On Fri, 10/09 11:25, charlie.song wrote: > At 2015-10-08 23:37:02, "Fam Zheng" <[email protected]> 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 thread pool parallism may reorder non-overlapping requests, but it shouldn't cause any reordering of overlapping requests like the case in your io pattern. QEMU ensures that. Do you see this with aio=native? Fam > > > > >> > >> 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 [email protected] > >> 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 [email protected] > 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 [email protected] More majordomo info at http://vger.kernel.org/majordomo-info.html
