Skip to contentSkip to conversation

hw/riscv: microchip_pfsoc: Correct DDR memory map

PATCH Superseded Bin Meng 2 participants 4 messages Nov 1, 2020 RISC-V

v3 of 4 Superseded Latest is v4 →

lore
Share & tools
lore

Bin Meng

author
#0 lore raw

PATCH v3 1/1 hw/riscv: microchip_pfsoc: Correct DDR memory map

From: Bin Meng <[email protected]>

When system memory is larger than 1 GiB (high memory), PolarFire SoC maps it at address 0x10_0000_0000. Address 0xC000_0000 and above is aliased to the same 1 GiB low memory with different cache attributes.

At present QEMU maps the system memory contiguously from 0x8000_0000. This corrects the wrong QEMU logic. Note address 0x14_0000_0000 is the alias to the high memory, and even physical memory is only 1 GiB, the HSS codes still tries to probe the high memory alias address. It seems there is no issue on the real hardware, so we will have to take that into the consideration in our emulation. Due to this, we we increase the default system memory size to 2047 MiB (the largest ram size allowed when running on a 32-bit host) so that user gets notified an error when less than 2047 MiB is specified.

---
This patch should replace the following commit in Alistair's riscv-to-apply.next tree:
https://github.com/alistair23/qemu/commit/8c47c1e9df850a928b4b230240a950feabe6152f

Changes in v3:
- Change default ram size to 2047 MiB for 32-bit host

 hw/riscv/microchip_pfsoc.c         | 48 ++++++++++++++++++++++++++----
 include/hw/riscv/microchip_pfsoc.h |  5 +++-
 2 files changed, 46 insertions(+), 7 deletions(-)
hw/riscv/microchip_pfsoc.c +42 -6
index 44a84732ac..0bc17b3955 100644
@@ -121,7 +121,10 @@ static const struct MemmapEntry {
     [MICROCHIP_PFSOC_ENVM_CFG] =        { 0x20200000,     0x1000 },
     [MICROCHIP_PFSOC_ENVM_DATA] =       { 0x20220000,    0x20000 },
     [MICROCHIP_PFSOC_IOSCB] =           { 0x30000000, 0x10000000 },
-    [MICROCHIP_PFSOC_DRAM] =            { 0x80000000,        0x0 },
+    [MICROCHIP_PFSOC_DRAM_LO] =         { 0x80000000, 0x40000000 },
+    [MICROCHIP_PFSOC_DRAM_LO_ALIAS] =   { 0xc0000000, 0x40000000 },
+    [MICROCHIP_PFSOC_DRAM_HI] =       { 0x1000000000,        0x0 },
+    [MICROCHIP_PFSOC_DRAM_HI_ALIAS] = { 0x1400000000,        0x0 },
 };
 
 static void microchip_pfsoc_soc_instance_init(Object *obj)
@@ -424,7 +427,11 @@ static void microchip_icicle_kit_machine_init(MachineState *machine)
     const struct MemmapEntry *memmap = microchip_pfsoc_memmap;
     MicrochipIcicleKitState *s = MICROCHIP_ICICLE_KIT_MACHINE(machine);
     MemoryRegion *system_memory = get_system_memory();
-    MemoryRegion *main_mem = g_new(MemoryRegion, 1);
+    MemoryRegion *mem_low = g_new(MemoryRegion, 1);
+    MemoryRegion *mem_low_alias = g_new(MemoryRegion, 1);
+    MemoryRegion *mem_high = g_new(MemoryRegion, 1);
+    MemoryRegion *mem_high_alias = g_new(MemoryRegion, 1);
+    uint64_t mem_high_size;
     DriveInfo *dinfo = drive_get_next(IF_SD);
 
     /* Sanity check on RAM size */
@@ -441,10 +448,39 @@ static void microchip_icicle_kit_machine_init(MachineState *machine)
     qdev_realize(DEVICE(&s->soc), NULL, &error_abort);
 
