On Tue, May 17, 2022 at 07:30:41PM +1000, Dave Chinner wrote: > On Mon, May 16, 2022 at 09:35:14AM -0400, Brian Foster wrote: > > On Thu, May 12, 2022 at 08:24:01AM +1000, Dave Chinner wrote: > > > On Wed, May 11, 2022 at 11:46:23AM -0400, Brian Foster wrote: > > > > On Wed, May 11, 2022 at 08:05:23AM +1000, Dave Chinner wrote: > > > > > On Tue, May 10, 2022 at 12:02:12PM -0700, Darrick J. Wong wrote: > > > > > > On Tue, May 10, 2022 at 09:21:03AM +0300, Amir Goldstein wrote: > > > > > > > On Mon, May 9, 2022 at 9:20 PM Darrick J. Wong <djwong@xxxxxxxxxx> wrote: > > > > > > > > I think the upcoming nrext64 xfsprogs patches took in the first patch in > > > > > > > > that series. > > > > > > > > > > > > > > > > Question: Now that mkfs has a min logsize of 64MB, should we refuse > > > > > > > > upgrades for any filesystem with logsize < 64MB? > > > > > > > > > > > > > > I think that would make a lot of sense. We do need to reduce the upgrade > > > > > > > test matrix as much as we can, at least as a starting point. > > > > > > > Our customers would have started with at least 1TB fs, so should not > > > > > > > have a problem with minimum logsize on upgrade. > > > > > > > > > > > > > > BTW, in LSFMM, Ted had a session about "Resize patterns" regarding the > > > > > > > practice of users to start with a small fs and grow it, which is encouraged by > > > > > > > Cloud providers pricing model. > > > > > > > > > > > > > > I had asked Ted about the option to resize the ext4 journal and he replied > > > > > > > that in theory it could be done, because the ext4 journal does not need to be > > > > > > > contiguous. He thought that it was not the case for XFS though. > > > > > > > > > > > > It's theoretically possible, but I'd bet that making it work reliably > > > > > > will be difficult for an infrequent operation. The old log would probably > > > > > > have to clean itself, and then write a single transaction containing > > > > > > both the bnobt update to allocate the new log as well as an EFI to erase > > > > > > it. Then you write to the new log a single transaction containing the > > > > > > superblock and an EFI to free the old log. Then you update the primary > > > > > > super and force it out to disk, un-quiesce the log, and finish that EFI > > > > > > so that the old log gets freed. > > > > > > > > > > > > And then you have to go back and find the necessary parts that I missed. > > > > > > > > > > The new log transaction to say "the new log is over there" so log > > > > > recovery knows that the old log is being replaced and can go find > > > > > the new log and recover it to free the old log. > > > > > > > > > > IOWs, there's a heap of log recovery work needed, a new > > > > > intent/transaction type, futzing with feature bits because old > > > > > kernels won't be able to recovery such a operation, etc. > > > > > > > > > > Then there's interesting issues that haven't ever been considered, > > > > > like having a discontiguity in the LSN as we physically switch logs. > > > > > What cycle number does the new log start at? What happens to all the > > > > > head and tail tracking fields when we switch to the new log? What > > > > > about all the log items in the AIL which is ordered by LSN? What > > > > > about all the active log items that track a specific LSN for > > > > > recovery integrity purposes (e.g. inode allocation buffers)? What > > > > > about updating the reservation grant heads that track log space > > > > > usage? Updating all the static size calculations used by the log > > > > > code which has to be done before the new log can be written to via > > > > > iclogs. > > > > > > > ... > > > > > > > TBH, if one were to go through the trouble of making the log resizeable, > > > > I start to wonder whether it's worth starting with a format change that > > > > better accommodates future flexibility. For example, the internal log is > > > > already AG allocated space.. why not do something like assign it to an > > > > internal log inode attached to the sb? Then the log inode has the > > > > obvious capability to allocate or free (non-active log) extents at > > > > runtime through all the usual codepaths without disruption because the > > > > log itself only cares about a target device, block offset and size. We > > > > already know a bump of the log cycle count is sufficient for consistency > > > > across a clean mount cycle because repair has been zapping clean logs by > > > > default as such since pretty much forever. > > > > > > Putting the log in an inode isn't needed for allocate/free of raw > > > extents - we already do that with grow/shrink for the tail of an AG. > > > Hence I'm not sure what virtually mapping the log actually gains us > > > over just a raw extent pointed to by the superblock? > > > > > > > I'm not sure what virtually mapping the log has to do with anything..? > > That's what adding the inode BMBT indirection to point at the log > blocks is - it's virtually mapping the physical block location. > > > That seems like something to consider for the future if there's value to > > it.. > > Virtual mapping as in "memory mapping the log" is a different > problem altogether. We've talked about that several times in the > context