* [PATCH 0/2] AMD Address Translation Library
@ 2023-08-02 18:55 Yazen Ghannam
2023-08-02 18:55 ` [PATCH 1/2] platform/x86/amd: Introduce " Yazen Ghannam
2023-08-02 18:55 ` [PATCH 2/2] EDAC/amd64: Use new " Yazen Ghannam
0 siblings, 2 replies; 15+ messages in thread
From: Yazen Ghannam @ 2023-08-02 18:55 UTC (permalink / raw)
To: bp, linux-edac, hdegoede, markgross, platform-driver-x86
Cc: linux-kernel, avadhut.naik, mario.limonciello, Yazen Ghannam
Hi all,
This set adds a new library to do AMD-specific address translation. The
first use case is for translating a Unified Memory Controller (UMC)
"Normalized" address to a system physical address. Another use case will
be to do a similar translation for certain CXL configurations. The only
user is EDAC at the moment. But this can be used in MCA and CXL
subsystems too. Patches coming soon...
Since this code is very much implementation-specific, I thought it'd be
appropriate to have it as a "platform driver". Having the option to
build as a module helps with development, but this will likely be
'built-in' to use for MCA and CXL in production.
Patch 1 adds the new code. This includes support for all current AMD
Zen-based systems with a couple of exceptions noted in the commit
message.
The code is based on AMD reference code. Much of this is arbitrary bit
arithmetic. But I tried my best to make clarifying comments and to
restructure the code to be easier to follow.
Also, I purposefully avoided "over-optimizing" for the same reason, and
also to leverage compile-time checks for bitfields, etc. For example,
there are many uses of FIELD_GET(), and this requires a constant
expression as input.
The reference code underwent a major refactor. Therefore, this latest
set is fresh start. I figure it's best to match the latest reference
rather than submit another revision based on old code that will need to
be refactored anyway.
There are many code paths that are reused between various interleaving
modes and Data Fabric revisions. And these aren't easily decoupled. So
run time checks are used for code flow rather than function pointers,
etc.
All the code is added within a single patch. Mostly, this was done to
get the "whole picture" of how things fit together. But I can break this
up into separate patches for each Data Fabric revision, if needed. I
also want to avoid taking the old code and incrementally refactoring.
Since the old code no longer matches the reference, I think it's simpler
to just add the new and delete the old.
Patch 2 removes the old code and switches the AMD64 EDAC module to use
the new code.
Previous set:
https://lore.kernel.org/r/20220127204115.384161-1-yazen.ghannam@amd.com
Notable changes from previous set:
1) Move code out of EDAC.
2) Another major refactor based on refactored reference code.
3) Addition of DF4 and DF4.5 support.
Thanks,
Yazen
Yazen Ghannam (2):
platform/x86/amd: Introduce AMD Address Translation Library
EDAC/amd64: Use new AMD Address Translation Library
MAINTAINERS | 7 +
drivers/edac/Kconfig | 1 +
drivers/edac/amd64_edac.c | 282 +--------
drivers/platform/x86/amd/Kconfig | 1 +
drivers/platform/x86/amd/Makefile | 1 +
drivers/platform/x86/amd/atl/Kconfig | 20 +
drivers/platform/x86/amd/atl/Makefile | 18 +
drivers/platform/x86/amd/atl/access.c | 107 ++++
drivers/platform/x86/amd/atl/core.c | 212 +++++++
drivers/platform/x86/amd/atl/dehash.c | 459 ++++++++++++++
drivers/platform/x86/amd/atl/denormalize.c | 644 ++++++++++++++++++++
drivers/platform/x86/amd/atl/internal.h | 307 ++++++++++
drivers/platform/x86/amd/atl/map.c | 659 +++++++++++++++++++++
drivers/platform/x86/amd/atl/reg_fields.h | 603 +++++++++++++++++++
drivers/platform/x86/amd/atl/system.c | 282 +++++++++
drivers/platform/x86/amd/atl/umc.c | 53 ++
include/linux/amd-atl.h | 18 +
17 files changed, 3398 insertions(+), 276 deletions(-)
create mode 100644 drivers/platform/x86/amd/atl/Kconfig
create mode 100644 drivers/platform/x86/amd/atl/Makefile
create mode 100644 drivers/platform/x86/amd/atl/access.c
create mode 100644 drivers/platform/x86/amd/atl/core.c
create mode 100644 drivers/platform/x86/amd/atl/dehash.c
create mode 100644 drivers/platform/x86/amd/atl/denormalize.c
create mode 100644 drivers/platform/x86/amd/atl/internal.h
create mode 100644 drivers/platform/x86/amd/atl/map.c
create mode 100644 drivers/platform/x86/amd/atl/reg_fields.h
create mode 100644 drivers/platform/x86/amd/atl/system.c
create mode 100644 drivers/platform/x86/amd/atl/umc.c
create mode 100644 include/linux/amd-atl.h
--
2.34.1
^ permalink raw reply [flat|nested] 15+ messages in thread
* [PATCH 1/2] platform/x86/amd: Introduce AMD Address Translation Library
2023-08-02 18:55 [PATCH 0/2] AMD Address Translation Library Yazen Ghannam
@ 2023-08-02 18:55 ` Yazen Ghannam
2023-08-07 20:44 ` Yazen Ghannam
2023-08-02 18:55 ` [PATCH 2/2] EDAC/amd64: Use new " Yazen Ghannam
1 sibling, 1 reply; 15+ messages in thread
From: Yazen Ghannam @ 2023-08-02 18:55 UTC (permalink / raw)
To: bp, linux-edac, hdegoede, markgross, platform-driver-x86
Cc: linux-kernel, avadhut.naik, mario.limonciello, Yazen Ghannam
AMD Zen-based systems report memory errors through Machine Check banks
representing Unified Memory Controllers (UMCs). The address value
reported for DRAM ECC errors is a "normalized address" that is relative
to the UMC. This normalized address must be converted to a system
physical address to be usable by the OS.
Support for this address translation was introduced to the MCA subsystem
with Zen1 systems. The code was later moved to the AMD64 EDAC module,
since this was the only user of the code at the time.
However, there are uses for this translation outside of EDAC. The system
physical address can be used in MCA for preemptive page offlining as done
in some MCA notifier functions. Also, this translation is needed as the
basis of similar functionality needed for some CXL configurations on AMD
systems.
Introduce a common address translation library that can be used for
multiple subsystems including MCA, EDAC, and CXL.
Include support for UMC normalized to system physical address
translation for current CPU systems.
Future development to include:
- DF4.5 Non-power-of-2 interleaving modes.
- Heterogeneous CPU+GPU system support.
- CXL translation support.
- Caching of common intermediate values and results.
- Leverage UEFI PRM methods as alternate backends to existing native
code.
Signed-off-by: Yazen Ghannam <yazen.ghannam@amd.com>
---
MAINTAINERS | 7 +
drivers/platform/x86/amd/Kconfig | 1 +
drivers/platform/x86/amd/Makefile | 1 +
drivers/platform/x86/amd/atl/Kconfig | 20 +
drivers/platform/x86/amd/atl/Makefile | 18 +
drivers/platform/x86/amd/atl/access.c | 107 ++++
drivers/platform/x86/amd/atl/core.c | 212 +++++++
drivers/platform/x86/amd/atl/dehash.c | 459 ++++++++++++++
drivers/platform/x86/amd/atl/denormalize.c | 644 ++++++++++++++++++++
drivers/platform/x86/amd/atl/internal.h | 307 ++++++++++
drivers/platform/x86/amd/atl/map.c | 659 +++++++++++++++++++++
drivers/platform/x86/amd/atl/reg_fields.h | 603 +++++++++++++++++++
drivers/platform/x86/amd/atl/system.c | 282 +++++++++
drivers/platform/x86/amd/atl/umc.c | 53 ++
include/linux/amd-atl.h | 18 +
15 files changed, 3391 insertions(+)
create mode 100644 drivers/platform/x86/amd/atl/Kconfig
create mode 100644 drivers/platform/x86/amd/atl/Makefile
create mode 100644 drivers/platform/x86/amd/atl/access.c
create mode 100644 drivers/platform/x86/amd/atl/core.c
create mode 100644 drivers/platform/x86/amd/atl/dehash.c
create mode 100644 drivers/platform/x86/amd/atl/denormalize.c
create mode 100644 drivers/platform/x86/amd/atl/internal.h
create mode 100644 drivers/platform/x86/amd/atl/map.c
create mode 100644 drivers/platform/x86/amd/atl/reg_fields.h
create mode 100644 drivers/platform/x86/amd/atl/system.c
create mode 100644 drivers/platform/x86/amd/atl/umc.c
create mode 100644 include/linux/amd-atl.h
diff --git a/MAINTAINERS b/MAINTAINERS
index d516295978a4..323df95c9cf2 100644
--- a/MAINTAINERS
+++ b/MAINTAINERS
@@ -891,6 +891,13 @@ Q: https://patchwork.kernel.org/project/linux-rdma/list/
F: drivers/infiniband/hw/efa/
F: include/uapi/rdma/efa-abi.h
+AMD ADDRESS TRANSLATION LIBRARY (ATL)
+M: Yazen Ghannam <Yazen.Ghannam@amd.com>
+L: platform-driver-x86@vger.kernel.org
+S: Supported
+F: drivers/platform/x86/amd/atl
+F: include/linux/amd-atl.h
+
AMD CDX BUS DRIVER
M: Nipun Gupta <nipun.gupta@amd.com>
M: Nikhil Agarwal <nikhil.agarwal@amd.com>
diff --git a/drivers/platform/x86/amd/Kconfig b/drivers/platform/x86/amd/Kconfig
index d9685aef0887..7651a3f3f77f 100644
--- a/drivers/platform/x86/amd/Kconfig
+++ b/drivers/platform/x86/amd/Kconfig
@@ -4,6 +4,7 @@
#
source "drivers/platform/x86/amd/pmf/Kconfig"
+source "drivers/platform/x86/amd/atl/Kconfig"
config AMD_PMC
tristate "AMD SoC PMC driver"
diff --git a/drivers/platform/x86/amd/Makefile b/drivers/platform/x86/amd/Makefile
index 65732f0a3913..fe396d3aeeb8 100644
--- a/drivers/platform/x86/amd/Makefile
+++ b/drivers/platform/x86/amd/Makefile
@@ -9,3 +9,4 @@ obj-$(CONFIG_AMD_PMC) += amd-pmc.o
amd_hsmp-y := hsmp.o
obj-$(CONFIG_AMD_HSMP) += amd_hsmp.o
obj-$(CONFIG_AMD_PMF) += pmf/
+obj-$(CONFIG_AMD_ATL) += atl/
diff --git a/drivers/platform/x86/amd/atl/Kconfig b/drivers/platform/x86/amd/atl/Kconfig
new file mode 100644
index 000000000000..c863bf442ade
--- /dev/null
+++ b/drivers/platform/x86/amd/atl/Kconfig
@@ -0,0 +1,20 @@
+# SPDX-License-Identifier: GPL-2.0-or-later
+#
+# AMD Address Translation Library Kconfig
+#
+# Copyright (c) 2023, Advanced Micro Devices, Inc.
+# All Rights Reserved.
+#
+# Author: Yazen Ghannam <Yazen.Ghannam@amd.com>
+
+
+config AMD_ATL
+ tristate "AMD Address Translation Library"
+ depends on AMD_NB
+ help
+ This library includes support for implementation-specific
+ address translation procedures needed for various error
+ handling cases.
+
+ Enable this option if using DRAM ECC on Zen-based systems
+ and OS-based error handling.
diff --git a/drivers/platform/x86/amd/atl/Makefile b/drivers/platform/x86/amd/atl/Makefile
new file mode 100644
index 000000000000..e62faefa6dc1
--- /dev/null
+++ b/drivers/platform/x86/amd/atl/Makefile
@@ -0,0 +1,18 @@
+# SPDX-License-Identifier: GPL-2.0-or-later
+#
+# AMD Address Translation Library Makefile
+#
+# Copyright (c) 2023, Advanced Micro Devices, Inc.
+# All Rights Reserved.
+#
+# Author: Yazen Ghannam <Yazen.Ghannam@amd.com>
+
+amd_atl-y := access.o
+amd_atl-y += core.o
+amd_atl-y += dehash.o
+amd_atl-y += denormalize.o
+amd_atl-y += map.o
+amd_atl-y += system.o
+amd_atl-y += umc.o
+
+obj-$(CONFIG_AMD_ATL) += amd_atl.o
diff --git a/drivers/platform/x86/amd/atl/access.c b/drivers/platform/x86/amd/atl/access.c
new file mode 100644
index 000000000000..888d47d498e0
--- /dev/null
+++ b/drivers/platform/x86/amd/atl/access.c
@@ -0,0 +1,107 @@
+// SPDX-License-Identifier: GPL-2.0-or-later
+/*
+ * AMD Address Translation Library
+ *
+ * access.c : DF Indirect Access functions
+ *
+ * Copyright (c) 2023, Advanced Micro Devices, Inc.
+ * All Rights Reserved.
+ *
+ * Author: Yazen Ghannam <Yazen.Ghannam@amd.com>
+ */
+
+#include "internal.h"
+
+/* Protect the PCI config register pairs used for DF indirect access. */
+static DEFINE_MUTEX(df_indirect_mutex);
+
+/*
+ * Data Fabric Indirect Access uses FICAA/FICAD.
+ *
+ * Fabric Indirect Configuration Access Address (FICAA): Constructed based
+ * on the device's Instance Id and the PCI function and register offset of
+ * the desired register.
+ *
+ * Fabric Indirect Configuration Access Data (FICAD): There are FICAD LO
+ * and FICAD HI registers but so far we only need the LO register.
+ *
+ * Use Instance Id 0xFF to indicate a broadcast read.
+ */
+#define DF_BROADCAST 0xFF
+
+#define DF_FICAA_INST_EN BIT(0)
+#define DF_FICAA_REG_NUM GENMASK(10, 1)
+#define DF_FICAA_FUNC_NUM GENMASK(13, 11)
+#define DF_FICAA_INST_ID GENMASK(23, 16)
+
+/* Register field changed in new systems. */
+#define DF_FICAA_REG_NUM_LEGACY GENMASK(10, 2)
+
+static int __df_indirect_read(u16 node, u8 func, u16 reg, u8 instance_id, u32 *lo)
+{
+ u32 ficaa_addr = 0x8C, ficad_addr = 0xB8;
+ struct pci_dev *F4;
+ int err = -ENODEV;
+ u32 ficaa = 0;
+
+ if (node >= amd_nb_num())
+ goto out;
+
+ F4 = node_to_amd_nb(node)->link;
+ if (!F4)
+ goto out;
+
+ /* Enable instance-specific access. */
+ if (instance_id != DF_BROADCAST) {
+ ficaa |= FIELD_PREP(DF_FICAA_INST_EN, 1);
+ ficaa |= FIELD_PREP(DF_FICAA_INST_ID, instance_id);
+ }
+
+ /*
+ * The two least-significant bits are masked when inputing the
+ * register offset to FICAA.
+ */
+ reg >>= 2;
+
+ if (df_cfg.flags.legacy_ficaa) {
+ ficaa_addr = 0x5C;
+ ficad_addr = 0x98;
+
+ ficaa |= FIELD_PREP(DF_FICAA_REG_NUM_LEGACY, reg);
+ } else {
+ ficaa |= FIELD_PREP(DF_FICAA_REG_NUM, reg);
+ }
+
+ ficaa |= FIELD_PREP(DF_FICAA_FUNC_NUM, func);
+
+ mutex_lock(&df_indirect_mutex);
+
+ err = pci_write_config_dword(F4, ficaa_addr, ficaa);
+ if (err) {
+ pr_warn("Error writing DF Indirect FICAA, FICAA=0x%x\n", ficaa);
+ goto out_unlock;
+ }
+
+ err = pci_read_config_dword(F4, ficad_addr, lo);
+ if (err)
+ pr_warn("Error reading DF Indirect FICAD LO, FICAA=0x%x.\n", ficaa);
+
+ atl_debug("node=%u inst=0x%x func=0x%x reg=0x%x val=0x%x",
+ node, instance_id, func, reg << 2, *lo);
+
+out_unlock:
+ mutex_unlock(&df_indirect_mutex);
+
+out:
+ return err;
+}
+
+int df_indirect_read_instance(u16 node, u8 func, u16 reg, u8 instance_id, u32 *lo)
+{
+ return __df_indirect_read(node, func, reg, instance_id, lo);
+}
+
+int df_indirect_read_broadcast(u16 node, u8 func, u16 reg, u32 *lo)
+{
+ return __df_indirect_read(node, func, reg, DF_BROADCAST, lo);
+}
diff --git a/drivers/platform/x86/amd/atl/core.c b/drivers/platform/x86/amd/atl/core.c
new file mode 100644
index 000000000000..8c997c7ae8a6
--- /dev/null
+++ b/drivers/platform/x86/amd/atl/core.c
@@ -0,0 +1,212 @@
+// SPDX-License-Identifier: GPL-2.0-or-later
+/*
+ * AMD Address Translation Library
+ *
+ * core.c : Module init and base translation functions
+ *
+ * Copyright (c) 2023, Advanced Micro Devices, Inc.
+ * All Rights Reserved.
+ *
+ * Author: Yazen Ghannam <Yazen.Ghannam@amd.com>
+ */
+
+#include <linux/module.h>
+
+#include "internal.h"
+
+struct df_config df_cfg __read_mostly;
+
+static int addr_over_limit(struct addr_ctx *ctx)
+{
+ u64 dram_limit_addr;
+
+ if (df_cfg.rev >= DF4)
+ dram_limit_addr = FIELD_GET(DF4_DRAM_LIMIT_ADDR, ctx->map.limit);
+ else
+ dram_limit_addr = FIELD_GET(DF2_DRAM_LIMIT_ADDR, ctx->map.limit);
+
+ dram_limit_addr <<= DF_DRAM_BASE_LIMIT_LSB;
+ dram_limit_addr |= GENMASK(DF_DRAM_BASE_LIMIT_LSB - 1, 0);
+
+ /* Is calculated system address above DRAM limit address? */
+ if (ctx->ret_addr > dram_limit_addr)
+ return -EINVAL;
+
+ return 0;
+}
+
+static bool legacy_hole_en(struct addr_ctx *ctx)
+{
+ u32 reg = ctx->map.base;
+
+ if (df_cfg.rev >= DF4)
+ reg = ctx->map.ctl;
+
+ return FIELD_GET(DF_LEGACY_MMIO_HOLE_EN, reg);
+}
+
+static int add_legacy_hole(struct addr_ctx *ctx)
+{
+ u32 dram_hole_base;
+ u8 func = 0;
+
+ if (!legacy_hole_en(ctx))
+ return 0;
+
+ if (df_cfg.rev >= DF4)
+ func = 7;
+
+ if (df_indirect_read_broadcast(ctx->node_id, func, 0x104, &dram_hole_base))
+ return -EINVAL;
+
+ dram_hole_base &= DF_DRAM_HOLE_BASE_MASK;
+
+ if (ctx->ret_addr >= dram_hole_base)
+ ctx->ret_addr += (BIT_ULL(32) - dram_hole_base);
+
+ return 0;
+}
+
+static u64 get_base_addr(struct addr_ctx *ctx)
+{
+ u64 base_addr;
+
+ if (df_cfg.rev >= DF4)
+ base_addr = FIELD_GET(DF4_BASE_ADDR, ctx->map.base);
+ else
+ base_addr = FIELD_GET(DF2_BASE_ADDR, ctx->map.base);
+
+ return base_addr << DF_DRAM_BASE_LIMIT_LSB;
+}
+
+static int add_base_and_hole(struct addr_ctx *ctx)
+{
+ ctx->ret_addr += get_base_addr(ctx);
+
+ if (add_legacy_hole(ctx))
+ return -EINVAL;
+
+ return 0;
+}
+
+static bool late_hole_remove(struct addr_ctx *ctx)
+{
+ if (df_cfg.rev == DF3p5)
+ return true;
+
+ if (df_cfg.rev == DF4)
+ return true;
+
+ if (ctx->map.intlv_mode == DF3_6CHAN)
+ return true;
+
+ return false;
+}
+
+int norm_to_sys_addr(u8 socket_id, u8 die_id, u8 cs_inst_id, u64 *addr)
+{
+ struct addr_ctx ctx;
+
+ if (df_cfg.rev == UNKNOWN)
+ return -EINVAL;
+
+ memset(&ctx, 0, sizeof(ctx));
+
+ /* We start from the normalized address */
+ ctx.ret_addr = *addr;
+ ctx.inst_id = cs_inst_id;
+
+ if (determine_node_id(&ctx, socket_id, die_id)) {
+ pr_warn("Failed to determine Node ID");
+ return -EINVAL;
+ }
+
+ if (get_address_map(&ctx)) {
+ pr_warn("Failed to get address maps");
+ return -EINVAL;
+ }
+
+ if (denormalize_address(&ctx)) {
+ pr_warn("Failed to denormalize address");
+ return -EINVAL;
+ }
+
+ if (!late_hole_remove(&ctx) && add_base_and_hole(&ctx)) {
+ pr_warn("Failed to add DRAM base address and hole");
+ return -EINVAL;
+ }
+
+ if (dehash_address(&ctx)) {
+ pr_warn("Failed to dehash address");
+ return -EINVAL;
+ }
+
+ if (late_hole_remove(&ctx) && add_base_and_hole(&ctx)) {
+ pr_warn("Failed to add DRAM base address and hole");
+ return -EINVAL;
+ }
+
+ if (addr_over_limit(&ctx)) {
+ pr_warn("Calculated address is over limit");
+ return -EINVAL;
+ }
+
+ *addr = ctx.ret_addr;
+ return 0;
+}
+
+static void check_for_legacy_df_access(void)
+{
+ /*
+ * All Zen-based systems before Family 19h use the legacy
+ * DF Indirect Access (FICAA/FICAD) offsets.
+ */
+ if (boot_cpu_data.x86 < 0x19) {
+ df_cfg.flags.legacy_ficaa = true;
+ return;
+ }
+
+ /* All systems after Family 19h use the current offsets. */
+ if (boot_cpu_data.x86 > 0x19)
+ return;
+
+ /* Some Family 19h systems use the legacy offsets. */
+ switch (boot_cpu_data.x86_model) {
+ case 0x00 ... 0x0f:
+ case 0x20 ... 0x5f:
+ df_cfg.flags.legacy_ficaa = true;
+ }
+}
+
+static int __init amd_atl_init(void)
+{
+ if (boot_cpu_data.x86_vendor != X86_VENDOR_AMD)
+ return -ENODEV;
+
+ check_for_legacy_df_access();
+
+ /*
+ * Not sure if this should return an error code.
+ * That may prevent other modules from loading.
+ *
+ * For now, don't fail out. The translation function
+ * will do a check and return early if the DF revision
+ * is not set.
+ */
+ if (get_df_system_info()) {
+ pr_warn("Failed to get DF information");
+ df_cfg.rev = UNKNOWN;
+ }
+
+ pr_info("AMD Address Translation Library initialized");
+ return 0;
+}
+
+static void __exit amd_atl_exit(void)
+{
+}
+
+module_init(amd_atl_init);
+module_exit(amd_atl_exit);
+
+MODULE_LICENSE("GPL");
diff --git a/drivers/platform/x86/amd/atl/dehash.c b/drivers/platform/x86/amd/atl/dehash.c
new file mode 100644
index 000000000000..e501f2e918d7
--- /dev/null
+++ b/drivers/platform/x86/amd/atl/dehash.c
@@ -0,0 +1,459 @@
+// SPDX-License-Identifier: GPL-2.0-or-later
+/*
+ * AMD Address Translation Library
+ *
+ * dehash.c : Functions to account for hashing bits
+ *
+ * Copyright (c) 2023, Advanced Micro Devices, Inc.
