Message ID | 20200922104932.46384-1-mreitz@redhat.com |
---|---|
Headers | show |
Series | block/export: Allow exporting BDSs via FUSE | expand |
On Tue, Sep 22, 2020 at 12:49:12PM +0200, Max Reitz wrote: > Based-on: <20200907182011.521007-1-kwolf@redhat.com> > (“block/export: Add infrastructure and QAPI for block exports”) > > (Though its patch 16 needs a s/= \&blk_exp_nbd/= drv/ on top.) > > v1: https://lists.nongnu.org/archive/html/qemu-block/2019-12/msg00451.html > > Branch: https://github.com/XanClic/qemu.git fuse-exports-v2 > Branch: https://git.xanclic.moe/XanClic/qemu.git fuse-exports-v2 > > > Hi, > > Ever since I found out that you can mount FUSE filesystems on regular > files (not just directories), I had the idea of adding FUSE block > exports to qemu where you can export block nodes as raw images. The > best thing is that you’d be able to mount an image on itself, so > whatever format it may be in, qemu lets it appear as a raw image (and > you can then use regular tools like dd on it). > > The performance is quite bad so far, but we can always try to improve it > if the need arises. For now I consider it mostly a cute feature to get > easy access to the raw contents of image files in any format (without > requiring root rights). Aside from the iotests, so you forsee any particular use cases where this feature is desirable / important ? Looking at it from a security POV, I'm not thrilled about the idea of granting QEMU permission to use the mount syscall for seccomp or SELinux. IOW, I expect this feature won't be something we want to expose in QEMU guests managed by libvirt, which would limit how widely it can be used. QEMU can export NBD. Would it make sense to do this as an NBD client ? There's already https://libguestfs.org/nbdfuse.1.html but IIUC that exposes it as a file within a dir. Presumably it is not too hard to make it support exposing it directly as a file too. I wonder how performance compares between your native FUSE impl in QEMU vs NBD FUSE ? Regards, Daniel -- |: https://berrange.com -o- https://www.flickr.com/photos/dberrange :| |: https://libvirt.org -o- https://fstop138.berrange.com :| |: https://entangle-photo.org -o- https://www.instagram.com/dberrange :|
On 22.09.20 17:58, Daniel P. Berrangé wrote: > On Tue, Sep 22, 2020 at 12:49:12PM +0200, Max Reitz wrote: >> Based-on: <20200907182011.521007-1-kwolf@redhat.com> >> (“block/export: Add infrastructure and QAPI for block exports”) >> >> (Though its patch 16 needs a s/= \&blk_exp_nbd/= drv/ on top.) >> >> v1: https://lists.nongnu.org/archive/html/qemu-block/2019-12/msg00451.html >> >> Branch: https://github.com/XanClic/qemu.git fuse-exports-v2 >> Branch: https://git.xanclic.moe/XanClic/qemu.git fuse-exports-v2 >> >> >> Hi, >> >> Ever since I found out that you can mount FUSE filesystems on regular >> files (not just directories), I had the idea of adding FUSE block >> exports to qemu where you can export block nodes as raw images. The >> best thing is that you’d be able to mount an image on itself, so >> whatever format it may be in, qemu lets it appear as a raw image (and >> you can then use regular tools like dd on it). >> >> The performance is quite bad so far, but we can always try to improve it >> if the need arises. For now I consider it mostly a cute feature to get >> easy access to the raw contents of image files in any format (without >> requiring root rights). > > Aside from the iotests, so you forsee any particular use cases > where this feature is desirable / important ? No. I implemented this feature for fun last year (when I realized that FUSE allows regular files to be mount points), and I got positive reactions. I assumed others would find it as nice as me to be able to quickly access an image file without requiring root rights (and then device file accesses), or setting up an NBD chain. (Though it should be noted that when I first came up with the feature, nbdfuse did not exist yet.) (It should also be noted that my original idea was to have a new executable qemu-blkfuse that would basically allow you to invoke “qemu-blkfuse $img”, and then $img would appear as a raw image. To me, that appeared very useful because it was so simple. I admit that the current proposal, which relies on the storage-daemon, has none of that simplicity. But if that’s the problem that prevents this from being considered useful, I’m sure we (I) can figure something out. Perhaps a simple script, bundled with qemu, that can generate -blockdev invocations based on the result of file(1).) > Looking at it from a security POV, I'm not thrilled about the > idea of granting QEMU permission to use the mount syscall for > seccomp or SELinux. IOW, I expect this feature won't be something > we want to expose in QEMU guests managed by libvirt, which would > limit how widely it can be used. I don’t expect this to be used through QEMU very much, but through the storage daemon. I assume that for the storage daemon, the permissions can effectively be fine-tuned for each export, because you can “just” launch another instance. > QEMU can export NBD. Would it make sense to do this as an NBD > client ? There's already https://libguestfs.org/nbdfuse.1.html > but IIUC that exposes it as a file within a dir. Presumably > it is not too hard to make it support exposing it directly as > a file too. I don’t like that idea very much, because my main gripe with the current state of my proposal is that it’s more cumbersome than “qemu-blkfuse $img”. Adding more indirections won’t make it simpler. > I wonder how performance compares between your native FUSE > impl in QEMU vs NBD FUSE ? Last year, I tried various ways of improving performance and nothing really amounted to much. So in the end I settled for a simple and naive implementation, for it to be improved in case anyone cares for it. Max