of DAX to avoid at least one level of memcpy() that the log > writes currently do on PMEM systems. > Ok.. we're getting a bit pedantic here. I realize that adding an extent to an inode creates a virtual mapping, but nothing in this prospective operation uses that mapping. The inode is used purely as a container to hold a reference to an extent across a log quiesce. I do think creating a virtually mapped log is an interesting future consideration related to the aforementioned format change (i.e. hanging a log inode off the sb), but that's a separate topic.. > > > > That potentially reduces log reallocation to a switchover algorithm that > > > > could run at mount time. I.e., a new prospective log extent is allocated > > > > at runtime (and maybe flagged with an xattr or something). The next > > > > mount identifies a new/prospective log, requires/verifies that the old > > > > log is clean, selects the new log extent (based on some currently > > > > undefined selection algorithm) and seeds it with the appropriate cycle > > > > count via synchronous transactions that release any currently inactive > > > > extent(s) from the log inode. Any failure along the way sticks with the > > > > old log and releases the still inactive new extent, if it happens to > > > > exist. We already do this sort of stale resource clean up for other > > > > things like unlinked inodes and stale COW blocks, so the general premise > > > > exists.. hm? > > > > > > That seems like just as much special case infrastructure as a custom > > > log record type to run it online, and overall the algorithm isn't > > > much different. And the online change doesn't require an on-disk > > > format change... > > > > > > > It seems a lot more simple to me, assuming it's viable, but then again I > > don't see a clear enough description of what you're comparing against to > > reason about it with certainty. I also don't think it necessarily > > depends on a format change. I.e., perhaps an online variant could look > > something like the following (pseudo algorithm): > > > > 1. Userspace growfs feeds in a tmpfile or some such donor inode via > > ioctl() that maps the new/prospective log as a data extent. Assume the > > extent is ideally sized/allocated and disk content is properly formatted > > as a valid log with an elevated cycle count for the purpose of this > > sequence. > > It would have to be a single contiguous extent and correctly > aligned to sunit/swidth as the original log would have been. > Otherwise stuff like log stripe unit IO optimisations get broken. > Yeah, and this really could all happen on the kernel side rather than userspace. I just wanted to keep the example simple and factor out as many implementation details as possible to focus the discussion. > > 2. Quiesce the log, freeze the fs, xfs_log_unmount() the active log and > > xfs_log_mount() the new range based on the extent mapped by the donor > > inode. > > > > 3. Commit changes to the new log range to unmap the current extent (i.e. > > the new log) from the donor inode and map the old log range, keeping the > > in-core inode pinned/locked so it cannot be written back. > > So a specialised swap extent operation? Currently we can swap > individual extents between inodes (the rmap extent swap code), but I > don't think we can't swap an inode extent with a bare, unreferenced > "free space" extent right now. > I suppose you could logically view it as a specialized extent swap, but it's really just an inode unmap and map. > > At this point > > the only changes on disk are to the donor's data extent (still mapped by > > the inode). > > And potentially rmap btree updates because the type and owner of > both extents change. There's also the need to mark the old log > extent as unwritten, because if we crash iand recovery doesn't free > the donor inode, we don't want to risk exposing the old contents of > the log to userspace. > rmap already refers to the active log, so we have to deal with rmap updates regardless (if enabled). The inode should be unlinked (or start unlinked as a tmpfile) before the old log is quiesced such that the extent mapping is never accessible to userspace. We could still mark the old log extent as unwritten if we wanted to be extra safe, of course. > Also, Because rmap updates are deferred, the changes will span > multiple transactions via intents, so this will make use of the CIL > and, potentially, commit to the log and insert items into the AIL. > > If we get items in the AIL, we are open to metadata writeback > starting while we are still doing the log changeover. e.g. memory > pressure occurs, direct reclaim causes xfs_reclaim_inodes_nr() to > run which calls xfs_ail_push_all(mp->m_ail).... > Good point. Committing a transaction and simply holding the inode lock may not be sufficient to prevent writeback. I'm not sure this is so difficult a problem to address, though... The bmap map/unmap operations are also intent based dfops (i.e. used by reflink remap). All the switchover really needs at this step is to put something in the prospective log that ensures the mapping updates occur in the event of a crash after the superblock update. All that requires is getting the intents committed, which