We can no longer reliably determine the number of hash table matches,
since we'll stop counting after the first one.
Signed-off-by: Zygo Blaxell <bees@furryterror.org>
We were doing a `LOGICAL_INO` ioctl on every _block_ of a matching extent,
just to see how long it takes. It takes a while!
This could be modified to do an ioctl with the `IGNORE_OFFSET` flag,
once per new extent, but the kernel bug was fixed a long time ago, so
we can start removing all the toxic extent code.
Signed-off-by: Zygo Blaxell <bees@furryterror.org>
When we have multiple possible matches for a block, we proceed in three
phases:
1. retrieve each match's extent refs and put them in a list,
2. iterate over the list converting viable block matches into range matches,
3. sort and flatten the list of range matches into a non-overlapping
list of ranges that cover all duplicate blocks exactly once.
The separation of phase 1 and 2 creates a performance issue when there
are many block matches in phase 1, and all the range matches in phase
2 are the same length. Even though we might quickly find the longest
possible matching range early in phase 2, we first extract all of the
extent refs from every possible matching block in phase 1, even though
most of those refs will never be used.
Fix this by moving the extent ref retrieval in phase 1 into a single
loop in phase 2, and stop looping over matching blocks as soon as any
dedupe range is created. This avoids iterating over a large list of
blocks with expensive `LOGICAL_INO` ioctls in an attempt to improve the
match when there is no hope of improvement, e.g. when all match ranges
are 4K and the content is extremely prevalent in the data.
If we find a matched block that is part of a short matching range,
we can replace it with a block that is part of a long matching range,
because there is a good chance we will find a matching hash block in
the long range by looking up hashes after the end of the short range.
In that case, overlapping dedupe ranges covering both blocks in the
target extent will be inserted into the dedupe list, and the longest
matches will be selected at phase 3. This usually provides a similar
result to that of the loop in phase 1, but _much_ more efficiently.
Some operations are left in phase 1, but they are all using internal
functions, not ioctls.
Signed-off-by: Zygo Blaxell <bees@furryterror.org>
A laundry list of problems fixed:
* Track which physical blocks have been read recently without making
any changes, and don't read them again.
* Separate dedupe, split, and hole-punching operations into distinct
planning and execution phases.
* Keep the longest dedupe from overlapping dedupe matches, and flatten
them into non-overlapping operations.
* Don't scan extents that have blocks already in the hash table.
We can't (yet) touch such an extent without making unreachable space.
Let them go.
* Give better information in the scan summary visualization: show dedupe
range start and end points (<ddd>), matching blocks (=), copy blocks
(+), zero blocks (0), inserted blocks (.), unresolved match blocks
(M), should-have-been-inserted-but-for-some-reason-wasn't blocks (i),
and there's-a-bug-we-didn't-do-this-one blocks (#).
* Drop cached data from extents that have been inserted into the hash
table without modification.
* Rewrite the hole punching for uncompressed extents, which apparently
hasn't worked properly since the beginning.
Nuisance dedupe elimination:
* Don't do more than 100 dedupe, copy, or hole-punch operations per
extent ref.
* Don't split an extent or punch a hole unless dedupe would save at
least half of the extent ref's size.
* Write a "skip:" summary showing the planned work when nuisance
dedupe elimination decides to skip an extent.
Signed-off-by: Zygo Blaxell <bees@furryterror.org>
Add a master switch to turn off the entire MultiLock infrastructure for
testing, without having to remove and add all the individual entry points.
Signed-off-by: Zygo Blaxell <bees@furryterror.org>
This prevents the storms of exceptions that occur when a subvol is
deleted. We simply treat the entire tree as if it was empty.
Signed-off-by: Zygo Blaxell <bees@furryterror.org>
try_lock allows specification of a different Task to be run instead of
the current Task when the lock is busy.
