[PATCH] linux-user: update ppoll/pselect6 timeout on EINTR

[PATCH] linux-user: update ppoll/pselect6 timeout on EINTR

[Sun Haoyu via qemu development]

Resolves: https://gitlab.com/qemu-project/qemu/-/work_items/3343

Linux kernel writes back the remaining timeout when raw ppoll/pselect6
syscalls are interrupted by signals (-EINTR).

However, QEMU keeps the original timeout and loses the remaining time.

Both do_ppoll() and do_pselect6() now write back the timeout on -EINTR
to match the kernel.

Tested with the issue reproducer.

Signed-off-by: Sun Haoyu <shyliuli@aosc.io>
---
 linux-user/syscall.c | 10 ++++++----
 1 file changed, 6 insertions(+), 4 deletions(-)

diff --git a/linux-user/syscall.c b/linux-user/syscall.c
index 064bc604c9..f049223772 100644
--- a/linux-user/syscall.c
+++ b/linux-user/syscall.c
@@ -1529,16 +1529,18 @@ static abi_long do_pselect6(abi_long arg1, abi_long arg2, abi_long arg3,
         if (efd_addr && copy_to_user_fdset(efd_addr, &efds, n)) {
             return -TARGET_EFAULT;
         }
+    }
+    if (((ret == -TARGET_EINTR) || !is_error(ret)) && ts_addr) {
         if (time64) {
-            if (ts_addr && host_to_target_timespec64(ts_addr, &ts)) {
+            if (host_to_target_timespec64(ts_addr, &ts)) {
                 return -TARGET_EFAULT;
             }
         } else {
-            if (ts_addr && host_to_target_timespec(ts_addr, &ts)) {
+            if (host_to_target_timespec(ts_addr, &ts)) {
                 return -TARGET_EFAULT;
             }
         }
-    }
+     }
     return ret;
 }
 #endif
@@ -1606,7 +1608,7 @@ static abi_long do_ppoll(abi_long arg1, abi_long arg2, abi_long arg3,
         if (set) {
             finish_sigsuspend_mask(ret);
         }
-        if (!is_error(ret) && arg3) {
+        if ((ret == -TARGET_EINTR || !is_error(ret)) && arg3) {
             if (time64) {
                 if (host_to_target_timespec64(arg3, timeout_ts)) {
                     return -TARGET_EFAULT;
-- 
2.53.0

Re: [PATCH] linux-user: update ppoll/pselect6 timeout on EINTR

[Peter Maydell]

On Tue, 17 Mar 2026 at 09:52, Sun Haoyu <shyliuli@aosc.io> wrote:
>
> Resolves: https://gitlab.com/qemu-project/qemu/-/work_items/3343
>
> Linux kernel writes back the remaining timeout when raw ppoll/pselect6
> syscalls are interrupted by signals (-EINTR).
>
> However, QEMU keeps the original timeout and loses the remaining time.
>
> Both do_ppoll() and do_pselect6() now write back the timeout on -EINTR
> to match the kernel.
>
> Tested with the issue reproducer.

Hi; thanks for sending in this patch. I have a couple of thoughts:

(1) if you look at the Linux kernel code for select/pselect/etc,
it doesn't actually make the "update the timespec" handling
specific to "operation succeeded or hit EINTR", it just does
it unconditionally:
https://elixir.bootlin.com/linux/v6.19.8/source/fs/select.c#L295
So I think we could do that too. (There are some early-exit
error cases where the host kernel doesn't get as far as
calling poll_select_finish(), but for those the timespec
struct won't have been changed by the host kernel. So it
will be harmless that we round-tripped it through the
target_to_host_foo and host_to_target_foo conversions.)

(2) this patch handles pselect6() and ppoll(), but do
we also need to do this in do_select() ?

thanks
-- PMM