Skip to contentSkip to conversation

Fixes for running U-boot under QEMU/KVM

PATCH Superseded Ard Biesheuvel 3 participants 19 messages Jun 6, 2020 – Jun 8, 2020 Arm (32-bit)

v1 of 3 Superseded Latest is v3 →

lore
Share & tools
lore
#0 lore raw

cover Fixes for running U-boot under QEMU/KVM

This series fixes a number of issues that exist in the QEMU/mach-virt port of u-boot, and that prevent it from executing correctly under virtualization (as opposed to TCG emulation)

As the Linux EFI subsystem maintainer, I am looking to increase test coverage for the EFI related changes that are under development for Linux, and one of the things I plan to do is start using U-boot as test firmware for boot testing. This can be done under TCG emulation, but given how loosely TCG implements the architecture, it is better to test under virtualization as well.

With these changes applied, u-boot can boot Linux in EFI mode under KVM.

Cc Tom Rini <[email protected]>
Cc Sughosh Ganu <[email protected]>
Cc Heinrich Schuchardt <[email protected]>

Ard Biesheuvel (5):
arm: enable allocate-on-read for LPAE's DCACHE_WRITEBACK
arm: qemu: enable LPAE on 32-bit
arm: qemu: implement enable_caches()
arm: qemu: disable the EFI workaround for older GRUB
arm: qemu: override flash accessors to use virtualizable instructions

 arch/arm/include/asm/system.h       |  2 +-
 board/emulation/qemu-arm/qemu-arm.c | 62 ++++++++++++++++++++
 configs/qemu_arm_defconfig          |  2 +
 include/configs/qemu-arm.h          |  1 +
 4 files changed, 66 insertions(+), 1 deletion(-)
Signature · 1 line
2.26.2
#1 lore raw

PATCH 1/5 arm: enable allocate-on-read for LPAE's DCACHE_WRITEBACK

The LPAE version of DCACHE_WRITEBACK is currently defined as no-allocate for both reads and writes, which deviates from the non-LPAE definition, and mostly defeats the purpose of enabling the caches in the first place.

So align LPAE with !LPAE, and enable allocate-on-read.

---
 arch/arm/include/asm/system.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
arch/arm/include/asm/system.h +1 -1
index 7a40b56acdca..21b26557d28b 100644
@@ -445,7 +445,7 @@ static inline void set_dacr(unsigned int val)
  * Memory types
  */
 #define MEMORY_ATTRIBUTES	((0x00 << (0 * 8)) | (0x88 << (1 * 8)) | \
-				 (0xcc << (2 * 8)) | (0xff << (3 * 8)))
+				 (0xee << (2 * 8)) | (0xff << (3 * 8)))
 
 /* options available for data cache on each page */
 enum dcache_option {
Signature · 1 line
2.26.2
#5 lore raw

PATCH 2/5 arm: qemu: enable LPAE on 32-bit

In reply to Ard Biesheuvel

QEMU's mach-virt machine only supports selecting CPU models that implement the virtualization extensions, and are therefore guaranteed to support LPAE as well.

Initially, QEMU would not allow emulating these CPUs running in HYP mode (or EL2, for AArch64), but today, it also contains a complete implementation of the virtualization extensions themselves.

This means we could be running U-Boot in HYP mode, in which case the LPAE long descriptor page table format is the only format that is supported. If we are not running in HYP mode, we can use either.

So let's enable CONFIG_ARMV7_LPAE for qemu_arm_defconfig so that we get the best support for running with the MMU and caches enabled at any privilege level.

---
 configs/qemu_arm_defconfig | 1 +
 1 file changed, 1 insertion(+)
configs/qemu_arm_defconfig +1 -0
index a8473988bd76..75bdce7708c7 100644
@@ -1,5 +1,6 @@
 CONFIG_ARM=y
 CONFIG_ARM_SMCCC=y
+CONFIG_ARMV7_LPAE=y
 CONFIG_ARCH_QEMU=y
 CONFIG_ENV_SIZE=0x40000
 CONFIG_ENV_SECT_SIZE=0x40000
Signature · 1 line
2.26.2
#11 lore raw

PATCH 3/5 arm: qemu: implement enable_caches()

In reply to Ard Biesheuvel

Add an override for enable_caches to enable the I and D caches, along with the cached 1:1 mapping of all of DRAM. This is needed for running U-Boot under virtualization with QEMU/kvm.

---
 board/emulation/qemu-arm/qemu-arm.c | 7 +++++++
 1 file changed, 7 insertions(+)
board/emulation/qemu-arm/qemu-arm.c +7 -0
index 69e8ef46f1f5..1b0d543b93c1 100644
@@ -4,6 +4,7 @@
  */
 
 #include <common.h>
+#include <cpu_func.h>
 #include <dm.h>
 #include <fdtdec.h>
 #include <init.h>
@@ -94,6 +95,12 @@ void *board_fdt_blob_setup(void)
 	return (void *)CONFIG_SYS_SDRAM_BASE;
 }
 
+void enable_caches(void)
+{
+	 icache_enable();
+	 dcache_enable();
+}
+
 #if defined(CONFIG_EFI_RNG_PROTOCOL)
 #include <efi_loader.h>
 #include <efi_rng.h>
Signature · 1 line
2.26.2
#14 lore raw

PATCH 4/5 arm: qemu: disable the EFI workaround for older GRUB

In reply to Ard Biesheuvel

The QEMU/mach-virt targeted port of u-boot currently only runs on QEMU under TCG emulation, which does not model the caches at all, and so no users can exist that are relying on the GRUB hack for EFI boot.

We will shortly enable support for running under KVM, but the GRUB hack (which disables all caches without doing cache cleaning by VA during ExitBootServices()) is likely to cause more problems than it solves, given that KVM hosts require correct maintenance if they incorporate non-architected system caches.

So let's disable the GRUB hack by default on the QEMU/mach-virt port.

