| Message ID | 20170329011637.89377-2-bjsdjshi@linux.vnet.ibm.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
On 29.03.2017 03:16, Dong Jia Shi wrote: > raw_open() expects the caller always passing in the right actual > @options parameter. But when trying to applying snapshot on a RBD > image, bdrv_snapshot_goto() calls raw_open() (by calling the > bdrv_open callback on the BlockDriver) with a NULL @options, and > that will result in a Segmentation fault. > > For the other non-raw format drivers, it also makes sense to passing > in the actual options, althought they don't trigger the problem so > far. > > Let's prepare a @options by adding the "file" key-value pair to a > copy of the actual options that were given for the node (i.e. > bs->options), and pass it to the callback. > > BlockDriver.bdrv_open() expects bs->file to be NULL and just > overwrites it with the result from bdrv_open_child(). If that > bdrv_open_child() fails, the field becomes NULL. While we are at > it, we also correct the cleanning up action for a call failure of > BlockDriver.bdrv_open() by replacing bdrv_unref() with > bdrv_unref_child(). > > Suggested-by: Max Reitz <mreitz@redhat.com> > Signed-off-by: Dong Jia Shi <bjsdjshi@linux.vnet.ibm.com> > --- > block/snapshot.c | 14 ++++++++++++-- > 1 file changed, 12 insertions(+), 2 deletions(-) > > diff --git a/block/snapshot.c b/block/snapshot.c > index bf5c2ca..281626c 100644 > --- a/block/snapshot.c > +++ b/block/snapshot.c > @@ -27,6 +27,7 @@ > #include "block/block_int.h" > #include "qapi/error.h" > #include "qapi/qmp/qerror.h" > +#include "qapi/qmp/qstring.h" > > QemuOptsList internal_snapshot_opts = { > .name = "snapshot", > @@ -189,11 +190,20 @@ int bdrv_snapshot_goto(BlockDriverState *bs, > } > > if (bs->file) { > + QDict *options = qdict_clone_shallow(bs->options); > + QDict *file_options; > + > + qdict_extract_subqdict(options, &file_options, "file."); > + QDECREF(file_options); > + qdict_put(options, "file", > + qstring_from_str(bdrv_get_node_name(bs->file->bs))); > + > drv->bdrv_close(bs); > ret = bdrv_snapshot_goto(bs->file->bs, snapshot_id); > - open_ret = drv->bdrv_open(bs, NULL, bs->open_flags, NULL); > + open_ret = drv->bdrv_open(bs, options, bs->open_flags, NULL); > + QDECREF(options); > if (open_ret < 0) { > - bdrv_unref(bs->file->bs); > + bdrv_unref_child(bs, bs->file); > bs->drv = NULL; > return open_ret; > } I just noticed another issue (sorry I did not before...): In drv->bdrv_open(), the block driver will generally overwrite bs->file without looking at it because it assumes that it's NULL. That means we should probably actually make sure it's NULL because otherwise we will the child BDS will have a reference count that is 1 too high. (bdrv_open_inherit() (ultimately called from bdrv_open_child()) invokes bdrv_ref() for a child BDS specified by node-name reference.) That means we should unconditionally invoke bdrv_unref_child(bs, bs->file) before calling drv->bdrv_open(). But we have to wrap everything in bdrv_ref()/bdrv_unref() so the BDS isn't deleted in the meantime. So I think it should look something like this: BlockDriverState *file; QDict *options = ...; QDict *file_options; file = bs->file->bs; /* Prevent it from getting deleted when detached from bs */ bdrv_ref(file); qdict_extract_subqdict(...); QDECREF(file_options); qdict_put(..., qstring_from_str(bdrv_get_node_name(file))); drv->bdrv_close(bs); bdrv_unref_child(bs, bs->file); bs->file = NULL; ret = bdrv_snapshot_goto(file, snapshot_id); open_ret = drv->bdrv_open(...); if (open_ret < 0) { bdrv_unref(file); bs->drv = NULL; return open_ret; } assert(bs->file->bs == file); bdrv_unref(file); Max
* Max Reitz <mreitz@redhat.com> [2017-03-29 23:07:22 +0200]: > On 29.03.2017 03:16, Dong Jia Shi wrote: > > raw_open() expects the caller always passing in the right actual > > @options parameter. But when trying to applying snapshot on a RBD > > image, bdrv_snapshot_goto() calls raw_open() (by calling the > > bdrv_open callback on the BlockDriver) with a NULL @options, and > > that will result in a Segmentation fault. > > > > For the other non-raw format drivers, it also makes sense to passing > > in the actual options, althought they don't trigger the problem so > > far. > > > > Let's prepare a @options by adding the "file" key-value pair to a > > copy of the actual options that were given for the node (i.e. > > bs->options), and pass it to the callback. > > > > BlockDriver.bdrv_open() expects bs->file to be NULL and just > > overwrites it with the result from bdrv_open_child(). If that > > bdrv_open_child() fails, the field becomes NULL. While we are at > > it, we also correct the cleanning up action for a call failure of > > BlockDriver.bdrv_open() by replacing bdrv_unref() with > > bdrv_unref_child(). > > > > Suggested-by: Max Reitz <mreitz@redhat.com> > > Signed-off-by: Dong Jia Shi <bjsdjshi@linux.vnet.ibm.com> > > --- > > block/snapshot.c | 14 ++++++++++++-- > > 1 file changed, 12 insertions(+), 2 deletions(-) > > > > diff --git a/block/snapshot.c b/block/snapshot.c > > index bf5c2ca..281626c 100644 > > --- a/block/snapshot.c > > +++ b/block/snapshot.c > > @@ -27,6 +27,7 @@ > > #include "block/block_int.h" > > #include "qapi/error.h" > > #include "qapi/qmp/qerror.h" > > +#include "qapi/qmp/qstring.h" > > > > QemuOptsList internal_snapshot_opts = { > > .name = "snapshot", > > @@ -189,11 +190,20 @@ int bdrv_snapshot_goto(BlockDriverState *bs, > > } > > > > if (bs->file) { > > + QDict *options = qdict_clone_shallow(bs->options); > > + QDict *file_options; > > + > > + qdict_extract_subqdict(options, &file_options, "file."); > > + QDECREF(file_options); > > + qdict_put(options, "file", > > + qstring_from_str(bdrv_get_node_name(bs->file->bs))); > > + > > drv->bdrv_close(bs); > > ret = bdrv_snapshot_goto(bs->file->bs, snapshot_id); > > - open_ret = drv->bdrv_open(bs, NULL, bs->open_flags, NULL); > > + open_ret = drv->bdrv_open(bs, options, bs->open_flags, NULL); > > + QDECREF(options); > > if (open_ret < 0) { > > - bdrv_unref(bs->file->bs); > > + bdrv_unref_child(bs, bs->file); > > bs->drv = NULL; > > return open_ret; > > } > > I just noticed another issue (sorry I did not before...): Hi Max, No need for sorry. > > In drv->bdrv_open(), the block driver will generally overwrite bs->file > without looking at it because it assumes that it's NULL. That means we > should probably actually make sure it's NULL because otherwise we will > the child BDS will have a reference count that is 1 too high. > > (bdrv_open_inherit() (ultimately called from bdrv_open_child()) invokes > bdrv_ref() for a child BDS specified by node-name reference.) > > That means we should unconditionally invoke > bdrv_unref_child(bs, bs->file) before calling drv->bdrv_open(). But we > have to wrap everything in bdrv_ref()/bdrv_unref() so the BDS isn't > deleted in the meantime. Understood. > > So I think it should look something like this: > > BlockDriverState *file; > QDict *options = ...; > QDict *file_options; > > file = bs->file->bs; > /* Prevent it from getting deleted when detached from bs */ > bdrv_ref(file); > > qdict_extract_subqdict(...); > QDECREF(file_options); > qdict_put(..., qstring_from_str(bdrv_get_node_name(file))); > > drv->bdrv_close(bs); > bdrv_unref_child(bs, bs->file); > bs->file = NULL; > > ret = bdrv_snapshot_goto(file, snapshot_id); > open_ret = drv->bdrv_open(...); Here, I think we'd still need a: QDECREF(options); > if (open_ret < 0) { > bdrv_unref(file); > bs->drv = NULL; > return open_ret; > } > > assert(bs->file->bs == file); > bdrv_unref(file); It looks convoluted, but I don't see a better solution. So I will prepare a new patch based on your code. Thanks for the help! > > > Max >
diff --git a/block/snapshot.c b/block/snapshot.c index bf5c2ca..281626c 100644 --- a/block/snapshot.c +++ b/block/snapshot.c @@ -27,6 +27,7 @@ #include "block/block_int.h" #include "qapi/error.h" #include "qapi/qmp/qerror.h" +#include "qapi/qmp/qstring.h" QemuOptsList internal_snapshot_opts = { .name = "snapshot", @@ -189,11 +190,20 @@ int bdrv_snapshot_goto(BlockDriverState *bs, } if (bs->file) { + QDict *options = qdict_clone_shallow(bs->options); + QDict *file_options; + + qdict_extract_subqdict(options, &file_options, "file."); + QDECREF(file_options); + qdict_put(options, "file", + qstring_from_str(bdrv_get_node_name(bs->file->bs))); + drv->bdrv_close(bs); ret = bdrv_snapshot_goto(bs->file->bs, snapshot_id); - open_ret = drv->bdrv_open(bs, NULL, bs->open_flags, NULL); + open_ret = drv->bdrv_open(bs, options, bs->open_flags, NULL); + QDECREF(options); if (open_ret < 0) { - bdrv_unref(bs->file->bs); + bdrv_unref_child(bs, bs->file); bs->drv = NULL; return open_ret; }
raw_open() expects the caller always passing in the right actual @options parameter. But when trying to applying snapshot on a RBD image, bdrv_snapshot_goto() calls raw_open() (by calling the bdrv_open callback on the BlockDriver) with a NULL @options, and that will result in a Segmentation fault. For the other non-raw format drivers, it also makes sense to passing in the actual options, althought they don't trigger the problem so far. Let's prepare a @options by adding the "file" key-value pair to a copy of the actual options that were given for the node (i.e. bs->options), and pass it to the callback. BlockDriver.bdrv_open() expects bs->file to be NULL and just overwrites it with the result from bdrv_open_child(). If that bdrv_open_child() fails, the field becomes NULL. While we are at it, we also correct the cleanning up action for a call failure of BlockDriver.bdrv_open() by replacing bdrv_unref() with bdrv_unref_child(). Suggested-by: Max Reitz <mreitz@redhat.com> Signed-off-by: Dong Jia Shi <bjsdjshi@linux.vnet.ibm.com> --- block/snapshot.c | 14 ++++++++++++-- 1 file changed, 12 insertions(+), 2 deletions(-)