+ * All Rights Reserved.
+ *
+ * Author: Yazen Ghannam <Yazen.Ghannam@amd.com>
+ */
+
+#include "internal.h"
+
+static int df2_dehash_addr(struct addr_ctx *ctx)
+{
+ u8 hashed_bit, intlv_bit;
+
+ /* Assert that interleave bit is 8 or 9. */
+ if (ctx->map.intlv_bit_pos != 8 && ctx->map.intlv_bit_pos != 9) {
+ pr_warn("%s: Invalid interleave bit: %u",
+ __func__, ctx->map.intlv_bit_pos);
+ return -EINVAL;
+ }
+
+ /* Assert that die and socket interleaving are disabled. */
+ if (ctx->map.num_intlv_dies > 1) {
+ pr_warn("%s: Invalid number of interleave dies: %u",
+ __func__, ctx->map.num_intlv_dies);
+ return -EINVAL;
+ }
+
+ if (ctx->map.num_intlv_sockets > 1) {
+ pr_warn("%s: Invalid number of interleave sockets: %u",
+ __func__, ctx->map.num_intlv_sockets);
+ return -EINVAL;
+ }
+
+ if (ctx->map.intlv_bit_pos == 8)
+ intlv_bit = FIELD_GET(BIT_ULL(8), ctx->ret_addr);
+ else
+ intlv_bit = FIELD_GET(BIT_ULL(9), ctx->ret_addr);
+
+ hashed_bit = intlv_bit;
+ hashed_bit ^= FIELD_GET(BIT_ULL(12), ctx->ret_addr);
+ hashed_bit ^= FIELD_GET(BIT_ULL(18), ctx->ret_addr);
+ hashed_bit ^= FIELD_GET(BIT_ULL(21), ctx->ret_addr);
+ hashed_bit ^= FIELD_GET(BIT_ULL(30), ctx->ret_addr);
+
+ if (hashed_bit != intlv_bit) {
+ if (ctx->map.intlv_bit_pos == 8)
+ ctx->ret_addr ^= BIT_ULL(8);
+ else
+ ctx->ret_addr ^= BIT_ULL(9);
+ }
+
+ return 0;
+}
+
+static int df3_dehash_addr(struct addr_ctx *ctx)
+{
+ bool hash_ctl_64k, hash_ctl_2M, hash_ctl_1G;
+ u8 hashed_bit, intlv_bit;
+
+ /* Assert that interleave bit is 8 or 9. */
+ if (ctx->map.intlv_bit_pos != 8 && ctx->map.intlv_bit_pos != 9) {
+ pr_warn("%s: Invalid interleave bit: %u",
+ __func__, ctx->map.intlv_bit_pos);
+ return -EINVAL;
+ }
+
+ /* Assert that die and socket interleaving are disabled. */
+ if (ctx->map.num_intlv_dies > 1) {
+ pr_warn("%s: Invalid number of interleave dies: %u",
+ __func__, ctx->map.num_intlv_dies);
+ return -EINVAL;
+ }
+
+ if (ctx->map.num_intlv_sockets > 1) {
+ pr_warn("%s: Invalid number of interleave sockets: %u",
+ __func__, ctx->map.num_intlv_sockets);
+ return -EINVAL;
+ }
+
+ hash_ctl_64k = FIELD_GET(DF3_HASH_CTL_64K, ctx->map.ctl);
+ hash_ctl_2M = FIELD_GET(DF3_HASH_CTL_2M, ctx->map.ctl);
+ hash_ctl_1G = FIELD_GET(DF3_HASH_CTL_1G, ctx->map.ctl);
+
+ if (ctx->map.intlv_bit_pos == 8)
+ intlv_bit = FIELD_GET(BIT_ULL(8), ctx->ret_addr);
+ else
+ intlv_bit = FIELD_GET(BIT_ULL(9), ctx->ret_addr);
+
+ hashed_bit = intlv_bit;
+ hashed_bit ^= FIELD_GET(BIT_ULL(14), ctx->ret_addr);
+ hashed_bit ^= FIELD_GET(BIT_ULL(18), ctx->ret_addr) & hash_ctl_64k;
+ hashed_bit ^= FIELD_GET(BIT_ULL(23), ctx->ret_addr) & hash_ctl_2M;
+ hashed_bit ^= FIELD_GET(BIT_ULL(32), ctx->ret_addr) & hash_ctl_1G;
+
+ if (hashed_bit != intlv_bit) {
+ if (ctx->map.intlv_bit_pos == 8)
+ ctx->ret_addr ^= BIT_ULL(8);
+ else
+ ctx->ret_addr ^= BIT_ULL(9);
+ }
+
+ /* Calculation complete for 2 channels. Continue for 4 and 8 channels. */
+ if (ctx->map.intlv_mode == DF3_COD4_2CHAN_HASH)
+ return 0;
+
+ intlv_bit = FIELD_GET(BIT_ULL(12), ctx->ret_addr);
+
+ hashed_bit = intlv_bit;
+ hashed_bit ^= FIELD_GET(BIT_ULL(16), ctx->ret_addr) & hash_ctl_64k;
+ hashed_bit ^= FIELD_GET(BIT_ULL(21), ctx->ret_addr) & hash_ctl_2M;
+ hashed_bit ^= FIELD_GET(BIT_ULL(30), ctx->ret_addr) & hash_ctl_1G;
+
+ if (hashed_bit != intlv_bit)
+ ctx->ret_addr ^= BIT_ULL(12);
+
+ /* Calculation complete for 4 channels. Continue for 8 channels. */
+ if (ctx->map.intlv_mode == DF3_COD2_4CHAN_HASH)
+ return 0;
+
+ intlv_bit = FIELD_GET(BIT_ULL(13), ctx->ret_addr);
+
+ hashed_bit = intlv_bit;
+ hashed_bit ^= FIELD_GET(BIT_ULL(17), ctx->ret_addr) & hash_ctl_64k;
+ hashed_bit ^= FIELD_GET(BIT_ULL(22), ctx->ret_addr) & hash_ctl_2M;
+ hashed_bit ^= FIELD_GET(BIT_ULL(31), ctx->ret_addr) & hash_ctl_1G;
+
+ if (hashed_bit != intlv_bit)
+ ctx->ret_addr ^= BIT_ULL(13);
+
+ return 0;
+}
+
+static int df3_6chan_dehash_addr(struct addr_ctx *ctx)
+{
+ u8 intlv_bit_pos = ctx->map.intlv_bit_pos;
+ u8 hashed_bit, intlv_bit, num_intlv_bits;
+ bool hash_ctl_2M, hash_ctl_1G;
+
+ if (ctx->map.intlv_mode != DF3_6CHAN)
+ return -EINVAL;
+
+ num_intlv_bits = ilog2(ctx->map.num_intlv_chan) + 1;
+
+ hash_ctl_2M = FIELD_GET(DF3_HASH_CTL_2M, ctx->map.ctl);
+ hash_ctl_1G = FIELD_GET(DF3_HASH_CTL_1G, ctx->map.ctl);
+
+ intlv_bit = atl_get_bit(intlv_bit_pos, ctx->ret_addr);
+
+ hashed_bit = intlv_bit;
+ hashed_bit ^= atl_get_bit((intlv_bit_pos + num_intlv_bits), ctx->ret_addr);
+ hashed_bit ^= FIELD_GET(BIT_ULL(23), ctx->ret_addr) & hash_ctl_2M;
+ hashed_bit ^= FIELD_GET(BIT_ULL(32), ctx->ret_addr) & hash_ctl_1G;
+
+ if (hashed_bit != intlv_bit)
+ ctx->ret_addr ^= BIT_ULL(intlv_bit_pos);
+
+ intlv_bit_pos++;
+ intlv_bit = atl_get_bit(intlv_bit_pos, ctx->ret_addr);
+
+ hashed_bit = intlv_bit;
+ hashed_bit ^= FIELD_GET(BIT_ULL(21), ctx->ret_addr) & hash_ctl_2M;
+ hashed_bit ^= FIELD_GET(BIT_ULL(30), ctx->ret_addr) & hash_ctl_1G;
+
+ if (hashed_bit != intlv_bit)
+ ctx->ret_addr ^= BIT_ULL(intlv_bit_pos);
+
+ intlv_bit_pos++;
+ intlv_bit = atl_get_bit(intlv_bit_pos, ctx->ret_addr);
+
+ hashed_bit = intlv_bit;
+ hashed_bit ^= FIELD_GET(BIT_ULL(22), ctx->ret_addr) & hash_ctl_2M;
+ hashed_bit ^= FIELD_GET(BIT_ULL(31), ctx->ret_addr) & hash_ctl_1G;
+
+ if (hashed_bit != intlv_bit)
+ ctx->ret_addr ^= BIT_ULL(intlv_bit_pos);
+
+ return 0;
+}
+
+static int df4_dehash_addr(struct addr_ctx *ctx)
+{
+ bool hash_ctl_64k, hash_ctl_2M, hash_ctl_1G;
+ u8 hashed_bit, intlv_bit;
+
+ /* Assert that interleave bit is 8. */
+ if (ctx->map.intlv_bit_pos != 8) {
+ pr_warn("%s: Invalid interleave bit: %u",
+ __func__, ctx->map.intlv_bit_pos);
+ return -EINVAL;
+ }
+
+ /* Assert that die interleaving is disabled. */
+ if (ctx->map.num_intlv_dies > 1) {
+ pr_warn("%s: Invalid number of interleave dies: %u",
+ __func__, ctx->map.num_intlv_dies);
+ return -EINVAL;
+ }
+
+ /* Assert that no more than 2 sockets are interleaved. */
+ if (ctx->map.num_intlv_sockets > 2) {
+ pr_warn("%s: Invalid number of interleave sockets: %u",
+ __func__, ctx->map.num_intlv_sockets);
+ return -EINVAL;
+ }
+
+ hash_ctl_64k = FIELD_GET(DF4_HASH_CTL_64K, ctx->map.ctl);
+ hash_ctl_2M = FIELD_GET(DF4_HASH_CTL_2M, ctx->map.ctl);
+ hash_ctl_1G = FIELD_GET(DF4_HASH_CTL_1G, ctx->map.ctl);
+
+ intlv_bit = FIELD_GET(BIT_ULL(8), ctx->ret_addr);
+
+ hashed_bit = intlv_bit;
+ hashed_bit ^= FIELD_GET(BIT_ULL(16), ctx->ret_addr) & hash_ctl_64k;
+ hashed_bit ^= FIELD_GET(BIT_ULL(21), ctx->ret_addr) & hash_ctl_2M;
+ hashed_bit ^= FIELD_GET(BIT_ULL(30), ctx->ret_addr) & hash_ctl_1G;
+
+ if (ctx->map.num_intlv_sockets == 1)
+ hashed_bit ^= FIELD_GET(BIT_ULL(14), ctx->ret_addr);
+
+ if (hashed_bit != intlv_bit)
+ ctx->ret_addr ^= BIT_ULL(8);
+
+ /*
+ * Hashing is possible with socket interleaving, so check the total number
+ * of channels in the system rather than DRAM map interleaving mode.
+ *
+ * Calculation complete for 2 channels. Continue for 4, 8, and 16 channels.
+ */
+ if (ctx->map.total_intlv_chan <= 2)
+ return 0;
+
+ intlv_bit = FIELD_GET(BIT_ULL(12), ctx->ret_addr);
+
+ hashed_bit = intlv_bit;
+ hashed_bit ^= FIELD_GET(BIT_ULL(17), ctx->ret_addr) & hash_ctl_64k;
+ hashed_bit ^= FIELD_GET(BIT_ULL(22), ctx->ret_addr) & hash_ctl_2M;
+ hashed_bit ^= FIELD_GET(BIT_ULL(31), ctx->ret_addr) & hash_ctl_1G;
+
+ if (hashed_bit != intlv_bit)
+ ctx->ret_addr ^= BIT_ULL(12);
+
+ /* Calculation complete for 4 channels. Continue for 8 and 16 channels. */
+ if (ctx->map.total_intlv_chan <= 4)
+ return 0;
+
+ intlv_bit = FIELD_GET(BIT_ULL(13), ctx->ret_addr);
+
+ hashed_bit = intlv_bit;
+ hashed_bit ^= FIELD_GET(BIT_ULL(18), ctx->ret_addr) & hash_ctl_64k;
+ hashed_bit ^= FIELD_GET(BIT_ULL(23), ctx->ret_addr) & hash_ctl_2M;
+ hashed_bit ^= FIELD_GET(BIT_ULL(32), ctx->ret_addr) & hash_ctl_1G;
+
+ if (hashed_bit != intlv_bit)
+ ctx->ret_addr ^= BIT_ULL(13);
+
+ /* Calculation complete for 8 channels. Continue for 16 channels. */
+ if (ctx->map.total_intlv_chan <= 8)
+ return 0;
+
+ intlv_bit = FIELD_GET(BIT_ULL(14), ctx->ret_addr);
+
+ hashed_bit = intlv_bit;
+ hashed_bit ^= FIELD_GET(BIT_ULL(19), ctx->ret_addr) & hash_ctl_64k;
+ hashed_bit ^= FIELD_GET(BIT_ULL(24), ctx->ret_addr) & hash_ctl_2M;
+ hashed_bit ^= FIELD_GET(BIT_ULL(33), ctx->ret_addr) & hash_ctl_1G;
+
+ if (hashed_bit != intlv_bit)
+ ctx->ret_addr ^= BIT_ULL(14);
+
+ return 0;
+}
+
+static int df4p5_dehash_addr(struct addr_ctx *ctx)
+{
+ bool hash_ctl_64k, hash_ctl_2M, hash_ctl_1G, hash_ctl_1T;
+ u8 hashed_bit, intlv_bit;
+ u64 rehash_vector;
+
+ /* Assert that interleave bit is 8. */
+ if (ctx->map.intlv_bit_pos != 8) {
+ pr_warn("%s: Invalid interleave bit: %u",
+ __func__, ctx->map.intlv_bit_pos);
+ return -EINVAL;
+ }
+
+ /* Assert that die interleaving is disabled. */
+ if (ctx->map.num_intlv_dies > 1) {
+ pr_warn("%s: Invalid number of interleave dies: %u",
+ __func__, ctx->map.num_intlv_dies);
+ return -EINVAL;
+ }
+
+ /* Assert that no more than 2 sockets are interleaved. */
+ if (ctx->map.num_intlv_sockets > 2) {
+ pr_warn("%s: Invalid number of interleave sockets: %u",
+ __func__, ctx->map.num_intlv_sockets);
+ return -EINVAL;
+ }
+
+ hash_ctl_64k = FIELD_GET(DF4_HASH_CTL_64K, ctx->map.ctl);
+ hash_ctl_2M = FIELD_GET(DF4_HASH_CTL_2M, ctx->map.ctl);
+ hash_ctl_1G = FIELD_GET(DF4_HASH_CTL_1G, ctx->map.ctl);
+ hash_ctl_1T = FIELD_GET(DF4_HASH_CTL_1T, ctx->map.ctl);
+
+ /*
+ * Generate a unique address to determine which bits
+ * need to be dehashed.
+ *
+ * Start with a contiguous bitmask for the total
+ * number of channels starting at bit 8.
+ *
+ * Then make a gap in the proper place based on
+ * interleave mode.
+ */
+ rehash_vector = ctx->map.total_intlv_chan - 1;
+ rehash_vector <<= 8;
+
+ if (ctx->map.intlv_mode == DF4p5_NPS2_4CHAN_1K_HASH ||
+ ctx->map.intlv_mode == DF4p5_NPS1_8CHAN_1K_HASH ||
+ ctx->map.intlv_mode == DF4p5_NPS1_16CHAN_1K_HASH)
+ rehash_vector = expand_bits(10, 2, rehash_vector);
+ else
+ rehash_vector = expand_bits(9, 3, rehash_vector);
+
+ if (rehash_vector & BIT_ULL(8)) {
+ intlv_bit = FIELD_GET(BIT_ULL(8), ctx->ret_addr);
+
+ hashed_bit = intlv_bit;
+ hashed_bit ^= FIELD_GET(BIT_ULL(16), ctx->ret_addr) & hash_ctl_64k;
+ hashed_bit ^= FIELD_GET(BIT_ULL(21), ctx->ret_addr) & hash_ctl_2M;
+ hashed_bit ^= FIELD_GET(BIT_ULL(30), ctx->ret_addr) & hash_ctl_1G;
+ hashed_bit ^= FIELD_GET(BIT_ULL(40), ctx->ret_addr) & hash_ctl_1T;
+
+ if (hashed_bit != intlv_bit)
+ ctx->ret_addr ^= BIT_ULL(8);
+ }
+
+ if (rehash_vector & BIT_ULL(9)) {
+ intlv_bit = FIELD_GET(BIT_ULL(9), ctx->ret_addr);
+
+ hashed_bit = intlv_bit;
+ hashed_bit ^= FIELD_GET(BIT_ULL(17), ctx->ret_addr) & hash_ctl_64k;
+ hashed_bit ^= FIELD_GET(BIT_ULL(22), ctx->ret_addr) & hash_ctl_2M;
+ hashed_bit ^= FIELD_GET(BIT_ULL(31), ctx->ret_addr) & hash_ctl_1G;
+ hashed_bit ^= FIELD_GET(BIT_ULL(41), ctx->ret_addr) & hash_ctl_1T;
+
+ if (hashed_bit != intlv_bit)
+ ctx->ret_addr ^= BIT_ULL(9);
+ }
+
+ if (rehash_vector & BIT_ULL(12)) {
+ intlv_bit = FIELD_GET(BIT_ULL(12), ctx->ret_addr);
+
+ hashed_bit = intlv_bit;
+ hashed_bit ^= FIELD_GET(BIT_ULL(18), ctx->ret_addr) & hash_ctl_64k;
+ hashed_bit ^= FIELD_GET(BIT_ULL(23), ctx->ret_addr) & hash_ctl_2M;
+ hashed_bit ^= FIELD_GET(BIT_ULL(32), ctx->ret_addr) & hash_ctl_1G;
+ hashed_bit ^= FIELD_GET(BIT_ULL(42), ctx->ret_addr) & hash_ctl_1T;
+
+ if (hashed_bit != intlv_bit)
+ ctx->ret_addr ^= BIT_ULL(12);
+ }
+
+ if (rehash_vector & BIT_ULL(13)) {
+ intlv_bit = FIELD_GET(BIT_ULL(13), ctx->ret_addr);
+
+ hashed_bit = intlv_bit;
+ hashed_bit ^= FIELD_GET(BIT_ULL(19), ctx->ret_addr) & hash_ctl_64k;
+ hashed_bit ^= FIELD_GET(BIT_ULL(24), ctx->ret_addr) & hash_ctl_2M;
+ hashed_bit ^= FIELD_GET(BIT_ULL(33), ctx->ret_addr) & hash_ctl_1G;
+ hashed_bit ^= FIELD_GET(BIT_ULL(43), ctx->ret_addr) & hash_ctl_1T;
+
+ if (hashed_bit != intlv_bit)
+ ctx->ret_addr ^= BIT_ULL(13);
+ }
+
+ if (rehash_vector & BIT_ULL(14)) {
+ intlv_bit = FIELD_GET(BIT_ULL(14), ctx->ret_addr);
+
+ hashed_bit = intlv_bit;
+ hashed_bit ^= FIELD_GET(BIT_ULL(20), ctx->ret_addr) & hash_ctl_64k;
+ hashed_bit ^= FIELD_GET(BIT_ULL(25), ctx->ret_addr) & hash_ctl_2M;
+ hashed_bit ^= FIELD_GET(BIT_ULL(34), ctx->ret_addr) & hash_ctl_1G;
+ hashed_bit ^= FIELD_GET(BIT_ULL(44), ctx->ret_addr) & hash_ctl_1T;
+
+ if (hashed_bit != intlv_bit)
+ ctx->ret_addr ^= BIT_ULL(14);
+ }
+
+ return 0;
+}
+
+int dehash_address(struct addr_ctx *ctx)
+{
+ switch (ctx->map.intlv_mode) {
+ /* No hashing cases. */
+ case NONE:
+ case NOHASH_2CHAN:
+ case NOHASH_4CHAN:
+ case NOHASH_8CHAN:
+ case NOHASH_16CHAN:
+ case NOHASH_32CHAN:
+ /* Hashing bits handled earlier during CS ID calculation. */
+ case DF4_NPS4_3CHAN_HASH:
+ case DF4_NPS2_5CHAN_HASH:
+ case DF4_NPS2_6CHAN_HASH:
+ case DF4_NPS1_10CHAN_HASH:
+ case DF4_NPS1_12CHAN_HASH:
+ case DF4p5_NPS2_6CHAN_1K_HASH:
+ case DF4p5_NPS2_6CHAN_2K_HASH:
+ case DF4p5_NPS1_10CHAN_1K_HASH:
+ case DF4p5_NPS1_10CHAN_2K_HASH:
+ case DF4p5_NPS1_12CHAN_1K_HASH:
+ case DF4p5_NPS1_12CHAN_2K_HASH:
+ case DF4p5_NPS0_24CHAN_1K_HASH:
+ case DF4p5_NPS0_24CHAN_2K_HASH:
+ /* No hash physical address bits, so nothing to do. */
+ case DF4p5_NPS4_3CHAN_1K_HASH:
+ case DF4p5_NPS4_3CHAN_2K_HASH:
+ case DF4p5_NPS2_5CHAN_1K_HASH:
+ case DF4p5_NPS2_5CHAN_2K_HASH:
+ return 0;
+
+ case DF2_2CHAN_HASH:
+ return df2_dehash_addr(ctx);
+
+ case DF3_COD4_2CHAN_HASH:
+ case DF3_COD2_4CHAN_HASH:
+ case DF3_COD1_8CHAN_HASH:
+ return df3_dehash_addr(ctx);
+
+ case DF3_6CHAN:
+ return df3_6chan_dehash_addr(ctx);
+
+ case DF4_NPS4_2CHAN_HASH:
+ case DF4_NPS2_4CHAN_HASH:
+ case DF4_NPS1_8CHAN_HASH:
+ return df4_dehash_addr(ctx);
+
+ case DF4p5_NPS4_2CHAN_1K_HASH:
+ case DF4p5_NPS4_2CHAN_2K_HASH:
+ case DF4p5_NPS2_4CHAN_2K_HASH:
+ case DF4p5_NPS2_4CHAN_1K_HASH:
+ case DF4p5_NPS1_8CHAN_1K_HASH:
+ case DF4p5_NPS1_8CHAN_2K_HASH:
+ case DF4p5_NPS1_16CHAN_1K_HASH:
+ case DF4p5_NPS1_16CHAN_2K_HASH:
+ return df4p5_dehash_addr(ctx);
+
+ default:
+ ATL_BAD_INTLV_MODE(ctx->map.intlv_mode);
+ return -EINVAL;
+ }
+}
diff --git a/drivers/platform/x86/amd/atl/denormalize.c b/drivers/platform/x86/amd/atl/denormalize.c
new file mode 100644
index 000000000000..fe1480c8e0d8
--- /dev/null
+++ b/drivers/platform/x86/amd/atl/denormalize.c
@@ -0,0 +1,644 @@
+// SPDX-License-Identifier: GPL-2.0-or-later
+/*
+ * AMD Address Translation Library
+ *
+ * denormalize.c : Functions to account for interleaving bits
+ *
+ * Copyright (c) 2023, Advanced Micro Devices, Inc.
+ * All Rights Reserved.