     /* Register RAM */
-    memory_region_init_ram(main_mem, NULL, "microchip.icicle.kit.ram",
-                           machine->ram_size, &error_fatal);
+    memory_region_init_ram(mem_low, NULL, "microchip.icicle.kit.ram_low",
+                           memmap[MICROCHIP_PFSOC_DRAM_LO].size,
+                           &error_fatal);
+    memory_region_init_alias(mem_low_alias, NULL,
+                             "microchip.icicle.kit.ram_low.alias",
+                             mem_low, 0,
+                             memmap[MICROCHIP_PFSOC_DRAM_LO_ALIAS].size);
+    memory_region_add_subregion(system_memory,
+                                memmap[MICROCHIP_PFSOC_DRAM_LO].base,
+                                mem_low);
+    memory_region_add_subregion(system_memory,
+                                memmap[MICROCHIP_PFSOC_DRAM_LO_ALIAS].base,
+                                mem_low_alias);
+
+    /*
+     * Map 1 GiB high memory because HSS will do memory test against the high
+     * memory address range regardless of physical memory installed.
+     *
+     * See memory_tests() in mss_ddr.c in the HSS source code.
+     */
+    mem_high_size = machine->ram_size - 1 * GiB;
+
+    memory_region_init_ram(mem_high, NULL, "microchip.icicle.kit.ram_high",
+                           mem_high_size, &error_fatal);
+    memory_region_init_alias(mem_high_alias, NULL,
+                             "microchip.icicle.kit.ram_high.alias",
+                             mem_high, 0, mem_high_size);
+    memory_region_add_subregion(system_memory,
+                                memmap[MICROCHIP_PFSOC_DRAM_HI].base,
+                                mem_high);
     memory_region_add_subregion(system_memory,
-                                memmap[MICROCHIP_PFSOC_DRAM].base, main_mem);
+                                memmap[MICROCHIP_PFSOC_DRAM_HI_ALIAS].base,
+                                mem_high_alias);
 
     /* Load the firmware */
     riscv_find_and_load_firmware(machine, BIOS_FILENAME, RESET_VECTOR, NULL);
@@ -470,7 +506,7 @@ static void microchip_icicle_kit_machine_class_init(ObjectClass *oc, void *data)
                    MICROCHIP_PFSOC_COMPUTE_CPU_COUNT;
     mc->min_cpus = MICROCHIP_PFSOC_MANAGEMENT_CPU_COUNT + 1;
     mc->default_cpus = mc->min_cpus;
-    mc->default_ram_size = 1 * GiB;
+    mc->default_ram_size = 2047 * MiB;
 }
 
 static const TypeInfo microchip_icicle_kit_machine_typeinfo = {
include/hw/riscv/microchip_pfsoc.h +4 -1
index f34a6b3fd7..db77e9c84a 100644
@@ -105,7 +105,10 @@ enum {
     MICROCHIP_PFSOC_ENVM_CFG,
     MICROCHIP_PFSOC_ENVM_DATA,
     MICROCHIP_PFSOC_IOSCB,
-    MICROCHIP_PFSOC_DRAM,
+    MICROCHIP_PFSOC_DRAM_LO,
+    MICROCHIP_PFSOC_DRAM_LO_ALIAS,
+    MICROCHIP_PFSOC_DRAM_HI,
+    MICROCHIP_PFSOC_DRAM_HI_ALIAS
 };
 
 enum {
Signature · 1 line
2.25.1
#1 lore raw
Quoted 7 lines · Bin Meng

From: Bin Meng <[email protected]>

When system memory is larger than 1 GiB (high memory), PolarFire SoC maps it at address 0x10_0000_0000. Address 0xC000_0000 and above is aliased to the same 1 GiB low memory with different cache attributes.

Quoted 9 lines

At present QEMU maps the system memory contiguously from 0x8000_0000. This corrects the wrong QEMU logic. Note address 0x14_0000_0000 is the alias to the high memory, and even physical memory is only 1 GiB, the HSS codes still tries to probe the high memory alias address. It seems there is no issue on the real hardware, so we will have to take that into the consideration in our emulation. Due to this, we we increase the default system memory size to 2047 MiB (the largest ram size allowed when running on a 32-bit host) so that user gets notified an error when less than 2047 MiB is specified.

Is this better than just not supporting 32-bit hosts? Or could we make this number even lower (as low as possible that still works with HSS)?

