Skip to contentSkip to conversation

arm/arm64: Prepare for target-efi

PATCH Superseded Andrew Jones 3 participants 25 messages Apr 20, 2021 – Apr 27, 2021

Subsystems (2)

v2 of 3 Superseded Latest is v3 →

lore
Share & tools
lore
#0 lore raw

v2 cover arm/arm64: Prepare for target-efi

v2:
- Addressed all comments from Nikos and Alex
- The biggest changes are
* dropping the weird persistent map stuff that I never liked by taking Alex's suggestion to just create the idmap early
* adding mem_region_add() to clean up memory region adding code, also improved the assignment of region fields
- Also, while we found that we still have a memory map assumption (3G-4G reserved for virtual memory allocation), I only make that assumption clear. I've left removing it for an additional patch for another day.
- Added psci_invoke_none() to use prior to the PSCI method being set
- Added r-b's for each patch given, unless the commit changed too much
- Didn't take Alex's suggestion to use x5 for stacktop when calling setup from start. I prefer explicitly loading it again.

This series is a collection of patches derived from [1] that pave the way for new targets, e.g. target-efi[2]. These patches mostly address the elimination of memory map assumptions and they shouldn't have any functional changes. The last two patches are a couple of patches not related to the memory map, but they also prepare for bare metal targets. I tossed them in since I don't think they should be too controversial. This patch series is also available here [3].

[1] https://github.com/rhdrjones/kvm-unit-tests/commits/target-efi
[2] https://www.youtube.com/watch?v=kvaufVrL0J0
[3] https://gitlab.com/rhdrjones/kvm-unit-tests/-/commits/efiprep

Thanks,
drew

Andrew Jones (8):
arm/arm64: Reorganize cstart assembler arm/arm64: Move setup_vm into setup
pci-testdev: ioremap regions
arm/arm64: mmu: Stop mapping an assumed IO region arm/arm64: mmu: Remove memory layout assumptions arm/arm64: setup: Consolidate memory layout assumptions
chr-testdev: Silently fail init
arm/arm64: psci: don't assume method is hvc

 arm/cstart.S           |  92 ++++++++++++++--------
 arm/cstart64.S         |  45 ++++++++---
 arm/flat.lds           |  23 ++++++
 arm/selftest.c         |  34 ++------
 lib/arm/asm/io.h       |   6 ++
 lib/arm/asm/mmu.h      |   1 +
 lib/arm/asm/page.h     |   2 +
 lib/arm/asm/psci.h     |  10 ++-
 lib/arm/asm/setup.h    |   7 +-
 lib/arm/mmu.c          |  53 +++++++++----
 lib/arm/psci.c         |  37 +++++++--
 lib/arm/setup.c        | 175 ++++++++++++++++++++++++-----------------
 lib/arm64/asm/io.h     |   6 ++
 lib/arm64/asm/mmu.h    |   1 +
 lib/arm64/asm/page.h   |   2 +
 lib/chr-testdev.c      |   5 +-
 lib/pci-host-generic.c |   5 +-
 lib/pci-host-generic.h |   4 +-
 lib/pci-testdev.c      |   4 +
 19 files changed, 333 insertions(+), 179 deletions(-)
Signature · 1 line
2.30.2
#1 lore raw

PATCH v2 1/8 arm/arm64: Reorganize cstart assembler

Move secondary_entry helper functions out of .init and into .text, since secondary_entry isn't run at at "init" time. Actually, anything that is used after init time should be in .text, as we may not include .init in some build configurations.

Reviewed-by Nikos Nikoleris <[email protected]>

---
 arm/cstart.S   | 66 +++++++++++++++++++++++++++++---------------------
 arm/cstart64.S | 18 ++++++++------
 2 files changed, 49 insertions(+), 35 deletions(-)
arm/cstart.S +38 -28
index d88a98362940..b2c0ba061cd5 100644
@@ -96,32 +96,7 @@ start:
 	bl	exit
 	b	halt
 
-
-.macro set_mode_stack mode, stack
-	add	\stack, #S_FRAME_SIZE
-	msr	cpsr_c, #(\mode | PSR_I_BIT | PSR_F_BIT)
-	isb
-	mov	sp, \stack
-.endm
-
-exceptions_init:
-	mrc	p15, 0, r2, c1, c0, 0	@ read SCTLR
-	bic	r2, #CR_V		@ SCTLR.V := 0
-	mcr	p15, 0, r2, c1, c0, 0	@ write SCTLR
-	ldr	r2, =vector_table
-	mcr	p15, 0, r2, c12, c0, 0	@ write VBAR
-
-	mrs	r2, cpsr
-
-	/* first frame reserved for svc mode */
-	set_mode_stack	UND_MODE, r0
-	set_mode_stack	ABT_MODE, r0
-	set_mode_stack	IRQ_MODE, r0
-	set_mode_stack	FIQ_MODE, r0
-
-	msr	cpsr_cxsf, r2		@ back to svc mode
-	isb
-	mov	pc, lr
+.text
 
 enable_vfp:
 	/* Enable full access to CP10 and CP11: */
@@ -133,8 +108,6 @@ enable_vfp:
 	vmsr	fpexc, r0
 	mov	pc, lr
 
-.text
-
 .global get_mmu_off
 get_mmu_off:
 	ldr	r0, =auxinfo
@@ -235,6 +208,43 @@ asm_mmu_disable:
 
 	mov     pc, lr
 
+/*
+ * Vectors
+ */
+
+.macro set_mode_stack mode, stack
+	add	\stack, #S_FRAME_SIZE
+	msr	cpsr_c, #(\mode | PSR_I_BIT | PSR_F_BIT)
+	isb
+	mov	sp, \stack
+.endm
+
+/*
+ * exceptions_init
+ *
+ * Input r0 is the stack top, which is the exception stacks base
+ */
+exceptions_init:
+	mrc	p15, 0, r2, c1, c0, 0	@ read SCTLR
+	bic	r2, #CR_V		@ SCTLR.V := 0
+	mcr	p15, 0, r2, c1, c0, 0	@ write SCTLR
+	ldr	r2, =vector_table
+	mcr	p15, 0, r2, c12, c0, 0	@ write VBAR
+
+	mrs	r2, cpsr
+
+	/*
+	 * The first frame is reserved for svc mode
+	 */
+	set_mode_stack	UND_MODE, r0
+	set_mode_stack	ABT_MODE, r0
+	set_mode_stack	IRQ_MODE, r0
+	set_mode_stack	FIQ_MODE, r0
+
+	msr	cpsr_cxsf, r2		@ back to svc mode
+	isb
+	mov	pc, lr
+
 /*
  * Vector stubs
  * Simplified version of the Linux kernel implementation
arm/cstart64.S +11 -7
index 0a85338bcdae..7963e1fea979 100644
@@ -109,13 +109,6 @@ start:
 	bl	exit
 	b	halt
 
-exceptions_init:
-	adrp	x4, vector_table
-	add	x4, x4, :lo12:vector_table
-	msr	vbar_el1, x4
-	isb
-	ret
-
 .text
 
 .globl get_mmu_off
@@ -251,6 +244,17 @@ asm_mmu_disable:
 
 /*
  * Vectors
+ */
+
+exceptions_init:
+	adrp	x4, vector_table
+	add	x4, x4, :lo12:vector_table
+	msr	vbar_el1, x4
+	isb
+	ret
+
+/*
+ * Vector stubs
  * Adapted from arch/arm64/kernel/entry.S
  */
 .macro vector_stub, name, vec
Signature · 1 line
2.30.2
Reviewed-by on v2 1/8
#2 lore raw

Re: [PATCH kvm-unit-tests v2 1/8] arm/arm64: Reorganize cstart assembler

Hi Drew,

Quoted 129 lines · Andrew Jones

Move secondary_entry helper functions out of .init and into .text, since secondary_entry isn't run at at "init" time. Actually, anything that is used after init time should be in .text, as we may not include .init in some build configurations.

Reviewed-by Nikos Nikoleris <[email protected]>

---
 arm/cstart.S   | 66 +++++++++++++++++++++++++++++---------------------
 arm/cstart64.S | 18 ++++++++------
 2 files changed, 49 insertions(+), 35 deletions(-)
arm/cstart.S +38 -28
index d88a98362940..b2c0ba061cd5 100644
@@ -96,32 +96,7 @@ start:
 	bl	exit
 	b	halt
 
-
-.macro set_mode_stack mode, stack
-	add	\stack, #S_FRAME_SIZE
-	msr	cpsr_c, #(\mode | PSR_I_BIT | PSR_F_BIT)
-	isb
-	mov	sp, \stack
-.endm
-
-exceptions_init:
-	mrc	p15, 0, r2, c1, c0, 0	@ read SCTLR
-	bic	r2, #CR_V		@ SCTLR.V := 0
-	mcr	p15, 0, r2, c1, c0, 0	@ write SCTLR
-	ldr	r2, =vector_table
-	mcr	p15, 0, r2, c12, c0, 0	@ write VBAR
-
-	mrs	r2, cpsr
-
-	/* first frame reserved for svc mode */
-	set_mode_stack	UND_MODE, r0
-	set_mode_stack	ABT_MODE, r0
-	set_mode_stack	IRQ_MODE, r0
-	set_mode_stack	FIQ_MODE, r0
-
-	msr	cpsr_cxsf, r2		@ back to svc mode
-	isb
-	mov	pc, lr
+.text
 
 enable_vfp:
 	/* Enable full access to CP10 and CP11: */
@@ -133,8 +108,6 @@ enable_vfp:
 	vmsr	fpexc, r0
 	mov	pc, lr
 
-.text
-
 .global get_mmu_off
 get_mmu_off:
 	ldr	r0, =auxinfo
@@ -235,6 +208,43 @@ asm_mmu_disable:
 
 	mov     pc, lr
 
+/*
+ * Vectors
+ */
+
+.macro set_mode_stack mode, stack
+	add	\stack, #S_FRAME_SIZE
+	msr	cpsr_c, #(\mode | PSR_I_BIT | PSR_F_BIT)
+	isb
+	mov	sp, \stack
+.endm
+
+/*
+ * exceptions_init
+ *
+ * Input r0 is the stack top, which is the exception stacks base
+ */
+exceptions_init:
+	mrc	p15, 0, r2, c1, c0, 0	@ read SCTLR
+	bic	r2, #CR_V		@ SCTLR.V := 0
+	mcr	p15, 0, r2, c1, c0, 0	@ write SCTLR
+	ldr	r2, =vector_table
+	mcr	p15, 0, r2, c12, c0, 0	@ write VBAR
+
+	mrs	r2, cpsr
+
+	/*
+	 * The first frame is reserved for svc mode
+	 */
+	set_mode_stack	UND_MODE, r0
+	set_mode_stack	ABT_MODE, r0
+	set_mode_stack	IRQ_MODE, r0
+	set_mode_stack	FIQ_MODE, r0
+
+	msr	cpsr_cxsf, r2		@ back to svc mode
+	isb
+	mov	pc, lr
+
 /*
  * Vector stubs
  * Simplified version of the Linux kernel implementation
arm/cstart64.S +4 -7
index 0a85338bcdae..7963e1fea979 100644
@@ -109,13 +109,6 @@ start:
 	bl	exit
 	b	halt
 
-exceptions_init:
-	adrp	x4, vector_table
-	add	x4, x4, :lo12:vector_table
-	msr	vbar_el1, x4
-	isb
-	ret
-
 .text
 
 .globl get_mmu_off
@@ -251,6 +244,17 @@ asm_mmu_disable:
 
 /*
  * Vectors
+ */
+
+exceptions_init:
+	adrp	x4, vector_table

arm/cstart64.S asm_mmu_disable: · line 260

Quoted 10 lines
arm/cstart64.S +7 -0
+	add	x4, x4, :lo12:vector_table
+	msr	vbar_el1, x4
+	isb
+	ret
+
+/*
+ * Vector stubs
  * Adapted from arch/arm64/kernel/entry.S
  */
 .macro vector_stub, name, vec

The diff looks nice and clean, exactly what you want from a straightforward move. I checked that the new code matches the old one, but for arm I haven't checked that the code is correct, because I'm not familiar with arm assembly. I also compiled for arm and arm64:

Thanks,

Alex

#3 lore raw

PATCH v2 2/8 arm/arm64: Move setup_vm into setup

In reply to Andrew Jones

Consolidate our setup calls to reduce the amount we need to do from init::start. Also remove a couple of pointless comments from setup().

---
 arm/cstart.S    | 6 ------
 arm/cstart64.S  | 5 -----
 lib/arm/setup.c | 7 +++++--
 3 files changed, 5 insertions(+), 13 deletions(-)
arm/cstart.S +0 -6
index b2c0ba061cd5..bf3c78157e6a 100644
@@ -81,12 +81,7 @@ start:
 	/* complete setup */
 	pop	{r0-r1}
 	bl	setup
-	bl	get_mmu_off
-	cmp	r0, #0
-	bne	1f
-	bl	setup_vm
 
-1:
 	/* run the test */
 	ldr	r0, =__argc
 	ldr	r0, [r0]
@@ -108,7 +103,6 @@ enable_vfp:
 	vmsr	fpexc, r0
 	mov	pc, lr
 
-.global get_mmu_off
 get_mmu_off:
 	ldr	r0, =auxinfo
 	ldr	r0, [r0, #4]
arm/cstart64.S +0 -5
index 7963e1fea979..27251fe8b5cd 100644
@@ -93,11 +93,7 @@ start:
 
 	/* complete setup */
 	bl	setup				// x0 is the addr of the dtb
-	bl	get_mmu_off
-	cbnz	x0, 1f
-	bl	setup_vm
 
-1:
 	/* run the test */
 	adrp	x0, __argc
 	ldr	w0, [x0, :lo12:__argc]
@@ -111,7 +107,6 @@ start:
 
 .text
 
-.globl get_mmu_off
 get_mmu_off:
 	adrp	x0, auxinfo
 	ldr	x0, [x0, :lo12:auxinfo + 8]
lib/arm/setup.c +5 -2
index 751ba980000a..9c16f6004e9f 100644
@@ -16,6 +16,8 @@
 #include <alloc.h>
 #include <alloc_phys.h>
 #include <alloc_page.h>
+#include <vmalloc.h>
+#include <auxinfo.h>
 #include <argv.h>
 #include <asm/thread_info.h>
 #include <asm/setup.h>
@@ -233,7 +235,6 @@ void setup(const void *fdt)
 		freemem += initrd_size;
 	}
 
-	/* call init functions */
 	mem_init(PAGE_ALIGN((unsigned long)freemem));
 	cpu_init();
 
@@ -243,7 +244,6 @@ void setup(const void *fdt)
 	/* mem_init must be called before io_init */
 	io_init();
 
-	/* finish setup */
 	timer_save_state();
 
 	ret = dt_get_bootargs(&bootargs);
@@ -256,4 +256,7 @@ void setup(const void *fdt)
 		memcpy(env, initrd, initrd_size);
 		setup_env(env, initrd_size);
 	}
+
+	if (!(auxinfo.flags & AUXINFO_MMU_OFF))
+		setup_vm();
 }
Signature · 1 line
2.30.2
#4 lore raw

PATCH v2 3/8 pci-testdev: ioremap regions

In reply to Andrew Jones

Don't assume the physical addresses used with PCI have already been identity mapped.

---
 lib/pci-host-generic.c | 5 ++---
 lib/pci-host-generic.h | 4 ++--
 lib/pci-testdev.c      | 4 ++++
 3 files changed, 8 insertions(+), 5 deletions(-)
lib/pci-host-generic.c +2 -3
index 818150dc0a66..de93b8feac39 100644
@@ -122,7 +122,7 @@ static struct pci_host_bridge *pci_dt_probe(void)
 		      sizeof(host->addr_space[0]) * nr_addr_spaces);
 	assert(host != NULL);
 
-	host->start		= base.addr;
+	host->start		= ioremap(base.addr, base.size);
 	host->size		= base.size;
 	host->bus		= bus;
 	host->bus_max		= bus_max;
@@ -279,8 +279,7 @@ phys_addr_t pci_host_bridge_get_paddr(u64 pci_addr)
 
 static void __iomem *pci_get_dev_conf(struct pci_host_bridge *host, int devfn)
 {
-	return (void __iomem *)(unsigned long)
-		host->start + (devfn << PCI_ECAM_DEVFN_SHIFT);
+	return (void __iomem *)host->start + (devfn << PCI_ECAM_DEVFN_SHIFT);
 }
 
 u8 pci_config_readb(pcidevaddr_t dev, u8 off)
lib/pci-host-generic.h +2 -2
index fd30e7c74ed8..0ffe6380ec8f 100644
@@ -18,8 +18,8 @@ struct pci_addr_space {
 };
 
 struct pci_host_bridge {
-	phys_addr_t		start;
-	phys_addr_t		size;
+	void __iomem		*start;
+	size_t			size;
 	int			bus;
 	int			bus_max;
 	int			nr_addr_spaces;
lib/pci-testdev.c +4 -0
index 039bb44781c1..4f2e5663b2d6 100644
@@ -185,7 +185,11 @@ int pci_testdev(void)
 	mem = ioremap(addr, PAGE_SIZE);
 
 	addr = pci_bar_get_addr(&pci_dev, 1);
+#if defined(__i386__) || defined(__x86_64__)
 	io = (void *)(unsigned long)addr;
+#else
+	io = ioremap(addr, PAGE_SIZE);
+#endif
 
 	nr_tests += pci_testdev_all(mem, &pci_testdev_mem_ops);
 	nr_tests += pci_testdev_all(io, &pci_testdev_io_ops);
Signature · 1 line
2.30.2
#5 lore raw

Re: [PATCH kvm-unit-tests v2 3/8] pci-testdev: ioremap regions

On Tue, 20 Apr 2021 20:59:57 +0200
Andrew Jones <[email protected]> wrote:

Hi,

Quoted 25 lines

Don't assume the physical addresses used with PCI have already been identity mapped.

---
 lib/pci-host-generic.c | 5 ++---
 lib/pci-host-generic.h | 4 ++--
 lib/pci-testdev.c      | 4 ++++
 3 files changed, 8 insertions(+), 5 deletions(-)
lib/pci-host-generic.c +1 -1
index 818150dc0a66..de93b8feac39 100644
@@ -122,7 +122,7 @@ static struct pci_host_bridge *pci_dt_probe(void)
 		      sizeof(host->addr_space[0]) * nr_addr_spaces);
 	assert(host != NULL);
 
-	host->start		= base.addr;
+	host->start		= ioremap(base.addr, base.size);
 	host->size		= base.size;
 	host->bus		= bus;
 	host->bus_max		= bus_max;

lib/pci-host-generic.c phys_addr_t pci_host_bridge_get_paddr(u64 pci_addr) · line 282

Quoted 6 lines
lib/pci-host-generic.c +1 -2
@@ -279,8 +279,7 @@ phys_addr_t pci_host_bridge_get_paddr(u64 pci_addr)
 
 static void __iomem *pci_get_dev_conf(struct pci_host_bridge *host, int devfn)
 {
-	return (void __iomem *)(unsigned long)
-		host->start + (devfn << PCI_ECAM_DEVFN_SHIFT);
+	return (void __iomem *)host->start + (devfn << PCI_ECAM_DEVFN_SHIFT);

But host->start's type is now exactly "void __iomem *", so why the cast? And are we OK with doing pointer arithmetic on a void pointer?