+ *
+ * Author: Yazen Ghannam <Yazen.Ghannam@amd.com>
+ */
+
+#include "internal.h"
+
+/*
+ * Returns the Destination Fabric ID. This is the first (lowest)
+ * CS Fabric ID used within a DRAM Address map.
+ */
+static u16 get_dst_fabric_id(struct addr_ctx *ctx)
+{
+ switch (df_cfg.rev) {
+ case DF2:
+ return FIELD_GET(DF2_DST_FABRIC_ID, ctx->map.limit);
+ case DF3:
+ return FIELD_GET(DF3_DST_FABRIC_ID, ctx->map.limit);
+ case DF3p5:
+ return FIELD_GET(DF3p5_DST_FABRIC_ID, ctx->map.limit);
+ case DF4:
+ return FIELD_GET(DF4_DST_FABRIC_ID, ctx->map.ctl);
+ case DF4p5:
+ return FIELD_GET(DF4p5_DST_FABRIC_ID, ctx->map.ctl);
+ default:
+ ATL_BAD_DF_REV;
+ return 0;
+ }
+}
+
+/*
+ * Make a contiguous gap in address for N bits starting at bit P.
+ *
+ * Example:
+ * address bits: [20:0]
+ * # of interleave bits (n): 3
+ * starting interleave bit (p): 8
+ *
+ * expanded address bits: [20+n : n+p][n+p-1 : p][p-1 : 0]
+ * [23 : 11][10 : 8][7 : 0]
+ */
+static u64 make_space_for_cs_id_at_intlv_bit(struct addr_ctx *ctx)
+{
+ return expand_bits(ctx->map.intlv_bit_pos,
+ ctx->map.total_intlv_bits,
+ ctx->ret_addr);
+}
+
+/*
+ * Make two gaps in address for N bits.
+ * First gap is a single bit at bit P.
+ * Second gap is the remaining N-1 bits at bit 12.
+ *
+ * Example:
+ * address bits: [20:0]
+ * # of interleave bits (n): 3
+ * starting interleave bit (p): 8
+ *
+ * First gap
+ * expanded address bits: [20+1 : p+1][p][p-1 : 0]
+ * [21 : 9][8][7 : 0]
+ *
+ * Second gap uses result from first.
+ * r = n - 1; remaining interleave bits
+ * expanded address bits: [21+r : 12+r][12+r-1: 12][11 : 0]
+ * [23 : 14][13 : 12][11 : 0]
+ */
+static u64 make_space_for_cs_id_split_2_1(struct addr_ctx *ctx)
+{
+ /* Make a single space at the interleave bit. */
+ u64 denorm_addr = expand_bits(ctx->map.intlv_bit_pos, 1, ctx->ret_addr);
+
+ /* Done if there's only a single interleave bit. */
+ if (ctx->map.total_intlv_bits <= 1)
+ return denorm_addr;
+
+ /* Make spaces for the remaining interleave bits starting at bit 12. */
+ return expand_bits(12, ctx->map.total_intlv_bits - 1, denorm_addr);
+}
+
+/*
+ * Take the current calculated address and shift enough bits in the middle
+ * to make a gap where the interleave bits will be inserted.
+ */
+static u64 make_space_for_cs_id(struct addr_ctx *ctx)
+{
+ switch (ctx->map.intlv_mode) {
+ case NOHASH_2CHAN:
+ case NOHASH_4CHAN:
+ case NOHASH_8CHAN:
+ case NOHASH_16CHAN:
+ case NOHASH_32CHAN:
+ case DF2_2CHAN_HASH:
+ return make_space_for_cs_id_at_intlv_bit(ctx);
+
+ case DF3_COD4_2CHAN_HASH:
+ case DF3_COD2_4CHAN_HASH:
+ case DF3_COD1_8CHAN_HASH:
+ case DF4_NPS4_2CHAN_HASH:
+ case DF4_NPS2_4CHAN_HASH:
+ case DF4_NPS1_8CHAN_HASH:
+ case DF4p5_NPS4_2CHAN_1K_HASH:
+ case DF4p5_NPS4_2CHAN_2K_HASH:
+ case DF4p5_NPS2_4CHAN_2K_HASH:
+ case DF4p5_NPS1_8CHAN_2K_HASH:
+ case DF4p5_NPS1_16CHAN_2K_HASH:
+ return make_space_for_cs_id_split_2_1(ctx);
+
+ case DF4p5_NPS2_4CHAN_1K_HASH:
+ //TODO
+ case DF4p5_NPS1_8CHAN_1K_HASH:
+ //TODO
+ case DF4p5_NPS1_16CHAN_1K_HASH:
+ //TODO
+ default:
+ ATL_BAD_INTLV_MODE(ctx->map.intlv_mode);
+ return ~0ULL;
+ }
+}
+
+static u16 get_cs_id_df2(struct addr_ctx *ctx)
+{
+ u8 num_socket_intlv_bits = ilog2(ctx->map.num_intlv_sockets);
+ u8 num_die_intlv_bits = ilog2(ctx->map.num_intlv_dies);
+ u8 num_intlv_bits;
+ u16 cs_id, mask;
+
+ cs_id = ctx->cs_fabric_id - get_dst_fabric_id(ctx);
+
+ /* Channel interleave bits */
+ num_intlv_bits = order_base_2(ctx->map.num_intlv_chan);
+ mask = GENMASK(num_intlv_bits - 1, 0);
+ cs_id &= mask;
+
+ /* Die interleave bits */
+ if (num_die_intlv_bits) {
+ u16 die_bits;
+
+ mask = GENMASK(num_die_intlv_bits - 1, 0);
+ die_bits = ctx->cs_fabric_id & df_cfg.die_id_mask;
+ die_bits >>= df_cfg.die_id_shift;
+
+ cs_id |= (die_bits & mask) << num_intlv_bits;
+ num_intlv_bits += num_die_intlv_bits;
+ }
+
+ /* Socket interleave bits */
+ if (num_socket_intlv_bits) {
+ u16 socket_bits;
+
+ mask = GENMASK(num_socket_intlv_bits - 1, 0);
+ socket_bits = ctx->cs_fabric_id & df_cfg.socket_id_mask;
+ socket_bits >>= df_cfg.socket_id_shift;
+
+ cs_id |= (socket_bits & mask) << num_intlv_bits;
+ }
+
+ return cs_id;
+}
+
+static u16 get_cs_id_df4(struct addr_ctx *ctx)
+{
+ /*
+ * Start with the original component mask and the number of interleave
+ * bits for the channels in this map.
+ */
+ u8 num_intlv_bits = ilog2(ctx->map.num_intlv_chan);
+ u16 mask = df_cfg.component_id_mask;
+
+ u16 socket_bits;
+
+ /* Set the derived CS ID to the input CS Fabric ID. */
+ u16 cs_id = ctx->cs_fabric_id & mask;
+
+ /*
+ * Subtract the "base" Destination Fabric ID.
+ * This accounts for systems with disabled Coherent Stations.
+ */
+ cs_id -= get_dst_fabric_id(ctx) & mask;
+
+ /*
+ * Generate and use a new mask based on the number of bits
+ * needed for channel interleaving in this map.
+ */
+ mask = GENMASK(num_intlv_bits - 1, 0);
+ cs_id &= mask;
+
+ /* Done if socket interleaving is not enabled. */
+ if (ctx->map.num_intlv_sockets <= 1)
+ return cs_id;
+
+ /*
+ * Figure out how many bits are needed for the number of
+ * interleaved sockets. And shift the derived CS ID to account
+ * for these.
+ */
+ num_intlv_bits = ilog2(ctx->map.num_intlv_sockets);
+ cs_id <<= num_intlv_bits;
+
+ /* Generate a new mask for the socket interleaving bits. */
+ mask = GENMASK(num_intlv_bits - 1, 0);
+
+ /* Get the socket interleave bits from the original CS Fabric ID. */
+ socket_bits = (ctx->cs_fabric_id & df_cfg.socket_id_mask) >> df_cfg.socket_id_shift;
+
+ /* Apply the appropriate socket bits to the derived CS ID. */
+ cs_id |= socket_bits & mask;
+
+ return cs_id;
+}
+
+/*
+ * Derive the correct CS ID that represents the interleave bits
+ * used within the system physical address. This accounts for the
+ * interleave mode, number of interleaved channels/dies/sockets, and
+ * other system/mode-specific bit swizzling.
+ */
+static u16 calculate_cs_id(struct addr_ctx *ctx)
+{
+ switch (ctx->map.intlv_mode) {
+ case NOHASH_2CHAN:
+ case NOHASH_4CHAN:
+ case NOHASH_8CHAN:
+ case NOHASH_16CHAN:
+ case NOHASH_32CHAN:
+ case DF3_COD4_2CHAN_HASH:
+ case DF3_COD2_4CHAN_HASH:
+ case DF3_COD1_8CHAN_HASH:
+ case DF2_2CHAN_HASH:
+ return get_cs_id_df2(ctx);
+
+ case DF4_NPS4_2CHAN_HASH:
+ case DF4_NPS2_4CHAN_HASH:
+ case DF4_NPS1_8CHAN_HASH:
+ case DF4p5_NPS4_2CHAN_1K_HASH:
+ case DF4p5_NPS4_2CHAN_2K_HASH:
+ case DF4p5_NPS2_4CHAN_2K_HASH:
+ case DF4p5_NPS1_8CHAN_2K_HASH:
+ case DF4p5_NPS1_16CHAN_2K_HASH:
+ return get_cs_id_df4(ctx);
+
+ /* CS ID is simply the CS Fabric ID adjusted by the Destination Fabric ID. */
+ case DF4p5_NPS2_4CHAN_1K_HASH:
+ case DF4p5_NPS1_8CHAN_1K_HASH:
+ case DF4p5_NPS1_16CHAN_1K_HASH:
+ return ctx->cs_fabric_id - get_dst_fabric_id(ctx);
+
+ default:
+ ATL_BAD_INTLV_MODE(ctx->map.intlv_mode);
+ return ~0;
+ }
+}
+
+static u64 insert_cs_id_at_intlv_bit(struct addr_ctx *ctx, u64 denorm_addr, u16 cs_id)
+{
+ return denorm_addr | (cs_id << ctx->map.intlv_bit_pos);
+}
+
+static u64 insert_cs_id_split_2_1(struct addr_ctx *ctx, u64 denorm_addr, u16 cs_id)
+{
+ /* Insert cs_id[0] at the interleave bit. */
+ denorm_addr |= (cs_id & BIT(0)) << ctx->map.intlv_bit_pos;
+
+ /* Insert cs_id[2:1] at bit 12. */
+ denorm_addr |= (cs_id & GENMASK(2, 1)) << 11;
+
+ return denorm_addr;
+}
+
+static u64 insert_cs_id_split_2_2(struct addr_ctx *ctx, u64 denorm_addr, u16 cs_id)
+{
+ /* Insert cs_id[1:0] at bit 8. */
+ denorm_addr |= (cs_id & GENMASK(1, 0)) << 8;
+
+ /*
+ * Insert cs_id[n:2] at bit 12. 'n' could be 2 or 3.
+ * Grab both because bit 3 will be clear if unused.
+ */
+ denorm_addr |= (cs_id & GENMASK(3, 2)) << 10;
+
+ return denorm_addr;
+}
+
+static u64 insert_cs_id(struct addr_ctx *ctx, u64 denorm_addr, u16 cs_id)
+{
+ switch (ctx->map.intlv_mode) {
+ case NOHASH_2CHAN:
+ case NOHASH_4CHAN:
+ case NOHASH_8CHAN:
+ case NOHASH_16CHAN:
+ case NOHASH_32CHAN:
+ case DF2_2CHAN_HASH:
+ return insert_cs_id_at_intlv_bit(ctx, denorm_addr, cs_id);
+
+ case DF3_COD4_2CHAN_HASH:
+ case DF3_COD2_4CHAN_HASH:
+ case DF3_COD1_8CHAN_HASH:
+ case DF4_NPS4_2CHAN_HASH:
+ case DF4_NPS2_4CHAN_HASH:
+ case DF4_NPS1_8CHAN_HASH:
+ case DF4p5_NPS4_2CHAN_1K_HASH:
+ case DF4p5_NPS4_2CHAN_2K_HASH:
+ case DF4p5_NPS2_4CHAN_2K_HASH:
+ case DF4p5_NPS1_8CHAN_2K_HASH:
+ case DF4p5_NPS1_16CHAN_2K_HASH:
+ return insert_cs_id_split_2_1(ctx, denorm_addr, cs_id);
+
+ case DF4p5_NPS2_4CHAN_1K_HASH:
+ case DF4p5_NPS1_8CHAN_1K_HASH:
+ case DF4p5_NPS1_16CHAN_1K_HASH:
+ return insert_cs_id_split_2_2(ctx, denorm_addr, cs_id);
+
+ default:
+ ATL_BAD_INTLV_MODE(ctx->map.intlv_mode);
+ return ~0ULL;
+ }
+}
+
+static u16 get_logical_cs_fabric_id(struct addr_ctx *ctx)
+{
+ u16 cs_component_id, log_cs_fabric_id;
+
+ /* Start with the physical CS Fabric ID. */
+ u16 phys_cs_fabric_id = ctx->cs_fabric_id;
+
+ /* Skip logical ID lookup if remapping is disabled. */
+ if (!FIELD_GET(DF4_REMAP_EN, ctx->map.ctl) &&
+ ctx->map.intlv_mode != DF3_6CHAN)
+ return phys_cs_fabric_id;
+
+ /* Mask off the Node ID bits to get the "local" Component ID. */
+ cs_component_id = phys_cs_fabric_id & df_cfg.component_id_mask;
+
+ /*
+ * Search the list of logical Component IDs for the one that
+ * matches this physical Component ID.
+ */
+ for (log_cs_fabric_id = 0; log_cs_fabric_id < MAX_CS_CHANNELS; log_cs_fabric_id++) {
+ if (ctx->map.remap_array[log_cs_fabric_id] == cs_component_id)
+ break;
+ }
+
+ if (log_cs_fabric_id == MAX_CS_CHANNELS)
+ pr_warn("%s: CS remap entry not found.", __func__);
+
+ /* Get the Node ID bits from the physical and apply to the logical. */
+ return (phys_cs_fabric_id & df_cfg.node_id_mask) | log_cs_fabric_id;
+}
+
+static int denorm_addr_common(struct addr_ctx *ctx)
+{
+ u64 denorm_addr;
+ u16 cs_id;
+
+ /*
+ * Convert the original physical CS Fabric ID to a logical value.
+ * This is required for non-power-of-two and other interleaving modes.
+ */
+ ctx->cs_fabric_id = get_logical_cs_fabric_id(ctx);
+
+ denorm_addr = make_space_for_cs_id(ctx);
+ cs_id = calculate_cs_id(ctx);
+ ctx->ret_addr = insert_cs_id(ctx, denorm_addr, cs_id);
+ return 0;
+}
+
+static int denorm_addr_df3_6chan(struct addr_ctx *ctx)
+{
+ u16 cs_id = ctx->cs_fabric_id & df_cfg.component_id_mask;
+ u8 total_intlv_bits = ctx->map.total_intlv_bits;
+ u8 low_bit, intlv_bit = ctx->map.intlv_bit_pos;
+ u64 msb_intlv_bits, temp_addr_a, temp_addr_b;
+ u8 np2_bits = ctx->map.np2_bits;
+
+ if (ctx->map.intlv_mode != DF3_6CHAN)
+ return -EINVAL;
+
+ /*
+ * 'np2_bits' holds the number of bits needed to cover the
+ * amount of memory (rounded up) in this map using 64K chunks.
+ *
+ * Example:
+ * Total memory in map: 6GB
+ * Rounded up to next power-of-2: 8GB
+ * Number of 64K chunks: 0x20000
+ * np2_bits = log2(# of chunks): 17
+ *
+ * Get the two most-significant interleave bits from the
+ * input address based on the following:
+ *
+ * [15 + np2_bits - total_intlv_bits : 14 + np2_bits - total_intlv_bits]
+ */
+ low_bit = 14 + np2_bits - total_intlv_bits;
+ msb_intlv_bits = ctx->ret_addr >> low_bit;
+ msb_intlv_bits &= 0x3;
+
+ /*
+ * If MSB are 11b, then logical CS ID is 6 or 7.
+ * Need to adjust based on the mod3 result.
+ */
+ if (msb_intlv_bits == 3) {
+ u8 addr_mod, phys_addr_msb, msb_cs_id;
+
+ /* Get the remaining interleave bits from the input address. */
+ temp_addr_b = GENMASK_ULL(low_bit - 1, intlv_bit) & ctx->ret_addr;
+ temp_addr_b >>= intlv_bit;
+
+ /* Calculate the logical CS offset based on mod3. */
+ addr_mod = temp_addr_b % 3;
+
+ /* Get CS ID bits [2:1]. */
+ msb_cs_id = (cs_id >> 1) & 0x3;
+
+ /* Get the bit that starts the physical address bits. */
+ phys_addr_msb = (intlv_bit + np2_bits + 1);
+ phys_addr_msb &= BIT(0);
+ phys_addr_msb++;
+ phys_addr_msb *= 3 - addr_mod + msb_cs_id;
+ phys_addr_msb %= 3;
+
+ /* Move the physical address MSB to the correct place. */
+ temp_addr_b |= phys_addr_msb << (low_bit - total_intlv_bits - intlv_bit);
+
+ /* Generate a new CS ID as follows: cs_id = [1, 1, cs_id[0]] */
+ cs_id &= BIT(0);
+ cs_id |= GENMASK(2, 1);
+ } else {
+ temp_addr_b = GENMASK_ULL(63, intlv_bit) & ctx->ret_addr;
+ temp_addr_b >>= intlv_bit;
+ }
+
+ temp_addr_a = GENMASK_ULL(intlv_bit - 1, 0) & ctx->ret_addr;
+ temp_addr_b <<= intlv_bit + total_intlv_bits;
+
+ ctx->ret_addr = temp_addr_a | temp_addr_b;
+ ctx->ret_addr |= cs_id << intlv_bit;
+ return 0;
+}
+
+static int denorm_addr_df4_np2(struct addr_ctx *ctx)
+{
+ bool hash_ctl_64k, hash_ctl_2M, hash_ctl_1G;
+ u16 group, group_offset, log_cs_offset;
+ unsigned int mod_value, shift_value;
+ u16 mask = df_cfg.component_id_mask;
+ u64 temp_addr_a, temp_addr_b;
+ u8 hash_pa8, hashed_bit;
+
+ switch (ctx->map.intlv_mode) {
+ case DF4_NPS4_3CHAN_HASH:
+ mod_value = 3;
+ shift_value = 13;
+ break;
+ case DF4_NPS2_6CHAN_HASH:
+ mod_value = 3;
+ shift_value = 12;
+ break;
+ case DF4_NPS1_12CHAN_HASH:
+ mod_value = 3;
+ shift_value = 11;
+ break;
+ case DF4_NPS2_5CHAN_HASH:
+ mod_value = 5;
+ shift_value = 13;
+ break;
+ case DF4_NPS1_10CHAN_HASH:
+ mod_value = 5;
+ shift_value = 12;
+ break;
+ default:
+ ATL_BAD_INTLV_MODE(ctx->map.intlv_mode);
+ return -EINVAL;
+ };
+
+ if (ctx->map.num_intlv_sockets == 1) {
+ hash_pa8 = BIT_ULL(shift_value) & ctx->ret_addr;
+ temp_addr_a = remove_bits(shift_value, shift_value, ctx->ret_addr);
+ } else {
+ hash_pa8 = (ctx->cs_fabric_id & df_cfg.socket_id_mask);
+ hash_pa8 >>= df_cfg.socket_id_shift;
+ temp_addr_a = ctx->ret_addr;
+ }
+
+ /* Make a gap for the real bit [8]. */
+ temp_addr_a = expand_bits(8, 1, temp_addr_a);
+
+ /* Make an additional gap for bits [13:12], as appropriate.*/
+ if (ctx->map.intlv_mode == DF4_NPS2_6CHAN_HASH ||
+ ctx->map.intlv_mode == DF4_NPS1_10CHAN_HASH) {
+ temp_addr_a = expand_bits(13, 1, temp_addr_a);
+ } else if (ctx->map.intlv_mode == DF4_NPS1_12CHAN_HASH) {
+ temp_addr_a = expand_bits(12, 2, temp_addr_a);
+ }
+
+ /* Keep bits [13:0]. */
+ temp_addr_a &= GENMASK_ULL(13, 0);
+
+ /* Get the appropriate high bits. */
+ shift_value += 1 - ilog2(ctx->map.num_intlv_sockets);
+ temp_addr_b = GENMASK_ULL(63, shift_value) & ctx->ret_addr;
+ temp_addr_b >>= shift_value;
+ temp_addr_b *= mod_value;
+
+ /*
+ * Coherent Stations are divided into groups.
+ *
+ * Multiples of 3 (mod3) are divided into quadrants.
+ * e.g. NP4_3CHAN -> [0, 1, 2] [6, 7, 8]
+ * [3, 4, 5] [9, 10, 11]
+ *
+ * Multiples of 5 (mod5) are divided into sides.
+ * e.g. NP2_5CHAN -> [0, 1, 2, 3, 4] [5, 6, 7, 8, 9]
+ */
+
+ /*
+ * Calculate the logical offset for the CS within its DRAM Address map.
+ * e.g. if map includes [5, 6, 7, 8, 9] and target instance is '8', then
+ * log_cs_offset = 8 - 5 = 3
+ */
+ log_cs_offset = (ctx->cs_fabric_id & mask) - (get_dst_fabric_id(ctx) & mask);
+
+ /*
+ * Figure out the group number.
+ *
+ * Following above example,
+ * log_cs_offset = 3
+ * mod_value = 5
+ * group = 3 / 5 = 0
+ */
+ group = log_cs_offset / mod_value;
+
+ /*
+ * Figure out the offset within the group.
