Module Name: src
Committed By: rin
Date: Thu Apr 15 00:00:46 UTC 2021
Modified Files:
src/sys/arch/powerpc/ibm4xx: pmap.c
Log Message:
Use uvm_km_alloc(9) with UVM_KMF_NOWAIT flag in pte_enter(), in order not to
sleep in pmap_enter(9) and pmap_kenter_pa(9), which can result in dead lock.
In most cases, pmap_enter(9) is used with PMAP_CANFAIL flag. In this case,
if pte_enter() fails due to uvm_km_alloc(9), we can safely return ENOMEM.
UVM layer will take care of it.
uvm_km_alloc(9) fails for pmap_enter(9) without PMAP_CANFAIL or for
pmap_kenter_pa(9), we have no choice but to panic there.
However, my testbeds for 403 and 405 survive more than a week at least
without hitting this panic.
To generate a diff of this commit:
cvs rdiff -u -r1.97 -r1.98 src/sys/arch/powerpc/ibm4xx/pmap.c
Please note that diffs are not public domain; they are subject to the
copyright notices on the relevant files.
Modified files:
Index: src/sys/arch/powerpc/ibm4xx/pmap.c
diff -u src/sys/arch/powerpc/ibm4xx/pmap.c:1.97 src/sys/arch/powerpc/ibm4xx/pmap.c:1.98
--- src/sys/arch/powerpc/ibm4xx/pmap.c:1.97 Wed Apr 14 23:45:11 2021
+++ src/sys/arch/powerpc/ibm4xx/pmap.c Thu Apr 15 00:00:46 2021
@@ -1,4 +1,4 @@
-/* $NetBSD: pmap.c,v 1.97 2021/04/14 23:45:11 rin Exp $ */
+/* $NetBSD: pmap.c,v 1.98 2021/04/15 00:00:46 rin Exp $ */
/*
* Copyright 2001 Wasabi Systems, Inc.
@@ -67,7 +67,7 @@
*/
#include <sys/cdefs.h>
-__KERNEL_RCSID(0, "$NetBSD: pmap.c,v 1.97 2021/04/14 23:45:11 rin Exp $");
+__KERNEL_RCSID(0, "$NetBSD: pmap.c,v 1.98 2021/04/15 00:00:46 rin Exp $");
#ifdef _KERNEL_OPT
#include "opt_ddb.h"
@@ -236,11 +236,15 @@ pte_enter(struct pmap *pm, vaddr_t va, u
if (!pm->pm_ptbl[seg]) {
/* Don't allocate a page to clear a non-existent mapping. */
if (!pte)
+ return 1;
+
+ vaddr_t km = uvm_km_alloc(kernel_map, PAGE_SIZE, 0,
+ UVM_KMF_WIRED | UVM_KMF_ZERO | UVM_KMF_NOWAIT);
+
+ if (__predict_false(km == 0))
return 0;
- /* Allocate a page XXXX this will sleep! */
- pm->pm_ptbl[seg] =
- (uint *)uvm_km_alloc(kernel_map, PAGE_SIZE, 0,
- UVM_KMF_WIRED | UVM_KMF_ZERO);
+
+ pm->pm_ptbl[seg] = (u_int *)km;
}
oldpte = pm->pm_ptbl[seg][ptn];
pm->pm_ptbl[seg][ptn] = pte;
@@ -862,7 +866,7 @@ pmap_enter(struct pmap *pm, vaddr_t va,
if (!pmap_enter_pv(pm, va, pa, flags)) {
/* Could not enter pv on a managed page */
- return 1;
+ return ENOMEM;
}
/* Now set attributes. */
@@ -880,7 +884,12 @@ pmap_enter(struct pmap *pm, vaddr_t va,
s = splvm();
/* Insert page into page table. */
- pte_enter(pm, va, tte);
+ if (__predict_false(!pte_enter(pm, va, tte))) {
+ if (__predict_false((flags & PMAP_CANFAIL) == 0))
+ panic("%s: pte_enter", __func__);
+ splx(s);
+ return ENOMEM;
+ }
/* If this is a real fault, enter it in the tlb */
if (tte && ((flags & PMAP_WIRED) == 0)) {
@@ -969,7 +978,8 @@ pmap_kenter_pa(vaddr_t va, paddr_t pa, v
s = splvm();
/* Insert page into page table. */
- pte_enter(pm, va, tte);
+ if (__predict_false(!pte_enter(pm, va, tte)))
+ panic("%s: pte_enter", __func__);
splx(s);
}
@@ -978,7 +988,7 @@ pmap_kremove(vaddr_t va, vsize_t len)
{
while (len > 0) {
- pte_enter(pmap_kernel(), va, 0);
+ (void)pte_enter(pmap_kernel(), va, 0); /* never fail */
va += PAGE_SIZE;
len -= PAGE_SIZE;
}