---
 configs/qemu_arm_defconfig | 1 +
 1 file changed, 1 insertion(+)
configs/qemu_arm_defconfig +1 -0
index 75bdce7708c7..1d2b4437cb07 100644
@@ -47,3 +47,4 @@ CONFIG_USB=y
 CONFIG_DM_USB=y
 CONFIG_USB_EHCI_HCD=y
 CONFIG_USB_EHCI_PCI=y
+# CONFIG_EFI_GRUB_ARM32_WORKAROUND is not set
Signature · 1 line
2.26.2
#16 lore raw

PATCH 5/5 arm: qemu: override flash accessors to use virtualizable instructions

In reply to Ard Biesheuvel

Some instructions in the ARM ISA have multiple output registers, such as ldrd/ldp (load pair), where two registers are loaded from memory, but also ldr with indexing, where the memory base register is incremented as well when the value is loaded to the destination register.

MMIO emulation under KVM is based on using the architecturally defined syndrome information that is provided when an exception is taken to the hypervisor. This syndrome information describes whether the instruction that triggered the exception is a load or a store, what the faulting address was, and which register was the destination register.

This syndrome information can only describe one destination register, and when the trapping instruction is one with multiple outputs, KVM throws an error like

kvm [615929]: Data abort outside memslots with no valid syndrome info

on the host and kills the QEMU process with the following error:

U-Boot 2020.07-rc3-00208-g88bd5b179360-dirty (Jun 06 2020 - 11:59:22 +0200)

DRAM: 1 GiB
Flash: error: kvm run failed Function not implemented R00=00000001 R01=00000040 R02=7ee0ce20 R03=00000000 R04=7ffd9eec R05=00000004 R06=7ffda3f8 R07=00000055 R08=7ffd9eec R09=7ef0ded0 R10=7ee0ce20 R11=00000000 R12=00000004 R13=7ee0cdf8 R14=00000000 R15=7ff72d08 PSR=200001d3 --C- A svc32
QEMU: Terminated

This means that, in order to run U-Boot in QEMU under KVM, we need to avoid such instructions when accessing emulated devices. For the flash in particular, which is a hybrid between a ROM (backed by a memslot) when in array mode, and an emulated MMIO device (when in write mode), we need to take care to only use instructions that KVM can deal with when they trap.

So override the flash accessors that are used when running on QEMU. Note that the write accessors are included for completeness, but the read accessors are the ones that need this special care.

---
 board/emulation/qemu-arm/qemu-arm.c | 55 ++++++++++++++++++++
 include/configs/qemu-arm.h          |  1 +
 2 files changed, 56 insertions(+)
board/emulation/qemu-arm/qemu-arm.c +55 -0
index 1b0d543b93c1..32e18fd8b985 100644
@@ -142,3 +142,58 @@ efi_status_t platform_get_rng_device(struct udevice **dev)
 	return EFI_SUCCESS;
 }
 #endif /* CONFIG_EFI_RNG_PROTOCOL */
+
+#ifdef CONFIG_ARM64
+#define __W	"w"
+#else
+#define __W
+#endif
+
+void flash_write8(u8 value, void *addr)
+{
+	asm("strb %" __W "1, %0" : "=m"(*(u8 *)addr) : "r"(value));
+}
+
+void flash_write16(u16 value, void *addr)
+{
+	asm("strh %" __W "1, %0" : "=m"(*(u16 *)addr) : "r"(value));
+}
+
+void flash_write32(u32 value, void *addr)
+{
+	asm("str %" __W "1, %0" : "=m"(*(u32 *)addr) : "r"(value));
+}
+
+void flash_write64(u64 value, void *addr)
+{
+	BUG(); /* FLASH_CFI_64BIT is not implemented by QEMU */
+}
+
+u8 flash_read8(void *addr)
+{
+	u8 ret;
+
+	asm("ldrb %" __W "0, %1" : "=r"(ret) : "m"(*(u8 *)addr));
+	return ret;
+}
+
+u16 flash_read16(void *addr)
+{
+	u16 ret;
+
+	asm("ldrh %" __W "0, %1" : "=r"(ret) : "m"(*(u16 *)addr));
+	return ret;
+}
+
+u32 flash_read32(void *addr)
+{
+	u32 ret;
+
+	asm("ldr %" __W "0, %1" : "=r"(ret) : "m"(*(u32 *)addr));
+	return ret;
+}
+
+u64 flash_read64(void *addr)
+{
+	BUG(); /* FLASH_CFI_64BIT is not implemented by QEMU */
+}
include/configs/qemu-arm.h +1 -0
index 1ef75a87836b..bc8b7c5c1238 100644
@@ -53,5 +53,6 @@
 #define CONFIG_SYS_MAX_FLASH_BANKS	2
 #endif
 #define CONFIG_SYS_MAX_FLASH_SECT	256 /* Sector: 256K, Bank: 64M */
+#define CONFIG_CFI_FLASH_USE_WEAK_ACCESSORS
 
 #endif /* __CONFIG_H */
Signature · 1 line
2.26.2
#2 lore raw

PATCH 1/5 arm: enable allocate-on-read for LPAE's DCACHE_WRITEBACK

In reply to Ard Biesheuvel
Quoted 5 lines · Ard Biesheuvel

The LPAE version of DCACHE_WRITEBACK is currently defined as no-allocate for both reads and writes, which deviates from the non-LPAE definition, and mostly defeats the purpose of enabling the caches in the first place.

So align LPAE with !LPAE, and enable allocate-on-read.

Hello Ard,

thanks for analyzing why booting Linux on QEMU fails in some scenarios.

Do you know where in U-Boot is the value for !LPAE is defined?

arch/arm/include/asm/system.h static inline void set_dacr(unsigned int val) · line 446

Quoted 13 lines
---
 arch/arm/include/asm/system.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
arch/arm/include/asm/system.h +0 -0
index 7a40b56acdca..21b26557d28b 100644
@@ -445,7 +445,7 @@ static inline void set_dacr(unsigned int val)
  * Memory types
  */

To me the lines below look like black magic.

In the comment above the definition, please, add a reference explaining where the values are defined and a comment explaining why the actual values are chosen.