Quoted 23 lines
lib/pci-host-generic.c +0 -0
 }
 
 u8 pci_config_readb(pcidevaddr_t dev, u8 off)
lib/pci-host-generic.h +2 -2
index fd30e7c74ed8..0ffe6380ec8f 100644
@@ -18,8 +18,8 @@ struct pci_addr_space {
 };
 
 struct pci_host_bridge {
-	phys_addr_t		start;
-	phys_addr_t		size;
+	void __iomem		*start;
+	size_t			size;
 	int			bus;
 	int			bus_max;
 	int			nr_addr_spaces;
lib/pci-testdev.c +0 -0
index 039bb44781c1..4f2e5663b2d6 100644

lib/pci-testdev.c int pci_testdev(void) · line 192

Quoted 8 lines
lib/pci-testdev.c +4 -0
@@ -185,7 +185,11 @@ int pci_testdev(void)
 	mem = ioremap(addr, PAGE_SIZE);
 
 	addr = pci_bar_get_addr(&pci_dev, 1);
+#if defined(__i386__) || defined(__x86_64__)
 	io = (void *)(unsigned long)addr;
+#else
+	io = ioremap(addr, PAGE_SIZE);
+#endif

I am bit puzzled: For anything but x86 ioremap() is implemented like the first statement, so why do we differentiate here? Shouldn't either one of the statements be fine, for all architectures?

Cheers,
Andre

Quoted 3 lines
lib/pci-testdev.c +0 -0
 	nr_tests += pci_testdev_all(mem, &pci_testdev_mem_ops);
 	nr_tests += pci_testdev_all(io, &pci_testdev_io_ops);
#6 lore raw

Re: [PATCH kvm-unit-tests v2 3/8] pci-testdev: ioremap regions

Quoted 31 lines · Andre Przywara

On Tue, 20 Apr 2021 20:59:57 +0200
Andrew Jones <[email protected]> wrote:

Hi,

Quoted 26 lines

Don't assume the physical addresses used with PCI have already been identity mapped.

---
 lib/pci-host-generic.c | 5 ++---
 lib/pci-host-generic.h | 4 ++--
 lib/pci-testdev.c      | 4 ++++
 3 files changed, 8 insertions(+), 5 deletions(-)
lib/pci-host-generic.c +1 -1
index 818150dc0a66..de93b8feac39 100644
@@ -122,7 +122,7 @@ static struct pci_host_bridge *pci_dt_probe(void)
 		      sizeof(host->addr_space[0]) * nr_addr_spaces);
 	assert(host != NULL);
 
-	host->start		= base.addr;
+	host->start		= ioremap(base.addr, base.size);
 	host->size		= base.size;
 	host->bus		= bus;
 	host->bus_max		= bus_max;
@@ -279,8 +279,7 @@ phys_addr_t pci_host_bridge_get_paddr(u64 pci_addr)
 
Quoted 7 lines

lib/pci-host-generic.c phys_addr_t pci_host_bridge_get_paddr(u64 pci_addr) · line 282

Quoted 5 lines
lib/pci-host-generic.c +1 -2
 static void __iomem *pci_get_dev_conf(struct pci_host_bridge *host, int devfn)
 {
-	return (void __iomem *)(unsigned long)
-		host->start + (devfn << PCI_ECAM_DEVFN_SHIFT);
+	return (void __iomem *)host->start + (devfn << PCI_ECAM_DEVFN_SHIFT);

But host->start's type is now exactly "void __iomem *", so why the cast?

Only because I didn't think to remove it. Will do for v3.

Quoted 1 line

And are we OK with doing pointer arithmetic on a void pointer?

I'm pretty sure we have other cases, but if you'd prefer I can create a local char* for the arithmetic and then return it as a void*. (Assuming that's what you're suggesting I do.)

Quoted 26 lines
Quoted 26 lines
lib/pci-host-generic.c +0 -0
 }
 
 u8 pci_config_readb(pcidevaddr_t dev, u8 off)
lib/pci-host-generic.h +2 -2
index fd30e7c74ed8..0ffe6380ec8f 100644
@@ -18,8 +18,8 @@ struct pci_addr_space {
 };
 
 struct pci_host_bridge {
-	phys_addr_t		start;
-	phys_addr_t		size;
+	void __iomem		*start;
+	size_t			size;
 	int			bus;
 	int			bus_max;
 	int			nr_addr_spaces;
lib/pci-testdev.c +0 -0
index 039bb44781c1..4f2e5663b2d6 100644
@@ -185,7 +185,11 @@ int pci_testdev(void)
 	mem = ioremap(addr, PAGE_SIZE);
 
Quoted 10 lines

lib/pci-testdev.c int pci_testdev(void) · line 192

Quoted 6 lines
lib/pci-testdev.c +4 -0
 	addr = pci_bar_get_addr(&pci_dev, 1);
+#if defined(__i386__) || defined(__x86_64__)
 	io = (void *)(unsigned long)addr;
+#else
+	io = ioremap(addr, PAGE_SIZE);
+#endif

I am bit puzzled: For anything but x86 ioremap() is implemented like the first statement, so why do we differentiate here? Shouldn't either one of the statements be fine, for all architectures?

The addresses in this context are pio. So x86 should use them verbatim, but other architectures that don't have pio will need to avoid them or remap them and use them with fake pio instructions (e.g. inb/outb wrappers for readb/writeb).

Thanks,
drew

#7 lore raw

PATCH v2 4/8 arm/arm64: mmu: Stop mapping an assumed IO region

In reply to Andrew Jones

By providing a proper ioremap function, we can just rely on devices calling it for each region they need (as they already do) instead of mapping a big assumed I/O range. We don't require the MMU to be enabled at the time of the ioremap. In that case, we add the mapping to the identity map anyway. This allows us to call setup_vm after io_init. Why don't we just call setup_vm before io_init, I hear you ask? Well, that's because tests like sieve want to start with the MMU off, later call setup_vm, and all the while have working I/O. Some unit tests are just really demanding...

---
 lib/arm/asm/io.h     |  6 ++++++
 lib/arm/asm/mmu.h    |  1 +
 lib/arm/asm/page.h   |  2 ++
 lib/arm/mmu.c        | 37 +++++++++++++++++++++++++++----------
 lib/arm64/asm/io.h   |  6 ++++++
 lib/arm64/asm/mmu.h  |  1 +
 lib/arm64/asm/page.h |  2 ++
 7 files changed, 45 insertions(+), 10 deletions(-)
lib/arm/asm/io.h +6 -0
index ba3b0b2412ad..e4caa6ff5d1e 100644
@@ -77,6 +77,12 @@ static inline void __raw_writel(u32 val, volatile void __iomem *addr)
 		     : "r" (val));
 }
 
+#define ioremap ioremap
+static inline void __iomem *ioremap(phys_addr_t phys_addr, size_t size)
+{
+	return __ioremap(phys_addr, size);
+}
+
 #define virt_to_phys virt_to_phys
 static inline phys_addr_t virt_to_phys(const volatile void *x)
 {
lib/arm/asm/mmu.h +1 -0
index 122874b8aebe..d88a4f16df42 100644
@@ -12,6 +12,7 @@
 #define PTE_SHARED		L_PTE_SHARED
 #define PTE_AF			PTE_EXT_AF
 #define PTE_WBWA		L_PTE_MT_WRITEALLOC
+#define PTE_UNCACHED		L_PTE_MT_UNCACHED
 
 /* See B3.18.7 TLB maintenance operations */
 
lib/arm/asm/page.h +2 -0
index 1fb5cd26ac66..8eb4a883808e 100644
@@ -47,5 +47,7 @@ typedef struct { pteval_t pgprot; } pgprot_t;
 extern phys_addr_t __virt_to_phys(unsigned long addr);
 extern unsigned long __phys_to_virt(phys_addr_t addr);
 
+extern void *__ioremap(phys_addr_t phys_addr, size_t size);
+
 #endif /* !__ASSEMBLY__ */
 #endif /* _ASMARM_PAGE_H_ */
lib/arm/mmu.c +27 -10
index 15eef007f256..ee0c79142ba1 100644
@@ -11,6 +11,7 @@
 #include <asm/mmu.h>
 #include <asm/setup.h>
 #include <asm/page.h>
+#include <asm/io.h>
 
 #include "alloc_page.h"
 #include "vmalloc.h"
@@ -157,9 +158,8 @@ void mmu_set_range_sect(pgd_t *pgtable, uintptr_t virt_offset,
 void *setup_mmu(phys_addr_t phys_end)
 {
 	uintptr_t code_end = (uintptr_t)&etext;
-	struct mem_region *r;
 
-	/* 0G-1G = I/O, 1G-3G = identity, 3G-4G = vmalloc */
+	/* 3G-4G region is reserved for vmalloc, cap phys_end at 3G */
 	if (phys_end > (3ul << 30))
 		phys_end = 3ul << 30;
 
@@ -170,14 +170,8 @@ void *setup_mmu(phys_addr_t phys_end)
 			"Unsupported translation granule %ld\n", PAGE_SIZE);
 #endif
 
-	mmu_idmap = alloc_page();
-
-	for (r = mem_regions; r->end; ++r) {
-		if (!(r->flags & MR_F_IO))
-			continue;
-		mmu_set_range_sect(mmu_idmap, r->start, r->start, r->end,
-				   __pgprot(PMD_SECT_UNCACHED | PMD_SECT_USER));
-	}
+	if (!mmu_idmap)
+		mmu_idmap = alloc_page();
 
 	/* armv8 requires code shared between EL1 and EL0 to be read-only */
 	mmu_set_range_ptes(mmu_idmap, PHYS_OFFSET,
@@ -192,6 +186,29 @@ void *setup_mmu(phys_addr_t phys_end)
 	return mmu_idmap;
 }
 
+void __iomem *__ioremap(phys_addr_t phys_addr, size_t size)
+{
+	phys_addr_t paddr_aligned = phys_addr & PAGE_MASK;
+	phys_addr_t paddr_end = PAGE_ALIGN(phys_addr + size);
+	pgprot_t prot = __pgprot(PTE_UNCACHED | PTE_USER);
+	pgd_t *pgtable;
+
+	assert(sizeof(long) == 8 || !(phys_addr >> 32));
+
+	if (mmu_enabled()) {
+		pgtable = current_thread_info()->pgtable;
+	} else {
+		if (!mmu_idmap)
+			mmu_idmap = alloc_page();
+		pgtable = mmu_idmap;
+	}
+
+	mmu_set_range_ptes(pgtable, paddr_aligned, paddr_aligned,
+			   paddr_end, prot);
+
+	return (void __iomem *)(unsigned long)phys_addr;
+}
+
 phys_addr_t __virt_to_phys(unsigned long addr)
 {
 	if (mmu_enabled()) {
lib/arm64/asm/io.h +6 -0
index e0a03b250d5b..be19f471c0fa 100644
@@ -71,6 +71,12 @@ static inline u64 __raw_readq(const volatile void __iomem *addr)
 	return val;
 }
 
+#define ioremap ioremap
+static inline void __iomem *ioremap(phys_addr_t phys_addr, size_t size)
+{
+	return __ioremap(phys_addr, size);
+}
+
 #define virt_to_phys virt_to_phys
 static inline phys_addr_t virt_to_phys(const volatile void *x)
 {
lib/arm64/asm/mmu.h +1 -0
index 72d75eafc882..72371b2d9fe3 100644
@@ -8,6 +8,7 @@
 #include <asm/barrier.h>
 
 #define PMD_SECT_UNCACHED	PMD_ATTRINDX(MT_DEVICE_nGnRE)
+#define PTE_UNCACHED		PTE_ATTRINDX(MT_DEVICE_nGnRE)
 #define PTE_WBWA		PTE_ATTRINDX(MT_NORMAL)
 
 static inline void flush_tlb_all(void)
lib/arm64/asm/page.h +2 -0
index ae4484b22114..d0fac6ea563d 100644
@@ -72,5 +72,7 @@ typedef struct { pteval_t pgprot; } pgprot_t;
 extern phys_addr_t __virt_to_phys(unsigned long addr);
 extern unsigned long __phys_to_virt(phys_addr_t addr);
 
+extern void *__ioremap(phys_addr_t phys_addr, size_t size);
+
 #endif /* !__ASSEMBLY__ */
 #endif /* _ASMARM64_PAGE_H_ */
Signature · 1 line
2.30.2
#8 lore raw

Re: [PATCH kvm-unit-tests v2 4/8] arm/arm64: mmu: Stop mapping an assumed IO region

Hi Drew,

Quoted 103 lines · Andrew Jones

By providing a proper ioremap function, we can just rely on devices calling it for each region they need (as they already do) instead of mapping a big assumed I/O range. We don't require the MMU to be enabled at the time of the ioremap. In that case, we add the mapping to the identity map anyway. This allows us to call setup_vm after io_init. Why don't we just call setup_vm before io_init, I hear you ask? Well, that's because tests like sieve want to start with the MMU off, later call setup_vm, and all the while have working I/O. Some unit tests are just really demanding...

---
 lib/arm/asm/io.h     |  6 ++++++
 lib/arm/asm/mmu.h    |  1 +
 lib/arm/asm/page.h   |  2 ++
 lib/arm/mmu.c        | 37 +++++++++++++++++++++++++++----------
 lib/arm64/asm/io.h   |  6 ++++++
 lib/arm64/asm/mmu.h  |  1 +
 lib/arm64/asm/page.h |  2 ++
 7 files changed, 45 insertions(+), 10 deletions(-)
lib/arm/asm/io.h +6 -0
index ba3b0b2412ad..e4caa6ff5d1e 100644
@@ -77,6 +77,12 @@ static inline void __raw_writel(u32 val, volatile void __iomem *addr)
 		     : "r" (val));
 }
 
+#define ioremap ioremap
+static inline void __iomem *ioremap(phys_addr_t phys_addr, size_t size)
+{
+	return __ioremap(phys_addr, size);
+}
+
 #define virt_to_phys virt_to_phys
 static inline phys_addr_t virt_to_phys(const volatile void *x)
 {
lib/arm/asm/mmu.h +1 -0
index 122874b8aebe..d88a4f16df42 100644
@@ -12,6 +12,7 @@
 #define PTE_SHARED		L_PTE_SHARED
 #define PTE_AF			PTE_EXT_AF
 #define PTE_WBWA		L_PTE_MT_WRITEALLOC
+#define PTE_UNCACHED		L_PTE_MT_UNCACHED
 
 /* See B3.18.7 TLB maintenance operations */
 
lib/arm/asm/page.h +2 -0
index 1fb5cd26ac66..8eb4a883808e 100644
@@ -47,5 +47,7 @@ typedef struct { pteval_t pgprot; } pgprot_t;
 extern phys_addr_t __virt_to_phys(unsigned long addr);
 extern unsigned long __phys_to_virt(phys_addr_t addr);
 
+extern void *__ioremap(phys_addr_t phys_addr, size_t size);
+
 #endif /* !__ASSEMBLY__ */
 #endif /* _ASMARM_PAGE_H_ */
lib/arm/mmu.c +4 -10
index 15eef007f256..ee0c79142ba1 100644
@@ -11,6 +11,7 @@
 #include <asm/mmu.h>
 #include <asm/setup.h>
 #include <asm/page.h>
+#include <asm/io.h>
 
 #include "alloc_page.h"
 #include "vmalloc.h"
@@ -157,9 +158,8 @@ void mmu_set_range_sect(pgd_t *pgtable, uintptr_t virt_offset,
 void *setup_mmu(phys_addr_t phys_end)
 {
 	uintptr_t code_end = (uintptr_t)&etext;
-	struct mem_region *r;
 
-	/* 0G-1G = I/O, 1G-3G = identity, 3G-4G = vmalloc */
+	/* 3G-4G region is reserved for vmalloc, cap phys_end at 3G */
 	if (phys_end > (3ul << 30))
 		phys_end = 3ul << 30;
 
@@ -170,14 +170,8 @@ void *setup_mmu(phys_addr_t phys_end)
 			"Unsupported translation granule %ld\n", PAGE_SIZE);
 #endif
 
-	mmu_idmap = alloc_page();
-
-	for (r = mem_regions; r->end; ++r) {
-		if (!(r->flags & MR_F_IO))
-			continue;
-		mmu_set_range_sect(mmu_idmap, r->start, r->start, r->end,
-				   __pgprot(PMD_SECT_UNCACHED | PMD_SECT_USER));
-	}
+	if (!mmu_idmap)
+		mmu_idmap = alloc_page();
 
 	/* armv8 requires code shared between EL1 and EL0 to be read-only */
 	mmu_set_range_ptes(mmu_idmap, PHYS_OFFSET,

lib/arm/mmu.c void *setup_mmu(phys_addr_t phys_end) · line 193

Quoted 8 lines
lib/arm/mmu.c +5 -0
@@ -192,6 +186,29 @@ void *setup_mmu(phys_addr_t phys_end)
 	return mmu_idmap;
 }
 
+void __iomem *__ioremap(phys_addr_t phys_addr, size_t size)
+{
+	phys_addr_t paddr_aligned = phys_addr & PAGE_MASK;
+	phys_addr_t paddr_end = PAGE_ALIGN(phys_addr + size);
+	pgprot_t prot = __pgprot(PTE_UNCACHED | PTE_USER);

From ARM DDI 0487G.a, page B-171:

"Hardware does not prevent speculative instruction fetches from a memory location with any of the Device memory attributes unless the memory location is also marked as execute-never for all Exception levels.
*Note*
This means that to prevent speculative instruction fetches from memory locations with Device memory attributes, any location that is assigned any Device memory type must also be marked as execute-never for all Exception levels. Failure to mark a memory location with any Device memory attribute as execute-never for all Exception levels is a programming error."

I think that should also be PTE_UXN | PTE_PXN (the kernel defines it the same way). Otherwise looks good.

