On Sep 18, 2021, at 21:49, Theodore Ts'o <tytso(a)mit.edu&gt; wrote: On Fri, Sep 17, 2021 at 04:06:55PM -0500, Eric Blake wrote: > TL;DR summary: is there documented best practices for parallel access > to the same inode of an ext2 filesystem from multiple threads? > > First, a meta-question: is there a publicly archived mailing list for > questions on e2fsprogs? The README merely mentions Ted's email > address, and http://e2fsprogs.sourceforge.net/ is silent on contact > information, although with some googling, I found at least > https://patchwork.ozlabs.org/project/linux-ext4/patch/20201205045856.8953... > which suggests linux-ext4(a)vger.kernel.org as a worthwhile list to > mention on the web page. Yes, discussions and patches relating to e2fsprogs take place on linux-ext4(a)vger.kernel.org. (Just as xfsprogs patches and discussions are sent to linux-xfs(a)vger.kernel.org.) > Now, on to my real reason for writing. The nbdkit project is using > the ext2fs library to provide an ext2/3/4 filter on top of any data > being served over NBD (Network Block Device protocol) in userspace: > https://libguestfs.org/nbdkit-ext2-filter.1.html > > Searching for the word 'thread' or 'concurrent' in libext2fs.info came > up with no hits, so I'm going off of minimal documentation, and mostly > what I can ascertain from existing examples (of which I'm not seeing > very many). Historically, libext2fs and e2fsprogs had no pthreads or concurrency access at all. This is because e2fsprogs predates Linux having pthreads support at all. This is _starting_ to change, but more on that in a little. > Right now, the nbdkit filter forces some rather strict serialization > in order to be conservatively safe: for every client that wants to > connect, the nbdkit filter calls ext2fs_open(), then eventually > ext2fs_file_open2(), then exposes the contents of that one extracted > file over NBD one operation at a time, then closes everything back > down before accepting a second client. But we'd LOVE to add some > parallelization; the NBD protocol allows multiple clients, as well as > out-of-order processing of requests from a single client. > > Right away, I already know that calling ext2fs_open() more than once > on the same file system is a recipe for disaster (it is equivalent to > mounting the same block device at once through more than one bare > metal OS, and won't work). So I've got a proposal for how to rework > the nbdkit code to open the file system exactly once and share that > handle among multiple NBD clients: > https://listman.redhat.com/archives/libguestfs/2021-May/msg00028.html So you are apparently calling ext2fs_open() before forking, and then you want to use the ext2fs handle from separate processes. Is that correct? That's not going to work if you are going to try to modify the file system from different processes simultaneously. That's because the libext2fs using a writeback cache. After the fork, each process has its own copy of the wrteiback cache. If you are using threads, older versions of libext2fs don't do any locking before modifying data structures internal to the ext2_fs file handle. So if two threads simultaneously try to use the "ext2_fs fs" handle, they might try to access the block allocation bitmap (for example) at the same time, without locking, and so bad things will happen. You can do your own locking to make sure only one thread is trying to use the fs handle at a time, at which point you should be fine. So you can have multiple clients accessing the file system without having to open the file system, open a file, and then close the file and close the file system before accepting the next client. But only one client can be using the ext2_fs handle at a time, and if you want to share any libext2fs data structure across multiple threads, appropriate read/write locking would be needed. > Is it okay to have two concurrent handles open to the same inode, or > do I need to implement a hash map on my end so that two NBD clients > requesting access to the same file within the ext2 filesystem share a > single inode? If concurrent handles are supported, what mechanism can > I use to ensure that a flush performed on one handle will be visible > for reading from the other handle, as ext2fs_file_flush does not seem > to be strong enough? You could have two threads sharing the same file handle, with locking so that only one thread is using a file handle at a time. Also, since we don't have an analogue for pread(2) and pwrite(2), each thread would have to assume that the fseek position may have changed by some other thread, so after it grabbed the file system lock, and then the per-file lock, it would need to call ext2fs_file_llseek() to make sure file's position is at a known location before calling ext2fs_file_read() or ext2fs_file_write(). (The reason why ext2fs_file_flush() is not strong enough is because that will force writeback, but it doesn't invaludate any cached information about the file's inode or extent tree structure. So if the file inode or extent tree is getting modified by some other thread out from under it, you're going to have a bad time.) > Finally, I see with > https://patchwork.ozlabs.org/project/linux-ext4/patch/20201205045856.8953... > that you recently added EXT2_FLAG_THREADS, as well as > CHANNEL_FLAGS_THREADS. I think it should be fairly straightforward to > tweak my nbdkit custom IO manager to advertise CHANNEL_FLAGS_THREADS > (as the NBD protocol really DOES support parallel outstanding IO > requests), and then add EXT2_FLAG_THREADS into the flags I pss to > ext2fs_file_open2(), to try and get ext2fs to take advantage of > parallel access to the underlying storage (regardless of whether the > clients are parallel coming into ext2fs). Are there any concurrency > issues I should be aware of on that front when updating my code? So this is the _beginning_ of adding threaded support into libext2fs. At the moment, we now have locking for the unix_io.c data structures. This allows multiple