llx.\n", addr, phys);
59 } else
60 phys = addr;
42 61
43// Convert a GPU physical address to CPU virtual address via the PRAMIN window 62 if (!phys)
44// @return A dereferencable address, or 0 (an invalid physical address) on err
45void __iomem *phy2PRAMIN(struct nvdebug_state* g, uint64_t phy) {
46 int off = vram2PRAMIN(g, phy);
47 if (off == -ERANGE)
48 printk(KERN_ERR "[nvdebug] Page table walk off end of PRAMIN!\n");
49 if (off < 0)
50 return 0; 63 return 0;
51 return g->regs + NV_PRAMIN + vram2PRAMIN(g, phy);
52}
53 64
54/* FIXME 65 return phys_to_virt(addr);
55void __iomem *off2BAR2(struct nvdebug_state* g, uint32_t off) {
56 return g->bar2 + off;
57} 66}
58*/
59 67
60// Internal helper for search_page_directory(). 68// Internal helper for search_page_directory().
61uint64_t search_page_directory_subtree(struct nvdebug_state *g, 69uint64_t search_page_directory_subtree(struct nvdebug_state *g,
62 void __iomem *pde_offset, 70 uintptr_t pde_addr,
63 void __iomem *(*off2addr)(struct nvdebug_state*, uint64_t), 71 enum PD_TARGET pde_target,
64 uint64_t addr_to_find, 72 uint64_t addr_to_find,
65 uint32_t level) { 73 uint32_t level) {
66 uint64_t res, i; 74 uint64_t res, i;
67 void __iomem *next; 75 void __iomem *pde_kern;
68 page_dir_entry_t entry; 76 page_dir_entry_t entry;
69 if (level > sizeof(NV_MMU_PT_V2_SZ)) 77 if (level > sizeof(NV_MMU_PT_V2_SZ))
70 return 0; 78 return 0;
71 // Hack to workaround PDE0 being double-size and strangely formatted 79 // Hack to workaround PDE0 being double-size and strangely formatted
72 if (NV_MMU_PT_V2_ENTRY_SZ[level] == 16) 80 if (NV_MMU_PT_V2_ENTRY_SZ[level] == 16)
73 pde_offset += 8; 81 pde_addr += 8;
74 entry.raw_w = readq(pde_offset); 82 // Translate a VID_MEM/SYS_MEM-space address to something kernel-accessible
83 pde_kern = pd_deref(g, pde_addr, pde_target);
84 if (IS_ERR_OR_NULL(pde_kern)) {
85 printk(KERN_ERR "[nvdebug] %s: Unable to resolve %#lx in GPU %s to a kernel-accessible address. Error %ld.\n", __func__, pde_addr, pd_target_to_text(pde_target), PTR_ERR(pde_kern));
86 return 0;
87 }
88 // Read the page directory entry (a pointer to another directory, or a PTE)
89 entry.raw_w = readq(pde_kern);
75 // If we reached an invalid (unpopulated) PDE, walk back up the tree 90 // If we reached an invalid (unpopulated) PDE, walk back up the tree
76 if (entry.target == PD_AND_TARGET_INVALID) 91 if (entry.target == PD_AND_TARGET_INVALID)
77 return 0; 92 return 0;
78 // Succeed when we reach a PTE with the address we want 93 // Succeed when we reach a PTE with the address we want
79 if (entry.is_pte) { 94 if (entry.is_pte) {
80 // TODO: Handle huge pages here 95 // TODO: Handle huge pages here
81 printk(KERN_INFO "[nvdebug] PTE for phy addr %#018llx, ap '%s', vol '%d', priv '%d', ro '%d', no_atomics '%d' (raw: %#018llx)\n", ((u64)entry.addr_w) << 12, pd_target_to_text(entry.target), entry.is_volatile, entry.is_privileged, entry.is_readonly, entry.atomics_disabled, entry.raw_w); 96 printk_debug(KERN_INFO "[nvdebug] PTE for phy addr %#018llx, ap '%s', vol '%d', priv '%d', ro '%d', no_atomics '%d' (raw: %#018llx)\n", ((u64)entry.addr_w) << 12, pd_target_to_text(entry.target), entry.is_volatile, entry.is_privileged, entry.is_readonly, entry.atomics_disabled, entry.raw_w);
82 return (uint64_t)entry.addr << 12 == addr_to_find; 97 return (uint64_t)entry.addr << 12 == addr_to_find;
83 } 98 }
84 printk(KERN_INFO "[nvdebug] Found PDE pointing to %#018llx in ap '%s' vol '%d' at lvl %d (raw: %#018llx)\n", ((u64)entry.addr_w) << 12, pd_target_to_text(entry.target), entry.is_volatile, level, entry.raw_w); 99 printk_debug(KERN_INFO "[nvdebug] Found PDE pointing to %#018llx in ap '%s' vol '%d' at lvl %d (raw: %#018llx)\n", ((u64)entry.addr_w) << 12, pd_target_to_text(entry.target), entry.is_volatile, level, entry.raw_w);
85 // Depth-first search of the page table 100 // Depth-first search of the page table
86 for (i = 0; i < NV_MMU_PT_V2_SZ[level + 1]; i++) { 101 for (i = 0; i < NV_MMU_PT_V2_SZ[level + 1]; i++) {
87 next = off2addr(g, ((uint64_t)entry.addr << 12) + NV_MMU_PT_V2_ENTRY_SZ[level + 1] * i); 102 uint64_t next = ((uint64_t)entry.addr << 12) + NV_MMU_PT_V2_ENTRY_SZ[level + 1] * i;
88 // off2addr can fail 103 res = search_page_directory_subtree(g, next, entry.target, addr_to_find, level + 1);
89 if (!next || !entry.addr_w) {
90 printk(KERN_ERR "[nvdebug] %s: Unable to resolve GPU PA to CPU PA\n", __func__);
91 return 0;
92 }
93 res = search_page_directory_subtree(g, next, off2addr, addr_to_find, level + 1);
94 if (res) 104 if (res)
95 return res | (i << NV_MMU_PT_V2_LSB[level + 1]); 105 return res | (i << NV_MMU_PT_V2_LSB[level + 1]);
96 } 106 }
97 return 0; 107 return 0;
98} 108}
99 109
100/* GPU Physical address -> Virtual address ("reverse" translation) 110/* GPU Physical address -> Virtual address ("reverse" translation) for V2 tables
111
112 Depth-first search a page directory of the GPU MMU for where a particular
113 physical address is mapped. Upon finding a mapping, the virtual address is
114 returned.