Thanks,

Alex

Quoted 62 lines
lib/arm/mmu.c +18 -0
+	pgd_t *pgtable;
+
+	assert(sizeof(long) == 8 || !(phys_addr >> 32));
+
+	if (mmu_enabled()) {
+		pgtable = current_thread_info()->pgtable;
+	} else {
+		if (!mmu_idmap)
+			mmu_idmap = alloc_page();
+		pgtable = mmu_idmap;
+	}
+
+	mmu_set_range_ptes(pgtable, paddr_aligned, paddr_aligned,
+			   paddr_end, prot);
+
+	return (void __iomem *)(unsigned long)phys_addr;
+}
+
 phys_addr_t __virt_to_phys(unsigned long addr)
 {
 	if (mmu_enabled()) {
lib/arm64/asm/io.h +6 -0
index e0a03b250d5b..be19f471c0fa 100644
@@ -71,6 +71,12 @@ static inline u64 __raw_readq(const volatile void __iomem *addr)
 	return val;
 }
 
+#define ioremap ioremap
+static inline void __iomem *ioremap(phys_addr_t phys_addr, size_t size)
+{
+	return __ioremap(phys_addr, size);
+}
+
 #define virt_to_phys virt_to_phys
 static inline phys_addr_t virt_to_phys(const volatile void *x)
 {
lib/arm64/asm/mmu.h +1 -0
index 72d75eafc882..72371b2d9fe3 100644
@@ -8,6 +8,7 @@
 #include <asm/barrier.h>
 
 #define PMD_SECT_UNCACHED	PMD_ATTRINDX(MT_DEVICE_nGnRE)
+#define PTE_UNCACHED		PTE_ATTRINDX(MT_DEVICE_nGnRE)
 #define PTE_WBWA		PTE_ATTRINDX(MT_NORMAL)
 
 static inline void flush_tlb_all(void)
lib/arm64/asm/page.h +2 -0
index ae4484b22114..d0fac6ea563d 100644
@@ -72,5 +72,7 @@ typedef struct { pteval_t pgprot; } pgprot_t;
 extern phys_addr_t __virt_to_phys(unsigned long addr);
 extern unsigned long __phys_to_virt(phys_addr_t addr);
 
+extern void *__ioremap(phys_addr_t phys_addr, size_t size);
+
 #endif /* !__ASSEMBLY__ */
 #endif /* _ASMARM64_PAGE_H_ */
#9 lore raw

Re: [PATCH kvm-unit-tests v2 4/8] arm/arm64: mmu: Stop mapping an assumed IO region

Quoted 127 lines · Alexandru Elisei

Hi Drew,

Quoted 103 lines · Andrew Jones

By providing a proper ioremap function, we can just rely on devices calling it for each region they need (as they already do) instead of mapping a big assumed I/O range. We don't require the MMU to be enabled at the time of the ioremap. In that case, we add the mapping to the identity map anyway. This allows us to call setup_vm after io_init. Why don't we just call setup_vm before io_init, I hear you ask? Well, that's because tests like sieve want to start with the MMU off, later call setup_vm, and all the while have working I/O. Some unit tests are just really demanding...

---
 lib/arm/asm/io.h     |  6 ++++++
 lib/arm/asm/mmu.h    |  1 +
 lib/arm/asm/page.h   |  2 ++
 lib/arm/mmu.c        | 37 +++++++++++++++++++++++++++----------
 lib/arm64/asm/io.h   |  6 ++++++
 lib/arm64/asm/mmu.h  |  1 +
 lib/arm64/asm/page.h |  2 ++
 7 files changed, 45 insertions(+), 10 deletions(-)
lib/arm/asm/io.h +6 -0
index ba3b0b2412ad..e4caa6ff5d1e 100644
@@ -77,6 +77,12 @@ static inline void __raw_writel(u32 val, volatile void __iomem *addr)
 		     : "r" (val));
 }
 
+#define ioremap ioremap
+static inline void __iomem *ioremap(phys_addr_t phys_addr, size_t size)
+{
+	return __ioremap(phys_addr, size);
+}
+
 #define virt_to_phys virt_to_phys
 static inline phys_addr_t virt_to_phys(const volatile void *x)
 {
lib/arm/asm/mmu.h +1 -0
index 122874b8aebe..d88a4f16df42 100644
@@ -12,6 +12,7 @@
 #define PTE_SHARED		L_PTE_SHARED
 #define PTE_AF			PTE_EXT_AF
 #define PTE_WBWA		L_PTE_MT_WRITEALLOC
+#define PTE_UNCACHED		L_PTE_MT_UNCACHED
 
 /* See B3.18.7 TLB maintenance operations */
 
lib/arm/asm/page.h +2 -0
index 1fb5cd26ac66..8eb4a883808e 100644
@@ -47,5 +47,7 @@ typedef struct { pteval_t pgprot; } pgprot_t;
 extern phys_addr_t __virt_to_phys(unsigned long addr);
 extern unsigned long __phys_to_virt(phys_addr_t addr);
 
+extern void *__ioremap(phys_addr_t phys_addr, size_t size);
+
 #endif /* !__ASSEMBLY__ */
 #endif /* _ASMARM_PAGE_H_ */
lib/arm/mmu.c +4 -10
index 15eef007f256..ee0c79142ba1 100644
@@ -11,6 +11,7 @@
 #include <asm/mmu.h>
 #include <asm/setup.h>
 #include <asm/page.h>
+#include <asm/io.h>
 
 #include "alloc_page.h"
 #include "vmalloc.h"
@@ -157,9 +158,8 @@ void mmu_set_range_sect(pgd_t *pgtable, uintptr_t virt_offset,
 void *setup_mmu(phys_addr_t phys_end)
 {
 	uintptr_t code_end = (uintptr_t)&etext;
-	struct mem_region *r;
 
-	/* 0G-1G = I/O, 1G-3G = identity, 3G-4G = vmalloc */
+	/* 3G-4G region is reserved for vmalloc, cap phys_end at 3G */
 	if (phys_end > (3ul << 30))
 		phys_end = 3ul << 30;
 
@@ -170,14 +170,8 @@ void *setup_mmu(phys_addr_t phys_end)
 			"Unsupported translation granule %ld\n", PAGE_SIZE);
 #endif
 
-	mmu_idmap = alloc_page();
-
-	for (r = mem_regions; r->end; ++r) {
-		if (!(r->flags & MR_F_IO))
-			continue;
-		mmu_set_range_sect(mmu_idmap, r->start, r->start, r->end,
-				   __pgprot(PMD_SECT_UNCACHED | PMD_SECT_USER));
-	}
+	if (!mmu_idmap)
+		mmu_idmap = alloc_page();
 
 	/* armv8 requires code shared between EL1 and EL0 to be read-only */
 	mmu_set_range_ptes(mmu_idmap, PHYS_OFFSET,

lib/arm/mmu.c void *setup_mmu(phys_addr_t phys_end) · line 193

Quoted 8 lines
lib/arm/mmu.c +5 -0
@@ -192,6 +186,29 @@ void *setup_mmu(phys_addr_t phys_end)
 	return mmu_idmap;
 }
 
+void __iomem *__ioremap(phys_addr_t phys_addr, size_t size)
+{
+	phys_addr_t paddr_aligned = phys_addr & PAGE_MASK;
+	phys_addr_t paddr_end = PAGE_ALIGN(phys_addr + size);
+	pgprot_t prot = __pgprot(PTE_UNCACHED | PTE_USER);

From ARM DDI 0487G.a, page B-171:

"Hardware does not prevent speculative instruction fetches from a memory location with any of the Device memory attributes unless the memory location is also marked as execute-never for all Exception levels.
*Note*
This means that to prevent speculative instruction fetches from memory locations with Device memory attributes, any location that is assigned any Device memory type must also be marked as execute-never for all Exception levels. Failure to mark a memory location with any Device memory attribute as execute-never for all Exception levels is a programming error."

Quoted 2 lines

I think that should also be PTE_UXN | PTE_PXN (the kernel defines it the same way). Otherwise looks good.

Will fix for v3.

Thanks,
drew

#10 lore raw

PATCH v2 5/8 arm/arm64: mmu: Remove memory layout assumptions

In reply to Andrew Jones

Rather than making too many assumptions about the memory layout in mmu code, just set up the page tables per the memory regions (which means putting all the memory layout assumptions in setup). To ensure we get the right default flags set we need to split the primary region into two regions for code and data.

We still only expect the primary regions to be present, but the next patch will remove that assumption too.

(Unfortunately we still have an assumption in setup_mmu. We assume
the range 3G-4G is available for the virtual memory allocator. We'll need to remove that assumption as well with another patch in order to support arbitrary memory maps.)

---
 lib/arm/asm/setup.h |  1 +
 lib/arm/mmu.c       | 26 +++++++++++++++-----------
 lib/arm/setup.c     | 29 +++++++++++++++++++++--------
 3 files changed, 37 insertions(+), 19 deletions(-)
lib/arm/asm/setup.h +1 -0
index c8afb2493f8d..210c14f818fb 100644
@@ -15,6 +15,7 @@ extern int nr_cpus;
 
 #define MR_F_PRIMARY		(1U << 0)
 #define MR_F_IO			(1U << 1)
+#define MR_F_CODE		(1U << 2)
 #define MR_F_UNKNOWN		(1U << 31)
 
 struct mem_region {
lib/arm/mmu.c +15 -11
index ee0c79142ba1..4e3cf37e33d0 100644
@@ -20,8 +20,6 @@
 
 #include <linux/compiler.h>
 
-extern unsigned long etext;
-
 pgd_t *mmu_idmap;
 
 /* CPU 0 starts with disabled MMU */
@@ -157,7 +155,7 @@ void mmu_set_range_sect(pgd_t *pgtable, uintptr_t virt_offset,
 
 void *setup_mmu(phys_addr_t phys_end)
 {
-	uintptr_t code_end = (uintptr_t)&etext;
+	struct mem_region *r;
 
 	/* 3G-4G region is reserved for vmalloc, cap phys_end at 3G */
 	if (phys_end > (3ul << 30))
@@ -173,14 +171,20 @@ void *setup_mmu(phys_addr_t phys_end)
 	if (!mmu_idmap)
 		mmu_idmap = alloc_page();
 
-	/* armv8 requires code shared between EL1 and EL0 to be read-only */
-	mmu_set_range_ptes(mmu_idmap, PHYS_OFFSET,
-		PHYS_OFFSET, code_end,
-		__pgprot(PTE_WBWA | PTE_RDONLY | PTE_USER));
-
-	mmu_set_range_ptes(mmu_idmap, code_end,
-		code_end, phys_end,
-		__pgprot(PTE_WBWA | PTE_USER));
+	for (r = mem_regions; r->end; ++r) {
+		if (r->flags & MR_F_IO) {
+			continue;
+		} else if (r->flags & MR_F_CODE) {
+			assert_msg(r->flags & MR_F_PRIMARY, "Unexpected code region");
+			/* armv8 requires code shared between EL1 and EL0 to be read-only */
+			mmu_set_range_ptes(mmu_idmap, r->start, r->start, r->end,
+					   __pgprot(PTE_WBWA | PTE_USER | PTE_RDONLY));
+		} else {
+			assert_msg(r->flags & MR_F_PRIMARY, "Unexpected data region");
+			mmu_set_range_ptes(mmu_idmap, r->start, r->start, r->end,
+					   __pgprot(PTE_WBWA | PTE_USER));
+		}
+	}
 
 	mmu_enable(mmu_idmap);
 	return mmu_idmap;
lib/arm/setup.c +21 -8
index 9c16f6004e9f..7db308b70744 100644
@@ -31,6 +31,7 @@
 #define NR_INITIAL_MEM_REGIONS 16
 
 extern unsigned long stacktop;
+extern unsigned long etext;
 
 struct timer_state __timer_state;
 
@@ -88,10 +89,12 @@ unsigned int mem_region_get_flags(phys_addr_t paddr)
 
 static void mem_init(phys_addr_t freemem_start)
 {
+	phys_addr_t code_end = (phys_addr_t)(unsigned long)&etext;
 	struct dt_pbus_reg regs[NR_INITIAL_MEM_REGIONS];
-	struct mem_region primary, mem = {
+	struct mem_region mem = {
 		.start = (phys_addr_t)-1,
 	};
+	struct mem_region *primary = NULL;
 	phys_addr_t base, top;
 	int nr_regs, nr_io = 0, i;
 
@@ -110,8 +113,6 @@ static void mem_init(phys_addr_t freemem_start)
 	nr_regs = dt_get_memory_params(regs, NR_INITIAL_MEM_REGIONS - nr_io);
 	assert(nr_regs > 0);
 
-	primary = (struct mem_region){ 0 };
-
 	for (i = 0; i < nr_regs; ++i) {
 		struct mem_region *r = &mem_regions[nr_io + i];
 
@@ -123,7 +124,7 @@ static void mem_init(phys_addr_t freemem_start)
 		 */
 		if (freemem_start >= r->start && freemem_start < r->end) {
 			r->flags |= MR_F_PRIMARY;
-			primary = *r;
+			primary = r;
 		}
 
 		/*
@@ -135,13 +136,25 @@ static void mem_init(phys_addr_t freemem_start)
 		if (r->end > mem.end)
 			mem.end = r->end;
 	}
-	assert(primary.end != 0);
+	assert(primary);
 	assert(!(mem.start & ~PHYS_MASK) && !((mem.end - 1) & ~PHYS_MASK));
 
-	__phys_offset = primary.start;	/* PHYS_OFFSET */
-	__phys_end = primary.end;	/* PHYS_END */
+	__phys_offset = primary->start;	/* PHYS_OFFSET */
+	__phys_end = primary->end;	/* PHYS_END */
+
+	/* Split the primary region into two regions; code and data */
+	mem_regions[nr_io + i] = (struct mem_region){
+		.start = code_end,
+		.end = primary->end,
+		.flags = MR_F_PRIMARY,
+	};
+	*primary = (struct mem_region){
+		.start = primary->start,
+		.end = code_end,
+		.flags = MR_F_PRIMARY | MR_F_CODE,
+	};
 
-	phys_alloc_init(freemem_start, primary.end - freemem_start);
+	phys_alloc_init(freemem_start, __phys_end - freemem_start);
 	phys_alloc_set_minimum_alignment(SMP_CACHE_BYTES);
 
 	phys_alloc_get_unused(&base, &top);
Signature · 1 line
2.30.2
Reviewed-by on v2 5/8
#11 lore raw

Re: [PATCH kvm-unit-tests v2 5/8] arm/arm64: mmu: Remove memory layout assumptions

Hi Drew,

Quoted 134 lines · Andrew Jones

Rather than making too many assumptions about the memory layout in mmu code, just set up the page tables per the memory regions (which means putting all the memory layout assumptions in setup). To ensure we get the right default flags set we need to split the primary region into two regions for code and data.

We still only expect the primary regions to be present, but the next patch will remove that assumption too.

(Unfortunately we still have an assumption in setup_mmu. We assume
the range 3G-4G is available for the virtual memory allocator. We'll need to remove that assumption as well with another patch in order to support arbitrary memory maps.)

---
 lib/arm/asm/setup.h |  1 +
 lib/arm/mmu.c       | 26 +++++++++++++++-----------
 lib/arm/setup.c     | 29 +++++++++++++++++++++--------
 3 files changed, 37 insertions(+), 19 deletions(-)
lib/arm/asm/setup.h +1 -0
index c8afb2493f8d..210c14f818fb 100644
@@ -15,6 +15,7 @@ extern int nr_cpus;
 
 #define MR_F_PRIMARY		(1U << 0)
 #define MR_F_IO			(1U << 1)
+#define MR_F_CODE		(1U << 2)
 #define MR_F_UNKNOWN		(1U << 31)
 
 struct mem_region {
lib/arm/mmu.c +15 -11
index ee0c79142ba1..4e3cf37e33d0 100644
@@ -20,8 +20,6 @@
 
 #include <linux/compiler.h>
 
-extern unsigned long etext;
-
 pgd_t *mmu_idmap;
 
 /* CPU 0 starts with disabled MMU */
@@ -157,7 +155,7 @@ void mmu_set_range_sect(pgd_t *pgtable, uintptr_t virt_offset,
 
 void *setup_mmu(phys_addr_t phys_end)
 {
-	uintptr_t code_end = (uintptr_t)&etext;
+	struct mem_region *r;
 
 	/* 3G-4G region is reserved for vmalloc, cap phys_end at 3G */
 	if (phys_end > (3ul << 30))
@@ -173,14 +171,20 @@ void *setup_mmu(phys_addr_t phys_end)
 	if (!mmu_idmap)
 		mmu_idmap = alloc_page();
 
-	/* armv8 requires code shared between EL1 and EL0 to be read-only */
-	mmu_set_range_ptes(mmu_idmap, PHYS_OFFSET,
-		PHYS_OFFSET, code_end,
-		__pgprot(PTE_WBWA | PTE_RDONLY | PTE_USER));
-
-	mmu_set_range_ptes(mmu_idmap, code_end,
-		code_end, phys_end,
-		__pgprot(PTE_WBWA | PTE_USER));
+	for (r = mem_regions; r->end; ++r) {
+		if (r->flags & MR_F_IO) {
+			continue;
+		} else if (r->flags & MR_F_CODE) {
+			assert_msg(r->flags & MR_F_PRIMARY, "Unexpected code region");
+			/* armv8 requires code shared between EL1 and EL0 to be read-only */
+			mmu_set_range_ptes(mmu_idmap, r->start, r->start, r->end,
+					   __pgprot(PTE_WBWA | PTE_USER | PTE_RDONLY));
+		} else {
+			assert_msg(r->flags & MR_F_PRIMARY, "Unexpected data region");
+			mmu_set_range_ptes(mmu_idmap, r->start, r->start, r->end,
+					   __pgprot(PTE_WBWA | PTE_USER));
+		}
+	}
 
 	mmu_enable(mmu_idmap);
 	return mmu_idmap;
lib/arm/setup.c +5 -5
index 9c16f6004e9f..7db308b70744 100644
@@ -31,6 +31,7 @@
 #define NR_INITIAL_MEM_REGIONS 16
 
 extern unsigned long stacktop;
+extern unsigned long etext;
 
 struct timer_state __timer_state;
 
@@ -88,10 +89,12 @@ unsigned int mem_region_get_flags(phys_addr_t paddr)
 
 static void mem_init(phys_addr_t freemem_start)
 {
+	phys_addr_t code_end = (phys_addr_t)(unsigned long)&etext;
 	struct dt_pbus_reg regs[NR_INITIAL_MEM_REGIONS];
-	struct mem_region primary, mem = {
+	struct mem_region mem = {
 		.start = (phys_addr_t)-1,
 	};
+	struct mem_region *primary = NULL;
 	phys_addr_t base, top;
 	int nr_regs, nr_io = 0, i;
 
@@ -110,8 +113,6 @@ static void mem_init(phys_addr_t freemem_start)
 	nr_regs = dt_get_memory_params(regs, NR_INITIAL_MEM_REGIONS - nr_io);
 	assert(nr_regs > 0);
 
-	primary = (struct mem_region){ 0 };
-
 	for (i = 0; i < nr_regs; ++i) {
 		struct mem_region *r = &mem_regions[nr_io + i];
 
@@ -123,7 +124,7 @@ static void mem_init(phys_addr_t freemem_start)
 		 */
 		if (freemem_start >= r->start && freemem_start < r->end) {
 			r->flags |= MR_F_PRIMARY;
-			primary = *r;
+			primary = r;
 		}
 
 		/*
@@ -135,13 +136,25 @@ static void mem_init(phys_addr_t freemem_start)
 		if (r->end > mem.end)
 			mem.end = r->end;
 	}
-	assert(primary.end != 0);

lib/arm/setup.c static void mem_init(phys_addr_t freemem_start) · line 146

Quoted 10 lines
lib/arm/setup.c +6 -2
+	assert(primary);
 	assert(!(mem.start & ~PHYS_MASK) && !((mem.end - 1) & ~PHYS_MASK));
 
-	__phys_offset = primary.start;	/* PHYS_OFFSET */
-	__phys_end = primary.end;	/* PHYS_END */
+	__phys_offset = primary->start;	/* PHYS_OFFSET */
+	__phys_end = primary->end;	/* PHYS_END */
+
+	/* Split the primary region into two regions; code and data */
+	mem_regions[nr_io + i] = (struct mem_region){

Nitpick: you could change that to nr_io + nr_regs, to make it obsolutely obvious that the new region is appended to the array. This works because the static array is created with NR_INITIAL_MEM_REGIONS + 1 (but it does mean that we loose the empty region at the end if nr_io + nr_regs = NR_INITIAL_MEM_REGIONS).