Maybe this could be a starting point for the description:

"This constant is used define memory attribute encodings in a Long-descriptor format translation table entry for stage 1 translations. It is used to set the Memory Attribute Indirection Registers MAIR and HMAIR. For details see [1,2].

[1] MAIR0, Memory Attribute Indirection Register 0

https://developer.arm.com/docs/ddi0595/b/aarch32-system-registers/mair0/a/DDI0568A_b_armv8_r_supplement.pdf
[2] HMAIR0, Hyp Memory Attribute Indirection Register 0

https://developer.arm.com/docs/ddi0595/b/aarch32-system-registers/hmair0 "

Best regards

Heinrich

Quoted 7 lines
arch/arm/include/asm/system.h +1 -1
 #define MEMORY_ATTRIBUTES	((0x00 << (0 * 8)) | (0x88 << (1 * 8)) | \
-				 (0xcc << (2 * 8)) | (0xff << (3 * 8)))
+				 (0xee << (2 * 8)) | (0xff << (3 * 8)))
 
 /* options available for data cache on each page */
 enum dcache_option {
#6 lore raw

PATCH 2/5 arm: qemu: enable LPAE on 32-bit

In reply to Ard Biesheuvel
Quoted 3 lines · Ard Biesheuvel

QEMU's mach-virt machine only supports selecting CPU models that implement the virtualization extensions, and are therefore guaranteed to support LPAE as well.

I wonder why
qemu-system-arm -machine virt -cpu help
lists cortex-a9 (which is not LPAE enabled).

But when I try to use it I get
qemu-system-arm: mach-virt: CPU type cortex-a9-arm-cpu not supported This looks like a missing feature in QEMU.

The default CPU for machine=virt is arm,cortex-a15.

Quoted 30 lines

Initially, QEMU would not allow emulating these CPUs running in HYP mode (or EL2, for AArch64), but today, it also contains a complete implementation of the virtualization extensions themselves.

This means we could be running U-Boot in HYP mode, in which case the LPAE long descriptor page table format is the only format that is supported. If we are not running in HYP mode, we can use either.

So let's enable CONFIG_ARMV7_LPAE for qemu_arm_defconfig so that we get the best support for running with the MMU and caches enabled at any privilege level.

---
 configs/qemu_arm_defconfig | 1 +
 1 file changed, 1 insertion(+)
configs/qemu_arm_defconfig +1 -0
index a8473988bd76..75bdce7708c7 100644
@@ -1,5 +1,6 @@
 CONFIG_ARM=y
 CONFIG_ARM_SMCCC=y
+CONFIG_ARMV7_LPAE=y
 CONFIG_ARCH_QEMU=y
 CONFIG_ENV_SIZE=0x40000
 CONFIG_ENV_SECT_SIZE=0x40000
#7 lore raw

PATCH 2/5 arm: qemu: enable LPAE on 32-bit

Quoted 9 lines · Heinrich Schuchardt
Quoted 3 lines · Ard Biesheuvel

QEMU's mach-virt machine only supports selecting CPU models that implement the virtualization extensions, and are therefore guaranteed to support LPAE as well.

I wonder why
qemu-system-arm -machine virt -cpu help
lists cortex-a9 (which is not LPAE enabled).

Quoted 22 lines

But when I try to use it I get
qemu-system-arm: mach-virt: CPU type cortex-a9-arm-cpu not supported This looks like a missing feature in QEMU.

The default CPU for machine=virt is arm,cortex-a15.

Quoted 14 lines

Initially, QEMU would not allow emulating these CPUs running in HYP mode (or EL2, for AArch64), but today, it also contains a complete implementation of the virtualization extensions themselves.

This means we could be running U-Boot in HYP mode, in which case the LPAE long descriptor page table format is the only format that is supported. If we are not running in HYP mode, we can use either.

So let's enable CONFIG_ARMV7_LPAE for qemu_arm_defconfig so that we get the best support for running with the MMU and caches enabled at any privilege level.

You missed to CC the maintainer of QEMU ARM 'VIRT' BOARD. - We have scripts/get_maintainer.pl to find the maintainers.

Cc Tuomas Tynkkynen <[email protected]>

Best regards

Heinrich

Quoted 16 lines
Quoted 16 lines
---
 configs/qemu_arm_defconfig | 1 +
 1 file changed, 1 insertion(+)
configs/qemu_arm_defconfig +1 -0
index a8473988bd76..75bdce7708c7 100644
@@ -1,5 +1,6 @@
 CONFIG_ARM=y
 CONFIG_ARM_SMCCC=y
+CONFIG_ARMV7_LPAE=y
 CONFIG_ARCH_QEMU=y
 CONFIG_ENV_SIZE=0x40000
 CONFIG_ENV_SECT_SIZE=0x40000
#12 lore raw

PATCH 3/5 arm: qemu: implement enable_caches()

In reply to Ard Biesheuvel
Quoted 22 lines · Ard Biesheuvel

Add an override for enable_caches to enable the I and D caches, along with the cached 1:1 mapping of all of DRAM. This is needed for running U-Boot under virtualization with QEMU/kvm.

---
 board/emulation/qemu-arm/qemu-arm.c | 7 +++++++
 1 file changed, 7 insertions(+)
board/emulation/qemu-arm/qemu-arm.c +1 -0
index 69e8ef46f1f5..1b0d543b93c1 100644
@@ -4,6 +4,7 @@
  */
 
 #include <common.h>
+#include <cpu_func.h>
 #include <dm.h>
 #include <fdtdec.h>
 #include <init.h>

board/emulation/qemu-arm/qemu-arm.c void *board_fdt_blob_setup(void) · line 103

Quoted 9 lines
board/emulation/qemu-arm/qemu-arm.c +6 -0
@@ -94,6 +95,12 @@ void *board_fdt_blob_setup(void)
 	return (void *)CONFIG_SYS_SDRAM_BASE;
 }
 
+void enable_caches(void)
+{
+	 icache_enable();
+	 dcache_enable();
+}
+

For other ARM architectures I have seen:

int arch_cpu_init(void)
{
????????icache_enable();
??????? return 0;
}

void enable_caches(void)
{
??????? dcache_enable();
}

Some boards have

if (!icache_status())

	icache_enable();

others

#if !CONFIG_IS_ENABLED(SYS_ICACHE_OFF)
	icache_enable();

#endif

Tom could you, please, advice what is the correct way to do it.

