On Fri, 31 Jul 2015, Lukáš Czerner wrote:
> Date: Fri, 31 Jul 2015 12:56:55 +0200 (CEST)
> From: Lukáš Czerner <lczerner@xxxxxxxxxx>
> To: Jan Kara <jack@xxxxxxx>
> Cc: linux-ext4@xxxxxxxxxxxxxxx
> Subject: Re: [PATCH] jbd2: Limit number of reserved credits
>
> On Fri, 31 Jul 2015, Jan Kara wrote:
>
> > Date: Fri, 31 Jul 2015 12:46:04 +0200
> > From: Jan Kara <jack@xxxxxxx>
> > To: Lukáš Czerner <lczerner@xxxxxxxxxx>
> > Cc: Jan Kara <jack@xxxxxxx>, linux-ext4@xxxxxxxxxxxxxxx
> > Subject: Re: [PATCH] jbd2: Limit number of reserved credits
> >
> > On Fri 31-07-15 12:22:43, Lukáš Czerner wrote:
> > > On Fri, 31 Jul 2015, Jan Kara wrote:
> > >
> > > > Date: Fri, 31 Jul 2015 11:46:39 +0200
> > > > From: Jan Kara <jack@xxxxxxx>
> > > > To: Lukas Czerner <lczerner@xxxxxxxxxx>
> > > > Cc: linux-ext4@xxxxxxxxxxxxxxx, jack@xxxxxxx
> > > > Subject: Re: [PATCH] jbd2: Limit number of reserved credits
> > > >
> > > > Hello,
> > > >
> > > > On Fri 31-07-15 10:04:23, Lukas Czerner wrote:
> > > > > Currently there is no limitation on number of reserved credits we can
> > > > > ask for. If we ask for more reserved credits than 1/2 of maximum
> > > > > transaction size, or if total number of credits exceeds the maximum
> > > > > transaction size per operation (which is currently only possible with
> > > > > the former) we will spin forever in start_this_handle().
> > > > >
> > > > > Fix this by adding this limitation at the start of start_this_handle().
> > > > >
> > > > > This patch also removes the credit limitation 1/2 of maximum transaction
> > > > > size, since we really only want to limit the number of reserved credits.
> > > > > There is not much point to limit the credits if there is still space in
> > > > > the journal.
> > > > >
> > > > > This accidentally also fixes the online resize, where due to the
> > > > > limitation of the journal credits we're unable to grow file systems with
> > > > > 1k block size and size between 16M and 32M. It has been partially fixed
> > > > > by 2c869b262a10ca99cb866d04087d75311587a30c, but not entirely.
> > > > >
> > > > > Signed-off-by: Lukas Czerner <lczerner@xxxxxxxxxx>
> > > > > ---
> > > > >
> > > > > Honzo I think that this should be enough to remove the limitation of 1/2 of
> > > > > maximum transaction size for regular credits, but I might be missing
> > > > > something, please let me know. Also do you have any specific test case to
> > > > > exercise transaction reservation support - I've only ran xfstests.
> > > > >
> > > > > fs/jbd2/transaction.c | 22 +++++++++++++---------
> > > > > 1 file changed, 13 insertions(+), 9 deletions(-)
> > > > >
> > > > > diff --git a/fs/jbd2/transaction.c b/fs/jbd2/transaction.c
> > > > > index f3d0617..491a328 100644
> > > > > --- a/fs/jbd2/transaction.c
> > > > > +++ b/fs/jbd2/transaction.c
> > > > > @@ -262,20 +262,24 @@ static int start_this_handle(journal_t *journal, handle_t *handle,
> > > > > int rsv_blocks = 0;
> > > > > unsigned long ts = jiffies;
> > > > >
> > > > > + if (handle->h_rsv_handle)
> > > > > + rsv_blocks = handle->h_rsv_handle->h_buffer_credits;
> > > > > +
> > > > > /*
> > > > > - * 1/2 of transaction can be reserved so we can practically handle
> > > > > - * only 1/2 of maximum transaction size per operation
> > > > > + * Limit the number of reserved credits to 1/2 of maximum transaction
> > > > > + * size and limit the number of total credits to not exceed maximum
> > > > > + * transaction size per operation.