On Tue, Sep 22, 2020 at 04:58:38PM +0100, Daniel P. Berrangé wrote: > On Tue, Sep 22, 2020 at 12:49:12PM +0200, Max Reitz wrote: > > Based-on: <20200907182011.521007-1-kwolf@redhat.com> > > (“block/export: Add infrastructure and QAPI for block exports”) > > > > (Though its patch 16 needs a s/= \&blk_exp_nbd/= drv/ on top.) > > > > v1: https://lists.nongnu.org/archive/html/qemu-block/2019-12/msg00451.html > > > > Branch: https://github.com/XanClic/qemu.git fuse-exports-v2 > > Branch: https://git.xanclic.moe/XanClic/qemu.git fuse-exports-v2 > > > > > > Hi, > > > > Ever since I found out that you can mount FUSE filesystems on regular > > files (not just directories), I had the idea of adding FUSE block > > exports to qemu where you can export block nodes as raw images. The > > best thing is that you’d be able to mount an image on itself, so > > whatever format it may be in, qemu lets it appear as a raw image (and > > you can then use regular tools like dd on it). > > > > The performance is quite bad so far, but we can always try to improve it > > if the need arises. For now I consider it mostly a cute feature to get > > easy access to the raw contents of image files in any format (without > > requiring root rights). > > Aside from the iotests, so you forsee any particular use cases > where this feature is desirable / important ? Alice Frosi is working on a qemu-storage-daemon-based project where the FUSE export type is useful. In this case qemu-storage-daemon is used stand-alone without a guest or libvirt directly involved. The goal is just to export disk images and how they are consumed is the user's responsibility (processes, containers, guests). > Looking at it from a security POV, I'm not thrilled about the > idea of granting QEMU permission to use the mount syscall for > seccomp or SELinux. IOW, I expect this feature won't be something > we want to expose in QEMU guests managed by libvirt, which would > limit how widely it can be used. I have CCed Miklos Szeredi, the Linux FUSE maintainer, to check what the options are for unprivileged mounting of a FUSE file system: 1. libfuse invokes open("/dev/fuse") + mount() 2. libfuse spawns the fusermount3 suid root helper 3. Any other options? D-Bus? etc > QEMU can export NBD. Would it make sense to do this as an NBD > client ? There's already https://libguestfs.org/nbdfuse.1.html > but IIUC that exposes it as a file within a dir. Presumably > it is not too hard to make it support exposing it directly as > a file too. > > I wonder how performance compares between your native FUSE > impl in QEMU vs NBD FUSE ? NBD exports are useful for networks but shouldn't preclude other export types from being merged just because they can be implemented on top of NBD. Native export types are simpler to manage and have less performance overhead than stacking additional tools on top of NBD. Stefan
Am 22.09.2020 um 12:49 hat Max Reitz geschrieben: > Based-on: <20200907182011.521007-1-kwolf@redhat.com> > (“block/export: Add infrastructure and QAPI for block exports”) > > (Though its patch 16 needs a s/= \&blk_exp_nbd/= drv/ on top.) > > v1: https://lists.nongnu.org/archive/html/qemu-block/2019-12/msg00451.html > > Branch: https://github.com/XanClic/qemu.git fuse-exports-v2 > Branch: https://git.xanclic.moe/XanClic/qemu.git fuse-exports-v2 > > > Hi, > > Ever since I found out that you can mount FUSE filesystems on regular > files (not just directories), I had the idea of adding FUSE block > exports to qemu where you can export block nodes as raw images. The > best thing is that you’d be able to mount an image on itself, so > whatever format it may be in, qemu lets it appear as a raw image (and > you can then use regular tools like dd on it). > > The performance is quite bad so far, but we can always try to improve it > if the need arises. For now I consider it mostly a cute feature to get > easy access to the raw contents of image files in any format (without > requiring root rights). > > In this version (as opposed to v1 linked above), I integrated the FUSE > export code into Kevin’s proposed common infrastructure for block > exports. Patches 5-16, 19 and 20: Reviewed-by: Kevin Wolf <kwolf@redhat.com>
On 15.10.20 14:01, Kevin Wolf wrote: > Am 22.09.2020 um 12:49 hat Max Reitz geschrieben: >> Based-on: <20200907182011.521007-1-kwolf@redhat.com> >> (“block/export: Add infrastructure and QAPI for block exports”) >> >> (Though its patch 16 needs a s/= \&blk_exp_nbd/= drv/ on top.) >> >> v1: https://lists.nongnu.org/archive/html/qemu-block/2019-12/msg00451.html >> >> Branch: https://github.com/XanClic/qemu.git fuse-exports-v2 >> Branch: https://git.xanclic.moe/XanClic/qemu.git fuse-exports-v2 >> >> >> Hi, >> >> Ever since I found out that you can mount FUSE filesystems on regular >> files (not just directories), I had the idea of adding FUSE block >> exports to qemu where you can export block nodes as raw images. The >> best thing is that you’d be able to mount an image on itself, so >> whatever format it may be in, qemu lets it appear as a raw image (and >> you can then use regular tools like dd on it). >> >> The performance is quite bad so far, but we can always try to improve it >> if the need arises. For now I consider it mostly a cute feature to get >> easy access to the raw contents of image files in any format (without >> requiring root rights). >> >> In this version (as opposed to v1 linked above), I integrated the FUSE >> export code into Kevin’s proposed common infrastructure for block >> exports. > > Patches 5-16, 19 and 20: > > Reviewed-by: Kevin Wolf <kwolf@redhat.com> Thanks! So far, (just FYI,) I’ve tried to address your remarks, gathered them in a single fat commit, and pushed everything here: https://github.com/XanClic/qemu/tree/fuse-exports-next (There’s also something about two bugs in fuse_fallocate(), where there are these two I/O loops (one for blk_pwrite_zeroes(), one for blk_pdiscard()), where I both forgot to increment @offset (alongside decrementing @length). Oops.) (And I’ve renamed init_fuse() to init_exports_table(), because that got really confusing with fuse_init().) I’ll split that up, squash it into the respective patches, and send v3 when I get back from PTO. Max