threads to safely do read-only operations in parallel. But this is *all* that it allows. This was implemented as part of preparatory work to do parallel e2fsck. The strategy is that we will have several different threads reading from disjoint parts of the file system. So for example, one thread might be reading from block groups 0 -- 100. Another thread might be reading from block groups 101 -- 200. And so on. Each thread will have its own copy of struct e2fsck_struct, and when they are done they will merge their data to the global e2fsck_struct. If there are any inconsistencies that need to be fixed, such that the file system needs to be modified, this will require waiting until all of the threads are done, or other specialized locking inside e2fsck. Of course, in the "happy path", where the file system does not need any repairs, we won't need to do any special locking or waiting, since the workload will be read-only. So we do not have any concurrency support for allocating inodes, or allocating blocks, or assigning blocks to an inode's extent tree, etc. Nor do we currently have any plans to add more concureency support to libext2fs. To do this would require a huge amount of effort, and it would also require making a lot of changes to the underlying data structures. For example, even if we added locking to all of the various data structures hanging off of the ext2_fs handle, if two threads tried to open the same inode using ext2fs_file_open(), the two file handles are completely independent, and there is no way for one thread to do any kind of cache invalidation of another thread's file handle after it has modified the inode. The same is true if one thread is using a directory iterator while another process tries to modify that directory. > Obviously, when the kernel accesses an ext2/3/4 file system, it DOES > support full concurrency (separate user space processes can open > independent handles to the same file.... Yes, and that's because the kernel was designed with that in mind from the beginning. The ext2fs library was originally designed to support programs like e2fsck, mke2fs, and debugfs. None of these tools required concurrency, and as I've mentioned, at the time when libext2fs was first implemented, Linux didn't even *have* threads support. So concurrency wasn't even possible, even if it had been needed at that time. > process must be observed from another). But nbdkit is all about > accessing the data of an ext2 filesystem from userspace, without any > kernel bio involvement, and is thus reliant on whatever concurrency > guarantees the ext2progs library has (or lacks). The e2fsprogs library pretty much doesn't have any concurrency guarantees, sorry. I suspect you could create a layer on top of libext2fs which actually implemented a global inode cache ala the kernel, so that when two threads call something like ext2fs_iget() function, it works like the kernel's iget() function and they get the same inode structure, which is reference counted. Things like the directory iterator would have to be changed into something more like the functions exported by the kernel VFS layer, which would make this layer useless for e2fsck, but it would be more useful for a threaded client that wanted concurrent read/write access to the filesystem from mulitple threads. Cheers, - Ted

On Fri, Sep 17, 2021 at 04:06:55PM -0500, Eric Blake wrote: > TL;DR summary: is there documented best practices for parallel access > to the same inode of an ext2 filesystem from multiple threads? > > First, a meta-question: is there a publicly archived mailing list for > questions on e2fsprogs? The README merely mentions Ted's email > address, and http://e2fsprogs.sourceforge.net/ is silent on contact > information, although with some googling, I found at least > https://patchwork.ozlabs.org/project/linux-ext4/patch/20201205045856.8953... > which suggests linux-ext4(a)vger.kernel.org as a worthwhile list to > mention on the web page. Yes, discussions and patches relating to e2fsprogs take place on linux-ext4(a)vger.kernel.org. (Just as xfsprogs patches and discussions are sent to linux-xfs(a)vger.kernel.org.) > Now, on to my real reason for writing. The nbdkit project is using > the ext2fs library to provide an ext2/3/4 filter on top of any data > being served over NBD (Network Block Device protocol) in userspace: > https://libguestfs.org/nbdkit-ext2-filter.1.html > > Searching for the word 'thread' or 'concurrent' in libext2fs.info came > up with no hits, so I'm going off of minimal documentation, and mostly > what I can ascertain from existing examples (of which I'm not seeing > very many). Historically, libext2fs and e2fsprogs had no pthreads or concurrency access at all. This is because e2fsprogs predates Linux having pthreads support at all. This is _starting_ to change, but more on that in a little. > Right now, the nbdkit filter forces some rather strict serialization > in order to be conservatively safe: for every client that wants to > connect, the nbdkit filter calls ext2fs_open(), then eventually > ext2fs_file_open2(), then exposes the contents of that one extracted > file over NBD one operation at a time, then closes everything back > down before accepting a second client. But we'd LOVE to add some > parallelization; the NBD protocol allows multiple clients, as well as > out-of-order processing of requests from a single client. > > Right away, I already know that calling ext2fs_open() more than once > on the same file system is a recipe for disaster (it is equivalent to > mounting the same block device at once through more than one bare > metal OS, and won't work). So I've got a proposal for how to rework > the nbdkit code to open the file system exactly once and share that > handle among multiple NBD clients: > https://listman.redhat.com/archives/libguestfs/2021-May/msg00028.html

So you are apparently calling ext2fs_open() before forking, and then you want to use the ext2fs handle from separate processes. Is that correct?

