Skip to contentSkip to conversation

target/riscv: Ignore reserved bits in PTE for RV64

PATCH In review [email protected] 4 participants 5 messages Sep 26, 2019 – Oct 14, 2019 RISC-V

lore
Share & tools
lore
#0 lore raw

PATCH v6 1/1 target/riscv: Ignore reserved bits in PTE for RV64

From: Guo Ren <[email protected]>

Highest 10 bits of PTE are reserved in riscv-privileged, ref: [1], so we need to ignore them. They cannot be a part of ppn.

1: The RISC-V Instruction Set Manual, Volume II: Privileged Architecture
4.4 Sv39: Page-Based 39-bit Virtual-Memory System
4.5 Sv48: Page-Based 48-bit Virtual-Memory System

---
 target/riscv/cpu_bits.h   | 7 +++++++
 target/riscv/cpu_helper.c | 2 +-
 2 files changed, 8 insertions(+), 1 deletion(-)

Changelog V6:
- Add Reviewer: Alistair Francis

Changelog V5:
- Add Reviewer and Tester: Bin Meng

Changelog V4:
- Change title to Ignore not Bugfix
- Use PTE_PPN_MASK for RV32 and RV64

Changelog V3:
- Use UUL define for PTE_RESERVED
- Keep ppn >> PTE_PPN_SHIFT

Changelog V2:
- Bugfix pte destroyed cause boot fail
- Change to AND with a mask instead of shifting both directions

target/riscv/cpu_bits.h +7 -0
index e998348..399c2c6 100644
@@ -473,6 +473,13 @@
 /* Page table PPN shift amount */
 #define PTE_PPN_SHIFT       10
 
+/* Page table PPN mask */
+#if defined(TARGET_RISCV32)
+#define PTE_PPN_MASK        0xffffffffUL
+#elif defined(TARGET_RISCV64)
+#define PTE_PPN_MASK        0x3fffffffffffffULL
+#endif
+
 /* Leaf page shift amount */
 #define PGSHIFT             12
 
target/riscv/cpu_helper.c +1 -1
index 87dd6a6..9961b37 100644
@@ -261,7 +261,7 @@ restart:
 #elif defined(TARGET_RISCV64)
         target_ulong pte = ldq_phys(cs->as, pte_addr);
 #endif
-        hwaddr ppn = pte >> PTE_PPN_SHIFT;
+        hwaddr ppn = (pte & PTE_PPN_MASK) >> PTE_PPN_SHIFT;
 
         if (!(pte & PTE_V)) {
             /* Invalid PTE */
Signature · 1 line
2.7.4
#1 lore raw
Quoted 60 lines · On Wed, 25 Sep 2019 17:14:21 PDT (-0700), [email protected] wrote:

From: Guo Ren <[email protected]>

Highest 10 bits of PTE are reserved in riscv-privileged, ref: [1], so we need to ignore them. They cannot be a part of ppn.

1: The RISC-V Instruction Set Manual, Volume II: Privileged Architecture
4.4 Sv39: Page-Based 39-bit Virtual-Memory System
4.5 Sv48: Page-Based 48-bit Virtual-Memory System

---
 target/riscv/cpu_bits.h   | 7 +++++++
 target/riscv/cpu_helper.c | 2 +-
 2 files changed, 8 insertions(+), 1 deletion(-)

Changelog V6:
- Add Reviewer: Alistair Francis

Changelog V5:
- Add Reviewer and Tester: Bin Meng

Changelog V4:
- Change title to Ignore not Bugfix
- Use PTE_PPN_MASK for RV32 and RV64

Changelog V3:
- Use UUL define for PTE_RESERVED
- Keep ppn >> PTE_PPN_SHIFT

Changelog V2:
- Bugfix pte destroyed cause boot fail
- Change to AND with a mask instead of shifting both directions

target/riscv/cpu_bits.h +7 -0
index e998348..399c2c6 100644
@@ -473,6 +473,13 @@
 /* Page table PPN shift amount */
 #define PTE_PPN_SHIFT       10
 
+/* Page table PPN mask */
+#if defined(TARGET_RISCV32)
+#define PTE_PPN_MASK        0xffffffffUL
+#elif defined(TARGET_RISCV64)
+#define PTE_PPN_MASK        0x3fffffffffffffULL
+#endif
+
 /* Leaf page shift amount */
 #define PGSHIFT             12
 
target/riscv/cpu_helper.c +0 -0
index 87dd6a6..9961b37 100644

target/riscv/cpu_helper.c restart: · line 267

Quoted 8 lines
target/riscv/cpu_helper.c +1 -1
@@ -261,7 +261,7 @@ restart:
 #elif defined(TARGET_RISCV64)
         target_ulong pte = ldq_phys(cs->as, pte_addr);
 #endif
-        hwaddr ppn = pte >> PTE_PPN_SHIFT;
+        hwaddr ppn = (pte & PTE_PPN_MASK) >> PTE_PPN_SHIFT;
 
         if (!(pte & PTE_V)) {
             /* Invalid PTE */

I know I'm a bit late to the party here, but I don't like this. There's ample evidence that wrapping the physical address space is a bad idea, and just because the ISA allows implementations to do this doesn't mean we should.

author
#2 lore raw

The patch didn't wrap the physical address space directly, just follow the spec. I admit that I am trying to use the compliance specification to allow qemu to support some non-standard software.
But compliance specification and wrapping the physical address space are different things.
I'm preparing c910 pachset for linux riscv and you can question me there.

