Обсуждение: Re: pgsql: dshash: Add sequential scan support.

[Re-directing to -hackers]

On Fri, Mar 11, 2022 at 2:27 PM Andres Freund <andres@anarazel.de> wrote:
> On 2022-03-10 20:09:56 -0500, Tom Lane wrote:
> > Andres Freund <andres@anarazel.de> writes:
> > > dshash: Add sequential scan support.
> > > Add ability to scan all entries sequentially to dshash. The interface is
> > > similar but a bit different both from that of dynahash and simple dshash
> > > search functions. The most significant differences is that dshash's interfac
> > > always needs a call to dshash_seq_term when scan ends.
> >
> > Umm ... what about error recovery?  Or have you just cemented the
> > proposition that long-lived dshashes are unsafe?
>
> I don't think this commit made it worse. dshash_seq_term() releases an lwlock
> (which will be released in case of an error) and unsets
> hash_table->find_[exclusively_]locked. The latter weren't introduced by this
> patch, and are also set by dshash_find().
>
> I agree that ->find_[exclusively_]locked are problematic from an error
> recovery perspective.

Right, as seen in the build farm at [1].  Also reproducible with something like:

@@ -269,6 +269,14 @@ dsm_impl_posix(dsm_op op, dsm_handle handle, Size
request_size,
                return false;
        }

+       /* XXX random fault injection */
+       if (op == DSM_OP_ATTACH && random() < RAND_MAX / 8)
+       {
+               close(fd);
+               elog(ERROR, "chaos");
+               return false;
+       }
+

I must have thought that it was easy and practical to write no-throw
straight-line code and be sure to reach dshash_release_lock(), but I
concede that it was a bad idea: even dsa_get_address() can throw*, and
you're often likely to need to call that while accessing dshash
elements.  For example, in lookup_rowtype_tupdesc_internal(), there is
a sequence dshash_find(), ..., dsa_get_address(), ...,
dshash_release_lock(), and I must have considered the range of code
between find and release to be no-throw, but now I know that it is
not.

> It's per-backend state at least and just used for assertions. We could remove
> it. Or stop checking it in places where it could be set wrongly: dshash_find()
> and dshash_detach() couldn't check anymore, but the rest of the assertions
> would still be valid afaics?

Yeah, it's all for assertions... let's just remove it.  Those
assertions were useful to me at some stage in development but won't
hold as well as I thought, at least without widespread PG_FINALLY(),
which wouldn't be nice.

*dsa_get_address() might need to adjust the memory map with system
calls, which might fail.  If you think of DSA as not only an allocator
but also a poor man's user level virtual memory scheme to tide us over
until we get threads, then this is a pretty low level kind of
should-not-happen failure that is analogous on some level to SIGBUS or
SIGSEGV or something like that, and we should PANIC.  Then we could
claim that dsa_get_address() is no-throw.  At least, that was one
argument I had with myself while investigating that strange Solaris
shm_open() failure, but ... I lost the argument.  It's quite an
extreme position to take just to support these assertions, which are
of pretty limited value.

[1] https://www.postgresql.org/message-id/20220701232009.jcwxpl45bptaxv5n%40alap3.anarazel.de

At Mon, 4 Jul 2022 14:55:43 +1200, Thomas Munro <thomas.munro@gmail.com> wrote in 
> [Re-directing to -hackers]
> 
> On Fri, Mar 11, 2022 at 2:27 PM Andres Freund <andres@anarazel.de> wrote:
> > It's per-backend state at least and just used for assertions. We could remove
> > it. Or stop checking it in places where it could be set wrongly: dshash_find()
> > and dshash_detach() couldn't check anymore, but the rest of the assertions
> > would still be valid afaics?
> 
> Yeah, it's all for assertions... let's just remove it.  Those
> assertions were useful to me at some stage in development but won't
> hold as well as I thought, at least without widespread PG_FINALLY(),
> which wouldn't be nice.
> 
> *dsa_get_address() might need to adjust the memory map with system
> calls, which might fail.  If you think of DSA as not only an allocator
> but also a poor man's user level virtual memory scheme to tide us over
> until we get threads, then this is a pretty low level kind of
> should-not-happen failure that is analogous on some level to SIGBUS or
> SIGSEGV or something like that, and we should PANIC.  Then we could
> claim that dsa_get_address() is no-throw.  At least, that was one
> argument I had with myself while investigating that strange Solaris
> shm_open() failure, but ... I lost the argument.  It's quite an
> extreme position to take just to support these assertions, which are
> of pretty limited value.
> 
> [1] https://www.postgresql.org/message-id/20220701232009.jcwxpl45bptaxv5n%40alap3.anarazel.de

