From: sisshiki1969 <sisshiki@mac.com>

For now, qemu-x86_64 returns ENOMEM when mprotect() was called with an argument
len is 0 from a guest process.
This behavior is incompatible with the current Linux implementation,
which mprotect() with len = 0 does nothing and returns 0,
although it does not appear to be explicitly described in man.

This is due to the following function which always returns false if len = 0.

```C
static inline bool guest_range_valid_untagged(abi_ulong start, abi_ulong len)
{
    return len - 1 <= GUEST_ADDR_MAX && start <= GUEST_ADDR_MAX - len + 1;
}

```

This patch fix this incompatibility problem.

Signed-off-by: sisshiki1969 <sisshiki@mac.com>
---
 linux-user/mmap.c | 6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)

diff --git a/linux-user/mmap.c b/linux-user/mmap.c
index 28f3bc85ed..1ed79459ea 100644
--- a/linux-user/mmap.c
+++ b/linux-user/mmap.c
@@ -130,12 +130,12 @@ int target_mprotect(abi_ulong start, abi_ulong len, int target_prot)
     }
     len = TARGET_PAGE_ALIGN(len);
     end = start + len;
-    if (!guest_range_valid_untagged(start, len)) {
-        return -TARGET_ENOMEM;
-    }
     if (len == 0) {
         return 0;
     }
+    if (!guest_range_valid_untagged(start, len)) {
+        return -TARGET_ENOMEM;
+    }
 
     mmap_lock();
     host_start = start & qemu_host_page_mask;
-- 
2.25.1

On 10/6/22 08:38, Soichiro Isshiki wrote:
> From: sisshiki1969 <sisshiki@mac.com>
> 
> For now, qemu-x86_64 returns ENOMEM when mprotect() was called with an argument
> len is 0 from a guest process.
> This behavior is incompatible with the current Linux implementation,
> which mprotect() with len = 0 does nothing and returns 0,
> although it does not appear to be explicitly described in man.

You're right that the ordering of checks differs from the kernel.
The kernel has:

(1) validate prot !(growdown && growup)
(2) validate page aligned
(3) pass len == 0
(4) validate no wraparound
(5) validate prot for arch.
(6) validate vma valid.

while we have

(1) validate page aligned
(2) validate prot for arch
(3) validate vma valid
(4) pass len == 0.

My previous answer vs guest_range_valid_untagged is incorrect considering all of this: if 
start > GUEST_ADDR_MAX, that *should* fail vma valid, but the kernel will have returned 
success before that.

Although, sorta, this smells like a kernel bug.
Why should mprotect(-4096, 0, 0) succeed while mprotect(-4096, 4096, 0) fails?

But anyway, if we're going to fix len == 0 to match, we might as well fix all 3 test 
ordering bugs at the same time.


r~

On Thu, 6 Oct 2022 at 19:05, Soichiro Isshiki
<sisshiki@isshiki-clinic.com> wrote:
>
> From: sisshiki1969 <sisshiki@mac.com>
>
> For now, qemu-x86_64 returns ENOMEM when mprotect() was called with an argument
> len is 0 from a guest process.
> This behavior is incompatible with the current Linux implementation,
> which mprotect() with len = 0 does nothing and returns 0,
> although it does not appear to be explicitly described in man.
>
> This is due to the following function which always returns false if len = 0.
>
> ```C
> static inline bool guest_range_valid_untagged(abi_ulong start, abi_ulong len)
> {
>     return len - 1 <= GUEST_ADDR_MAX && start <= GUEST_ADDR_MAX - len + 1;
> }
>
> ```
>
> This patch fix this incompatibility problem.
>
> Signed-off-by: sisshiki1969 <sisshiki@mac.com>
> ---
>  linux-user/mmap.c | 6 +++---
>  1 file changed, 3 insertions(+), 3 deletions(-)
>
> diff --git a/linux-user/mmap.c b/linux-user/mmap.c
> index 28f3bc85ed..1ed79459ea 100644
> --- a/linux-user/mmap.c
> +++ b/linux-user/mmap.c
> @@ -130,12 +130,12 @@ int target_mprotect(abi_ulong start, abi_ulong len, int target_prot)
>      }
>      len = TARGET_PAGE_ALIGN(len);
>      end = start + len;
> -    if (!guest_range_valid_untagged(start, len)) {
> -        return -TARGET_ENOMEM;
> -    }
>      if (len == 0) {
>          return 0;
>      }
> +    if (!guest_range_valid_untagged(start, len)) {
> +        return -TARGET_ENOMEM;
> +    }
>
>      mmap_lock();
>      host_start = start & qemu_host_page_mask;