Quoted 5 lines
lib/arm/setup.c +5 -0
+		.start = code_end,
+		.end = primary->end,
+		.flags = MR_F_PRIMARY,
+	};
+	*primary = (struct mem_region){

lib/arm/setup.c static void mem_init(phys_addr_t freemem_start) · line 160

Quoted 10 lines
lib/arm/setup.c +5 -1
+		.start = primary->start,
+		.end = code_end,
+		.flags = MR_F_PRIMARY | MR_F_CODE,
+	};
 
-	phys_alloc_init(freemem_start, primary.end - freemem_start);
+	phys_alloc_init(freemem_start, __phys_end - freemem_start);
 	phys_alloc_set_minimum_alignment(SMP_CACHE_BYTES);
 
 	phys_alloc_get_unused(&base, &top);

The nitpick is a matter of preference, and the patch looks really good now, it was a lot easier to figure out what is going on:

Thanks,

Alex

#12 lore raw

PATCH v2 6/8 arm/arm64: setup: Consolidate memory layout assumptions

In reply to Andrew Jones

Keep as much memory layout assumptions as possible in init::start and a single setup function. This prepares us for calling setup() from different start functions which have been linked with different linker scripts. To do this, stacktop is only referenced from init::start, making freemem_start a parameter to setup(). We also split mem_init() into three parts, one that populates the mem regions per the DT, one that populates the mem regions per assumptions, and one that does the mem init. The concept of a primary region is dropped, but we add a sanity check for the absence of memory holes, because we don't know how to deal with them yet.

---
 arm/cstart.S        |   4 +-
 arm/cstart64.S      |   2 +
 arm/flat.lds        |  23 ++++++
 lib/arm/asm/setup.h |   8 +-
 lib/arm/mmu.c       |   2 -
 lib/arm/setup.c     | 173 ++++++++++++++++++++++++--------------------
 6 files changed, 127 insertions(+), 85 deletions(-)
arm/cstart.S +3 -1
index bf3c78157e6a..446966de350d 100644
@@ -80,7 +80,9 @@ start:
 
 	/* complete setup */
 	pop	{r0-r1}
-	bl	setup
+	mov	r3, #0
+	ldr	r2, =stacktop		@ r2,r3 is the base of free memory
+	bl	setup			@ r0 is the addr of the dtb
 
 	/* run the test */
 	ldr	r0, =__argc
arm/cstart64.S +2 -0
index 27251fe8b5cd..42ba3a3ca249 100644
@@ -92,6 +92,8 @@ start:
 	bl	exceptions_init
 
 	/* complete setup */
+	adrp	x1, stacktop
+	add	x1, x1, :lo12:stacktop		// x1 is the base of free memory
 	bl	setup				// x0 is the addr of the dtb
 
 	/* run the test */
arm/flat.lds +23 -0
index 6ed377c0eaa0..6fb459efb815 100644
@@ -1,3 +1,26 @@
+/*
+ * init::start will pass stacktop to setup() as the base of free memory.
+ * setup() will then move the FDT and initrd to that base before calling
+ * mem_init(). With those movements and this linker script, we'll end up
+ * having the following memory layout:
+ *
+ *    +----------------------+   <-- top of physical memory
+ *    |                      |
+ *    ~                      ~
+ *    |                      |
+ *    +----------------------+   <-- top of initrd
+ *    |                      |
+ *    +----------------------+   <-- top of FDT
+ *    |                      |
+ *    +----------------------+   <-- top of cpu0's stack
+ *    |                      |
+ *    +----------------------+   <-- top of text/data/bss sections
+ *    |                      |
+ *    |                      |
+ *    +----------------------+   <-- load address
+ *    |                      |
+ *    +----------------------+   <-- physical address 0x0
+ */
 
 SECTIONS
 {
lib/arm/asm/setup.h +4 -4
index 210c14f818fb..f0e70b119fb0 100644
@@ -13,9 +13,8 @@
 extern u64 cpus[NR_CPUS];	/* per-cpu IDs (MPIDRs) */
 extern int nr_cpus;
 
-#define MR_F_PRIMARY		(1U << 0)
-#define MR_F_IO			(1U << 1)
-#define MR_F_CODE		(1U << 2)
+#define MR_F_IO			(1U << 0)
+#define MR_F_CODE		(1U << 1)
 #define MR_F_UNKNOWN		(1U << 31)
 
 struct mem_region {
@@ -26,6 +25,7 @@ struct mem_region {
 extern struct mem_region *mem_regions;
 extern phys_addr_t __phys_offset, __phys_end;
 
+extern struct mem_region *mem_region_find(phys_addr_t paddr);
 extern unsigned int mem_region_get_flags(phys_addr_t paddr);
 
 #define PHYS_OFFSET		(__phys_offset)
@@ -35,6 +35,6 @@ extern unsigned int mem_region_get_flags(phys_addr_t paddr);
 #define L1_CACHE_BYTES		(1 << L1_CACHE_SHIFT)
 #define SMP_CACHE_BYTES		L1_CACHE_BYTES
 
-void setup(const void *fdt);
+void setup(const void *fdt, phys_addr_t freemem_start);
 
 #endif /* _ASMARM_SETUP_H_ */
lib/arm/mmu.c +0 -2
index 4e3cf37e33d0..10df98e7a955 100644
@@ -175,12 +175,10 @@ void *setup_mmu(phys_addr_t phys_end)
 		if (r->flags & MR_F_IO) {
 			continue;
 		} else if (r->flags & MR_F_CODE) {
-			assert_msg(r->flags & MR_F_PRIMARY, "Unexpected code region");
 			/* armv8 requires code shared between EL1 and EL0 to be read-only */
 			mmu_set_range_ptes(mmu_idmap, r->start, r->start, r->end,
 					   __pgprot(PTE_WBWA | PTE_USER | PTE_RDONLY));
 		} else {
-			assert_msg(r->flags & MR_F_PRIMARY, "Unexpected data region");
 			mmu_set_range_ptes(mmu_idmap, r->start, r->start, r->end,
 					   __pgprot(PTE_WBWA | PTE_USER));
 		}
lib/arm/setup.c +95 -78

237 lines not on this page · raw message

Signature · 1 line
2.30.2
#13 lore raw

Re: [PATCH kvm-unit-tests v2 6/8] arm/arm64: setup: Consolidate memory layout assumptions

Quoted 275 lines · Andrew Jones

Keep as much memory layout assumptions as possible in init::start and a single setup function. This prepares us for calling setup() from different start functions which have been linked with different linker scripts. To do this, stacktop is only referenced from init::start, making freemem_start a parameter to setup(). We also split mem_init() into three parts, one that populates the mem regions per the DT, one that populates the mem regions per assumptions, and one that does the mem init. The concept of a primary region is dropped, but we add a sanity check for the absence of memory holes, because we don't know how to deal with them yet.

---
 arm/cstart.S        |   4 +-
 arm/cstart64.S      |   2 +
 arm/flat.lds        |  23 ++++++
 lib/arm/asm/setup.h |   8 +-
 lib/arm/mmu.c       |   2 -
 lib/arm/setup.c     | 173 ++++++++++++++++++++++++--------------------
 6 files changed, 127 insertions(+), 85 deletions(-)
arm/cstart.S +3 -1
index bf3c78157e6a..446966de350d 100644
@@ -80,7 +80,9 @@ start:
 
 	/* complete setup */
 	pop	{r0-r1}
-	bl	setup
+	mov	r3, #0
+	ldr	r2, =stacktop		@ r2,r3 is the base of free memory
+	bl	setup			@ r0 is the addr of the dtb
 
 	/* run the test */
 	ldr	r0, =__argc
arm/cstart64.S +2 -0
index 27251fe8b5cd..42ba3a3ca249 100644
@@ -92,6 +92,8 @@ start:
 	bl	exceptions_init
 
 	/* complete setup */
+	adrp	x1, stacktop
+	add	x1, x1, :lo12:stacktop		// x1 is the base of free memory
 	bl	setup				// x0 is the addr of the dtb
 
 	/* run the test */
arm/flat.lds +23 -0
index 6ed377c0eaa0..6fb459efb815 100644
@@ -1,3 +1,26 @@
+/*
+ * init::start will pass stacktop to setup() as the base of free memory.
+ * setup() will then move the FDT and initrd to that base before calling
+ * mem_init(). With those movements and this linker script, we'll end up
+ * having the following memory layout:
+ *
+ *    +----------------------+   <-- top of physical memory
+ *    |                      |
+ *    ~                      ~
+ *    |                      |
+ *    +----------------------+   <-- top of initrd
+ *    |                      |
+ *    +----------------------+   <-- top of FDT
+ *    |                      |
+ *    +----------------------+   <-- top of cpu0's stack
+ *    |                      |
+ *    +----------------------+   <-- top of text/data/bss sections
+ *    |                      |
+ *    |                      |
+ *    +----------------------+   <-- load address
+ *    |                      |
+ *    +----------------------+   <-- physical address 0x0
+ */
 
 SECTIONS
 {
lib/arm/asm/setup.h +4 -4
index 210c14f818fb..f0e70b119fb0 100644
@@ -13,9 +13,8 @@
 extern u64 cpus[NR_CPUS];	/* per-cpu IDs (MPIDRs) */
 extern int nr_cpus;
 
-#define MR_F_PRIMARY		(1U << 0)
-#define MR_F_IO			(1U << 1)
-#define MR_F_CODE		(1U << 2)
+#define MR_F_IO			(1U << 0)
+#define MR_F_CODE		(1U << 1)
 #define MR_F_UNKNOWN		(1U << 31)
 
 struct mem_region {
@@ -26,6 +25,7 @@ struct mem_region {
 extern struct mem_region *mem_regions;
 extern phys_addr_t __phys_offset, __phys_end;
 
+extern struct mem_region *mem_region_find(phys_addr_t paddr);
 extern unsigned int mem_region_get_flags(phys_addr_t paddr);
 
 #define PHYS_OFFSET		(__phys_offset)
@@ -35,6 +35,6 @@ extern unsigned int mem_region_get_flags(phys_addr_t paddr);
 #define L1_CACHE_BYTES		(1 << L1_CACHE_SHIFT)
 #define SMP_CACHE_BYTES		L1_CACHE_BYTES
 
-void setup(const void *fdt);
+void setup(const void *fdt, phys_addr_t freemem_start);
 
 #endif /* _ASMARM_SETUP_H_ */
lib/arm/mmu.c +0 -2
index 4e3cf37e33d0..10df98e7a955 100644
@@ -175,12 +175,10 @@ void *setup_mmu(phys_addr_t phys_end)
 		if (r->flags & MR_F_IO) {
 			continue;
 		} else if (r->flags & MR_F_CODE) {
-			assert_msg(r->flags & MR_F_PRIMARY, "Unexpected code region");
 			/* armv8 requires code shared between EL1 and EL0 to be read-only */
 			mmu_set_range_ptes(mmu_idmap, r->start, r->start, r->end,
 					   __pgprot(PTE_WBWA | PTE_USER | PTE_RDONLY));
 		} else {
-			assert_msg(r->flags & MR_F_PRIMARY, "Unexpected data region");
 			mmu_set_range_ptes(mmu_idmap, r->start, r->start, r->end,
 					   __pgprot(PTE_WBWA | PTE_USER));
 		}
lib/arm/setup.c +73 -33
index 7db308b70744..a5ebec3c5a12 100644
@@ -28,9 +28,10 @@
 
 #include "io.h"
 
-#define NR_INITIAL_MEM_REGIONS 16
+#define MAX_DT_MEM_REGIONS	16
+#define NR_EXTRA_MEM_REGIONS	16
+#define NR_INITIAL_MEM_REGIONS	(MAX_DT_MEM_REGIONS + NR_EXTRA_MEM_REGIONS)
 
-extern unsigned long stacktop;
 extern unsigned long etext;
 
 struct timer_state __timer_state;
@@ -75,28 +76,68 @@ static void cpu_init(void)
 	set_cpu_online(0, true);
 }
 
-unsigned int mem_region_get_flags(phys_addr_t paddr)
+static void mem_region_add(struct mem_region *r)
+{
+	struct mem_region *r_next = mem_regions;
+	int i = 0;
+
+	for (; r_next->end; ++r_next, ++i)
+		;
+	assert(i != NR_INITIAL_MEM_REGIONS);
+
+	*r_next = *r;
+}
+
+static void mem_regions_get_dt_regions(void)
+{
+	struct dt_pbus_reg regs[MAX_DT_MEM_REGIONS];
+	int nr_regs, i;
+
+	nr_regs = dt_get_memory_params(regs, MAX_DT_MEM_REGIONS);
+	assert(nr_regs > 0);
+
+	for (i = 0; i < nr_regs; ++i) {
+		mem_region_add(&(struct mem_region){
+			.start = regs[i].addr,
+			.end = regs[i].addr + regs[i].size,
+		});
+	}
+}
+
+struct mem_region *mem_region_find(phys_addr_t paddr)
 {
 	struct mem_region *r;
 
-	for (r = mem_regions; r->end; ++r) {
+	for (r = mem_regions; r->end; ++r)
 		if (paddr >= r->start && paddr < r->end)
-			return r->flags;
-	}
+			return r;
+	return NULL;
+}
 
-	return MR_F_UNKNOWN;
+unsigned int mem_region_get_flags(phys_addr_t paddr)
+{
+	struct mem_region *r = mem_region_find(paddr);
+	return r ? r->flags : MR_F_UNKNOWN;
 }
 
-static void mem_init(phys_addr_t freemem_start)
+static void mem_regions_add_assumed(void)
 {
 	phys_addr_t code_end = (phys_addr_t)(unsigned long)&etext;
-	struct dt_pbus_reg regs[NR_INITIAL_MEM_REGIONS];
-	struct mem_region mem = {
-		.start = (phys_addr_t)-1,
+	struct mem_region *r;
+
+	r = mem_region_find(code_end - 1);
+	assert(r);
+
+	/* Split the region with the code into two regions; code and data */
+	mem_region_add(&(struct mem_region){
+		.start = code_end,
+		.end = r->end,
+	});
+	*r = (struct mem_region){
+		.start = r->start,
+		.end = code_end,
+		.flags = MR_F_CODE,
 	};
-	struct mem_region *primary = NULL;
-	phys_addr_t base, top;
-	int nr_regs, nr_io = 0, i;
 
 	/*
 	 * mach-virt I/O regions:
@@ -104,57 +145,47 @@ static void mem_init(phys_addr_t freemem_start)
 	 *   - 512M at 256G (arm64, arm uses highmem=off)
 	 *   - 512G at 512G (arm64, arm uses highmem=off)
 	 */
-	mem_regions[nr_io++] = (struct mem_region){ 0, (1ul << 30), MR_F_IO };
+	mem_region_add(&(struct mem_region){ 0, (1ul << 30), MR_F_IO });
 #ifdef __aarch64__
-	mem_regions[nr_io++] = (struct mem_region){ (1ul << 38), (1ul << 38) | (1ul << 29), MR_F_IO };
-	mem_regions[nr_io++] = (struct mem_region){ (1ul << 39), (1ul << 40), MR_F_IO };
+	mem_region_add(&(struct mem_region){ (1ul << 38), (1ul << 38) | (1ul << 29), MR_F_IO });
+	mem_region_add(&(struct mem_region){ (1ul << 39), (1ul << 40), MR_F_IO });
 #endif
+}
 
-	nr_regs = dt_get_memory_params(regs, NR_INITIAL_MEM_REGIONS - nr_io);
-	assert(nr_regs > 0);
-
-	for (i = 0; i < nr_regs; ++i) {
-		struct mem_region *r = &mem_regions[nr_io + i];
+static void mem_init(phys_addr_t freemem_start)
+{
+	phys_addr_t base, top;
+	struct mem_region *freemem, *r, mem = {
+		.start = (phys_addr_t)-1,
+	};
 
-		r->start = regs[i].addr;
-		r->end = regs[i].addr + regs[i].size;
+	freemem = mem_region_find(freemem_start);
+	assert(freemem && !(freemem->flags & (MR_F_IO | MR_F_CODE)));
 
-		/*
-		 * pick the region we're in for our primary region
-		 */
-		if (freemem_start >= r->start && freemem_start < r->end) {
-			r->flags |= MR_F_PRIMARY;
-			primary = r;
+	for (r = mem_regions; r->end; ++r) {
+		if (!(r->flags & MR_F_IO)) {
+			if (r->start < mem.start)
+				mem.start = r->start;
+			if (r->end > mem.end)
+				mem.end = r->end;
 		}
-
-		/*
-		 * set the lowest and highest addresses found,

lib/arm/setup.c static void mem_init(phys_addr_t freemem_start) · line 174

Quoted 10 lines
lib/arm/setup.c +1 -7
-		 * ignoring potential gaps
-		 */
-		if (r->start < mem.start)
-			mem.start = r->start;
-		if (r->end > mem.end)
-			mem.end = r->end;
 	}
-	assert(primary);
+	assert(mem.end);
 	assert(!(mem.start & ~PHYS_MASK) && !((mem.end - 1) & ~PHYS_MASK));

Eh, I promised Alex not to do this, but then didn't correct it quite right. This should be

  assert(!(mem.start & ~PHYS_MASK));
  if ((mem.end - 1) & ~PHYS_MASK)
     mem.end &= PHYS_MASK;
Quoted 82 lines
lib/arm/setup.c +18 -38
-	__phys_offset = primary->start;	/* PHYS_OFFSET */
-	__phys_end = primary->end;	/* PHYS_END */
+	/* Check for holes */
+	r = mem_region_find(mem.start);
+	while (r && r->end != mem.end)
+		r = mem_region_find(r->end);
+	assert(r);
 
-	/* Split the primary region into two regions; code and data */
-	mem_regions[nr_io + i] = (struct mem_region){
-		.start = code_end,
-		.end = primary->end,
-		.flags = MR_F_PRIMARY,
-	};
-	*primary = (struct mem_region){
-		.start = primary->start,
-		.end = code_end,
-		.flags = MR_F_PRIMARY | MR_F_CODE,
-	};
+	/* Ensure our selected freemem region is somewhere in our full range */
+	assert(freemem_start >= mem.start && freemem->end <= mem.end);
 
-	phys_alloc_init(freemem_start, __phys_end - freemem_start);
+	__phys_offset = mem.start;	/* PHYS_OFFSET */
+	__phys_end = mem.end;		/* PHYS_END */
+
+	phys_alloc_init(freemem_start, freemem->end - freemem_start);
 	phys_alloc_set_minimum_alignment(SMP_CACHE_BYTES);
 
 	phys_alloc_get_unused(&base, &top);
@@ -204,35 +235,17 @@ static void timer_save_state(void)
 	__timer_state.vtimer.irq_flags = fdt32_to_cpu(data[8]);
 }
 
-void setup(const void *fdt)
+void setup(const void *fdt, phys_addr_t freemem_start)
 {
-	void *freemem = &stacktop;
+	void *freemem;
 	const char *bootargs, *tmp;
 	u32 fdt_size;
 	int ret;
 
-	/*
-	 * Before calling mem_init we need to move the fdt and initrd
-	 * to safe locations. We move them to construct the memory
-	 * map illustrated below:
-	 *
-	 *    +----------------------+   <-- top of physical memory
-	 *    |                      |
-	 *    ~                      ~
-	 *    |                      |
-	 *    +----------------------+   <-- top of initrd
-	 *    |                      |
-	 *    +----------------------+   <-- top of FDT
-	 *    |                      |
-	 *    +----------------------+   <-- top of cpu0's stack
-	 *    |                      |
-	 *    +----------------------+   <-- top of text/data/bss sections,
-	 *    |                      |       see arm/flat.lds
-	 *    |                      |
-	 *    +----------------------+   <-- load address
-	 *    |                      |
-	 *    +----------------------+
-	 */
+	assert(sizeof(long) == 8 || freemem_start < (3ul << 30));
+	freemem = (void *)(unsigned long)freemem_start;
+
+	/* Move the FDT to the base of free memory */
 	fdt_size = fdt_totalsize(fdt);
 	ret = fdt_move(fdt, freemem, fdt_size);
 	assert(ret == 0);
@@ -240,6 +253,7 @@ void setup(const void *fdt)
 	assert(ret == 0);
 	freemem += fdt_size;
 
+	/* Move the initrd to the top of the FDT */
 	ret = dt_get_initrd(&tmp, &initrd_size);
 	assert(ret == 0 || ret == -FDT_ERR_NOTFOUND);
 	if (ret == 0) {

lib/arm/setup.c void setup(const void *fdt) · line 265

Quoted 4 lines
lib/arm/setup.c +1 -0
@@ -248,7 +262,10 @@ void setup(const void *fdt)
 		freemem += initrd_size;
 	}
 
+	mem_regions_get_dt_regions();

Since I have to respin for at least the mem.end assert, I think I'll rename this to mem_regions_add_dt_regions at the same time.