101 115
102 Depth-first search a page directory of the GPU MMU for where a particular 116 The page directory may be located in VID_MEM, SYS_MEM, or some combination of
103 physical address is mapped. Upon finding a mapping, the virtual address is 117 the two.
104 returned.
105 118
106 @param pde_offset Dereferenceable pointer to the start of the PDE3 entries 119 @param pd_config Page Directory configuration, containing pointer and
107 @param off2addr Func to convert VRAM phys addresses to valid CPU VAs 120 aperture for the start of the PDE3 entries
108 @param addr_to_find Physical address to reconstruct the virtual address of 121 @param addr_to_find Physical address to reconstruct the virtual address of
109 @return 0 on error, otherwise the virtual address at which addr_to_find is 122 @return 0 on error, otherwise the virtual address at which addr_to_find is
110 mapped into by this page table. (Zero is not a valid virtual address) 123 mapped into by this page table. (Zero is not a valid virtual address)
111*/ 124*/
112uint64_t search_page_directory(struct nvdebug_state *g, 125uint64_t search_page_directory(struct nvdebug_state *g,
113 void __iomem *pde_offset, 126 page_dir_config_t pd_config,
114 void __iomem *(*off2addr)(struct nvdebug_state*, uint64_t),
115 uint64_t addr_to_find) { 127 uint64_t addr_to_find) {
116 uint64_t res, i; 128 uint64_t res, i;
117 // Make sure that the query is page-aligned 129 // Make sure that the query is page-aligned
@@ -119,57 +131,62 @@ uint64_t search_page_directory(struct nvdebug_state *g,
119 printk(KERN_WARNING "[nvdebug] Attempting to search for unaligned address %llx in search_page_directory()!\n", addr_to_find); 131 printk(KERN_WARNING "[nvdebug] Attempting to search for unaligned address %llx in search_page_directory()!\n", addr_to_find);
120 return 0; 132 return 0;
121 } 133 }
122 printk(KERN_INFO "[nvdebug] Searching for addr %#018llx in page table with base %#018llx\n", (u64)addr_to_find, (u64)pde_offset); 134 printk(KERN_INFO "[nvdebug] Searching for addr %#018llx in page table with base %#018lx\n", addr_to_find, (uintptr_t)pd_config.page_dir << 12);
123 // Search the top-level page directory (PDE3) 135 // Search the top-level page directory (PDE3)
124 for (i = 0; i < NV_MMU_PT_V2_SZ[0]; i++) 136 for (i = 0; i < NV_MMU_PT_V2_SZ[0]; i++)
125 if ((res = search_page_directory_subtree(g, pde_offset + NV_MMU_PT_V2_ENTRY_SZ[0] * i, off2addr, addr_to_find, 0))) 137 if ((res = search_page_directory_subtree(g, ((uintptr_t)pd_config.page_dir << 12) + NV_MMU_PT_V2_ENTRY_SZ[0] * i, INST2PD_TARGET(pd_config.target), addr_to_find, 0)))
126 return (res & ~0xfff) | (i << NV_MMU_PT_V2_LSB[0]); 138 return (res & ~0xfff) | (i << NV_MMU_PT_V2_LSB[0]);
127 return 0; 139 return 0;
128} 140}
129 141
130/* GMMU Page Tables Version 1 142/* GPU Physical address -> Virtual address ("reverse" translation) for V1 tables
131 This page table only contains 2 levels and is used in the Fermi, Kepler, and 143 (See `search_page_directory()` for documentation.)
132 Maxwell architectures 144 */
133*/
134// Number of entries in the PDE and PTE levels
135static const int NV_MMU_PT_V1_SZ[2] = {512, 1<<13}; // 2<<13 is an educated guess!!!
136// Which bit index is the least significant in indexing each page level
137static const int NV_MMU_PT_V1_LSB[2] = {25, 12}; // 25 is an educated guess!!!
