From mboxrd@z Thu Jan 1 00:00:00 1970 Received: from eggs.gnu.org ([2001:4830:134:3::10]:51839) by lists.gnu.org with esmtp (Exim 4.71) (envelope-from ) id 1Zf3Js-0003Xt-Ls for qemu-devel@nongnu.org; Thu, 24 Sep 2015 06:00:10 -0400 Received: from Debian-exim by eggs.gnu.org with spam-scanned (Exim 4.71) (envelope-from ) id 1Zf3Jo-0001q5-Gv for qemu-devel@nongnu.org; Thu, 24 Sep 2015 06:00:04 -0400 Received: from [59.151.112.132] (port=54340 helo=heian.cn.fujitsu.com) by eggs.gnu.org with esmtp (Exim 4.71) (envelope-from ) id 1Zf3Jn-0001oe-LS for qemu-devel@nongnu.org; Thu, 24 Sep 2015 06:00:00 -0400 Message-ID: <5603C997.20600@cn.fujitsu.com> Date: Thu, 24 Sep 2015 17:59:51 +0800 From: Yang Hongyang MIME-Version: 1.0 References: <1442405768-23019-1-git-send-email-yanghy@cn.fujitsu.com> <1442405768-23019-2-git-send-email-yanghy@cn.fujitsu.com> <87a8sctg74.fsf@blackfin.pond.sub.org> <5603B5C8.7080201@cn.fujitsu.com> <87eghop2zt.fsf@blackfin.pond.sub.org> In-Reply-To: <87eghop2zt.fsf@blackfin.pond.sub.org> Content-Type: text/plain; charset="utf-8"; format=flowed Content-Transfer-Encoding: 7bit Subject: Re: [Qemu-devel] [PATCH v11 01/12] qmp: delete qemu opts when delete an object List-Id: List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , To: Markus Armbruster Cc: thuth@redhat.com, zhang.zhanghailiang@huawei.com, lizhijian@cn.fujitsu.com, jasowang@redhat.com, qemu-devel@nongnu.org, stefanha@redhat.com, Paolo Bonzini On 09/24/2015 05:42 PM, Markus Armbruster wrote: > Yang Hongyang writes: > >> On 09/24/2015 03:43 PM, Markus Armbruster wrote: >>> This has finally reached the front of my review queue. I apologize for >>> the loooong delay. >>> >>> Copying Paolo for another pair of eyeballs (he wrote this code). >>> >> [...] >>>> + >>>> + opts = qemu_opts_find(qemu_find_opts_err("object", NULL), id); >>>> + qemu_opts_del(opts); >>> >>> qemu_find_opts_err("object", &error_abort) please, because when it >>> fails, we want to die right away, not when the null pointer it returns >>> gets dereferenced. >> >> Thanks for the review. >> Jason, do you want me to propose a fix on top of this series or simply drop >> this for now because this patch is an independent bug fix and won't affect the >> other filter patch series. >> >>> >>> Same sloppiness in netdev_del_completion() and qmp_netdev_del(), not >>> your patch's fault. >>> >>> Elsewhere, we store the QemuOpts in the object just so we can delete it: >>> DeviceState, DriveInfo. Paolo, what do you think? >> >> I don't get it. Currently, only objects created at the beginning through >> QEMU command line will be stored in the QemuOpts, objects that created >> with object_add won't stored in QemuOpts. Do you mean for DeviceState, >> DriveInfo they store there QemuOpts explicity so that they can delete it? >> Why don't we just delete it from objects directly instead? > > Let me elaborate. Thanks very much for the elaboration. > > We have the same pattern in multiple places: some kind of object gets > configured via QemuOpts, and an object's QemuOpts need to stay around > until the object dies. > > Example 1: Block device backends > > DriveInfo has a member opts. > > drive_new() stores the QemuOpts in dinfo->opts. > > drive_info_del() destroys dinfo->opts. > > Note: DriveInfo member opts is always non-null. But not every > BlockBackend has a DriveInfo. > > Example 2: Device frontends > > DeviceState has a member opts. > > qdev_device_add() stores the QemuOpts in dev->opts. > > device_finalize() destroys dev->opts. > > Note: DeviceState member opts may be null (not every device is > created by qdev_device_add()). Fine, because qemu_opts_del(NULL) is > a no-op. > > Example 3: Character device backends > > CharDriverState has a member opts. > > qemu_chr_new_from_opts() stores the QemuOpts in chr->opts. > > qemu_chr_delete() destroys chr->opts. 1-3 store there ops in there own state, not in global ops group right? > > Example 4: Network device backends > > Two cases > > A. netdev > > qmp_netdev_add() does not store the QemuOpts. > > qmp_netdev_del() still needs to destroy it. It has to find it > somehow. Here's how it does it: > > opts = qemu_opts_find(qemu_find_opts_err("netdev", NULL), id); > if (!opts) { > error_setg(errp, "Device '%s' is not a netdev", id); > return; > } > > The !opts condition is a non-obvious way to test "not created > with -netdev", see commit 645c949. Note that the commit's claim > that qemu_opts_del(NULL) crashes is no longer true since commit > 4782183. > > B. Legacy net > > hmp_host_net_add() does not store the QemuOpts. I'm afraid it does store the QemuOpts, but not in it's own state. net/net.c: 1088 qemu_opt_set(opts, "type", device, &error_abort); This will store the QemuOpts, or am I misunderstood it? > > hmp_host_net_remove() still needs to destroy it. I can't see > where that happens, and I'm not sure it does. > > Example 5: Generic object > > object_create() does not store the QemuOpts. > > It still needs to be destroyed along with the object. It isn't, and > your patch fixes it. > > Personally, I find the technique in example 1-3 easier to understand > than the one in example 4-5. I agree that opts should not be used to determine not created something while there's case when something created but Opts not stored. > . > -- Thanks, Yang.