Re: [Qemu-devel] DAX can not work on virtual nvdimm device

[Ross Zwisler]

On Mon, Aug 29, 2016 at 03:54:10PM +0800, Xiao Guangrong wrote:
> 
> Hi Ross,
> 
> Sorry for the delay, i just returned back from KVM Forum.
> 
> On 08/20/2016 02:30 AM, Ross Zwisler wrote:
> > On Fri, Aug 19, 2016 at 07:59:29AM -0700, Dan Williams wrote:
> > > On Fri, Aug 19, 2016 at 4:19 AM, Xiao Guangrong
> > > <address@hidden> wrote:
> > > > 
> > > > Hi Dan,
> > > > 
> > > > Recently, Redhat reported that nvml test suite failed on QEMU/KVM,
> > > > more detailed info please refer to:
> > > >    https://bugzilla.redhat.com/show_bug.cgi?id=1365721
> > > > 
> > > > The reason for this bug is that the memory region created by mmap()
> > > > on the dax-based file was gone so that the region can not be found
> > > > in /proc/self/smaps during the runtime.
> > > > 
> > > > This is a simple way to trigger this issue:
> > > >    mount -o dax /dev/pmem0 /mnt/pmem/
> > > >    vim /mnt/pmem/xxx
> > > > then 'vim' is crashed due to segment fault.
> > > > 
> > > > This bug can be reproduced on your tree, the top commit is
> > > > 10d7902fa0e82b (dax: unmap/truncate on device shutdown), the kernel
> > > > configure file is attached.
> > > > 
> > > > Your thought or comment is highly appreciated.
> > > 
> > > I'm going to be offline until Tuesday, but I will investigate when I'm
> > > back.  In the meantime if Ross or Vishal had an opportunity to take a
> > > look I wouldn't say "no" :).
> > 
> > I haven't been able to reproduce this vim segfault.  I'm using QEMU v2.6.0,
> > and the kernel commit you mentioned, and your kernel config.
> > 
> > Here's my QEMU command line:
> > 
> > sudo ~/qemu/bin/qemu-system-x86_64 /var/lib/libvirt/images/alara.qcow2 \
> > -machine pc,nvdimm -m 8G,maxmem=100G,slots=100  -object \
> > memory-backend-file,id=mem1,share,mem-path=/dev/pmem0,size=8G -device \
> > nvdimm,memdev=mem1,id=nv1 -smp 6 -machine pc,accel=kvm
> > 
> > With this I'm able to mkfs the guest's /dev/pmem0, mount it with -o dax, and
> > write a file with vim.
> 
> Thanks for your test. That's strange...
> 
> > 
> > Can you reproduce your results with a pmem device created via a memmap 
> > kernel
> > command line parameter in the guest?  You'll need to update your kernel
> > config to enable CONFIG_X86_PMEM_LEGACY and CONFIG_X86_PMEM_LEGACY_DEVICE.
> > 
> 
> Okay, i tested it with mmap=6G!10G, it failed too. So it looks like it's a
> filesystem or DAX issue.
> 
> More precisely, i figured out the root case that read() returns a wrong value
> when it reaches the end of the file, following test case can trigger it:
> 
> #include <stdlib.h>
> #include <stdio.h>
> #include <unistd.h>
> #include <fcntl.h>
> #include <sys/types.h>
> #include <sys/stat.h>
> 
> int main(int argc, char *argv[])
> {
>       char *filename;
> 
>       if (argc < 2) {
>               printf("arg: filename.\n");
>               return -1;
>       }
> 
>       filename = argv[1];
>       printf("test on %s.\n", filename);
> 
>       int fd = open(filename, O_RDWR);
> 
>       if (fd < 0) {
>               perror("open");
>               return -1;
>       }
> 
>       int count = 0;
> 
>       while (1) {
>               ssize_t ret;
>               char buf;
> 
>               ret = read(fd, &buf, sizeof(buf));
>               if (ret < 0) {
>                       perror("READ");
>                       return -1;
>               }
> 
>               if (ret == 0)
>                       break;
>               if (ret != sizeof(buf)) {
>                       printf("Count %x Ret %lx sizeof(buf) %lx.\n",
>                               count, ret, sizeof(buf));
>                       return -1;
>               }
> 
>               count++;
>               printf("%c", buf);
>       }
> 
>       printf("\n Good Read.\n");
>       return 0;
> }
> 
> 
> 
> It will fail at "ret != sizeof(buf)", for example, the error output on my
> test env is:
>       Count 1000 Ret 22f84200 sizeof(buf) 1.

Can you please verify that you are using "usable" memory for your memmap?  All
the details are here:

https://nvdimm.wiki.kernel.org/how_to_choose_the_correct_memmap_kernel_parameter_for_pmem_on_your_system

My guess is that Boaz was correct, and that your memmap is off using addresses
that don't actually map to memory.

- Ross