page fault after wiring page

[Jeremiah Lott]

We've been seeing some panics and deadlocks that appear to be related to getting a page fault when accessing a page after it has been wired (on amd64).  All the ones we have seen are related to sysctl handlers that call sysctl_wire_old_buffer, then lock a mutex, then call SYSCTL_OUT.  When it does the copyout, it gets a page fault even though the page has been wired, sometimes causing it to sleep while holding a mutex or recurse on non-recursable mutexes.  Here are the two panics that are most easy to follow:

Sleeping thread (tid 100458, pid 2737) owns a non-sleepable lock
sched_switch() at 0xffffffff80603bf5 = sched_switch+0x146
mi_switch() at 0xffffffff805e8e15 = mi_switch+0x183
sleepq_switch() at 0xffffffff8061e6e7 = sleepq_switch+0xb1
sleepq_wait() at 0xffffffff8061f0ea = sleepq_wait+0x3d
_sx_slock_hard() at 0xffffffff805e7ca7 = _sx_slock_hard+0x41d
_sx_slock() at 0xffffffff805e7e32 = _sx_slock+0x3d
vm_map_lookup() at 0xffffffff807909e4 = vm_map_lookup+0x54
vm_fault() at 0xffffffff80786c20 = vm_fault+0x11c
trap_pfault() at 0xffffffff80844dd0 = trap_pfault+0xe1
trap() at 0xffffffff80845286 = trap+0x337
calltrap() at 0xffffffff80827f28 = calltrap+0x8
--- trap 0xc, rip = 0xffffffff8084296b, rsp = 0xffffff811391e7e0, rbp = 0xffffff811391e810 ---
copyout() at 0xffffffff8084296b = copyout+0x3b
sysctl_rtsock() at 0xffffffff806a5ef7 = sysctl_rtsock+0x499
sysctl_root() at 0xffffffff805eab9e = sysctl_root+0xea
userland_sysctl() at 0xffffffff805eae6e = userland_sysctl+0x14f
sysctl() at 0xffffffff805eb258 = sysctl+0x9a
amd64_syscall() at 0xffffffff80844065 = amd64_syscall+0x145
Xfast_syscall() at 0xffffffff8082821c = Xfast_syscall+0xfc

login: panic: _mtx_lock_sleep: recursed on non-recursive mutex process lock @ ../../../amd64/amd64/trap.c:731
cpuid = 0
KDB: stack backtrace:
gdb_trace_self_wrapper() at 0xffffffff8057e7ea = gdb_trace_self_wrapper+0x2a
kdb_backtrace() at 0xffffffff8062ffdc = kdb_backtrace+0x37
panic() at 0xffffffff805f89ca = panic+0x2ad
_mtx_lock_flags() at 0xffffffff805e9376 = _mtx_lock_flags
_mtx_lock_flags() at 0xffffffff805e9417 = _mtx_lock_flags+0xa1
trap_pfault() at 0xffffffff80880450 = trap_pfault+0xa1
trap() at 0xffffffff80880ac7 = trap+0x4b8
calltrap() at 0xffffffff80861af8 = calltrap+0x8
--- trap 0xc, rip = 0xffffffff8087de8b, rsp = 0xffffff807b7e9410, rbp = 0xffffff807b7e9440 ---
copyout() at 0xffffffff8087de8b = copyout+0x3b
sysctl_out_proc() at 0xffffffff805ed305 = sysctl_out_proc+0x16c
sysctl_root() at 0xffffffff80606141 = sysctl_root+0x13a
userland_sysctl() at 0xffffffff8060640a = userland_sysctl+0x14f
sysctl() at 0xffffffff806067f8 = sysctl+0x9a
amd64_syscall() at 0xffffffff8087f635 = amd64_syscall+0x145
Xfast_syscall() at 0xffffffff80861dec = Xfast_syscall+0xfc
--- syscall (202, FreeBSD ELF64, sysctl), rip = 0x801c12b0c, rsp = 0x7fffffffb768, rbp = 0x7fffffffb7b0 ---
--- curthread 0xffffff000465b000, tid 100142

After doing some instrumentation, I think I've figured out what is causing this.  It seems that when I am wiring the page, in some situations the page table entry is being changed from read-only -> read-write as well as being wired.  I haven't figured out the exact scenario that causes this, but I can definitely see it in my added trace.  Here is an example page table entry transition I am seeing in pmap_enter that is called as a result of the wire:

pmap_enter: origpte: 80000000ad201425 newpte: 80000000ad201607

This means that we are setting PG_W (wired) and PG_RW (read/write) in this pmap_enter operation.  Everytime I saw a page-fault after wiring it was immediately preceded by a transition like this (in the cases that did not page fault, the page table entry already had PG_RW set).  This made me suspect that a read-only version of the page table entry was cached in the TLB.  I noticed we invalidate in some situations in pmap_enter, but this transition is not one of them.  I was able to eliminate the panics by making this change:

diff --git a/src/sys/amd64/amd64/pmap.c b/src/sys/amd64/amd64/pmap.c
--- a/src/sys/amd64/amd64/pmap.c
+++ b/src/sys/amd64/amd64/pmap.c
@@ -3251,6 +3251,11 @@ validate:
                                if (opa != VM_PAGE_TO_PHYS(m) || ((origpte &
                                    PG_NX) == 0 && (newpte & PG_NX)))
                                        invlva = TRUE;
+                               if ((newpte & PG_W) &&
+                                   ((origpte & PG_RW) == 0) &&
+                                   (newpte & PG_RW)) {
+                                       invlva = TRUE;
+                               }
                       }
                        if ((origpte & (PG_M | PG_RW)) == (PG_M | PG_RW)) {
                                if ((origpte & PG_MANAGED) != 0)

I wanted to see if anyone has seen issues in this area, and if this fix seems appropriate.  I'm running 8.2, but I didn't see any obvious changes to pmap stuff in head which would change this behavior.  Thanks for any feedback,

  Jeremiah Lott
  Avere Systems