From: Andrew Morton Subject: Re: [Fwd: [Bug 3003] New: might_sleep warning when setting up IPSec with IPCOMP] Date: Fri, 2 Jul 2004 14:39:49 -0700 Sender: netdev-bounce@oss.sgi.com Message-ID: <20040702143949.21a50b74.akpm@osdl.org> References: <40E5A1B3.2020202@us.ibm.com> <40E5D326.5000509@us.ibm.com> Mime-Version: 1.0 Content-Type: text/plain; charset=US-ASCII Content-Transfer-Encoding: 7bit Cc: niv@us.ibm.com, jmorris@intercode.com.au, netdev@oss.sgi.com, christophe@saout.de, mjbligh@us.ibm.com Return-path: To: Nivedita Singhvi In-Reply-To: <40E5D326.5000509@us.ibm.com> Errors-to: netdev-bounce@oss.sgi.com List-Id: netdev.vger.kernel.org Nivedita Singhvi wrote: > > Nivedita Singhvi wrote: > > > James, Andrew, > > > > Looks like deflate_comp_init() is not calling > > __vmalloc in a kosher way. > > > Jul 2 18:39:04 websrv Debug: sleeping function called from invalid > > context at > > mm/slab.c:1994 > > Jul 2 18:39:04 websrv in_atomic():1, irqs_disabled():0 > > Jul 2 18:39:04 websrv [] dump_stack+0x17/0x20 > > Jul 2 18:39:04 websrv [] __might_sleep+0xb4/0xe0 > > Jul 2 18:39:04 websrv [] kmem_cache_alloc+0x5c/0x60 > > Jul 2 18:39:04 websrv [] __get_vm_area+0x20/0xf0 > > Jul 2 18:39:04 websrv [] get_vm_area+0x24/0x30 > > Jul 2 18:39:04 websrv [] __vmalloc+0x3c/0x100 > > Jul 2 18:39:04 websrv [] deflate_comp_init+0x4a/0xf0 > > Jul 2 18:39:04 websrv [] deflate_compress+0x24/0x80 > > Jul 2 18:39:04 websrv [] crypto_compress+0x24/0x30 > > Jul 2 18:39:04 websrv [] ipcomp_compress+0x6c/0x110 > > Jul 2 18:39:04 websrv [] ipcomp_output+0xc1/0x370 > > We are grabbing dst->xfrm lock in ipcomp_output(), > and have it held when we call ipcomp_compress(). > > Is that the issue? I don't have the crypto module > code, but in_atomic() will be true. > Yes, that is the issue. >From a design point-of-view, it is not idiomatic for deflate_compress() to be doing this magical lazy initialisation. It would be better if the client of the deflate.c code were to explicitly initialise the context before diving in and using it. /* * Lazy initialization to make interface simple without allocating * un-needed workspaces. Thus can be called in softirq context. */ static int deflate_comp_init(struct deflate_ctx *ctx) Well no. Those games with deflate_gfp() really need to go away. in_atomic() works OK if CONFIG_PREEMPT is enabled. But with CONFIG_PREEMPT=n, in_atomic() returns false inside spinlock. And in_atomic()'s return value is unaffected by local_irq_disable(). This all needs to be redesigned, sorry.