+ *
+ * Following above example,
+ * log_cs_offset = 3
+ * mod_value = 5
+ * group_offset = 3 % 5 = 3
+ */
+ group_offset = log_cs_offset % mod_value;
+
+ /* Adjust group_offset if the hashed bit [8] is set. */
+ if (hash_pa8) {
+ if (!group_offset)
+ group_offset = mod_value - 1;
+ else
+ group_offset--;
+ }
+
+ /* Add in the group offset to the high bits. */
+ temp_addr_b += group_offset;
+
+ /* Shift the high bits to the proper starting position. */
+ temp_addr_b <<= 14;
+
+ /* Combine the high and low bits together. */
+ ctx->ret_addr = temp_addr_a | temp_addr_b;
+
+ /* Account for hashing here instead of in dehash_address(). */
+ hash_ctl_64k = FIELD_GET(DF4_HASH_CTL_64K, ctx->map.ctl);
+ hash_ctl_2M = FIELD_GET(DF4_HASH_CTL_2M, ctx->map.ctl);
+ hash_ctl_1G = FIELD_GET(DF4_HASH_CTL_1G, ctx->map.ctl);
+
+ hashed_bit = !!hash_pa8;
+ hashed_bit ^= FIELD_GET(BIT_ULL(14), ctx->ret_addr);
+ hashed_bit ^= FIELD_GET(BIT_ULL(16), ctx->ret_addr) & hash_ctl_64k;
+ hashed_bit ^= FIELD_GET(BIT_ULL(21), ctx->ret_addr) & hash_ctl_2M;
+ hashed_bit ^= FIELD_GET(BIT_ULL(30), ctx->ret_addr) & hash_ctl_1G;
+
+ ctx->ret_addr |= hashed_bit << 8;
+
+ /* Done for 3 and 5 channel. */
+ if (ctx->map.intlv_mode == DF4_NPS4_3CHAN_HASH ||
+ ctx->map.intlv_mode == DF4_NPS2_5CHAN_HASH)
+ return 0;
+
+ /* Select the proper 'group' bit to use for Bit 13. */
+ if (ctx->map.intlv_mode == DF4_NPS1_12CHAN_HASH)
+ hashed_bit = !!(group & BIT(1));
+ else
+ hashed_bit = group & BIT(0);
+
+ hashed_bit ^= FIELD_GET(BIT_ULL(18), ctx->ret_addr) & hash_ctl_64k;
+ hashed_bit ^= FIELD_GET(BIT_ULL(23), ctx->ret_addr) & hash_ctl_2M;
+ hashed_bit ^= FIELD_GET(BIT_ULL(32), ctx->ret_addr) & hash_ctl_1G;
+
+ ctx->ret_addr |= hashed_bit << 13;
+
+ /* Done for 6 and 10 channel. */
+ if (ctx->map.intlv_mode != DF4_NPS1_12CHAN_HASH)
+ return 0;
+
+ hashed_bit = group & BIT(0);
+ hashed_bit ^= FIELD_GET(BIT_ULL(17), ctx->ret_addr) & hash_ctl_64k;
+ hashed_bit ^= FIELD_GET(BIT_ULL(22), ctx->ret_addr) & hash_ctl_2M;
+ hashed_bit ^= FIELD_GET(BIT_ULL(31), ctx->ret_addr) & hash_ctl_1G;
+
+ ctx->ret_addr |= hashed_bit << 12;
+ return 0;
+}
+
+static int denorm_addr_df4p5_np2(struct addr_ctx *ctx)
+{
+ //TODO
+ return -EINVAL;
+}
+
+int denormalize_address(struct addr_ctx *ctx)
+{
+ switch (ctx->map.intlv_mode) {
+ case NONE:
+ return 0;
+ case DF4_NPS4_3CHAN_HASH:
+ case DF4_NPS2_6CHAN_HASH:
+ case DF4_NPS1_12CHAN_HASH:
+ case DF4_NPS2_5CHAN_HASH:
+ case DF4_NPS1_10CHAN_HASH:
+ return denorm_addr_df4_np2(ctx);
+ case DF4p5_NPS0_24CHAN_1K_HASH:
+ case DF4p5_NPS4_3CHAN_1K_HASH:
+ case DF4p5_NPS2_6CHAN_1K_HASH:
+ case DF4p5_NPS1_12CHAN_1K_HASH:
+ case DF4p5_NPS2_5CHAN_1K_HASH:
+ case DF4p5_NPS1_10CHAN_1K_HASH:
+ case DF4p5_NPS4_3CHAN_2K_HASH:
+ case DF4p5_NPS2_6CHAN_2K_HASH:
+ case DF4p5_NPS1_12CHAN_2K_HASH:
+ case DF4p5_NPS0_24CHAN_2K_HASH:
+ case DF4p5_NPS2_5CHAN_2K_HASH:
+ case DF4p5_NPS1_10CHAN_2K_HASH:
+ return denorm_addr_df4p5_np2(ctx);
+ case DF3_6CHAN:
+ return denorm_addr_df3_6chan(ctx);
+ default:
+ return denorm_addr_common(ctx);
+ }
+}
diff --git a/drivers/platform/x86/amd/atl/internal.h b/drivers/platform/x86/amd/atl/internal.h
new file mode 100644
index 000000000000..f3888c8fd02d
--- /dev/null
+++ b/drivers/platform/x86/amd/atl/internal.h
@@ -0,0 +1,307 @@
+/* SPDX-License-Identifier: GPL-2.0 */
+/*
+ * AMD Address Translation Library
+ *
+ * internal.h : Helper functions and common defines
+ *
+ * Copyright (c) 2023, Advanced Micro Devices, Inc.
+ * All Rights Reserved.
+ *
+ * Author: Yazen Ghannam <Yazen.Ghannam@amd.com>
+ */
+
+#ifndef __AMD_ATL_INTERNAL_H__
+#define __AMD_ATL_INTERNAL_H__
+
+#include <asm/amd_nb.h>
+
+#include <linux/amd-atl.h>
+#include <linux/bitfield.h>
+#include <linux/bitops.h>
+
+#include "reg_fields.h"
+
+/* Maximum possible number of Coherent Stations within a single Data Fabric. */
+#define MAX_CS_CHANNELS 32
+
+/* PCI IDs for Genoa DF Function 0. */
+#define DF_FUNC0_ID_GENOA 0x14AD1022
+
+/* Shift needed for adjusting register values to true values. */
+#define DF_DRAM_BASE_LIMIT_LSB 28
+
+/*
+ * Glossary of acronyms used in address translation for Zen-based systems
+ *
+ * CCM = Cache Coherent Moderator
+ * COD = Cluster-on-Die
+ * CS = Coherent Station
+ * DF = Data Fabric
+ */
+
+enum df_revisions {
+ UNKNOWN,
+ DF2,
+ DF3,
+ DF3p5,
+ DF4,
+ DF4p5,
+};
+
+/* These are mapped 1:1 to the hardware values. Special cases are set at > 0x20. */
+enum intlv_modes {
+ NONE = 0x00,
+ NOHASH_2CHAN = 0x01,
+ NOHASH_4CHAN = 0x03,
+ NOHASH_8CHAN = 0x05,
+ DF3_6CHAN = 0x06,
+ NOHASH_16CHAN = 0x07,
+ NOHASH_32CHAN = 0x08,
+ DF3_COD4_2CHAN_HASH = 0x0C,
+ DF3_COD2_4CHAN_HASH = 0x0D,
+ DF3_COD1_8CHAN_HASH = 0x0E,
+ DF4_NPS4_2CHAN_HASH = 0x10,
+ DF4_NPS2_4CHAN_HASH = 0x11,
+ DF4_NPS1_8CHAN_HASH = 0x12,
+ DF4_NPS4_3CHAN_HASH = 0x13,
+ DF4_NPS2_6CHAN_HASH = 0x14,
+ DF4_NPS1_12CHAN_HASH = 0x15,
+ DF4_NPS2_5CHAN_HASH = 0x16,
+ DF4_NPS1_10CHAN_HASH = 0x17,
+ DF2_2CHAN_HASH = 0x21,
+ /* DF4.5 modes are all IntLvNumChan + 0x20 */
+ DF4p5_NPS1_16CHAN_1K_HASH = 0x2C,
+ DF4p5_NPS0_24CHAN_1K_HASH = 0x2E,
+ DF4p5_NPS4_2CHAN_1K_HASH = 0x30,
+ DF4p5_NPS2_4CHAN_1K_HASH = 0x31,
+ DF4p5_NPS1_8CHAN_1K_HASH = 0x32,
+ DF4p5_NPS4_3CHAN_1K_HASH = 0x33,
+ DF4p5_NPS2_6CHAN_1K_HASH = 0x34,
+ DF4p5_NPS1_12CHAN_1K_HASH = 0x35,
+ DF4p5_NPS2_5CHAN_1K_HASH = 0x36,
+ DF4p5_NPS1_10CHAN_1K_HASH = 0x37,
+ DF4p5_NPS4_2CHAN_2K_HASH = 0x40,
+ DF4p5_NPS2_4CHAN_2K_HASH = 0x41,
+ DF4p5_NPS1_8CHAN_2K_HASH = 0x42,
+ DF4p5_NPS1_16CHAN_2K_HASH = 0x43,
+ DF4p5_NPS4_3CHAN_2K_HASH = 0x44,
+ DF4p5_NPS2_6CHAN_2K_HASH = 0x45,
+ DF4p5_NPS1_12CHAN_2K_HASH = 0x46,
+ DF4p5_NPS0_24CHAN_2K_HASH = 0x47,
+ DF4p5_NPS2_5CHAN_2K_HASH = 0x48,
+ DF4p5_NPS1_10CHAN_2K_HASH = 0x49,
+};
+
+struct df_flags {
+ __u8 legacy_ficaa : 1,
+ genoa_quirk : 1,
+ __reserved_0 : 6;
+};
+
+struct df_config {
+ enum df_revisions rev;
+
+ /*
+ * These masks operate on the 16-bit CS IDs,
+ * e.g. Instance, Fabric, Destination, etc.
+ */
+ u16 component_id_mask;
+ u16 die_id_mask;
+ u16 node_id_mask;
+ u16 socket_id_mask;
+
+ /*
+ * Least-significant bit of Node ID portion of the
+ * system-wide CS Fabric ID.
+ */
+ u8 node_id_shift;
+
+ /*
+ * Least-significant bit of Die portion of the Node ID.
+ * Adjusted to include the Node ID shift in order to apply
+ * to the CS Fabric ID.
+ */
+ u8 die_id_shift;
+
+ /*
+ * Least-significant bit of Socket portion of the Node ID.
+ * Adjusted to include the Node ID shift in order to apply
+ * to the CS Fabric ID.
+ */
+ u8 socket_id_shift;
+
+ /* Number of DRAM Address maps visible in a CS. */
+ u8 num_cs_maps;
+
+ /* Global flags to handle special cases. */
+ struct df_flags flags;
+};
+
+extern struct df_config df_cfg;
+
+struct dram_addr_map {
+ /*
+ * Each DRAM Address Map can operate independently
+ * in different interleaving modes.
+ */
+ enum intlv_modes intlv_mode;
+
+ /* System-wide number for this address map. */
+ u8 num;
+
+ /* Raw register values */
+ u32 base;
+ u32 limit;
+ u32 ctl;
+ u32 intlv;
+
+ /*
+ * Logical to Physical CS Remapping array
+ *
+ * Index: Logical CS Instance ID
+ * Value: Physical CS Instance ID
+ *
+ * phys_cs_inst_id = remap_array[log_cs_inst_id]
+ */
+ u8 remap_array[MAX_CS_CHANNELS];
+
+ /*
+ * Number of bits covering DRAM Address map 0
+ * when interleaving is non-power-of-2.
+ *
+ * Used only for DF3_6CHAN.
+ */
+ u8 np2_bits;
+
+ /* Position of the 'interleave bit'. */
+ u8 intlv_bit_pos;
+ /* Number of channels interleaved in this map. */
+ u8 num_intlv_chan;
+ /* Number of dies interleaved in this map. */
+ u8 num_intlv_dies;
+ /* Number of sockets interleaved in this map. */
+ u8 num_intlv_sockets;
+ /*
+ * Total number of channels interleaved accounting
+ * for die and socket interleaving.
+ */
+ u8 total_intlv_chan;
+ /* Total bits needed to cover 'total_intlv_chan'. */
+ u8 total_intlv_bits;
+};
+
+struct addr_ctx {
+ u64 ret_addr;
+
+ struct dram_addr_map map;
+
+ /* AMD Node ID calculated from Socket and Die IDs. */
+ u8 node_id;
+
+ /*
+ * CS Instance ID
+ * Local ID used within a 'node'.
+ */
+ u16 inst_id;
+
+ /*
+ * CS Fabric ID
+ * System-wide ID that includes 'node' bits.
+ */
+ u16 cs_fabric_id;
+};
+
+int df_indirect_read_instance(u16 node, u8 func, u16 reg, u8 instance_id, u32 *lo);
+int df_indirect_read_broadcast(u16 node, u8 func, u16 reg, u32 *lo);
+
+int get_df_system_info(void);
+int determine_node_id(struct addr_ctx *ctx, u8 socket_num, u8 die_num);
+
+int get_address_map(struct addr_ctx *ctx);
+
+int denormalize_address(struct addr_ctx *ctx);
+int dehash_address(struct addr_ctx *ctx);
+
+int norm_to_sys_addr(u8 socket_id, u8 die_id, u8 cs_inst_id, u64 *addr);
+
+/*
+ * Helper to use test_bit() without needing to do
+ * a cast to "unsigned long *" everywhere.
+ *
+ * Use this for dynamic checks, i.e. when FIELD_GET()
+ * won't work.
+ */
+static inline bool atl_get_bit(u8 bit_num, u64 data)
+{
+ return test_bit(bit_num, (unsigned long *)&data);
+}
+
+/*
+ * Make a gap in 'data' that is 'num_bits' long starting at 'bit_num.
+ * e.g. data = 11111111'b
+ * bit_num = 3
+ * num_bits = 2
+ * result = 1111100111'b
+ */
+static inline u64 expand_bits(u8 bit_num, u8 num_bits, u64 data)
+{
+ u64 temp1, temp2;
+
+ /*
+ * Return the orginal data if the "space" needed is '0'.
+ * This helps avoid the need to check for '0' at each
+ * caller.
+ */
+ if (!num_bits)
+ return data;
+
+ if (!bit_num)
+ return data << num_bits;
+
+ temp1 = data & GENMASK_ULL(bit_num - 1, 0);
+
+ temp2 = data & GENMASK_ULL(63, bit_num);
+ temp2 <<= num_bits;
+
+ return temp1 | temp2;
+}
+
+/*
+ * Remove bits in 'data' between low_bit and high_bit inclusive.
+ * e.g. data = XXXYYZZZ'b
+ * low_bit = 3
+ * high_bit = 4
+ * result = XXXZZZ'b
+ */
+static inline u64 remove_bits(u8 low_bit, u8 high_bit, u64 data)
+{
+ u64 temp1, temp2;
+
+ if (high_bit >= 64 || low_bit >= 64 || low_bit > high_bit) {
+ pr_warn("%s: Invalid inputs: low_bit=%u high_bit=%u",
+ __func__, low_bit, high_bit);
+ return 0;
+ }
+
+ if (!low_bit)
+ return data >> (high_bit++);
+
+ temp1 = GENMASK_ULL(low_bit - 1, 0) & data;
+ temp2 = GENMASK_ULL(63, high_bit + 1) & data;
+ temp2 >>= high_bit - low_bit + 1;
+
+ return temp1 | temp2;
+}
+
+#define ATL_BAD_DF_REV \
+ ({ pr_warn("%s: Unrecognized DF rev: %u", \
+ __func__, df_cfg.rev); })
+
+#define ATL_BAD_INTLV_MODE(mode) \
+ ({pr_warn("%s: Unrecognized interleave mode: %u", \
+ __func__, mode); })
+
+#define atl_debug(fmt, arg...) \
+ pr_debug("AMD ATL: %s: " fmt, __func__, ##arg)
+
+#endif /* __AMD_ATL_INTERNAL_H__ */
diff --git a/drivers/platform/x86/amd/atl/map.c b/drivers/platform/x86/amd/atl/map.c
new file mode 100644
index 000000000000..05141da27028
--- /dev/null
+++ b/drivers/platform/x86/amd/atl/map.c
@@ -0,0 +1,659 @@
+// SPDX-License-Identifier: GPL-2.0-or-later
+/*
+ * AMD Address Translation Library
+ *
+ * map.c : Functions to read and decode DRAM address maps
+ *
+ * Copyright (c) 2023, Advanced Micro Devices, Inc.
+ * All Rights Reserved.
+ *
+ * Author: Yazen Ghannam <Yazen.Ghannam@amd.com>
+ */
+
+#include "internal.h"
+
+static int df2_get_intlv_mode(struct addr_ctx *ctx)
+{
+ ctx->map.intlv_mode = FIELD_GET(DF2_INTLV_NUM_CHAN, ctx->map.base);
+
+ if (ctx->map.intlv_mode == 8)
+ ctx->map.intlv_mode = DF2_2CHAN_HASH;
+
+ if (ctx->map.intlv_mode != NONE &&
+ ctx->map.intlv_mode != NOHASH_2CHAN &&
+ ctx->map.intlv_mode != DF2_2CHAN_HASH)
+ return -EINVAL;
+
+ return 0;
+}
+
+static int df3_get_intlv_mode(struct addr_ctx *ctx)
+{
+ ctx->map.intlv_mode = FIELD_GET(DF3_INTLV_NUM_CHAN, ctx->map.base);
+ return 0;
+}
+
+static int df3p5_get_intlv_mode(struct addr_ctx *ctx)
+{
+ ctx->map.intlv_mode = FIELD_GET(DF3p5_INTLV_NUM_CHAN, ctx->map.base);
+
+ if (ctx->map.intlv_mode == DF3_6CHAN)
+ return -EINVAL;
+
+ return 0;
+}
+
+static int df4_get_intlv_mode(struct addr_ctx *ctx)
+{
+ ctx->map.intlv_mode = FIELD_GET(DF4_INTLV_NUM_CHAN, ctx->map.intlv);
+
+ if (ctx->map.intlv_mode == DF3_COD4_2CHAN_HASH ||
+ ctx->map.intlv_mode == DF3_COD2_4CHAN_HASH ||
+ ctx->map.intlv_mode == DF3_COD1_8CHAN_HASH ||
+ ctx->map.intlv_mode == DF3_6CHAN)
+ return -EINVAL;
+
+ return 0;
+}
+
+static int df4p5_get_intlv_mode(struct addr_ctx *ctx)
+{
+ ctx->map.intlv_mode = FIELD_GET(DF4p5_INTLV_NUM_CHAN, ctx->map.intlv);
+
+ if (ctx->map.intlv_mode <= NOHASH_32CHAN)
+ return 0;
+
+ /*
+ * Modes matching the ranges above are returned as-is.
+ *
+ * All other modes are "fixed up" by adding 20h to make a unique value.
+ */
+ ctx->map.intlv_mode += 0x20;
+
+ return 0;
+}
+
+/*
+ * Some, but not all, cases have asserts.
+ * So use return values to indicate failure where needed.
+ */
+static int get_intlv_mode(struct addr_ctx *ctx)
+{
+ switch (df_cfg.rev) {
+ case DF2:
+ return df2_get_intlv_mode(ctx);
+ case DF3:
+ return df3_get_intlv_mode(ctx);
+ case DF3p5:
+ return df3p5_get_intlv_mode(ctx);
+ case DF4:
+ return df4_get_intlv_mode(ctx);
+ case DF4p5:
+ return df4p5_get_intlv_mode(ctx);
+ default:
+ ATL_BAD_DF_REV;
+ return -EINVAL;
+ }
+}
+
+static u64 get_hi_addr_offset(u32 reg_dram_offset)
+{
+ u8 shift = DF_DRAM_BASE_LIMIT_LSB;
+ u64 hi_addr_offset = 0;
+
+ switch (df_cfg.rev) {
+ case DF2:
+ hi_addr_offset = FIELD_GET(DF2_HI_ADDR_OFFSET, reg_dram_offset);
+ break;
+ case DF3:
+ case DF3p5:
+ hi_addr_offset = FIELD_GET(DF3_HI_ADDR_OFFSET, reg_dram_offset);
+ break;
+ case DF4:
+ case DF4p5:
+ hi_addr_offset = FIELD_GET(DF4_HI_ADDR_OFFSET, reg_dram_offset);
+ break;
+ default:
+ ATL_BAD_DF_REV;
+ }
+
+ return hi_addr_offset << shift;
+}
+
+static int get_dram_offset(struct addr_ctx *ctx, bool *enabled, u64 *norm_offset)
+{
+ u32 reg_dram_offset;
+ u8 map_num;
+
+ /* Should not be called for map 0. */
+ if (!ctx->map.num)
+ return -EINVAL;
+
+ /*
+ * DramOffset registers don't exist for map 0, so the base register
+ * actually refers to map 1.
+ * Adjust the map_num for the register offsets.
