At 2015-10-08 23:37:02, "Fam Zheng" <famz@xxxxxxxxxx> wrote: >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? I means the thread pool mechanism used by qemu-kvm to accelerate I/O processing.This is used by paio_submit (block/raw-posix.c) by default and with pool->max_threads = 64 as I know. (qemu-kvm version 1.5.3) > >> >> 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. Is there any proof or artical about the ordering of io_submit requests? > >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
