On 11/22/24 at 10:30am, David Hildenbrand wrote: > On 22.11.24 10:16, Baoquan He wrote: > > On 10/25/24 at 05:11pm, David Hildenbrand wrote: > > ......snip... > > > @@ -1482,6 +1470,10 @@ int vmcore_add_device_dump(struct vmcoredd_data *data) > > > return -EINVAL; > > > } > > > + /* We'll recheck under lock later. */ > > > + if (data_race(vmcore_opened)) > > > + return -EBUSY; > > > > Hi, > > > As I commented to patch 7, if vmcore is opened and closed after > > checking, do we need to give up any chance to add device dumping > > as below? > > > > fd = open(/proc/vmcore); > > ...do checking; > > close(fd); > > > > quit any device dump adding; > > > > run makedumpfile on s390; > > ->fd = open(/proc/vmcore); > > -> try to dump; > > ->close(fd); > > The only reasonable case where this could happen (with virtio_mem) would be > when you hotplug a virtio-mem device into a VM that is currently in the > kdump kernel. However, in this case, the device would not provide any memory > we want to dump: > > (1) The memory was not available to the 1st (crashed) kernel, because > the device got hotplugged later. > (2) Hotplugged virtio-mem devices show up with "no plugged memory", > meaning there wouldn't be even something to dump. > > Drivers will be loaded (as part of the kernel or as part of the initrd) > before any kdump action is happening. Similarly, just imagine your NIC > driver not being loaded when you start dumping to a network share ... > > This should similarly apply to vmcoredd providers. > > There is another concern I had at some point with changing the effective > /proc/vmcore size after someone already opened it, and might assume the size > will stay unmodified (IOW, the file was completely static before vmcoredd > showed up). > > So unless there is a real use case that requires tracking whether the file > is no longer open, to support modifying the vmcore afterwards, we should > keep it simple. > > I am not aware of any such cases, and my experiments with virtio_mem showed > that the driver get loaded extremely early from the initrd, compared to when > we actually start messing with /proc/vmcore from user space. Hmm, thanks for sharing your thoughts. I personally think if we could, we had better make code as robust as possible. Here, since we have already integrated the lock with one vmcore_mutex, whether the vmcoredd is added before or after /proc/vmcore opening/closing, it won't harm. The benefit is it works well with the current kwown case, virtio-mem probing, makedumpfile running. And it also works well with other possible cases, e.g if you doubt virtio-mem dumping and want to debug, you set rd.break or take any way to drop into emengency shell of kdump kernel, you can play it to poke virtio-mem module again and run makedumpfile manually or reversing the order or taking any testing. Then kernel implementation should not preset that user space have to run the makedumpfile after the much earlier virtio-mem probing. If we implement codes according to preset userspace scenario, that limit the future adapttion, IMHO.