is already the initial substep of transaction commit. IOW, I suspect all that may be necessary here is to run the intent creation step of defer finish and to roll the transaction. That would only log intents before the sb update and allow the transaction to commit after the sb update. Also note that the bmap deferred op completion already handles rmap updates, if enabled, so there's nothing additional to do there wrt to the inode. We would probably have to include rmapbt log record update intents in the transaction, however. > Hence if we end up with new metadata changes in place, we can't go > back to the old log - we've got changes that are only sequenced in > the new log, so we can only recover at this point from the new log. > If we can just log the intents before updating the superblock, nothing lands in the AIL. Since the operations are deferred by nature, the in-core changes aren't even made until the new log is active. > Yes, we can fix all these things, but we'll end up touching lots of > high level code in subtle ways to make this all work, and I can't > help but think this will end up being a bit fragile as a result. > Perhaps, but I'd prefer to accurately pin down what may or may not be involved before getting into that. I expect anything that implements a log switch to be some degree of complex, invasive and potentially fragile. > > The superblock has not been updated, so any crash/recovery > > will see the original log in the quiesced state, recover as normal and > > process the donor file as an unlinked inode. > > > > (The filesystem at this point has two recoverable logs. The original log > > is unmodified since the quiesce and still referenced by the superblock. > > It can be recovered in the event of a log switch failure. The new log > > still resides in an inode extent mapping, but contains valid log records > > to exchange its own extent mapping with the old log range). > > Noted, but as per above, the potential for metadata writeback to > occur before we've switch logs means we can't just go back to the > old log. hence to make sure this works, we'll need new interlocks to > ensure operations like metadata writeback can't occur while the log > swap transactions are being performed. > > > 4. Sync update the superblock to point at the new log range and unpin > > the donor inode. > > A crash after this point sees the new log in the sb and > > recovers the changes to map the old log into the donor inode. The old > > log range is returned to free space upon unlinked inode processing of > > the donor inode. > > Yeah, the moment we write the superblock, we have to be prepared to > recover from the new log. That's why I think that both the old > log and the new log should just have an identical "log change" > record written to them after the unmount record. That simply records > the location/size of both the old log and the new log and we can > recover from that no matter which side of the superblock write we > crash on. > > i.e. we have the same change record information in both logs and the > recovery action we need to perform only differs by the old log > having to first: > > - update the superblock to point at the new log > - switch to the new log > - restart log recovery on the new log. > The prospect of a single "log termination" or "log chain/move" type record sounds a lot more elegant to me than the initial train of thought earlier in the thread of landing the alloc/free transactions to the old log and the new log and having to validate that recovery works properly at each step. I still don't have enough context to grok how the custom record part provides recovery/consistency and simplifies the implementation.. > > I'm not totally sure that works without prototyping some basic form of > > it, but it seems generally sane to me. > > I think we could make it work, but it my gut tells me it is the > wrong approach. > Fair enough. TBH, I'm not looking to argue or convince anybody about what approach XFS should or should not take. I'm interested in fleshing out the idea so viability can be assessed and the tradeoffs accurately represented. > > There are a lot of implementation > > details to work through as well of course, such as whether the kernel > > allocates the prospective log extent before the quiesce (and seeds it), > > whether to support the ability to allocate the new in-core log subsystem > > before tearing down the old (for more reliable failure handling and so > > it can be activated atomically), how to format changes into the new log, > > etc. etc., but that's far more detail than this is intended to cover... > > *nod* > > I'm not doubting it could be done, but I feel that it requires > far more high level filesystem structures to be manipulated in > intricate ways than I'm comfortable with. > > My main reservation is that using inodes, BMBT and RMAP updates for > physically swapping the location of the log involves layers of the > filesystem that the log itself should not be interacting with. I'm > just not comfortable to tying low level physical log layout > manipulations to high level transactional filesystem metadata > modifications that rely on writing to the log for recovery. > As noted above, I'm not sure this really adds that much complexity. For reference, an updated version of the sequence might look something like: 1. Allocate tmpfile and new log extent. 