Signed-off-by: Zygo Blaxell <bees@furryterror.org>
Originally the limit was 2730 (64KiB worth of ref pointers). This limit
was a little too low for some common workloads, so it was then raised by
a factor of 256 to 699050, but there are a lot of problems with extent
counts that large. Most of those problems are memory usage and speed
problems, but some of them trigger subtle kernel MM issues.
699050 references is too many to be practical. Set the limit to 9999,
only 3-4x larger than the original 2730, to give up on deduplication
when each deduped ref reduces the amount of space by no more than 0.01%.
Signed-off-by: Zygo Blaxell <bees@furryterror.org>
This solves a third bad problem with bees reads:
3. The architecture above the read operations will issue read requests
for the same physical blocks over and over in a short period of time.
Fixing that properly requires rewriting the upper-level code, but a
simple small table of recent read requests can reduce the effect of the
problem by orders of magnitude.
Signed-off-by: Zygo Blaxell <bees@furryterror.org>
This solves some of the worst problems with bees reads:
1. The kernel readahead doesn't work. More precisely, it's much better
adapted for a very different use case: a single thread alternating
between reading a file sequentially and processing the data that was read.
bees has multiple threads which compete for access to IO and then issue
reads in random order immediately after the call to readahead. The kernel
uses idle ioprio scheduling for the readaheads, so the readaheads get
preempted by the random reads, or cancels the readaheads because the
data access pattern isn't sequential after the readahead was issued.
2. Seeking drives perform terribly with multiple competing readers,
especially with btrfs striped profiles where the iops are broken into
tiny stripe-sized pieces. At one point I intended to read the btrfs
device map and figure out which devices can be read in parallel, but to
make that useful, the user needs to have an array with multiple drives
in single profile, or 4+ drives in raid1 profile. In all other cases,
the elaborate calculations always return the same result: there can be
only one reader at a time.
This commit fixes both problems:
1. Don't use the kernel readahead. Use normal reads into a dummy
buffer instead.
2. Allow only one thread to readahead at any time. Once the read is
completed, the data is in the page cache, and all the random-order small
reads that bees does will hit the page cache, not a spinning disk.
In some cases we need to read two things close together, so add a
`bees_readahead_pair` which holds one lock across both reads.
Signed-off-by: Zygo Blaxell <bees@furryterror.org>
The hash table is read sequentially and from a single thread, so
the kernel's implementation of readahead is appropriate here.
Signed-off-by: Zygo Blaxell <bees@furryterror.org>
Commit c3b664fea5 ("context: don't forget
to retry locked extents") removed the critical return that prevents a
Task from processing an extent that is locked.
Put the return back.
Fixes: c3b664fea5 ("context: don't forget to retry locked extents")
Signed-off-by: Zygo Blaxell <bees@furryterror.org>
The kernel has not required a 16 MiB limit on dedupe requests since
v4.18-rc1 b67287682688 ("Btrfs: dedupe_file_range ioctl: remove 16MiB
restriction").
Kernels before v4.18 would truncate the request and return the size
actually deduped in `bytes_deduped`. Kernel v4.18 and later will loop
in the kernel until the entire request is satisfied (although still
in 16 MiB chunks, so larger extents will be split).
Modify the loop in userspace to measure the size the kernel actually
deduped, instead of assuming the kernel will only accept 16 MiB.
On current kernels this will always loop exactly once.
Since we now rely on `bytes_deduped`, make sure it has a sane value.
Signed-off-by: Zygo Blaxell <bees@furryterror.org>
These were added to crucible all the way back in 2018 (1beb61fb78
"crucible: error: record location of exception in what() message")
but it's even more useful in the stack tracer in bees.
Signed-off-by: Zygo Blaxell <bees@furryterror.org>
Since we'll never process more than BEES_MAX_EXTENT_REF_COUNT extent
references by definition, it follows that we should not allocate buffer
space for them when we perform the LOGICAL_INO ioctl.