138uint64_t search_v1_page_directory(struct nvdebug_state *g, 145uint64_t search_v1_page_directory(struct nvdebug_state *g,
139 void __iomem *pde_offset, 146 page_dir_config_t pd_config,
140 void __iomem *(*off2addr)(struct nvdebug_state*, uint64_t),
141 uint64_t addr_to_find) { 147 uint64_t addr_to_find) {
142 uint64_t j, i = 0; 148 uint64_t j, i = 0;
143 page_dir_entry_v1_t pde; 149 page_dir_entry_v1_t pde;
144 page_tbl_entry_v1_t pte; 150 page_tbl_entry_v1_t pte;
145 void __iomem *pte_offset; 151 uintptr_t pte_offset, pde_offset;
152 void __iomem *pte_addr, *pde_addr;
146 // For each PDE 153 // For each PDE
147 do { 154 do {
155 // Index the list of page directory entries
156 pde_offset = ((uint64_t)pd_config.page_dir << 12) + i * sizeof(page_dir_entry_v1_t);
157 // Convert the VID_MEM/SYS_MEM address to a kernel-accessible addr
158 pde_addr = pd_deref(g, pde_offset, INST2PD_TARGET(pd_config.target));
159 if (IS_ERR_OR_NULL(pde_addr)) {
160 printk(KERN_ERR "[nvdebug] %s: Unable to resolve %#lx in GPU %s to a kernel-accessible address. Error %ld.\n", __func__, pde_offset, pd_target_to_text(INST2PD_TARGET(pd_config.target)), -PTR_ERR(pde_addr));
161 return 0;
162 }
148 // readq doesn't seem to work on BAR0 163 // readq doesn't seem to work on BAR0
149 pde.raw = readl(pde_offset + i * sizeof(page_dir_entry_v1_t) + 4); 164 pde.raw = readl(pde_addr + 4);
150 pde.raw <<= 32; 165 pde.raw <<= 32;
151 pde.raw |= readl(pde_offset + i * sizeof(page_dir_entry_v1_t)); 166 pde.raw |= readl(pde_addr);
152 // Verify PDE is present 167 // Verify PDE is present
153 if (pde.target == PD_TARGET_INVALID && pde.alt_target == PD_TARGET_INVALID) 168 if (pde.target == PD_TARGET_INVALID && pde.alt_target == PD_TARGET_INVALID)
154 continue; 169 continue;
155 // Convert to a dereferencable pointer from CPU virtual address space
156 pte_offset = off2addr(g, (uint64_t)pde.alt_addr << 12);
157 if (!pte_offset)
158 continue;
159// printk(KERN_INFO "[nvdebug] Found %s PDE pointing to PTEs @ %llx in ap '%d' (raw: %llx)\n", pde.is_volatile ? "volatile" : "non-volatile", ((u64)pde.addr) << 12, pde.target, pde.raw); 170// printk(KERN_INFO "[nvdebug] Found %s PDE pointing to PTEs @ %llx in ap '%d' (raw: %llx)\n", pde.is_volatile ? "volatile" : "non-volatile", ((u64)pde.addr) << 12, pde.target, pde.raw);
160// printk(KERN_INFO "[nvdebug] Found %s PDE pointing to PTEs @ %llx in ap '%d' (raw: %llx)\n", pde.alt_is_volatile ? "volatile" : "non-volatile", ((u64)pde.alt_addr) << 12, pde.target, pde.raw); 171 printk_debug(KERN_INFO "[nvdebug] Found %s PDE pointing to PTEs @ %llx in ap '%d' (raw: %llx)\n", pde.alt_is_volatile ? "volatile" : "non-volatile", ((u64)pde.alt_addr) << 12, pde.alt_target, pde.raw);
161 // For each PTE 172 // For each PTE
162 for (j = 0; j < NV_MMU_PT_V1_SZ[1]; j++) { 173 for (j = 0; j < NV_MMU_PT_V1_SZ[1]; j++) {
163 // Don't overrun the PRAMIN window 174 // Index the list of page table entries starting at pde.alt_addr
164 if (pte_offset > NV_PRAMIN + g->regs + NV_PRAMIN_LEN) 175 pte_offset = ((uint64_t)pde.alt_addr << 12) + j * sizeof(page_tbl_entry_v1_t);
176 // Convert the VID_MEM/SYS_MEM address to a kernel-accessible addr
177 pte_addr = pd_deref(g, pte_offset, V12PD_TARGET(pde.alt_target));
178 if (IS_ERR_OR_NULL(pte_addr)) {
179 printk(KERN_ERR "[nvdebug] %s: Unable to resolve %#lx in GPU %s to a kernel-accessible address. Error %ld.\n", __func__, pte_offset, pd_target_to_text(V12PD_TARGET(pde.alt_target)), -PTR_ERR(pte_addr));
165 return 0; 180 return 0;
166 pte.raw = readl(pte_offset + j * sizeof(page_tbl_entry_v1_t) + 4); 181 }
182 // Read page table entry, avoiding readq
183 pte.raw = readl(pte_addr + 4);
167 pte.raw <<= 32; 184 pte.raw <<= 32;
168 pte.raw |= readl(pte_offset + j * sizeof(page_tbl_entry_v1_t)); 185 pte.raw |= readl(pte_addr);
169 // Skip non-present PTEs 186 // Skip non-present PTEs
170 if (!pte.is_present) 187 if (!pte.is_present)
171 continue; 188 continue;
172// printk(KERN_INFO "[nvdebug] PTE for phy addr %llx %s (raw: %llx)\n", ((u64)pte.addr) << 12, pte.is_present ? "present" : "non-present", pte.raw); 189 printk_debug(KERN_INFO "[nvdebug] PTE for phy addr %llx %s (raw: %llx)\n", ((u64)pte.addr) << 12, pte.is_present ? "present" : "non-present", pte.raw);
173 // If we find a matching PTE, return its virtual address 190 // If we find a matching PTE, return its virtual address
174 if ((uint64_t)pte.addr << 12 == addr_to_find) 191 if ((uint64_t)pte.addr << 12 == addr_to_find)
175 return i << NV_MMU_PT_V1_LSB[0] | j << NV_MMU_PT_V1_LSB[1]; 192 return i << NV_MMU_PT_V1_LSB[0] | j << NV_MMU_PT_V1_LSB[1];
@@ -178,9 +195,6 @@ uint64_t search_v1_page_directory(struct nvdebug_state *g,
178 return 0; 195 return 0;
179} 196}
180 197
181/* GMMU Page Tables Version 0
182 This page table only contains 2 levels and is used in the Tesla architecture
183*/
184/* *** UNTESTED *** 198/* *** UNTESTED ***
185#define NV_MMU_PT_V0_SZ 2048 199#define NV_MMU_PT_V0_SZ 2048
186#define NV_MMU_PT_V0_LSB 29 200#define NV_MMU_PT_V0_LSB 29
diff --git a/nvdebug.h b/nvdebug.h
index 567806d..eff1470 100644
--- a/nvdebug.h
+++ b/nvdebug.h
@@ -818,6 +818,14 @@ typedef union {
818} bar_config_block_t; 818} bar_config_block_t;
819 819
820/* BAR0 PRAMIN (Private RAM Instance) window configuration 820/* BAR0 PRAMIN (Private RAM Instance) window configuration
821 One of the oldest ways to access video memory on NVIDIA GPUs is by using
822 a configurable 1MB window into VRAM which is mapped into BAR0 (register)
823 space starting at offset NV_PRAMIN. This is still supported on NVIDIA GPUs
824 and appear to be used today to bootstrap page table configuration.