Quoted 9 lines
lib/arm/setup.c +2 -0
+	mem_regions_add_assumed();
 	mem_init(PAGE_ALIGN((unsigned long)freemem));
+
 	cpu_init();
 
 	/* cpu_init must be called before thread_info_init */
Signature · 1 line
2.30.2

Thanks,
drew

#14 lore raw

Re: [PATCH kvm-unit-tests v2 6/8] arm/arm64: setup: Consolidate memory layout assumptions

Quoted 10 lines · Andrew Jones

lib/arm/setup.c static void mem_init(phys_addr_t freemem_start) · line 174

Quoted 2 lines · Andrew Jones
lib/arm/setup.c +1 -0
+	assert(mem.end);
 	assert(!(mem.start & ~PHYS_MASK) && !((mem.end - 1) & ~PHYS_MASK));

Eh, I promised Alex not to do this, but then didn't correct it quite right. This should be

  assert(!(mem.start & ~PHYS_MASK));
  if ((mem.end - 1) & ~PHYS_MASK)
     mem.end &= PHYS_MASK;

I've changed this to

  assert(mem.end && !(mem.start & ~PHYS_MASK));
  mem.end &= PHYS_MASK;

for v3.

Thanks,
drew

#15 lore raw

Re: [PATCH kvm-unit-tests v2 6/8] arm/arm64: setup: Consolidate memory layout assumptions

Hi Drew,

Quoted 10 lines · Andrew Jones
Quoted 9 lines · Andrew Jones

lib/arm/setup.c static void mem_init(phys_addr_t freemem_start) · line 174

Quoted 2 lines · Andrew Jones
lib/arm/setup.c +1 -0
+	assert(mem.end);
 	assert(!(mem.start & ~PHYS_MASK) && !((mem.end - 1) & ~PHYS_MASK));

Eh, I promised Alex not to do this, but then didn't correct it quite right. This should be

  assert(!(mem.start & ~PHYS_MASK));
  if ((mem.end - 1) & ~PHYS_MASK)
     mem.end &= PHYS_MASK;
Quoted 6 lines

I've changed this to

  assert(mem.end && !(mem.start & ~PHYS_MASK));
  mem.end &= PHYS_MASK;

for v3.

Looks good,

Thanks,

Alex

#16 lore raw

Re: [PATCH kvm-unit-tests v2 6/8] arm/arm64: setup: Consolidate memory layout assumptions

In reply to Andrew Jones

Hi Drew,

Quoted 150 lines · Andrew Jones

Keep as much memory layout assumptions as possible in init::start and a single setup function. This prepares us for calling setup() from different start functions which have been linked with different linker scripts. To do this, stacktop is only referenced from init::start, making freemem_start a parameter to setup(). We also split mem_init() into three parts, one that populates the mem regions per the DT, one that populates the mem regions per assumptions, and one that does the mem init. The concept of a primary region is dropped, but we add a sanity check for the absence of memory holes, because we don't know how to deal with them yet.

---
 arm/cstart.S        |   4 +-
 arm/cstart64.S      |   2 +
 arm/flat.lds        |  23 ++++++
 lib/arm/asm/setup.h |   8 +-
 lib/arm/mmu.c       |   2 -
 lib/arm/setup.c     | 173 ++++++++++++++++++++++++--------------------
 6 files changed, 127 insertions(+), 85 deletions(-)
arm/cstart.S +3 -1
index bf3c78157e6a..446966de350d 100644
@@ -80,7 +80,9 @@ start:
 
 	/* complete setup */
 	pop	{r0-r1}
-	bl	setup
+	mov	r3, #0
+	ldr	r2, =stacktop		@ r2,r3 is the base of free memory
+	bl	setup			@ r0 is the addr of the dtb
 
 	/* run the test */
 	ldr	r0, =__argc
arm/cstart64.S +2 -0
index 27251fe8b5cd..42ba3a3ca249 100644
@@ -92,6 +92,8 @@ start:
 	bl	exceptions_init
 
 	/* complete setup */
+	adrp	x1, stacktop
+	add	x1, x1, :lo12:stacktop		// x1 is the base of free memory
 	bl	setup				// x0 is the addr of the dtb
 
 	/* run the test */
arm/flat.lds +23 -0
index 6ed377c0eaa0..6fb459efb815 100644
@@ -1,3 +1,26 @@
+/*
+ * init::start will pass stacktop to setup() as the base of free memory.
+ * setup() will then move the FDT and initrd to that base before calling
+ * mem_init(). With those movements and this linker script, we'll end up
+ * having the following memory layout:
+ *
+ *    +----------------------+   <-- top of physical memory
+ *    |                      |
+ *    ~                      ~
+ *    |                      |
+ *    +----------------------+   <-- top of initrd
+ *    |                      |
+ *    +----------------------+   <-- top of FDT
+ *    |                      |
+ *    +----------------------+   <-- top of cpu0's stack
+ *    |                      |
+ *    +----------------------+   <-- top of text/data/bss sections
+ *    |                      |
+ *    |                      |
+ *    +----------------------+   <-- load address
+ *    |                      |
+ *    +----------------------+   <-- physical address 0x0
+ */
 
 SECTIONS
 {
lib/arm/asm/setup.h +4 -4
index 210c14f818fb..f0e70b119fb0 100644
@@ -13,9 +13,8 @@
 extern u64 cpus[NR_CPUS];	/* per-cpu IDs (MPIDRs) */
 extern int nr_cpus;
 
-#define MR_F_PRIMARY		(1U << 0)
-#define MR_F_IO			(1U << 1)
-#define MR_F_CODE		(1U << 2)
+#define MR_F_IO			(1U << 0)
+#define MR_F_CODE		(1U << 1)
 #define MR_F_UNKNOWN		(1U << 31)
 
 struct mem_region {
@@ -26,6 +25,7 @@ struct mem_region {
 extern struct mem_region *mem_regions;
 extern phys_addr_t __phys_offset, __phys_end;
 
+extern struct mem_region *mem_region_find(phys_addr_t paddr);
 extern unsigned int mem_region_get_flags(phys_addr_t paddr);
 
 #define PHYS_OFFSET		(__phys_offset)
@@ -35,6 +35,6 @@ extern unsigned int mem_region_get_flags(phys_addr_t paddr);
 #define L1_CACHE_BYTES		(1 << L1_CACHE_SHIFT)
 #define SMP_CACHE_BYTES		L1_CACHE_BYTES
 
-void setup(const void *fdt);
+void setup(const void *fdt, phys_addr_t freemem_start);
 
 #endif /* _ASMARM_SETUP_H_ */
lib/arm/mmu.c +0 -2
index 4e3cf37e33d0..10df98e7a955 100644
@@ -175,12 +175,10 @@ void *setup_mmu(phys_addr_t phys_end)
 		if (r->flags & MR_F_IO) {
 			continue;
 		} else if (r->flags & MR_F_CODE) {
-			assert_msg(r->flags & MR_F_PRIMARY, "Unexpected code region");
 			/* armv8 requires code shared between EL1 and EL0 to be read-only */
 			mmu_set_range_ptes(mmu_idmap, r->start, r->start, r->end,
 					   __pgprot(PTE_WBWA | PTE_USER | PTE_RDONLY));
 		} else {
-			assert_msg(r->flags & MR_F_PRIMARY, "Unexpected data region");
 			mmu_set_range_ptes(mmu_idmap, r->start, r->start, r->end,
 					   __pgprot(PTE_WBWA | PTE_USER));
 		}
lib/arm/setup.c +3 -2
index 7db308b70744..a5ebec3c5a12 100644
@@ -28,9 +28,10 @@
 
 #include "io.h"
 
-#define NR_INITIAL_MEM_REGIONS 16
+#define MAX_DT_MEM_REGIONS	16
+#define NR_EXTRA_MEM_REGIONS	16
+#define NR_INITIAL_MEM_REGIONS	(MAX_DT_MEM_REGIONS + NR_EXTRA_MEM_REGIONS)
 
-extern unsigned long stacktop;
 extern unsigned long etext;
 
 struct timer_state __timer_state;
@@ -75,28 +76,68 @@ static void cpu_init(void)
 	set_cpu_online(0, true);
 }

lib/arm/setup.c static void cpu_init(void) · line 86

Quoted 10 lines
lib/arm/setup.c +8 -1
 
-unsigned int mem_region_get_flags(phys_addr_t paddr)
+static void mem_region_add(struct mem_region *r)
+{
+	struct mem_region *r_next = mem_regions;
+	int i = 0;
+
+	for (; r_next->end; ++r_next, ++i)
+		;
+	assert(i != NR_INITIAL_MEM_REGIONS);

Shouldn't that be i < NR_INITIAL_MEM_REGIONS? I think it conveys intention better, and also helps catch situations where we've set in other parts of the code the entire mem_regions array, from index 0 to index NR_INITIAL_MEM_REGIONS (at least), which is an error.

Quoted 123 lines
lib/arm/setup.c +62 -37
+
+	*r_next = *r;
+}
+
+static void mem_regions_get_dt_regions(void)
+{
+	struct dt_pbus_reg regs[MAX_DT_MEM_REGIONS];
+	int nr_regs, i;
+
+	nr_regs = dt_get_memory_params(regs, MAX_DT_MEM_REGIONS);
+	assert(nr_regs > 0);
+
+	for (i = 0; i < nr_regs; ++i) {
+		mem_region_add(&(struct mem_region){
+			.start = regs[i].addr,
+			.end = regs[i].addr + regs[i].size,
+		});
+	}
+}
+
+struct mem_region *mem_region_find(phys_addr_t paddr)
 {
 	struct mem_region *r;
 
-	for (r = mem_regions; r->end; ++r) {
+	for (r = mem_regions; r->end; ++r)
 		if (paddr >= r->start && paddr < r->end)
-			return r->flags;
-	}
+			return r;
+	return NULL;
+}
 
-	return MR_F_UNKNOWN;
+unsigned int mem_region_get_flags(phys_addr_t paddr)
+{
+	struct mem_region *r = mem_region_find(paddr);
+	return r ? r->flags : MR_F_UNKNOWN;
 }
 
-static void mem_init(phys_addr_t freemem_start)
+static void mem_regions_add_assumed(void)
 {
 	phys_addr_t code_end = (phys_addr_t)(unsigned long)&etext;
-	struct dt_pbus_reg regs[NR_INITIAL_MEM_REGIONS];
-	struct mem_region mem = {
-		.start = (phys_addr_t)-1,
+	struct mem_region *r;
+
+	r = mem_region_find(code_end - 1);
+	assert(r);
+
+	/* Split the region with the code into two regions; code and data */
+	mem_region_add(&(struct mem_region){
+		.start = code_end,
+		.end = r->end,
+	});
+	*r = (struct mem_region){
+		.start = r->start,
+		.end = code_end,
+		.flags = MR_F_CODE,
 	};
-	struct mem_region *primary = NULL;
-	phys_addr_t base, top;
-	int nr_regs, nr_io = 0, i;
 
 	/*
 	 * mach-virt I/O regions:
@@ -104,57 +145,47 @@ static void mem_init(phys_addr_t freemem_start)
 	 *   - 512M at 256G (arm64, arm uses highmem=off)
 	 *   - 512G at 512G (arm64, arm uses highmem=off)
 	 */
-	mem_regions[nr_io++] = (struct mem_region){ 0, (1ul << 30), MR_F_IO };
+	mem_region_add(&(struct mem_region){ 0, (1ul << 30), MR_F_IO });
 #ifdef __aarch64__
-	mem_regions[nr_io++] = (struct mem_region){ (1ul << 38), (1ul << 38) | (1ul << 29), MR_F_IO };
-	mem_regions[nr_io++] = (struct mem_region){ (1ul << 39), (1ul << 40), MR_F_IO };
+	mem_region_add(&(struct mem_region){ (1ul << 38), (1ul << 38) | (1ul << 29), MR_F_IO });
+	mem_region_add(&(struct mem_region){ (1ul << 39), (1ul << 40), MR_F_IO });
 #endif
+}
 
-	nr_regs = dt_get_memory_params(regs, NR_INITIAL_MEM_REGIONS - nr_io);
-	assert(nr_regs > 0);
-
-	for (i = 0; i < nr_regs; ++i) {
-		struct mem_region *r = &mem_regions[nr_io + i];
+static void mem_init(phys_addr_t freemem_start)
+{
+	phys_addr_t base, top;
+	struct mem_region *freemem, *r, mem = {
+		.start = (phys_addr_t)-1,
+	};
 
-		r->start = regs[i].addr;
-		r->end = regs[i].addr + regs[i].size;
+	freemem = mem_region_find(freemem_start);
+	assert(freemem && !(freemem->flags & (MR_F_IO | MR_F_CODE)));
 
-		/*
-		 * pick the region we're in for our primary region
-		 */
-		if (freemem_start >= r->start && freemem_start < r->end) {
-			r->flags |= MR_F_PRIMARY;
-			primary = r;
+	for (r = mem_regions; r->end; ++r) {
+		if (!(r->flags & MR_F_IO)) {
+			if (r->start < mem.start)
+				mem.start = r->start;
+			if (r->end > mem.end)
+				mem.end = r->end;
 		}
-
-		/*
-		 * set the lowest and highest addresses found,
-		 * ignoring potential gaps
-		 */
-		if (r->start < mem.start)
-			mem.start = r->start;
-		if (r->end > mem.end)
-			mem.end = r->end;
 	}
-	assert(primary);

lib/arm/setup.c static void mem_init(phys_addr_t freemem_start) · line 180

Quoted 10 lines
lib/arm/setup.c +6 -2
+	assert(mem.end);
 	assert(!(mem.start & ~PHYS_MASK) && !((mem.end - 1) & ~PHYS_MASK));
 
-	__phys_offset = primary->start;	/* PHYS_OFFSET */
-	__phys_end = primary->end;	/* PHYS_END */
+	/* Check for holes */
+	r = mem_region_find(mem.start);
+	while (r && r->end != mem.end)
+		r = mem_region_find(r->end);
+	assert(r);

It took me a while to figure this out, what it does is find all memory regions adjacent to each other and starting at mem.start and ending precisely at mem.end. Looks good.

Quoted 4 lines
lib/arm/setup.c +0 -3
-	/* Split the primary region into two regions; code and data */
-	mem_regions[nr_io + i] = (struct mem_region){
-		.start = code_end,

lib/arm/setup.c static void mem_init(phys_addr_t freemem_start) · line 183

Quoted 10 lines
lib/arm/setup.c +2 -8
-		.end = primary->end,
-		.flags = MR_F_PRIMARY,
-	};
-	*primary = (struct mem_region){
-		.start = primary->start,
-		.end = code_end,
-		.flags = MR_F_PRIMARY | MR_F_CODE,
-	};
+	/* Ensure our selected freemem region is somewhere in our full range */
+	assert(freemem_start >= mem.start && freemem->end <= mem.end);

That looks a bit strange, the comment refers to checking the freemem region, but we're actually checking is the subregion from freemem_start to freemem->end. I suppose it's because freemem_start comes after data, stack, fdt and initrd in the freemem region, and what we consider available memory is [freemem_start, freemem->end), right?

Thanks,

Alex

Quoted 70 lines
lib/arm/setup.c +14 -25
-	phys_alloc_init(freemem_start, __phys_end - freemem_start);
+	__phys_offset = mem.start;	/* PHYS_OFFSET */
+	__phys_end = mem.end;		/* PHYS_END */
+
+	phys_alloc_init(freemem_start, freemem->end - freemem_start);
 	phys_alloc_set_minimum_alignment(SMP_CACHE_BYTES);
 
 	phys_alloc_get_unused(&base, &top);
@@ -204,35 +235,17 @@ static void timer_save_state(void)
 	__timer_state.vtimer.irq_flags = fdt32_to_cpu(data[8]);
 }
 
-void setup(const void *fdt)
+void setup(const void *fdt, phys_addr_t freemem_start)
 {
-	void *freemem = &stacktop;
+	void *freemem;
 	const char *bootargs, *tmp;
 	u32 fdt_size;
 	int ret;
 
-	/*
-	 * Before calling mem_init we need to move the fdt and initrd
-	 * to safe locations. We move them to construct the memory
-	 * map illustrated below:
-	 *
-	 *    +----------------------+   <-- top of physical memory
-	 *    |                      |
-	 *    ~                      ~
-	 *    |                      |
-	 *    +----------------------+   <-- top of initrd
-	 *    |                      |
-	 *    +----------------------+   <-- top of FDT
-	 *    |                      |
-	 *    +----------------------+   <-- top of cpu0's stack
-	 *    |                      |
-	 *    +----------------------+   <-- top of text/data/bss sections,
-	 *    |                      |       see arm/flat.lds
-	 *    |                      |
-	 *    +----------------------+   <-- load address
-	 *    |                      |
-	 *    +----------------------+
-	 */
+	assert(sizeof(long) == 8 || freemem_start < (3ul << 30));
+	freemem = (void *)(unsigned long)freemem_start;
+
+	/* Move the FDT to the base of free memory */
 	fdt_size = fdt_totalsize(fdt);
 	ret = fdt_move(fdt, freemem, fdt_size);
 	assert(ret == 0);
@@ -240,6 +253,7 @@ void setup(const void *fdt)
 	assert(ret == 0);
 	freemem += fdt_size;
 
+	/* Move the initrd to the top of the FDT */
 	ret = dt_get_initrd(&tmp, &initrd_size);
 	assert(ret == 0 || ret == -FDT_ERR_NOTFOUND);
 	if (ret == 0) {
@@ -248,7 +262,10 @@ void setup(const void *fdt)
 		freemem += initrd_size;
 	}
 
+	mem_regions_get_dt_regions();
+	mem_regions_add_assumed();
 	mem_init(PAGE_ALIGN((unsigned long)freemem));
+
 	cpu_init();
 
 	/* cpu_init must be called before thread_info_init */
#17 lore raw

Re: [PATCH kvm-unit-tests v2 6/8] arm/arm64: setup: Consolidate memory layout assumptions

Quoted 154 lines · Alexandru Elisei

Hi Drew,

Quoted 151 lines · Andrew Jones

Keep as much memory layout assumptions as possible in init::start and a single setup function. This prepares us for calling setup() from different start functions which have been linked with different linker scripts. To do this, stacktop is only referenced from init::start, making freemem_start a parameter to setup(). We also split mem_init() into three parts, one that populates the mem regions per the DT, one that populates the mem regions per assumptions, and one that does the mem init. The concept of a primary region is dropped, but we add a sanity check for the absence of memory holes, because we don't know how to deal with them yet.

