[PATCH umr 1/3] Improve handling of non-standard page tables in AI+
Joseph Greathouse
Joseph.Greathouse at amd.com
Thu Jun 17 19:25:38 UTC 2021
Fixes handling of GPUVM page table decoding when not using 4-level
page tables with 512 entries per level. This includes:
- Calculating actual size of top-most PDB based on total VM range,
page table depth, and page table block size.
- Calculating size of PTB based on the page table block size
and the PDE0's block fragment size.
- Handling PTE offset and masks from from PDE0 with P-bit, normal
PTBs, or PTBs from a translate-further layer.
- When using a PTE with F bit to go one layer deeper, pull new
block fragment size out of that PTE to handle further-level PTBs
of non-standard sizes.
Signed-off-by: Joseph Greathouse <Joseph.Greathouse at amd.com>
---
src/lib/read_vram.c | 199 ++++++++++++++++++++++++++++++++++----------
1 file changed, 153 insertions(+), 46 deletions(-)
diff --git a/src/lib/read_vram.c b/src/lib/read_vram.c
index efcd081..049acd4 100644
--- a/src/lib/read_vram.c
+++ b/src/lib/read_vram.c
@@ -297,6 +297,26 @@ invalid_page:
return -1;
}
+/** round_up_pot -- Round up value to next power of two */
+static uint64_t round_up_pot(uint64_t x)
+{
+ uint64_t y = (64ULL * 1024 * 1024); // start at 64MiB
+ while (y < x)
+ y <<= 1;
+ return y;
+}
+
+static uint64_t log2_vm_size(uint64_t page_table_start_addr, uint64_t page_table_end_addr)
+{
+ uint64_t size_of_vm_bytes = page_table_end_addr - page_table_start_addr + 4096;
+ size_of_vm_bytes = round_up_pot(size_of_vm_bytes);
+ // Find the highest bit set to get an estimate for log2(size)
+ uint32_t vm_bits = 0;
+ while (size_of_vm_bytes >>= 1)
+ vm_bits++;
+ return vm_bits;
+}
+
/**
* umr_access_vram_ai - Access GPU mapped memory for GFX9+ platforms
*/
@@ -304,17 +324,19 @@ static int umr_access_vram_ai(struct umr_asic *asic, uint32_t vmid,
uint64_t address, uint32_t size,
void *dst, int write_en)
{
- uint64_t start_addr, page_table_start_addr, page_table_base_addr,
- page_table_block_size, pte_idx, pde_idx, pte_entry, pde_entry,
+ uint64_t start_addr, page_table_start_addr, page_table_end_addr, page_table_base_addr,
+ page_table_block_size, log2_ptb_entries, pte_idx, pde_idx, pte_entry, pde_entry,
pde_address, vm_fb_offset,
va_mask, offset_mask, system_aperture_low, system_aperture_high,
- fb_top, fb_bottom, pte_page_mask, agp_base, agp_bot, agp_top, prev_addr;
+ fb_top, fb_bottom, ptb_mask, pte_page_mask, agp_base, agp_bot, agp_top, prev_addr;
uint32_t chunk_size, tmp, pde0_block_fragment_size;
int pde_cnt, current_depth, page_table_depth, zfb, further;
struct {
uint32_t
mmVM_CONTEXTx_PAGE_TABLE_START_ADDR_LO32,
mmVM_CONTEXTx_PAGE_TABLE_START_ADDR_HI32,
+ mmVM_CONTEXTx_PAGE_TABLE_END_ADDR_LO32,
+ mmVM_CONTEXTx_PAGE_TABLE_END_ADDR_HI32,
mmVM_CONTEXTx_CNTL,
mmVM_CONTEXTx_PAGE_TABLE_BASE_ADDR_LO32,
mmVM_CONTEXTx_PAGE_TABLE_BASE_ADDR_HI32,
@@ -461,6 +483,12 @@ static int umr_access_vram_ai(struct umr_asic *asic, uint32_t vmid,
sprintf(buf, "mm%sVM_CONTEXT%" PRIu32 "_PAGE_TABLE_START_ADDR_HI32", regprefix, vmid);