Best regards

Heinrich

Quoted 4 lines
board/emulation/qemu-arm/qemu-arm.c +0 -0
 #if defined(CONFIG_EFI_RNG_PROTOCOL)
 #include <efi_loader.h>
 #include <efi_rng.h>
#15 lore raw

PATCH 4/5 arm: qemu: disable the EFI workaround for older GRUB

In reply to Ard Biesheuvel
Quoted 11 lines · Ard Biesheuvel

The QEMU/mach-virt targeted port of u-boot currently only runs on QEMU under TCG emulation, which does not model the caches at all, and so no users can exist that are relying on the GRUB hack for EFI boot.

We will shortly enable support for running under KVM, but the GRUB hack (which disables all caches without doing cache cleaning by VA during ExitBootServices()) is likely to cause more problems than it solves, given that KVM hosts require correct maintenance if they incorporate non-architected system caches.

Quoted 4 lines

So let's disable the GRUB hack by default on the QEMU/mach-virt port.

This patch could be merged with 2/5. You are changing the same defconfig.

Quoted 14 lines
---
 configs/qemu_arm_defconfig | 1 +
 1 file changed, 1 insertion(+)
configs/qemu_arm_defconfig +1 -0
index 75bdce7708c7..1d2b4437cb07 100644
@@ -47,3 +47,4 @@ CONFIG_USB=y
 CONFIG_DM_USB=y
 CONFIG_USB_EHCI_HCD=y
 CONFIG_USB_EHCI_PCI=y
+# CONFIG_EFI_GRUB_ARM32_WORKAROUND is not set
#17 lore raw

PATCH 5/5 arm: qemu: override flash accessors to use virtualizable instructions

In reply to Ard Biesheuvel
Quoted 70 lines · Ard Biesheuvel

Some instructions in the ARM ISA have multiple output registers, such as ldrd/ldp (load pair), where two registers are loaded from memory, but also ldr with indexing, where the memory base register is incremented as well when the value is loaded to the destination register.

MMIO emulation under KVM is based on using the architecturally defined syndrome information that is provided when an exception is taken to the hypervisor. This syndrome information describes whether the instruction that triggered the exception is a load or a store, what the faulting address was, and which register was the destination register.

This syndrome information can only describe one destination register, and when the trapping instruction is one with multiple outputs, KVM throws an error like

kvm [615929]: Data abort outside memslots with no valid syndrome info

on the host and kills the QEMU process with the following error:

U-Boot 2020.07-rc3-00208-g88bd5b179360-dirty (Jun 06 2020 - 11:59:22 +0200)

DRAM: 1 GiB
Flash: error: kvm run failed Function not implemented R00=00000001 R01=00000040 R02=7ee0ce20 R03=00000000 R04=7ffd9eec R05=00000004 R06=7ffda3f8 R07=00000055 R08=7ffd9eec R09=7ef0ded0 R10=7ee0ce20 R11=00000000 R12=00000004 R13=7ee0cdf8 R14=00000000 R15=7ff72d08 PSR=200001d3 --C- A svc32
QEMU: Terminated

This means that, in order to run U-Boot in QEMU under KVM, we need to avoid such instructions when accessing emulated devices. For the flash in particular, which is a hybrid between a ROM (backed by a memslot) when in array mode, and an emulated MMIO device (when in write mode), we need to take care to only use instructions that KVM can deal with when they trap.

So override the flash accessors that are used when running on QEMU. Note that the write accessors are included for completeness, but the read accessors are the ones that need this special care.

---
 board/emulation/qemu-arm/qemu-arm.c | 55 ++++++++++++++++++++
 include/configs/qemu-arm.h          |  1 +
 2 files changed, 56 insertions(+)
board/emulation/qemu-arm/qemu-arm.c +15 -0
index 1b0d543b93c1..32e18fd8b985 100644
@@ -142,3 +142,58 @@ efi_status_t platform_get_rng_device(struct udevice **dev)
 	return EFI_SUCCESS;
 }
 #endif /* CONFIG_EFI_RNG_PROTOCOL */
+
+#ifdef CONFIG_ARM64
+#define __W	"w"
+#else
+#define __W
+#endif
+
+void flash_write8(u8 value, void *addr)
+{
+	asm("strb %" __W "1, %0" : "=m"(*(u8 *)addr) : "r"(value));
+}
+
+void flash_write16(u16 value, void *addr)
+{
+	asm("strh %" __W "1, %0" : "=m"(*(u16 *)addr) : "r"(value));

board/emulation/qemu-arm/qemu-arm.c efi_status_t platform_get_rng_device(struct udevice **dev) · line 169

Quoted 10 lines
board/emulation/qemu-arm/qemu-arm.c +10 -0
+}
+
+void flash_write32(u32 value, void *addr)
+{
+	asm("str %" __W "1, %0" : "=m"(*(u32 *)addr) : "r"(value));
+}
+
+void flash_write64(u64 value, void *addr)
+{
+	BUG(); /* FLASH_CFI_64BIT is not implemented by QEMU */

Why should this BUG() on aarch64?
Why is panicking in U-Boot better than crashing in QEMU? Why can't this be realized as two 32bit writes?

This seems to be wrong:
drivers/mtd/cfi_flash.c:179:
/* No architectures currently implement __raw_readq() */

I find definitions for all architectures.

So shouldn't you correct cfi_flash.c and override __arch_getq() and __arch_putq() instead?