+ */
+ map_num = ctx->map.num - 1;
+
+ if (df_cfg.rev >= DF4) {
+ /* Read D18F7x140 (DramOffset) */
+ if (df_indirect_read_instance(ctx->node_id, 7, 0x140 + (4 * map_num),
+ ctx->inst_id, ®_dram_offset))
+ return -EINVAL;
+
+ } else {
+ /* Read D18F0x1B4 (DramOffset) */
+ if (df_indirect_read_instance(ctx->node_id, 0, 0x1B4 + (4 * map_num),
+ ctx->inst_id, ®_dram_offset))
+ return -EINVAL;
+ }
+
+ if (!FIELD_GET(DF_HI_ADDR_OFFSET_EN, reg_dram_offset))
+ return 0;
+
+ *enabled = true;
+ *norm_offset = get_hi_addr_offset(reg_dram_offset);
+
+ return 0;
+}
+
+static int df3_6ch_get_dram_addr_map(struct addr_ctx *ctx)
+{
+ u16 dst_fabric_id = FIELD_GET(DF3_DST_FABRIC_ID, ctx->map.limit);
+ u8 i, j, shift = 4, mask = 0xF;
+ u32 reg, offset = 0x60;
+ u16 dst_node_id;
+
+ /* Get Socket 1 register. */
+ if (dst_fabric_id & df_cfg.socket_id_mask)
+ offset = 0x68;
+
+ /* Read D18F0x06{0,8} (DF::Skt0CsTargetRemap0)/(DF::Skt0CsTargetRemap1) */
+ if (df_indirect_read_broadcast(ctx->node_id, 0, offset, ®))
+ return -EINVAL;
+
+ /* Save 8 remap entries. */
+ for (i = 0, j = 0; i < 8; i++, j++)
+ ctx->map.remap_array[i] = (reg >> (j * shift)) & mask;
+
+ dst_node_id = dst_fabric_id & df_cfg.node_id_mask;
+ dst_node_id >>= df_cfg.node_id_shift;
+
+ /* Read D18F2x090 (DF::Np2ChannelConfig) */
+ if (df_indirect_read_broadcast(dst_node_id, 2, 0x90, ®))
+ return -EINVAL;
+
+ ctx->map.np2_bits = FIELD_GET(DF_LOG2_ADDR_64K_SPACE0, reg);
+ return 0;
+}
+
+static int df2_get_dram_addr_map(struct addr_ctx *ctx)
+{
+ /* Read D18F0x110 (DramBaseAddress). */
+ if (df_indirect_read_instance(ctx->node_id, 0, 0x110 + (8 * ctx->map.num),
+ ctx->inst_id, &ctx->map.base))
+ return -EINVAL;
+
+ /* Read D18F0x114 (DramLimitAddress). */
+ if (df_indirect_read_instance(ctx->node_id, 0, 0x114 + (8 * ctx->map.num),
+ ctx->inst_id, &ctx->map.limit))
+ return -EINVAL;
+
+ return 0;
+}
+
+static int df3_get_dram_addr_map(struct addr_ctx *ctx)
+{
+ if (df2_get_dram_addr_map(ctx))
+ return -EINVAL;
+
+ /* Read D18F0x3F8 (DfGlobalCtl). */
+ if (df_indirect_read_instance(ctx->node_id, 0, 0x3F8,
+ ctx->inst_id, &ctx->map.ctl))
+ return -EINVAL;
+
+ return 0;
+}
+
+static int df4_get_dram_addr_map(struct addr_ctx *ctx)
+{
+ u8 remap_sel, i, j, shift = 4, mask = 0xF;
+ u32 remap_reg;
+
+ /* Read D18F7xE00 (DramBaseAddress). */
+ if (df_indirect_read_instance(ctx->node_id, 7, 0xE00 + (16 * ctx->map.num),
+ ctx->inst_id, &ctx->map.base))
+ return -EINVAL;
+
+ /* Read D18F7xE04 (DramLimitAddress). */
+ if (df_indirect_read_instance(ctx->node_id, 7, 0xE04 + (16 * ctx->map.num),
+ ctx->inst_id, &ctx->map.limit))
+ return -EINVAL;
+
+ /* Read D18F7xE08 (DramAddressCtl). */
+ if (df_indirect_read_instance(ctx->node_id, 7, 0xE08 + (16 * ctx->map.num),
+ ctx->inst_id, &ctx->map.ctl))
+ return -EINVAL;
+
+ /* Read D18F7xE0C (DramAddressIntlv). */
+ if (df_indirect_read_instance(ctx->node_id, 7, 0xE0C + (16 * ctx->map.num),
+ ctx->inst_id, &ctx->map.intlv))
+ return -EINVAL;
+
+ /* Check if Remap Enable bit is valid. */
+ if (!FIELD_GET(DF4_REMAP_EN, ctx->map.ctl))
+ return 0;
+
+ /* Fill with bogus values, because '0' is a valid value. */
+ memset(&ctx->map.remap_array, 0xFF, sizeof(ctx->map.remap_array));
+
+ /* Get Remap registers. */
+ remap_sel = FIELD_GET(DF4_REMAP_SEL, ctx->map.ctl);
+
+ /* Read D18F7x180 (CsTargetRemap0A). */
+ if (df_indirect_read_instance(ctx->node_id, 7, 0x180 + (8 * remap_sel),
+ ctx->inst_id, &remap_reg))
+ return -EINVAL;
+
+ /* Save first 8 remap entries. */
+ for (i = 0, j = 0; i < 8; i++, j++)
+ ctx->map.remap_array[i] = (remap_reg >> (j * shift)) & mask;
+
+ /* Read D18F7x184 (CsTargetRemap0B). */
+ if (df_indirect_read_instance(ctx->node_id, 7, 0x184 + (8 * remap_sel),
+ ctx->inst_id, &remap_reg))
+ return -EINVAL;
+
+ /* Save next 8 remap entries. */
+ for (i = 8, j = 0; i < 16; i++, j++)
+ ctx->map.remap_array[i] = (remap_reg >> (j * shift)) & mask;
+
+ return 0;
+}
+
+static int df4p5_get_dram_addr_map(struct addr_ctx *ctx)
+{
+ u8 remap_sel, i, j, shift = 5, mask = 0x1F;
+ u32 remap_reg;
+
+ /* Read D18F7x200 (DramBaseAddress). */
+ if (df_indirect_read_instance(ctx->node_id, 7, 0x200 + (16 * ctx->map.num),
+ ctx->inst_id, &ctx->map.base))
+ return -EINVAL;
+
+ /* Read D18F7x204 (DramLimitAddress). */
+ if (df_indirect_read_instance(ctx->node_id, 7, 0x204 + (16 * ctx->map.num),
+ ctx->inst_id, &ctx->map.limit))
+ return -EINVAL;
+
+ /* Read D18F7x208 (DramAddressCtl). */
+ if (df_indirect_read_instance(ctx->node_id, 7, 0x208 + (16 * ctx->map.num),
+ ctx->inst_id, &ctx->map.ctl))
+ return -EINVAL;
+
+ /* Read D18F7x20C (DramAddressIntlv). */
+ if (df_indirect_read_instance(ctx->node_id, 7, 0x20C + (16 * ctx->map.num),
+ ctx->inst_id, &ctx->map.intlv))
+ return -EINVAL;
+
+ /* Check if Remap Enable bit is valid. */
+ if (!FIELD_GET(DF4_REMAP_EN, ctx->map.ctl))
+ return 0;
+
+ /* Fill with bogus values, because '0' is a valid value. */
+ memset(&ctx->map.remap_array, 0xFF, sizeof(ctx->map.remap_array));
+
+ /* Get Remap registers. */
+ remap_sel = FIELD_GET(DF4_REMAP_SEL, ctx->map.ctl);
+
+ /* Read D18F7x180 (CsTargetRemap0A). */
+ if (df_indirect_read_instance(ctx->node_id, 7, 0x180 + (24 * remap_sel),
+ ctx->inst_id, &remap_reg))
+ return -EINVAL;
+
+ /* Save first 6 remap entries. */
+ for (i = 0, j = 0; i < 6; i++, j++)
+ ctx->map.remap_array[i] = (remap_reg >> (j * shift)) & mask;
+
+ /* Read D18F7x184 (CsTargetRemap0B). */
+ if (df_indirect_read_instance(ctx->node_id, 7, 0x184 + (24 * remap_sel),
+ ctx->inst_id, &remap_reg))
+ return -EINVAL;
+
+ /* Save next 6 remap entries. */
+ for (i = 6, j = 0; i < 12; i++, j++)
+ ctx->map.remap_array[i] = (remap_reg >> (j * shift)) & mask;
+
+ /* Read D18F7x188 (CsTargetRemap0C). */
+ if (df_indirect_read_instance(ctx->node_id, 7, 0x188 + (24 * remap_sel),
+ ctx->inst_id, &remap_reg))
+ return -EINVAL;
+
+ /* Save next 6 remap entries. */
+ for (i = 12, j = 0; i < 18; i++, j++)
+ ctx->map.remap_array[i] = (remap_reg >> (j * shift)) & mask;
+
+ return 0;
+}
+
+static int get_dram_addr_map(struct addr_ctx *ctx)
+{
+ switch (df_cfg.rev) {
+ case DF2:
+ return df2_get_dram_addr_map(ctx);
+ case DF3:
+ case DF3p5:
+ return df3_get_dram_addr_map(ctx);
+ case DF4:
+ return df4_get_dram_addr_map(ctx);
+ case DF4p5:
+ return df4p5_get_dram_addr_map(ctx);
+ default:
+ ATL_BAD_DF_REV;
+ return -EINVAL;
+ }
+}
+
+static int lookup_cs_fabric_id(struct addr_ctx *ctx)
+{
+ u32 reg;
+
+ /* Read D18F0x50 (FabricBlockInstanceInformation3). */
+ if (df_indirect_read_instance(ctx->node_id, 0, 0x50, ctx->inst_id, ®))
+ return -EINVAL;
+
+ if (df_cfg.rev < DF4p5)
+ ctx->cs_fabric_id = FIELD_GET(DF2_CS_FABRIC_ID, reg);
+ else
+ ctx->cs_fabric_id = FIELD_GET(DF4p5_CS_FABRIC_ID, reg);
+
+ return 0;
+}
+
+static int get_cs_fabric_id(struct addr_ctx *ctx)
+{
+ /* TODO: Add special path for DF3.5 heterogeneous systems. */
+ return lookup_cs_fabric_id(ctx);
+}
+
+static int find_normalized_offset(struct addr_ctx *ctx, u64 *norm_offset)
+{
+ u64 last_offset = 0;
+ bool enabled;
+
+ for (ctx->map.num = 1; ctx->map.num < df_cfg.num_cs_maps; ctx->map.num++) {
+ enabled = false;
+
+ if (get_dram_offset(ctx, &enabled, norm_offset))
+ return -EINVAL;
+
+ /* Continue search if this map's offset is not enabled. */
+ if (!enabled)
+ continue;
+
+ /* Enabled offsets should never be 0. */
+ if (*norm_offset == 0)
+ return -EINVAL;
+
+ /* Offsets should always increase from one map to the next. */
+ if (*norm_offset <= last_offset)
+ return -EINVAL;
+
+ /* Match if this map's offset is less than the current calculated address. */
+ if (ctx->ret_addr >= *norm_offset)
+ break;
+
+ last_offset = *norm_offset;
+ }
+
+ /*
+ * Finished search without finding a match.
+ * Reset to map 0 and no offset.
+ */
+ if (ctx->map.num >= df_cfg.num_cs_maps) {
+ ctx->map.num = 0;
+ *norm_offset = 0;
+ }
+
+ return 0;
+}
+
+static bool valid_map(struct addr_ctx *ctx)
+{
+ if (df_cfg.rev >= DF4)
+ return FIELD_GET(DF_ADDR_RANGE_VAL, ctx->map.ctl);
+
+ return FIELD_GET(DF_ADDR_RANGE_VAL, ctx->map.base);
+}
+
+static int get_address_map_common(struct addr_ctx *ctx)
+{
+ u64 norm_offset = 0;
+
+ if (get_cs_fabric_id(ctx))
+ return -EINVAL;
+
+ if (find_normalized_offset(ctx, &norm_offset))
+ return -EINVAL;
+
+ if (get_dram_addr_map(ctx))
+ return -EINVAL;
+
+ if (!valid_map(ctx))
+ return -EINVAL;
+
+ ctx->ret_addr -= norm_offset;
+
+ return 0;
+}
+
+static u8 get_num_intlv_chan(enum intlv_modes intlv_mode)
+{
+ switch (intlv_mode) {
+ case NONE:
+ return 1;
+ case NOHASH_2CHAN:
+ case DF2_2CHAN_HASH:
+ case DF3_COD4_2CHAN_HASH:
+ case DF4_NPS4_2CHAN_HASH:
+ case DF4p5_NPS4_2CHAN_1K_HASH:
+ case DF4p5_NPS4_2CHAN_2K_HASH:
+ return 2;
+ case DF4_NPS4_3CHAN_HASH:
+ case DF4p5_NPS4_3CHAN_1K_HASH:
+ case DF4p5_NPS4_3CHAN_2K_HASH:
+ return 3;
+ case NOHASH_4CHAN:
+ case DF3_COD2_4CHAN_HASH:
+ case DF4_NPS2_4CHAN_HASH:
+ case DF4p5_NPS2_4CHAN_1K_HASH:
+ case DF4p5_NPS2_4CHAN_2K_HASH:
+ return 4;
+ case DF4_NPS2_5CHAN_HASH:
+ case DF4p5_NPS2_5CHAN_1K_HASH:
+ case DF4p5_NPS2_5CHAN_2K_HASH:
+ return 5;
+ case DF3_6CHAN:
+ case DF4_NPS2_6CHAN_HASH:
+ case DF4p5_NPS2_6CHAN_1K_HASH:
+ case DF4p5_NPS2_6CHAN_2K_HASH:
+ return 6;
+ case NOHASH_8CHAN:
+ case DF3_COD1_8CHAN_HASH:
+ case DF4_NPS1_8CHAN_HASH:
+ case DF4p5_NPS1_8CHAN_1K_HASH:
+ case DF4p5_NPS1_8CHAN_2K_HASH:
+ return 8;
+ case DF4_NPS1_10CHAN_HASH:
+ case DF4p5_NPS1_10CHAN_1K_HASH:
+ case DF4p5_NPS1_10CHAN_2K_HASH:
+ return 10;
+ case DF4_NPS1_12CHAN_HASH:
+ case DF4p5_NPS1_12CHAN_1K_HASH:
+ case DF4p5_NPS1_12CHAN_2K_HASH:
+ return 12;
+ case NOHASH_16CHAN:
+ case DF4p5_NPS1_16CHAN_1K_HASH:
+ case DF4p5_NPS1_16CHAN_2K_HASH:
+ return 16;
+ case DF4p5_NPS0_24CHAN_1K_HASH:
+ case DF4p5_NPS0_24CHAN_2K_HASH:
+ return 24;
+ case NOHASH_32CHAN:
+ return 32;
+ default:
+ ATL_BAD_INTLV_MODE(intlv_mode);
+ return 0;
+ }
+}
+
+static void calculate_intlv_bits(struct addr_ctx *ctx)
+{
+ ctx->map.num_intlv_chan = get_num_intlv_chan(ctx->map.intlv_mode);
+
+ ctx->map.total_intlv_chan = ctx->map.num_intlv_chan;
+ ctx->map.total_intlv_chan *= ctx->map.num_intlv_dies;
+ ctx->map.total_intlv_chan *= ctx->map.num_intlv_sockets;
+
+ /*
+ * Get the number of bits needed to cover this many channels.
+ * order_base_2() rounds up automatically.
+ */
+ ctx->map.total_intlv_bits = order_base_2(ctx->map.total_intlv_chan);
+}
+
+static u8 get_intlv_bit_pos(struct addr_ctx *ctx)
+{
+ u8 addr_sel = 0;
+
+ switch (df_cfg.rev) {
+ case DF2:
+ addr_sel = FIELD_GET(DF2_INTLV_ADDR_SEL, ctx->map.base);
+ break;
+ case DF3:
+ case DF3p5:
+ addr_sel = FIELD_GET(DF3_INTLV_ADDR_SEL, ctx->map.base);
+ break;
+ case DF4:
+ case DF4p5:
+ addr_sel = FIELD_GET(DF4_INTLV_ADDR_SEL, ctx->map.intlv);
+ break;
+ default:
+ ATL_BAD_DF_REV;
+ break;
+ }
+
+ /* Add '8' to get the 'interleave bit position'. */
+ return addr_sel + 8;
+}
+
+static u8 get_num_intlv_dies(struct addr_ctx *ctx)
+{
+ u8 dies = 0;
+
+ switch (df_cfg.rev) {
+ case DF2:
+ dies = FIELD_GET(DF2_INTLV_NUM_DIES, ctx->map.limit);
+ break;
+ case DF3:
+ dies = FIELD_GET(DF3_INTLV_NUM_DIES, ctx->map.base);
+ break;
+ case DF3p5:
+ dies = FIELD_GET(DF3p5_INTLV_NUM_DIES, ctx->map.base);
+ break;
+ case DF4:
+ case DF4p5:
+ dies = FIELD_GET(DF4_INTLV_NUM_DIES, ctx->map.intlv);
+ break;
+ default:
+ ATL_BAD_DF_REV;
+ break;
+ }
+
+ /* Register value is log2, e.g. 0 -> 1 die, 1 -> 2 dies, etc. */
+ return 1 << dies;
+}
+
+static u8 get_num_intlv_sockets(struct addr_ctx *ctx)
+{
+ u8 sockets = 0;
+
+ switch (df_cfg.rev) {
+ case DF2:
+ sockets = FIELD_GET(DF2_INTLV_NUM_SOCKETS, ctx->map.limit);
+ break;
+ case DF3:
+ case DF3p5:
+ sockets = FIELD_GET(DF2_INTLV_NUM_SOCKETS, ctx->map.base);
+ break;
+ case DF4:
+ case DF4p5:
+ sockets = FIELD_GET(DF4_INTLV_NUM_SOCKETS, ctx->map.intlv);
+ break;
+ default:
+ ATL_BAD_DF_REV;
+ break;
+ }
+
+ /* Register value is log2, e.g. 0 -> 1 sockets, 1 -> 2 sockets, etc. */
+ return 1 << sockets;
+}
+
+static int get_global_map_data(struct addr_ctx *ctx)
+{
+ if (get_intlv_mode(ctx)) {
+ pr_warn("Failed to get interleave mode");
+ return -EINVAL;
+ }
+
+ if (ctx->map.intlv_mode == DF3_6CHAN &&
+ df3_6ch_get_dram_addr_map(ctx))
+ return -EINVAL;
+
+ ctx->map.intlv_bit_pos = get_intlv_bit_pos(ctx);
+ ctx->map.num_intlv_dies = get_num_intlv_dies(ctx);
+ ctx->map.num_intlv_sockets = get_num_intlv_sockets(ctx);
+ calculate_intlv_bits(ctx);
+
+ return 0;
+}
+
+static void dump_address_map(struct dram_addr_map *map)
+{
+ u8 i;
+
+ atl_debug("intlv_mode=0x%x", map->intlv_mode);
+ atl_debug("num=0x%x", map->num);
+ atl_debug("base=0x%x", map->base);
+ atl_debug("limit=0x%x", map->limit);
+ atl_debug("ctl=0x%x", map->ctl);
+ atl_debug("intlv=0x%x", map->intlv);
+
+ for (i = 0; i < MAX_CS_CHANNELS; i++)
+ atl_debug("remap_array[%u]=0x%x", i, map->remap_array[i]);
+
+ atl_debug("intlv_bit_pos=%u", map->intlv_bit_pos);
+ atl_debug("num_intlv_chan=%u", map->num_intlv_chan);
+ atl_debug("num_intlv_dies=%u", map->num_intlv_dies);
+ atl_debug("num_intlv_sockets=%u", map->num_intlv_sockets);
+ atl_debug("total_intlv_chan=%u", map->total_intlv_chan);
+ atl_debug("total_intlv_bits=%u", map->total_intlv_bits);
+}
+
+int get_address_map(struct addr_ctx *ctx)
+{
+ int ret = 0;
+
+ /* TODO: Add special path for DF3.5 heterogeneous systems. */
+ ret = get_address_map_common(ctx);
+ if (ret)
+ return ret;
+
+ if (get_global_map_data(ctx))
+ return -EINVAL;
+
+ dump_address_map(&ctx->map);
+
+ return ret;
+}
diff --git a/drivers/platform/x86/amd/atl/reg_fields.h b/drivers/platform/x86/amd/atl/reg_fields.h
new file mode 100644
index 000000000000..d48470e12498
--- /dev/null
+++ b/drivers/platform/x86/amd/atl/reg_fields.h
@@ -0,0 +1,603 @@
+/* SPDX-License-Identifier: GPL-2.0 */
+/*
+ * AMD Address Translation Library
+ *
+ * reg_fields.h : Register field definitions
+ *
+ * Copyright (c) 2023, Advanced Micro Devices, Inc.
+ * All Rights Reserved.
+ *
+ * Author: Yazen Ghannam <Yazen.Ghannam@amd.com>
+ */
+
+/*
+ * Notes on naming:
+ * 1) Use "DF_" prefix for fields that are the same for all revisions.
+ * 2) Use "DFx_" prefix for fields that differ between revisions.
+ * a) "x" is the first major revision where the new field appears.
+ * b) E.g., if DF2 and DF3 have the same field, then call it DF2.
+ * c) E.g., if DF3p5 and DF4 have the same field, then call it DF4.