Alistair

Quoted 116 lines

---
This patch should replace the following commit in Alistair's riscv-to-apply.next tree:
https://github.com/alistair23/qemu/commit/8c47c1e9df850a928b4b230240a950feabe6152f

Changes in v3:
- Change default ram size to 2047 MiB for 32-bit host

 hw/riscv/microchip_pfsoc.c         | 48 ++++++++++++++++++++++++++----
 include/hw/riscv/microchip_pfsoc.h |  5 +++-
 2 files changed, 46 insertions(+), 7 deletions(-)
hw/riscv/microchip_pfsoc.c +42 -6
index 44a84732ac..0bc17b3955 100644
@@ -121,7 +121,10 @@ static const struct MemmapEntry {
     [MICROCHIP_PFSOC_ENVM_CFG] =        { 0x20200000,     0x1000 },
     [MICROCHIP_PFSOC_ENVM_DATA] =       { 0x20220000,    0x20000 },
     [MICROCHIP_PFSOC_IOSCB] =           { 0x30000000, 0x10000000 },
-    [MICROCHIP_PFSOC_DRAM] =            { 0x80000000,        0x0 },
+    [MICROCHIP_PFSOC_DRAM_LO] =         { 0x80000000, 0x40000000 },
+    [MICROCHIP_PFSOC_DRAM_LO_ALIAS] =   { 0xc0000000, 0x40000000 },
+    [MICROCHIP_PFSOC_DRAM_HI] =       { 0x1000000000,        0x0 },
+    [MICROCHIP_PFSOC_DRAM_HI_ALIAS] = { 0x1400000000,        0x0 },
 };
 
 static void microchip_pfsoc_soc_instance_init(Object *obj)
@@ -424,7 +427,11 @@ static void microchip_icicle_kit_machine_init(MachineState *machine)
     const struct MemmapEntry *memmap = microchip_pfsoc_memmap;
     MicrochipIcicleKitState *s = MICROCHIP_ICICLE_KIT_MACHINE(machine);
     MemoryRegion *system_memory = get_system_memory();
-    MemoryRegion *main_mem = g_new(MemoryRegion, 1);
+    MemoryRegion *mem_low = g_new(MemoryRegion, 1);
+    MemoryRegion *mem_low_alias = g_new(MemoryRegion, 1);
+    MemoryRegion *mem_high = g_new(MemoryRegion, 1);
+    MemoryRegion *mem_high_alias = g_new(MemoryRegion, 1);
+    uint64_t mem_high_size;
     DriveInfo *dinfo = drive_get_next(IF_SD);
 
     /* Sanity check on RAM size */
@@ -441,10 +448,39 @@ static void microchip_icicle_kit_machine_init(MachineState *machine)
     qdev_realize(DEVICE(&s->soc), NULL, &error_abort);
 
     /* Register RAM */
-    memory_region_init_ram(main_mem, NULL, "microchip.icicle.kit.ram",
-                           machine->ram_size, &error_fatal);
+    memory_region_init_ram(mem_low, NULL, "microchip.icicle.kit.ram_low",
+                           memmap[MICROCHIP_PFSOC_DRAM_LO].size,
+                           &error_fatal);
+    memory_region_init_alias(mem_low_alias, NULL,
+                             "microchip.icicle.kit.ram_low.alias",
+                             mem_low, 0,
+                             memmap[MICROCHIP_PFSOC_DRAM_LO_ALIAS].size);
+    memory_region_add_subregion(system_memory,
+                                memmap[MICROCHIP_PFSOC_DRAM_LO].base,
+                                mem_low);
+    memory_region_add_subregion(system_memory,
+                                memmap[MICROCHIP_PFSOC_DRAM_LO_ALIAS].base,
+                                mem_low_alias);
+
+    /*
+     * Map 1 GiB high memory because HSS will do memory test against the high
+     * memory address range regardless of physical memory installed.
+     *
+     * See memory_tests() in mss_ddr.c in the HSS source code.
+     */
+    mem_high_size = machine->ram_size - 1 * GiB;
+
+    memory_region_init_ram(mem_high, NULL, "microchip.icicle.kit.ram_high",
+                           mem_high_size, &error_fatal);
+    memory_region_init_alias(mem_high_alias, NULL,
+                             "microchip.icicle.kit.ram_high.alias",
+                             mem_high, 0, mem_high_size);
+    memory_region_add_subregion(system_memory,
+                                memmap[MICROCHIP_PFSOC_DRAM_HI].base,
+                                mem_high);
     memory_region_add_subregion(system_memory,
-                                memmap[MICROCHIP_PFSOC_DRAM].base, main_mem);
+                                memmap[MICROCHIP_PFSOC_DRAM_HI_ALIAS].base,
+                                mem_high_alias);
 
