mempool: fix slow allocation of large mempools

Message ID 20200109132742.15828-1-olivier.matz@6wind.com (mailing list archive)
State Superseded, archived
Delegated to: David Marchand
Headers
Series mempool: fix slow allocation of large mempools |

Checks

Context Check Description
ci/checkpatch success coding style OK
ci/iol-intel-Performance fail Performance Testing issues
ci/iol-testing fail Testing issues
ci/iol-mellanox-Performance fail Performance Testing issues
ci/iol-nxp-Performance fail Performance Testing issues
ci/travis-robot warning Travis build: failed
ci/Intel-compilation success Compilation OK

Commit Message

Olivier Matz Jan. 9, 2020, 1:27 p.m. UTC
When allocating a mempool which is larger than the largest
available area, it can take a lot of time:

a- the mempool calculate the required memory size, and tries
   to allocate it, it fails
b- then it tries to allocate the largest available area (this
   does not request new huge pages)
c- add this zone to the mempool, this triggers the allocation
   of a mem hdr, which request a new huge page
d- back to a- until mempool is populated or until there is no
   more memory

This can take a lot of time to finally fail (several minutes): in step
a- it takes all available hugepages on the system, then release them
after it fails.

The problem appeared with commit eba11e364614 ("mempool: reduce wasted
space on populate"), because smaller chunks are now allowed. Previously,
it had to be at least one page size, which is not the case in step b-.

To fix this, implement our own way to allocate the largest available
area instead of using the feature from memzone: if an allocation fails,
try to divide the size by 2 and retry. When the requested size falls
below min_chunk_size, stop and return an error.

Fixes: eba11e364614 ("mempool: reduce wasted space on populate")
Cc: stable@dpdk.org

Signed-off-by: Olivier Matz <olivier.matz@6wind.com>
---
 lib/librte_mempool/rte_mempool.c | 29 ++++++++++++-----------------
 1 file changed, 12 insertions(+), 17 deletions(-)
  

Comments

Burakov, Anatoly Jan. 9, 2020, 1:57 p.m. UTC | #1
On 09-Jan-20 1:27 PM, Olivier Matz wrote:
> When allocating a mempool which is larger than the largest
> available area, it can take a lot of time:
> 
> a- the mempool calculate the required memory size, and tries
>     to allocate it, it fails
> b- then it tries to allocate the largest available area (this
>     does not request new huge pages)
> c- add this zone to the mempool, this triggers the allocation
>     of a mem hdr, which request a new huge page
> d- back to a- until mempool is populated or until there is no
>     more memory
> 
> This can take a lot of time to finally fail (several minutes): in step
> a- it takes all available hugepages on the system, then release them
> after it fails.
> 
> The problem appeared with commit eba11e364614 ("mempool: reduce wasted
> space on populate"), because smaller chunks are now allowed. Previously,
> it had to be at least one page size, which is not the case in step b-.
> 
> To fix this, implement our own way to allocate the largest available
> area instead of using the feature from memzone: if an allocation fails,
> try to divide the size by 2 and retry. When the requested size falls
> below min_chunk_size, stop and return an error.
> 
> Fixes: eba11e364614 ("mempool: reduce wasted space on populate")
> Cc: stable@dpdk.org
> 
> Signed-off-by: Olivier Matz <olivier.matz@6wind.com>
> ---

I don't particularly like the idea of working around this issue as 
opposed to fixing it memzone-side, but since there's currently no plan 
to address this in memzone allocator, this should work much better than 
before.