---
 arm/cstart.S        |   4 +-
 arm/cstart64.S      |   2 +
 arm/flat.lds        |  23 ++++++
 lib/arm/asm/setup.h |   8 +-
 lib/arm/mmu.c       |   2 -
 lib/arm/setup.c     | 173 ++++++++++++++++++++++++--------------------
 6 files changed, 127 insertions(+), 85 deletions(-)
arm/cstart.S +3 -1
index bf3c78157e6a..446966de350d 100644
@@ -80,7 +80,9 @@ start:
 
 	/* complete setup */
 	pop	{r0-r1}
-	bl	setup
+	mov	r3, #0
+	ldr	r2, =stacktop		@ r2,r3 is the base of free memory
+	bl	setup			@ r0 is the addr of the dtb
 
 	/* run the test */
 	ldr	r0, =__argc
arm/cstart64.S +2 -0
index 27251fe8b5cd..42ba3a3ca249 100644
@@ -92,6 +92,8 @@ start:
 	bl	exceptions_init
 
 	/* complete setup */
+	adrp	x1, stacktop
+	add	x1, x1, :lo12:stacktop		// x1 is the base of free memory
 	bl	setup				// x0 is the addr of the dtb
 
 	/* run the test */
arm/flat.lds +23 -0
index 6ed377c0eaa0..6fb459efb815 100644
@@ -1,3 +1,26 @@
+/*
+ * init::start will pass stacktop to setup() as the base of free memory.
+ * setup() will then move the FDT and initrd to that base before calling
+ * mem_init(). With those movements and this linker script, we'll end up
+ * having the following memory layout:
+ *
+ *    +----------------------+   <-- top of physical memory
+ *    |                      |
+ *    ~                      ~
+ *    |                      |
+ *    +----------------------+   <-- top of initrd
+ *    |                      |
+ *    +----------------------+   <-- top of FDT
+ *    |                      |
+ *    +----------------------+   <-- top of cpu0's stack
+ *    |                      |
+ *    +----------------------+   <-- top of text/data/bss sections
+ *    |                      |
+ *    |                      |
+ *    +----------------------+   <-- load address
+ *    |                      |
+ *    +----------------------+   <-- physical address 0x0
+ */
 
 SECTIONS
 {
lib/arm/asm/setup.h +4 -4
index 210c14f818fb..f0e70b119fb0 100644
@@ -13,9 +13,8 @@
 extern u64 cpus[NR_CPUS];	/* per-cpu IDs (MPIDRs) */
 extern int nr_cpus;
 
-#define MR_F_PRIMARY		(1U << 0)
-#define MR_F_IO			(1U << 1)
-#define MR_F_CODE		(1U << 2)
+#define MR_F_IO			(1U << 0)
+#define MR_F_CODE		(1U << 1)
 #define MR_F_UNKNOWN		(1U << 31)
 
 struct mem_region {
@@ -26,6 +25,7 @@ struct mem_region {
 extern struct mem_region *mem_regions;
 extern phys_addr_t __phys_offset, __phys_end;
 
+extern struct mem_region *mem_region_find(phys_addr_t paddr);
 extern unsigned int mem_region_get_flags(phys_addr_t paddr);
 
 #define PHYS_OFFSET		(__phys_offset)
@@ -35,6 +35,6 @@ extern unsigned int mem_region_get_flags(phys_addr_t paddr);
 #define L1_CACHE_BYTES		(1 << L1_CACHE_SHIFT)
 #define SMP_CACHE_BYTES		L1_CACHE_BYTES
 
-void setup(const void *fdt);
+void setup(const void *fdt, phys_addr_t freemem_start);
 
 #endif /* _ASMARM_SETUP_H_ */
lib/arm/mmu.c +0 -2
index 4e3cf37e33d0..10df98e7a955 100644
@@ -175,12 +175,10 @@ void *setup_mmu(phys_addr_t phys_end)
 		if (r->flags & MR_F_IO) {
 			continue;
 		} else if (r->flags & MR_F_CODE) {
-			assert_msg(r->flags & MR_F_PRIMARY, "Unexpected code region");
 			/* armv8 requires code shared between EL1 and EL0 to be read-only */
 			mmu_set_range_ptes(mmu_idmap, r->start, r->start, r->end,
 					   __pgprot(PTE_WBWA | PTE_USER | PTE_RDONLY));
 		} else {
-			assert_msg(r->flags & MR_F_PRIMARY, "Unexpected data region");
 			mmu_set_range_ptes(mmu_idmap, r->start, r->start, r->end,
 					   __pgprot(PTE_WBWA | PTE_USER));
 		}
lib/arm/setup.c +3 -2
index 7db308b70744..a5ebec3c5a12 100644
@@ -28,9 +28,10 @@
 
 #include "io.h"
 
-#define NR_INITIAL_MEM_REGIONS 16
+#define MAX_DT_MEM_REGIONS	16
+#define NR_EXTRA_MEM_REGIONS	16
+#define NR_INITIAL_MEM_REGIONS	(MAX_DT_MEM_REGIONS + NR_EXTRA_MEM_REGIONS)
 
-extern unsigned long stacktop;
 extern unsigned long etext;
 
 struct timer_state __timer_state;
@@ -75,28 +76,68 @@ static void cpu_init(void)
 	set_cpu_online(0, true);
 }
 
Quoted 11 lines

lib/arm/setup.c static void cpu_init(void) · line 86

Quoted 9 lines
lib/arm/setup.c +8 -1
-unsigned int mem_region_get_flags(phys_addr_t paddr)
+static void mem_region_add(struct mem_region *r)
+{
+	struct mem_region *r_next = mem_regions;
+	int i = 0;
+
+	for (; r_next->end; ++r_next, ++i)
+		;
+	assert(i != NR_INITIAL_MEM_REGIONS);

Shouldn't that be i < NR_INITIAL_MEM_REGIONS?

Indeed. Will fix.

Quoted 155 lines

I think it conveys intention better,
and also helps catch situations where we've set in other parts of the code the entire mem_regions array, from index 0 to index NR_INITIAL_MEM_REGIONS (at least), which is an error.

Quoted 123 lines
lib/arm/setup.c +62 -37
+
+	*r_next = *r;
+}
+
+static void mem_regions_get_dt_regions(void)
+{
+	struct dt_pbus_reg regs[MAX_DT_MEM_REGIONS];
+	int nr_regs, i;
+
+	nr_regs = dt_get_memory_params(regs, MAX_DT_MEM_REGIONS);
+	assert(nr_regs > 0);
+
+	for (i = 0; i < nr_regs; ++i) {
+		mem_region_add(&(struct mem_region){
+			.start = regs[i].addr,
+			.end = regs[i].addr + regs[i].size,
+		});
+	}
+}
+
+struct mem_region *mem_region_find(phys_addr_t paddr)
 {
 	struct mem_region *r;
 
-	for (r = mem_regions; r->end; ++r) {
+	for (r = mem_regions; r->end; ++r)
 		if (paddr >= r->start && paddr < r->end)
-			return r->flags;
-	}
+			return r;
+	return NULL;
+}
 
-	return MR_F_UNKNOWN;
+unsigned int mem_region_get_flags(phys_addr_t paddr)
+{
+	struct mem_region *r = mem_region_find(paddr);
+	return r ? r->flags : MR_F_UNKNOWN;
 }
 
-static void mem_init(phys_addr_t freemem_start)
+static void mem_regions_add_assumed(void)
 {
 	phys_addr_t code_end = (phys_addr_t)(unsigned long)&etext;
-	struct dt_pbus_reg regs[NR_INITIAL_MEM_REGIONS];
-	struct mem_region mem = {
-		.start = (phys_addr_t)-1,
+	struct mem_region *r;
+
+	r = mem_region_find(code_end - 1);
+	assert(r);
+
+	/* Split the region with the code into two regions; code and data */
+	mem_region_add(&(struct mem_region){
+		.start = code_end,
+		.end = r->end,
+	});
+	*r = (struct mem_region){
+		.start = r->start,
+		.end = code_end,
+		.flags = MR_F_CODE,
 	};
-	struct mem_region *primary = NULL;
-	phys_addr_t base, top;
-	int nr_regs, nr_io = 0, i;
 
 	/*
 	 * mach-virt I/O regions:
@@ -104,57 +145,47 @@ static void mem_init(phys_addr_t freemem_start)
 	 *   - 512M at 256G (arm64, arm uses highmem=off)
 	 *   - 512G at 512G (arm64, arm uses highmem=off)
 	 */
-	mem_regions[nr_io++] = (struct mem_region){ 0, (1ul << 30), MR_F_IO };
+	mem_region_add(&(struct mem_region){ 0, (1ul << 30), MR_F_IO });
 #ifdef __aarch64__
-	mem_regions[nr_io++] = (struct mem_region){ (1ul << 38), (1ul << 38) | (1ul << 29), MR_F_IO };
-	mem_regions[nr_io++] = (struct mem_region){ (1ul << 39), (1ul << 40), MR_F_IO };
+	mem_region_add(&(struct mem_region){ (1ul << 38), (1ul << 38) | (1ul << 29), MR_F_IO });
+	mem_region_add(&(struct mem_region){ (1ul << 39), (1ul << 40), MR_F_IO });
 #endif
+}
 
-	nr_regs = dt_get_memory_params(regs, NR_INITIAL_MEM_REGIONS - nr_io);
-	assert(nr_regs > 0);
-
-	for (i = 0; i < nr_regs; ++i) {
-		struct mem_region *r = &mem_regions[nr_io + i];
+static void mem_init(phys_addr_t freemem_start)
+{
+	phys_addr_t base, top;
+	struct mem_region *freemem, *r, mem = {
+		.start = (phys_addr_t)-1,
+	};
 
-		r->start = regs[i].addr;
-		r->end = regs[i].addr + regs[i].size;
+	freemem = mem_region_find(freemem_start);
+	assert(freemem && !(freemem->flags & (MR_F_IO | MR_F_CODE)));
 
-		/*
-		 * pick the region we're in for our primary region
-		 */
-		if (freemem_start >= r->start && freemem_start < r->end) {
-			r->flags |= MR_F_PRIMARY;
-			primary = r;
+	for (r = mem_regions; r->end; ++r) {
+		if (!(r->flags & MR_F_IO)) {
+			if (r->start < mem.start)
+				mem.start = r->start;
+			if (r->end > mem.end)
+				mem.end = r->end;
 		}
-
-		/*
-		 * set the lowest and highest addresses found,
-		 * ignoring potential gaps
-		 */
-		if (r->start < mem.start)
-			mem.start = r->start;
-		if (r->end > mem.end)
-			mem.end = r->end;
 	}
-	assert(primary);

lib/arm/setup.c static void mem_init(phys_addr_t freemem_start) · line 180

Quoted 10 lines
lib/arm/setup.c +6 -2
+	assert(mem.end);
 	assert(!(mem.start & ~PHYS_MASK) && !((mem.end - 1) & ~PHYS_MASK));
 
-	__phys_offset = primary->start;	/* PHYS_OFFSET */
-	__phys_end = primary->end;	/* PHYS_END */
+	/* Check for holes */
+	r = mem_region_find(mem.start);
+	while (r && r->end != mem.end)
+		r = mem_region_find(r->end);
+	assert(r);

It took me a while to figure this out, what it does is find all memory regions adjacent to each other and starting at mem.start and ending precisely at mem.end. Looks good.

Quoted 12 lines
lib/arm/setup.c +0 -11
-	/* Split the primary region into two regions; code and data */
-	mem_regions[nr_io + i] = (struct mem_region){
-		.start = code_end,
-		.end = primary->end,
-		.flags = MR_F_PRIMARY,
-	};
-	*primary = (struct mem_region){
-		.start = primary->start,
-		.end = code_end,
-		.flags = MR_F_PRIMARY | MR_F_CODE,
-	};
Quoted 8 lines

lib/arm/setup.c static void mem_init(phys_addr_t freemem_start) · line 183

Quoted 2 lines
lib/arm/setup.c +2 -0
+	/* Ensure our selected freemem region is somewhere in our full range */
+	assert(freemem_start >= mem.start && freemem->end <= mem.end);

That looks a bit strange, the comment refers to checking the freemem region, but we're actually checking is the subregion from freemem_start to freemem->end. I suppose it's because freemem_start comes after data, stack, fdt and initrd in the freemem region, and what we consider available memory is [freemem_start, freemem->end), right?

Right. I'll change the comment to say 'freemem range' rather than region.

Thanks,
drew

#18 lore raw

PATCH v2 7/8 chr-testdev: Silently fail init

In reply to Andrew Jones

If there's no virtio-console / chr-testdev configured, then the user probably didn't want them. Just silently fail rather than stating the obvious.

---
 lib/chr-testdev.c | 5 +----
 1 file changed, 1 insertion(+), 4 deletions(-)
lib/chr-testdev.c +1 -4
index 6890f63c8b29..b3c641a833fe 100644
@@ -54,11 +54,8 @@ void chr_testdev_init(void)
 	int ret;
 
 	vcon = virtio_bind(VIRTIO_ID_CONSOLE);
-	if (vcon == NULL) {
-		printf("%s: %s: can't find a virtio-console\n",
-				__func__, TESTDEV_NAME);
+	if (vcon == NULL)
 		return;
-	}
 
 	ret = vcon->config->find_vqs(vcon, 2, vqs, NULL, io_names);
 	if (ret < 0) {
Signature · 1 line
2.30.2
#19 lore raw

PATCH v2 8/8 arm/arm64: psci: don't assume method is hvc

In reply to Andrew Jones

The method can also be smc and it will be when running on bare metal.

---
 arm/cstart.S       | 22 ++++++++++++++++++++++
 arm/cstart64.S     | 22 ++++++++++++++++++++++
 arm/selftest.c     | 34 +++++++---------------------------
 lib/arm/asm/psci.h | 10 ++++++++--
 lib/arm/psci.c     | 37 +++++++++++++++++++++++++++++--------
 lib/arm/setup.c    |  2 ++
 6 files changed, 90 insertions(+), 37 deletions(-)
arm/cstart.S +22 -0
index 446966de350d..2401d92cdadc 100644
@@ -95,6 +95,28 @@ start:
 
 .text
 
+/*
+ * psci_invoke_hvc / psci_invoke_smc
+ *
+ * Inputs:
+ *   r0 -- function_id
+ *   r1 -- arg0
+ *   r2 -- arg1
+ *   r3 -- arg2
+ *
+ * Outputs:
+ *   r0 -- return code
+ */
+.globl psci_invoke_hvc
+psci_invoke_hvc:
+	hvc	#0
+	mov	pc, lr
+
+.globl psci_invoke_smc
+psci_invoke_smc:
+	smc	#0
+	mov	pc, lr
+
 enable_vfp:
 	/* Enable full access to CP10 and CP11: */
 	mov	r0, #(3 << 22 | 3 << 20)
arm/cstart64.S +22 -0
index 42ba3a3ca249..7610e28f06dd 100644
@@ -109,6 +109,28 @@ start:
 
 .text
 
+/*
+ * psci_invoke_hvc / psci_invoke_smc
+ *
+ * Inputs:
+ *   x0 -- function_id
+ *   x1 -- arg0
+ *   x2 -- arg1
+ *   x3 -- arg2
+ *
+ * Outputs:
+ *   x0 -- return code
+ */
+.globl psci_invoke_hvc
+psci_invoke_hvc:
+	hvc	#0
+	ret
+
+.globl psci_invoke_smc
+psci_invoke_smc:
+	smc	#0
+	ret
+
 get_mmu_off:
 	adrp	x0, auxinfo
 	ldr	x0, [x0, :lo12:auxinfo + 8]
arm/selftest.c +7 -27
index 4495b161cdd5..9f459ed3d571 100644
@@ -400,33 +400,13 @@ static void check_vectors(void *arg __unused)
 	exit(report_summary());
 }
 
-static bool psci_check(void)
+static void psci_print(void)
 {
-	const struct fdt_property *method;
-	int node, len, ver;
-
-	node = fdt_node_offset_by_compatible(dt_fdt(), -1, "arm,psci-0.2");
-	if (node < 0) {
-		printf("PSCI v0.2 compatibility required\n");
-		return false;
-	}
-
-	method = fdt_get_property(dt_fdt(), node, "method", &len);
-	if (method == NULL) {
-		printf("bad psci device tree node\n");
-		return false;
-	}
-
-	if (len < 4 || strcmp(method->data, "hvc") != 0) {
-		printf("psci method must be hvc\n");
-		return false;
-	}
-
-	ver = psci_invoke(PSCI_0_2_FN_PSCI_VERSION, 0, 0, 0);
-	printf("PSCI version %d.%d\n", PSCI_VERSION_MAJOR(ver),
-				       PSCI_VERSION_MINOR(ver));
-
-	return true;
+	int ver = psci_invoke(PSCI_0_2_FN_PSCI_VERSION, 0, 0, 0);
+	report_info("PSCI version: %d.%d", PSCI_VERSION_MAJOR(ver),
+					  PSCI_VERSION_MINOR(ver));
+	report_info("PSCI method: %s", psci_invoke == psci_invoke_hvc ?
+				       "hvc" : "smc");
 }
 
 static void cpu_report(void *data __unused)
@@ -465,7 +445,7 @@ int main(int argc, char **argv)
 
 	} else if (strcmp(argv[1], "smp") == 0) {
 
-		report(psci_check(), "PSCI version");
+		psci_print();
 		on_cpus(cpu_report, NULL);
 		while (!cpumask_full(&ready))
 			cpu_relax();
lib/arm/asm/psci.h +8 -2
index 7b956bf5987d..2820c0a3afc7 100644
@@ -3,8 +3,14 @@
 #include <libcflat.h>
 #include <linux/psci.h>
 