825
826 Why is it mapped at a location called NVIDIA Private RAM Instance? Because
827 this used to point to the entirety of intance RAM, which was seperate from
828 VRAM on older NVIDIA GPUs.
821 829
822 BASE : Base of window >> 16 in [TARGET] virtual address space 830 BASE : Base of window >> 16 in [TARGET] virtual address space
823 TARGET : Which address space BASE points into 831 TARGET : Which address space BASE points into
@@ -843,7 +851,7 @@ typedef union {
843typedef union { 851typedef union {
844 struct { 852 struct {
845 uint32_t target:2; 853 uint32_t target:2;
846 uint32_t vol:1; 854 uint32_t is_volatile:1;
847 uint32_t padding0:1; 855 uint32_t padding0:1;
848 uint32_t fault_replay_tex:1; 856 uint32_t fault_replay_tex:1;
849 uint32_t fault_replay_gcc:1; 857 uint32_t fault_replay_gcc:1;
@@ -853,6 +861,10 @@ typedef union {
853 uint32_t page_dir_lo:20; 861 uint32_t page_dir_lo:20;
854 uint32_t page_dir_hi:32; 862 uint32_t page_dir_hi:32;
855 } __attribute__((packed)); 863 } __attribute__((packed));
864 struct {
865 uint32_t pad:12;
866 uint64_t page_dir:52; // Confirmed working on Xavier and tama
867 } __attribute__((packed));
856 uint64_t raw; 868 uint64_t raw;
857} page_dir_config_t; 869} page_dir_config_t;
858 870
@@ -888,6 +900,14 @@ typedef union {
888 The following arrays merely represent different projections of Fig. 1, and 900 The following arrays merely represent different projections of Fig. 1, and
889 only one is strictly needed to reconstruct all the others. However, due to 901 only one is strictly needed to reconstruct all the others. However, due to
890 the complexity of page tables, we include all of these to aid in readability. 902 the complexity of page tables, we include all of these to aid in readability.
903
904 Support: Pascal, Volta, Turing, Ampere, Ada, Ampere, Hopper*, Blackwell*
905 Note: *Hopper introduces Version 3 Page Tables, but is backwards-compatible.
906 The newer version adds a PD4 level to support 57-bit virtual
907 addresses, and slightly shifts the PDE and PTE fields.
908
909 See also: gp100-mmu-format.pdf in open-gpu-doc. In open-gpu-kernel-modules
910 this is synonymously the "NEW" and "VER2" layout.
891*/ 911*/
892// How many nodes/entries per level in V2 of NVIDIA's page table format 912// How many nodes/entries per level in V2 of NVIDIA's page table format
893static const int NV_MMU_PT_V2_SZ[5] = {4, 512, 512, 256, 512}; 913static const int NV_MMU_PT_V2_SZ[5] = {4, 512, 512, 256, 512};
@@ -907,6 +927,12 @@ enum PD_TARGET {
907 PTE_AND_TARGET_SYS_MEM_COHERENT = 5, // b101 927 PTE_AND_TARGET_SYS_MEM_COHERENT = 5, // b101
908 PTE_AND_TARGET_SYS_MEM_NONCOHERENT = 7, // b111 928 PTE_AND_TARGET_SYS_MEM_NONCOHERENT = 7, // b111
909}; 929};
930// The low bit is unset on page directory (PD) targets
931#define IS_PD_TARGET(target) (!(target & 0x1u))
932// Convert from an enum INST_TARGET to an enum PD_TARGET
933#define INST2PD_TARGET(target) ((target & 0x2) ? (target << 1) : (!target) << 1)
934// Convert from an enum V1_PD_TARGET to an enum PD_TARGET
935#define V12PD_TARGET(target) (target << 1)
910static inline const char *pd_target_to_text(enum PD_TARGET t) { 936static inline const char *pd_target_to_text(enum PD_TARGET t) {
911 switch (t) { 937 switch (t) {
912 case PD_AND_TARGET_INVALID: 938 case PD_AND_TARGET_INVALID:
@@ -928,13 +954,10 @@ static inline const char *pd_target_to_text(enum PD_TARGET t) {
928} 954}
929 955
930// Page Directory Entry/Page Table Entry V2 type 956// Page Directory Entry/Page Table Entry V2 type
931// Note: As the meaning of target (bits 2:1) changes depending on if the entry 957// Note: As the meaning of target (bits 2:1) at a PDE-level changes if the
932// is a PTE or not, this combines them into a single target field to 958// entry is a large-page PTE or not. To simply the logic, we combine them
933// simplify comparisons. 959// into a single target field to simplify comparisons.