Quoted 19 lines
board/emulation/qemu-arm/qemu-arm.c +19 -0
+}
+
+u8 flash_read8(void *addr)
+{
+	u8 ret;
+
+	asm("ldrb %" __W "0, %1" : "=r"(ret) : "m"(*(u8 *)addr));
+	return ret;
+}
+
+u16 flash_read16(void *addr)
+{
+	u16 ret;
+
+	asm("ldrh %" __W "0, %1" : "=r"(ret) : "m"(*(u16 *)addr));
+	return ret;
+}
+
+u32 flash_read32(void *addr)

board/emulation/qemu-arm/qemu-arm.c efi_status_t platform_get_rng_device(struct udevice **dev) · line 198

Quoted 10 lines
board/emulation/qemu-arm/qemu-arm.c +10 -0
+{
+	u32 ret;
+
+	asm("ldr %" __W "0, %1" : "=r"(ret) : "m"(*(u32 *)addr));
+	return ret;
+}
+
+u64 flash_read64(void *addr)
+{
+	BUG(); /* FLASH_CFI_64BIT is not implemented by QEMU */

Same here.

Best regards

Heinrich

Quoted 13 lines
board/emulation/qemu-arm/qemu-arm.c +1 -0
+}
include/configs/qemu-arm.h +1 -0
index 1ef75a87836b..bc8b7c5c1238 100644
@@ -53,5 +53,6 @@
 #define CONFIG_SYS_MAX_FLASH_BANKS	2
 #endif
 #define CONFIG_SYS_MAX_FLASH_SECT	256 /* Sector: 256K, Bank: 64M */
+#define CONFIG_CFI_FLASH_USE_WEAK_ACCESSORS
 
 #endif /* __CONFIG_H */
#18 lore raw

PATCH 5/5 arm: qemu: override flash accessors to use virtualizable instructions

Quoted 76 lines · Heinrich Schuchardt
Quoted 74 lines · Ard Biesheuvel

Some instructions in the ARM ISA have multiple output registers, such as ldrd/ldp (load pair), where two registers are loaded from memory, but also ldr with indexing, where the memory base register is incremented as well when the value is loaded to the destination register.

MMIO emulation under KVM is based on using the architecturally defined syndrome information that is provided when an exception is taken to the hypervisor. This syndrome information describes whether the instruction that triggered the exception is a load or a store, what the faulting address was, and which register was the destination register.

This syndrome information can only describe one destination register, and when the trapping instruction is one with multiple outputs, KVM throws an error like

kvm [615929]: Data abort outside memslots with no valid syndrome info

on the host and kills the QEMU process with the following error:

U-Boot 2020.07-rc3-00208-g88bd5b179360-dirty (Jun 06 2020 - 11:59:22 +0200)

DRAM: 1 GiB
Flash: error: kvm run failed Function not implemented R00=00000001 R01=00000040 R02=7ee0ce20 R03=00000000 R04=7ffd9eec R05=00000004 R06=7ffda3f8 R07=00000055 R08=7ffd9eec R09=7ef0ded0 R10=7ee0ce20 R11=00000000 R12=00000004 R13=7ee0cdf8 R14=00000000 R15=7ff72d08 PSR=200001d3 --C- A svc32
QEMU: Terminated

This means that, in order to run U-Boot in QEMU under KVM, we need to avoid such instructions when accessing emulated devices. For the flash in particular, which is a hybrid between a ROM (backed by a memslot) when in array mode, and an emulated MMIO device (when in write mode), we need to take care to only use instructions that KVM can deal with when they trap.

So override the flash accessors that are used when running on QEMU. Note that the write accessors are included for completeness, but the read accessors are the ones that need this special care.

---
 board/emulation/qemu-arm/qemu-arm.c | 55 ++++++++++++++++++++
 include/configs/qemu-arm.h          |  1 +
 2 files changed, 56 insertions(+)
board/emulation/qemu-arm/qemu-arm.c +19 -0
index 1b0d543b93c1..32e18fd8b985 100644
@@ -142,3 +142,58 @@ efi_status_t platform_get_rng_device(struct udevice **dev)
      return EFI_SUCCESS;
 }
 #endif /* CONFIG_EFI_RNG_PROTOCOL */
+
+#ifdef CONFIG_ARM64
+#define __W  "w"
+#else
+#define __W
+#endif
+
+void flash_write8(u8 value, void *addr)
+{
+     asm("strb %" __W "1, %0" : "=m"(*(u8 *)addr) : "r"(value));
+}
+
+void flash_write16(u16 value, void *addr)
+{
+     asm("strh %" __W "1, %0" : "=m"(*(u16 *)addr) : "r"(value));
+}
+
+void flash_write32(u32 value, void *addr)
+{
Quoted 8 lines

board/emulation/qemu-arm/qemu-arm.c efi_status_t platform_get_rng_device(struct udevice **dev) · line 169

Quoted 6 lines
board/emulation/qemu-arm/qemu-arm.c +6 -0
+     asm("str %" __W "1, %0" : "=m"(*(u32 *)addr) : "r"(value));
+}
+
+void flash_write64(u64 value, void *addr)
+{
+     BUG(); /* FLASH_CFI_64BIT is not implemented by QEMU */

Why should this BUG() on aarch64?

QEMU's CFI emulation does not implement 8 byte width, so there is no point in implementing this accessor, as it will never be called anyway.

The BUG() is there to ensure that *if* QEMU ever does get support for 8 byte wide CFI flash, we notice immediately, rather than having to debug weird failures.

Quoted 1 line

Why is panicking in U-Boot better than crashing in QEMU?

Because U-boot crashes in the guest, while QEMU crashes in the host.

Quoted 2 lines

Why can't this be realized as two 32bit writes?

It could be but there is no point: QEMU will never exercise this code path anyway.

Quoted 9 lines

This seems to be wrong:
drivers/mtd/cfi_flash.c:179:
/* No architectures currently implement __raw_readq() */

I find definitions for all architectures.

So shouldn't you correct cfi_flash.c and override __arch_getq() and __arch_putq() instead?

If anyone wants to fix 8 byte CFI, they are welcome to do so, but it is a separate issue.