+ */
+
+/*
+ * Coherent Station Fabric ID
+ *
+ * Access type: Instance
+ *
+ * Register
+ * Rev Fieldname Bits
+ *
+ * D18F0x50 [Fabric Block Instance Information 3]
+ * DF2 BlockFabricId [19:8]
+ * DF3 BlockFabricId [19:8]
+ * DF3p5 BlockFabricId [19:8]
+ * DF4 BlockFabricId [19:8]
+ * DF4p5 BlockFabricId [15:8]
+ */
+#define DF2_CS_FABRIC_ID GENMASK(19, 8)
+#define DF4p5_CS_FABRIC_ID GENMASK(15, 8)
+
+/*
+ * Component ID Mask
+ *
+ * Access type: Broadcast
+ *
+ * Register
+ * Rev Fieldname Bits
+ *
+ * DF2 N/A
+ *
+ * D18F1x208 [System Fabric ID Mask 0]
+ * DF3 ComponentIdMask [9:0]
+ *
+ * D18F1x150 [System Fabric ID Mask 0]
+ * DF3p5 ComponentIdMask [15:0]
+ *
+ * D18F4x1B0 [System Fabric ID Mask 0]
+ * DF4 ComponentIdMask [15:0]
+ * DF4p5 ComponentIdMask [15:0]
+ */
+#define DF3_COMPONENT_ID_MASK GENMASK(9, 0)
+#define DF4_COMPONENT_ID_MASK GENMASK(15, 0)
+
+/*
+ * Destination Fabric ID
+ *
+ * Access type: Instance
+ *
+ * Register
+ * Rev Fieldname Bits
+ *
+ * D18F0x114 [DRAM Limit Address]
+ * DF2 DstFabricID [7:0]
+ * DF3 DstFabricID [9:0]
+ * DF3 DstFabricID [11:0]
+ *
+ * D18F7xE08 [DRAM Address Control]
+ * DF4 DstFabricID [27:16]
+ *
+ * D18F7x208 [DRAM Address Control]
+ * DF4p5 DstFabricID [23:16]
+ */
+#define DF2_DST_FABRIC_ID GENMASK(7, 0)
+#define DF3_DST_FABRIC_ID GENMASK(9, 0)
+#define DF3p5_DST_FABRIC_ID GENMASK(11, 0)
+#define DF4_DST_FABRIC_ID GENMASK(27, 16)
+#define DF4p5_DST_FABRIC_ID GENMASK(23, 16)
+
+/*
+ * Die ID Mask
+ *
+ * Access type: Broadcast
+ *
+ * Register
+ * Rev Fieldname Bits
+ *
+ * D18F1x208 [System Fabric ID Mask]
+ * DF2 DieIdMask [15:8]
+ *
+ * D18F1x20C [System Fabric ID Mask 1]
+ * DF3 DieIdMask [18:16]
+ *
+ * D18F1x158 [System Fabric ID Mask 2]
+ * DF3p5 DieIdMask [15:0]
+ *
+ * D18F4x1B8 [System Fabric ID Mask 2]
+ * DF4 DieIdMask [15:0]
+ * DF4p5 DieIdMask [15:0]
+ */
+#define DF2_DIE_ID_MASK GENMASK(15, 8)
+#define DF3_DIE_ID_MASK GENMASK(18, 16)
+#define DF4_DIE_ID_MASK GENMASK(15, 0)
+
+/*
+ * Die ID Shift
+ *
+ * Access type: Broadcast
+ *
+ * Register
+ * Rev Fieldname Bits
+ *
+ * D18F1x208 [System Fabric ID Mask]
+ * DF2 DieIdShift [27:24]
+ *
+ * DF3 N/A
+ * DF3p5 N/A
+ * DF4 N/A
+ * DF4p5 N/A
+ */
+#define DF2_DIE_ID_SHIFT GENMASK(27, 24)
+
+/*
+ * DRAM Address Range Valid
+ *
+ * Access type: Instance
+ *
+ * Register
+ * Rev Fieldname Bits
+ *
+ * D18F0x110 [DRAM Base Address]
+ * DF2 AddrRngVal [0]
+ * DF3 AddrRngVal [0]
+ * DF3p5 AddrRngVal [0]
+ *
+ * D18F7xE08 [DRAM Address Control]
+ * DF4 AddrRngVal [0]
+ *
+ * D18F7x208 [DRAM Address Control]
+ * DF4p5 AddrRngVal [0]
+ */
+#define DF_ADDR_RANGE_VAL BIT(0)
+
+/*
+ * DRAM Base Address
+ *
+ * Access type: Instance
+ *
+ * Register
+ * Rev Fieldname Bits
+ *
+ * D18F0x110 [DRAM Base Address]
+ * DF2 DramBaseAddr [31:12]
+ * DF3 DramBaseAddr [31:12]
+ * DF3p5 DramBaseAddr [31:12]
+ *
+ * D18F7xE00 [DRAM Base Address]
+ * DF4 DramBaseAddr [27:0]
+ *
+ * D18F7x200 [DRAM Base Address]
+ * DF4p5 DramBaseAddr [27:0]
+ */
+#define DF2_BASE_ADDR GENMASK(31, 12)
+#define DF4_BASE_ADDR GENMASK(27, 0)
+
+/*
+ * DRAM Hole Base
+ *
+ * Access type: Broadcast
+ *
+ * Register
+ * Rev Fieldname Bits
+ *
+ * D18F0x104 [DRAM Hole Control]
+ * DF2 DramHoleBase [31:24]
+ * DF3 DramHoleBase [31:24]
+ * DF3p5 DramHoleBase [31:24]
+ *
+ * D18F7x104 [DRAM Hole Control]
+ * DF4 DramHoleBase [31:24]
+ * DF4p5 DramHoleBase [31:24]
+ */
+#define DF_DRAM_HOLE_BASE_MASK GENMASK(31, 24)
+
+/*
+ * DRAM Limit Address
+ *
+ * Access type: Instance
+ *
+ * Register
+ * Rev Fieldname Bits
+ *
+ * D18F0x114 [DRAM Limit Address]
+ * DF2 DramLimitAddr [31:12]
+ * DF3 DramLimitAddr [31:12]
+ * DF3p5 DramLimitAddr [31:12]
+ *
+ * D18F7xE04 [DRAM Limit Address]
+ * DF4 DramLimitAddr [27:0]
+ *
+ * D18F7x204 [DRAM Limit Address]
+ * DF4p5 DramLimitAddr [27:0]
+ */
+#define DF2_DRAM_LIMIT_ADDR GENMASK(31, 12)
+#define DF4_DRAM_LIMIT_ADDR GENMASK(27, 0)
+
+/*
+ * Hash Interleave Controls
+ *
+ * Access type: Instance
+ *
+ * Register
+ * Rev Fieldname Bits
+ *
+ * DF2 N/A
+ *
+ * D18F0x3F8 [DF Global Control]
+ * DF3 GlbHashIntlvCtl64K [20]
+ * GlbHashIntlvCtl2M [21]
+ * GlbHashIntlvCtl1G [22]
+ *
+ * DF3p5 GlbHashIntlvCtl64K [20]
+ * GlbHashIntlvCtl2M [21]
+ * GlbHashIntlvCtl1G [22]
+ *
+ * D18F7xE08 [DRAM Address Control]
+ * DF4 HashIntlvCtl64K [8]
+ * HashIntlvCtl2M [9]
+ * HashIntlvCtl1G [10]
+ *
+ * D18F7x208 [DRAM Address Control]
+ * DF4p5 HashIntlvCtl4K [7]
+ * HashIntlvCtl64K [8]
+ * HashIntlvCtl2M [9]
+ * HashIntlvCtl1G [10]
+ * HashIntlvCtl1T [15]
+ */
+#define DF3_HASH_CTL_64K BIT(20)
+#define DF3_HASH_CTL_2M BIT(21)
+#define DF3_HASH_CTL_1G BIT(22)
+#define DF4_HASH_CTL_4K BIT(7)
+#define DF4_HASH_CTL_64K BIT(8)
+#define DF4_HASH_CTL_2M BIT(9)
+#define DF4_HASH_CTL_1G BIT(10)
+#define DF4_HASH_CTL_1T BIT(15)
+
+/*
+ * High Address Offset
+ *
+ * Access type: Instance
+ *
+ * Register
+ * Rev Fieldname Bits
+ *
+ * D18F0x1B4 [DRAM Offset]
+ * DF2 HiAddrOffset [31:20]
+ * DF3 HiAddrOffset [31:12]
+ * DF3p5 HiAddrOffset [31:12]
+ *
+ * D18F7x140 [DRAM Offset]
+ * DF4 HiAddrOffset [24:1]
+ * DF4p5 HiAddrOffset [24:1]
+ */
+#define DF2_HI_ADDR_OFFSET GENMASK(31, 20)
+#define DF3_HI_ADDR_OFFSET GENMASK(31, 12)
+#define DF4_HI_ADDR_OFFSET GENMASK(24, 1)
+
+/*
+ * High Address Offset Enable
+ *
+ * Access type: Instance
+ *
+ * Register
+ * Rev Fieldname Bits
+ *
+ * D18F0x1B4 [DRAM Offset]
+ * DF2 HiAddrOffsetEn [0]
+ * DF3 HiAddrOffsetEn [0]
+ * DF3p5 HiAddrOffsetEn [0]
+ *
+ * D18F7x140 [DRAM Offset]
+ * DF4 HiAddrOffsetEn [0]
+ * DF4p5 HiAddrOffsetEn [0]
+ */
+#define DF_HI_ADDR_OFFSET_EN BIT(0)
+
+/*
+ * Interleave Address Select
+ *
+ * Access type: Instance
+ *
+ * Register
+ * Rev Fieldname Bits
+ *
+ * D18F0x110 [DRAM Base Address]
+ * DF2 IntLvAddrSel [10:8]
+ * DF3 IntLvAddrSel [11:9]
+ * DF3p5 IntLvAddrSel [11:9]
+ *
+ * D18F7xE0C [DRAM Address Interleave]
+ * DF4 IntLvAddrSel [2:0]
+ *
+ * D18F7x20C [DRAM Address Interleave]
+ * DF4p5 IntLvAddrSel [2:0]
+ */
+#define DF2_INTLV_ADDR_SEL GENMASK(10, 8)
+#define DF3_INTLV_ADDR_SEL GENMASK(11, 9)
+#define DF4_INTLV_ADDR_SEL GENMASK(2, 0)
+
+/*
+ * Interleave Number of Channels
+ *
+ * Access type: Instance
+ *
+ * Register
+ * Rev Fieldname Bits
+ *
+ * D18F0x110 [DRAM Base Address]
+ * DF2 IntLvNumChan [7:4]
+ * DF3 IntLvNumChan [5:2]
+ * DF3p5 IntLvNumChan [6:2]
+ *
+ * D18F7xE0C [DRAM Address Interleave]
+ * DF4 IntLvNumChan [8:4]
+ *
+ * D18F7x20C [DRAM Address Interleave]
+ * DF4p5 IntLvNumChan [9:4]
+ */
+#define DF2_INTLV_NUM_CHAN GENMASK(7, 4)
+#define DF3_INTLV_NUM_CHAN GENMASK(5, 2)
+#define DF3p5_INTLV_NUM_CHAN GENMASK(6, 2)
+#define DF4_INTLV_NUM_CHAN GENMASK(8, 4)
+#define DF4p5_INTLV_NUM_CHAN GENMASK(9, 4)
+
+/*
+ * Interleave Number of Dies
+ *
+ * Access type: Instance
+ *
+ * Register
+ * Rev Fieldname Bits
+ *
+ * D18F0x114 [DRAM Limit Address]
+ * DF2 IntLvNumDies [11:10]
+ *
+ * D18F0x110 [DRAM Base Address]
+ * DF3 IntLvNumDies [7:6]
+ * DF3p5 IntLvNumDies [7]
+ *
+ * D18F7xE0C [DRAM Address Interleave]
+ * DF4 IntLvNumDies [13:12]
+ *
+ * D18F7x20C [DRAM Address Interleave]
+ * DF4p5 IntLvNumDies [13:12]
+ */
+#define DF2_INTLV_NUM_DIES GENMASK(11, 10)
+#define DF3_INTLV_NUM_DIES GENMASK(7, 6)
+#define DF3p5_INTLV_NUM_DIES BIT(7)
+#define DF4_INTLV_NUM_DIES GENMASK(13, 12)
+
+/*
+ * Interleave Number of Sockets
+ *
+ * Access type: Instance
+ *
+ * Register
+ * Rev Fieldname Bits
+ *
+ * D18F0x114 [DRAM Limit Address]
+ * DF2 IntLvNumSockets [8]
+ *
+ * D18F0x110 [DRAM Base Address]
+ * DF3 IntLvNumSockets [8]
+ * DF3p5 IntLvNumSockets [8]
+ *
+ * D18F7xE0C [DRAM Address Interleave]
+ * DF4 IntLvNumSockets [18]
+ *
+ * D18F7x20C [DRAM Address Interleave]
+ * DF4p5 IntLvNumSockets [18]
+ */
+#define DF2_INTLV_NUM_SOCKETS BIT(8)
+#define DF4_INTLV_NUM_SOCKETS BIT(18)
+
+/*
+ * Legacy MMIO Hole Enable
+ *
+ * Access type: Instance
+ *
+ * Register
+ * Rev Fieldname Bits
+ *
+ * D18F0x110 [DRAM Base Address]
+ * DF2 LgcyMmioHoleEn [1]
+ * DF3 LgcyMmioHoleEn [1]
+ * DF3p5 LgcyMmioHoleEn [1]
+ *
+ * D18F7xE08 [DRAM Address Control]
+ * DF4 LgcyMmioHoleEn [1]
+ *
+ * D18F7x208 [DRAM Address Control]
+ * DF4p5 LgcyMmioHoleEn [1]
+ */
+#define DF_LEGACY_MMIO_HOLE_EN BIT(1)
+
+/*
+ * Log2 Address 64K Space 0
+ *
+ * Access type: Instance
+ *
+ * Register
+ * Rev Fieldname Bits
+ *
+ * DF2 N/A
+ *
+ * D18F2x90 [Non-power-of-2 channel Configuration Register for CS DRAM Address Maps]
+ * DF3 Log2Addr64KSpace0 [5:0]
+ *
+ * DF3p5 N/A
+ * DF4 N/A
+ * DF4p5 N/A
+ */
+#define DF_LOG2_ADDR_64K_SPACE0 GENMASK(5, 0)
+
+/*
+ * Major Revision
+ *
+ * Access type: Broadcast
+ *
+ * Register
+ * Rev Fieldname Bits
+ *
+ * DF2 N/A
+ * DF3 N/A
+ * DF3p5 N/A
+ *
+ * D18F0x040 [Fabric Block Instance Count]
+ * DF4 MajorRevision [27:24]
+ * DF4p5 MajorRevision [27:24]
+ */
+#define DF_MAJOR_REVISION GENMASK(27, 24)
+
+/*
+ * Minor Revision
+ *
+ * Access type: Broadcast
+ *
+ * Register
+ * Rev Fieldname Bits
+ *
+ * DF2 N/A
+ * DF3 N/A
+ * DF3p5 N/A
+ *
+ * D18F0x040 [Fabric Block Instance Count]
+ * DF4 MinorRevision [23:16]
+ * DF4p5 MinorRevision [23:16]
+ */
+#define DF_MINOR_REVISION GENMASK(23, 16)
+
+/*
+ * Node ID Mask
+ *
+ * Access type: Broadcast
+ *
+ * Register
+ * Rev Fieldname Bits
+ *
+ * DF2 N/A
+ *
+ * D18F1x208 [System Fabric ID Mask 0]
+ * DF3 NodeIdMask [25:16]
+ *
+ * D18F1x150 [System Fabric ID Mask 0]
+ * DF3p5 NodeIdMask [31:16]
+ *
+ * D18F4x1B0 [System Fabric ID Mask 0]
+ * DF4 NodeIdMask [31:16]
+ * DF4p5 NodeIdMask [31:16]
+ */
+#define DF3_NODE_ID_MASK GENMASK(25, 16)
+#define DF4_NODE_ID_MASK GENMASK(31, 16)
+
+/*
+ * Node ID Shift
+ *
+ * Access type: Broadcast
+ *
+ * Register
+ * Rev Fieldname Bits
+ *
+ * DF2 N/A
+ *
+ * D18F1x20C [System Fabric ID Mask 1]
+ * DF3 NodeIdShift [3:0]
+ *
+ * D18F1x154 [System Fabric ID Mask 1]
+ * DF3p5 NodeIdShift [3:0]
+ *
+ * D18F4x1B4 [System Fabric ID Mask 1]
+ * DF4 NodeIdShift [3:0]
+ * DF4p5 NodeIdShift [3:0]
+ */
+#define DF3_NODE_ID_SHIFT GENMASK(3, 0)
+
+/*
+ * Remap Enable
+ *
+ * Access type: Instance
+ *
+ * Register
+ * Rev Fieldname Bits
+ *
+ * DF2 N/A
+ * DF3 N/A
+ * DF3p5 N/A
+ *
+ * D18F7xE08 [DRAM Address Control]
+ * DF4 RemapEn [4]
+ *
+ * D18F7x208 [DRAM Address Control]
+ * DF4p5 RemapEn [4]
+ */
+#define DF4_REMAP_EN BIT(4)
+
+/*
+ * Remap Select
+ *
+ * Access type: Instance
+ *
+ * Register
+ * Rev Fieldname Bits
+ *
+ * DF2 N/A
+ * DF3 N/A
+ * DF3p5 N/A
+ *
+ * D18F7xE08 [DRAM Address Control]
+ * DF4 RemapSel [7:5]
+ *
+ * D18F7x208 [DRAM Address Control]
+ * DF4p5 RemapSel [6:5]
+ */
+#define DF4_REMAP_SEL GENMASK(7, 5)
+#define DF4p5_REMAP_SEL GENMASK(6, 5)
+
+/*
+ * Socket ID Mask
+ *
+ * Access type: Broadcast
+ *
+ * Register
+ * Rev Fieldname Bits
+ *
+ * D18F1x208 [System Fabric ID Mask]
+ * DF2 SocketIdMask [23:16]
+ *
+ * D18F1x20C [System Fabric ID Mask 1]
+ * DF3 SocketIdMask [26:24]
+ *
+ * D18F1x158 [System Fabric ID Mask 2]
+ * DF3p5 SocketIdMask [31:16]
+ *
+ * D18F4x1B8 [System Fabric ID Mask 2]
+ * DF4 SocketIdMask [31:16]
+ * DF4p5 SocketIdMask [31:16]
+ */
+#define DF2_SOCKET_ID_MASK GENMASK(23, 16)
+#define DF3_SOCKET_ID_MASK GENMASK(26, 24)
+#define DF4_SOCKET_ID_MASK GENMASK(31, 16)
+
+/*
+ * Socket ID Shift
+ *
+ * Access type: Broadcast
+ *
+ * Register
+ * Rev Fieldname Bits
+ *
+ * D18F1x208 [System Fabric ID Mask]
+ * DF2 SocketIdShift [31:28]
+ *
+ * D18F1x20C [System Fabric ID Mask 1]
+ * DF3 SocketIdShift [9:8]
+ *
+ * D18F1x158 [System Fabric ID Mask 2]
+ * DF3p5 SocketIdShift [11:8]
+ *
+ * D18F4x1B4 [System Fabric ID Mask 1]
+ * DF4 SocketIdShift [11:8]
+ * DF4p5 SocketIdShift [11:8]
+ */
+#define DF2_SOCKET_ID_SHIFT GENMASK(31, 28)
+#define DF3_SOCKET_ID_SHIFT GENMASK(9, 8)
+#define DF4_SOCKET_ID_SHIFT GENMASK(11, 8)
diff --git a/drivers/platform/x86/amd/atl/system.c b/drivers/platform/x86/amd/atl/system.c
new file mode 100644
index 000000000000..86488138e120
--- /dev/null
+++ b/drivers/platform/x86/amd/atl/system.c
@@ -0,0 +1,282 @@
+// SPDX-License-Identifier: GPL-2.0-or-later
+/*
+ * AMD Address Translation Library
+ *
+ * system.c : Functions to read and save system-wide data
+ *
+ * Copyright (c) 2023, Advanced Micro Devices, Inc.
+ * All Rights Reserved.
+ *
+ * Author: Yazen Ghannam <Yazen.Ghannam@amd.com>
+ */
+
+#include "internal.h"
+
+int determine_node_id(struct addr_ctx *ctx, u8 socket_id, u8 die_id)
+{
+ u16 socket_id_bits, die_id_bits;
+
+ if (socket_id > 0 && df_cfg.socket_id_mask == 0) {
+ pr_warn("%s: Invalid socket inputs: socket_id=%u socket_id_mask=0x%x",
+ __func__, socket_id, df_cfg.socket_id_mask);
+ return -EINVAL;
+ }
+
+ /* Do each step independently to avoid shift out-of-bounds issues. */
+ socket_id_bits = socket_id;
+ socket_id_bits <<= df_cfg.socket_id_shift;
+ socket_id_bits &= df_cfg.socket_id_mask;
+
+ if (die_id > 0 && df_cfg.die_id_mask == 0) {
+ pr_warn("%s: Invalid die inputs: die_id=%u die_id_mask=0x%x",
+ __func__, die_id, df_cfg.die_id_mask);
+ return -EINVAL;
+ }
+
+ /* Do each step independently to avoid shift out-of-bounds issues. */
+ die_id_bits = die_id;
+ die_id_bits <<= df_cfg.die_id_shift;
+ die_id_bits &= df_cfg.die_id_mask;
+
+ ctx->node_id = (socket_id_bits | die_id_bits) >> df_cfg.node_id_shift;
+ return 0;
+}
+
+static void df2_get_masks_shifts(u32 mask0)
+{
+ df_cfg.socket_id_shift = FIELD_GET(DF2_SOCKET_ID_SHIFT, mask0);
+ df_cfg.socket_id_mask = FIELD_GET(DF2_SOCKET_ID_MASK, mask0);
+ df_cfg.die_id_shift = FIELD_GET(DF2_DIE_ID_SHIFT, mask0);
+ df_cfg.die_id_mask = FIELD_GET(DF2_DIE_ID_MASK, mask0);
+ df_cfg.node_id_shift = df_cfg.die_id_shift;
+ df_cfg.node_id_mask = df_cfg.socket_id_mask | df_cfg.die_id_mask;
+ df_cfg.component_id_mask = ~df_cfg.node_id_mask;
+}
+
+static void df3_get_masks_shifts(u32 mask0, u32 mask1)
+{
+ df_cfg.component_id_mask = FIELD_GET(DF3_COMPONENT_ID_MASK, mask0);
+ df_cfg.node_id_mask = FIELD_GET(DF3_NODE_ID_MASK, mask0);
+
+ df_cfg.node_id_shift = FIELD_GET(DF3_NODE_ID_SHIFT, mask1);
+ df_cfg.socket_id_shift = FIELD_GET(DF3_SOCKET_ID_SHIFT, mask1);
+ df_cfg.socket_id_mask = FIELD_GET(DF3_SOCKET_ID_MASK, mask1);
+ df_cfg.die_id_mask = FIELD_GET(DF3_DIE_ID_MASK, mask1);
+}
+
+static void df3p5_get_masks_shifts(u32 mask0, u32 mask1, u32 mask2)
+{
+ df_cfg.component_id_mask = FIELD_GET(DF4_COMPONENT_ID_MASK, mask0);
+ df_cfg.node_id_mask = FIELD_GET(DF4_NODE_ID_MASK, mask0);
+
+ df_cfg.node_id_shift = FIELD_GET(DF3_NODE_ID_SHIFT, mask1);
+ df_cfg.socket_id_shift = FIELD_GET(DF4_SOCKET_ID_SHIFT, mask1);
+
+ df_cfg.socket_id_mask = FIELD_GET(DF4_SOCKET_ID_MASK, mask2);
+ df_cfg.die_id_mask = FIELD_GET(DF4_DIE_ID_MASK, mask2);
+}
+
+static void df4_get_masks_shifts(u32 mask0, u32 mask1, u32 mask2)
+{
+ df3p5_get_masks_shifts(mask0, mask1, mask2);
+
+ if (!(df_cfg.flags.genoa_quirk && df_cfg.socket_id_shift == 1))
+ return;
+
+ df_cfg.socket_id_shift = 0;
+ df_cfg.socket_id_mask = 1;
+ df_cfg.die_id_shift = 0;
+ df_cfg.die_id_mask = 0;
+ df_cfg.node_id_shift = 8;
+ df_cfg.node_id_mask = 0x100;
+}
+
+static int df4_get_fabric_id_mask_registers(void)
+{
+ u32 mask0, mask1, mask2;
+
+ /* Read D18F4x1B0 (SystemFabricIdMask0) */
+ if (df_indirect_read_broadcast(0, 4, 0x1B0, &mask0))
+ return -EINVAL;
+
+ /* Read D18F4x1B4 (SystemFabricIdMask1) */
+ if (df_indirect_read_broadcast(0, 4, 0x1B4, &mask1))
+ return -EINVAL;
+
+ /* Read D18F4x1B8 (SystemFabricIdMask2) */
+ if (df_indirect_read_broadcast(0, 4, 0x1B8, &mask2))
+ return -EINVAL;
+
+ df4_get_masks_shifts(mask0, mask1, mask2);
+ return 0;
+}
+
+static int df4_determine_df_rev(u32 reg)
+{
+ df_cfg.rev = FIELD_GET(DF_MINOR_REVISION, reg) < 5 ? DF4 : DF4p5;
+
+ /* Check for special cases or quirks based on Device/Vendor IDs.*/
+
+ /* Read D18F0x000 (DeviceVendorId0) */
+ if (df_indirect_read_broadcast(0, 0, 0, ®))
+ return -EINVAL;
+
+ if (reg == DF_FUNC0_ID_GENOA)
+ df_cfg.flags.genoa_quirk = 1;
+
+ return df4_get_fabric_id_mask_registers();
+}
+
+static int determine_df_rev_legacy(void)
+{
+ u32 fabric_id_mask0, fabric_id_mask1, fabric_id_mask2;
+
+ /*
+ * Check for DF3.5.
+ *
+ * Component ID Mask must be non-zero. Register D18F1x150 is
+ * reserved pre-DF3.5, so value will be Read-as-Zero.
+ */
+
+ /* Read D18F1x150 (SystemFabricIdMask0). */
+ if (df_indirect_read_broadcast(0, 1, 0x150, &fabric_id_mask0))
+ return -EINVAL;
+
+ if (FIELD_GET(DF4_COMPONENT_ID_MASK, fabric_id_mask0)) {
+ df_cfg.rev = DF3p5;
+
+ /* Read D18F1x154 (SystemFabricIdMask1) */
+ if (df_indirect_read_broadcast(0, 1, 0x154, &fabric_id_mask1))
+ return -EINVAL;
+
+ /* Read D18F1x158 (SystemFabricIdMask2) */
+ if (df_indirect_read_broadcast(0, 1, 0x158, &fabric_id_mask2))
+ return -EINVAL;
+
+ df3p5_get_masks_shifts(fabric_id_mask0, fabric_id_mask1, fabric_id_mask2);
+ return 0;
+ }
+
+ /*
+ * Check for DF3.
+ *
+ * Component ID Mask must be non-zero. Field is Read-as-Zero on DF2.
+ */
+
+ /* Read D18F1x208 (SystemFabricIdMask). */
+ if (df_indirect_read_broadcast(0, 1, 0x208, &fabric_id_mask0))
+ return -EINVAL;
+
+ if (FIELD_GET(DF3_COMPONENT_ID_MASK, fabric_id_mask0)) {
+ df_cfg.rev = DF3;
+
+ /* Read D18F1x20C (SystemFabricIdMask1) */
+ if (df_indirect_read_broadcast(0, 1, 0x20C, &fabric_id_mask1))
+ return -EINVAL;
+
+ df3_get_masks_shifts(fabric_id_mask0, fabric_id_mask1);
+ return 0;
+ }
+
+ /* Default to DF2. */
+ df_cfg.rev = DF2;
+ df2_get_masks_shifts(fabric_id_mask0);
+ return 0;
+}
+
+static int determine_df_rev(void)
+{
+ u32 reg;
+ u8 rev;
+
+ if (df_cfg.rev != UNKNOWN)
+ return 0;
+
+ /* Read D18F0x40 (FabricBlockInstanceCount). */
+ if (df_indirect_read_broadcast(0, 0, 0x40, ®))
+ return -EINVAL;
+
+ /*
+ * Revision fields added for DF4 and later.
+ *
+ * Major revision of '0' is found pre-DF4. Field is Read-as-Zero.
+ */
+ rev = FIELD_GET(DF_MAJOR_REVISION, reg);
+ if (!rev)
+ return determine_df_rev_legacy();
+
+ /*
+ * Fail out for major revisions other than '4'.
+ *
+ * Explicit support should be added for newer systems to avoid issues.