There is some evidence (particularly
https://github.com/Zygo/bees/issues/260#issuecomment-1627598058) that
the kernel is subjecting the page cache to a lot of disruption when
trying allocate large buffers for LOGICAL_INO.
Signed-off-by: Zygo Blaxell <bees@furryterror.org>
Apparently reinterpret_cast<uint64_t> sign-extends 32-bit pointers.
This is OK when running on a 32-bit kernel that will truncate the pointer
to 32 bits, but when running on a 64-bit kernel, the extra bits are
interpreted as part of the (now very invalid) address.
Use <uintptr_t> instead, which is unsigned, integer, and the same word
size as the arch's pointer type. Ordinary numeric conversion can take
it from there, filling the rest of the word with zeros.
Signed-off-by: Zygo Blaxell <bees@furryterror.org>
The bug is:
v6.3-rc6: f349b15e183d mm: vmalloc: avoid warn_alloc noise caused by fatal signal
The fixes are:
v6.4: 95a301eefa82 mm/vmalloc: do not output a spurious warning when huge vmalloc() fails
v6.3.10: c189994b5dd3 mm/vmalloc: do not output a spurious warning when huge vmalloc() fails
The bug has been backported to LTS, but the fix has not:
v6.2.11: 61334bc29781 mm: vmalloc: avoid warn_alloc noise caused by fatal signal
v6.1.24: ef6bd8f64ce0 mm: vmalloc: avoid warn_alloc noise caused by fatal signal
v5.15.107: a184df0de132 mm: vmalloc: avoid warn_alloc noise caused by fatal signal
Signed-off-by: Zygo Blaxell <bees@furryterror.org>
There was a bug in kernel 6.3 where LOGICAL_INO with IGNORE_OFFSET
sometimes fails to ignore the offset. That bug is now fixed, but
LOGICAL_INO still returns 0 refs much more often than seems appropriate.
This is most likely because bees frequently deletes extents while there
is still work waiting for them in Task queues. In this case, LOGICAL_INO
correctly returns an empty list, because every reference to some extent
is deleted, but the new extent tree with that extent removed is not yet
committed in btrfs.
Add a DEBUG-level log message and an event counter to track these events.
In the absence of a kernel bug, the debug message may indicate CPU time
was wasted performing a search whose outcome could have been predicted.
Signed-off-by: Zygo Blaxell <bees@furryterror.org>
Toxic extents are much less of a problem now than they were in kernels
before 5.7. Downgrade the log message level to reflect their lesser
importance.
Signed-off-by: Zygo Blaxell <bees@furryterror.org>
The critical kernel bugs in send have been fixed for years.
The limitations that remain aren't bugs, and bees has no sustainable
workaround for them.
Also update copyright year range.
Signed-off-by: Zygo Blaxell <bees@furryterror.org>
We check the result of transid_max_nocache(), but not the result of
transid_max(). The latter is a computed result that is even more likely
to be wrong[citation needed].
Signed-off-by: Zygo Blaxell <bees@furryterror.org>
At least one user was significantly confused by "designed for large
filesystems".
The btrfs send workarounds aren't new any more.
Signed-off-by: Zygo Blaxell <bees@furryterror.org>
Clarify that "too large" and "too small" are some distance away from each other.
The Goldilocks zone is _wide_.
The interval between cache drops is now shorter.
Signed-off-by: Zygo Blaxell <bees@furryterror.org>
Each object contains a 16 MiB buffer, which is very heavy for some
malloc implementations.
Keep the objects in a Pool so that their buffers are only allocated and
deallocated once in the process lifetime.
Signed-off-by: Zygo Blaxell <bees@furryterror.org>
Some malloc implementations will try to mmap() and munmap() large buffers
every time they are used, causing a severe loss of performance.
Nothing ever overrode the virtual methods, and there was no virtual
destructor, so they cause compiler warnings at build time when used with
a template that tries to delete pointers to them.