     /* Load the firmware */
     riscv_find_and_load_firmware(machine, BIOS_FILENAME, RESET_VECTOR, NULL);
@@ -470,7 +506,7 @@ static void microchip_icicle_kit_machine_class_init(ObjectClass *oc, void *data)
                    MICROCHIP_PFSOC_COMPUTE_CPU_COUNT;
     mc->min_cpus = MICROCHIP_PFSOC_MANAGEMENT_CPU_COUNT + 1;
     mc->default_cpus = mc->min_cpus;
-    mc->default_ram_size = 1 * GiB;
+    mc->default_ram_size = 2047 * MiB;
 }
 
 static const TypeInfo microchip_icicle_kit_machine_typeinfo = {
include/hw/riscv/microchip_pfsoc.h +4 -2
index f34a6b3fd7..db77e9c84a 100644
@@ -105,7 +105,10 @@ enum {
     MICROCHIP_PFSOC_ENVM_CFG,
     MICROCHIP_PFSOC_ENVM_DATA,
     MICROCHIP_PFSOC_IOSCB,
-    MICROCHIP_PFSOC_DRAM,
+    MICROCHIP_PFSOC_DRAM_LO,
+    MICROCHIP_PFSOC_DRAM_LO_ALIAS,
+    MICROCHIP_PFSOC_DRAM_HI,
+    MICROCHIP_PFSOC_DRAM_HI_ALIAS
 };
 
 enum {
--

2.25.1

author
#2 lore raw

Hi Alistair,

Quoted 19 lines · Alistair Francis
Quoted 7 lines · Bin Meng

From: Bin Meng <[email protected]>

When system memory is larger than 1 GiB (high memory), PolarFire SoC maps it at address 0x10_0000_0000. Address 0xC000_0000 and above is aliased to the same 1 GiB low memory with different cache attributes.

Quoted 9 lines

At present QEMU maps the system memory contiguously from 0x8000_0000. This corrects the wrong QEMU logic. Note address 0x14_0000_0000 is the alias to the high memory, and even physical memory is only 1 GiB, the HSS codes still tries to probe the high memory alias address. It seems there is no issue on the real hardware, so we will have to take that into the consideration in our emulation. Due to this, we we increase the default system memory size to 2047 MiB (the largest ram size allowed when running on a 32-bit host) so that user gets notified an error when less than 2047 MiB is specified.

Quoted 1 line

Is this better than just not supporting 32-bit hosts? Or could we make

I am not sure if we have a general rule about discontinuing 32-bit hosts support, i.e.: deprecating 32-bit hosts at some time?

Quoted 2 lines

this number even lower (as low as possible that still works with HSS)?

Sure I will figure this out and set this number to meet the minium requirement of HSS.

Regards,
Bin

#3 lore raw
Quoted 25 lines · Bin Meng

Hi Alistair,

Quoted 19 lines · Alistair Francis
Quoted 7 lines · Bin Meng

From: Bin Meng <[email protected]>

When system memory is larger than 1 GiB (high memory), PolarFire SoC maps it at address 0x10_0000_0000. Address 0xC000_0000 and above is aliased to the same 1 GiB low memory with different cache attributes.

Quoted 9 lines

At present QEMU maps the system memory contiguously from 0x8000_0000. This corrects the wrong QEMU logic. Note address 0x14_0000_0000 is the alias to the high memory, and even physical memory is only 1 GiB, the HSS codes still tries to probe the high memory alias address. It seems there is no issue on the real hardware, so we will have to take that into the consideration in our emulation. Due to this, we we increase the default system memory size to 2047 MiB (the largest ram size allowed when running on a 32-bit host) so that user gets notified an error when less than 2047 MiB is specified.

Quoted 1 line

Is this better than just not supporting 32-bit hosts? Or could we make

Quoted 8 lines

I am not sure if we have a general rule about discontinuing 32-bit hosts support, i.e.: deprecating 32-bit hosts at some time?

Quoted 2 lines

this number even lower (as low as possible that still works with HSS)?

Sure I will figure this out and set this number to meet the minium requirement of HSS.

Thanks, that's probably the best bet for both 32 and 64-bit hosts then.

Alistair

Quoted 3 lines

Regards,
Bin