registers.mmVM_CONTEXTx_PAGE_TABLE_START_ADDR_HI32 = umr_read_reg_by_name_by_ip(asic, hub, buf);
page_table_start_addr |= (uint64_t)registers.mmVM_CONTEXTx_PAGE_TABLE_START_ADDR_HI32 << 44;
+ sprintf(buf, "mm%sVM_CONTEXT%" PRIu32 "_PAGE_TABLE_END_ADDR_LO32", regprefix, vmid);
+ registers.mmVM_CONTEXTx_PAGE_TABLE_END_ADDR_LO32 = umr_read_reg_by_name_by_ip(asic, hub, buf);
+ page_table_end_addr = (uint64_t)registers.mmVM_CONTEXTx_PAGE_TABLE_END_ADDR_LO32 << 12;
+ sprintf(buf, "mm%sVM_CONTEXT%" PRIu32 "_PAGE_TABLE_END_ADDR_HI32", regprefix, vmid);
+ registers.mmVM_CONTEXTx_PAGE_TABLE_END_ADDR_HI32 = umr_read_reg_by_name_by_ip(asic, hub, buf);
+ page_table_end_addr |= (uint64_t)registers.mmVM_CONTEXTx_PAGE_TABLE_END_ADDR_HI32 << 44;
sprintf(buf, "mm%sVM_CONTEXT%" PRIu32 "_CNTL", regprefix, vmid);
tmp = registers.mmVM_CONTEXTx_CNTL = umr_read_reg_by_name_by_ip(asic, hub, buf);
@@ -495,6 +523,8 @@ static int umr_access_vram_ai(struct umr_asic *asic, uint32_t vmid,
asic->mem_funcs.vm_message(
"mm%sVM_CONTEXT%" PRIu32 "_PAGE_TABLE_START_ADDR_LO32=0x%" PRIx32 "\n"
"mm%sVM_CONTEXT%" PRIu32 "_PAGE_TABLE_START_ADDR_HI32=0x%" PRIx32 "\n"
+ "mm%sVM_CONTEXT%" PRIu32 "_PAGE_TABLE_END_ADDR_LO32=0x%" PRIx32 "\n"
+ "mm%sVM_CONTEXT%" PRIu32 "_PAGE_TABLE_END_ADDR_HI32=0x%" PRIx32 "\n"
"mm%sVM_CONTEXT%" PRIu32 "_PAGE_TABLE_BASE_ADDR_LO32=0x%" PRIx32 "\n"
"mm%sVM_CONTEXT%" PRIu32 "_PAGE_TABLE_BASE_ADDR_HI32=0x%" PRIx32 "\n"
"mm%sVM_CONTEXT%" PRIu32 "_CNTL=0x%" PRIx32 "\n"
@@ -513,6 +543,8 @@ static int umr_access_vram_ai(struct umr_asic *asic, uint32_t vmid,
"mm%sMC_VM_AGP_TOP=0x%" PRIx32 "\n",
regprefix, vmid, registers.mmVM_CONTEXTx_PAGE_TABLE_START_ADDR_LO32,
regprefix, vmid, registers.mmVM_CONTEXTx_PAGE_TABLE_START_ADDR_HI32,
+ regprefix, vmid, registers.mmVM_CONTEXTx_PAGE_TABLE_END_ADDR_LO32,
+ regprefix, vmid, registers.mmVM_CONTEXTx_PAGE_TABLE_END_ADDR_HI32,
regprefix, vmid, registers.mmVM_CONTEXTx_PAGE_TABLE_BASE_ADDR_LO32,
regprefix, vmid, registers.mmVM_CONTEXTx_PAGE_TABLE_BASE_ADDR_HI32,
regprefix, vmid, registers.mmVM_CONTEXTx_CNTL,
@@ -535,10 +567,6 @@ static int umr_access_vram_ai(struct umr_asic *asic, uint32_t vmid,
// transform page_table_base
page_table_base_addr -= vm_fb_offset;
- // convert some defaults to actual values AFTER printing out to user
- // page_table_block_size of 0 means 9 (512 entries)
- if (!page_table_block_size)
- page_table_block_size = 9;
pde0_block_fragment_size = 0;
if (vmid == 0) {
@@ -593,7 +621,9 @@ static int umr_access_vram_ai(struct umr_asic *asic, uint32_t vmid,
// defaults in case we have to bail out before fully decoding to a PTE
pde_cnt = 0;
+ ptb_mask = (1ULL << 9) - 1;
pte_page_mask = (1ULL << 12) - 1;
+ log2_ptb_entries = 9;
further = 0;
if (page_table_depth >= 1) {
@@ -608,8 +638,23 @@ static int umr_access_vram_ai(struct umr_asic *asic, uint32_t vmid,
// AI+ supports more than 1 level of PDEs so we iterate for all of the depths
pde_address = pde_fields.pte_base_addr;
- // TODO: Should "page_table_block_size" just be 9 to account for potential PTB1 selectors?
