Timothy Pearson (tpearson@raptorengineeringinc.com) just uploaded a new patch set to gerrit, which you can find at http://review.coreboot.org/12036
-gerrit
commit 0caa25c08b3c46ade8e800eae4b2e98d8abc6e13 Author: Timothy Pearson tpearson@raptorengineeringinc.com Date: Sat Aug 8 20:29:27 2015 -0500
northbridge/amd/amdfam10: Add Family 15h cache partitioning support
Certain workloads may evict too many lines of other cores from the L3 cache if configured as one monolithic shared cache region.
Forcibly partition L3 cache to improve performance.
Change-Id: Ie4e28dd886aaa1c586b0919c5fe87ef1696f47e9 Signed-off-by: Timothy Pearson tpearson@raptorengineeringinc.com --- src/northbridge/amd/amdfam10/northbridge.c | 94 ++++++++++++++++++++++++++++++ 1 file changed, 94 insertions(+)
diff --git a/src/northbridge/amd/amdfam10/northbridge.c b/src/northbridge/amd/amdfam10/northbridge.c index 17dabae..ced64aa 100644 --- a/src/northbridge/amd/amdfam10/northbridge.c +++ b/src/northbridge/amd/amdfam10/northbridge.c @@ -1805,9 +1805,103 @@ static void detect_and_enable_probe_filter(device_t dev) } }
+static void detect_and_enable_cache_partitioning(device_t dev) +{ + uint8_t i; + uint32_t dword; + + if (is_fam15h()) { + printk(BIOS_DEBUG, "Enabling L3 cache partitioning\n"); + + uint32_t f5x80; + uint8_t cu_enabled; + uint8_t compute_unit_count = 0; + + uint32_t f3xe8; + uint8_t dual_node = 0; + + for (i = 0; i < sysconf.nodes; i++) { + device_t f3x_dev = dev_find_slot(0, PCI_DEVFN(0x18 + i, 3)); + device_t f4x_dev = dev_find_slot(0, PCI_DEVFN(0x18 + i, 4)); + device_t f5x_dev = dev_find_slot(0, PCI_DEVFN(0x18 + i, 5)); + + f3xe8 = pci_read_config32(f3x_dev, 0xe8); + + /* Check for dual node capability */ + if (f3xe8 & 0x20000000) + dual_node = 1; + + /* Determine the number of active compute units on this node */ + f5x80 = pci_read_config32(f5x_dev, 0x80); + cu_enabled = f5x80 & 0xf; + if (cu_enabled == 0x1) + compute_unit_count = 1; + if (cu_enabled == 0x3) + compute_unit_count = 2; + if (cu_enabled == 0x7) + compute_unit_count = 3; + if (cu_enabled == 0xf) + compute_unit_count = 4; + + /* Disable BAN mode */ + dword = pci_read_config32(f3x_dev, 0x1b8); + dword &= ~(0x7 << 19); /* L3BanMode = 0x0 */ + pci_write_config32(f3x_dev, 0x1b8, dword); + + /* Set up cache mapping */ + dword = pci_read_config32(f4x_dev, 0x1d4); + if (compute_unit_count == 1) { + dword |= 0xf; /* ComputeUnit0SubCacheEn = 0xf */ + } + if (compute_unit_count == 2) { + dword &= ~(0xf << 4); /* ComputeUnit1SubCacheEn = 0xc */ + dword |= (0xc << 4); + dword &= ~0xf; /* ComputeUnit0SubCacheEn = 0x3 */ + dword |= 0x3; + } + if (compute_unit_count == 3) { + dword &= ~(0xf << 8); /* ComputeUnit2SubCacheEn = 0x8 */ + dword |= (0x8 << 8); + dword &= ~(0xf << 4); /* ComputeUnit1SubCacheEn = 0x4 */ + dword |= (0x4 << 4); + dword &= ~0xf; /* ComputeUnit0SubCacheEn = 0x3 */ + dword |= 0x3; + } + if (compute_unit_count == 4) { + dword &= ~(0xf << 12); /* ComputeUnit3SubCacheEn = 0x8 */ + dword |= (0x8 << 12); + dword &= ~(0xf << 8); /* ComputeUnit2SubCacheEn = 0x4 */ + dword |= (0x4 << 8); + dword &= ~(0xf << 4); /* ComputeUnit1SubCacheEn = 0x2 */ + dword |= (0x2 << 4); + dword &= ~0xf; /* ComputeUnit0SubCacheEn = 0x1 */ + dword |= 0x1; + } + pci_write_config32(f4x_dev, 0x1d4, dword); + + /* Enable cache partitioning */ + pci_write_config32(f4x_dev, 0x1d4, dword); + if (compute_unit_count == 1) { + dword &= ~(0xf << 26); /* MaskUpdateForComputeUnit = 0x1 */ + dword |= (0x1 << 26); + } else if (compute_unit_count == 2) { + dword &= ~(0xf << 26); /* MaskUpdateForComputeUnit = 0x3 */ + dword |= (0x3 << 26); + } else if (compute_unit_count == 3) { + dword &= ~(0xf << 26); /* MaskUpdateForComputeUnit = 0x7 */ + dword |= (0x7 << 26); + } else if (compute_unit_count == 4) { + dword |= (0xf << 26); /* MaskUpdateForComputeUnit = 0xf */ + } + pci_write_config32(f4x_dev, 0x1d4, dword); + } + } +} + static void cpu_bus_init(device_t dev) { detect_and_enable_probe_filter(dev); + detect_and_enable_cache_partitioning(dev); initialize_cpus(dev->link_list); #if CONFIG_AMD_SB_CIMX sb_After_Pci_Init();