Skip to contentSkip to conversation

Fix some PMP implementation

PATCH Superseded Zong Li 3 participants 8 messages Jul 21, 2020 – Jul 23, 2020 RISC-V

v3 of 6 Superseded Latest is v6 →

lore
Share & tools
lore

Zong Li

author
#0 lore raw

v3 cover Fix some PMP implementation

This patch set contains the fixes for wrong index of pmpcfg CSR on rv64, and the pmp range in CSR function table. After 3rd version of this patch series, we also fix the PMP wrong checking problem.

Changed in v3:
- Refine the implementation. Suggested by Bin Meng.
- Add fix for PMP wrong checking

Changed in v2:
- Move out the shifting operation from loop. Suggested by Bin Meng.

Zong Li (3):
target/riscv: Fix the range of pmpcfg of CSR funcion table target/riscv/pmp.c: Fix the index offset on RV64 target/riscv: Fix the translation of physical address

 target/riscv/cpu_helper.c | 3 ++-
 target/riscv/csr.c        | 2 +-
 target/riscv/pmp.c        | 8 ++++++++
 3 files changed, 11 insertions(+), 2 deletions(-)
Signature · 1 line
2.27.0
author
#1 lore raw

PATCH v3 1/3 target/riscv: Fix the range of pmpcfg of CSR funcion table

The range of Physical Memory Protection should be from CSR_PMPCFG0 to CSR_PMPCFG3, not to CSR_PMPADDR9.

---
 target/riscv/csr.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
target/riscv/csr.c +1 -1
index ac01c835e1..6a96a01b1c 100644
@@ -1353,7 +1353,7 @@ static riscv_csr_operations csr_ops[CSR_TABLE_SIZE] = {
     [CSR_MTINST] =              { hmode,   read_mtinst,      write_mtinst     },
 
     /* Physical Memory Protection */
-    [CSR_PMPCFG0  ... CSR_PMPADDR9] =  { pmp,   read_pmpcfg,  write_pmpcfg   },
+    [CSR_PMPCFG0  ... CSR_PMPCFG3]   = { pmp,   read_pmpcfg,  write_pmpcfg   },
     [CSR_PMPADDR0 ... CSR_PMPADDR15] = { pmp,   read_pmpaddr, write_pmpaddr  },
 
     /* Performance Counters */
Signature · 1 line
2.27.0
author
#2 lore raw

PATCH v3 2/3 target/riscv/pmp.c: Fix the index offset on RV64

In reply to Zong Li

On RV64, the reg_index is 2 (pmpcfg2 CSR) after the seventh pmp entry, it is not 1 (pmpcfg1 CSR) like RV32. In the original implementation, the second parameter of pmp_write_cfg is "reg_index * sizeof(target_ulong)", and we get the the result which is started from 16 if reg_index is 2, but we expect that it should be started from 8. Separate the implementation for RV32 and RV64 respectively.

Changed in v3:
- Refine the implementation. Suggested by Bin Meng.

Changed in v2:
- Move out the shifting operation from loop. Suggested by Bin Meng.

---
 target/riscv/pmp.c | 8 ++++++++
 1 file changed, 8 insertions(+)
target/riscv/pmp.c +8 -0
index 2a2b9f5363..f2d50bace5 100644
@@ -318,6 +318,10 @@ void pmpcfg_csr_write(CPURISCVState *env, uint32_t reg_index,
         return;
     }
 
+#if defined(TARGET_RISCV64)
+    reg_index >>= 1;
+#endif
+
     for (i = 0; i < sizeof(target_ulong); i++) {
         cfg_val = (val >> 8 * i)  & 0xff;
         pmp_write_cfg(env, (reg_index * sizeof(target_ulong)) + i,
@@ -335,6 +339,10 @@ target_ulong pmpcfg_csr_read(CPURISCVState *env, uint32_t reg_index)
     target_ulong cfg_val = 0;
     target_ulong val = 0;
 
+#if defined(TARGET_RISCV64)
+    reg_index >>= 1;
+#endif
+
     for (i = 0; i < sizeof(target_ulong); i++) {
         val = pmp_read_cfg(env, (reg_index * sizeof(target_ulong)) + i);
         cfg_val |= (val << (i * 8));
Signature · 1 line
2.27.0
author
#5 lore raw

PATCH v3 3/3 target/riscv: Fix the translation of physical address

In reply to Zong Li

The real physical address should add the 12 bits page offset. It also causes the PMP wrong checking due to the minimum granularity of PMP is 4 byte, but we always get the physical address which is 4KB alignment, that means, we always use the start address of the page to check PMP for all addresses which in the same page.