934// Support: Pascal, Volta, Turing, Ampere, Ada 960#define TARGET_PEER 1
935//
936// V3 introduced with Hopper, but Hopper and Blackwell also support V2
937//
938typedef union { 961typedef union {
939 // Page Directory Entry (PDE) 962 // Page Directory Entry (PDE)
940 struct { 963 struct {
@@ -965,21 +988,74 @@ typedef union {
965 uint64_t raw_w; 988 uint64_t raw_w;
966} page_dir_entry_t; 989} page_dir_entry_t;
967 990
968// Page Directory Entry/Page Table Entry V1 type 991/* GMMU Page Tables Version 1
969// Support: Fermi, Kepler, Maxwell 992 These page tables contain 2 levels and are used in the Fermi, Kepler, and
993 Maxwell architectures to support a 40-bit virtual address space.
994
995 Version 1 Page Tables may be configured to support either 64 KiB or 128 KiB
996 large pages. Table addressing differs between the modes---even if the table
997 contains no large pages. The format for 4 KiB pages in each mode is shown
998 below.
999
1000 V1 of NVIDIA's page table format uses 1 level of PDEs and a level of PTEs.
1001 How the virtual address is sliced to yield an index into each level and a
1002 page offset is shown by Fig 1 and Fig 2 (for 64 KiB and 128 KiB large page
1003 modes respectively).
1004
1005 == Figure 1: 64 KiB mode ==
1006 Page Offset (12 bits) <----------------------------------+
1007 Page Table Entry (PTE) (13 bits) <--------------+ |
1008 Page Directory Entry (PDE) (13 bits) <-+ | |
1009 ^ ^ ^
1010 Virtual address: [39, 25] [24, 12] [11, 0]
1011
1012 == Figure 2: 128 KiB mode ==
1013 Page Offset (12 bits) <----------------------------------+
1014 Page Table Entry (PTE) (14 bits) <--------------+ |
1015 Page Directory Entry (PDE) (12 bits) <-+ | |
1016 ^ ^ ^
1017 Virtual address: [39, 26] [25, 12] [11, 0]
1018
1019
1020 Support: Fermi, Kepler, Maxwell, Pascal*
1021 Note: *Pascal introduces Version 2 Page Tables, but is backwards-compatible.
1022 Note: We only implement the 64-KiB-large-page mode in nvdebug.
1023
1024 See also: mm_gk20a.c in nvgpu (Jetson GPU driver) and kern_gmmu_fmt_gm10x.c
1025 in open-gpu-kernel-modules (open-source NVRM variant). This is
1026 synonymously the "VER1" and unversioned layout in
1027 open-gpu-kernel-modules, with some differences noted in Appdx 1.
1028
1029 == Appdx 1 ==
1030 In open-gpu-kernel-modules, the unversioned MMU layout adds:
1031 - Bit 35: NV_MMU_PTE_LOCK synonym for NV_MMU_PTE_ATOMIC_DISABLE
1032 - Bit 62: NV_MMU_PTE_READ_DISABLE overlapping NV_MMU_PTE_COMPTAGLINE
1033 - Bit 63: NV_MMU_PTE_WRITE_DISABLE overlapping NV_MMU_PTE_COMPTAGLINE
1034 And removes:
1035 - Bit 40, 41, 42, 43 from NV_MMU_PTE_KIND
1036 The PDE layouts are identical. Given that the unversioned defines seem to
1037 predate renaming and/or field extension/relocation, they are likely artifacts
1038 from the page table development process, and have no meaning now.
1039*/
1040// Number of entries in the PDE and PTE levels
1041static const int NV_MMU_PT_V1_SZ[2] = {8192, 8192};
1042// Which bit index is the least significant in indexing each page level
1043static const int NV_MMU_PT_V1_LSB[2] = {25, 12};
1044
1045// V1 Page Directory Entry target
970enum V1_PD_TARGET { 1046enum V1_PD_TARGET {
971 PD_TARGET_INVALID = 0, 1047 PD_TARGET_INVALID = 0,
972 PD_TARGET_VID_MEM = 1, 1048 PD_TARGET_VID_MEM = 1,
973 PD_TARGET_SYS_MEM_COHERENT = 2, 1049 PD_TARGET_SYS_MEM_COHERENT = 2,
974 PD_TARGET_SYS_MEM_NONCOHERENT = 3, 1050 PD_TARGET_SYS_MEM_NONCOHERENT = 3,
975}; 1051};
976// Page Directory Entry (PDE) 1052// V1 Page Directory Entry (PDE)
977typedef union { 1053typedef union {
978// Large page fields 1054// Large page fields
979 struct { 1055 struct {
980// 0:32 1056// 0:32
981 enum V1_PD_TARGET target:2; 1057 enum V1_PD_TARGET target:2;
982 uint32_t padding0:2; 1058 uint32_t padding0:2; // Documented as "PDE_SIZE"?
983 uint64_t addr:28; // May be wider? 1059 uint64_t addr:28; // May be wider?