-extern int psci_invoke(unsigned long function_id, unsigned long arg0,
-		       unsigned long arg1, unsigned long arg2);
+typedef int (*psci_invoke_fn)(unsigned long function_id, unsigned long arg0,
+			      unsigned long arg1, unsigned long arg2);
+extern psci_invoke_fn psci_invoke;
+extern int psci_invoke_hvc(unsigned long function_id, unsigned long arg0,
+			   unsigned long arg1, unsigned long arg2);
+extern int psci_invoke_smc(unsigned long function_id, unsigned long arg0,
+			   unsigned long arg1, unsigned long arg2);
+extern void psci_set_conduit(void);
 extern int psci_cpu_on(unsigned long cpuid, unsigned long entry_point);
 extern void psci_system_reset(void);
 extern int cpu_psci_cpu_boot(unsigned int cpu);
lib/arm/psci.c +29 -8
index 936c83948b6a..168786dcf792 100644
@@ -6,22 +6,23 @@
  *
  * This work is licensed under the terms of the GNU LGPL, version 2.
  */
+#include <devicetree.h>
 #include <asm/psci.h>
 #include <asm/setup.h>
 #include <asm/page.h>
 #include <asm/smp.h>
 
-__attribute__((noinline))
-int psci_invoke(unsigned long function_id, unsigned long arg0,
-		unsigned long arg1, unsigned long arg2)
+extern void halt(void);
+
+static int psci_invoke_none(unsigned long function_id, unsigned long arg0,
+			    unsigned long arg1, unsigned long arg2)
 {
-	asm volatile(
-		"hvc #0"
-	: "+r" (function_id)
-	: "r" (arg0), "r" (arg1), "r" (arg2));
-	return function_id;
+	printf("No PSCI method configured! Can't invoke...\n");
+	return PSCI_RET_NOT_PRESENT;
 }
 
+psci_invoke_fn psci_invoke = psci_invoke_none;
+
 int psci_cpu_on(unsigned long cpuid, unsigned long entry_point)
 {
 #ifdef __arm__
@@ -56,3 +57,23 @@ void psci_system_off(void)
 	int err = psci_invoke(PSCI_0_2_FN_SYSTEM_OFF, 0, 0, 0);
 	printf("CPU%d unable to do system off (error = %d)\n", smp_processor_id(), err);
 }
+
+void psci_set_conduit(void)
+{
+	const void *fdt = dt_fdt();
+	const struct fdt_property *method;
+	int node, len;
+
+	node = fdt_node_offset_by_compatible(fdt, -1, "arm,psci-0.2");
+	assert_msg(node >= 0, "PSCI v0.2 compatibility required");
+
+	method = fdt_get_property(fdt, node, "method", &len);
+	assert(method != NULL && len == 4);
+
+	if (strcmp(method->data, "hvc") == 0)
+		psci_invoke = psci_invoke_hvc;
+	else if (strcmp(method->data, "smc") == 0)
+		psci_invoke = psci_invoke_smc;
+	else
+		assert_msg(false, "Unknown PSCI conduit: %s", method->data);
+}
lib/arm/setup.c +2 -0
index a5ebec3c5a12..07d52d2e5fe6 100644
@@ -25,6 +25,7 @@
 #include <asm/processor.h>
 #include <asm/smp.h>
 #include <asm/timer.h>
+#include <asm/psci.h>
 
 #include "io.h"
 
@@ -266,6 +267,7 @@ void setup(const void *fdt, phys_addr_t freemem_start)
 	mem_regions_add_assumed();
 	mem_init(PAGE_ALIGN((unsigned long)freemem));
 
+	psci_set_conduit();
 	cpu_init();
 
 	/* cpu_init must be called before thread_info_init */
Signature · 1 line
2.30.2
#20 lore raw

Re: [PATCH kvm-unit-tests v2 8/8] arm/arm64: psci: don't assume method is hvc

Quoted 138 lines · Andrew Jones

The method can also be smc and it will be when running on bare metal.

---
 arm/cstart.S       | 22 ++++++++++++++++++++++
 arm/cstart64.S     | 22 ++++++++++++++++++++++
 arm/selftest.c     | 34 +++++++---------------------------
 lib/arm/asm/psci.h | 10 ++++++++--
 lib/arm/psci.c     | 37 +++++++++++++++++++++++++++++--------
 lib/arm/setup.c    |  2 ++
 6 files changed, 90 insertions(+), 37 deletions(-)
arm/cstart.S +22 -0
index 446966de350d..2401d92cdadc 100644
@@ -95,6 +95,28 @@ start:
 
 .text
 
+/*
+ * psci_invoke_hvc / psci_invoke_smc
+ *
+ * Inputs:
+ *   r0 -- function_id
+ *   r1 -- arg0
+ *   r2 -- arg1
+ *   r3 -- arg2
+ *
+ * Outputs:
+ *   r0 -- return code
+ */
+.globl psci_invoke_hvc
+psci_invoke_hvc:
+	hvc	#0
+	mov	pc, lr
+
+.globl psci_invoke_smc
+psci_invoke_smc:
+	smc	#0
+	mov	pc, lr
+
 enable_vfp:
 	/* Enable full access to CP10 and CP11: */
 	mov	r0, #(3 << 22 | 3 << 20)
arm/cstart64.S +22 -0
index 42ba3a3ca249..7610e28f06dd 100644
@@ -109,6 +109,28 @@ start:
 
 .text
 
+/*
+ * psci_invoke_hvc / psci_invoke_smc
+ *
+ * Inputs:
+ *   x0 -- function_id
+ *   x1 -- arg0
+ *   x2 -- arg1
+ *   x3 -- arg2
+ *
+ * Outputs:
+ *   x0 -- return code
+ */
+.globl psci_invoke_hvc
+psci_invoke_hvc:
+	hvc	#0
+	ret
+
+.globl psci_invoke_smc
+psci_invoke_smc:
+	smc	#0
+	ret
+
 get_mmu_off:
 	adrp	x0, auxinfo
 	ldr	x0, [x0, :lo12:auxinfo + 8]
arm/selftest.c +7 -27
index 4495b161cdd5..9f459ed3d571 100644
@@ -400,33 +400,13 @@ static void check_vectors(void *arg __unused)
 	exit(report_summary());
 }
 
-static bool psci_check(void)
+static void psci_print(void)
 {
-	const struct fdt_property *method;
-	int node, len, ver;
-
-	node = fdt_node_offset_by_compatible(dt_fdt(), -1, "arm,psci-0.2");
-	if (node < 0) {
-		printf("PSCI v0.2 compatibility required\n");
-		return false;
-	}
-
-	method = fdt_get_property(dt_fdt(), node, "method", &len);
-	if (method == NULL) {
-		printf("bad psci device tree node\n");
-		return false;
-	}
-
-	if (len < 4 || strcmp(method->data, "hvc") != 0) {
-		printf("psci method must be hvc\n");
-		return false;
-	}
-
-	ver = psci_invoke(PSCI_0_2_FN_PSCI_VERSION, 0, 0, 0);
-	printf("PSCI version %d.%d\n", PSCI_VERSION_MAJOR(ver),
-				       PSCI_VERSION_MINOR(ver));
-
-	return true;
+	int ver = psci_invoke(PSCI_0_2_FN_PSCI_VERSION, 0, 0, 0);
+	report_info("PSCI version: %d.%d", PSCI_VERSION_MAJOR(ver),
+					  PSCI_VERSION_MINOR(ver));
+	report_info("PSCI method: %s", psci_invoke == psci_invoke_hvc ?
+				       "hvc" : "smc");
 }
 
 static void cpu_report(void *data __unused)
@@ -465,7 +445,7 @@ int main(int argc, char **argv)
 
 	} else if (strcmp(argv[1], "smp") == 0) {
 
-		report(psci_check(), "PSCI version");
+		psci_print();
 		on_cpus(cpu_report, NULL);
 		while (!cpumask_full(&ready))
 			cpu_relax();
lib/arm/asm/psci.h +0 -0
index 7b956bf5987d..2820c0a3afc7 100644
@@ -3,8 +3,14 @@
 #include <libcflat.h>
 #include <linux/psci.h>

lib/arm/asm/psci.h line 12

Quoted 10 lines
lib/arm/asm/psci.h +7 -2
 
-extern int psci_invoke(unsigned long function_id, unsigned long arg0,
-		       unsigned long arg1, unsigned long arg2);
+typedef int (*psci_invoke_fn)(unsigned long function_id, unsigned long arg0,
+			      unsigned long arg1, unsigned long arg2);
+extern psci_invoke_fn psci_invoke;
+extern int psci_invoke_hvc(unsigned long function_id, unsigned long arg0,
+			   unsigned long arg1, unsigned long arg2);
+extern int psci_invoke_smc(unsigned long function_id, unsigned long arg0,
+			   unsigned long arg1, unsigned long arg2);

Hmm, I forgot to change function_id to 'unsigned int'.

Quoted 12 lines
lib/arm/asm/psci.h +1 -0
+extern void psci_set_conduit(void);
 extern int psci_cpu_on(unsigned long cpuid, unsigned long entry_point);
 extern void psci_system_reset(void);
 extern int cpu_psci_cpu_boot(unsigned int cpu);
lib/arm/psci.c +0 -0
index 936c83948b6a..168786dcf792 100644
@@ -6,22 +6,23 @@
  *
  * This work is licensed under the terms of the GNU LGPL, version 2.
  */

lib/arm/psci.c line 15

Quoted 10 lines
lib/arm/psci.c +2 -3
+#include <devicetree.h>
 #include <asm/psci.h>
 #include <asm/setup.h>
 #include <asm/page.h>
 #include <asm/smp.h>
 
-__attribute__((noinline))
-int psci_invoke(unsigned long function_id, unsigned long arg0,
-		unsigned long arg1, unsigned long arg2)
+extern void halt(void);

And, this is a left over from an earlier version of psci_invoke_none.

Will fix these things for v3.

Quoted 65 lines
lib/arm/psci.c +27 -5
+
+static int psci_invoke_none(unsigned long function_id, unsigned long arg0,
+			    unsigned long arg1, unsigned long arg2)
 {
-	asm volatile(
-		"hvc #0"
-	: "+r" (function_id)
-	: "r" (arg0), "r" (arg1), "r" (arg2));
-	return function_id;
+	printf("No PSCI method configured! Can't invoke...\n");
+	return PSCI_RET_NOT_PRESENT;
 }
 
+psci_invoke_fn psci_invoke = psci_invoke_none;
+
 int psci_cpu_on(unsigned long cpuid, unsigned long entry_point)
 {
 #ifdef __arm__
@@ -56,3 +57,23 @@ void psci_system_off(void)
 	int err = psci_invoke(PSCI_0_2_FN_SYSTEM_OFF, 0, 0, 0);
 	printf("CPU%d unable to do system off (error = %d)\n", smp_processor_id(), err);
 }
+
+void psci_set_conduit(void)
+{
+	const void *fdt = dt_fdt();
+	const struct fdt_property *method;
+	int node, len;
+
+	node = fdt_node_offset_by_compatible(fdt, -1, "arm,psci-0.2");
+	assert_msg(node >= 0, "PSCI v0.2 compatibility required");
+
+	method = fdt_get_property(fdt, node, "method", &len);
+	assert(method != NULL && len == 4);
+
+	if (strcmp(method->data, "hvc") == 0)
+		psci_invoke = psci_invoke_hvc;
+	else if (strcmp(method->data, "smc") == 0)
+		psci_invoke = psci_invoke_smc;
+	else
+		assert_msg(false, "Unknown PSCI conduit: %s", method->data);
+}
lib/arm/setup.c +2 -0
index a5ebec3c5a12..07d52d2e5fe6 100644
@@ -25,6 +25,7 @@
 #include <asm/processor.h>
 #include <asm/smp.h>
 #include <asm/timer.h>
+#include <asm/psci.h>
 
 #include "io.h"
 
@@ -266,6 +267,7 @@ void setup(const void *fdt, phys_addr_t freemem_start)
 	mem_regions_add_assumed();
 	mem_init(PAGE_ALIGN((unsigned long)freemem));
 
+	psci_set_conduit();
 	cpu_init();
 
 	/* cpu_init must be called before thread_info_init */
Signature · 1 line
2.30.2

Thanks,
drew

#21 lore raw

Re: [PATCH kvm-unit-tests v2 8/8] arm/arm64: psci: don't assume method is hvc

For v3, I've done the following changes (inline)

Quoted 52 lines · Andrew Jones
Quoted 51 lines · Andrew Jones

The method can also be smc and it will be when running on bare metal.

---
 arm/cstart.S       | 22 ++++++++++++++++++++++
 arm/cstart64.S     | 22 ++++++++++++++++++++++
 arm/selftest.c     | 34 +++++++---------------------------
 lib/arm/asm/psci.h | 10 ++++++++--
 lib/arm/psci.c     | 37 +++++++++++++++++++++++++++++--------
 lib/arm/setup.c    |  2 ++
 6 files changed, 90 insertions(+), 37 deletions(-)
arm/cstart.S +22 -0
index 446966de350d..2401d92cdadc 100644
@@ -95,6 +95,28 @@ start:
 
 .text
 
+/*
+ * psci_invoke_hvc / psci_invoke_smc
+ *
+ * Inputs:
+ *   r0 -- function_id
+ *   r1 -- arg0
+ *   r2 -- arg1
+ *   r3 -- arg2
+ *
+ * Outputs:
+ *   r0 -- return code
+ */
+.globl psci_invoke_hvc
+psci_invoke_hvc:
+	hvc	#0
+	mov	pc, lr
+
+.globl psci_invoke_smc
+psci_invoke_smc:
+	smc	#0
+	mov	pc, lr
+
 enable_vfp:
 	/* Enable full access to CP10 and CP11: */
 	mov	r0, #(3 << 22 | 3 << 20)
arm/cstart64.S +0 -0
index 42ba3a3ca249..7610e28f06dd 100644
@@ -109,6 +109,28 @@ start:
 
Quoted 7 lines

arm/cstart64.S start: · line 116

Quoted 7 lines
arm/cstart64.S +5 -0
 .text
 
+/*
+ * psci_invoke_hvc / psci_invoke_smc
+ *
+ * Inputs:
+ *   x0 -- function_id

changed this comment to be 'w0 -- function_id'

Quoted 81 lines
Quoted 81 lines
arm/cstart64.S +17 -0
+ *   x1 -- arg0
+ *   x2 -- arg1
+ *   x3 -- arg2
+ *
+ * Outputs:
+ *   x0 -- return code
+ */
+.globl psci_invoke_hvc
+psci_invoke_hvc:
+	hvc	#0
+	ret
+
+.globl psci_invoke_smc
+psci_invoke_smc:
+	smc	#0
+	ret
+
 get_mmu_off:
 	adrp	x0, auxinfo
 	ldr	x0, [x0, :lo12:auxinfo + 8]
arm/selftest.c +7 -27
index 4495b161cdd5..9f459ed3d571 100644
@@ -400,33 +400,13 @@ static void check_vectors(void *arg __unused)
 	exit(report_summary());
 }
 
-static bool psci_check(void)
+static void psci_print(void)
 {
-	const struct fdt_property *method;
-	int node, len, ver;
-
-	node = fdt_node_offset_by_compatible(dt_fdt(), -1, "arm,psci-0.2");
-	if (node < 0) {
-		printf("PSCI v0.2 compatibility required\n");
-		return false;
-	}
-
-	method = fdt_get_property(dt_fdt(), node, "method", &len);
-	if (method == NULL) {
-		printf("bad psci device tree node\n");
-		return false;
-	}
-
-	if (len < 4 || strcmp(method->data, "hvc") != 0) {
-		printf("psci method must be hvc\n");
-		return false;
-	}
-
-	ver = psci_invoke(PSCI_0_2_FN_PSCI_VERSION, 0, 0, 0);
-	printf("PSCI version %d.%d\n", PSCI_VERSION_MAJOR(ver),
-				       PSCI_VERSION_MINOR(ver));
-
-	return true;
+	int ver = psci_invoke(PSCI_0_2_FN_PSCI_VERSION, 0, 0, 0);
+	report_info("PSCI version: %d.%d", PSCI_VERSION_MAJOR(ver),
+					  PSCI_VERSION_MINOR(ver));
+	report_info("PSCI method: %s", psci_invoke == psci_invoke_hvc ?
+				       "hvc" : "smc");
 }
 
 static void cpu_report(void *data __unused)
@@ -465,7 +445,7 @@ int main(int argc, char **argv)
 
 	} else if (strcmp(argv[1], "smp") == 0) {
 
-		report(psci_check(), "PSCI version");
+		psci_print();
 		on_cpus(cpu_report, NULL);
 		while (!cpumask_full(&ready))
 			cpu_relax();
lib/arm/asm/psci.h +0 -0
index 7b956bf5987d..2820c0a3afc7 100644
@@ -3,8 +3,14 @@
 #include <libcflat.h>
 #include <linux/psci.h>
 
Quoted 9 lines

lib/arm/asm/psci.h line 12

Quoted 9 lines
lib/arm/asm/psci.h +7 -2
-extern int psci_invoke(unsigned long function_id, unsigned long arg0,
-		       unsigned long arg1, unsigned long arg2);
+typedef int (*psci_invoke_fn)(unsigned long function_id, unsigned long arg0,
+			      unsigned long arg1, unsigned long arg2);
+extern psci_invoke_fn psci_invoke;
+extern int psci_invoke_hvc(unsigned long function_id, unsigned long arg0,
+			   unsigned long arg1, unsigned long arg2);
+extern int psci_invoke_smc(unsigned long function_id, unsigned long arg0,
+			   unsigned long arg1, unsigned long arg2);

The prototypes are now

long invoke_fn(unsigned int function_id, unsigned long arg0,
unsigned long arg1, unsigned long arg2)

Notice the return value changed to long and the function_id to unsigned int.

I also improved the commit message by adding the following:

The method can be smc in addition to hvc, and it will be when running on bare metal. Additionally, we move the invocations to assembly so we don't have to rely on compiler assumptions. We also fix the prototype of psci_invoke. It should return long, not int, and function_id should be an unsigned int, not an unsigned long.

Thanks,
drew

#22 lore raw

Re: [PATCH kvm-unit-tests v2 8/8] arm/arm64: psci: don't assume method is hvc

Hi Drew,

Quoted 156 lines · Andrew Jones

For v3, I've done the following changes (inline)

Quoted 52 lines · Andrew Jones
Quoted 51 lines · Andrew Jones

The method can also be smc and it will be when running on bare metal.