---
 target/riscv/cpu_helper.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)
target/riscv/cpu_helper.c +2 -1
index 75d2ae3434..08b069f0c9 100644
@@ -543,7 +543,8 @@ restart:
             /* for superpage mappings, make a fake leaf PTE for the TLB's
                benefit. */
             target_ulong vpn = addr >> PGSHIFT;
-            *physical = (ppn | (vpn & ((1L << ptshift) - 1))) << PGSHIFT;
+            *physical = ((ppn | (vpn & ((1L << ptshift) - 1))) << PGSHIFT) |
+                        (addr & ~TARGET_PAGE_MASK);
 
             /* set permissions on the TLB entry */
             if ((pte & PTE_R) || ((pte & PTE_X) && mxr)) {
Signature · 1 line
2.27.0
#3 lore raw

Re: [PATCH v3 2/3] target/riscv/pmp.c: Fix the index offset on RV64

In reply to Zong Li

Hi Zong,

Quoted 9 lines · Zong Li

On RV64, the reg_index is 2 (pmpcfg2 CSR) after the seventh pmp entry, it is not 1 (pmpcfg1 CSR) like RV32. In the original implementation, the second parameter of pmp_write_cfg is "reg_index * sizeof(target_ulong)", and we get the the result which is started from 16 if reg_index is 2, but we expect that it should be started from 8. Separate the implementation for RV32 and RV64 respectively.

Quoted 7 lines

Changed in v3:
- Refine the implementation. Suggested by Bin Meng.

Changed in v2:
- Move out the shifting operation from loop. Suggested by Bin Meng.

As I mentioned previously, these changelog should go after --- below. It should not appear in the commit message.

Quoted 20 lines
---
 target/riscv/pmp.c | 8 ++++++++
 1 file changed, 8 insertions(+)
target/riscv/pmp.c +4 -0
index 2a2b9f5363..f2d50bace5 100644
@@ -318,6 +318,10 @@ void pmpcfg_csr_write(CPURISCVState *env, uint32_t reg_index,
         return;
     }
 
+#if defined(TARGET_RISCV64)
+    reg_index >>= 1;
+#endif
+
     for (i = 0; i < sizeof(target_ulong); i++) {
         cfg_val = (val >> 8 * i)  & 0xff;
         pmp_write_cfg(env, (reg_index * sizeof(target_ulong)) + i,

target/riscv/pmp.c target_ulong pmpcfg_csr_read(CPURISCVState *env, uint32_t reg_index) · line 344

Quoted 6 lines
target/riscv/pmp.c +3 -0
@@ -335,6 +339,10 @@ target_ulong pmpcfg_csr_read(CPURISCVState *env, uint32_t reg_index)
     target_ulong cfg_val = 0;
     target_ulong val = 0;
 
+#if defined(TARGET_RISCV64)
+    reg_index >>= 1;
+#endif

We should also move the following:

    trace_pmpcfg_csr_read(env->mhartid, reg_index, cfg_val);

before shifting reg_index. Otherwise it traces the wrong pmpcfg CSR read.

target/riscv/pmp.c target_ulong pmpcfg_csr_read(CPURISCVState *env, uint32_t reg_index)

Quoted 5 lines
target/riscv/pmp.c +1 -1
+
     for (i = 0; i < sizeof(target_ulong); i++) {
         val = pmp_read_cfg(env, (reg_index * sizeof(target_ulong)) + i);
         cfg_val |= (val << (i * 8));
--

Regards,
Bin

#6 lore raw

Re: [PATCH v3 3/3] target/riscv: Fix the translation of physical address

In reply to Zong Li
Quoted 27 lines · Zong Li

The real physical address should add the 12 bits page offset. It also causes the PMP wrong checking due to the minimum granularity of PMP is 4 byte, but we always get the physical address which is 4KB alignment, that means, we always use the start address of the page to check PMP for all addresses which in the same page.

---
 target/riscv/cpu_helper.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)
target/riscv/cpu_helper.c +2 -2
index 75d2ae3434..08b069f0c9 100644
@@ -543,7 +543,8 @@ restart:
             /* for superpage mappings, make a fake leaf PTE for the TLB's
                benefit. */
             target_ulong vpn = addr >> PGSHIFT;
-            *physical = (ppn | (vpn & ((1L << ptshift) - 1))) << PGSHIFT;
+            *physical = ((ppn | (vpn & ((1L << ptshift) - 1))) << PGSHIFT) |
+                        (addr & ~TARGET_PAGE_MASK);
 
             /* set permissions on the TLB entry */
             if ((pte & PTE_R) || ((pte & PTE_X) && mxr)) {
--
Quoted 1 line

