[Qemu-devel] [PATCH 0/2] linux-user: minor epoll_wait fixes

[Qemu-devel] [PATCH 0/2] linux-user: minor epoll_wait fixes

[Peter Maydell]

This patchset fixes a couple of minor problems in epoll_wait
and epoll_pwait:
 * we weren't checking an argument for sanity, so we would
   fail EFAULT rather than EINVAL for cases like negative
   lengths (thus failing an LTP testcase)
 * we were using alloca() to allocate an array whose length
   is set by the guest

Peter Maydell (2):
  linux-user: Check for bad event numbers in epoll_wait
  linux-user: Don't use alloca() for epoll_wait's epoll event array

 linux-user/syscall.c      | 22 ++++++++++++++++++----
 linux-user/syscall_defs.h |  3 +++
 2 files changed, 21 insertions(+), 4 deletions(-)

-- 
1.9.1

[Qemu-devel] [PATCH 1/2] linux-user: Check for bad event numbers in epoll_wait

[Peter Maydell]

The kernel checks that the maxevents parameter to epoll_wait
is non-negative and not larger than EP_MAX_EVENTS. Add this
check to our implementation, so that:
 * we fail these cases EINVAL rather than EFAULT
 * we don't pass negative or overflowing values to the
   lock_user() size calculation

Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
---
This fixes an LTP epoll_wait03 test case that checks a
negative max_events argument.
---
 linux-user/syscall.c      | 5 +++++
 linux-user/syscall_defs.h | 3 +++
 2 files changed, 8 insertions(+)

diff --git a/linux-user/syscall.c b/linux-user/syscall.c
index fe72abe..3552295 100644
--- a/linux-user/syscall.c
+++ b/linux-user/syscall.c
@@ -11024,6 +11024,11 @@ abi_long do_syscall(void *cpu_env, int num, abi_long arg1,
         int maxevents = arg3;
         int timeout = arg4;
 
+        if (maxevents <= 0 || maxevents > TARGET_EP_MAX_EVENTS) {
+            ret = -TARGET_EINVAL;
+            break;
+        }
+
         target_ep = lock_user(VERIFY_WRITE, arg2,
                               maxevents * sizeof(struct target_epoll_event), 1);
         if (!target_ep) {
diff --git a/linux-user/syscall_defs.h b/linux-user/syscall_defs.h
index d9dea0e..dbf6a38 100644
--- a/linux-user/syscall_defs.h
+++ b/linux-user/syscall_defs.h
@@ -2585,6 +2585,9 @@ struct target_epoll_event {
     abi_uint events;
     target_epoll_data_t data;
 } TARGET_EPOLL_PACKED;
+
+#define TARGET_EP_MAX_EVENTS (INT_MAX / sizeof(struct target_epoll_event))
+
 #endif
 struct target_rlimit64 {
     uint64_t rlim_cur;
-- 
1.9.1

[Qemu-devel] [PATCH 2/2] linux-user: Don't use alloca() for epoll_wait's epoll event array

[Peter Maydell]

The epoll event array which epoll_wait() allocates has a size
determined by the guest which could potentially be quite large.
Use g_try_new() rather than alloca() so that we can fail more
cleanly if the guest hands us an oversize value. (ENOMEM is
not a documented return value for epoll_wait() but in practice
some kernel configurations can return it -- see for instance
sys_oabi_epoll_wait() on ARM.)

This rearrangement includes fixing a bug where we were
incorrectly passing a negative length to unlock_user() in
the error-exit codepath.

Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
---
 linux-user/syscall.c | 17 +++++++++++++----
 1 file changed, 13 insertions(+), 4 deletions(-)

diff --git a/linux-user/syscall.c b/linux-user/syscall.c
index 3552295..721d7b1 100644
--- a/linux-user/syscall.c
+++ b/linux-user/syscall.c
@@ -11035,7 +11035,12 @@ abi_long do_syscall(void *cpu_env, int num, abi_long arg1,
             goto efault;
         }
 
-        ep = alloca(maxevents * sizeof(struct epoll_event));
+        ep = g_try_new(struct epoll_event, maxevents);
+        if (!ep) {
+            unlock_user(target_ep, arg2, 0);
+            ret = -TARGET_ENOMEM;
+            break;
+        }
 
         switch (num) {
 #if defined(TARGET_NR_epoll_pwait)
@@ -11053,8 +11058,8 @@ abi_long do_syscall(void *cpu_env, int num, abi_long arg1,
                 target_set = lock_user(VERIFY_READ, arg5,
                                        sizeof(target_sigset_t), 1);
                 if (!target_set) {
-                    unlock_user(target_ep, arg2, 0);
-                    goto efault;
+                    ret = -TARGET_EFAULT;
+                    break;
                 }
                 target_to_host_sigset(set, target_set);
                 unlock_user(target_set, arg5, 0);
@@ -11082,8 +11087,12 @@ abi_long do_syscall(void *cpu_env, int num, abi_long arg1,
                 target_ep[i].events = tswap32(ep[i].events);
                 target_ep[i].data.u64 = tswap64(ep[i].data.u64);
             }
+            unlock_user(target_ep, arg2,
+                        ret * sizeof(struct target_epoll_event));
+        } else {
+            unlock_user(target_ep, arg2, 0);
         }
-        unlock_user(target_ep, arg2, ret * sizeof(struct target_epoll_event));
+        g_free(ep);
         break;
     }
 #endif
-- 
1.9.1

Re: [Qemu-devel] [PATCH 2/2] linux-user: Don't use alloca() for epoll_wait's epoll event array

[Peter Maydell]

Ping? It looks like patch 1/2 of this series got into the
recent linux-user pullreq, but this one (2/2) didn't. Do
you want a resend as a standalone patch?

thanks
-- PMM

On 18 July 2016 at 15:36, Peter Maydell <peter.maydell@linaro.org> wrote:
> The epoll event array which epoll_wait() allocates has a size
> determined by the guest which could potentially be quite large.
> Use g_try_new() rather than alloca() so that we can fail more
> cleanly if the guest hands us an oversize value. (ENOMEM is
> not a documented return value for epoll_wait() but in practice
> some kernel configurations can return it -- see for instance
> sys_oabi_epoll_wait() on ARM.)
>
> This rearrangement includes fixing a bug where we were
> incorrectly passing a negative length to unlock_user() in
> the error-exit codepath.
>
> Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
> ---
>  linux-user/syscall.c | 17 +++++++++++++----
>  1 file changed, 13 insertions(+), 4 deletions(-)
>
> diff --git a/linux-user/syscall.c b/linux-user/syscall.c
> index 3552295..721d7b1 100644
> --- a/linux-user/syscall.c
> +++ b/linux-user/syscall.c
> @@ -11035,7 +11035,12 @@ abi_long do_syscall(void *cpu_env, int num, abi_long arg1,
>              goto efault;
>          }
>
> -        ep = alloca(maxevents * sizeof(struct epoll_event));
> +        ep = g_try_new(struct epoll_event, maxevents);
> +        if (!ep) {
> +            unlock_user(target_ep, arg2, 0);
> +            ret = -TARGET_ENOMEM;
> +            break;
> +        }
>
>          switch (num) {
>  #if defined(TARGET_NR_epoll_pwait)
> @@ -11053,8 +11058,8 @@ abi_long do_syscall(void *cpu_env, int num, abi_long arg1,
>                  target_set = lock_user(VERIFY_READ, arg5,
>                                         sizeof(target_sigset_t), 1);
>                  if (!target_set) {
> -                    unlock_user(target_ep, arg2, 0);
> -                    goto efault;
> +                    ret = -TARGET_EFAULT;
> +                    break;
>                  }
>                  target_to_host_sigset(set, target_set);
>                  unlock_user(target_set, arg5, 0);
> @@ -11082,8 +11087,12 @@ abi_long do_syscall(void *cpu_env, int num, abi_long arg1,
>                  target_ep[i].events = tswap32(ep[i].events);
>                  target_ep[i].data.u64 = tswap64(ep[i].data.u64);
>              }
> +            unlock_user(target_ep, arg2,
> +                        ret * sizeof(struct target_epoll_event));
> +        } else {
> +            unlock_user(target_ep, arg2, 0);
>          }
> -        unlock_user(target_ep, arg2, ret * sizeof(struct target_epoll_event));
> +        g_free(ep);
>          break;
>      }
>  #endif
> --
> 1.9.1