Signed-off-by: Zygo Blaxell <bees@furryterror.org>
ProgressTracker was only freeing memory for work items when they reach
the head of the work tracking queue. If the first work item takes
hours to complete, and thousands of items are processed every second,
this leads to millions of completed items tracked in memory at a time,
wasting gigabytes of system RAM.
Rewrite ProgressHolderState methods to keep only incomplete work items
in memory, regardless of the order in which they are added or removed.
Also fix the unit tests which were relying on the memory leak to work,
and add test cases for code coverage.
Signed-off-by: Zygo Blaxell <bees@furryterror.org>
If the send workaround is enabled, it is possible for two threads (a
thread running the crawl_new task, and a thread attempting to apply the
send workaround) to access the same RootFetcher object at the same time.
That never ends well.
Give each function its own BtrfsRootFetcher object.
Fixes: https://github.com/Zygo/bees/issues/250
Signed-off-by: Zygo Blaxell <bees@furryterror.org>
With SIGTERM and fast exit, the trickle writeback is less important.
We don't want to flood people's IO subsystems with continuous writes.
This really should be configurable at runtime.
Signed-off-by: Zygo Blaxell <bees@furryterror.org>
Do rebuild bees-version.cc if libcrucible changes.
Don't rebuild bees-version.cc if it doesn't change.
Also use the standard suffix for new files.
Signed-off-by: Zygo Blaxell <bees@furryterror.org>
crucible::VERSION doesn't make much sense now that libcrucible no
longer exists as a shared library. Nothing ever referenced it, so
it can go away.
Signed-off-by: Zygo Blaxell <bees@furryterror.org>
According to ioctl_iflags(2):
The type of the argument given to the FS_IOC_GETFLAGS and
FS_IOC_SETFLAGS operations is int *, notwithstanding the
implication in the kernel source file include/uapi/linux/fs.h
that the argument is long *.
So this code doesn't work on be64 machines.
Also, Valgrind complains about it.
Signed-off-by: Zygo Blaxell <bees@furryterror.org>
A subtle distinction, and not one that is particularly relevant to bees,
but it does make toolchains complain.
Signed-off-by: Zygo Blaxell <bees@furryterror.org>
Another instance of the pattern where we derived a crucible class
from a btrfs struct. Make it an automatic variable instead.
Signed-off-by: Zygo Blaxell <bees@furryterror.org>
This was fixed in
7f660f50b lib: fs: stop using libbtrfs-dev helper functions to re-enable buffer length checks
but apparently some copies live on.
Signed-off-by: Zygo Blaxell <bees@furryterror.org>
These tools are obsolete. fiemap was a thin wrapper around FIEMAP,
but FIEMAP is not useful on btrfs. fiewalk was a thin wrapper around
BtrfsExtentWalker, but development on BtrfsExtentWalker has been
abandoned.
Signed-off-by: Zygo Blaxell <bees@furryterror.org>
When a hash table write fails, we skip over the write throttling because
we didn't report that we successfully wrote an extent. This can be bad
if the filesystem is full and the allocations for writes are burning a
lot of CPU time searching for free space.
We also don't retry the write later on since we assume the extent is
clean after a write attempt whether it was successful or not, so the
extent might not be written out later when writes are possible again.
Check whether a hash extent is dirty, and always throttle after
attempting the write.
If a write fails, leave the extent dirty so we attempt to write it out
the next time flush cycles through the hash table. During shutdown
this will reattempt each failing write once, after that the updated hash
table data will be dropped.
Signed-off-by: Zygo Blaxell <bees@furryterror.org>
Calling 'bees -m4' should not call 'std::terminate()', but it does.
Use catch_all instead. It will still pass the exit value to return
from main.
Signed-off-by: Zygo Blaxell <bees@furryterror.org>
We only use BtrfsExtentInfo when it's exactly equivalent to the
base, so drop the derived class.
While we're here, fix BtrfsExtentSame::add so it uses a btrfs-compatible
uint64_t instead of an off_t.
