From: Martin Pieuchot <mpi@grenadille.net>
Subject: pdaemon, interrupts and ci_idepth
To: tech@openbsd.org
Date: Sat, 19 Oct 2024 15:23:32 +0200

When a system is OOM and has only its reserved pages available, it is
not helpful to let interrupt handlers steal such memory from the page
daemon.  

I've seen iwm(4)'s interrupt handler allocate memory reserved for the
pagedaemon because checking against `curproc' is not good enough.  So
the diff below use the MD `ci_idepth' variable to prevent such scenario.
If this is acceptable, I'd be happy to rename `ci_intrdepth' used by
some architectures to `ci_idepth'.

Ok?

diff --git sys/uvm/uvm_pmemrange.c sys/uvm/uvm_pmemrange.c
index 0dff15f6cee..611d6d2785d 100644
--- sys/uvm/uvm_pmemrange.c
+++ sys/uvm/uvm_pmemrange.c
@@ -967,7 +967,8 @@ again:
 	}
 
 	if ((uvmexp.free <= (uvmexp.reserve_pagedaemon + count)) &&
-	    (curproc != uvm.pagedaemon_proc) && (curproc != syncerproc)) {
+	    (((curproc != uvm.pagedaemon_proc) && (curproc != syncerproc)) ||
+	     (curcpu()->ci_idepth > 0))) {
 		uvm_unlock_fpageq();
 		if (flags & UVM_PLA_WAITOK) {
 			uvm_wait("uvm_pmr_getpages");
@@ -2095,7 +2096,7 @@ uvm_wait_pla(paddr_t low, paddr_t high, paddr_t size, int failok)
 	struct uvm_pmalloc pma;
 	const char *wmsg = "pmrwait";
 
-	if (curproc == uvm.pagedaemon_proc) {
+	if (curproc == uvm.pagedaemon_proc && curcpu()->ci_idepth == 0) {
 		/*
 		 * This is not that uncommon when the pagedaemon is trying
 		 * to flush out a large mmapped file. VOP_WRITE will circle