2.27.0

I made the same change for our CHERI fork a few months ago but forgot to send the patch upstream (despite marking the commit as a candidate for upstreaming). Sorry about the duplicated debugging work!
(https://github.com/CTSRD-CHERI/qemu/commit/61c8e3f2c0fd4965ec3f316146d1751fae673c12)

author
#4 lore raw

Re: [PATCH v3 2/3] target/riscv/pmp.c: Fix the index offset on RV64

In reply to Bin Meng
Quoted 15 lines · Bin Meng

Hi Zong,

Quoted 11 lines · Zong Li

On RV64, the reg_index is 2 (pmpcfg2 CSR) after the seventh pmp entry, it is not 1 (pmpcfg1 CSR) like RV32. In the original implementation, the second parameter of pmp_write_cfg is "reg_index * sizeof(target_ulong)", and we get the the result which is started from 16 if reg_index is 2, but we expect that it should be started from 8. Separate the implementation for RV32 and RV64 respectively.

Quoted 9 lines
Quoted 5 lines

Changed in v3:
- Refine the implementation. Suggested by Bin Meng.

Changed in v2:
- Move out the shifting operation from loop. Suggested by Bin Meng.

As I mentioned previously, these changelog should go after --- below. It should not appear in the commit message.

OK, remove it in the next version.

Quoted 29 lines
Quoted 20 lines
---
 target/riscv/pmp.c | 8 ++++++++
 1 file changed, 8 insertions(+)
target/riscv/pmp.c +4 -0
index 2a2b9f5363..f2d50bace5 100644
@@ -318,6 +318,10 @@ void pmpcfg_csr_write(CPURISCVState *env, uint32_t reg_index,
         return;
     }
 
+#if defined(TARGET_RISCV64)
+    reg_index >>= 1;
+#endif
+
     for (i = 0; i < sizeof(target_ulong); i++) {
         cfg_val = (val >> 8 * i)  & 0xff;
         pmp_write_cfg(env, (reg_index * sizeof(target_ulong)) + i,

target/riscv/pmp.c target_ulong pmpcfg_csr_read(CPURISCVState *env, uint32_t reg_index) · line 344

Quoted 6 lines
target/riscv/pmp.c +3 -0
@@ -335,6 +339,10 @@ target_ulong pmpcfg_csr_read(CPURISCVState *env, uint32_t reg_index)
     target_ulong cfg_val = 0;
     target_ulong val = 0;
 
+#if defined(TARGET_RISCV64)
+    reg_index >>= 1;
+#endif

We should also move the following:

Quoted 3 lines
    trace_pmpcfg_csr_read(env->mhartid, reg_index, cfg_val);

before shifting reg_index. Otherwise it traces the wrong pmpcfg CSR read.

Yes, thanks for the reminding, Fix it in the next version.

Quoted 9 lines

target/riscv/pmp.c target_ulong pmpcfg_csr_read(CPURISCVState *env, uint32_t reg_index)

Quoted 6 lines
target/riscv/pmp.c +1 -1
+
     for (i = 0; i < sizeof(target_ulong); i++) {
         val = pmp_read_cfg(env, (reg_index * sizeof(target_ulong)) + i);
         cfg_val |= (val << (i * 8));
--

Regards,
Bin

author
#7 lore raw

Re: [PATCH v3 3/3] target/riscv: Fix the translation of physical address

In reply to Alexander Richardson
Quoted 26 lines · Alexander Richardson
Quoted 24 lines · Zong Li

The real physical address should add the 12 bits page offset. It also causes the PMP wrong checking due to the minimum granularity of PMP is 4 byte, but we always get the physical address which is 4KB alignment, that means, we always use the start address of the page to check PMP for all addresses which in the same page.

---
 target/riscv/cpu_helper.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)
target/riscv/cpu_helper.c +2 -1
index 75d2ae3434..08b069f0c9 100644
@@ -543,7 +543,8 @@ restart:
             /* for superpage mappings, make a fake leaf PTE for the TLB's
                benefit. */
             target_ulong vpn = addr >> PGSHIFT;
-            *physical = (ppn | (vpn & ((1L << ptshift) - 1))) << PGSHIFT;
+            *physical = ((ppn | (vpn & ((1L << ptshift) - 1))) << PGSHIFT) |
+                        (addr & ~TARGET_PAGE_MASK);
 
Quoted 9 lines
Quoted 4 lines
target/riscv/cpu_helper.c +0 -1
             /* set permissions on the TLB entry */
             if ((pte & PTE_R) || ((pte & PTE_X) && mxr)) {
--

2.27.0

I made the same change for our CHERI fork a few months ago but forgot to send the patch upstream (despite marking the commit as a candidate for upstreaming). Sorry about the duplicated debugging work!
(https://github.com/CTSRD-CHERI/qemu/commit/61c8e3f2c0fd4965ec3f316146d1751fae673c12)

No, problem.