Quoted 68 lines · Palmer Dabbelt
Quoted 66 lines · On Wed, 25 Sep 2019 17:14:21 PDT (-0700), [email protected] wrote:

From: Guo Ren <[email protected]>

Highest 10 bits of PTE are reserved in riscv-privileged, ref: [1], so we need to ignore them. They cannot be a part of ppn.

1: The RISC-V Instruction Set Manual, Volume II: Privileged Architecture
4.4 Sv39: Page-Based 39-bit Virtual-Memory System
4.5 Sv48: Page-Based 48-bit Virtual-Memory System

---
 target/riscv/cpu_bits.h   | 7 +++++++
 target/riscv/cpu_helper.c | 2 +-
 2 files changed, 8 insertions(+), 1 deletion(-)

Changelog V6:
- Add Reviewer: Alistair Francis

Changelog V5:
- Add Reviewer and Tester: Bin Meng

Changelog V4:
- Change title to Ignore not Bugfix
- Use PTE_PPN_MASK for RV32 and RV64

Changelog V3:
- Use UUL define for PTE_RESERVED
- Keep ppn >> PTE_PPN_SHIFT

Changelog V2:
- Bugfix pte destroyed cause boot fail
- Change to AND with a mask instead of shifting both directions

target/riscv/cpu_bits.h +7 -0
index e998348..399c2c6 100644
@@ -473,6 +473,13 @@
 /* Page table PPN shift amount */
 #define PTE_PPN_SHIFT       10
 
+/* Page table PPN mask */
+#if defined(TARGET_RISCV32)
+#define PTE_PPN_MASK        0xffffffffUL
+#elif defined(TARGET_RISCV64)
+#define PTE_PPN_MASK        0x3fffffffffffffULL
+#endif
+
 /* Leaf page shift amount */
 #define PGSHIFT             12
 
target/riscv/cpu_helper.c +1 -1
index 87dd6a6..9961b37 100644
@@ -261,7 +261,7 @@ restart:
 #elif defined(TARGET_RISCV64)
         target_ulong pte = ldq_phys(cs->as, pte_addr);
 #endif
-        hwaddr ppn = pte >> PTE_PPN_SHIFT;
+        hwaddr ppn = (pte & PTE_PPN_MASK) >> PTE_PPN_SHIFT;
 
Quoted 6 lines

target/riscv/cpu_helper.c restart: · line 267

Quoted 2 lines
target/riscv/cpu_helper.c +0 -0
         if (!(pte & PTE_V)) {
             /* Invalid PTE */

I know I'm a bit late to the party here, but I don't like this. There's ample evidence that wrapping the physical address space is a bad idea, and just because the ISA allows implementations to do this doesn't mean we should.

--
Best Regards
Guo Ren

ML: https://lore.kernel.org/linux-csky/

#3 lore raw

There is nowhere in the spec that ever says what hardware has to do if any of those reserved bits are non-zero. Hardware is certainly not required to ignore them and treat the PTE as being valid (which is what this patch does). I'd argue that since only buggy code would ever set these bits, QEMU should treat any PTE with them set as being invalid so that programmers can realize they've made a mistake.

Jonathan

Quoted 92 lines · Guo Ren

The patch didn't wrap the physical address space directly, just follow the spec. I admit that I am trying to use the compliance specification to allow qemu to support some non-standard software.
But compliance specification and wrapping the physical address space are different things.
I'm preparing c910 pachset for linux riscv and you can question me there.

Quoted 68 lines · Palmer Dabbelt
Quoted 66 lines · On Wed, 25 Sep 2019 17:14:21 PDT (-0700), [email protected] wrote:

From: Guo Ren <[email protected]>

Highest 10 bits of PTE are reserved in riscv-privileged, ref: [1], so we need to ignore them. They cannot be a part of ppn.