Re: [Qemu-devel] [PATCH 2/2] linux-user: Don't use alloca() for epoll_wait's epoll event array

[Riku Voipio]

On Tue, Oct 04, 2016 at 02:09:33PM +0100, Peter Maydell wrote:
> Ping? It looks like patch 1/2 of this series got into the
> recent linux-user pullreq, but this one (2/2) didn't. Do
> you want a resend as a standalone patch?

No need, I've pulled it from patchwork, for some reason this didn't get to
my mailbox... 

> On 18 July 2016 at 15:36, Peter Maydell <peter.maydell@linaro.org> wrote:
> > The epoll event array which epoll_wait() allocates has a size
> > determined by the guest which could potentially be quite large.
> > Use g_try_new() rather than alloca() so that we can fail more
> > cleanly if the guest hands us an oversize value. (ENOMEM is
> > not a documented return value for epoll_wait() but in practice
> > some kernel configurations can return it -- see for instance
> > sys_oabi_epoll_wait() on ARM.)
> >
> > This rearrangement includes fixing a bug where we were
> > incorrectly passing a negative length to unlock_user() in
> > the error-exit codepath.
> >
> > Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
> > ---
> >  linux-user/syscall.c | 17 +++++++++++++----
> >  1 file changed, 13 insertions(+), 4 deletions(-)
> >
> > diff --git a/linux-user/syscall.c b/linux-user/syscall.c
> > index 3552295..721d7b1 100644
> > --- a/linux-user/syscall.c
> > +++ b/linux-user/syscall.c
> > @@ -11035,7 +11035,12 @@ abi_long do_syscall(void *cpu_env, int num, abi_long arg1,
> >              goto efault;
> >          }
> >
> > -        ep = alloca(maxevents * sizeof(struct epoll_event));
> > +        ep = g_try_new(struct epoll_event, maxevents);
> > +        if (!ep) {
> > +            unlock_user(target_ep, arg2, 0);
> > +            ret = -TARGET_ENOMEM;
> > +            break;
> > +        }
> >
> >          switch (num) {
> >  #if defined(TARGET_NR_epoll_pwait)
> > @@ -11053,8 +11058,8 @@ abi_long do_syscall(void *cpu_env, int num, abi_long arg1,
> >                  target_set = lock_user(VERIFY_READ, arg5,
> >                                         sizeof(target_sigset_t), 1);
> >                  if (!target_set) {
> > -                    unlock_user(target_ep, arg2, 0);
> > -                    goto efault;
> > +                    ret = -TARGET_EFAULT;
> > +                    break;
> >                  }
> >                  target_to_host_sigset(set, target_set);
> >                  unlock_user(target_set, arg5, 0);
> > @@ -11082,8 +11087,12 @@ abi_long do_syscall(void *cpu_env, int num, abi_long arg1,
> >                  target_ep[i].events = tswap32(ep[i].events);
> >                  target_ep[i].data.u64 = tswap64(ep[i].data.u64);
> >              }
> > +            unlock_user(target_ep, arg2,
> > +                        ret * sizeof(struct target_epoll_event));
> > +        } else {
> > +            unlock_user(target_ep, arg2, 0);
> >          }
> > -        unlock_user(target_ep, arg2, ret * sizeof(struct target_epoll_event));
> > +        g_free(ep);
> >          break;
> >      }
> >  #endif
> > --
> > 1.9.1