11:28 a.m.
Hello,
the attached patch adds support for copy-on-write images: creating them,
deleting them, identifying them.
Mirek
11:58 a.m.
On Tue, Dec 09, 2008 at 05:28:57PM +0000, Miloslav Trma?? wrote:
the attached patch adds support for copy-on-write images: creating
them,
deleting them, identifying them.
Nice - thanks for working on this feature ! I'll put a few comments
inline....
Index: docs/formatstorage.html.in
===================================================================
RCS file: /data/cvs/libvirt/docs/formatstorage.html.in,v
retrieving revision 1.5
diff -u -p -r1.5 formatstorage.html.in
--- docs/formatstorage.html.in 4 Dec 2008 14:51:57 -0000 1.5
+++ docs/formatstorage.html.in 9 Dec 2008 17:25:18 -0000
@@ -269,6 +269,13 @@
contains the MAC (eg SELinux) label string.
<span class="since">Since 0.4.1</span>
</dd>
+ <dt><code>cow-base</code></dt>
+ <dd>Provides the key of a base volume. This element is optional, and
+ supported only in some volume types and formats. If it is present,
+ this volume is a copy-on-write volume that stores only changes to the
+ base volume. Changes to the base volume may make this volume
+ inconsistent.
+ </dd>
I'm not a fan of having '-' or '_' in XML element names, so how
about naming it
<backingstore>/some/path</backingstore>
We will also want to make use of this feature to allow LVM snapshots.
ie, creating an LVM volume with a <backingstore> element would take
an LVM snapshot of this backingstore. The 'allocation' would be the
amount of LVM storage allocated, and the 'capacity' would just match
the original volume capacity.
Index: src/storage_backend_fs.c
===================================================================
RCS file: /data/cvs/libvirt/src/storage_backend_fs.c,v
retrieving revision 1.22
diff -u -p -r1.22 storage_backend_fs.c
--- src/storage_backend_fs.c 17 Nov 2008 11:19:33 -0000 1.22
+++ src/storage_backend_fs.c 9 Dec 2008 17:25:18 -0000
+static int
+cowGetCowBase(virConnectPtr conn,
+ char **res,
+ const unsigned char *buf,
+ size_t buf_size)
+{
+ size_t len;
+
+ *res = NULL;
+ if (buf_size < 4+4+1024)
+ return GET_COW_BASE_INVALID;
+ if (buf[4+4] == '\0')
+ return GET_COW_BASE_OK;
+
+ len = 1024;
+ if (VIR_ALLOC_N(*res, len + 1) < 0) {
+ virStorageReportError(conn, VIR_ERR_NO_MEMORY,
+ _("copy-on-write base path"));
+ return GET_COW_BASE_ERROR;
+ }
+ memcpy(*res, buf + 4+4, len);
+ (*res)[len] = '\0';
+ if (VIR_REALLOC_N(*res, strlen(*res) + 1) < 0) {
+ /* Ignore failure */
+ }
+ return GET_COW_BASE_OK;
+}
In general, this entire patch has alot of whitespace damage - I
think your editor has inserted TABs instead of spaces - could you
resend with whitespace fixed as per the 'HACKING' file.
+
+static int
+qcowXGetCowBase(virConnectPtr conn,
+ char **res,
+ const unsigned char *buf,
+ size_t buf_size)
+{
+ unsigned long long offset;
+ unsigned long size;
+
+ *res = NULL;
+ if (buf_size < 4+4+8+4)
+ return GET_COW_BASE_INVALID;
+ offset = (((unsigned long long)buf[4+4] << 56)
+ | ((unsigned long long)buf[4+4+1] << 48)
+ | ((unsigned long long)buf[4+4+2] << 40)
+ | ((unsigned long long)buf[4+4+3] << 32)
+ | ((unsigned long long)buf[4+4+4] << 24)
+ | ((unsigned long long)buf[4+4+5] << 16)
+ | ((unsigned long long)buf[4+4+6] << 8)
+ | buf[4+4+7]);
+ if (offset > buf_size)
+ return GET_COW_BASE_INVALID;
+ size = ((buf[4+4+8] << 24)
+ | (buf[4+4+8+1] << 16)
+ | (buf[4+4+8+2] << 8)
+ | buf[4+4+8+3]);
+ if (size == 0)
+ return GET_COW_BASE_OK;
+ if (offset + size > buf_size || offset + size < offset)
+ return GET_COW_BASE_INVALID;
+ if (size + 1 == 0)
+ return GET_COW_BASE_INVALID;
It'd be useful to have comments against each of these 'if'
checks to indicate what named header field is being checked
in each place.
