From mboxrd@z Thu Jan 1 00:00:00 1970 From: Cong Wang Subject: Re: [net-next PATCH v4 3/3] net: reserve ports for applications using fixed port numbers Date: Thu, 18 Feb 2010 00:07:08 +0800 Message-ID: <4B7C142C.9040707@redhat.com> References: <1266271241-6293-1-git-send-email-opurdila@ixiacom.com> <201002161306.29708.opurdila@ixiacom.com> <4B7A9852.5020105@redhat.com> <201002161625.22495.opurdila@ixiacom.com> Mime-Version: 1.0 Content-Type: text/plain; charset=ISO-8859-1; format=flowed Content-Transfer-Encoding: 7bit Cc: David Miller , Linux Kernel Network Developers , Linux Kernel Developers , Neil Horman , Eric Dumazet To: Octavian Purdila Return-path: In-Reply-To: <201002161625.22495.opurdila@ixiacom.com> Sender: linux-kernel-owner@vger.kernel.org List-Id: netdev.vger.kernel.org Octavian Purdila wrote: > On Tuesday 16 February 2010 15:06:26 you wrote: >> Octavian Purdila wrote: >>> On Tuesday 16 February 2010 11:37:04 you wrote: >>>>> BUILD_BUG_ON(sizeof(struct inet_skb_parm) > sizeof(dummy_skb->cb)); >>>>> >>>>> + sysctl_local_reserved_ports = kzalloc(65536 / 8, GFP_KERNEL); >>>>> + if (!sysctl_local_reserved_ports) >>>>> + goto out; >>>>> + >>>> I think we should also consider the ports in ip_local_port_range, >>>> since we can only reserve the ports in that range. >>> That is subject to changes at runtime, which means we will have to >>> readjust the bitmap at runtime which introduces the need for additional >>> synchronization operations which I would rather avoid. >> Why? As long as the bitmap is global, this will not be hard. >> > > For the more important point see bellow, but with regard to reallocation, this > means we need to at least use rcu_read_lock() in the fast path to avoid races > between freeing the old bitmap and doing a read in progress. > > Granted, that is a light operation, but would it makes things so much more > complicated just so that we save one memory page (assuming the range is the > default [32000 64000] one). Why not just allocate the bitmap for all ports? 65535/8 bytes are needed. > >> Consider that if one user writes a port number which is beyond >> the ip_local_port_range into ip_local_reserved_ports, we should >> not accept this, because it doesn't make any sense. But with your >> patch, we do. >> > > I think it should be allowed. I see ip_local_reserved_ports and ip_local_range > as independent settings that can be change at any time. According to the original purpose, they are not. > > That way I can flag port 8080 even if the current range is [32000, 64000] and > then later I can expand the range to [1024, 64000] without loosing the 8080 > reservation. Then its meaning is changed, bind(0) will never have chance to get 8080, thus reserving 8080 for this purpose fails. I want to always keep its original meaning, if the local_port_range goes out, then local_reserved_port should be empty at the same time, you have to reset it after changing local_port_range.