At 2015-10-09 12:16:03, "Fam Zheng" <[email protected]> wrote: >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? We see this with both aio=threads and aio=native. Are you sure "QEMU ensures the ordering of overlapping requests"? I can provide the example code if needed.
> >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
