Re: uma_zalloc_arg: zone "256" with non-sleepable exclusive rw ifnet_rw @ /usr/src/sys/net/if.c:414

From: Sergey Kandaurov <pluknet_at_gmail.com>
Date: Wed, 19 Jan 2011 17:51:59 +0300
On 18 January 2011 22:04, John Baldwin <jhb_at_freebsd.org> wrote:
> On Tuesday, January 18, 2011 1:22:24 pm Sergey Kandaurov wrote:
>> On 18 January 2011 17:54, John Baldwin <jhb_at_freebsd.org> wrote:
>> > On Monday, January 17, 2011 12:55:26 pm Sergey Kandaurov wrote:
>> >> Hi,
>> >>
>> >> I see this "malloc with non-sleepable" on current during boot.
>> >> It's strange that I don't see it if I boot via pxe/nfs.
>> >>
>> >> if_alloc() calls ifindex_alloc_locked() under IFNET_WLOCK() which
>> >> might call if_grow().
>> >> Looks like a regression from r196553.
>> >
>> > I'm guessing that ifindex_alloc() should drop the lock and retry the
>> > allocation after calling if_grow()?  This compiles, but I haven't booted
> it
>> > yet:
>>
>> vnet_if_init() calls if_grow() without lock.
>
> So it does. :(  I've added locking to the sysinit to handle this:

Seems a bit more work there. The vnet_if_init() sysinit cannot use ifnet
locking since it runs before another sysinit initialized those ifnet locks.
On the other side it's safe to call if_grow() from vnet_if_init() w/o locking:

db> bt
Tracing pid 0 tid 100000 td 0xffffffff80ccffc0
_sx_xlock_hard() at _sx_xlock_hard+0xa0
_sx_xlock() at _sx_xlock+0xbb
vnet_if_init() at vnet_if_init+0x4a
mi_startup() at mi_startup+0x77
btext() at btext+0x2c

After said certain locking changes I came to the next problem.
A modified ifindex_alloc_locked() function doesn't update/increment V_if_index
in successful if_grow() case, so it ends up in quick memory exhaustion:

bce0: <Broadcom NetXtreme II BCM5709 1000Base-T (C0)> mem
0x92000000-0x93ffffff irq 28 at device 0.0 on pci4
panic: kmem_malloc(-2147483648): kmem_map too small: 1110716416 total allocated
cpuid = 0
KDB: enter: panic
[ thread pid 0 tid 100000 ]
Stopped at      kdb_enter+0x3d: movq    $0,0x700140(%rip)
db> bt
Tracing pid 0 tid 100000 td 0xffffffff80ccffc0
kdb_enter() at kdb_enter+0x3d
panic() at panic+0x180
kmem_malloc() at kmem_malloc+0x25d
uma_large_malloc() at uma_large_malloc+0x4a
malloc() at malloc+0x15d
if_grow() at if_grow+0x98
if_alloc() at if_alloc+0xd8
bce_attach() at bce_attach+0x18de
[...]
> show malloc
   ifnet            2      1048578K           25

So I added V_if_index increment into if_grow() itself.
At least it boots now :)

There are 18 visible interfaces in the test system, so it looks appropriate
minus one extra elm.
        ifnet    19 16420K       -       36  128,256,512,1024,2048,4096
db> show all ifnet
                bce0 ifp=0xfffffe0019426000
                igb0 ifp=0xfffffe0019425800
                igb1 ifp=0xfffffe0019425000
                igb2 ifp=0xfffffe0016caf800
                igb3 ifp=0xfffffe00195d3800
                igb4 ifp=0xfffffe00195d2800
                igb5 ifp=0xfffffe00195d2000
              usbus0 ifp=0xfffffe0019426800
              usbus1 ifp=0xfffffe00195d3000
              usbus2 ifp=0xfffffe0019902800
              usbus3 ifp=0xfffffe0019902000
              usbus4 ifp=0xfffffe0019901800
              usbus5 ifp=0xfffffe0019901000
              usbus6 ifp=0xfffffe0019900800
              usbus7 ifp=0xfffffe0019900000
                igb6 ifp=0xfffffe0019d99000
                igb7 ifp=0xfffffe0019d98800
                 lo0 ifp=0xfffffe0019d98000

%%%
Index: sys/net/if.c
===================================================================
--- sys/net/if.c        (revision 217503)
+++ sys/net/if.c        (working copy)
_at__at_ -266,6 +266,7 _at__at_ ifindex_alloc_locked(u_short *idxp)

        IFNET_WLOCK_ASSERT();

+retry:
        /*
         * Try to find an empty slot below V_if_index.  If we fail, take the
         * next slot.
_at__at_ -278,10 +279,11 _at__at_ ifindex_alloc_locked(u_short *idxp)
        /* Catch if_index overflow. */
        if (idx < 1)
                return (ENOSPC);
-       if (idx > V_if_index)
-               V_if_index = idx;
-       if (V_if_index >= V_if_indexlim)
+       if (idx > V_if_index) {
                if_grow();
+               goto retry;
+       }
+       V_if_index = idx;
        *idxp = idx;
        return (0);
 }
_at__at_ -351,7 +353,10 _at__at_ vnet_if_init(const void *unused __unused)

        TAILQ_INIT(&V_ifnet);
        TAILQ_INIT(&V_ifg_head);
+       IFNET_LOCK_INIT();    /* XXX */
+       IFNET_WLOCK();
        if_grow();                              /* create initial table */
+       IFNET_WUNLOCK();
        vnet_if_clone_init();
 }
 VNET_SYSINIT(vnet_if_init, SI_SUB_INIT_IF, SI_ORDER_FIRST, vnet_if_init,
_at__at_ -362,7 +367,6 _at__at_ static void
 if_init(void *dummy __unused)
 {

-       IFNET_LOCK_INIT();
        if_clone_init();
 }
 SYSINIT(interfaces, SI_SUB_INIT_IF, SI_ORDER_SECOND, if_init, NULL);
_at__at_ -385,16 +389,26 _at__at_ VNET_SYSUNINIT(vnet_if_uninit, SI_SUB_INIT_IF, SI_
 static void
 if_grow(void)
 {
+       int oldlim;
        u_int n;
        struct ifindex_entry *e;

-       V_if_indexlim <<= 1;
-       n = V_if_indexlim * sizeof(*e);
+       IFNET_WLOCK_ASSERT();
+       oldlim = V_if_indexlim;
+       IFNET_WUNLOCK();
+       n = (oldlim << 1) * sizeof(*e);
        e = malloc(n, M_IFNET, M_WAITOK | M_ZERO);
+       IFNET_WLOCK();
+       if (V_if_indexlim != oldlim) {
+               free(e, M_IFNET);
+               return;
+       }
        if (V_ifindex_table != NULL) {
                memcpy((caddr_t)e, (caddr_t)V_ifindex_table, n/2);
                free((caddr_t)V_ifindex_table, M_IFNET);
        }
+       V_if_indexlim <<= 1;
+       V_if_index++;
        V_ifindex_table = e;
 }

%%%

-- 
wbr,
pluknet
Received on Wed Jan 19 2011 - 13:52:01 UTC

This archive was generated by hypermail 2.4.0 : Wed May 19 2021 - 11:40:10 UTC