Cc'ing Richard -- is this the right fix, or would it be better instead
to make guest_range_valid_untagged() correctly handle a zero-length
range ?

thanks
-- PMM

On 10/6/22 11:13, Peter Maydell wrote:
> On Thu, 6 Oct 2022 at 19:05, Soichiro Isshiki
> <sisshiki@isshiki-clinic.com> wrote:
>>
>> From: sisshiki1969 <sisshiki@mac.com>
>>
>> For now, qemu-x86_64 returns ENOMEM when mprotect() was called with an argument
>> len is 0 from a guest process.
>> This behavior is incompatible with the current Linux implementation,
>> which mprotect() with len = 0 does nothing and returns 0,
>> although it does not appear to be explicitly described in man.
>>
>> This is due to the following function which always returns false if len = 0.
>>
>> ```C
>> static inline bool guest_range_valid_untagged(abi_ulong start, abi_ulong len)
>> {
>>      return len - 1 <= GUEST_ADDR_MAX && start <= GUEST_ADDR_MAX - len + 1;
>> }
>>
...
> Cc'ing Richard -- is this the right fix, or would it be better instead
> to make guest_range_valid_untagged() correctly handle a zero-length
> range ?

I think fixing the range check might be best.


r~

Thank you for your response.

Yes, we can also modify guest_range_valid_untagged() like this:

 static inline bool guest_range_valid_untagged(abi_ulong start, abi_ulong
len)
 {
-    return len - 1 <= GUEST_ADDR_MAX && start <= GUEST_ADDR_MAX - len + 1;
+   return !len || len - 1 <= GUEST_ADDR_MAX && start <= GUEST_ADDR_MAX -
len + 1;
 }

But actually, guest_range_valid_untagged() is called from several sites
other than target_mprotect().
(1) target_mmap() in bsd-user
(2) target_madvise() in linux-user
(3) target_mmap() in linux-user
(4) target_munmap() in linux-user
(5) access_ok_untagged() in linux-user/qemu.h
(6) target_mremap() in linux-user
(7) do_shmat() in linux-user/syscall.c

(1)-(5) have explicit guards for the condition of len = 0 in front of
calling  guest_range_valid_untagged().
(1) https://gitlab.com/qemu-project/qemu/-/blob/master/bsd-user/mmap.c#L477
(2)
https://gitlab.com/qemu-project/qemu/-/blob/master/linux-user/mmap.c#L900
(3)
https://gitlab.com/qemu-project/qemu/-/blob/master/linux-user/mmap.c#L456
(4)
https://gitlab.com/qemu-project/qemu/-/blob/master/linux-user/mmap.c#L724
(5)
https://gitlab.com/qemu-project/qemu/-/blob/master/linux-user/qemu.h#L176

But I'm not sure whether this change is correct for (6) and (7).

2022年10月7日(金) 3:31 Richard Henderson <richard.henderson@linaro.org>:

> On 10/6/22 11:13, Peter Maydell wrote:
> > On Thu, 6 Oct 2022 at 19:05, Soichiro Isshiki
> > <sisshiki@isshiki-clinic.com> wrote:
> >>
> >> From: sisshiki1969 <sisshiki@mac.com>
> >>
> >> For now, qemu-x86_64 returns ENOMEM when mprotect() was called with an
> argument
> >> len is 0 from a guest process.
> >> This behavior is incompatible with the current Linux implementation,
> >> which mprotect() with len = 0 does nothing and returns 0,
> >> although it does not appear to be explicitly described in man.
> >>
> >> This is due to the following function which always returns false if len
> = 0.
> >>
> >> ```C
> >> static inline bool guest_range_valid_untagged(abi_ulong start,
> abi_ulong len)
> >> {
> >>      return len - 1 <= GUEST_ADDR_MAX && start <= GUEST_ADDR_MAX - len
> + 1;
> >> }
> >>
> ...
> > Cc'ing Richard -- is this the right fix, or would it be better instead
> > to make guest_range_valid_untagged() correctly handle a zero-length
> > range ?
>
> I think fixing the range check might be best.
>
>
> r~
>