Signed-off-by: Zygo Blaxell <bees@furryterror.org>
BEESTOOLONG was always reporting a size of zero, and the offset of the
end of the readahead region. Report the original size instead (and also
in BEESTRACE and BEESNOTE).
Signed-off-by: Zygo Blaxell <bees@furryterror.org>
Drop the crawl_restart counter, it doesn't happen here (or anywhere else).
Add the crawl_again counter for extents that are restarted due to an
extent-level lock.
Signed-off-by: Zygo Blaxell <bees@furryterror.org>
libcrucible can deal with the Linux kernel and/or libc's thread name
limitations. No need to duplicate that work in bees.
Signed-off-by: Zygo Blaxell <bees@furryterror.org>
It turns out I've been using pthread_setname_np wrong the whole time:
* on Linux, the thread name length is 15 characters.
TASK_COMM_LEN is 16 bytes, and the last one is always 0.
This is now hardcoded in many places and cannot be changed.
* pthread_setname_np doesn't return -errno, so DIE_IF_MINUS_ERRNO
was the wrong macro. On the other hand, we never want to do anything
differently when pthread_setname_np fails, so we never needed to
check the return value.
Also, libc silently ignores attempts to set the thread name when it is too
long. That's almost certainly a libc bug, but libc probably suppresses
the error result for the same reasons I ignore the error result.
Wrap the pthread_setname function with a C++ std::string overload that
truncates the argument at 15 characters, so we at least get the first
part of the task name in the thread name field. Later commits can deal
with making the bees thread names shorter.
Also wrap pthread_getname for symmetry.
Signed-off-by: Zygo Blaxell <bees@furryterror.org>
The caller of scan_forward has to stop advancing the BeesFileCrawl
position when an extent lock blocks a scan, so that it will resume
from the same position when the Task is scheduled again; otherwise,
bees simply skips over the extent and leave it incompletely deduped.
Signed-off-by: Zygo Blaxell <bees@furryterror.org>
Restart crawl_more (and update crawl roots and flush FD caches) every
time the transid changes, and only when the transid changes, but
not more often than a reasonable minimum poll interval.
Clean up the log message: use the proper thread name and remove
the wildly inaccurate estimate of when crawl will resume.
Signed-off-by: Zygo Blaxell <bees@furryterror.org>
We don't need to cache 65536 extent maps, especially if each one
can have almost 700K references.
Valgrind's massif tool points to the extent map cache as a very
large memory allocator, but test runs with memcg disagree.
Signed-off-by: Zygo Blaxell <bees@furryterror.org>
If we have loadavg targeting enabled, there may be no worker threads
available to respond to new subvols, so we should not bother updating
the subvols list.
Put insert_new_crawl into a Task so it only executes when a worker
is available.
Signed-off-by: Zygo Blaxell <bees@furryterror.org>
On large filesystems where the min_transid of all subvols gets stuck at 0,
bees may lose the ability to effectively track recent data. A secondary sort
by max_transid will allow scanning newer subvols that were created after bees
started running on the filesystem, but before bees completed the first scan
of all subvols.
On the other hand, the secondary sort does a reverse version of the
sequential scan mode, and the sequential scan mode is simply awful.
Disable it for now.
Signed-off-by: Zygo Blaxell <bees@furryterror.org>
Also attempted to clarify the descriptions of the modes based on
feedback and questions from users over the years.
Signed-off-by: Zygo Blaxell <bees@furryterror.org>
Split each scan mode into two distinct phases:
1. A heavy discovery phase, where we search the entire filesystem
for something (new items in subvol trees in this case).
2. A light consuming phase, where we fetch extents to dedupe
from places that we found in the discovery phase.
Part 1 recomputes the subvol ordering every time there is a new transid.
For some scan modes this computation is quite expensive, far too costly
to pay for every extent, so we do it no more than once per transaction.
Part 2 is run every time a worker thread hits the crawl_more Task.
