vlserver: Return error when growing beyond 2 GiB

vlserver: Return error when growing beyond 2 GiB

In the vlserver, when we add a new vlentry or extent block, we grow
the VLDB by doing something like this:

    vital_header.eofPtr += sizeof(item);

Since we don't check for overflow, and all of our offset-related
variables are signed 32-bit integers, this can cause some odd behavior
if we try to grow the database to be over 2 GiB in size.

To avoid this, change the two places in vlserver code that grow the
database to use a new function, grow_eofPtr(), which checks for 31-bit
overflow. If we are about to overflow, log a message and return an
error.

See the following for a specific example of our "odd behavior" when we
overflow the 2 GiB limit in the VLDB:

With 1 extent block, we can create 14509076 vlentries successfully. On
the 14509077th vlentry, we'll attempt to write the entry to offset
2147483560 (0x7FFFFFA8). Since a vlentry is 148 bytes long, we'll
write all the way through offset 2147483707 (0x8000003B), which is
over the 31-bit limit.

In the udisk subsystem, this results in writing to page numbers
2097151, and -2097152 (since our ubik pages are 1k, and going over the
31-bit limit causes us to treat offsets as negative). These pages
start at physical offsets 2147482688 (0x7FFFFC40) and -2147483584
(-0x7FFFFFC0) in our vldb.DB0 (where offset is page*1024+64).

Modifying each of these pages involves reading in the existing page
first, modifying the parts we are changing, and writing it back. This
works just fine for 2097151, but of course fails for -2097152. The
latter fails in DReadBuffer when eventually our pread() fails with
EINVAL, and causes ubik to log the message:

    Ubik: Error reading database file: errno=22

But when DReadBuffer fails, DReadBufferForWrite assumes this is due to
EOF, and just creates a new buffer for the given page (DNewBuffer).
So, the udisk_write() call ultimately succeeds.

When we go to flush the dirty data to disk when committing the
transaction, after we have successfully written the transaction log,
DFlush() fails for the -2097152 page when the pwrite() call eventually
fails with EINVAL, causing ubik to panic, logging the messages:

    Ubik PANIC:
    Writing Ubik DB modifications

When the vlserver gets restarted by bosserver, we then process the
transaction log, and perform the operations in the log before starting
up (ReplayLog). The log records the actual data we wrote, not split
into pages, and the log-replaying code writes directly to the db
usying uphys_write instead of udisk_write. So, because of this, the
write actually succeeds when replaying the log, since we just write
148 bytes to offset 2147483624 (0x7FFFFFE8), and no negative offsets
are used.

The vlserver will then be able to run, but will be unable to read that
newly-created vlentry, since it involves reading a ubik page beyond
the 31-bit boundary. That means trying to lookup that entry will fail
with i/o errors, and as well as any entry on the same hash chains as
the new entry (since the new entry will be added to the head of the
hash chain). Listing all entries in the database will also just show
an empty database, since our vital_header.eofPtr will be negative, and
we determine EOF by comparing our current blockindex to the value in
eofPtr.

Change-Id: Ie0b7ac61f9121fa265686449efbae8e18edb1896
Reviewed-on: https://gerrit.openafs.org/14180
Tested-by: BuildBot <buildbot@rampaginggeek.com>
Reviewed-by: Benjamin Kaduk <kaduk@mit.edu>
Reviewed-by: Cheyenne Wills <cwills@sinenomine.net>