> > > > > */
> > > > > - if (WARN_ON(blocks > journal->j_max_transaction_buffers / 2)) {
> > > > > - printk(KERN_ERR "JBD2: %s wants too many credits (%d > %d)\n",
> > > > > - current->comm, blocks,
> > > > > - journal->j_max_transaction_buffers / 2);
> > > > > + if ((rsv_blocks > journal->j_max_transaction_buffers / 2) ||
> > > > > + (rsv_blocks + blocks > journal->j_max_transaction_buffers)) {
> > > > > + printk(KERN_ERR "JBD2: %s wants too many credits "
> > > > > + "credits:%d rsv_credits:%d max:%d\n",
> > > > > + current->comm, blocks, rsv_blocks,
> > > > > + journal->j_max_transaction_buffers);
> > > > > + WARN_ON(1);
> > > > > return -ENOSPC;
> > > > > }
> > > >
> > > > Well, the trouble with this is the following: The currently running
> > > > transaction has X reserved credits and Y normal credits. We know X+Y <=
> > > > journal->j_max_transaction_buffers. Now you request additional A reserved
> > > > and B normal credits. Suppose we cannot fit in the current transaction -
> > > > i.e., X+Y+A+B > journal->j_max_transaction_buffers. The only thing we can do
> > > > is to push running transaction to commit and start a new one. However, the
> > > > new transaction will also have X reserved credits - you inherit reserved
> > > > credits from the previous transaction until they are converted to normal
> > > > credits. So if X+A+B is still > journal->j_max_transaction_buffers, you
> > > > still cannot start current handle and you'd have to wait until someone
> > > > converts his reserved credits.
> > >
> > > Ok I understand, but isn't this true either way ? If anything the
> > > limit might make it worse in that case because if
> > >
> > > X+A+B is still > journal->j_max_transaction_buffers
> > >
> > > in the new case without the limit then it's definitely true for the
> > > case with the limit as well. The number of reserved credits is
> > > limited in both cases so it's not really a factor, is it ?
> > >
> > > Yes in the limitless case it might happen that we have so much
> > > normal credits that we can't fit in the reserved credits so we have
> > > to commit and start a new one, but that's true in both cases only
> > > with the limit it will happen sooner and possible more often because
> > > we just have less space to work with.
> > >
> > > Sorry if I am asking dumb questions, but I am trying to understand
> > > how is this supposed to work.
> > >
> > > And above all that limitation we're talking about is a hard limit
> > > which you're not supposed to hit ever. Only if something is really
> > > wrong and is asking for a handle with way too much credits...that's
> > > not what can normally happen. So what's the problem again ?
> >
> > Thanks for correcting me! I was conflating two different conditions in the
> > transaction handling code. So with the change you propose, it would be only
> > possible that starting of large handles would keep pushing transactions to
> > commit because it couldn't fit the handle into the running transaction
> > because of reserved credits. So if we wanted to relieve the condition as
> > you suggest, we'd also need to modify the logic in
> > add_transaction_credits() to wait on j_wait_reserved in case number of
> > reserved credits of current trans + number of credits requested for the handle
> > is too big. But that looks doable...
>
> Ah, right. Thanks, I'll resend the patch.
One more question tough. Since we already check for available blocks
in the transaction in add_transaction_credits() with:
needed = atomic_add_return(total, &t->t_outstanding_credits);
if (needed > journal->j_max_transaction_buffers) {
...
then we know that we have enough space, so we just need to make sure
that number of reserved credits is limited to 1/2 of
j_max_transaction_buffers right ? So nothing more is needed except of
maybe making sure that this is still true while we wait in case the
reserved credits exceed 1/2 of the transaction. Something like this
in the "if (needed > journal->j_max_transaction_buffers / 2)"
condition should be enough ?
wait_event(journal->j_wait_reserved,
atomic_read(&journal->j_reserved_credits) + total
<= journal->j_max_transaction_buffers);
though I am not entirely sure this is necessary.
Thanks!
-Lukas
>
> -Lukas
>
> >
> > Honza
> >
> > >
> > > Thanks!
> > > -Lukas
> > >
> > >
> > > >
> > > > However these waits will create journal stalls causing possible performance
> > > > issues and also introduce a lock dependency - suddently you are not allowed
> > > > to acquire locks ranking above transaction start before starting a reserved
> > > > handle (as these locks can be held by processes being stuck waiting for
> > > > reserved credits to convert).
> > > >
> > > > So overall halving the maximum allowed credits seemed like the least
> > > > painful solution to the problem.
> > > >
> > > > Honza
> > > > >
> > > > > - if (handle->h_rsv_handle)
> > > > > - rsv_blocks = handle->h_rsv_handle->h_buffer_credits;
> > > > > -
> > > > > alloc_transaction:
> > > > > if (!journal->j_running_transaction) {
> > > > > /*
> > > > > --
> > > > > 1.8.3.1
> > > > >
> > > >
> >