2. Quiesce old log, freeze fs, tear down old in-core log structures and init the new. 3. Allocate a transaction (against the new log) to unmap the current (new log) extent of the donor inode and map in the old log extent. Create the intents and roll the transaction (but do not commit/finish). 4. Update the superblock to point at the new log. 5. Commit/finish the update transaction to perform the mapping changes. The general idea is the same as before.. a crash before the sync update tosses the prospective log on unlinked inode processing and a crash after it processes the remapping intents and frees up the old log extent in the same way. I'm sure there are some caveats or warts buried in there (and obviously it still involves a transaction), but if it fundamentally works it actually may not be that much new code at all for the switchover piece. The core functionality it depends on is the ability to run xfs_defer_finish() in multiple steps. I think this is kind of how the underlying mechanism works already, so may just involve some refactoring to carry over pending dfops state across multiple calls. > We can - and IMO should - run the physical location swap the > entirely within the log code itself without involving transactions, > log items, the CIL, AIL, etc. The log is capable of writing it's own > records (example: log unmount records), hence we can implement the > physical swap in a recoverable manner at that level without > involving an high level transactions or log items at all. > > If we separate the transactional allocation/freeing of extents from > the exchange of the log location, then we don't have to try to work > out how to run a recoverable transaction that creates a > discontinuity in the mechanism that provides transaction recovery > and is completely recoverable itself. > > The high level filesystem operation can provide the log with a new > extent that is already owned by the log and will remain that way > through the quiesce process. With that, the log location swap > doesn't need any other interactions with the transaction subsystem - > We just need to be able to write to the old log directly, to > initialise the new log with the same information, and to write the > superblock directly. The log already writes the superblock directly > as part of log recovery, so doing it as part of a log swap isn't an > issue. > I think I follow the approach you're describing of burying extent ownership into the log itself with a custom/open-coded log record. I conceptualize this as sort of an xfs_log_remount() mechanism from an old range to a new one. It sounds sensible enough, but I'm not sure I follow what the low level switchover sequence looks like, particularly how crash recovery is handled. Could you run through the key steps that provide the basic atomicity of the operation? For example, I'll just take a stab and assume it goes something like: 1. Allocate the new log extent and commit it in a transaction with a LOGCHANGE record that refers to the old range and new log range. 2. Do some sort of (partial?) quiesce/freeze that drains the AIL but doesn't lose the opencoded LOGCHANGE intent record (and ensures it isn't lost by covering etc.)..?? Tear down old log subsystem and init new one. 3. Format the new log appropriately (cycle, etc.) with a copy (essentially a relog) of the LOGCHANGE intent. 4. Sync update the superblock to point at the new log. 5. Free the old log extent via transaction that also includes a LOGCHANGE intent done (to the new log). So essentially a crash after step 2 recovers the original log and restarts the log change op at the new log format step (since recovery can't confirm whether it occurred) or a crash after step 4 recovers the new log and frees the old based on whether the intent done record exists or not. Is that the right idea? If not, can you swizzle it into something that is? If so, the part that is not so clear to me is what the magic looks like to commit the extent alloc/frees with the open-coded log record(s) without conflicting with quiesce (i.e. managing consistency with the transactional part vs. the open coded record write part), or otherwise managing consistency between the extent alloc and existence of the custom record. Brian > Then we can hand the old extent back to the high level code and it > can free it. Once it is freed, we can queisce the log again and the > new unmount record effectively cancels the log swap record at the > start of the log as it is no longer between the head and tail. > > Job done, and at that point the filesystem can be resumed on the new > log. > > If you want to work out the high level transactional allocation, > quiesce and freeing side of things, I can write the low level "log > change" record and recovery code taht swaps the location of the log. > I think if we separate the two sets of operations and isolate them > within the appropriate layers, everything becomes a lot easier and > it doesn't need to fall on just one person to do everything... > > Cheers, > > Dave. > -- > Dave Chinner > david@xxxxxxxxxxxxx >