984// 32:63 1060// 32:63
985 uint32_t padding2:3; 1061 uint32_t padding2:3;
@@ -998,45 +1074,58 @@ typedef union {
998 } __attribute__((packed)); 1074 } __attribute__((packed));
999 uint64_t raw; 1075 uint64_t raw;
1000} page_dir_entry_v1_t; 1076} page_dir_entry_v1_t;
1001// Page Table Entry (PTE) 1077
1002// Reconstructed from info in Jetson nvgpu driver 1078// V1 Page Table Entry (PTE)
1003typedef union { 1079typedef union {
1004 struct { 1080 struct {
1005// 0:32 1081// 0:32
1006 bool is_present:1; 1082 bool is_present:1;
1007 bool is_privileged:1; 1083 bool is_privileged:1;
1008 bool is_readonly:1; 1084 bool is_readonly:1;
1009 uint32_t padding0:1; 1085 bool is_encrypted:1;
1010 uint64_t addr:28; 1086 uint64_t addr:28;
1011// 32:63 1087// 32:63
1012 bool is_volatile:1; 1088 bool is_volatile:1;
1013 enum INST_TARGET:2; 1089 enum INST_TARGET:2;
1014 uint32_t padding1:1; 1090 bool atomics_disabled:1;
1015 uint32_t kind:8; 1091 uint32_t kind:8;
1016 uint32_t comptag:17; 1092 uint32_t comptag:20;
1017 uint32_t padding2:1;
1018 bool is_read_disabled:1;
1019 bool is_write_disabled:1;
1020 } __attribute__((packed)); 1093 } __attribute__((packed));
1021 uint64_t raw; 1094 uint64_t raw;
1022} page_tbl_entry_v1_t; 1095} page_tbl_entry_v1_t;
1023//enum V0_PDE_TYPE {NOT_PRESENT = 0, PAGE_64K = 1, PAGE_16K = 2, PAGE_4K = 3}; 1096
1024//enum V0_PDE_SIZE {PDE_SZ_128K = 0, PDE_SZ_32K = 1, PDE_SZ_16K = 2, PDE_SZ_8K = 3}; 1097/* GMMU Page Tables Version 0
1025//static const int V0_PDE_SIZE2NUM[4] = {128*1024, 32*1024, 16*1024, 8*1024}; 1098 This page table contains 2 levels to support a 40-bit virtual address space,
1026/* PDE V0 (nv50/Tesla) 1099 and is used in the Tesla (2.0?) architecture.
1100
1101 It is unclear what NVIDIA calls this page table layout. It predates V1, so we
1102 call it V0.
1103
1104 See also: https://envytools.readthedocs.io/en/latest/hw/memory/g80-vm.html
1105 */
1106/*
1107// What size pages are in the pointed-to page table?
1108enum V0_PDE_TYPE {NOT_PRESENT = 0, PAGE_64K = 1, PAGE_16K = 2, PAGE_4K = 3};
1109// How large is the pointed-to page table?
1110enum V0_PDE_SIZE {PDE_SZ_128K = 0, PDE_SZ_32K = 1, PDE_SZ_16K = 2, PDE_SZ_8K = 3};
1111// Given a page table size, how many entries does it have?
1112static const int V0_PDE_SIZE2NUM[4] = {128*1024, 32*1024, 16*1024, 8*1024};
1113
1114// PDE V0 (nv50/Tesla)
1027typedef union { 1115typedef union {
1028 struct { 1116 struct {
1029 enum V1_PDE_TYPE type:2; 1117 enum V0_PDE_TYPE type:2;
1030 enum INST_TARGET target:2; 1118 enum INST_TARGET target:2;
1031 uint32_t padding0:1; 1119 uint32_t padding0:1;
1032 enum V1_PDE_SIZE sublevel_size:2; 1120 enum V0_PDE_SIZE sublevel_size:2;
1033 uint32_t padding1:5; 1121 uint32_t padding1:5;
1034 uint32_t addr:28; 1122 uint32_t addr:28;
1035 uint32_t padding2:24; 1123 uint32_t padding2:24;
1036 } __attribute__((packed)); 1124 } __attribute__((packed));
1037 uint64_t raw; 1125 uint64_t raw;
1038} page_dir_entry_v1_t;*/ 1126} page_dir_entry_v0_t;
1039/* PTE V0 (nv50) 1127
1128// PTE V0 (nv50) for small pages
1040typedef union { 1129typedef union {
1041 struct { 1130 struct {
1042 bool is_present:1; 1131 bool is_present:1;
@@ -1055,7 +1144,8 @@ typedef union {
1055 uint32_t padding5:1; 1144 uint32_t padding5:1;
1056 } __attribute__((packed)); 1145 } __attribute__((packed));
1057 uint64_t raw; 1146 uint64_t raw;
1058} page_tbl_entry_v1_t;*/ 1147} page_tbl_entry_v0_t;
1148*/
1059 1149
1060// TODO(jbakita): Maybe put the above GPU types in a different file. 1150// TODO(jbakita): Maybe put the above GPU types in a different file.