+ */
+ if (rev == 4)
+ return df4_determine_df_rev(reg);
+
+ return -EINVAL;
+}
+
+static void get_num_maps(void)
+{
+ switch (df_cfg.rev) {
+ case DF2:
+ case DF3:
+ case DF3p5:
+ df_cfg.num_cs_maps = 2;
+ break;
+ case DF4:
+ df_cfg.num_cs_maps = 4;
+ break;
+ case DF4p5:
+ df_cfg.num_cs_maps = 4;
+ break;
+ default:
+ ATL_BAD_DF_REV;
+ }
+}
+
+static void apply_node_id_shift(void)
+{
+ if (df_cfg.rev == DF2)
+ return;
+
+ df_cfg.die_id_shift = df_cfg.node_id_shift;
+ df_cfg.die_id_mask <<= df_cfg.node_id_shift;
+ df_cfg.socket_id_mask <<= df_cfg.node_id_shift;
+ df_cfg.socket_id_shift += df_cfg.node_id_shift;
+}
+
+static void dump_df_cfg(void)
+{
+ atl_debug("rev=0x%x", df_cfg.rev);
+
+ atl_debug("component_id_mask=0x%x", df_cfg.component_id_mask);
+ atl_debug("die_id_mask=0x%x", df_cfg.die_id_mask);
+ atl_debug("node_id_mask=0x%x", df_cfg.node_id_mask);
+ atl_debug("socket_id_mask=0x%x", df_cfg.socket_id_mask);
+
+ atl_debug("die_id_shift=0x%x", df_cfg.die_id_shift);
+ atl_debug("node_id_shift=0x%x", df_cfg.node_id_shift);
+ atl_debug("socket_id_shift=0x%x", df_cfg.socket_id_shift);
+
+ atl_debug("num_cs_maps=%u", df_cfg.num_cs_maps);
+
+ atl_debug("flags.legacy_ficaa=%u", df_cfg.flags.legacy_ficaa);
+ atl_debug("flags.genoa_quirk=%u", df_cfg.flags.genoa_quirk);
+}
+
+int get_df_system_info(void)
+{
+ if (determine_df_rev()) {
+ pr_warn("Failed to determine DF Revision");
+ return -EINVAL;
+ }
+
+ apply_node_id_shift();
+
+ get_num_maps();
+
+ dump_df_cfg();
+
+ return 0;
+}
diff --git a/drivers/platform/x86/amd/atl/umc.c b/drivers/platform/x86/amd/atl/umc.c
new file mode 100644
index 000000000000..25c60d752fcb
--- /dev/null
+++ b/drivers/platform/x86/amd/atl/umc.c
@@ -0,0 +1,53 @@
+// SPDX-License-Identifier: GPL-2.0-or-later
+/*
+ * AMD Address Translation Library
+ *
+ * umc.c : Unified Memory Controller (UMC) topology helpers
+ *
+ * Copyright (c) 2023, Advanced Micro Devices, Inc.
+ * All Rights Reserved.
+ *
+ * Author: Yazen Ghannam <Yazen.Ghannam@amd.com>
+ */
+
+#include "internal.h"
+
+static u8 get_socket_id(struct mce *m)
+{
+ return m->socketid;
+}
+
+static u8 get_die_id(struct mce *m)
+{
+ /*
+ * For CPUs, this is the AMD Node ID modulo the number
+ * of AMD Nodes per socket.
+ */
+ return topology_die_id(m->extcpu) % amd_get_nodes_per_socket();
+}
+
+static u64 get_norm_addr(struct mce *m)
+{
+ return m->addr;
+}
+
+#define UMC_CHANNEL_NUM GENMASK(31, 20)
+static u8 get_cs_inst_id(struct mce *m)
+{
+ return FIELD_GET(UMC_CHANNEL_NUM, m->ipid);
+}
+
+int amd_umc_mca_addr_to_sys_addr(struct mce *m, u64 *sys_addr)
+{
+ u8 cs_inst_id = get_cs_inst_id(m);
+ u8 socket_id = get_socket_id(m);
+ u64 addr = get_norm_addr(m);
+ u8 die_id = get_die_id(m);
+
+ if (norm_to_sys_addr(socket_id, die_id, cs_inst_id, &addr))
+ return -EINVAL;
+
+ *sys_addr = addr;
+ return 0;
+}
+EXPORT_SYMBOL_GPL(amd_umc_mca_addr_to_sys_addr);
diff --git a/include/linux/amd-atl.h b/include/linux/amd-atl.h
new file mode 100644
index 000000000000..ff6ed886c99e
--- /dev/null
+++ b/include/linux/amd-atl.h
@@ -0,0 +1,18 @@
+/* SPDX-License-Identifier: GPL-2.0-only */
+/*
+ * AMD Address Translation Library
+ *
+ * Copyright (c) 2023, Advanced Micro Devices, Inc.
+ * All Rights Reserved.
+ *
+ * Author: Yazen Ghannam <Yazen.Ghannam@amd.com>
+ */
+
+#ifndef _AMD_ATL_H
+#define _AMD_ATL_H
+
+#include <uapi/asm/mce.h>
+
+int amd_umc_mca_addr_to_sys_addr(struct mce *m, u64 *sys_addr);
+
+#endif /* _AMD_ATL_H */
--
2.34.1
^ permalink raw reply related [flat|nested] 15+ messages in thread
* [PATCH 2/2] EDAC/amd64: Use new AMD Address Translation Library
2023-08-02 18:55 [PATCH 0/2] AMD Address Translation Library Yazen Ghannam
2023-08-02 18:55 ` [PATCH 1/2] platform/x86/amd: Introduce " Yazen Ghannam
@ 2023-08-02 18:55 ` Yazen Ghannam
1 sibling, 0 replies; 15+ messages in thread
From: Yazen Ghannam @ 2023-08-02 18:55 UTC (permalink / raw)
To: bp, linux-edac, hdegoede, markgross, platform-driver-x86
Cc: linux-kernel, avadhut.naik, mario.limonciello, Yazen Ghannam
Remove old address translation code and use the new AMD Address
Translation Library.
Use "imply" in Kconfig so that the "AMD_ATL" config option takes the
value of "EDAC_AMD64" as its default.
Signed-off-by: Yazen Ghannam <yazen.ghannam@amd.com>
---
drivers/edac/Kconfig | 1 +
drivers/edac/amd64_edac.c | 282 +-------------------------------------
2 files changed, 7 insertions(+), 276 deletions(-)
diff --git a/drivers/edac/Kconfig b/drivers/edac/Kconfig
index 110e99b86a66..87feeb24b3ff 100644
--- a/drivers/edac/Kconfig
+++ b/drivers/edac/Kconfig
@@ -78,6 +78,7 @@ config EDAC_GHES
config EDAC_AMD64
tristate "AMD64 (Opteron, Athlon64)"
depends on AMD_NB && EDAC_DECODE_MCE
+ imply AMD_ATL
help
Support for error detection and correction of DRAM ECC errors on
the AMD64 families (>= K8) of memory controllers.
diff --git a/drivers/edac/amd64_edac.c b/drivers/edac/amd64_edac.c
index e3b59c488ed1..86059834446f 100644
--- a/drivers/edac/amd64_edac.c
+++ b/drivers/edac/amd64_edac.c
@@ -1,6 +1,7 @@
// SPDX-License-Identifier: GPL-2.0-only
#include "amd64_edac.h"
#include <asm/amd_nb.h>
+#include <linux/amd-atl.h>
static struct edac_pci_ctl_info *pci_ctl;
@@ -1043,281 +1044,6 @@ static int fixup_node_id(int node_id, struct mce *m)
return nid - gpu_node_map.base_node_id + 1;
}
-/* Protect the PCI config register pairs used for DF indirect access. */
-static DEFINE_MUTEX(df_indirect_mutex);
-
-/*
- * Data Fabric Indirect Access uses FICAA/FICAD.
- *
- * Fabric Indirect Configuration Access Address (FICAA): Constructed based
- * on the device's Instance Id and the PCI function and register offset of
- * the desired register.
- *
- * Fabric Indirect Configuration Access Data (FICAD): There are FICAD LO
- * and FICAD HI registers but so far we only need the LO register.
- *
- * Use Instance Id 0xFF to indicate a broadcast read.
- */
-#define DF_BROADCAST 0xFF
-static int __df_indirect_read(u16 node, u8 func, u16 reg, u8 instance_id, u32 *lo)
-{
- struct pci_dev *F4;
- u32 ficaa;
- int err = -ENODEV;
-
- if (node >= amd_nb_num())
- goto out;
-
- F4 = node_to_amd_nb(node)->link;
- if (!F4)
- goto out;
-
- ficaa = (instance_id == DF_BROADCAST) ? 0 : 1;
- ficaa |= reg & 0x3FC;
- ficaa |= (func & 0x7) << 11;
- ficaa |= instance_id << 16;
-
- mutex_lock(&df_indirect_mutex);
-
- err = pci_write_config_dword(F4, 0x5C, ficaa);
- if (err) {
- pr_warn("Error writing DF Indirect FICAA, FICAA=0x%x\n", ficaa);
- goto out_unlock;
- }
-
- err = pci_read_config_dword(F4, 0x98, lo);
- if (err)
- pr_warn("Error reading DF Indirect FICAD LO, FICAA=0x%x.\n", ficaa);
-
-out_unlock:
- mutex_unlock(&df_indirect_mutex);
-
-out:
- return err;
-}
-
-static int df_indirect_read_instance(u16 node, u8 func, u16 reg, u8 instance_id, u32 *lo)
-{
- return __df_indirect_read(node, func, reg, instance_id, lo);
-}
-
-static int df_indirect_read_broadcast(u16 node, u8 func, u16 reg, u32 *lo)
-{
- return __df_indirect_read(node, func, reg, DF_BROADCAST, lo);
-}
-
-struct addr_ctx {
- u64 ret_addr;
- u32 tmp;
- u16 nid;
- u8 inst_id;
-};
-
-static int umc_normaddr_to_sysaddr(u64 norm_addr, u16 nid, u8 umc, u64 *sys_addr)
-{
- u64 dram_base_addr, dram_limit_addr, dram_hole_base;
-
- u8 die_id_shift, die_id_mask, socket_id_shift, socket_id_mask;
- u8 intlv_num_dies, intlv_num_chan, intlv_num_sockets;
- u8 intlv_addr_sel, intlv_addr_bit;
- u8 num_intlv_bits, hashed_bit;
- u8 lgcy_mmio_hole_en, base = 0;
- u8 cs_mask, cs_id = 0;
- bool hash_enabled = false;
-
- struct addr_ctx ctx;
-
- memset(&ctx, 0, sizeof(ctx));
-
- /* Start from the normalized address */
- ctx.ret_addr = norm_addr;
-
- ctx.nid = nid;
- ctx.inst_id = umc;
-
- /* Read D18F0x1B4 (DramOffset), check if base 1 is used. */
- if (df_indirect_read_instance(nid, 0, 0x1B4, umc, &ctx.tmp))
- goto out_err;
-
- /* Remove HiAddrOffset from normalized address, if enabled: */
- if (ctx.tmp & BIT(0)) {
- u64 hi_addr_offset = (ctx.tmp & GENMASK_ULL(31, 20)) << 8;
-
- if (norm_addr >= hi_addr_offset) {
- ctx.ret_addr -= hi_addr_offset;
- base = 1;
- }
- }
-
- /* Read D18F0x110 (DramBaseAddress). */
- if (df_indirect_read_instance(nid, 0, 0x110 + (8 * base), umc, &ctx.tmp))
- goto out_err;
-
- /* Check if address range is valid. */
- if (!(ctx.tmp & BIT(0))) {
- pr_err("%s: Invalid DramBaseAddress range: 0x%x.\n",
- __func__, ctx.tmp);
- goto out_err;
- }
-
- lgcy_mmio_hole_en = ctx.tmp & BIT(1);
- intlv_num_chan = (ctx.tmp >> 4) & 0xF;
- intlv_addr_sel = (ctx.tmp >> 8) & 0x7;
- dram_base_addr = (ctx.tmp & GENMASK_ULL(31, 12)) << 16;
-
- /* {0, 1, 2, 3} map to address bits {8, 9, 10, 11} respectively */
- if (intlv_addr_sel > 3) {
- pr_err("%s: Invalid interleave address select %d.\n",
- __func__, intlv_addr_sel);
- goto out_err;
- }
-
- /* Read D18F0x114 (DramLimitAddress). */
- if (df_indirect_read_instance(nid, 0, 0x114 + (8 * base), umc, &ctx.tmp))
- goto out_err;
-
- intlv_num_sockets = (ctx.tmp >> 8) & 0x1;
- intlv_num_dies = (ctx.tmp >> 10) & 0x3;
- dram_limit_addr = ((ctx.tmp & GENMASK_ULL(31, 12)) << 16) | GENMASK_ULL(27, 0);
-
- intlv_addr_bit = intlv_addr_sel + 8;
-
- /* Re-use intlv_num_chan by setting it equal to log2(#channels) */
- switch (intlv_num_chan) {
- case 0: intlv_num_chan = 0; break;
- case 1: intlv_num_chan = 1; break;
- case 3: intlv_num_chan = 2; break;
- case 5: intlv_num_chan = 3; break;
- case 7: intlv_num_chan = 4; break;
-
- case 8: intlv_num_chan = 1;
- hash_enabled = true;
- break;
- default:
- pr_err("%s: Invalid number of interleaved channels %d.\n",
- __func__, intlv_num_chan);
- goto out_err;
- }
-
- num_intlv_bits = intlv_num_chan;
-
- if (intlv_num_dies > 2) {
- pr_err("%s: Invalid number of interleaved nodes/dies %d.\n",
- __func__, intlv_num_dies);
- goto out_err;
- }
-
- num_intlv_bits += intlv_num_dies;
-
- /* Add a bit if sockets are interleaved. */
- num_intlv_bits += intlv_num_sockets;
-
- /* Assert num_intlv_bits <= 4 */
- if (num_intlv_bits > 4) {
- pr_err("%s: Invalid interleave bits %d.\n",
- __func__, num_intlv_bits);
- goto out_err;
- }
-
- if (num_intlv_bits > 0) {
- u64 temp_addr_x, temp_addr_i, temp_addr_y;
- u8 die_id_bit, sock_id_bit, cs_fabric_id;
-
- /*
- * Read FabricBlockInstanceInformation3_CS[BlockFabricID].
- * This is the fabric id for this coherent slave. Use
- * umc/channel# as instance id of the coherent slave
- * for FICAA.
- */
- if (df_indirect_read_instance(nid, 0, 0x50, umc, &ctx.tmp))
- goto out_err;
-
- cs_fabric_id = (ctx.tmp >> 8) & 0xFF;
- die_id_bit = 0;
-
- /* If interleaved over more than 1 channel: */
- if (intlv_num_chan) {
- die_id_bit = intlv_num_chan;
- cs_mask = (1 << die_id_bit) - 1;
- cs_id = cs_fabric_id & cs_mask;
- }
-
- sock_id_bit = die_id_bit;
-
- /* Read D18F1x208 (SystemFabricIdMask). */
- if (intlv_num_dies || intlv_num_sockets)
- if (df_indirect_read_broadcast(nid, 1, 0x208, &ctx.tmp))
- goto out_err;
-
- /* If interleaved over more than 1 die. */
- if (intlv_num_dies) {
- sock_id_bit = die_id_bit + intlv_num_dies;
- die_id_shift = (ctx.tmp >> 24) & 0xF;
- die_id_mask = (ctx.tmp >> 8) & 0xFF;
-
- cs_id |= ((cs_fabric_id & die_id_mask) >> die_id_shift) << die_id_bit;
- }
-
- /* If interleaved over more than 1 socket. */
- if (intlv_num_sockets) {
- socket_id_shift = (ctx.tmp >> 28) & 0xF;
- socket_id_mask = (ctx.tmp >> 16) & 0xFF;
-
- cs_id |= ((cs_fabric_id & socket_id_mask) >> socket_id_shift) << sock_id_bit;
- }
-
- /*
- * The pre-interleaved address consists of XXXXXXIIIYYYYY
- * where III is the ID for this CS, and XXXXXXYYYYY are the
- * address bits from the post-interleaved address.
- * "num_intlv_bits" has been calculated to tell us how many "I"
- * bits there are. "intlv_addr_bit" tells us how many "Y" bits
- * there are (where "I" starts).
- */
- temp_addr_y = ctx.ret_addr & GENMASK_ULL(intlv_addr_bit - 1, 0);
- temp_addr_i = (cs_id << intlv_addr_bit);
- temp_addr_x = (ctx.ret_addr & GENMASK_ULL(63, intlv_addr_bit)) << num_intlv_bits;
- ctx.ret_addr = temp_addr_x | temp_addr_i | temp_addr_y;
- }
-
- /* Add dram base address */
- ctx.ret_addr += dram_base_addr;
-
- /* If legacy MMIO hole enabled */
- if (lgcy_mmio_hole_en) {
- if (df_indirect_read_broadcast(nid, 0, 0x104, &ctx.tmp))
- goto out_err;
-
- dram_hole_base = ctx.tmp & GENMASK(31, 24);
- if (ctx.ret_addr >= dram_hole_base)
- ctx.ret_addr += (BIT_ULL(32) - dram_hole_base);
- }
-
- if (hash_enabled) {
- /* Save some parentheses and grab ls-bit at the end. */
- hashed_bit = (ctx.ret_addr >> 12) ^
- (ctx.ret_addr >> 18) ^
- (ctx.ret_addr >> 21) ^
- (ctx.ret_addr >> 30) ^
- cs_id;
-
- hashed_bit &= BIT(0);
-
- if (hashed_bit != ((ctx.ret_addr >> intlv_addr_bit) & BIT(0)))
- ctx.ret_addr ^= BIT(intlv_addr_bit);
- }
-
- /* Is calculated system address is above DRAM limit address? */
- if (ctx.ret_addr > dram_limit_addr)
- goto out_err;
-
- *sys_addr = ctx.ret_addr;
- return 0;
-
-out_err:
- return -EINVAL;
-}
-
static int get_channel_from_ecc_syndrome(struct mem_ctl_info *, u16);
/*
@@ -3098,7 +2824,11 @@ static void decode_umc_error(int node_id, struct mce *m)
pvt->ops->get_err_info(m, &err);
- if (umc_normaddr_to_sysaddr(m->addr, pvt->mc_node_id, err.channel, &sys_addr)) {
+ if (!IS_REACHABLE(CONFIG_AMD_ATL)) {
+ amd64_warn("AMD Address Translation Library not available");
+ err.err_code = ERR_NORM_ADDR;
+ goto log_error;
+ } else if (amd_umc_mca_addr_to_sys_addr(m, &sys_addr)) {
err.err_code = ERR_NORM_ADDR;
goto log_error;
}
--
2.34.1
^ permalink raw reply related [flat|nested] 15+ messages in thread
* Re: [PATCH 1/2] platform/x86/amd: Introduce AMD Address Translation Library
2023-08-02 18:55 ` [PATCH 1/2] platform/x86/amd: Introduce " Yazen Ghannam
@ 2023-08-07 20:44 ` Yazen Ghannam
2023-08-08 3:17 ` Limonciello, Mario
0 siblings, 1 reply; 15+ messages in thread
From: Yazen Ghannam @ 2023-08-07 20:44 UTC (permalink / raw)
To: bp, linux-edac, hdegoede, markgross, platform-driver-x86, Luck, Tony
Cc: yazen.ghannam, linux-kernel, avadhut.naik, mario.limonciello
On 8/2/2023 2:55 PM, Yazen Ghannam wrote:
> AMD Zen-based systems report memory errors through Machine Check banks
> representing Unified Memory Controllers (UMCs). The address value
> reported for DRAM ECC errors is a "normalized address" that is relative
> to the UMC. This normalized address must be converted to a system
> physical address to be usable by the OS.
>
> Support for this address translation was introduced to the MCA subsystem
> with Zen1 systems. The code was later moved to the AMD64 EDAC module,
> since this was the only user of the code at the time.
>
> However, there are uses for this translation outside of EDAC. The system
> physical address can be used in MCA for preemptive page offlining as done
> in some MCA notifier functions. Also, this translation is needed as the
> basis of similar functionality needed for some CXL configurations on AMD
> systems.
>
> Introduce a common address translation library that can be used for
> multiple subsystems including MCA, EDAC, and CXL.
>
> Include support for UMC normalized to system physical address
> translation for current CPU systems.
>
> Future development to include:
> - DF4.5 Non-power-of-2 interleaving modes.
> - Heterogeneous CPU+GPU system support.
> - CXL translation support.
> - Caching of common intermediate values and results.
> - Leverage UEFI PRM methods as alternate backends to existing native
> code.
>
> Signed-off-by: Yazen Ghannam <yazen.ghannam@amd.com>
> ---
> MAINTAINERS | 7 +
> drivers/platform/x86/amd/Kconfig | 1 +
> drivers/platform/x86/amd/Makefile | 1 +
> drivers/platform/x86/amd/atl/Kconfig | 20 +
> drivers/platform/x86/amd/atl/Makefile | 18 +
> drivers/platform/x86/amd/atl/access.c | 107 ++++
> drivers/platform/x86/amd/atl/core.c | 212 +++++++
> drivers/platform/x86/amd/atl/dehash.c | 459 ++++++++++++++
> drivers/platform/x86/amd/atl/denormalize.c | 644 ++++++++++++++++++++
> drivers/platform/x86/amd/atl/internal.h | 307 ++++++++++
> drivers/platform/x86/amd/atl/map.c | 659 +++++++++++++++++++++
> drivers/platform/x86/amd/atl/reg_fields.h | 603 +++++++++++++++++++
> drivers/platform/x86/amd/atl/system.c | 282 +++++++++
> drivers/platform/x86/amd/atl/umc.c | 53 ++
> include/linux/amd-atl.h | 18 +
>
Hi all,
I'd like to get feedback on the most appropriate place for this code.
I want to move this out of EDAC, since it's not really an EDAC feature.
And it needs to be used by subsystems other than EDAC.
I thought x86 Platform Drivers, because the code is very
platform-specific. And there are already some AMD platform drivers. But
there isn't any platform control or management for this translation.
It's just reading registers and calculating values. So it's not really a
"platform driver" in the sense that it manages platform-specific behavior.
Another option is for this code to be in arch/x86/ras/. But I would like
the option for this code to be built as a module, at least for debug and
development. And I don't know that modules, nor platform-specific code,
should be in arch/.
Currently, I think this could go in drivers/ras/. This address
translation is needed for RAS use cases, so making it a part of "RAS
Infrastructure" may make the most sense.
Boris, Tony, (and others) what do you think?
Thanks,
Yazen
^ permalink raw reply [flat|nested] 15+ messages in thread
* Re: [PATCH 1/2] platform/x86/amd: Introduce AMD Address Translation Library
2023-08-07 20:44 ` Yazen Ghannam
@ 2023-08-08 3:17 ` Limonciello, Mario
2023-08-08 12:10 ` Borislav Petkov
0 siblings, 1 reply; 15+ messages in thread
From: Limonciello, Mario @ 2023-08-08 3:17 UTC (permalink / raw)
To: Yazen Ghannam, bp, linux-edac, hdegoede, markgross,
platform-driver-x86, Luck, Tony
Cc: linux-kernel, avadhut.naik
On 8/7/2023 3:44 PM, Yazen Ghannam wrote:
> On 8/2/2023 2:55 PM, Yazen Ghannam wrote:
>> AMD Zen-based systems report memory errors through Machine Check banks
>> representing Unified Memory Controllers (UMCs). The address value
>> reported for DRAM ECC errors is a "normalized address" that is relative
>> to the UMC. This normalized address must be converted to a system
>> physical address to be usable by the OS.
>>
>> Support for this address translation was introduced to the MCA subsystem
>> with Zen1 systems. The code was later moved to the AMD64 EDAC module,
>> since this was the only user of the code at the time.
>>
>> However, there are uses for this translation outside of EDAC. The system
>> physical address can be used in MCA for preemptive page offlining as done
>> in some MCA notifier functions. Also, this translation is needed as the
>> basis of similar functionality needed for some CXL configurations on AMD
>> systems.
>>
>> Introduce a common address translation library that can be used for
>> multiple subsystems including MCA, EDAC, and CXL.
>>
>> Include support for UMC normalized to system physical address
>> translation for current CPU systems.
>>
>> Future development to include:
>> - DF4.5 Non-power-of-2 interleaving modes.
>> - Heterogeneous CPU+GPU system support.
>> - CXL translation support.
>> - Caching of common intermediate values and results.
>> - Leverage UEFI PRM methods as alternate backends to existing native
>> code.