- va_mask = ((uint64_t)511 << ((page_table_depth)*9 + (12 + pde0_block_fragment_size + page_table_block_size)));
+ /*
+ * Size of the first PDB depends on the total coverage of the
+ * page table and the PAGE_TABLE_BLOCK_SIZE.
+ * Entire table takes ceil(log2(total_vm_size)) bits
+ * All PDBs except the first one take 9 bits each
+ * The PTB covers at least 2 MiB (21 bits)
+ * And PAGE_TABLE_BLOCK_SIZE is log2(num 2MiB ranges PTB covers)
+ * As such, the formula for the size of the first PDB is:
+ * PDB1, PDB0, etc. PTB covers at least 2 MiB
+ * Block size can make it cover more
+ * total_vm_bits - (9 * num_middle_pdbs) - (page_table_block_size + 21)
+ */
+ int total_vm_bits = log2_vm_size(page_table_start_addr, page_table_end_addr);
+ int top_pdb_bits = total_vm_bits - (9 * (page_table_depth - 1)) - (page_table_block_size + 21);
+
+ va_mask = (1ULL << top_pdb_bits) - 1;
+ va_mask <<= (total_vm_bits - top_pdb_bits);
if ((asic->options.no_fold_vm_decode || memcmp(&pde_fields, &pde_array[pde_cnt], sizeof pde_fields)) && asic->options.verbose)
asic->mem_funcs.vm_message("BASE=0x%016" PRIx64 ", VA=0x%012" PRIx64 ", PBA==0x%012" PRIx64 ", V=%" PRIu64 ", S=%" PRIu64 ", C=%" PRIu64 ", P=%" PRIu64 "\n",
@@ -624,14 +669,19 @@ static int umr_access_vram_ai(struct umr_asic *asic, uint32_t vmid,
current_depth = page_table_depth;
while (current_depth) {
- pde_idx = address >> (9 * (current_depth - 1) + page_table_block_size + 12);
- // mask only 9 bits
- if (current_depth != page_table_depth)
- pde_idx &= (1ULL << 9) - 1;
-
-
- // TODO: redo va_mask
- va_mask = ((uint64_t)511 << ((page_table_depth - pde_cnt)*9 + (12 + pde0_block_fragment_size + page_table_block_size)));
+ // Every middle PDB has 512 entries, so shift a further 9 bits
+ // for every layer beyond the first one.
+ int amount_to_shift = (total_vm_bits - top_pdb_bits);
+ amount_to_shift -= ((page_table_depth - current_depth)*9);
+ pde_idx = address >> amount_to_shift;
+
+ // Middle layers need the upper bits masked out after the right-shift.
+ // For the top-most layer, the va_mask is set above the while loop,
+ // so we can skip re-setting it here.
