Mauricio Faria de Oliveira <mauricfo@xxxxxxxxxxxxxxxxxx> writes:
> Currently, aio-nr is incremented in steps of 'num_possible_cpus() * 8'
> for io_setup(nr_events, ..) with 'nr_events < num_possible_cpus() * 4':
>
> ioctx_alloc()
> ...
> nr_events = max(nr_events, num_possible_cpus() * 4);
> nr_events *= 2;
> ...
> ctx->max_reqs = nr_events;
> ...
> aio_nr += ctx->max_reqs;
> ....
>
> This limits the number of aio contexts actually available to much less
> than aio-max-nr, and is increasingly worse with greater number of CPUs.
>
> For example, with 64 CPUs, only 256 aio contexts are actually available
> (with aio-max-nr = 65536) because the increment is 512 in that scenario.
>
> Note: 65536 [max aio contexts] / (64*4*2) [increment per aio context]
> is 128, but make it 256 (double) as counting against 'aio-max-nr * 2':
>
> ioctx_alloc()
> ...
> if (aio_nr + nr_events > (aio_max_nr * 2UL) ||
> ...
> goto err_ctx;
> ...
>
> This patch uses the original value of nr_events (from userspace) to
> increment aio-nr and count against aio-max-nr, which resolves those.
>
> Signed-off-by: Mauricio Faria de Oliveira <mauricfo@xxxxxxxxxxxxxxxxxx>
> Reported-by: Lekshmi C. Pillai <lekshmi.cpillai@xxxxxxxxxx>
> Tested-by: Lekshmi C. Pillai <lekshmi.cpillai@xxxxxxxxxx>
> Tested-by: Paul Nguyen <nguyenp@xxxxxxxxxx>
Thanks for your persistence in re-posting this. The fix looks good to
me. Ben, can you queue this up?
Reviewed-by: Jeff Moyer <jmoyer@xxxxxxxxxx>
> ---
> fs/aio.c | 19 ++++++++++++-------
> 1 file changed, 12 insertions(+), 7 deletions(-)
>
> diff --git a/fs/aio.c b/fs/aio.c
> index f52d925ee259..3908480d7ccd 100644
> --- a/fs/aio.c
> +++ b/fs/aio.c
> @@ -441,10 +441,9 @@ static int aio_migratepage(struct address_space *mapping, struct page *new,
> #endif
> };
>
> -static int aio_setup_ring(struct kioctx *ctx)
> +static int aio_setup_ring(struct kioctx *ctx, unsigned int nr_events)
> {
> struct aio_ring *ring;
> - unsigned nr_events = ctx->max_reqs;
> struct mm_struct *mm = current->mm;
> unsigned long size, unused;
> int nr_pages;
> @@ -707,6 +706,12 @@ static struct kioctx *ioctx_alloc(unsigned nr_events)
> int err = -ENOMEM;
>
> /*
> + * Store the original nr_events -- what userspace passed to io_setup(),
> + * for counting against the global limit -- before it changes.
> + */
> + unsigned int max_reqs = nr_events;
> +
> + /*
> * We keep track of the number of available ringbuffer slots, to prevent
> * overflow (reqs_available), and we also use percpu counters for this.
> *
> @@ -724,14 +729,14 @@ static struct kioctx *ioctx_alloc(unsigned nr_events)
> return ERR_PTR(-EINVAL);
> }
>
> - if (!nr_events || (unsigned long)nr_events > (aio_max_nr * 2UL))
> + if (!nr_events || (unsigned long)max_reqs > aio_max_nr)
> return ERR_PTR(-EAGAIN);
>
> ctx = kmem_cache_zalloc(kioctx_cachep, GFP_KERNEL);
> if (!ctx)
> return ERR_PTR(-ENOMEM);
>
> - ctx->max_reqs = nr_events;
> + ctx->max_reqs = max_reqs;
>
> spin_lock_init(&ctx->ctx_lock);
> spin_lock_init(&ctx->completion_lock);
> @@ -753,7 +758,7 @@ static struct kioctx *ioctx_alloc(unsigned nr_events)
> if (!ctx->cpu)
> goto err;
>
> - err = aio_setup_ring(ctx);
> + err = aio_setup_ring(ctx, nr_events);
> if (err < 0)
> goto err;
>
> @@ -764,8 +769,8 @@ static struct kioctx *ioctx_alloc(unsigned nr_events)
>
> /* limit the number of system wide aios */
> spin_lock(&aio_nr_lock);
> - if (aio_nr + nr_events > (aio_max_nr * 2UL) ||
> - aio_nr + nr_events < aio_nr) {
> + if (aio_nr + ctx->max_reqs > aio_max_nr ||
> + aio_nr + ctx->max_reqs < aio_nr) {
> spin_unlock(&aio_nr_lock);
> err = -EAGAIN;
> goto err_ctx;