>>
>> Signed-off-by: Yazen Ghannam <yazen.ghannam@amd.com>
>> ---
>> MAINTAINERS | 7 +
>> drivers/platform/x86/amd/Kconfig | 1 +
>> drivers/platform/x86/amd/Makefile | 1 +
>> drivers/platform/x86/amd/atl/Kconfig | 20 +
>> drivers/platform/x86/amd/atl/Makefile | 18 +
>> drivers/platform/x86/amd/atl/access.c | 107 ++++
>> drivers/platform/x86/amd/atl/core.c | 212 +++++++
>> drivers/platform/x86/amd/atl/dehash.c | 459 ++++++++++++++
>> drivers/platform/x86/amd/atl/denormalize.c | 644 ++++++++++++++++++++
>> drivers/platform/x86/amd/atl/internal.h | 307 ++++++++++
>> drivers/platform/x86/amd/atl/map.c | 659 +++++++++++++++++++++
>> drivers/platform/x86/amd/atl/reg_fields.h | 603 +++++++++++++++++++
>> drivers/platform/x86/amd/atl/system.c | 282 +++++++++
>> drivers/platform/x86/amd/atl/umc.c | 53 ++
>> include/linux/amd-atl.h | 18 +
>>
>
> Hi all,
>
> I'd like to get feedback on the most appropriate place for this code.
>
> I want to move this out of EDAC, since it's not really an EDAC feature.
> And it needs to be used by subsystems other than EDAC.
>
> I thought x86 Platform Drivers, because the code is very
> platform-specific. And there are already some AMD platform drivers. But
> there isn't any platform control or management for this translation.
> It's just reading registers and calculating values. So it's not really a
> "platform driver" in the sense that it manages platform-specific behavior.
>
> Another option is for this code to be in arch/x86/ras/. But I would like
> the option for this code to be built as a module, at least for debug and
> development. And I don't know that modules, nor platform-specific code,
> should be in arch/.
>
> Currently, I think this could go in drivers/ras/. This address
> translation is needed for RAS use cases, so making it a part of "RAS
> Infrastructure" may make the most sense.
>
> Boris, Tony, (and others) what do you think?
Given it's 'library code' to be used by a bunch of things and also want
to be able to use a module, what about putting it in lib/? There's
plenty of library code there as tristate.
^ permalink raw reply [flat|nested] 15+ messages in thread
* Re: [PATCH 1/2] platform/x86/amd: Introduce AMD Address Translation Library
2023-08-08 3:17 ` Limonciello, Mario
@ 2023-08-08 12:10 ` Borislav Petkov
2023-08-08 14:07 ` Yazen Ghannam
0 siblings, 1 reply; 15+ messages in thread
From: Borislav Petkov @ 2023-08-08 12:10 UTC (permalink / raw)
To: Limonciello, Mario, Yazen Ghannam, linux-edac, hdegoede,
markgross, platform-driver-x86, Luck, Tony
Cc: linux-kernel, avadhut.naik
On August 8, 2023 5:17:33 AM GMT+02:00, "Limonciello, Mario" <mario.limonciello@amd.com> wrote:
>Given it's 'library code' to be used by a bunch of things and also want to be able to use a module, what about putting it in lib/? There's plenty of library code there as tristate.
It is x86-specific so not in there. Also, it might be used by multiple things so you want it as a separate "translation" service which is called by other modules.
Thx.
--
Sent from a small device: formatting sucks and brevity is inevitable.
^ permalink raw reply [flat|nested] 15+ messages in thread
* Re: [PATCH 1/2] platform/x86/amd: Introduce AMD Address Translation Library
2023-08-08 12:10 ` Borislav Petkov
@ 2023-08-08 14:07 ` Yazen Ghannam
2023-08-08 14:20 ` Borislav Petkov
2023-08-09 14:38 ` Hans de Goede
0 siblings, 2 replies; 15+ messages in thread
From: Yazen Ghannam @ 2023-08-08 14:07 UTC (permalink / raw)
To: Borislav Petkov, Limonciello, Mario, linux-edac, hdegoede,
markgross, platform-driver-x86, Luck, Tony
Cc: yazen.ghannam, linux-kernel, avadhut.naik
On 8/8/2023 8:10 AM, Borislav Petkov wrote:
> On August 8, 2023 5:17:33 AM GMT+02:00, "Limonciello, Mario" <mario.limonciello@amd.com> wrote:
>> Given it's 'library code' to be used by a bunch of things and also want to be able to use a module, what about putting it in lib/? There's plenty of library code there as tristate.
>
> It is x86-specific so not in there. Also, it might be used by multiple things so you want it as a separate "translation" service which is called by other modules.
>
There are modules in arch/x86, so I guess that's not an issue (not sure
what I was thinking).
This code is AMD-specific, and there's vendor-specific code in arch/x86
already. So it may be fine to have this in arch/x86/ras.
But just because there's existing vendor-specific code in arch/x86
doesn't necessarily mean we should keep adding to it. So is this okay?
Thanks,
Yazen
^ permalink raw reply [flat|nested] 15+ messages in thread
* Re: [PATCH 1/2] platform/x86/amd: Introduce AMD Address Translation Library
2023-08-08 14:07 ` Yazen Ghannam
@ 2023-08-08 14:20 ` Borislav Petkov
2023-08-08 14:28 ` Yazen Ghannam
2023-08-09 14:38 ` Hans de Goede
1 sibling, 1 reply; 15+ messages in thread
From: Borislav Petkov @ 2023-08-08 14:20 UTC (permalink / raw)
To: Yazen Ghannam
Cc: Limonciello, Mario, linux-edac, hdegoede, markgross,
platform-driver-x86, Luck, Tony, linux-kernel, avadhut.naik
On Tue, Aug 08, 2023 at 10:07:19AM -0400, Yazen Ghannam wrote:
> But just because there's existing vendor-specific code in arch/x86 doesn't
> necessarily mean we should keep adding to it. So is this okay?
I'd say since it is not so close to the arch/ part, drivers/ras/ is more
fitting IMO. Just like mce_amd is in drivers/edac/.
Which reminds me: why aren't you adding it to
drivers/edac/mce_amd.c
?
That is an MCE decoder *already*.
Thx.
--
Regards/Gruss,
Boris.
https://people.kernel.org/tglx/notes-about-netiquette
^ permalink raw reply [flat|nested] 15+ messages in thread
* Re: [PATCH 1/2] platform/x86/amd: Introduce AMD Address Translation Library
2023-08-08 14:20 ` Borislav Petkov
@ 2023-08-08 14:28 ` Yazen Ghannam
2023-08-08 14:37 ` Borislav Petkov
0 siblings, 1 reply; 15+ messages in thread
From: Yazen Ghannam @ 2023-08-08 14:28 UTC (permalink / raw)
To: Borislav Petkov
Cc: yazen.ghannam, Limonciello, Mario, linux-edac, hdegoede,
markgross, platform-driver-x86, Luck, Tony, linux-kernel,
avadhut.naik
On 8/8/2023 10:20 AM, Borislav Petkov wrote:
> On Tue, Aug 08, 2023 at 10:07:19AM -0400, Yazen Ghannam wrote:
>> But just because there's existing vendor-specific code in arch/x86 doesn't
>> necessarily mean we should keep adding to it. So is this okay?
>
> I'd say since it is not so close to the arch/ part, drivers/ras/ is more
> fitting IMO. Just like mce_amd is in drivers/edac/.
>
Okay, then I'll move this set to drivers/ras, if there are no objections
from others.
> Which reminds me: why aren't you adding it to
>
> drivers/edac/mce_amd.c
>
> ?
>
> That is an MCE decoder *already*.
>
Because this isn't intended to be only for MCA errors. The translation
code is related to the AMD Data Fabric. And it'll be a common back-end
for memory errors coming from MCA and CXL.
Thanks,
Yazen
^ permalink raw reply [flat|nested] 15+ messages in thread
* Re: [PATCH 1/2] platform/x86/amd: Introduce AMD Address Translation Library
2023-08-08 14:28 ` Yazen Ghannam
@ 2023-08-08 14:37 ` Borislav Petkov
2023-08-08 15:18 ` Yazen Ghannam
0 siblings, 1 reply; 15+ messages in thread
From: Borislav Petkov @ 2023-08-08 14:37 UTC (permalink / raw)
To: Yazen Ghannam
Cc: Limonciello, Mario, linux-edac, hdegoede, markgross,
platform-driver-x86, Luck, Tony, linux-kernel, avadhut.naik
On Tue, Aug 08, 2023 at 10:28:51AM -0400, Yazen Ghannam wrote:
> Because this isn't intended to be only for MCA errors. The translation code
> is related to the AMD Data Fabric. And it'll be a common back-end for memory
> errors coming from MCA and CXL.
But EDAC is not only about memory errors. Why not extend this into
something which does other RAS functionality instead of doing a second
one which is more or less related?
mce_amd is already loaded on the system, why add a second module if it
can be part of the first one just the same?
Strictly speaking, this all should've been drivers/ras/ from the very
beginning and all EDAC should move there but that's going to be madness
to do now.
Thx.
--
Regards/Gruss,
Boris.
https://people.kernel.org/tglx/notes-about-netiquette
^ permalink raw reply [flat|nested] 15+ messages in thread
* Re: [PATCH 1/2] platform/x86/amd: Introduce AMD Address Translation Library
2023-08-08 14:37 ` Borislav Petkov
@ 2023-08-08 15:18 ` Yazen Ghannam
2023-08-08 15:58 ` Borislav Petkov
0 siblings, 1 reply; 15+ messages in thread
From: Yazen Ghannam @ 2023-08-08 15:18 UTC (permalink / raw)
To: Borislav Petkov
Cc: yazen.ghannam, Limonciello, Mario, linux-edac, hdegoede,
markgross, platform-driver-x86, Luck, Tony, linux-kernel,
avadhut.naik
On 8/8/2023 10:37 AM, Borislav Petkov wrote:
> On Tue, Aug 08, 2023 at 10:28:51AM -0400, Yazen Ghannam wrote:
>> Because this isn't intended to be only for MCA errors. The translation code
>> is related to the AMD Data Fabric. And it'll be a common back-end for memory
>> errors coming from MCA and CXL.
>
> But EDAC is not only about memory errors. Why not extend this into
> something which does other RAS functionality instead of doing a second
> one which is more or less related?
>
> mce_amd is already loaded on the system, why add a second module if it
> can be part of the first one just the same?
>
I think it would be better to avoid dependencies between independent things.
For example, amd_smn_read() is mostly used in amd64_edac. EDAC was the
original user of SMN accesses, and all the SMN stuff could have been
included in EDAC. However, SMN is not specifically for EDAC, so it was
added to amd_nb.c to be commonly available. Currently, SMN accesses are
done in other modules. I don't think it would have been a good idea to
force other modules or subsystems to require EDAC to be used.
This is my reasoning for a separate, independent module for the
translation. EDAC is the first user of this. But there will be future
code that can leverage this, like CXL, and even the MCE subsystem. And,
yes, mce_amd may be already loaded, but this isn't a given. A person may
want MCE and CXL support without wanting to use EDAC.
Furthermore, some things using the translation will be built-in, so the
translation module will need to be built-in. And it seems unnecessary to
require all of mce_amd to be built-in just for the translation part.
> Strictly speaking, this all should've been drivers/ras/ from the very
> beginning and all EDAC should move there but that's going to be madness
> to do now.
>
I agree. And I don't think much of the existing things in EDAC should be
moved out. But this is new code, so there's an opportunity to have it in
a more appropriate place.
And, thinking on it more, this could be another example for future
"common RAS" functionality. Isn't that why the CEC is in drivers/ras? It
seems like things go into EDAC because it's thought of as the de facto
RAS location. But why have something in EDAC if it doesn't provide EDAC
functionality? Other RAS things, like AER, APEI, etc., don't live in EDAC.
Thanks,
Yazen
^ permalink raw reply [flat|nested] 15+ messages in thread
* Re: [PATCH 1/2] platform/x86/amd: Introduce AMD Address Translation Library
2023-08-08 15:18 ` Yazen Ghannam
@ 2023-08-08 15:58 ` Borislav Petkov
2023-08-08 16:24 ` Yazen Ghannam
0 siblings, 1 reply; 15+ messages in thread
From: Borislav Petkov @ 2023-08-08 15:58 UTC (permalink / raw)
To: Yazen Ghannam
Cc: Limonciello, Mario, linux-edac, hdegoede, markgross,
platform-driver-x86, Luck, Tony, linux-kernel, avadhut.naik
On Tue, Aug 08, 2023 at 11:18:07AM -0400, Yazen Ghannam wrote:
> I think it would be better to avoid dependencies between independent things.
If they really are independent then I guess. Not that it all ends up in
a twisty dependency where you wish you should've merged the two
together. So think about all deps before you design this - it needs to
handle all cases without hackery.
> For example, amd_smn_read() is mostly used in amd64_edac. EDAC was the
> original user of SMN accesses, and all the SMN stuff could have been
> included in EDAC. However, SMN is not specifically for EDAC, so it was added
> to amd_nb.c to be commonly available. Currently, SMN accesses are done in
> other modules. I don't think it would have been a good idea to force other
> modules or subsystems to require EDAC to be used.
What does that have to do with this? SMN access is generic and should be
in amd_nb.c as it is needed by other stuff. EDAC, RAS, whatever are all
users of that thing.
> This is my reasoning for a separate, independent module for the translation.
> EDAC is the first user of this. But there will be future code that can
> leverage this, like CXL, and even the MCE subsystem. And, yes, mce_amd may
> be already loaded, but this isn't a given. A person may want MCE and CXL
> support without wanting to use EDAC.
Is that a real use case or just a hypothetical thing?
> Furthermore, some things using the translation will be built-in, so the
> translation module will need to be built-in.
This sounds weird.
> I agree. And I don't think much of the existing things in EDAC should be
> moved out. But this is new code, so there's an opportunity to have it in a
> more appropriate place.
>
> And, thinking on it more, this could be another example for future "common
> RAS" functionality. Isn't that why the CEC is in drivers/ras?
It is there because it doesn't need EDAC at all. If your translation
doesn't need EDAC and EDAC is going to be only a user of it, then good.
But if you're going to have to need the MCA error decoded by EDAC and
then the error translation done by this thing, then you'd need to
synchronize between the two. I'm not saying it is impossible - it should
be well thought out first though before you go coding.
> It seems like things go into EDAC because it's thought of as the de
> facto RAS location. But why have something in EDAC if it doesn't
> provide EDAC functionality? Other RAS things, like AER, APEI, etc.,
> don't live in EDAC.
AER is part of PCI so we haven't considereed tying it into EDAC. And
there wasn't any desire to do so.
As to APEI, there's ghes_edac...
--
Regards/Gruss,
Boris.
https://people.kernel.org/tglx/notes-about-netiquette
^ permalink raw reply [flat|nested] 15+ messages in thread
* Re: [PATCH 1/2] platform/x86/amd: Introduce AMD Address Translation Library
2023-08-08 15:58 ` Borislav Petkov
@ 2023-08-08 16:24 ` Yazen Ghannam
0 siblings, 0 replies; 15+ messages in thread
From: Yazen Ghannam @ 2023-08-08 16:24 UTC (permalink / raw)
To: Borislav Petkov
Cc: yazen.ghannam, Limonciello, Mario, linux-edac, hdegoede,
markgross, platform-driver-x86, Luck, Tony, linux-kernel,
avadhut.naik
On 8/8/2023 11:58 AM, Borislav Petkov wrote:
> On Tue, Aug 08, 2023 at 11:18:07AM -0400, Yazen Ghannam wrote:
>> I think it would be better to avoid dependencies between independent things.
>
> If they really are independent then I guess. Not that it all ends up in
> a twisty dependency where you wish you should've merged the two
> together. So think about all deps before you design this - it needs to
> handle all cases without hackery.
>
I agree. My goal with this is to avoid the hackery. :)
I guess we'll see how it goes...
>> For example, amd_smn_read() is mostly used in amd64_edac. EDAC was the
>> original user of SMN accesses, and all the SMN stuff could have been
>> included in EDAC. However, SMN is not specifically for EDAC, so it was added
>> to amd_nb.c to be commonly available. Currently, SMN accesses are done in
>> other modules. I don't think it would have been a good idea to force other
>> modules or subsystems to require EDAC to be used.
>
> What does that have to do with this? SMN access is generic and should be
> in amd_nb.c as it is needed by other stuff. EDAC, RAS, whatever are all
> users of that thing.
>
Right, but that's my point. The translation is "generic" and tied to the
Data Fabric which is the central "thing" in an modern AMD SoC. Anything
that needs memory address translation for the fabric will use it.
>> This is my reasoning for a separate, independent module for the translation.
>> EDAC is the first user of this. But there will be future code that can
>> leverage this, like CXL, and even the MCE subsystem. And, yes, mce_amd may
>> be already loaded, but this isn't a given. A person may want MCE and CXL
>> support without wanting to use EDAC.
>
> Is that a real use case or just a hypothetical thing?
>
Real. There are actually two use cases for CXL. First is memory error
reporting and page offline which is analogous to MCA. Second is general
memory online/offline support which will be used for hot-plug/hot-swap
cases.
In actuality, both CXL cases need the same functionality. Take an AMD
Data Fabric "normalized" address and translate it to a system physical
address for the OS to take action.
>> Furthermore, some things using the translation will be built-in, so the
>> translation module will need to be built-in.
>
> This sounds weird.
>
Yes, and this is my intention for using the "imply" Kconfig thing you
see in the second patch. Any config options that need this code will
"imply" it. And the default option for this code will take the strictest
setting of all the options that "imply" it.
>> I agree. And I don't think much of the existing things in EDAC should be
>> moved out. But this is new code, so there's an opportunity to have it in a
>> more appropriate place.
>>
>> And, thinking on it more, this could be another example for future "common
>> RAS" functionality. Isn't that why the CEC is in drivers/ras?
>
> It is there because it doesn't need EDAC at all. If your translation
> doesn't need EDAC and EDAC is going to be only a user of it, then good.
>
Yep, that's the intent.
> But if you're going to have to need the MCA error decoded by EDAC and
> then the error translation done by this thing, then you'd need to
> synchronize between the two. I'm not saying it is impossible - it should
> be well thought out first though before you go coding.
>
Yes, I agree. That's the reason for this code to take a raw struct mce,
and do the translation independent of EDAC, etc. (for the MCA case).
The same will be the goal for CXL. This code will take a raw struct
pci_dev, or whatever, and do the translation independent of other CXL code.
>> It seems like things go into EDAC because it's thought of as the de
>> facto RAS location. But why have something in EDAC if it doesn't
>> provide EDAC functionality? Other RAS things, like AER, APEI, etc.,
>> don't live in EDAC.
>
> AER is part of PCI so we haven't considereed tying it into EDAC. And
> there wasn't any desire to do so.
>
> As to APEI, there's ghes_edac...
>
True. Though the intent for ghes_edac is for it to provide the EDAC
functionality.
The other features that don't need EDAC functionality aren't in EDAC.
Going back to the dependency concerns... there are a lot of
inter-dependencies between RAS code in various subsystems. Maybe some of
these can be streamlined by moving common things to drivers/ras? But I
guess this can be a discussion for a later time...
Thanks,
Yazen
^ permalink raw reply [flat|nested] 15+ messages in thread
* Re: [PATCH 1/2] platform/x86/amd: Introduce AMD Address Translation Library
2023-08-08 14:07 ` Yazen Ghannam
2023-08-08 14:20 ` Borislav Petkov
@ 2023-08-09 14:38 ` Hans de Goede
2023-08-09 15:05 ` Yazen Ghannam
1 sibling, 1 reply; 15+ messages in thread
From: Hans de Goede @ 2023-08-09 14:38 UTC (permalink / raw)
To: Yazen Ghannam, Borislav Petkov, Limonciello, Mario, linux-edac,
markgross, platform-driver-x86, Luck, Tony
Cc: linux-kernel, avadhut.naik
Hi,
On 8/8/23 16:07, Yazen Ghannam wrote:
> On 8/8/2023 8:10 AM, Borislav Petkov wrote:
>> On August 8, 2023 5:17:33 AM GMT+02:00, "Limonciello, Mario" <mario.limonciello@amd.com> wrote:
>>> Given it's 'library code' to be used by a bunch of things and also want to be able to use a module, what about putting it in lib/? There's plenty of library code there as tristate.
>>
>> It is x86-specific so not in there. Also, it might be used by multiple things so you want it as a separate "translation" service which is called by other modules.
>>
>
> There are modules in arch/x86, so I guess that's not an issue (not sure what I was thinking).
Right, I think arch/x86 makes sense.
As for putting this under drivers/platform/x86/amd as initially suggested. That means that any code which wants to select this because it needs the functions also must select X86_PLATFORM_DEVICES, so IMHO it is better to put it in another place.
Regards,
Hans
^ permalink raw reply [flat|nested] 15+ messages in thread
* Re: [PATCH 1/2] platform/x86/amd: Introduce AMD Address Translation Library
2023-08-09 14:38 ` Hans de Goede
@ 2023-08-09 15:05 ` Yazen Ghannam
0 siblings, 0 replies; 15+ messages in thread
From: Yazen Ghannam @ 2023-08-09 15:05 UTC (permalink / raw)
To: Hans de Goede, Borislav Petkov, Limonciello, Mario, linux-edac,
markgross, platform-driver-x86, Luck, Tony
Cc: yazen.ghannam, linux-kernel, avadhut.naik
On 8/9/2023 10:38 AM, Hans de Goede wrote:
> Hi,
>
> On 8/8/23 16:07, Yazen Ghannam wrote:
>> On 8/8/2023 8:10 AM, Borislav Petkov wrote:
>>> On August 8, 2023 5:17:33 AM GMT+02:00, "Limonciello, Mario" <mario.limonciello@amd.com> wrote:
>>>> Given it's 'library code' to be used by a bunch of things and also want to be able to use a module, what about putting it in lib/? There's plenty of library code there as tristate.
>>>
>>> It is x86-specific so not in there. Also, it might be used by multiple things so you want it as a separate "translation" service which is called by other modules.
>>>
>>
>> There are modules in arch/x86, so I guess that's not an issue (not sure what I was thinking).
>
> Right, I think arch/x86 makes sense.
>
> As for putting this under drivers/platform/x86/amd as initially suggested. That means that any code which wants to select this because it needs the functions also must select X86_PLATFORM_DEVICES, so IMHO it is better to put it in another place.
>
Good point, and thanks for your feedback!
Thanks,
Yazen
^ permalink raw reply [flat|nested] 15+ messages in thread
end of thread, other threads:[~2023-08-09 15:06 UTC | newest]
Thread overview: 15+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2023-08-02 18:55 [PATCH 0/2] AMD Address Translation Library Yazen Ghannam
2023-08-02 18:55 ` [PATCH 1/2] platform/x86/amd: Introduce " Yazen Ghannam
2023-08-07 20:44 ` Yazen Ghannam
2023-08-08 3:17 ` Limonciello, Mario
2023-08-08 12:10 ` Borislav Petkov
2023-08-08 14:07 ` Yazen Ghannam
2023-08-08 14:20 ` Borislav Petkov
2023-08-08 14:28 ` Yazen Ghannam
2023-08-08 14:37 ` Borislav Petkov
2023-08-08 15:18 ` Yazen Ghannam
2023-08-08 15:58 ` Borislav Petkov
2023-08-08 16:24 ` Yazen Ghannam
2023-08-09 14:38 ` Hans de Goede
2023-08-09 15:05 ` Yazen Ghannam
2023-08-02 18:55 ` [PATCH 2/2] EDAC/amd64: Use new " Yazen Ghannam
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).