From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: from mail-pf0-x242.google.com (mail-pf0-x242.google.com [IPv6:2607:f8b0:400e:c00::242]) (using TLSv1.2 with cipher ECDHE-RSA-AES128-GCM-SHA256 (128/128 bits)) (No client certificate requested) by lists.ozlabs.org (Postfix) with ESMTPS id 3x4QRh5QThzDr9C for ; Sat, 8 Jul 2017 19:00:32 +1000 (AEST) Received: by mail-pf0-x242.google.com with SMTP id q85so7592301pfq.2 for ; Sat, 08 Jul 2017 02:00:32 -0700 (PDT) Date: Sat, 8 Jul 2017 19:00:12 +1000 From: Nicholas Piggin To: Gautham R Shenoy Cc: Michael Ellerman , Michael Neuling , Vaidyanathan Srinivasan , Shilpasri G Bhat , "Rafael J. Wysocki" , Akshay Adiga , linuxppc-dev@lists.ozlabs.org, linux-kernel@vger.kernel.org, linux-pm@vger.kernel.org Subject: Re: [PATCH 4/5] powernv:idle: Move initialization of sibling pacas to pnv_alloc_idle_core_states Message-ID: <20170708190012.6c697975@roar.ozlabs.ibm.com> In-Reply-To: <20170707150416.GD8913@in.ibm.com> References: <1499272696-28751-1-git-send-email-ego@linux.vnet.ibm.com> <1499272696-28751-5-git-send-email-ego@linux.vnet.ibm.com> <20170707011609.76c51329@roar.ozlabs.ibm.com> <20170707150416.GD8913@in.ibm.com> MIME-Version: 1.0 Content-Type: text/plain; charset=US-ASCII List-Id: Linux on PowerPC Developers Mail List List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , On Fri, 7 Jul 2017 20:34:16 +0530 Gautham R Shenoy wrote: > On Fri, Jul 07, 2017 at 01:16:09AM +1000, Nicholas Piggin wrote: > > On Wed, 5 Jul 2017 22:08:15 +0530 > > "Gautham R. Shenoy" wrote: > > > > > From: "Gautham R. Shenoy" > > > > > > On POWER9 DD1, in order to get around a hardware issue, we store in > > > every CPU thread's paca the paca pointers of all its siblings. > > > > > > Move this code into pnv_alloc_idle_core_states() soon after the space > > > for saving the sibling pacas is allocated. > > > > > > Signed-off-by: Gautham R. Shenoy > > > > > - if (cpu_has_feature(CPU_FTR_POWER9_DD1)) { > > > - int cpu; > > > - > > > - pr_info("powernv: idle: Saving PACA pointers of all CPUs in their thread sibling PACA\n"); > > > - for_each_possible_cpu(cpu) { > > > - int base_cpu = cpu_first_thread_sibling(cpu); > > > - int idx = cpu_thread_in_core(cpu); > > > - int i; > > > - > > > > You could move the thread_sibling_pacas allocation to here? > > > > Speaking of which... core_idle_state and thread_sibling_pacas are > > allocated with kmalloc_node... What happens if we take an SLB miss > > in the idle wakeup code on these guys? Nothing good I think. Perhaps > > we should put them into the pacas or somewhere in bolted memory. > > Yes, though the SLB miss hasn't yet been encountered in practise so > far! Considering it's a node-affine allocation, it may actually be possible to hit in practice on very large memory systems in practice. > While one can define thread_sibling_pacas in PACA, it doesn't make > sense to allocate space for core_idle_state in PACA since the > allocated value of the secondary threads will never be used. Well, same for core_idle_state, although that's smaller. > What is the right way to ensure that these allocations fall in the > bolted range ? I'm not sure, I guess the memblock allocator is not up anymore at this point. I think we'd have to move it earlier. You could allocate another array of them along side the paca allocation. Thanks, Nick