Acked-by: Anatoly Burakov <anatoly.burakov@intel.com>
  
Ali Alnubani Jan. 9, 2020, 4:06 p.m. UTC | #2
Hi Olivier,

> -----Original Message-----
> From: dev <dev-bounces@dpdk.org> On Behalf Of Olivier Matz
> Sent: Thursday, January 9, 2020 3:28 PM
> To: dev@dpdk.org
> Cc: Andrew Rybchenko <arybchenko@solarflare.com>; Anatoly Burakov
> <anatoly.burakov@intel.com>; stable@dpdk.org
> Subject: [dpdk-dev] [PATCH] mempool: fix slow allocation of large mempools
> 
> When allocating a mempool which is larger than the largest available area, it
> can take a lot of time:
> 
> a- the mempool calculate the required memory size, and tries
>    to allocate it, it fails
> b- then it tries to allocate the largest available area (this
>    does not request new huge pages)
> c- add this zone to the mempool, this triggers the allocation
>    of a mem hdr, which request a new huge page
> d- back to a- until mempool is populated or until there is no
>    more memory
> 
> This can take a lot of time to finally fail (several minutes): in step
> a- it takes all available hugepages on the system, then release them after it
> fails.
> 
> The problem appeared with commit eba11e364614 ("mempool: reduce
> wasted space on populate"), because smaller chunks are now allowed.
> Previously, it had to be at least one page size, which is not the case in step b-.
> 
> To fix this, implement our own way to allocate the largest available area
> instead of using the feature from memzone: if an allocation fails, try to divide
> the size by 2 and retry. When the requested size falls below min_chunk_size,
> stop and return an error.
> 
> Fixes: eba11e364614 ("mempool: reduce wasted space on populate")
> Cc: stable@dpdk.org
> 
> Signed-off-by: Olivier Matz <olivier.matz@6wind.com>
> ---

Testpmd (testpmd -n4 -- -i) fails to start after applying this patch with:
"""
EAL: Error - exiting with code: 1
  Cause: Creation of mbuf pool for socket 0 failed: File exists
"""

This is why the check ci/iol-mellanox-Performance is failing (not sure if the other tests are failing for the same reason).

Regards,
Ali
  
Olivier Matz Jan. 9, 2020, 5:27 p.m. UTC | #3
Hi Ali,

On Thu, Jan 09, 2020 at 04:06:53PM +0000, Ali Alnubani wrote:
> Hi Olivier,
> 
> > -----Original Message-----
> > From: dev <dev-bounces@dpdk.org> On Behalf Of Olivier Matz
> > Sent: Thursday, January 9, 2020 3:28 PM
> > To: dev@dpdk.org
> > Cc: Andrew Rybchenko <arybchenko@solarflare.com>; Anatoly Burakov
> > <anatoly.burakov@intel.com>; stable@dpdk.org
> > Subject: [dpdk-dev] [PATCH] mempool: fix slow allocation of large mempools
> > 
> > When allocating a mempool which is larger than the largest available area, it
> > can take a lot of time:
> > 
> > a- the mempool calculate the required memory size, and tries
> >    to allocate it, it fails
> > b- then it tries to allocate the largest available area (this
> >    does not request new huge pages)
> > c- add this zone to the mempool, this triggers the allocation
> >    of a mem hdr, which request a new huge page
> > d- back to a- until mempool is populated or until there is no
> >    more memory
> > 
> > This can take a lot of time to finally fail (several minutes): in step
> > a- it takes all available hugepages on the system, then release them after it
> > fails.
> > 
> > The problem appeared with commit eba11e364614 ("mempool: reduce
> > wasted space on populate"), because smaller chunks are now allowed.
> > Previously, it had to be at least one page size, which is not the case in step b-.
> > 
> > To fix this, implement our own way to allocate the largest available area
> > instead of using the feature from memzone: if an allocation fails, try to divide
> > the size by 2 and retry. When the requested size falls below min_chunk_size,
> > stop and return an error.
> > 
> > Fixes: eba11e364614 ("mempool: reduce wasted space on populate")
> > Cc: stable@dpdk.org
> > 
> > Signed-off-by: Olivier Matz <olivier.matz@6wind.com>
> > ---
> 
> Testpmd (testpmd -n4 -- -i) fails to start after applying this patch with:
> """
> EAL: Error - exiting with code: 1
>   Cause: Creation of mbuf pool for socket 0 failed: File exists
> """
> 
> This is why the check ci/iol-mellanox-Performance is failing (not sure if the other tests are failing for the same reason).

Thanks for the report.