+ if (current_depth != page_table_depth) {
+ pde_idx &= 511;
+ va_mask = (uint64_t)511 << amount_to_shift;
+ }
// read PDE entry
prev_addr = pde_address + pde_idx * 8;
@@ -671,9 +721,18 @@ static int umr_access_vram_ai(struct umr_asic *asic, uint32_t vmid,
pde_fields.pte = (pde_entry >> 54) & 1;
if (current_depth == 1) {
pde0_block_fragment_size = pde_fields.frag_size;
- // page_table_block_size is the number of entries in a PTB that spans 2MB
- page_table_block_size = 21 - (12 + pde0_block_fragment_size);
- pte_page_mask = (1ULL << (12 + pde0_block_fragment_size)) - 1;
+ /*
+ * page_table_block_size is the number of 2MiB regions covered by a PTB
+ * If we set it to 0, then PTB cover 2 MiB
+ * If it's 9 PTB cover 1024 MiB
+ * pde0_block_fragment_size tells us how many 4 KiB regions each PTE covers
+ * If it's 0 PTEs cover 4 KiB
+ * If it's 9 PTEs cover 2 MiB
+ * So the number of PTEs in a PTB is 2^(9+ptbs-pbfs)
+ */
+ log2_ptb_entries = (9 + (page_table_block_size - pde0_block_fragment_size));
+ ptb_mask = (1ULL << log2_ptb_entries) - 1;
+ pte_page_mask = (1ULL << (pde0_block_fragment_size + 12)) - 1;
if (asic->options.verbose)
asic->mem_funcs.vm_message("pde0.pte = %u\npde0.block_fragment_size = %u\npage_table_block_size = %u\n",
(unsigned)pde_fields.pte,
@@ -723,9 +782,13 @@ static int umr_access_vram_ai(struct umr_asic *asic, uint32_t vmid,
pde_address = pde_fields.pte_base_addr;
}
- // read PTE selector (to select from PTB0)
- // TODO: support for page_table_block_size > 9
- pte_idx = (address >> (12 + pde0_block_fragment_size)) & ((1ULL << page_table_block_size) - 1);
+ // If we fall through to here, we are pointing into PTB, so pull out
+ // the index and mask.
+ // At minimum, each PTE is 4 KiB (12 bits)
+ // PDE0.BFS tells us how many of these 4 KiB page each PTE covers
+ // So add those bits in.
+ // We also calculated the PTE mask up above, to know how many PTEs are in this PTB
+ pte_idx = (address >> (12 + pde0_block_fragment_size)) & ptb_mask;
pte_further:
// now read PTE entry for this page
prev_addr = pde_fields.pte_base_addr + pte_idx*8;
@@ -778,20 +841,74 @@ pde_is_pte:
pte_fields.fragment,
pte_fields.further);
- if (pte_fields.further) {
- if (page_table_block_size == 9) {
- // this case doesn't make sense unless we support PTBS > 9
- asic->mem_funcs.vm_message("[ERROR]: PTE.further is set and *CNTL.PAGE_TABLE_BLOCK_SIZE is 9...\n");
- return -1;
+ // How many bits in the address are used to index into the PTB?
+ // If further is set, that means we jumped back to pde_is_pte,
+ // and the va_mask was properly set down there.
+ if (!further) {
+ // total_vm_bits are all the bits in the VM space
+ // We want to ignore the top-most PDB, which uses top_pdb_bits
+ // We also want to ignore lower PDBs, which use 9 bits each
+ int bits_to_use = total_vm_bits - top_pdb_bits - (9 * (page_table_depth - 1));
+
+ // At a minimum, we want to ignore the bottom 12 bits for a 4 KiB page
+ int lower_bits_to_ignore = 12;
+
+ if (pde_fields.pte) {
+ // We are in here because we're in PDE0 with P bit. So we don't want
+ // to skip the 9 bits from PDB0.
+ bits_to_use += 9;
+
+ // If the P bit is set, we are coming from PDE0, thus this entry
+ // covers the whole page_table_block_size, instead of the PDE0.BFS.
+ // So we want to ignore those bits in the address.