1: The RISC-V Instruction Set Manual, Volume II: Privileged Architecture
4.4 Sv39: Page-Based 39-bit Virtual-Memory System
4.5 Sv48: Page-Based 48-bit Virtual-Memory System

---
 target/riscv/cpu_bits.h   | 7 +++++++
 target/riscv/cpu_helper.c | 2 +-
 2 files changed, 8 insertions(+), 1 deletion(-)

Changelog V6:
- Add Reviewer: Alistair Francis

Changelog V5:
- Add Reviewer and Tester: Bin Meng

Changelog V4:
- Change title to Ignore not Bugfix
- Use PTE_PPN_MASK for RV32 and RV64

Changelog V3:
- Use UUL define for PTE_RESERVED
- Keep ppn >> PTE_PPN_SHIFT

Changelog V2:
- Bugfix pte destroyed cause boot fail
- Change to AND with a mask instead of shifting both directions

target/riscv/cpu_bits.h +7 -0
index e998348..399c2c6 100644
@@ -473,6 +473,13 @@
 /* Page table PPN shift amount */
 #define PTE_PPN_SHIFT       10
 
+/* Page table PPN mask */
+#if defined(TARGET_RISCV32)
+#define PTE_PPN_MASK        0xffffffffUL
+#elif defined(TARGET_RISCV64)
+#define PTE_PPN_MASK        0x3fffffffffffffULL
+#endif
+
 /* Leaf page shift amount */
 #define PGSHIFT             12
 
target/riscv/cpu_helper.c +1 -1
index 87dd6a6..9961b37 100644
@@ -261,7 +261,7 @@ restart:
 #elif defined(TARGET_RISCV64)
         target_ulong pte = ldq_phys(cs->as, pte_addr);
 #endif
-        hwaddr ppn = pte >> PTE_PPN_SHIFT;
+        hwaddr ppn = (pte & PTE_PPN_MASK) >> PTE_PPN_SHIFT;
 
Quoted 6 lines

target/riscv/cpu_helper.c restart: · line 267

Quoted 2 lines
target/riscv/cpu_helper.c +0 -0
         if (!(pte & PTE_V)) {
             /* Invalid PTE */

I know I'm a bit late to the party here, but I don't like this. There's ample evidence that wrapping the physical address space is a bad idea, and just because the ISA allows implementations to do this doesn't mean we should.

--
Best Regards
Guo Ren

ML: https://lore.kernel.org/linux-csky/

#4 lore raw
In reply to Palmer Dabbelt
Quoted 68 lines · Palmer Dabbelt
Quoted 66 lines · On Wed, 25 Sep 2019 17:14:21 PDT (-0700), [email protected] wrote:

From: Guo Ren <[email protected]>

Highest 10 bits of PTE are reserved in riscv-privileged, ref: [1], so we need to ignore them. They cannot be a part of ppn.

1: The RISC-V Instruction Set Manual, Volume II: Privileged Architecture
4.4 Sv39: Page-Based 39-bit Virtual-Memory System
4.5 Sv48: Page-Based 48-bit Virtual-Memory System

---
 target/riscv/cpu_bits.h   | 7 +++++++
 target/riscv/cpu_helper.c | 2 +-
 2 files changed, 8 insertions(+), 1 deletion(-)

Changelog V6:
- Add Reviewer: Alistair Francis

Changelog V5:
- Add Reviewer and Tester: Bin Meng

Changelog V4:
- Change title to Ignore not Bugfix
- Use PTE_PPN_MASK for RV32 and RV64

Changelog V3:
- Use UUL define for PTE_RESERVED
- Keep ppn >> PTE_PPN_SHIFT

Changelog V2:
- Bugfix pte destroyed cause boot fail
- Change to AND with a mask instead of shifting both directions

target/riscv/cpu_bits.h +7 -0
index e998348..399c2c6 100644
@@ -473,6 +473,13 @@
 /* Page table PPN shift amount */
 #define PTE_PPN_SHIFT       10
 
+/* Page table PPN mask */
+#if defined(TARGET_RISCV32)
+#define PTE_PPN_MASK        0xffffffffUL
+#elif defined(TARGET_RISCV64)
+#define PTE_PPN_MASK        0x3fffffffffffffULL
+#endif
+
 /* Leaf page shift amount */
 #define PGSHIFT             12
 
target/riscv/cpu_helper.c +1 -1
index 87dd6a6..9961b37 100644
@@ -261,7 +261,7 @@ restart:
 #elif defined(TARGET_RISCV64)
         target_ulong pte = ldq_phys(cs->as, pte_addr);
 #endif
-        hwaddr ppn = pte >> PTE_PPN_SHIFT;
+        hwaddr ppn = (pte & PTE_PPN_MASK) >> PTE_PPN_SHIFT;
 
Quoted 6 lines

target/riscv/cpu_helper.c restart: · line 267

Quoted 2 lines
target/riscv/cpu_helper.c +0 -0
         if (!(pte & PTE_V)) {
             /* Invalid PTE */

I know I'm a bit late to the party here, but I don't like this. There's ample evidence that wrapping the physical address space is a bad idea, and just because the ISA allows implementations to do this doesn't mean we should.

I think this is ok as the spec specifically states that "These reserved bits may also be used to facilitate research experimentation." and as QEMU is generally used for developing new features it makes sense to allow guests to set these bit and we just ignore them.

Software should always set these to zero, so the worst outcome of doing this is that QEMU will hid software bugs if people set the bits, but I don't see that as huge downside.

Alistair