+static int
+vmdk4GetCowBase(virConnectPtr conn,
+ char **res,
+ const unsigned char *buf,
+ size_t buf_size)
+{
+ static const char prefix[] = "parentFileNameHint=\"";
+
+ char desc[20*512 + 1], *start, *end;
+ size_t len;
+
+ *res = NULL;
+ if (buf_size <= 0x200)
+ return GET_COW_BASE_INVALID;
+ /* FIXME? the default buf_size is not large enough to contain the complete
+ descriptor. */
+ len = buf_size - 0x200;
Do you know how large it needs to be ? There's no particular reason
for the current size, except that it was big enough for our needs
at the time. We can increase it to suit....
+
+/**
+ * Return an absolute path corresponding to PATH, which is absolute or relative
+ * to the directory containing BASE_FILE, or NULL on error
+ */
+static char *absolutePathFromBaseFile(const char *base_file, const char *path)
+{
+ size_t base_size, path_size;
+ char *res, *p;
+
+ if (*path == '/')
+ return strdup(path);
+
+ base_size = strlen(base_file) + 1;
+ path_size = strlen(path) + 1;
+ if (VIR_ALLOC_N(res, base_size - 1 + path_size) < 0)
+ return NULL;
+ memcpy(res, base_file, base_size);
+ p = strrchr(res, '/');
+ if (p != NULL)
+ p++;
+ else
+ p = res;
+ memcpy(p, path, path_size);
+ if (VIR_REALLOC_N(res, (p + path_size) - res) < 0) {
+ /* Ignore failure */
+ }
+ return res;
+}
Seeems like we could just call libc 'realpath' function here ?
@@ -832,7 +1029,9 @@ virStorageBackendFileSystemVolCreate(vir
#if HAVE_QEMU_IMG
const char *type;
char size[100];
- const char *imgargv[7];
+ const char *imgargv[9];
+ size_t i;
+ virStorageVolDefPtr cowBase;
if ((type = virStorageVolFormatFileSystemTypeToString(vol->target.format)) ==
NULL) {
virStorageReportError(conn, VIR_ERR_INTERNAL_ERROR,
@@ -840,17 +1039,55 @@ virStorageBackendFileSystemVolCreate(vir
vol->target.format);
return -1;
}
+ if (vol->target.cowBase == NULL)
+ cowBase = NULL;
+ else {
+ const struct FileTypeInfo *info;
+
+ info = virStorageBackendFileSystemVolFormatToInfo(conn,
+ vol->target.format);
+ if (info == NULL) {
+ virStorageReportError(conn, VIR_ERR_INTERNAL_ERROR,
+ _("unknown storage vol type %d"),
+ vol->target.format);
+ return -1;
+ }
+ if (info->getCowBase == NULL) {
+ virStorageReportError(conn, VIR_ERR_INTERNAL_ERROR,
+ _("copy-on-write not supported with "
+ "format %s"), type);
+ return -1;
+ }
+ cowBase = virStorageVolDefFindByKey(pool, vol->target.cowBase);
+ if (cowBase == NULL) {
+ virStorageReportError(conn, VIR_ERR_NO_STORAGE_VOL,
+ _("unknown copy-on-write base volume "
+ "key %s"), vol->target.cowBase);
+ return -1;
+ }
+ if (cowBase->target.format != vol->target.format) {
+ virStorageReportError(conn, VIR_ERR_INTERNAL_ERROR,
+ _("copy-on-write with mixed file formats "
+ "not supported"));
+ return -1;
IIRC there's no restricton on backing file type for QCow files. You
can even make a qcow file that's backed by a physical device.
So we should remove the virStorageVolDefFindByKey() call and format
check, and merely do 'access(vol->target.cowBase, R_OK)' on it.
*/
static int
virStorageBackendFileSystemVolDelete(virConnectPtr conn,
- virStoragePoolObjPtr pool ATTRIBUTE_UNUSED,
+ virStoragePoolObjPtr pool,
virStorageVolDefPtr vol,
unsigned int flags ATTRIBUTE_UNUSED)
{
+ size_t i;
+
+ for (i = 0; i < pool->volumes.count; i++) {
+ const char *cowBase;
+
+ cowBase = pool->volumes.objs[i]->target.cowBase;
+ if (cowBase != NULL && STREQ(cowBase, vol->key)) {
+ virStorageReportError(conn, VIR_ERR_INTERNAL_ERROR,
+ _("volume %s is used by volume %s"),
+ vol->key, pool->volumes.objs[i]->key);
+ return -1;
+ }
+ }
As a general rule libvirt does not implement policy like this, leave it
upto the caller whether they want to imeplment such policies. In addition
we can't rely on the the cow base file being part of the same storage pool
as the volume being deleted - it might not be part of any pool in fact.