It simply pulls one extent from the first crawler off a sorted list,
removing the crawler from the list when the crawler runs out of data.
Part 1 creates a new structure and swaps it into place, while Part 2
continues to run using the previous strucuture. Neither of these
need to block the other, so they don't.
The separate class and base pointer also make it easer to add new scan
modes that are not based on subvol trees or that don't use BeesCrawl.
While we're here, fix up some method visibility in BeesRoots.
Signed-off-by: Zygo Blaxell <bees@furryterror.org>
Set the constructor's default scan mode to an invalid mode, so if we
change the default, we don't have to update two places.
Signed-off-by: Zygo Blaxell <bees@furryterror.org>
Crawl mode 3 'recent' prioritizes data from new updates to previously
scanned subvols over subvols that have not been completely scanned yet.
If no such new data exists, falls back to a variation of 'lockstep'
scan mode.
This enables us to keep up with new data as it arrives, a key weakness
of all the other scan modes, and worth violating our unwritten "no new
scan modes until we have extent-tree dedupe working" policy for.
Signed-off-by: Zygo Blaxell <bees@furryterror.org>
Inode-oriented scan workers must do all of their work sequentially,
so it's counterproductive to spawn a Task to do a background dedupe.
Signed-off-by: Zygo Blaxell <bees@furryterror.org>
When two Tasks attempt to lock the same extent, append the later Task
to the earlier Task's post-exec work queue. This will guarantee that
all Tasks which attempt to manipulate the same extent will execute
sequentially, and free up threads to process other extents.
Similarly, if two scanner threads operate on the same inode, any dedupe
they perform will lock out other scanner threads in btrfs. Avoid this
by serializing Task objects that reference the same file.
This does theoretically use an unbounded amount of memory, but in practice
a Task that encounters a contended extent or inode quickly stops spawning
new Tasks that might increase the queue size, and all Tasks that might
contend for the same lock(s) end up on a single FIFO queue.
Note that the scope of inode locks is intentionally global, i.e. when
an inode is locked, it locks every inode with the same number in every
subvol. This avoids significant lock contention and task queue growth
when the same inode with the same file extents appear in snapshots.
Fixes: https://github.com/Zygo/bees/issues/158
Signed-off-by: Zygo Blaxell <bees@furryterror.org>
Split crawlers into two separate Tasks:
1. a Task which locates the next inode with a new data extent.
2. a Task which scans every new extent in that inode.
This simplifies some lock contention and execution ordering issues.
Files are read sequentially. Workers dynamically scale up or
down as needed, without creating thousands of deferred Task objects.
Workers obtain inode locks for different inodes in btrfs, so they
can work in parallel instead of waiting for each other.
This change in behavior comes with new names for the worker Tasks:
"crawl_master" is now "crawl_more", the singular Task which
creates inode-scanning Tasks.
"crawl_<subvol>" is now "crawl_<subvol>_<inode>".
Signed-off-by: Zygo Blaxell <bees@furryterror.org>
This was done on the development branch three years ago, and
has been creating annoying merge conflicts ever since. Sync
up the branches so they have the same names for these.
Signed-off-by: Zygo Blaxell <bees@furryterror.org>
Drop the cache since we no longer have to open a file every time we
check a subvol's status.
Also stop counting workaround events at the root level twice.
Signed-off-by: Zygo Blaxell <bees@furryterror.org>
btrfs-tree provides classes for low-level access to btrfs tree objects.
An item class is provided to decode polymorphic btrfs item fields.
Several tree classes provide forward and backward iteration over raw
object items at different tree levels.
A csum tree class provides convenient access to csums by bytenr,
supporting all current btrfs csum types.
Wrapper classes for inode and subvol items provide direct access to
btrfs metadata fields without clumsy stat() wrappers or ioctls.
Signed-off-by: Zygo Blaxell <bees@furryterror.org>
Zygo Blaxell