FWIW, the discussion above is convincing to me and the patch looks good.

regards.

-- 
Kyotaro Horiguchi
NTT Open Source Software Center

Hi,

On 2022-07-04 14:55:43 +1200, Thomas Munro wrote:
> Right, as seen in the build farm at [1].  Also reproducible with something like:
>
> @@ -269,6 +269,14 @@ dsm_impl_posix(dsm_op op, dsm_handle handle, Size
> request_size,
>                 return false;
>         }
>
> +       /* XXX random fault injection */
> +       if (op == DSM_OP_ATTACH && random() < RAND_MAX / 8)
> +       {
> +               close(fd);
> +               elog(ERROR, "chaos");
> +               return false;
> +       }
> +
>
> I must have thought that it was easy and practical to write no-throw
> straight-line code and be sure to reach dshash_release_lock(), but I
> concede that it was a bad idea: even dsa_get_address() can throw*, and
> you're often likely to need to call that while accessing dshash
> elements.  For example, in lookup_rowtype_tupdesc_internal(), there is
> a sequence dshash_find(), ..., dsa_get_address(), ...,
> dshash_release_lock(), and I must have considered the range of code
> between find and release to be no-throw, but now I know that it is
> not.

Yea - I'd go as far as saying that it's almost never feasible.


> > It's per-backend state at least and just used for assertions. We could remove
> > it. Or stop checking it in places where it could be set wrongly: dshash_find()
> > and dshash_detach() couldn't check anymore, but the rest of the assertions
> > would still be valid afaics?
>
> Yeah, it's all for assertions... let's just remove it.  Those
> assertions were useful to me at some stage in development but won't
> hold as well as I thought, at least without widespread PG_FINALLY(),
> which wouldn't be nice.

Hm. I'd be inclined to at least add a few more
Assert(!LWLockHeldByMe[InMode]()) style assertions. E.g. to
dshash_find_or_insert().



> @@ -572,13 +552,8 @@ dshash_release_lock(dshash_table *hash_table, void *entry)
>      size_t        partition_index = PARTITION_FOR_HASH(item->hash);
>
>      Assert(hash_table->control->magic == DSHASH_MAGIC);
> -    Assert(hash_table->find_locked);
> -    Assert(LWLockHeldByMeInMode(PARTITION_LOCK(hash_table, partition_index),
> -                                hash_table->find_exclusively_locked
> -                                ? LW_EXCLUSIVE : LW_SHARED));
> +    Assert(LWLockHeldByMe(PARTITION_LOCK(hash_table, partition_index)));
>
> -    hash_table->find_locked = false;
> -    hash_table->find_exclusively_locked = false;
>      LWLockRelease(PARTITION_LOCK(hash_table, partition_index));
>  }

This LWLockHeldByMe() doesn't add much - the LWLockRelease() will error out if
we don't hold the lock.

Greetings,

Andres Freund

On Tue, Jul 5, 2022 at 8:54 AM Andres Freund <andres@anarazel.de> wrote:
> On 2022-07-04 14:55:43 +1200, Thomas Munro wrote:
> > > It's per-backend state at least and just used for assertions. We could remove
> > > it. Or stop checking it in places where it could be set wrongly: dshash_find()
> > > and dshash_detach() couldn't check anymore, but the rest of the assertions
> > > would still be valid afaics?
> >
> > Yeah, it's all for assertions... let's just remove it.  Those
> > assertions were useful to me at some stage in development but won't
> > hold as well as I thought, at least without widespread PG_FINALLY(),
> > which wouldn't be nice.
>
> Hm. I'd be inclined to at least add a few more
> Assert(!LWLockHeldByMe[InMode]()) style assertions. E.g. to
> dshash_find_or_insert().

