glenn@sun.uucp (Glenn C. Skinner) (08/05/86)
Index: /sys/h/mbuf.h 4.3BSD
Description:
The MCLGET macro defined in /sys/h/mbuf.h invokes the m_clalloc
routine defined in /sys/sys/uipc_mbuf.c. This latter routine
expects to be called at splimp, but the macro makes no provisions
to do so. Moreover, the macro is invoked in sosend(), defined in
/sys/sys/uipc_socket.c, while the processor level is at spl0.
Repeat-By:
I haven't (yet) observed crashes due to this problem. However,
one would expect to see it manifest during periods of intense
network activity when processes are sending sufficiently many
large packets back and forth to exceed the initial cluster mbuf
allocation.
Fix:
An obvious patch is to modify MCLGET's definition to surround
the m_clalloc call with the proper spl calls:
#define MCLGET(m) \
{ struct mbuf *p; \
+ int ms = splimp(); \
if (mclfree == 0) \
(void)m_clalloc(1, MPG_CLUSTERS, M_DONTWAIT); \
+ splx(ms); \
MCLALLOC(p, 1); \
if (p) { \
(m)->m_off = (int)p - (int)(m); \
(m)->m_len = CLBYTES; \
} \
}
However, this fix isn't very satisfactory, because the MCLALLOC
macro is called immediately afterward, and the first thing it
does is to go to splimp. Better would be to revise the definitions
of both macros to have MCLGET go to splimp immediately and not
splx until after the MCLALLOC invocation. MCLALLOC itself then
would refrain from spl calls. Other callers would have to take
this change into account.
-- Glenn Skinner, SMI