Quoted 50 lines
Quoted 19 lines
board/emulation/qemu-arm/qemu-arm.c +19 -0
+}
+
+u8 flash_read8(void *addr)
+{
+     u8 ret;
+
+     asm("ldrb %" __W "0, %1" : "=r"(ret) : "m"(*(u8 *)addr));
+     return ret;
+}
+
+u16 flash_read16(void *addr)
+{
+     u16 ret;
+
+     asm("ldrh %" __W "0, %1" : "=r"(ret) : "m"(*(u16 *)addr));
+     return ret;
+}
+
+u32 flash_read32(void *addr)

board/emulation/qemu-arm/qemu-arm.c efi_status_t platform_get_rng_device(struct udevice **dev) · line 198

Quoted 10 lines
board/emulation/qemu-arm/qemu-arm.c +10 -0
+{
+     u32 ret;
+
+     asm("ldr %" __W "0, %1" : "=r"(ret) : "m"(*(u32 *)addr));
+     return ret;
+}
+
+u64 flash_read64(void *addr)
+{
+     BUG(); /* FLASH_CFI_64BIT is not implemented by QEMU */

Same here.

Best regards

Heinrich

Quoted 13 lines
board/emulation/qemu-arm/qemu-arm.c +1 -0
+}
include/configs/qemu-arm.h +1 -0
index 1ef75a87836b..bc8b7c5c1238 100644
@@ -53,5 +53,6 @@
 #define CONFIG_SYS_MAX_FLASH_BANKS   2
 #endif
 #define CONFIG_SYS_MAX_FLASH_SECT    256 /* Sector: 256K, Bank: 64M */
+#define CONFIG_CFI_FLASH_USE_WEAK_ACCESSORS
 
 #endif /* __CONFIG_H */
#3 lore raw

PATCH 1/5 arm: enable allocate-on-read for LPAE's DCACHE_WRITEBACK

In reply to Heinrich Schuchardt
Quoted 6 lines · Heinrich Schuchardt
Quoted 4 lines · Ard Biesheuvel

The LPAE version of DCACHE_WRITEBACK is currently defined as no-allocate for both reads and writes, which deviates from the non-LPAE definition, and mostly defeats the purpose of enabling the caches in the first place.

Quoted 8 lines
Quoted 1 line

So align LPAE with !LPAE, and enable allocate-on-read.

Hello Ard,

thanks for analyzing why booting Linux on QEMU fails in some scenarios.

Do you know where in U-Boot is the value for !LPAE is defined?

Non-LPAE ARMV7A has (in arch/arm/include/asm/system.h)

DCACHE_WRITETHROUGH = DCACHE_OFF | TTB_SECT_C_MASK, DCACHE_WRITEBACK = DCACHE_WRITETHROUGH | TTB_SECT_B_MASK, DCACHE_WRITEALLOC = DCACHE_WRITEBACK | TTB_SECT_TEX(1),

and so DCACHE_WRITEBACK has the C and B bits set in the block descriptor, and the TEX field set to 0x0

G5.7.2 in the ARM ARM (DDI0487E.a) describes this as

Outer and Inner Write-Back, Read-Allocate
No Write-Allocate

DCACHE_WRITEALLOC has the C and B bits set in the block descriptor, and the TEX field set to 0x1, which is described as

Outer and Inner Write-Back, Read-Allocate Write-Allocate

Quoted 27 lines

arch/arm/include/asm/system.h static inline void set_dacr(unsigned int val) · line 446

Quoted 13 lines
---
 arch/arm/include/asm/system.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
arch/arm/include/asm/system.h +0 -0
index 7a40b56acdca..21b26557d28b 100644
@@ -445,7 +445,7 @@ static inline void set_dacr(unsigned int val)
  * Memory types
  */

To me the lines below look like black magic.

In the comment above the definition, please, add a reference explaining where the values are defined and a comment explaining why the actual values are chosen.

Maybe this could be a starting point for the description:

"This constant is used define memory attribute encodings in a Long-descriptor format translation table entry for stage 1 translations. It is used to set the Memory Attribute Indirection Registers MAIR and HMAIR. For details see [1,2].

Quoted 7 lines

[1] MAIR0, Memory Attribute Indirection Register 0

https://developer.arm.com/docs/ddi0595/b/aarch32-system-registers/mair0/a/DDI0568A_b_armv8_r_supplement.pdf
[2] HMAIR0, Hyp Memory Attribute Indirection Register 0

https://developer.arm.com/docs/ddi0595/b/aarch32-system-registers/hmair0 "

Better refer to the ARM ARM for the A profile here (not R). [DDI0487E.a]

So the memory types are indexed: four fields of MAIR are populated with the four chosen memory types:

[0] Device-nGnrnE
[1] Outer and Inner Write-Through, Read-Allocate No Write-Allocate
[2] Outer and Inner Write-Back, Read-Allocate No Write-Allocate
[3] Outer and Inner Write-Back, Read-Allocate Write-Allocate

and the enum just selects one of these fields:

DCACHE_OFF = TTB_SECT | TTB_SECT_MAIR(0) | TTB_SECT_XN_MASK, DCACHE_WRITETHROUGH = TTB_SECT | TTB_SECT_MAIR(1), DCACHE_WRITEBACK = TTB_SECT | TTB_SECT_MAIR(2), DCACHE_WRITEALLOC = TTB_SECT | TTB_SECT_MAIR(3),

BTW it seems DCACHE_WRITETHROUGH is also incorrect: this should be 0xaa for read-allocate as well.

Quoted 8 lines
Quoted 7 lines
arch/arm/include/asm/system.h +1 -1
 #define MEMORY_ATTRIBUTES    ((0x00 << (0 * 8)) | (0x88 << (1 * 8)) | \
-                              (0xcc << (2 * 8)) | (0xff << (3 * 8)))
+                              (0xee << (2 * 8)) | (0xff << (3 * 8)))
 
 /* options available for data cache on each page */
 enum dcache_option {
#4 lore raw

PATCH 1/5 arm: enable allocate-on-read for LPAE's DCACHE_WRITEBACK

Quoted 82 lines · Ard Biesheuvel
Quoted 6 lines · Heinrich Schuchardt
Quoted 4 lines · Ard Biesheuvel

The LPAE version of DCACHE_WRITEBACK is currently defined as no-allocate for both reads and writes, which deviates from the non-LPAE definition, and mostly defeats the purpose of enabling the caches in the first place.