IMHO, this block of code can just be omitted
Regards,
Daniel
--
|: Red Hat, Engineering, London -o- http://people.redhat.com/berrange/ :|
|: http://libvirt.org -o- http://virt-manager.org -o- http://ovirt.org :|
|: http://autobuild.org -o- http://search.cpan.org/~danberr/ :|
|: GnuPG: 7D3B9505 -o- F3C9 553F A1DA 4AC2 5648 23C1 B3DF F742 7D3B 9505 :|
8:54 a.m.
Daniel,
thanks for the review.
The attached patch should fix everything you pointed out, except for
this comment:
Daniel P. Berrange píše v St 10. 12. 2008 v 17:58 +0000:
> +
> +/**
> + * Return an absolute path corresponding to PATH, which is absolute or relative
> + * to the directory containing BASE_FILE, or NULL on error
> + */
> +static char *absolutePathFromBaseFile(const char *base_file, const char *path)
> +{
> + size_t base_size, path_size;
> + char *res, *p;
> +
> + if (*path == '/')
> + return strdup(path);
> +
> + base_size = strlen(base_file) + 1;
> + path_size = strlen(path) + 1;
> + if (VIR_ALLOC_N(res, base_size - 1 + path_size) < 0)
> + return NULL;
> + memcpy(res, base_file, base_size);
> + p = strrchr(res, '/');
> + if (p != NULL)
> + p++;
> + else
> + p = res;
> + memcpy(p, path, path_size);
> + if (VIR_REALLOC_N(res, (p + path_size) - res) < 0) {
> + /* Ignore failure */
> + }
> + return res;
> +}
Seeems like we could just call libc 'realpath' function here ?
realpath()
is relative to the current directory, not to "path".
Mirek
12:09 p.m.
On Mon, Jan 12, 2009 at 09:19:02AM +0000, Miloslav Trmač wrote:
Hello,
here's a version of the patch updated to apply against current CVS.
Mirek
Patch looks reasonable, and has been reviewed. Shall we apply?
Rich.
--
Richard Jones, Emerging Technologies, Red Hat http://et.redhat.com/~rjones
virt-top is 'top' for virtual machines. Tiny program with many
powerful monitoring features, net stats, disk stats, logging, etc.
http://et.redhat.com/~rjones/virt-top
9:53 a.m.
On Mon, Jan 12, 2009 at 09:19:02AM +0000, Miloslav Trma?? wrote:
Hello,
here's a version of the patch updated to apply against current CVS.
I've realized there's a slight problem with our idea to add a simple
<backingStore>/some/path</backingStore>
element to the volume '<target>' section. Specifically when creating
a new image it is desirable to explicitly specify the format of the
backing store. Without this, QEMU will probe backing store format
and this opens a security problem - if the backing store was a raw
file, the guest could have written data into it, such that QEMU will
mis-probe it as QCow, and thus potentially be able to compromise
abuse it to read any file on the host.
I thus think it is better to have the backing store info outside
the target block, as a top level item, allowing the exact same
child elements are 'target' does. As an example, a QCow2 image,
with a raw backing store would appear as:
<volume>
<name>OtherDemo.img</name>
<key>/var/lib/libvirt/images/OtherDemo.img</key>
<source>
</source>
<capacity>5242880000</capacity>
<allocation>294912</allocation>
<target>
<path>/var/lib/libvirt/images/OtherDemo.img</path>
<format type='qcow2'/>
<permissions>
<mode>0644</mode>
<owner>0</owner>
<group>0</group>
<label>unconfined_u:object_r:virt_image_t:s0</label>
</permissions>
</target>
<backingStore>
<path>/var/lib/libvirt/images/XennerDemo.img</path>
<format type='raw'/>
<permissions>
<mode>0444</mode>
<owner>0</owner>
<group>0</group>
<label>system_u:object_r:virt_image_t:s0</label>
</permissions>
</backingStore>
</volume>
I've updated your patch to work in this way and am in the process
of making it support LVM based COW/snapshotting too. I'll post
a complete patch with all this in soon..
Daniel
--
|: Red Hat, Engineering, London -o- http://people.redhat.com/berrange/ :|
|: http://libvirt.org -o- http://virt-manager.org -o- http://ovirt.org :|
|: http://autobuild.org -o- http://search.cpan.org/~danberr/ :|
|: GnuPG: 7D3B9505 -o- F3C9 553F A1DA 4AC2 5648 23C1 B3DF F742 7D3B 9505 :|
5791
days inactive
5827
days old
5 comments
3 participants
participants (3)
-
Daniel P. Berrange -
Miloslav Trmač -
Richard W.M. Jones