>>>>> "Tom" == Tom Lane <tgl@sss.pgh.pa.us> writes:
Tom> That doesn't work, unless you insist that the first backend Tom> can't exit its transaction until all the
otherones are done. Tom> Which introduces its own possibilities for deadlock --- but Tom> even worse, how does the
firstbackend *know* that the other Tom> ones are done? You're right back where you started: it has Tom> to be
possibleto tell which backends have share-locked a Tom> particular row.
Is a count a solution ?
The first backend gets the S lock on the row - I'm assuming you plan
to do it by recording it on the tuple and not in a shared memory lock
table, which means that you might have to unnecessarily write an
unmodified page if its buffer pool frame is stolen.
The problem is that on commit time, you must carefully decrement the
count value of shared locks on any tuple that you own. This can be
accomplished by having each backend keep track of the list of files
and TIDs for any rows for which it acquired S locks. Is this the same
way that pgsql releases the X locks ?
Bruce, I don't disagree that MVCC has the very nice property that
writers don't block readers. However, I don't buy that 2-phase
locking, with lock escalation is either unworkable because of too many
locks, or causes any extra pain for the user application (apart from
the fact that writers not blocking readers gives you more concurrency
at some very minor overhead of not being strictly serializable).
--
Pip-pip
Sailesh
http://www.cs.berkeley.edu/~sailesh