Skip to contentSkip to conversation

target/riscv/pmp: Fix bug preventing

PATCH In review Chris Williams 3 participants 4 messages Oct 6, 2019 – Oct 15, 2019 RISC-V

lore
Share & tools
lore
#0 lore raw

PATCH v2 1/1 target/riscv/pmp: Fix bug preventing

Hello. I hope you don't mind me resubmitting this patch. Please let me know if I've formatted it incorrectly or if it needs more explanation. My previous attempt probably wasn't formatted quite right. This is my first time contributing to Qemu, so I'm keen to get it right - thanks.

This fixes an issue that prevents a RISC-V CPU from executing instructions immediately from the base address of a PMP TOR region.

When jumping to an instruction in a PMP TOR region, pmp_hart_has_privs() is called to validate the access. If this instruction is the very first word of a PMP TOR region, at address 0 relative to the start address of the region, then the access will fail. This is because pmp_hart_has_privs() is called with size 0 to perform this validation, causing this check...

e = pmp_is_in_range(env, i, addr + size - 1);

... to fail, as (addr + size - 1) falls below the base address of the PMP region. Really, the access should succeed. For example, if I have a region spanning 0x80d96000 to 0x88d95fff and the CPU jumps to 0x80d96000, then:

s = 0x80d96000
e = 0x80d95fff

And the validation fails. The size check proposed below catches these zero-size instruction fetch access probes. The word alignment in pmpaddr{0-15} and earlier instruction alignment checks should prevent the execution of instructions over the upper boundary of the PMP region, though I'm happy to give this more attention if this is a concern.

Signed-off-by Chris Williams <[email protected] <mailto:[email protected]>>
target/riscv/pmp.c +0 -0
index d4f1007109..9308672e20 100644
@@ -235,8 +235,9 @@ bool pmp_hart_has_privs(CPURISCVState *env, target_ulong

addr,
     /* 1.10 draft priv spec states there is an implicit order from low to high */
     for (i = 0; i < MAX_RISCV_PMPS; i++) {
+        /* catch zero-size instruction checks */
         s = pmp_is_in_range(env, i, addr);
-        e = pmp_is_in_range(env, i, addr + size - 1);
+        e = pmp_is_in_range(env, i, (size == 0) ? addr : addr + size - 1);

         /* partially inside */
         if ((s + e) == 1) {

#1 lore raw
Quoted 2 lines · Chris Williams

Hello. I hope you don't mind me resubmitting this patch. Please let me know if

Not at all!

Thanks for the patch.

In future if people don't respond you can just keep pinging your patch.

Quoted 3 lines

I've formatted it incorrectly or if it needs more explanation. My previous attempt probably wasn't formatted quite right. This is my first time contributing to Qemu, so I'm keen to get it right - thanks.

This whole paragraph should not be in the commit. It can go below the line though.

Also please use git format-patch to format the patch and then git send-email to send the patch. There is a whole heap of detail here:
https://wiki.qemu.org/Contribute/SubmitAPatch

Quoted 16 lines

This fixes an issue that prevents a RISC-V CPU from executing instructions immediately from the base address of a PMP TOR region.

When jumping to an instruction in a PMP TOR region, pmp_hart_has_privs() is called to validate the access. If this instruction is the very first word of a PMP TOR region, at address 0 relative to the start address of the region, then the access will fail. This is because pmp_hart_has_privs() is called with size 0 to perform this validation, causing this check...

e = pmp_is_in_range(env, i, addr + size - 1);

... to fail, as (addr + size - 1) falls below the base address of the PMP region. Really, the access should succeed. For example, if I have a region spanning 0x80d96000 to 0x88d95fff and the CPU jumps to 0x80d96000, then:

Quoted 8 lines

s = 0x80d96000
e = 0x80d95fff

And the validation fails. The size check proposed below catches these zero-size instruction fetch access probes. The word alignment in pmpaddr{0-15} and earlier instruction alignment checks should prevent the execution of instructions over the upper boundary of the PMP region, though I'm happy to give this more attention if this is a concern.

This seems like a similar issue to this patch as well:
https://lore.kernel.org/qemu-devel/[email protected]/

From that discussion:

"In general, size 0 means "unknown size". In this case, the one tlb lookup is going to be used by lots of instructions -- everything that fits on the page."

