| Message ID | 20220803144446.20723-1-hreitz@redhat.com |
|---|---|
| Headers | show |
| Series | block: Keep auto_backing_file post-migration | expand |
Am 03.08.2022 um 16:44 hat Hanna Reitz geschrieben: > Hi, > > https://gitlab.com/qemu-project/qemu/-/issues/1117 reports the following > issue: > > Say you have a VM with a backing chain of images where the image > metadata contains json:{} backing file strings, which however will be > resolved to simple plain filenames when opened[1]. > > So when these images are opened, bs->auto_backing_file is first read > directly from the image header, and will thus contain a json:{} > filename. The backing image is opened based off of this filename, and > bdrv_refresh_filename() simplfies the filename as shown[1]. We then > update bs->auto_backing_file from bs->backing->bs->filename, so both are > equal. > > It is quite important that both are equal, because > bdrv_backing_overridden() checks whether the backing file has been > changed from the default by comparing bs->auto_backing_file to > bs->backing->bs->filename. > > Because we did set bs->auto_backing_file from bs->backing->bs->filename, > both are equal, the backing file is not considered overridden, and > bdrv_refresh_filename(bs) will not consider it necessary to generate a > json:{} filename for the overlay. > > Then the VM is migrated. > > The destination side invokes bdrv_invalidate_cache(), which by qcow2 and > qed is implemented by closing the image and opening it. This re-reads > the backing file string from disk, resetting bs->auto_backing_file. > Now, it will contains the json:{} filename again and thus differ from > bs->backing->bs->filename. > > Consequentially, a subsequent bdrv_refresh_filename(bs) will find that > the overlay’s backing file has been overridden and generate a json:{} > filename, which isn’t great. > > This series fixes that by having qcow2’s and qed’s image-open operations > not overwrite bs->auto_backing_file unless something has changed since > the last time we read the backing filename from the metadata. > > > Now, generating a json:{} filename can be a nuisance but shouldn’t be a > real problem. The actual problem reported in 1117 comes later, namely > when creating a snapshot overlay post-migration. This overlay image > will have a json:{} backing filename in its image metadata, which > contains a 'backing' key[2]. > > 'qemu-img info' uses the BDRV_O_NO_BACKING flag to open images, which > conflicts with those backing options: With that flag, nobody processes > those options, and that’s an error. Therefore, you can’t run 'qemu-img > info --backing-chain' on that overlay image. > > That part of the issue is not fixed in this series, however. I’ll send > a separate RFC series for it, because I’m honstly not quite certain how > it should be fixed. > > > [1] Example: > json:{"driver": "qcow2", > "file": {"driver": "file", "filename": "img.qcow2"}} > Will generally be “resolved” by bdrv_refresh_filename() to > "img.qcow2" > > [2] That it contains a 'backing' key is only natural, because the reason > why bdrv_refresh_filename() decided to generate a json:{} filename > for the image is because it considered the backing file overridden. > Hence it must put the actual backing file options into a 'backing' > object in the json:{} filename. > > > Hanna Reitz (3): > block/qcow2: Keep auto_backing_file if possible > block/qed: Keep auto_backing_file if possible > iotests/backing-file-invalidation: Add new test Thanks, applied to the block branch. Kevin
Hi, https://gitlab.com/qemu-project/qemu/-/issues/1117 reports the following issue: Say you have a VM with a backing chain of images where the image metadata contains json:{} backing file strings, which however will be resolved to simple plain filenames when opened[1]. So when these images are opened, bs->auto_backing_file is first read directly from the image header, and will thus contain a json:{} filename. The backing image is opened based off of this filename, and bdrv_refresh_filename() simplfies the filename as shown[1]. We then update bs->auto_backing_file from bs->backing->bs->filename, so both are equal. It is quite important that both are equal, because bdrv_backing_overridden() checks whether the backing file has been changed from the default by comparing bs->auto_backing_file to bs->backing->bs->filename. Because we did set bs->auto_backing_file from bs->backing->bs->filename, both are equal, the backing file is not considered overridden, and bdrv_refresh_filename(bs) will not consider it necessary to generate a json:{} filename for the overlay. Then the VM is migrated. The destination side invokes bdrv_invalidate_cache(), which by qcow2 and qed is implemented by closing the image and opening it. This re-reads the backing file string from disk, resetting bs->auto_backing_file. Now, it will contains the json:{} filename again and thus differ from bs->backing->bs->filename. Consequentially, a subsequent bdrv_refresh_filename(bs) will find that the overlay’s backing file has been overridden and generate a json:{} filename, which isn’t great. This series fixes that by having qcow2’s and qed’s image-open operations not overwrite bs->auto_backing_file unless something has changed since the last time we read the backing filename from the metadata. Now, generating a json:{} filename can be a nuisance but shouldn’t be a real problem. The actual problem reported in 1117 comes later, namely when creating a snapshot overlay post-migration. This overlay image will have a json:{} backing filename in its image metadata, which contains a 'backing' key[2]. 'qemu-img info' uses the BDRV_O_NO_BACKING flag to open images, which conflicts with those backing options: With that flag, nobody processes those options, and that’s an error. Therefore, you can’t run 'qemu-img info --backing-chain' on that overlay image. That part of the issue is not fixed in this series, however. I’ll send a separate RFC series for it, because I’m honstly not quite certain how it should be fixed. [1] Example: json:{"driver": "qcow2", "file": {"driver": "file", "filename": "img.qcow2"}} Will generally be “resolved” by bdrv_refresh_filename() to "img.qcow2" [2] That it contains a 'backing' key is only natural, because the reason why bdrv_refresh_filename() decided to generate a json:{} filename for the image is because it considered the backing file overridden. Hence it must put the actual backing file options into a 'backing' object in the json:{} filename. Hanna Reitz (3): block/qcow2: Keep auto_backing_file if possible block/qed: Keep auto_backing_file if possible iotests/backing-file-invalidation: Add new test block/qcow2.c | 21 ++- block/qed.c | 15 +- .../tests/backing-file-invalidation | 152 ++++++++++++++++++ .../tests/backing-file-invalidation.out | 5 + 4 files changed, 184 insertions(+), 9 deletions(-) create mode 100755 tests/qemu-iotests/tests/backing-file-invalidation create mode 100644 tests/qemu-iotests/tests/backing-file-invalidation.out