Quoted 8 lines
Quoted 1 line

So align LPAE with !LPAE, and enable allocate-on-read.

Hello Ard,

thanks for analyzing why booting Linux on QEMU fails in some scenarios.

Do you know where in U-Boot is the value for !LPAE is defined?

Non-LPAE ARMV7A has (in arch/arm/include/asm/system.h)

DCACHE_WRITETHROUGH = DCACHE_OFF | TTB_SECT_C_MASK, DCACHE_WRITEBACK = DCACHE_WRITETHROUGH | TTB_SECT_B_MASK, DCACHE_WRITEALLOC = DCACHE_WRITEBACK | TTB_SECT_TEX(1),

and so DCACHE_WRITEBACK has the C and B bits set in the block descriptor, and the TEX field set to 0x0

G5.7.2 in the ARM ARM (DDI0487E.a) describes this as

Outer and Inner Write-Back, Read-Allocate
No Write-Allocate

DCACHE_WRITEALLOC has the C and B bits set in the block descriptor, and the TEX field set to 0x1, which is described as

Outer and Inner Write-Back, Read-Allocate Write-Allocate

Quoted 27 lines

arch/arm/include/asm/system.h static inline void set_dacr(unsigned int val) · line 446

Quoted 13 lines
---
 arch/arm/include/asm/system.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
arch/arm/include/asm/system.h +0 -0
index 7a40b56acdca..21b26557d28b 100644
@@ -445,7 +445,7 @@ static inline void set_dacr(unsigned int val)
  * Memory types
  */

To me the lines below look like black magic.

In the comment above the definition, please, add a reference explaining where the values are defined and a comment explaining why the actual values are chosen.

Maybe this could be a starting point for the description:

"This constant is used define memory attribute encodings in a Long-descriptor format translation table entry for stage 1 translations. It is used to set the Memory Attribute Indirection Registers MAIR and HMAIR. For details see [1,2].

Quoted 7 lines

[1] MAIR0, Memory Attribute Indirection Register 0

https://developer.arm.com/docs/ddi0595/b/aarch32-system-registers/mair0/a/DDI0568A_b_armv8_r_supplement.pdf
[2] HMAIR0, Hyp Memory Attribute Indirection Register 0

https://developer.arm.com/docs/ddi0595/b/aarch32-system-registers/hmair0 "

Better refer to the ARM ARM for the A profile here (not R). [DDI0487E.a]

So the memory types are indexed: four fields of MAIR are populated with the four chosen memory types:

[0] Device-nGnrnE
[1] Outer and Inner Write-Through, Read-Allocate No Write-Allocate
[2] Outer and Inner Write-Back, Read-Allocate No Write-Allocate
[3] Outer and Inner Write-Back, Read-Allocate Write-Allocate

and the enum just selects one of these fields:

Quoted 7 lines

DCACHE_OFF = TTB_SECT | TTB_SECT_MAIR(0) | TTB_SECT_XN_MASK, DCACHE_WRITETHROUGH = TTB_SECT | TTB_SECT_MAIR(1), DCACHE_WRITEBACK = TTB_SECT | TTB_SECT_MAIR(2), DCACHE_WRITEALLOC = TTB_SECT | TTB_SECT_MAIR(3),

BTW it seems DCACHE_WRITETHROUGH is also incorrect: this should be 0xaa for read-allocate as well.

This description would give a good starting point for other developers when supplied as a comment for MEMORY_ATTRIBUTES.

Best regards

Heinrich

Quoted 10 lines
Quoted 10 lines
Quoted 9 lines
arch/arm/include/asm/system.h +1 -1
 #define MEMORY_ATTRIBUTES    ((0x00 << (0 * 8)) | (0x88 << (1 * 8)) | \
-                              (0xcc << (2 * 8)) | (0xff << (3 * 8)))
+                              (0xee << (2 * 8)) | (0xff << (3 * 8)))
 
 /* options available for data cache on each page */
 enum dcache_option {
#8 lore raw

PATCH 2/5 arm: qemu: enable LPAE on 32-bit

In reply to Heinrich Schuchardt
Quoted 7 lines · Heinrich Schuchardt
Quoted 5 lines · Heinrich Schuchardt
Quoted 3 lines · Ard Biesheuvel

QEMU's mach-virt machine only supports selecting CPU models that implement the virtualization extensions, and are therefore guaranteed to support LPAE as well.

Quoted 8 lines
Quoted 8 lines

I wonder why
qemu-system-arm -machine virt -cpu help
lists cortex-a9 (which is not LPAE enabled).

But when I try to use it I get
qemu-system-arm: mach-virt: CPU type cortex-a9-arm-cpu not supported This looks like a missing feature in QEMU.

This is not a missing feature. The virt board uses PSCI for powerdown and reset, and to bring up secondary cores.
PSCI requires the HVC instruction, which is only available if the virt extensions are implemented.

So emulating CPUs without the virt extensions would require a replacement for PSCI to be implemented as well, which seems rather pointless to me.

Quoted 4 lines
Quoted 4 lines

The default CPU for machine=virt is arm,cortex-a15.

Thanks.

Quoted 13 lines
Quoted 13 lines
Quoted 13 lines

Initially, QEMU would not allow emulating these CPUs running in HYP mode (or EL2, for AArch64), but today, it also contains a complete implementation of the virtualization extensions themselves.

This means we could be running U-Boot in HYP mode, in which case the LPAE long descriptor page table format is the only format that is supported. If we are not running in HYP mode, we can use either.

So let's enable CONFIG_ARMV7_LPAE for qemu_arm_defconfig so that we get the best support for running with the MMU and caches enabled at any privilege level.

Quoted 7 lines
Quoted 1 line
Quoted 1 line

You missed to CC the maintainer of QEMU ARM 'VIRT' BOARD. - We have scripts/get_maintainer.pl to find the maintainers.

Cc Tuomas Tynkkynen <[email protected]>

Apologies. I will cc Tuomas for v2.