Richard's last comment seems like a better fix:

"You certainly could do

    if (size == 0) {
        size = -(addr | TARGET_PAGE_MASK);
    }

to assume that all bytes from addr to the end of the page are accessed. That would avoid changing too much of the rest of the logic.

That said, this code will continue to not work for mis-aligned boundaries."

So I don't think this is the correct solution. I'm not sure if Dayeol is planning on sending a follow up version. If not feel free to send it.

Quoted 2 lines
Signed-off-by Chris Williams <[email protected] <mailto:[email protected]>>

It looks like this is a HTML patch, also ensure all patches are just plain text, git send-email will do this.

Thanks for the patch though! Please send any more fixes that you find :)

Alistair

Quoted 19 lines
target/riscv/pmp.c +2 -1
index d4f1007109..9308672e20 100644
@@ -235,8 +235,9 @@ bool pmp_hart_has_privs(CPURISCVState *env, target_ulong
 addr,
     /* 1.10 draft priv spec states there is an implicit order
          from low to high */
     for (i = 0; i < MAX_RISCV_PMPS; i++) {
+        /* catch zero-size instruction checks */
         s = pmp_is_in_range(env, i, addr);
-        e = pmp_is_in_range(env, i, addr + size - 1);
+        e = pmp_is_in_range(env, i, (size == 0) ? addr : addr + size - 1);
 
         /* partially inside */
         if ((s + e) == 1) {
#2 lore raw

Hi, Alistair,

Thank you for reminding me.
I already had the local patch, so I re-submitted the patch. Please let me know if that's fair enough (or you have any other comments)

Thanks,

Dayeol

Quoted 118 lines · Alistair Francis
Quoted 2 lines · Chris Williams

Hello. I hope you don't mind me resubmitting this patch. Please let me

know if

Not at all!

Thanks for the patch.

In future if people don't respond you can just keep pinging your patch.

Quoted 1 line

I've formatted it incorrectly or if it needs more explanation. My

previous

Quoted 2 lines

attempt probably wasn't formatted quite right. This is my first time contributing to Qemu, so I'm keen to get it right - thanks.

This whole paragraph should not be in the commit. It can go below the line though.

Also please use git format-patch to format the patch and then git send-email to send the patch. There is a whole heap of detail here:
https://wiki.qemu.org/Contribute/SubmitAPatch

Quoted 2 lines

This fixes an issue that prevents a RISC-V CPU from executing

instructions

Quoted 3 lines

immediately from the base address of a PMP TOR region.

When jumping to an instruction in a PMP TOR region, pmp_hart_has_privs()

is

Quoted 1 line

called to validate the access. If this instruction is the very first

word of a

Quoted 1 line

PMP TOR region, at address 0 relative to the start address of the

region, then

Quoted 1 line

the access will fail. This is because pmp_hart_has_privs() is called

with size

Quoted 6 lines

0 to perform this validation, causing this check...

e = pmp_is_in_range(env, i, addr + size - 1);

... to fail, as (addr + size - 1) falls below the base address of the PMP region. Really, the access should succeed. For example, if I have a

region

Quoted 6 lines

spanning 0x80d96000 to 0x88d95fff and the CPU jumps to 0x80d96000, then:

s = 0x80d96000
e = 0x80d95fff

And the validation fails. The size check proposed below catches these

zero-size

Quoted 3 lines

instruction fetch access probes. The word alignment in pmpaddr{0-15} and earlier instruction alignment checks should prevent the execution of instructions over the upper boundary of the PMP region, though I'm happy

to give

Quoted 1 line

this more attention if this is a concern.

This seems like a similar issue to this patch as well:

https://lore.kernel.org/qemu-devel/[email protected]/

From that discussion:

"In general, size 0 means "unknown size". In this case, the one tlb lookup is
going to be used by lots of instructions -- everything that fits on the page."

Richard's last comment seems like a better fix:

"You certainly could do

    if (size == 0) {
        size = -(addr | TARGET_PAGE_MASK);
    }

to assume that all bytes from addr to the end of the page are accessed. That
would avoid changing too much of the rest of the logic.

That said, this code will continue to not work for mis-aligned boundaries."