I should have retested after my "little rework"... :)
I'll send a v2 with this fix:

  --- a/lib/librte_mempool/rte_mempool.c
  +++ b/lib/librte_mempool/rte_mempool.c
  @@ -572,7 +572,7 @@ rte_mempool_populate_default(struct rte_mempool *mp)
   
                          max_alloc_size = RTE_MIN(max_alloc_size,
                                                  (size_t)mem_size) / 2;
  -               } while (max_alloc_size >= min_chunk_size);
  +               } while (mz == NULL && max_alloc_size >= min_chunk_size);
   
                  if (mz == NULL) {
                          ret = -rte_errno;

Olivier
  
Andrew Rybchenko Jan. 10, 2020, 9:53 a.m. UTC | #4
On 1/9/20 4:27 PM, Olivier Matz wrote:
> When allocating a mempool which is larger than the largest
> available area, it can take a lot of time:
> 
> a- the mempool calculate the required memory size, and tries
>    to allocate it, it fails
> b- then it tries to allocate the largest available area (this
>    does not request new huge pages)
> c- add this zone to the mempool, this triggers the allocation
>    of a mem hdr, which request a new huge page
> d- back to a- until mempool is populated or until there is no
>    more memory
> 
> This can take a lot of time to finally fail (several minutes): in step
> a- it takes all available hugepages on the system, then release them
> after it fails.
> 
> The problem appeared with commit eba11e364614 ("mempool: reduce wasted
> space on populate"), because smaller chunks are now allowed. Previously,
> it had to be at least one page size, which is not the case in step b-.
> 
> To fix this, implement our own way to allocate the largest available
> area instead of using the feature from memzone: if an allocation fails,
> try to divide the size by 2 and retry. When the requested size falls
> below min_chunk_size, stop and return an error.
> 
> Fixes: eba11e364614 ("mempool: reduce wasted space on populate")
> Cc: stable@dpdk.org
> 
> Signed-off-by: Olivier Matz <olivier.matz@6wind.com>

LGTM except already mentioned bug with missing mz == NULL to retry loop.
Plus one minor question below.

> ---
>  lib/librte_mempool/rte_mempool.c | 29 ++++++++++++-----------------
>  1 file changed, 12 insertions(+), 17 deletions(-)
> 
> diff --git a/lib/librte_mempool/rte_mempool.c b/lib/librte_mempool/rte_mempool.c
> index bda361ce6..03c8d984c 100644
> --- a/lib/librte_mempool/rte_mempool.c
> +++ b/lib/librte_mempool/rte_mempool.c
> @@ -481,6 +481,7 @@ rte_mempool_populate_default(struct rte_mempool *mp)
>  	unsigned mz_id, n;
>  	int ret;
>  	bool need_iova_contig_obj;
> +	size_t max_alloc_size = SIZE_MAX;
>  
>  	ret = mempool_ops_alloc_once(mp);
>  	if (ret != 0)
> @@ -560,30 +561,24 @@ rte_mempool_populate_default(struct rte_mempool *mp)
>  		if (min_chunk_size == (size_t)mem_size)
>  			mz_flags |= RTE_MEMZONE_IOVA_CONTIG;
>  
> -		mz = rte_memzone_reserve_aligned(mz_name, mem_size,
> +		/* Allocate a memzone, retrying with a smaller area on ENOMEM */
> +		do {
> +			mz = rte_memzone_reserve_aligned(mz_name,
> +				RTE_MIN((size_t)mem_size, max_alloc_size),
>  				mp->socket_id, mz_flags, align);
>  
> -		/* don't try reserving with 0 size if we were asked to reserve
> -		 * IOVA-contiguous memory.
> -		 */
> -		if (min_chunk_size < (size_t)mem_size && mz == NULL) {
> -			/* not enough memory, retry with the biggest zone we
> -			 * have
> -			 */
> -			mz = rte_memzone_reserve_aligned(mz_name, 0,
> -					mp->socket_id, mz_flags, align);
> -		}
> +			if (mz == NULL && rte_errno != ENOMEM)
> +				break;
> +
> +			max_alloc_size = RTE_MIN(max_alloc_size,
> +						(size_t)mem_size) / 2;

Does it make sense to make max_alloc_size multiple of
min_chunk_size here? I think it could help to waste less
memory space.