#9 lore raw

PATCH 2/5 arm: qemu: enable LPAE on 32-bit

Am June 7, 2020 8:59:00 AM UTC schrieb Ard Biesheuvel <[email protected]>:

Quoted 16 lines
Quoted 5 lines · Heinrich Schuchardt
Quoted 3 lines · Heinrich Schuchardt
Quoted 2 lines · Ard Biesheuvel

QEMU's mach-virt machine only supports selecting CPU models that implement the virtualization extensions, and are therefore

guaranteed

Quoted 8 lines
Quoted 8 lines
Quoted 1 line

to support LPAE as well.

I wonder why
qemu-system-arm -machine virt -cpu help
lists cortex-a9 (which is not LPAE enabled).

But when I try to use it I get
qemu-system-arm: mach-virt: CPU type cortex-a9-arm-cpu not

Quoted 8 lines

supported

Quoted 2 lines
Quoted 2 lines

This looks like a missing feature in QEMU.

This is not a missing feature. The virt board uses PSCI for powerdown and reset, and to bring up secondary cores.
PSCI requires the HVC instruction, which is only available if the virt extensions are implemented.

By missing feature I meant -cpu help output should be filtered according to the -machine value if provided.

Quoted 38 lines

So emulating CPUs without the virt extensions would require a replacement for PSCI to be implemented as well, which seems rather pointless to me.

Quoted 4 lines
Quoted 4 lines

The default CPU for machine=virt is arm,cortex-a15.

Thanks.

Quoted 2 lines
Quoted 2 lines
Quoted 2 lines

Initially, QEMU would not allow emulating these CPUs running in

HYP

Quoted 4 lines
Quoted 4 lines
Quoted 4 lines

mode (or EL2, for AArch64), but today, it also contains a complete implementation of the virtualization extensions themselves.

This means we could be running U-Boot in HYP mode, in which case

the

Quoted 4 lines
Quoted 4 lines
Quoted 4 lines

LPAE long descriptor page table format is the only format that is supported. If we are not running in HYP mode, we can use either.

So let's enable CONFIG_ARMV7_LPAE for qemu_arm_defconfig so that

we

Quoted 1 line
Quoted 1 line
Quoted 1 line

get the best support for running with the MMU and caches enabled

at

Quoted 9 lines
Quoted 3 lines
Quoted 3 lines

any privilege level.

You missed to CC the maintainer of QEMU ARM 'VIRT' BOARD. - We have scripts/get_maintainer.pl to find the maintainers.

Cc Tuomas Tynkkynen <[email protected]>

Apologies. I will cc Tuomas for v2.

#10 lore raw

PATCH 2/5 arm: qemu: enable LPAE on 32-bit

Quoted 22 lines · Heinrich Schuchardt

Am June 7, 2020 8:59:00 AM UTC schrieb Ard Biesheuvel <[email protected]>:

Quoted 20 lines
Quoted 5 lines · Heinrich Schuchardt
Quoted 3 lines · Heinrich Schuchardt
Quoted 2 lines · Ard Biesheuvel

QEMU's mach-virt machine only supports selecting CPU models that implement the virtualization extensions, and are therefore

guaranteed

Quoted 8 lines
Quoted 8 lines
Quoted 1 line

to support LPAE as well.

I wonder why
qemu-system-arm -machine virt -cpu help
lists cortex-a9 (which is not LPAE enabled).

But when I try to use it I get
qemu-system-arm: mach-virt: CPU type cortex-a9-arm-cpu not

supported

Quoted 2 lines
Quoted 2 lines

This looks like a missing feature in QEMU.

Quoted 7 lines
Quoted 4 lines

This is not a missing feature. The virt board uses PSCI for powerdown and reset, and to bring up secondary cores.
PSCI requires the HVC instruction, which is only available if the virt extensions are implemented.

By missing feature I meant -cpu help output should be filtered according to the -machine value if provided.

Ah fair enough. Yes, that would be useful.

Unfortunately, ARM does not permit me to contribute to QEMU, so hopefully someone else can take this on.

#13 lore raw

PATCH 3/5 arm: qemu: implement enable_caches()

In reply to Heinrich Schuchardt
Quoted 51 lines · Heinrich Schuchardt
Quoted 22 lines · Ard Biesheuvel

Add an override for enable_caches to enable the I and D caches, along with the cached 1:1 mapping of all of DRAM. This is needed for running U-Boot under virtualization with QEMU/kvm.

---
 board/emulation/qemu-arm/qemu-arm.c | 7 +++++++
 1 file changed, 7 insertions(+)
board/emulation/qemu-arm/qemu-arm.c +1 -0
index 69e8ef46f1f5..1b0d543b93c1 100644
@@ -4,6 +4,7 @@
  */
 
 #include <common.h>
+#include <cpu_func.h>
 #include <dm.h>
 #include <fdtdec.h>
 #include <init.h>

board/emulation/qemu-arm/qemu-arm.c void *board_fdt_blob_setup(void) · line 103

Quoted 9 lines
board/emulation/qemu-arm/qemu-arm.c +6 -0
@@ -94,6 +95,12 @@ void *board_fdt_blob_setup(void)
 	return (void *)CONFIG_SYS_SDRAM_BASE;
 }
 
+void enable_caches(void)
+{
+	 icache_enable();
+	 dcache_enable();
+}
+

For other ARM architectures I have seen:

int arch_cpu_init(void)
{
????????icache_enable();
??????? return 0;
}

void enable_caches(void)
{
??????? dcache_enable();
}

Some boards have

if (!icache_status())

	icache_enable();
Quoted 7 lines

others

#if !CONFIG_IS_ENABLED(SYS_ICACHE_OFF)
	icache_enable();

#endif

Tom could you, please, advice what is the correct way to do it.

Off-hand? Dealing with it per SoC/mach/board as we do today. Sometimes we need a dcache_disable() before the dcache_enable() as well to clear out previous mappings. So dropping this in here is probably the best choice for qemu-arm. Thanks!

Signature · 8 lines
Tom
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 659 bytes
Desc: not available
URL: <https://lists.denx.de/pipermail/u-boot/attachments/20200608/30cafd04/attachment.sig>