systems running with a heavy filesystem load. Tracking down this
bug was elusive because there were actually two problems. Sometimes
the in-memory check hash was wrong and sometimes the check hash
computed when doing the read was wrong. The occurrence of either
error caused a check-hash mismatch to be reported.
The first error was that the check hash in the in-memory cylinder
group was incorrect. This error was caused by the following
sequence of events:
- We read a cylinder-group buffer and the check hash is valid.
- We update its cg_time and cg_old_time which makes the in-memory
check-hash value invalid but we do not mark the cylinder group dirty.
- We do not make any other changes to the cylinder group, so we
never mark it dirty, thus do not write it out, and hence never
update the incorrect check hash for the in-memory buffer.
- Later, the buffer gets freed, but the page with the old incorrect
check hash is still in the VM cache.
- Later, we read the cylinder group again, and the first page with
the old check hash is still in the VM cache, but some other pages
are not, so we have to do a read.
- The read does not actually get the first page from disk, but rather
from the VM cache, resulting in the old check hash in the buffer.
- The value computed after doing the read does not match causing the
error to be printed.
The fix for this problem is to only set cg_time and cg_old_time as
the cylinder group is being written to disk. This keeps the in-memory
check-hash valid unless the cylinder group has had other modifications
which will require it to be written with a new check hash calculated.
It also requires that the check hash be recalculated in the in-memory
cylinder group when it is marked clean after doing a background write.
The second problem was that the check hash computed at the end of the
read was incorrect because the calculation of the check hash on
completion of the read was being done too soon.
- When a read completes we had the following sequence:
- bufdone()
-- b_ckhashcalc (calculates check hash)
-- bufdone_finish()
--- vfs_vmio_iodone() (replaces bogus pages with the cached ones)
- When we are reading a buffer where one or more pages are already
in memory (but not all pages, or we wouldn't be doing the read),
the I/O is done with bogus_page mapped in for the pages that exist
in the VM cache. This mapping is done to avoid corrupting the
cached pages if there is any I/O overrun. The vfs_vmio_iodone()
function is responsible for replacing the bogus_page(s) with the
cached ones. But we were calculating the check hash before the
bogus_page(s) were replaced. Hence, when we were calculating the
check hash, we were partly reading from bogus_page, which means
we calculated a bad check hash (e.g., because multiple pages have
been mapped to bogus_page, so its contents are indeterminate).
The second fix is to move the check-hash calculation from bufdone()
to bufdone_finish() after the call to vfs_vmio_iodone() so that it
computes the check hash over the correct set of pages.
With these two changes, the occasional cylinder-group check-hash
errors are gone.
Submitted by: David Pfitzner <dpfitzner@netflix.com>
Reviewed by: kib
Tested by: David Pfitzner
47806d1b93