So I don't think this is the correct solution. I'm not sure if Dayeol is planning on sending a follow up version. If not feel free to send it.

Quoted 2 lines

[email protected]>>

It looks like this is a HTML patch, also ensure all patches are just plain text, git send-email will do this.

Thanks for the patch though! Please send any more fixes that you find :)

Alistair

target/riscv/pmp.c bool pmp_hart_has_privs(CPURISCVState *env,

Quoted 6 lines
target/riscv/pmp.c +0 -0
index d4f1007109..9308672e20 100644
@@ -235,8 +235,9 @@ bool pmp_hart_has_privs(CPURISCVState *env,

target_ulong

target/riscv/pmp.c bool pmp_hart_has_privs(CPURISCVState *env, · line 241

Quoted 8 lines
target/riscv/pmp.c +2 -1
 addr,
     /* 1.10 draft priv spec states there is an implicit order
          from low to high */
     for (i = 0; i < MAX_RISCV_PMPS; i++) {
+        /* catch zero-size instruction checks */
         s = pmp_is_in_range(env, i, addr);
-        e = pmp_is_in_range(env, i, addr + size - 1);
+        e = pmp_is_in_range(env, i, (size == 0) ? addr : addr + size -

1);

Quoted 5 lines
target/riscv/pmp.c +0 -0
         /* partially inside */
         if ((s + e) == 1) {
#3 lore raw
In reply to Alistair Francis

Hi,
Oct 11, 2019, 15:18 by [email protected]:

Quoted 6 lines · Chris Williams

Also please use git format-patch to format the patch and then git send-email to send the patch. There is a whole heap of detail here:
https://wiki.qemu.org/Contribute/SubmitAPatch <https://wiki.qemu.org/Contribute/SubmitAPatch>

OK, I will do in future. I read the page but failed to get it right. Thanks for spotting my patch, and the advice, though.

Quoted 25 lines
Quoted 15 lines

This fixes an issue that prevents a RISC-V CPU from executing instructions immediately from the base address of a PMP TOR region.

When jumping to an instruction in a PMP TOR region, pmp_hart_has_privs() is called to validate the access. If this instruction is the very first word of a PMP TOR region, at address 0 relative to the start address of the region, then the access will fail. This is because pmp_hart_has_privs() is called with size 0 to perform this validation, causing this check...

e = pmp_is_in_range(env, i, addr + size - 1);

... to fail, as (addr + size - 1) falls below the base address of the PMP region. Really, the access should succeed. For example, if I have a region spanning 0x80d96000 to 0x88d95fff and the CPU jumps to 0x80d96000, then:

Quoted 9 lines

s = 0x80d96000
e = 0x80d95fff

And the validation fails. The size check proposed below catches these zero-size instruction fetch access probes. The word alignment in pmpaddr{0-15} and earlier instruction alignment checks should prevent the execution of instructions over the upper boundary of the PMP region, though I'm happy to give this more attention if this is a concern.

Quoted 3 lines

This seems like a similar issue to this patch as well:
https://lore.kernel.org/qemu-devel/[email protected]/ <https://lore.kernel.org/qemu-devel/[email protected]/>

Yes, it appears Dayeol and I have encountered the same issue.

Quoted 13 lines

From that discussion:

"In general, size 0 means "unknown size". In this case, the one tlb lookup is going to be used by lots of instructions -- everything that fits on the page."

Richard's last comment seems like a better fix:

"You certainly could do

if (size == 0) {
size = -(addr | TARGET_PAGE_MASK);
}

Quoted 9 lines

to assume that all bytes from addr to the end of the page are accessed. That would avoid changing too much of the rest of the logic.

That said, this code will continue to not work for mis-aligned boundaries."

So I don't think this is the correct solution. I'm not sure if Dayeol is planning on sending a follow up version. If not feel free to send it.

I'm happy for Dayeol to submit a better patch, if necessary.

Quoted 6 lines
Quoted 2 lines
Signed-off-by Chris Williams <[email protected] <mailto:[email protected]>>

It looks like this is a HTML patch, also ensure all patches are just plain text, git send-email will do this.

Yes, you're right: my webmail client isn't particularly neighborly with respect to Qemu's submission process.

C.