diff --git a/redox-rt/src/arch/aarch64.rs b/redox-rt/src/arch/aarch64.rs
index 265ea88816..58d2972944 100644
--- a/redox-rt/src/arch/aarch64.rs
+++ b/redox-rt/src/arch/aarch64.rs
@@ -476,5 +476,10 @@ pub unsafe fn arch_pre(stack: &mut SigStack, os: &mut SigArea) -> PosixStackt {
         flags: 0,                  // TODO
     }
 }
+pub fn arch_ret_to_sig(stack: &mut SigStack, control: &Sigcontrol) {
+    let orig_pc = core::mem::replace(&mut stack.regs.pc, __relibc_internal_sigentry as usize);
+    control.saved_ip.set(orig_pc);
+    control.saved_archdep_reg.set(stack.regs.x0);
+}
 pub(crate) static PROC_FD: SyncUnsafeCell<usize> = SyncUnsafeCell::new(usize::MAX);
 static PROC_CALL: [usize; 1] = [ProcCall::Sigdeq as usize];
diff --git a/redox-rt/src/arch/i686.rs b/redox-rt/src/arch/i686.rs
index cd71f604ad..64439fc9d6 100644
--- a/redox-rt/src/arch/i686.rs
+++ b/redox-rt/src/arch/i686.rs
@@ -371,6 +371,11 @@ pub unsafe fn arch_pre(stack: &mut SigStack, area: &mut SigArea) -> PosixStackt
         flags: 0, // TODO
     }
 }
+pub fn arch_ret_to_sig(stack: &mut SigStack, control: &Sigcontrol) {
+    let orig_eip = core::mem::replace(&mut stack.regs.eip, __relibc_internal_sigentry as usize);
+    control.saved_ip.set(orig_eip);
+    control.saved_archdep_reg.set(stack.regs.eflags);
+}
 #[unsafe(no_mangle)]
 pub unsafe fn manually_enter_trampoline() {
     let c = unsafe { &crate::Tcb::current().unwrap().os_specific.control };
diff --git a/redox-rt/src/arch/riscv64.rs b/redox-rt/src/arch/riscv64.rs
index 88e93fb046..ca802438c9 100644
--- a/redox-rt/src/arch/riscv64.rs
+++ b/redox-rt/src/arch/riscv64.rs
@@ -643,5 +643,12 @@ pub unsafe fn arch_pre(stack: &mut SigStack, area: &mut SigArea) -> PosixStackt
 
     get_sigaltstack(area, stack.regs.int_regs[1] as usize).into()
 }
+pub fn arch_ret_to_sig(stack: &mut SigStack, control: &Sigcontrol) {
+    let orig_pc = core::mem::replace(&mut stack.regs.pc, __relibc_internal_sigentry as u64);
+    control.saved_ip.set(orig_pc as usize);
+    control
+        .saved_archdep_reg
+        .set(stack.regs.int_regs[4] as usize); // t0
+}
 pub(crate) static PROC_FD: SyncUnsafeCell<usize> = SyncUnsafeCell::new(usize::MAX);
 static PROC_CALL: [usize; 1] = [ProcCall::Sigdeq as usize];
diff --git a/redox-rt/src/arch/x86_64.rs b/redox-rt/src/arch/x86_64.rs
index dc24394193..516f93b2f0 100644
--- a/redox-rt/src/arch/x86_64.rs
+++ b/redox-rt/src/arch/x86_64.rs
@@ -482,6 +482,14 @@ pub unsafe fn arch_pre(stack: &mut SigStack, area: &mut SigArea) -> PosixStackt
     get_sigaltstack(area, stack.regs.rsp).into()
 }
 
+/// Rearrange the restore-stack and sigarea in a way that makes it look like a new signal was
+/// immediately delivered after restoring the allowset to what it was prior to the original signal delivery.
+pub fn arch_ret_to_sig(stack: &mut SigStack, control: &Sigcontrol) {
+    let orig_rip = core::mem::replace(&mut stack.regs.rip, __relibc_internal_sigentry as usize);
+    control.saved_ip.set(orig_rip);
+    control.saved_archdep_reg.set(stack.regs.rflags);
+}
+
 pub(crate) static SUPPORTS_AVX: AtomicU8 = AtomicU8::new(0);
 