Yeah, I was wondering about that, but it needs to check the whole 128
element lock array.  Hmm, yeah that seems OK for assertion builds.
Since there were 6 places with I-hold-no-lock assertions, I shoved the
loop into a function so I could do:

-               Assert(!status->hash_table->find_locked);
+               assert_no_lock_held_by_me(hash_table);

> > +     Assert(LWLockHeldByMe(PARTITION_LOCK(hash_table, partition_index)));
> >
> > -     hash_table->find_locked = false;
> > -     hash_table->find_exclusively_locked = false;
> >       LWLockRelease(PARTITION_LOCK(hash_table, partition_index));

> This LWLockHeldByMe() doesn't add much - the LWLockRelease() will error out if
> we don't hold the lock.

Duh.  Removed.

Hi,

On 2022-07-05 11:20:54 +1200, Thomas Munro wrote:
> On Tue, Jul 5, 2022 at 8:54 AM Andres Freund <andres@anarazel.de> wrote:
> > > Yeah, it's all for assertions... let's just remove it.  Those
> > > assertions were useful to me at some stage in development but won't
> > > hold as well as I thought, at least without widespread PG_FINALLY(),
> > > which wouldn't be nice.
> >
> > Hm. I'd be inclined to at least add a few more
> > Assert(!LWLockHeldByMe[InMode]()) style assertions. E.g. to
> > dshash_find_or_insert().
> 
> Yeah, I was wondering about that, but it needs to check the whole 128
> element lock array.

I think it'd be ok to just check the current partition - yes, it'd not catch
cases where we're still holding a lock on another partition, but that's imo
not too bad?


> Hmm, yeah that seems OK for assertion builds.
> Since there were 6 places with I-hold-no-lock assertions, I shoved the
> loop into a function so I could do:
> 
> -               Assert(!status->hash_table->find_locked);
> +               assert_no_lock_held_by_me(hash_table);

I am a *bit* wary about the costs of that, even in assert builds - each of the
partition checks in the loop will in turn need to iterate through
held_lwlocks. But I guess we can also just later weaken them if it turns out
to be a problem.

Greetings,

Andres Freund

On Tue, Jul 5, 2022 at 11:25 AM Andres Freund <andres@anarazel.de> wrote:
> On 2022-07-05 11:20:54 +1200, Thomas Munro wrote:
> > Since there were 6 places with I-hold-no-lock assertions, I shoved the
> > loop into a function so I could do:
> >
> > -               Assert(!status->hash_table->find_locked);
> > +               assert_no_lock_held_by_me(hash_table);
>
> I am a *bit* wary about the costs of that, even in assert builds - each of the
> partition checks in the loop will in turn need to iterate through
> held_lwlocks. But I guess we can also just later weaken them if it turns out
> to be a problem.

Maybe we should add assertion support for arrays of locks, so we don't
need two levels of loop?  Something like the attached?

On Tue, Jul 5, 2022 at 3:21 PM Thomas Munro <thomas.munro@gmail.com> wrote:
> On Tue, Jul 5, 2022 at 11:25 AM Andres Freund <andres@anarazel.de> wrote:
> > On 2022-07-05 11:20:54 +1200, Thomas Munro wrote:
> > > Since there were 6 places with I-hold-no-lock assertions, I shoved the
> > > loop into a function so I could do:
> > >
> > > -               Assert(!status->hash_table->find_locked);
> > > +               assert_no_lock_held_by_me(hash_table);
> >
> > I am a *bit* wary about the costs of that, even in assert builds - each of the
> > partition checks in the loop will in turn need to iterate through
> > held_lwlocks. But I guess we can also just later weaken them if it turns out
> > to be a problem.
>
> Maybe we should add assertion support for arrays of locks, so we don't
> need two levels of loop?  Something like the attached?

Pushed.