Re: [PATCH 2/2] linux-user: implement epoll_pwait2 syscall

[Michael Tokarev]

On 1/8/26 20:43, Michael Tokarev wrote:
epoll_pwait2 is the same as epoll_pwait but with timeout being
(a pointer to) struct timespec instead of an integer.
Resolves: https://gitlab.com/qemu-project/qemu/-/issues/3210

I failed to realize there was a previous attempt to implement
the same syscall by Zixing Liu - it was even submitted through my tree,
but it was badly reviewed - I failed to notice a memory leak in there.

@@ -13654,8 +13666,13 @@ static abi_long do_syscall1(CPUArchState *cpu_env, int 
num, abi_long arg1,
                  }
              }
  
-            ret = get_errno(safe_epoll_pwait(epfd, ep, maxevents, timeout,
-                                             set, SIGSET_T_SIZE));
+            if (num == TARGET_NR_epoll_pwait) {
+                ret = safe_epoll_pwait(epfd, ep, maxevents, arg4,
+                                       set, SIGSET_T_SIZE);
+            } else {
+                ret = safe_epoll_pwait2(epfd, ep, maxevents, timeout_ts,
+                                        set, SIGSET_T_SIZE);
+            }


And this is wrong - I forgot get_errno() in the new version.
Fixed now:

@@ -13654,8 +13666,13 @@ static abi_long do_syscall1(CPUArchState 
*cpu_env, int num, abi_long arg1,
                 }
             }

-            ret = get_errno(safe_epoll_pwait(epfd, ep, maxevents, timeout,
-                                             set, SIGSET_T_SIZE));
+            if (num == TARGET_NR_epoll_pwait) {
+                ret = get_errno(safe_epoll_pwait(epfd, ep, maxevents, arg4,
+                                                 set, SIGSET_T_SIZE));
+            } else {
+                ret = get_errno(safe_epoll_pwait2(epfd, ep, maxevents, 
timeout_ts,
+                                                  set, SIGSET_T_SIZE));
+            }

             if (set) {
                 finish_sigsuspend_mask(ret);

The previous attempt were not successful, because it failed build on
armhf.  I checked why it failed -- it is rather fun, the failure to
build was because qemu carried old/incomplete linux-headers for 32bit
arm.  These headers weren't needed since v5.0.0, and were removed in
v10.0.0-2612-g99c6e970a4 "linux-headers: Remove the 32-bit arm headers" -
after what commit, the patch by Zixing Liu would've worked (modulo the
memory leak).

But I think my version is better, in the end (after fixing get_errno()).

Thanks,

/mjt