Messages in this thread | | | Date | Thu, 7 Sep 2006 23:26:28 -0700 | From | Stephane Eranian <> | Subject | Re: [PATCH 8/18] 2.6.17.9 perfmon2 patch for review: event sets and multiplexing support |
| |
Andrew,
On Wed, Sep 06, 2006 at 07:21:35PM -0700, Andrew Morton wrote: > On Wed, 6 Sep 2006 07:50:31 -0700 > Stephane Eranian <eranian@hpl.hp.com> wrote: > > > > > + > > > > + cachep = ctx->flags.mapset ? pfm_set_cachep : pfm_lg_set_cachep; > > > > + > > > > + new_set = kmem_cache_alloc(cachep, SLAB_ATOMIC); > > > > > > SLAB_ATOMIC is unreliable. Is it possible to use SLAB_KERNEL here? If > > > coms ecallers can sleep and others cannot then passing in the gfp_flags > > > would permit improvement here. > > > > > > > I made some changes and now I know I execute this part of the function > > with interrupts disabled, holding only the perfmon context lock. I assume > > SLAB_KERNEL means, we can sleep. I think I can make this change safely. > > > > > > > > > > > + if (ctx->flags.mapset) { > > > > + view_size = PAGE_ALIGN(sizeof(struct pfm_set_view)); > > > > + view = vmalloc(view_size); > > > > > > vmalloc() sleeps, so this _could_ have used SLAB_ATOMIC. > > > > > > > I am not sure I follow you here. Are you talking about eh kmem_cache_alloc() > > above? > > > > My logic was as follows: > > a) vmalloc() can sleep > > b) Stephane at some time tested this conde with > CONFIG_DEBUG_SPINLOCK_SLEEP and didn't get sleep-while-atomic warnings out of > that vmalloc(). > > c) Hence this code is never called under spinlock, or with local > interrupts disabled. > > d) Hence it is safe to convert the earlier SLAB_ATOMIC into SLAB_KERNEL. > > > If b) is false then it's the vmalloc() call which is incorrect, not the > SLAB_ATOMIC.
Looking at the code again, I now think that vmalloc is wrong. I have made some changes to lift the restrictions on interrupts being masked, but I still need to hold a spinlock. So I think, I need to replace vmalloc with kmalloc and SLAB_ATOMIC. Furthermore, I think I need to surround this with a pair of preempt_disable/preempt_enable (given the interrupts are unmasked).
Thanks.
--
-Stephane - To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/
| |