---
 arm/cstart.S       | 22 ++++++++++++++++++++++
 arm/cstart64.S     | 22 ++++++++++++++++++++++
 arm/selftest.c     | 34 +++++++---------------------------
 lib/arm/asm/psci.h | 10 ++++++++--
 lib/arm/psci.c     | 37 +++++++++++++++++++++++++++++--------
 lib/arm/setup.c    |  2 ++
 6 files changed, 90 insertions(+), 37 deletions(-)
arm/cstart.S +22 -0
index 446966de350d..2401d92cdadc 100644
@@ -95,6 +95,28 @@ start:
 
 .text
 
+/*
+ * psci_invoke_hvc / psci_invoke_smc
+ *
+ * Inputs:
+ *   r0 -- function_id
+ *   r1 -- arg0
+ *   r2 -- arg1
+ *   r3 -- arg2
+ *
+ * Outputs:
+ *   r0 -- return code
+ */
+.globl psci_invoke_hvc
+psci_invoke_hvc:
+	hvc	#0
+	mov	pc, lr
+
+.globl psci_invoke_smc
+psci_invoke_smc:
+	smc	#0
+	mov	pc, lr
+
 enable_vfp:
 	/* Enable full access to CP10 and CP11: */
 	mov	r0, #(3 << 22 | 3 << 20)
arm/cstart64.S +0 -0
index 42ba3a3ca249..7610e28f06dd 100644
@@ -109,6 +109,28 @@ start:
 
Quoted 7 lines

arm/cstart64.S start: · line 116

Quoted 7 lines
arm/cstart64.S +5 -0
 .text
 
+/*
+ * psci_invoke_hvc / psci_invoke_smc
+ *
+ * Inputs:
+ *   x0 -- function_id

changed this comment to be 'w0 -- function_id'

Quoted 81 lines
Quoted 81 lines
arm/cstart64.S +17 -0
+ *   x1 -- arg0
+ *   x2 -- arg1
+ *   x3 -- arg2
+ *
+ * Outputs:
+ *   x0 -- return code
+ */
+.globl psci_invoke_hvc
+psci_invoke_hvc:
+	hvc	#0
+	ret
+
+.globl psci_invoke_smc
+psci_invoke_smc:
+	smc	#0
+	ret
+
 get_mmu_off:
 	adrp	x0, auxinfo
 	ldr	x0, [x0, :lo12:auxinfo + 8]
arm/selftest.c +7 -27
index 4495b161cdd5..9f459ed3d571 100644
@@ -400,33 +400,13 @@ static void check_vectors(void *arg __unused)
 	exit(report_summary());
 }
 
-static bool psci_check(void)
+static void psci_print(void)
 {
-	const struct fdt_property *method;
-	int node, len, ver;
-
-	node = fdt_node_offset_by_compatible(dt_fdt(), -1, "arm,psci-0.2");
-	if (node < 0) {
-		printf("PSCI v0.2 compatibility required\n");
-		return false;
-	}
-
-	method = fdt_get_property(dt_fdt(), node, "method", &len);
-	if (method == NULL) {
-		printf("bad psci device tree node\n");
-		return false;
-	}
-
-	if (len < 4 || strcmp(method->data, "hvc") != 0) {
-		printf("psci method must be hvc\n");
-		return false;
-	}
-
-	ver = psci_invoke(PSCI_0_2_FN_PSCI_VERSION, 0, 0, 0);
-	printf("PSCI version %d.%d\n", PSCI_VERSION_MAJOR(ver),
-				       PSCI_VERSION_MINOR(ver));
-
-	return true;
+	int ver = psci_invoke(PSCI_0_2_FN_PSCI_VERSION, 0, 0, 0);
+	report_info("PSCI version: %d.%d", PSCI_VERSION_MAJOR(ver),
+					  PSCI_VERSION_MINOR(ver));
+	report_info("PSCI method: %s", psci_invoke == psci_invoke_hvc ?
+				       "hvc" : "smc");
 }
 
 static void cpu_report(void *data __unused)
@@ -465,7 +445,7 @@ int main(int argc, char **argv)
 
 	} else if (strcmp(argv[1], "smp") == 0) {
 
-		report(psci_check(), "PSCI version");
+		psci_print();
 		on_cpus(cpu_report, NULL);
 		while (!cpumask_full(&ready))
 			cpu_relax();
lib/arm/asm/psci.h +0 -0
index 7b956bf5987d..2820c0a3afc7 100644
@@ -3,8 +3,14 @@
 #include <libcflat.h>
 #include <linux/psci.h>
 
Quoted 9 lines

lib/arm/asm/psci.h line 12

Quoted 9 lines
lib/arm/asm/psci.h +7 -2
-extern int psci_invoke(unsigned long function_id, unsigned long arg0,
-		       unsigned long arg1, unsigned long arg2);
+typedef int (*psci_invoke_fn)(unsigned long function_id, unsigned long arg0,
+			      unsigned long arg1, unsigned long arg2);
+extern psci_invoke_fn psci_invoke;
+extern int psci_invoke_hvc(unsigned long function_id, unsigned long arg0,
+			   unsigned long arg1, unsigned long arg2);
+extern int psci_invoke_smc(unsigned long function_id, unsigned long arg0,
+			   unsigned long arg1, unsigned long arg2);

The prototypes are now

Quoted 5 lines

long invoke_fn(unsigned int function_id, unsigned long arg0,
unsigned long arg1, unsigned long arg2)

Notice the return value changed to long and the function_id to unsigned int.

Strictly speaking, arm always returns an unsigned long (32bits), but arm64 can return either an unsigned long (64bits) when using SMC64/HVC64, or an unsigned int (32bits) when using SMC32/HVC32. But, I did check C99 and it says that the original value is unchanged if it fits in the new type (section 6.3.1.3), so we can just cast the result to an unsigned if we ever want to use SMC32/HVC32 on arm64.

The declaration for psci_invoke_none should also be changed, but otherwise looks good.

Thanks,

Alex

Quoted 13 lines

I also improved the commit message by adding the following:

The method can be smc in addition to hvc, and it will be when running on bare metal. Additionally, we move the invocations to assembly so we don't have to rely on compiler assumptions. We also fix the prototype of psci_invoke. It should return long, not int, and function_id should be an unsigned int, not an unsigned long.

Thanks,
drew

#23 lore raw

Re: [PATCH kvm-unit-tests v2 8/8] arm/arm64: psci: don't assume method is hvc

Quoted 162 lines · Alexandru Elisei

Hi Drew,

Quoted 159 lines · Andrew Jones

For v3, I've done the following changes (inline)

Quoted 52 lines · Andrew Jones
Quoted 51 lines · Andrew Jones

The method can also be smc and it will be when running on bare metal.

---
 arm/cstart.S       | 22 ++++++++++++++++++++++
 arm/cstart64.S     | 22 ++++++++++++++++++++++
 arm/selftest.c     | 34 +++++++---------------------------
 lib/arm/asm/psci.h | 10 ++++++++--
 lib/arm/psci.c     | 37 +++++++++++++++++++++++++++++--------
 lib/arm/setup.c    |  2 ++
 6 files changed, 90 insertions(+), 37 deletions(-)
arm/cstart.S +22 -0
index 446966de350d..2401d92cdadc 100644
@@ -95,6 +95,28 @@ start:
 
 .text
 
+/*
+ * psci_invoke_hvc / psci_invoke_smc
+ *
+ * Inputs:
+ *   r0 -- function_id
+ *   r1 -- arg0
+ *   r2 -- arg1
+ *   r3 -- arg2
+ *
+ * Outputs:
+ *   r0 -- return code
+ */
+.globl psci_invoke_hvc
+psci_invoke_hvc:
+	hvc	#0
+	mov	pc, lr
+
+.globl psci_invoke_smc
+psci_invoke_smc:
+	smc	#0
+	mov	pc, lr
+
 enable_vfp:
 	/* Enable full access to CP10 and CP11: */
 	mov	r0, #(3 << 22 | 3 << 20)
arm/cstart64.S +0 -0
index 42ba3a3ca249..7610e28f06dd 100644
@@ -109,6 +109,28 @@ start:
 
Quoted 7 lines

arm/cstart64.S start: · line 116

Quoted 7 lines
arm/cstart64.S +5 -0
 .text
 
+/*
+ * psci_invoke_hvc / psci_invoke_smc
+ *
+ * Inputs:
+ *   x0 -- function_id

changed this comment to be 'w0 -- function_id'

Quoted 81 lines
Quoted 81 lines
arm/cstart64.S +17 -0
+ *   x1 -- arg0
+ *   x2 -- arg1
+ *   x3 -- arg2
+ *
+ * Outputs:
+ *   x0 -- return code
+ */
+.globl psci_invoke_hvc
+psci_invoke_hvc:
+	hvc	#0
+	ret
+
+.globl psci_invoke_smc
+psci_invoke_smc:
+	smc	#0
+	ret
+
 get_mmu_off:
 	adrp	x0, auxinfo
 	ldr	x0, [x0, :lo12:auxinfo + 8]
arm/selftest.c +7 -27
index 4495b161cdd5..9f459ed3d571 100644
@@ -400,33 +400,13 @@ static void check_vectors(void *arg __unused)
 	exit(report_summary());
 }
 
-static bool psci_check(void)
+static void psci_print(void)
 {
-	const struct fdt_property *method;
-	int node, len, ver;
-
-	node = fdt_node_offset_by_compatible(dt_fdt(), -1, "arm,psci-0.2");
-	if (node < 0) {
-		printf("PSCI v0.2 compatibility required\n");
-		return false;
-	}
-
-	method = fdt_get_property(dt_fdt(), node, "method", &len);
-	if (method == NULL) {
-		printf("bad psci device tree node\n");
-		return false;
-	}
-
-	if (len < 4 || strcmp(method->data, "hvc") != 0) {
-		printf("psci method must be hvc\n");
-		return false;
-	}
-
-	ver = psci_invoke(PSCI_0_2_FN_PSCI_VERSION, 0, 0, 0);
-	printf("PSCI version %d.%d\n", PSCI_VERSION_MAJOR(ver),
-				       PSCI_VERSION_MINOR(ver));
-
-	return true;
+	int ver = psci_invoke(PSCI_0_2_FN_PSCI_VERSION, 0, 0, 0);
+	report_info("PSCI version: %d.%d", PSCI_VERSION_MAJOR(ver),
+					  PSCI_VERSION_MINOR(ver));
+	report_info("PSCI method: %s", psci_invoke == psci_invoke_hvc ?
+				       "hvc" : "smc");
 }
 
 static void cpu_report(void *data __unused)
@@ -465,7 +445,7 @@ int main(int argc, char **argv)
 
 	} else if (strcmp(argv[1], "smp") == 0) {
 
-		report(psci_check(), "PSCI version");
+		psci_print();
 		on_cpus(cpu_report, NULL);
 		while (!cpumask_full(&ready))
 			cpu_relax();
lib/arm/asm/psci.h +0 -0
index 7b956bf5987d..2820c0a3afc7 100644
@@ -3,8 +3,14 @@
 #include <libcflat.h>
 #include <linux/psci.h>
 
Quoted 9 lines

lib/arm/asm/psci.h line 12

Quoted 9 lines
lib/arm/asm/psci.h +7 -2
-extern int psci_invoke(unsigned long function_id, unsigned long arg0,
-		       unsigned long arg1, unsigned long arg2);
+typedef int (*psci_invoke_fn)(unsigned long function_id, unsigned long arg0,
+			      unsigned long arg1, unsigned long arg2);
+extern psci_invoke_fn psci_invoke;
+extern int psci_invoke_hvc(unsigned long function_id, unsigned long arg0,
+			   unsigned long arg1, unsigned long arg2);
+extern int psci_invoke_smc(unsigned long function_id, unsigned long arg0,
+			   unsigned long arg1, unsigned long arg2);

The prototypes are now

long invoke_fn(unsigned int function_id, unsigned long arg0,
unsigned long arg1, unsigned long arg2)

Quoted 6 lines
Quoted 2 lines

Notice the return value changed to long and the function_id to unsigned int.

Strictly speaking, arm always returns an unsigned long (32bits), but arm64 can return either an unsigned long (64bits) when using SMC64/HVC64, or an unsigned int (32bits) when using SMC32/HVC32.

Hmm, where did you see that? Because section 5.1 of the SMC calling convention disagrees

"""
5.1 Error codes
Errors codes that are returned in R0, W0 and X0 are signed integers of the appropriate size:
* In AArch32:
o When using the SMC32/HVC32 calling convention, error codes, which are returned in R0, are 32-bit signed integers.
* In AArch64:
o When using the SMC64/HVC64 calling convention, error codes, which are returned in X0, are 64-bit signed integers.
o When using the SMC32/HVC32 calling convention, error codes, which are returned in W0, are 32-bit signed integers. X0[63:32] is UNDEFINED.
"""

And 5.2.2 from the Power State Coordination Interface manual

"""
5.2.2 Return error codes
Table 6 defines the values for error codes used with PSCI functions. All errors are considered to be 32-bit signed integers.
"""

Thanks,
drew

#24 lore raw

Re: [PATCH kvm-unit-tests v2 8/8] arm/arm64: psci: don't assume method is hvc

Hi Drew,

Quoted 171 lines · Andrew Jones
Quoted 159 lines · Alexandru Elisei

Hi Drew,

Quoted 156 lines · Andrew Jones

For v3, I've done the following changes (inline)

Quoted 52 lines · Andrew Jones
Quoted 51 lines · Andrew Jones

The method can also be smc and it will be when running on bare metal.

---
 arm/cstart.S       | 22 ++++++++++++++++++++++
 arm/cstart64.S     | 22 ++++++++++++++++++++++
 arm/selftest.c     | 34 +++++++---------------------------
 lib/arm/asm/psci.h | 10 ++++++++--
 lib/arm/psci.c     | 37 +++++++++++++++++++++++++++++--------
 lib/arm/setup.c    |  2 ++
 6 files changed, 90 insertions(+), 37 deletions(-)
arm/cstart.S +22 -0
index 446966de350d..2401d92cdadc 100644
@@ -95,6 +95,28 @@ start:
 
 .text
 
+/*
+ * psci_invoke_hvc / psci_invoke_smc
+ *
+ * Inputs:
+ *   r0 -- function_id
+ *   r1 -- arg0
+ *   r2 -- arg1
+ *   r3 -- arg2
+ *
+ * Outputs:
+ *   r0 -- return code
+ */
+.globl psci_invoke_hvc
+psci_invoke_hvc:
+	hvc	#0
+	mov	pc, lr
+
+.globl psci_invoke_smc
+psci_invoke_smc:
+	smc	#0
+	mov	pc, lr
+
 enable_vfp:
 	/* Enable full access to CP10 and CP11: */
 	mov	r0, #(3 << 22 | 3 << 20)
arm/cstart64.S +0 -0
index 42ba3a3ca249..7610e28f06dd 100644
@@ -109,6 +109,28 @@ start:
 
Quoted 7 lines

arm/cstart64.S start: · line 116

Quoted 7 lines
arm/cstart64.S +5 -0
 .text
 
+/*
+ * psci_invoke_hvc / psci_invoke_smc
+ *
+ * Inputs:
+ *   x0 -- function_id

changed this comment to be 'w0 -- function_id'

Quoted 81 lines
Quoted 81 lines
arm/cstart64.S +17 -0
+ *   x1 -- arg0
+ *   x2 -- arg1
+ *   x3 -- arg2
+ *
+ * Outputs:
+ *   x0 -- return code
+ */
+.globl psci_invoke_hvc
+psci_invoke_hvc:
+	hvc	#0
+	ret
+
+.globl psci_invoke_smc
+psci_invoke_smc:
+	smc	#0
+	ret
+
 get_mmu_off:
 	adrp	x0, auxinfo
 	ldr	x0, [x0, :lo12:auxinfo + 8]
arm/selftest.c +7 -27
index 4495b161cdd5..9f459ed3d571 100644
@@ -400,33 +400,13 @@ static void check_vectors(void *arg __unused)
 	exit(report_summary());
 }
 
-static bool psci_check(void)
+static void psci_print(void)
 {
-	const struct fdt_property *method;
-	int node, len, ver;
-
-	node = fdt_node_offset_by_compatible(dt_fdt(), -1, "arm,psci-0.2");
-	if (node < 0) {
-		printf("PSCI v0.2 compatibility required\n");
-		return false;
-	}
-
-	method = fdt_get_property(dt_fdt(), node, "method", &len);
-	if (method == NULL) {
-		printf("bad psci device tree node\n");
-		return false;
-	}
-
-	if (len < 4 || strcmp(method->data, "hvc") != 0) {
-		printf("psci method must be hvc\n");
-		return false;
-	}
-
-	ver = psci_invoke(PSCI_0_2_FN_PSCI_VERSION, 0, 0, 0);
-	printf("PSCI version %d.%d\n", PSCI_VERSION_MAJOR(ver),
-				       PSCI_VERSION_MINOR(ver));
-
-	return true;
+	int ver = psci_invoke(PSCI_0_2_FN_PSCI_VERSION, 0, 0, 0);
+	report_info("PSCI version: %d.%d", PSCI_VERSION_MAJOR(ver),
+					  PSCI_VERSION_MINOR(ver));
+	report_info("PSCI method: %s", psci_invoke == psci_invoke_hvc ?
+				       "hvc" : "smc");
 }
 
 static void cpu_report(void *data __unused)
@@ -465,7 +445,7 @@ int main(int argc, char **argv)
 
 	} else if (strcmp(argv[1], "smp") == 0) {
 
-		report(psci_check(), "PSCI version");
+		psci_print();
 		on_cpus(cpu_report, NULL);
 		while (!cpumask_full(&ready))
 			cpu_relax();
lib/arm/asm/psci.h +0 -0
index 7b956bf5987d..2820c0a3afc7 100644
@@ -3,8 +3,14 @@
 #include <libcflat.h>
 #include <linux/psci.h>
 
Quoted 9 lines

lib/arm/asm/psci.h line 12

Quoted 9 lines
lib/arm/asm/psci.h +7 -2
-extern int psci_invoke(unsigned long function_id, unsigned long arg0,
-		       unsigned long arg1, unsigned long arg2);
+typedef int (*psci_invoke_fn)(unsigned long function_id, unsigned long arg0,
+			      unsigned long arg1, unsigned long arg2);
+extern psci_invoke_fn psci_invoke;
+extern int psci_invoke_hvc(unsigned long function_id, unsigned long arg0,
+			   unsigned long arg1, unsigned long arg2);
+extern int psci_invoke_smc(unsigned long function_id, unsigned long arg0,
+			   unsigned long arg1, unsigned long arg2);

The prototypes are now

Quoted 8 lines
Quoted 5 lines

long invoke_fn(unsigned int function_id, unsigned long arg0,
unsigned long arg1, unsigned long arg2)

Notice the return value changed to long and the function_id to unsigned int.

Strictly speaking, arm always returns an unsigned long (32bits), but arm64 can return either an unsigned long (64bits) when using SMC64/HVC64, or an unsigned int (32bits) when using SMC32/HVC32.

Hmm, where did you see that? Because section 5.1 of the SMC calling convention disagrees

Quoted 13 lines

"""
5.1 Error codes
Errors codes that are returned in R0, W0 and X0 are signed integers of the appropriate size:
* In AArch32:
o When using the SMC32/HVC32 calling convention, error codes, which are returned in R0, are 32-bit signed integers.
* In AArch64:
o When using the SMC64/HVC64 calling convention, error codes, which are returned in X0, are 64-bit signed integers.
o When using the SMC32/HVC32 calling convention, error codes, which are returned in W0, are 32-bit signed integers. X0[63:32] is UNDEFINED.
"""

I made a mistake, I read "long" and thought unsigned integer instead of signed.

I went over the PSCI spec again, and PSCI_STAT_RESIDENCY and PSCI_STAT_COUNT return an uint64_t or uint32_t, depending on the convention. Those were introduced in PSCI 1.0, they are optional and none of the tests use them, so I guess we can ignore them for now and revisit the prototypes if we ever need to.

Thanks,

Alex

Quoted 12 lines

And 5.2.2 from the Power State Coordination Interface manual

"""
5.2.2 Return error codes
Table 6 defines the values for error codes used with PSCI functions. All errors are considered to be 32-bit signed integers.
"""

Thanks,
drew