1061 1151
@@ -1077,6 +1167,8 @@ struct nvdebug_state {
1077 struct gk20a *g; 1167 struct gk20a *g;
1078 // Pointer to PCI device needed for pci_iounmap 1168 // Pointer to PCI device needed for pci_iounmap
1079 struct pci_dev *pcid; 1169 struct pci_dev *pcid;
1170 // Pointer to generic device struct (both platform and pcie devices)
1171 struct device *dev;
1080}; 1172};
1081 1173
1082/*const struct runlist_funcs { 1174/*const struct runlist_funcs {
@@ -1152,13 +1244,11 @@ int preempt_runlist(struct nvdebug_state *g, uint32_t rl_id);
1152void __iomem *phy2PRAMIN(struct nvdebug_state* g, uint64_t phy); 1244void __iomem *phy2PRAMIN(struct nvdebug_state* g, uint64_t phy);
1153uint64_t search_page_directory( 1245uint64_t search_page_directory(
1154 struct nvdebug_state *g, 1246 struct nvdebug_state *g,
1155 void __iomem *pde_offset, 1247 page_dir_config_t pd_config,
1156 void __iomem *(*off2addr)(struct nvdebug_state*, uint64_t),
1157 uint64_t addr_to_find); 1248 uint64_t addr_to_find);
1158uint64_t search_v1_page_directory( 1249uint64_t search_v1_page_directory(
1159 struct nvdebug_state *g, 1250 struct nvdebug_state *g,
1160 void __iomem *pde_offset, 1251 page_dir_config_t pd_config,
1161 void __iomem *(*off2addr)(struct nvdebug_state*, uint64_t),
1162 uint64_t addr_to_find); 1252 uint64_t addr_to_find);
1163 1253
1164 1254
@@ -1252,4 +1342,4 @@ static inline void nvdebug_writeq(struct nvdebug_state *s, u32 r, u64 v) {
1252} 1342}
1253// Defined in bus.c 1343// Defined in bus.c
1254int addr_to_pramin_mut(struct nvdebug_state *g, uint64_t addr, enum INST_TARGET target); 1344int addr_to_pramin_mut(struct nvdebug_state *g, uint64_t addr, enum INST_TARGET target);
1255int get_bar2_pdb(struct nvdebug_state *g, void **pdb, bool *is_v2_pdb); 1345int get_bar2_pdb(struct nvdebug_state *g, page_dir_config_t* pd);
diff --git a/nvdebug_entry.c b/nvdebug_entry.c
index 0cf5344..68e4d71 100644
--- a/nvdebug_entry.c
+++ b/nvdebug_entry.c
@@ -109,6 +109,7 @@ int probe_and_cache_devices(void) {
109 g_nvdebug_state[i].chip_id = ids.chip_id; 109 g_nvdebug_state[i].chip_id = ids.chip_id;
110 g_nvdebug_state[i].pcid = NULL; 110 g_nvdebug_state[i].pcid = NULL;
111 g_nvdebug_state[i].bar3 = NULL; 111 g_nvdebug_state[i].bar3 = NULL;
112 g_nvdebug_state[i].dev = dev;
112 printk(KERN_INFO "[nvdebug] Chip ID %x (architecture %s) detected on platform bus and initialized.", 113 printk(KERN_INFO "[nvdebug] Chip ID %x (architecture %s) detected on platform bus and initialized.",
113 ids.chip_id, ARCH2NAME(ids.architecture)); 114 ids.chip_id, ARCH2NAME(ids.architecture));
114 i++; 115 i++;
@@ -131,6 +132,7 @@ int probe_and_cache_devices(void) {
131 if (!g_nvdebug_state[i].bar3) 132 if (!g_nvdebug_state[i].bar3)
132 g_nvdebug_state[i].bar3 = pci_iomap(pcid, 3, pci_resource_len(pcid, 3)/2); 133 g_nvdebug_state[i].bar3 = pci_iomap(pcid, 3, pci_resource_len(pcid, 3)/2);
133 g_nvdebug_state[i].pcid = pcid; 134 g_nvdebug_state[i].pcid = pcid;
135 g_nvdebug_state[i].dev = &pcid->dev;
134 ids.raw = nvdebug_readl(&g_nvdebug_state[i], NV_MC_BOOT_0); 136 ids.raw = nvdebug_readl(&g_nvdebug_state[i], NV_MC_BOOT_0);
135 if (ids.raw == -1) { 137 if (ids.raw == -1) {
136 pci_err(pcid, "[nvdebug] Unable to read config from Master Controller on this GPU\n"); 138 pci_err(pcid, "[nvdebug] Unable to read config from Master Controller on this GPU\n");
diff --git a/runlist.c b/runlist.c
index c725e77..2eee01c 100644
--- a/runlist.c
+++ b/runlist.c
@@ -9,7 +9,8 @@
9 9
10// Uncomment to, upon BAR2 access failure, return a PRAMIN-based runlist pointer 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!** 11// **If enabled, PRAMIN may not be otherwise used while walking the runlist!**
12#define FALLBACK_TO_PRAMIN 12// Runlists can only be printed on the Jetson TX2 if this is enabled.
13//#define FALLBACK_TO_PRAMIN
13 14
14/* Get runlist head and info (incl. length) 15/* Get runlist head and info (incl. length)
15 @param rl_id Which runlist to obtain? 16 @param rl_id Which runlist to obtain?