That's not going to work if you are going to try to modify the file system from different processes simultaneously. That's because the libext2fs using a writeback cache. After the fork, each process has its own copy of the wrteiback cache. If you are using threads, older versions of libext2fs don't do any locking before modifying data structures internal to the ext2_fs file handle. So if two threads simultaneously try to use the "ext2_fs fs" handle, they might try to access the block allocation bitmap (for example) at the same time, without locking, and so bad things will happen. You can do your own locking to make sure only one thread is trying to use the fs handle at a time, at which point you should be fine. So you can have multiple clients accessing the file system without having to open the file system, open a file, and then close the file and close the file system before accepting the next client. But only one client can be using the ext2_fs handle at a time, and if you want to share any libext2fs data structure across multiple threads, appropriate read/write locking would be needed.

> Is it okay to have two concurrent handles open to the same inode, or > do I need to implement a hash map on my end so that two NBD clients > requesting access to the same file within the ext2 filesystem share a > single inode? If concurrent handles are supported, what mechanism can > I use to ensure that a flush performed on one handle will be visible > for reading from the other handle, as ext2fs_file_flush does not seem > to be strong enough? You could have two threads sharing the same file handle, with locking so that only one thread is using a file handle at a time. Also, since we don't have an analogue for pread(2) and pwrite(2), each thread would have to assume that the fseek position may have changed by some other thread, so after it grabbed the file system lock, and then the per-file lock, it would need to call ext2fs_file_llseek() to make sure file's position is at a known location before calling ext2fs_file_read() or ext2fs_file_write(). (The reason why ext2fs_file_flush() is not strong enough is because that will force writeback, but it doesn't invaludate any cached information about the file's inode or extent tree structure. So if the file inode or extent tree is getting modified by some other thread out from under it, you're going to have a bad time.) > Finally, I see with > https://patchwork.ozlabs.org/project/linux-ext4/patch/20201205045856.8953... > that you recently added EXT2_FLAG_THREADS, as well as > CHANNEL_FLAGS_THREADS. I think it should be fairly straightforward to > tweak my nbdkit custom IO manager to advertise CHANNEL_FLAGS_THREADS > (as the NBD protocol really DOES support parallel outstanding IO > requests), and then add EXT2_FLAG_THREADS into the flags I pss to > ext2fs_file_open2(), to try and get ext2fs to take advantage of > parallel access to the underlying storage (regardless of whether the > clients are parallel coming into ext2fs). Are there any concurrency > issues I should be aware of on that front when updating my code? So this is the _beginning_ of adding threaded support into libext2fs. At the moment, we now have locking for the unix_io.c data structures. This allows multiple threads to safely do read-only operations in parallel. But this is *all* that it allows. This was implemented as part of preparatory work to do parallel e2fsck. The strategy is that we will have several different threads reading from disjoint parts of the file system. So for example, one thread might be reading from block groups 0 -- 100. Another thread might be reading from block groups 101 -- 200. And so on. Each thread will have its own copy of struct e2fsck_struct, and when they are done they will merge their data to the global e2fsck_struct. If there are any inconsistencies that need to be fixed, such that the file system needs to be modified, this will require waiting until all of the threads are done, or other specialized locking inside e2fsck. Of course, in the "happy path", where the file system does not need any repairs, we won't need to do any special locking or waiting, since the workload will be read-only. So we do not have any concurrency support for allocating inodes, or allocating blocks, or assigning blocks to an inode's extent tree, etc. Nor do we currently have any plans to add more concureency support to libext2fs. To do this would require a huge amount of effort, and it would also require making a lot of changes to the underlying data structures. For example, even if we added locking to all of the various data structures hanging off of the ext2_fs handle, if two threads tried to open the same inode using ext2fs_file_open(), the two file handles are completely independent, and there is no way for one thread to do any kind of cache invalidation of another thread's file handle after it has modified the inode. The same is true if one thread is using a directory iterator while another process tries to modify that directory. > Obviously, when the kernel accesses an ext2/3/4 file system, it DOES > support full concurrency (separate user space processes can open > independent handles to the same file.... Yes, and that's because the kernel was designed with that in mind from the beginning. The ext2fs library was originally designed to support programs like e2fsck, mke2fs, and debugfs. None of these tools required concurrency, and as I've mentioned, at the time when libext2fs was first implemented, Linux didn't even *have* threads support. So concurrency wasn't even possible, even if it had been needed at that time. > process must be observed from another). But nbdkit is all about > accessing the data of an ext2 filesystem from userspace, without any > kernel bio involvement, and is thus reliant on whatever concurrency > guarantees the ext2progs library has (or lacks). The e2fsprogs library pretty much doesn't have any concurrency guarantees, sorry. I suspect you could create a layer on top of libext2fs which actually implemented a global inode cache ala the kernel, so that when two threads call something like ext2fs_iget() function, it works like the kernel's iget() function and they get the same inode structure, which is reference counted. Things like the directory iterator would have to be changed into something more like the functions exported by the kernel VFS layer, which would make this layer useless for e2fsck, but it would be more useful for a threaded client that wanted concurrent read/write access to the filesystem from mulitple threads.