+ lower_bits_to_ignore += page_table_block_size;
} else {
- pte_idx = (address >> 12) & ((1ULL << pde0_block_fragment_size) - 1);
- pte_page_mask = (1ULL << 12) - 1;
-
- // grab PTE base address from the PTE that has the F bit set.
- pde_fields.pte_base_addr = pte_fields.page_base_addr;
- further = 1;
- goto pte_further;
+ // If we are at an actual PTE, then based on PDE0.BFS, we want to ignore
+ // some of the lowest bits.
+ // If PDE0.BFS=0, the bottom 12 bits are used to index within the page
+ // If PDE0.BFS=9, the bottom 21 bits are used to index within the page
+ // etc. These are the bits we want to ignore, and we already put 12 in.
+ lower_bits_to_ignore += pde0_block_fragment_size;
}
+
+ va_mask = (1 << bits_to_use) - 1;
+ int mask_to_ignore = (1 << lower_bits_to_ignore) - 1;
+ va_mask = va_mask & ~mask_to_ignore;
+ }
+
+ uint32_t pte_block_fragment_size = 0;
+ if (pte_fields.further) {
+ // Going to go one more layer deep, so now we need the Further-PTE's
+ // block_fragment_size. This tells us how many 4K pages each
+ // last-layer-PTE covers.
+ pte_block_fragment_size = (pte_entry >> 59) & 0x1F;
+
+ // Each entry covers the Further-PTE.block_fragment_size numbesr
+ // of 4K pages so we can potentially ignore some low-order bits.
+ int last_level_ptb_bits = 12 + pte_block_fragment_size;
+ pte_idx = address >> last_level_ptb_bits;
+
+ // The total size covered by the last-layer-PTB is a function of
+ // pde0_block_fragment_size, which tells us how many 4K entries the
+ // PTB covers.
+ // So number of bits needed to index the entries in the final PTE is:
+ uint32_t num_entry_bits = pde0_block_fragment_size - pte_block_fragment_size;
+ // Clamp the index to the new last-level PTB's size.
+ pte_idx &= ((1 << num_entry_bits) - 1);
+
+ uint32_t upper_mask = (1ULL << (12 + pde0_block_fragment_size)) - 1;
+ pte_page_mask = (1ULL << last_level_ptb_bits) - 1;
+ va_mask &= (upper_mask & ~pte_page_mask);
+
+ // grab PTE base address and other data from the PTE that has the F bit set.
+ pde_fields.frag_size = (pte_entry >> 59) & 0x1F;
+ pde_fields.pte_base_addr = pte_entry & 0xFFFFFFFFFFC0ULL;
+ pde_fields.valid = pte_entry & 1;
+ pde_fields.system = (pte_entry >> 1) & 1;
+ pde_fields.cache = (pte_entry >> 2) & 1;
+ pde_fields.pte = 0;
+ further = 1;
+ goto pte_further;
}
if (!pte_fields.system)
@@ -802,11 +919,10 @@ pde_is_pte:
// compute starting address
// this also accounts for PDE-is-PTE masking since current_depth > 0 at this point
- // if we are processing a PTE leaf node then the page size is 12 bits
if (!further)
offset_mask = (1ULL << ((current_depth * 9) + (12 + pde0_block_fragment_size))) - 1;
else
- offset_mask = (1ULL << 12) - 1; // offset masks are always 12-bits wide with PTE.further set
+ offset_mask = (1ULL << (12 + pte_block_fragment_size)) - 1;
start_addr = asic->mem_funcs.gpu_bus_to_cpu_address(asic, pte_fields.page_base_addr) + (address & offset_mask);
} else {
@@ -935,15 +1051,6 @@ invalid_page:
return -1;
}
-/** round_up_pot -- Round up value to next power of two */
-static uint64_t round_up_pot(uint64_t x)
-{
- uint64_t y = (64ULL * 1024 * 1024); // start at 64MiB
- while (y < x)
- y <<= 1;
- return y;
-}
-
/**
* umr_access_vram - Access GPU mapped memory
*
--
2.20.1
More information about the amd-gfx
mailing list