@@ -20,6 +21,7 @@ int get_runlist_iter(struct nvdebug_state *g, int rl_id, struct runlist_iter *rl
20 uint64_t runlist_iova; 21 uint64_t runlist_iova;
21 enum INST_TARGET runlist_target; 22 enum INST_TARGET runlist_target;
22 uint16_t runlist_len; 23 uint16_t runlist_len;
24 int err;
23#ifdef FALLBACK_TO_PRAMIN 25#ifdef FALLBACK_TO_PRAMIN
24 int off; 26 int off;
25#endif // FALLBACK_TO_PRAMIN 27#endif // FALLBACK_TO_PRAMIN
@@ -33,9 +35,9 @@ int get_runlist_iter(struct nvdebug_state *g, int rl_id, struct runlist_iter *rl
33 return -EIO; 35 return -EIO;
34 runlist_iova = ((uint64_t)rl.ptr) << 12; 36 runlist_iova = ((uint64_t)rl.ptr) << 12;
35 runlist_target = rl.target; 37 runlist_target = rl.target;
36 printk(KERN_INFO "[nvdebug] Runlist %d: %d entries @ %llx in %s (config raw: %#018llx)\n",
37 rl_id, rl.len, runlist_iova, target_to_text(rl.target), rl.raw);
38 runlist_len = rl.len; 38 runlist_len = rl.len;
39 printk(KERN_INFO "[nvdebug] Runlist %d for %x: %d entries @ %llx in %s (config raw: %#018llx)\n",
40 rl_id, g->chip_id, rl.len, runlist_iova, target_to_text(rl.target), rl.raw);
39 } else if (g->chip_id < NV_CHIP_ID_AMPERE) { 41 } else if (g->chip_id < NV_CHIP_ID_AMPERE) {
40 runlist_base_tu102_t base; 42 runlist_base_tu102_t base;
41 runlist_submit_tu102_t submit; 43 runlist_submit_tu102_t submit;
@@ -46,6 +48,8 @@ int get_runlist_iter(struct nvdebug_state *g, int rl_id, struct runlist_iter *rl
46 runlist_iova = ((uint64_t)base.ptr) << 12; 48 runlist_iova = ((uint64_t)base.ptr) << 12;
47 runlist_target = base.target; 49 runlist_target = base.target;
48 runlist_len = submit.len; 50 runlist_len = submit.len;
51 printk(KERN_INFO "[nvdebug] Runlist %d for %x: %d entries @ %llx in %s (config raw: %#018llx %#018llx)\n",
52 rl_id, g->chip_id, submit.len, runlist_iova, target_to_text(runlist_target), base.raw, submit.raw);
49 } 53 }
50 // Return early on an empty runlist 54 // Return early on an empty runlist
51 if (!runlist_len) 55 if (!runlist_len)
@@ -53,24 +57,25 @@ int get_runlist_iter(struct nvdebug_state *g, int rl_id, struct runlist_iter *rl
53 57
54 // If the runlist is in VID_MEM, search the BAR2/3 page tables for a mapping 58 // If the runlist is in VID_MEM, search the BAR2/3 page tables for a mapping
55 if (runlist_target == TARGET_VID_MEM) { 59 if (runlist_target == TARGET_VID_MEM) {
56 void __iomem *bar2_page_dir;
57 bool pdb_is_ver2;
58 uint64_t runlist_bar_vaddr; 60 uint64_t runlist_bar_vaddr;
61 page_dir_config_t pd_config;
59 62
60 if (get_bar2_pdb(g, &bar2_page_dir, &pdb_is_ver2) < 0) 63 if ((err = get_bar2_pdb(g, &pd_config)) < 0)
61 return -EIO; 64 goto attempt_pramin_access;
62 65
63 if (pdb_is_ver2) 66 if (pd_config.is_ver2)
64 runlist_bar_vaddr = search_page_directory(g, bar2_page_dir, phy2PRAMIN, runlist_iova); 67 runlist_bar_vaddr = search_page_directory(g, pd_config, runlist_iova);
65 else 68 else
66 runlist_bar_vaddr = search_v1_page_directory(g, bar2_page_dir, phy2PRAMIN, runlist_iova); 69 runlist_bar_vaddr = search_v1_page_directory(g, pd_config, runlist_iova);
67 if (!runlist_bar_vaddr) { 70 if (!runlist_bar_vaddr) {
68 printk(KERN_WARNING "[nvdebug] Unable to find runlist mapping in BAR2/3 page tables.\n"); 71 printk(KERN_WARNING "[nvdebug] Unable to find runlist %d mapping in BAR2/3 page tables for %x.\n", rl_id, g->chip_id);
72 err = -EOPNOTSUPP;
69 goto attempt_pramin_access; 73 goto attempt_pramin_access;
70 } 74 }
71 printk(KERN_INFO "[nvdebug] Runlist @ %llx in BAR2 virtual address space.\n", runlist_bar_vaddr); 75
76 printk(KERN_INFO "[nvdebug] Runlist %d for %x @ %llx in BAR2 virtual address space.\n", rl_id, g->chip_id, runlist_bar_vaddr);
72 if (!g->bar2) { 77 if (!g->bar2) {
73 printk(KERN_WARNING "[nvdebug] BAR2/3 not mapped.\n"); 78 printk(KERN_WARNING "[nvdebug] BAR2/3 not mapped for %x.\n", g->chip_id);
74 return -ENODEV; 79 return -ENODEV;
75 } 80 }
76 rl_iter->curr_entry = g->bar2 + runlist_bar_vaddr; 81 rl_iter->curr_entry = g->bar2 + runlist_bar_vaddr;
@@ -91,7 +96,7 @@ attempt_pramin_access:
91 rl_iter->len = runlist_len; 96 rl_iter->len = runlist_len;
92 return 0; 97 return 0;
93#else 98#else
94 return -EOPNOTSUPP; 99 return err;
95#endif // FALLBACK_TO_PRAMIN 100#endif // FALLBACK_TO_PRAMIN
96} 101}
97 102
generated by cgit v1.2.2 (git 2.25.0) at 2024-10-31 00:08:56 -0400