 // __relibc will be prepended to the name, so no_mangle is fine
diff --git a/redox-rt/src/signal.rs b/redox-rt/src/signal.rs
index 31a316069e..b29f9218e7 100644
--- a/redox-rt/src/signal.rs
+++ b/redox-rt/src/signal.rs
@@ -242,15 +242,20 @@ unsafe fn inner(stack: &mut SigStack) {
     );
     core::sync::atomic::compiler_fence(Ordering::Acquire);
 
-    // Update allowset again.
+    // Update allowset again, and obtain the set of pending unblocked signals with this new
+    // allowset. If this is nonempty, we must deliver the signal here rather than wait for it
+    // asynchronously, as shown e.g. in https://gitlab.redox-os.org/redox-os/relibc/-/issues/239.
+    // We do this even if signals_were_disabled beforehand (where the caller hence explicitly
+    // must have jumped to this trampoline).
 
     let new_mask = stack.old_mask;
     let old_mask = get_allowset_raw(&os.control.word);
 
-    let _pending_when_restored_mask = set_allowset_raw(&os.control.word, old_mask, new_mask);
-
-    // TODO: If resetting the sigmask caused signals to be unblocked, then should they be delivered
-    // here? And would it be possible to tail-call-optimize that?
+    let pending_unblocked = {
+        let thread_pending = set_allowset_raw(&os.control.word, old_mask, new_mask);
+        let proc_pending = PROC_CONTROL_STRUCT.pending.load(Ordering::Relaxed);
+        (thread_pending | proc_pending) & new_mask
+    };
 
     unsafe { (*os.arch.get()).last_sig_was_restart = shall_restart };
 
@@ -259,8 +264,18 @@ unsafe fn inner(stack: &mut SigStack) {
 
     // TODO: Support restoring uc_stack?
 
-    // And re-enable them again
-    if !signals_were_disabled {
+    if pending_unblocked > 0 {
+        // If there were signals visible at the time we checked (otherwise it can be considered
+        // "asynchronous" and will be delivered the next time an EINTR is seen, or after the
+        // post-preemption check in the kernel), we delay clearing the INHIBIT_DELIVERY bit, and
+        // rearrange the saved instr ptr field on the stack and the sc_saved_rip field, so that it
+        // returns to the start of the trampoline in a way that makes it look like a new signal was
+        // immediately delivered. There's no need to worry about the __crit_* sections since
+        // INHIBIT_DELIVERY is still set.
+        arch_ret_to_sig(stack, &os.control);
+    } else if !signals_were_disabled {
+        // Re-enable signals again, except in the case where they were disabled and the trampoline
+        // was explicitly jumped to when handling EINTR.
         core::sync::atomic::compiler_fence(Ordering::Release);
         control_flags.store(
             control_flags.load(Ordering::Relaxed) & !SigcontrolFlags::INHIBIT_DELIVERY.bits(),
@@ -747,7 +762,7 @@ pub fn currently_pending_blocked() -> u64 {
     let w0 = control.word[0].load(Ordering::Relaxed);
     let w1 = control.word[1].load(Ordering::Relaxed);
     let allow = (w0 >> 32) | ((w1 >> 32) << 32);
-    let thread_pending = (w0 & 0xffff_ffff) | ((w1 >> 32) & 0xffff_ffff);
+    let thread_pending = (w0 & 0xffff_ffff) | ((w1 & 0xffff_ffff) << 32);
     let proc_pending = PROC_CONTROL_STRUCT.pending.load(Ordering::Relaxed);
 
     core::sync::atomic::fence(Ordering::Acquire); // TODO: Correct ordering?
diff --git a/src/platform/redox/signal.rs b/src/platform/redox/signal.rs
index 5ae9e35f17..b0ff66b74d 100644
--- a/src/platform/redox/signal.rs
+++ b/src/platform/redox/signal.rs
@@ -223,13 +223,19 @@ impl PalSignal for Sys {
         set: Option<&sigset_t>,
         oset: Option<&mut sigset_t>,
     ) -> Result<(), Errno> {
-        Ok(match how {
+        match how {
+            _ if set.is_none() => {
+                if let Some(oset) = oset {
+                    *oset = redox_rt::signal::get_sigmask()?;
+                }
+            }
             SIG_SETMASK => redox_rt::signal::set_sigmask(set.copied(), oset)?,
             SIG_BLOCK => redox_rt::signal::or_sigmask(set.copied(), oset)?,
             SIG_UNBLOCK => redox_rt::signal::andn_sigmask(set.copied(), oset)?,
 
             _ => return Err(Errno(EINVAL)),
-        })
+        }
+        Ok(())
     }
 
     fn sigsuspend(mask: &sigset_t) -> Errno {