On 17 Feb 2021 at 18:53, Brian Foster wrote:

The blocks used for allocation btrees (bnobt and countbt) are
technically considered free space. This is because as free space is
used, allocbt blocks are removed and naturally become available for
traditional allocation. However, this means that a significant
portion of free space may consist of in-use btree blocks if free
space is severely fragmented.

On large filesystems with large perag reservations, this can lead to
a rare but nasty condition where a significant amount of physical
free space is available, but the majority of actual usable blocks
consist of in-use allocbt blocks. We have a record of a (~12TB, 32
AG) filesystem with multiple AGs in a state with ~2.5GB or so free
blocks tracked across ~300 total allocbt blocks, but effectively at
100% full because the the free space is entirely consumed by
refcountbt perag reservation.

Such a large perag reservation is by design on large filesystems.
The problem is that because the free space is so fragmented, this AG
contributes the 300 or so allocbt blocks to the global counters as
free space. If this pattern repeats across enough AGs, the
filesystem lands in a state where global block reservation can
outrun physical block availability. For example, a streaming
buffered write on the affected filesystem continues to allow delayed
allocation beyond the point where writeback starts to fail due to
physical block allocation failures. The expected behavior is for the
delalloc block reservation to fail gracefully with -ENOSPC before
physical block allocation failure is a possibility.

To address this problem, introduce a percpu counter to track the sum
of the allocbt block counters already tracked in the AGF. Use the
new counter to set these blocks aside at reservation time and thus
ensure they cannot be allocated until truly available. Since this is
only necessary when large reflink perag reservations are in place
and the counter requires a read of each AGF to fully populate, only
enforce on reflink enabled filesystems. This allows initialization
of the counter at ->pagf_init time because the refcountbt perag
reservation init code reads each AGF at mount time.

Note that the counter uses a small percpu batch size to allow the
allocation paths to keep the primary count accurate enough that the
reservation path doesn't ever need to lock and sum the counter.
Absolute accuracy is not required here, just that the counter
reflects the majority of unavailable blocks so the reservation path
fails first.

The changes look good to me from the perspective of logical correctness.

Reviewed-by: Chandan Babu R <redacted>

-- 
chandan