Recently upstream Xen added support for having xvd devices > 16. For the most part, this doesn't really concern libvirt, since for things like attach and detach we just pass it through and let xend worry about whether it is supported or not. The one place this breaks down is in the stats collecting code, where we need to figure out the device number so we can go digging in /sys for the statistics. To remedy this, I've re-written xenLinuxDomainDeviceID() to use regular expressions to figure out the device number from the name. The major advantage is that now xenLinuxDomainDeviceID() looks fairly identical to tools/python/xen/util/blkif.py (in the Xen sources), so that adding additional devices in the future should be much easier. It also reduces the size of the code, and, in my opinion, the code complexity. With this patch in place, I was able to get block statistics both on older style devices (/dev/xvda) and on the new, expanded devices (/dev/xvdaa).

- if (strlen (path) >= 4 && - STRPREFIX (path, "xvd")) { - /* Xen paravirt device handling */ - disk = (path[3] - 'a'); - if (disk < 0 || disk > 15) { - statsErrorFunc (conn, VIR_ERR_INVALID_ARG, __FUNCTION__, - "invalid path, device names must be in range xvda - xvdp", - domid); - return -1; - }

- return (majors[disk/2] * 256) + ((disk % 2) * 63) + part; + if (disk_re_match("/dev/sd[a-z]([1-9]|1[0-5])?$", mod_path, &part)) { + major = scsi_majors[(mod_path[7] - 'a') / 16]; + minor = ((mod_path[7] - 'a') % 16) * 16 + part; + retval = major * 256 + minor; } + else if (disk_re_match("/dev/sd[a-i][a-z]([1-9]|1[0-5])?$", + mod_path, &part)) { + major = scsi_majors[((mod_path[7] - 'a' + 1) * 26 + (mod_path[8] - 'a')) / 16]; + minor = (((mod_path[7] -'a' + 1) * 26 + (mod_path[8] - 'a')) % 16) + * 16 + part; + retval = major * 256 + minor; + } + else if (disk_re_match("/dev/hd[a-t]([1-9]|[1-5][0-9]|6[0-3])?", + mod_path, &part)) { + major = ide_majors[(mod_path[7] - 'a') / 2]; + minor = ((mod_path[7] - 'a') % 2) * 64 + part; + retval = major * 256 + minor; + } + else if (disk_re_match("/dev/xvd[a-p]([1-9]|1[0-5])?$", mod_path, &part)) + retval = (202 << 8) + ((mod_path[8] - 'a') << 4) + part; + else if (disk_re_match("/dev/xvd[q-z]([1-9]|1[0-5])?$", mod_path, &part)) + retval = (1 << 28) + ((mod_path[8] - 'a') << 8) + part; + else if (disk_re_match("/dev/xvd[a-i][a-z]([1-9]|1[0-5])?$", + mod_path, &part)) + retval = (1 << 28) + (((mod_path[8] - 'a' + 1) * 26 + (mod_path[9] - 'a')) << 8) + part; + else + statsErrorFunc (conn, VIR_ERR_INVALID_ARG, __FUNCTION__, + "unsupported path, use xvdN, hdN, or sdN", domid);

Daniel P. Berrange wrote: > On Tue, Aug 05, 2008 at 12:12:21PM +0200, Chris Lalancette wrote: >> Recently upstream Xen added support for having xvd devices > 16. For the most >> part, this doesn't really concern libvirt, since for things like attach and >> detach we just pass it through and let xend worry about whether it is supported >> or not. The one place this breaks down is in the stats collecting code, where >> we need to figure out the device number so we can go digging in /sys for the >> statistics. >> >> To remedy this, I've re-written xenLinuxDomainDeviceID() to use regular >> expressions to figure out the device number from the name. The major advantage >> is that now xenLinuxDomainDeviceID() looks fairly identical to >> tools/python/xen/util/blkif.py (in the Xen sources), so that adding additional >> devices in the future should be much easier. It also reduces the size of the >> code, and, in my opinion, the code complexity. >> >> With this patch in place, I was able to get block statistics both on older style >> devices (/dev/xvda) and on the new, expanded devices (/dev/xvdaa). > > This patch breaks the test suite for disk name -> device ID conversion. > The test suite also needs to have more tests added to cover the new > interesting boundary conditions for xvdXXX naming. OK. Well, there were 3 different problems with the test suite: 1) A number of tests were actually wrong. For instance, there is a DO_TEST("/dev/hdt", 23359); but Xen actually uses the encoding "major*256 + minor + part". So in this case, the major is 91, and the minor is 64 (according to http://www.lanana.org/docs/device-list/devices.txt), so that would be 23360. I've fixed the wrong tests now, and I'll re-submit it with the updated patch. 2) In the hda0 and hda64 case, the upstream Xen regex isn't tight enough. I've tightened it up in the libvirt patch, so these now pass. 3) The upstream Xen regex's allows /dev/sdi{w,x,y,z}, although they aren't legal devices. I've fixed up my regex to handle this. Attached is an updated patch with the above fixes both to my code and to the test suite. As far as the error reporting goes, I won't argue that my patch gives slightly less information. However, that being said, I have to believe that the most likely use of block statistics is something like:

virsh dumpxml <dom> ...see what devices are listed there virsh domblkstats <dom> <device> In which case the slightly less verbose error reporting won't matter a whole lot.

To retain the diagnostic Dan mentioned, you should be able to insert something like this just before the final "else": else if (disk_re_match("/dev/sd[a-z]([0-9])+$", mod_path, &part)) { > + else > + statsErrorFunc (conn, VIR_ERR_INVALID_ARG, __FUNCTION__, > + "unsupported path, use xvdN, hdN, or sdN", domid);

Jim Meyering wrote: > To retain the diagnostic Dan mentioned, you should be able to > insert something like this just before the final "else": > > else if (disk_re_match("/dev/sd[a-z]([0-9])+$", mod_path, &part)) { > >> + else >> + statsErrorFunc (conn, VIR_ERR_INVALID_ARG, __FUNCTION__, >> + "unsupported path, use xvdN, hdN, or sdN", domid); OK, another go, with all of Jim's concerns addressed. I did something like this last point (thanks for the idea Jim), except that I didn't use regex's but basic STRPREFIX() to get better error messages. Dan, is this better?

Recently upstream Xen added support for having xvd devices > 16. For the most part, this doesn't really concern libvirt, since for things like attach and detach we just pass it through and let xend worry about whether it is supported or not. The one place this breaks down is in the stats collecting code, where we need to figure out the device number so we can go digging in /sys for the statistics. To remedy this, I've re-written xenLinuxDomainDeviceID() to use regular expressions to figure out the device number from the name. The major advantage is that now xenLinuxDomainDeviceID() looks fairly identical to tools/python/xen/util/blkif.py (in the Xen sources), so that adding additional devices in the future should be much easier. It also reduces the size of the code, and, in my opinion, the code complexity. With this patch in place, I was able to get block statistics both on older style devices (/dev/xvda) and on the new, expanded devices (/dev/xvdaa).

Signed-off-by: Chris Lalancette <clalance(a)redhat.com&gt;

+ /* OK, we have a match; see if we have a partition */ + *part = 0; + if (pmatch[1].rm_so != -1) + *part = strtol(path + pmatch[1].rm_so, NULL, 10);