On Thu, 10/08 19:59, charlie.song wrote: > Dear KVM Developers: > I am Xiang Song from UCloud company. We currently encounter a weird phenomenon about Qemu-KVM IOthread. > We recently try to use Linux AIO from guest OS and find that the IOthread mechanism of Qemu-KVM will reorder I/O requests from guest OS > even when the AIO write requests are issued from a single thread in order. This does not happen on the host OS however. > We are not sure whether this is a feature of Qemu-KVM IOthread mechanism or a Bug. > > The testbd is as following: (the guest disk device cache is configured to writethrough.) > CPU: Intel(R) Xeon(R) CPU E5-2650 > QEMU version: 1.5.3 > Host/Guest Kernel: Both Linux 4.1.8 & Linux 2.6.32, OS type CentOS 6.5 > Simplified Guest OS qemu cmd: > /usr/libexec/qemu-kvm -machine rhel6.3.0,accel=kvm,usb=off -cpu kvm64 -smp 8,sockets=8,cores=1,threads=1 > -drive file=/var/lib/libvirt/images/song-disk.img,if=none,id=drive-virtio-disk0,format=qcow2,serial=UCLOUD_DISK_VDA,cache=writethrough > -device virtio-net-pci,netdev=hostnet0,id=net0,mac=52:54:00:22:d5:52,bus=pci.0,addr=0x4 You mentioned iothread above but it's not in your command line? > > The test code triggerring this phenomenon work as following: it use linux aio API to issue concurrent async write requests to a file. During exection it will > continuously write data into target test file. There are total 'X' jobs, and each job is assigned a job id JOB_ID which starts from 0. Each job will write 16 * 512 > Byte data into the target file at offset = JOB_ID * 512. (the data is repeated uint64_t JOB_ID). > There is only one thread handling 'X' jobs one by one through Linux AIO (io_submit) cmd. When handling jobs, it will continuously > issuing AIO requests without waiting for AIO Callbacks. When it finishes, the file should look like: > [0....0][1...1][2...2][3...3]...[X-1...X-1] > Then we use a check program to test the resulting file, it can continuously read the first 8 byte (uint64_t) of each sector and print it out. In normal cases, > it's output is like: > 0 1 2 3 .... X-1 > > Exec output: (Set X=32) > In our guest OS, the output is abnormal: 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 18 18 18 18 18 24 25 26 27 28 29 30 31. > It can be seen that job20~job24 are overwrited by job19. > In our host OS, the output is as expected, 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31. I'm not 100% sure but I don't think the returning of io_submit guarantees any ordering, usually you need to wait for the callback to ensure that. Fam > > > I can provide the example code if needed. > > Best regards, song > > 2015-10-08 > > > charlie.song > > -- > To unsubscribe from this list: send the line "unsubscribe kvm" in > the body of a message to majordomo@xxxxxxxxxxxxxxx > More majordomo info at http://vger.kernel.org/majordomo-info.html -- To unsubscribe from this list: send the line "unsubscribe kvm" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