> +		} while (max_alloc_size >= min_chunk_size);
> +
>  		if (mz == NULL) {
>  			ret = -rte_errno;
>  			goto fail;
>  		}
>  
> -		if (mz->len < min_chunk_size) {
> -			rte_memzone_free(mz);
> -			ret = -ENOMEM;
> -			goto fail;
> -		}
> -
>  		if (need_iova_contig_obj)
>  			iova = mz->iova;
>  		else
>
  
Olivier Matz Jan. 17, 2020, 8:45 a.m. UTC | #5
Hi,

On Fri, Jan 10, 2020 at 12:53:24PM +0300, Andrew Rybchenko wrote:
> On 1/9/20 4:27 PM, Olivier Matz wrote:
> > When allocating a mempool which is larger than the largest
> > available area, it can take a lot of time:
> > 
> > a- the mempool calculate the required memory size, and tries
> >    to allocate it, it fails
> > b- then it tries to allocate the largest available area (this
> >    does not request new huge pages)
> > c- add this zone to the mempool, this triggers the allocation
> >    of a mem hdr, which request a new huge page
> > d- back to a- until mempool is populated or until there is no
> >    more memory
> > 
> > This can take a lot of time to finally fail (several minutes): in step
> > a- it takes all available hugepages on the system, then release them
> > after it fails.
> > 
> > The problem appeared with commit eba11e364614 ("mempool: reduce wasted
> > space on populate"), because smaller chunks are now allowed. Previously,
> > it had to be at least one page size, which is not the case in step b-.
> > 
> > To fix this, implement our own way to allocate the largest available
> > area instead of using the feature from memzone: if an allocation fails,
> > try to divide the size by 2 and retry. When the requested size falls
> > below min_chunk_size, stop and return an error.
> > 
> > Fixes: eba11e364614 ("mempool: reduce wasted space on populate")
> > Cc: stable@dpdk.org
> > 
> > Signed-off-by: Olivier Matz <olivier.matz@6wind.com>
> 
> LGTM except already mentioned bug with missing mz == NULL to retry loop.
> Plus one minor question below.
> 
> > ---
> >  lib/librte_mempool/rte_mempool.c | 29 ++++++++++++-----------------
> >  1 file changed, 12 insertions(+), 17 deletions(-)
> > 
> > diff --git a/lib/librte_mempool/rte_mempool.c b/lib/librte_mempool/rte_mempool.c
> > index bda361ce6..03c8d984c 100644
> > --- a/lib/librte_mempool/rte_mempool.c
> > +++ b/lib/librte_mempool/rte_mempool.c
> > @@ -481,6 +481,7 @@ rte_mempool_populate_default(struct rte_mempool *mp)
> >  	unsigned mz_id, n;
> >  	int ret;
> >  	bool need_iova_contig_obj;
> > +	size_t max_alloc_size = SIZE_MAX;
> >  
> >  	ret = mempool_ops_alloc_once(mp);
> >  	if (ret != 0)
> > @@ -560,30 +561,24 @@ rte_mempool_populate_default(struct rte_mempool *mp)
> >  		if (min_chunk_size == (size_t)mem_size)
> >  			mz_flags |= RTE_MEMZONE_IOVA_CONTIG;
> >  
> > -		mz = rte_memzone_reserve_aligned(mz_name, mem_size,
> > +		/* Allocate a memzone, retrying with a smaller area on ENOMEM */
> > +		do {
> > +			mz = rte_memzone_reserve_aligned(mz_name,
> > +				RTE_MIN((size_t)mem_size, max_alloc_size),
> >  				mp->socket_id, mz_flags, align);
> >  
> > -		/* don't try reserving with 0 size if we were asked to reserve
> > -		 * IOVA-contiguous memory.
> > -		 */
> > -		if (min_chunk_size < (size_t)mem_size && mz == NULL) {
> > -			/* not enough memory, retry with the biggest zone we
> > -			 * have
> > -			 */
> > -			mz = rte_memzone_reserve_aligned(mz_name, 0,
> > -					mp->socket_id, mz_flags, align);
> > -		}
> > +			if (mz == NULL && rte_errno != ENOMEM)
> > +				break;
> > +
> > +			max_alloc_size = RTE_MIN(max_alloc_size,
> > +						(size_t)mem_size) / 2;
> 
> Does it make sense to make max_alloc_size multiple of
> min_chunk_size here? I think it could help to waste less
> memory space.

I don't think it's worth doing it: I agree it could avoid to waste
space, but it is only significant if max_alloc_size is in the same order
of size than min_chunk_size. And this would only happen when we are
running out of memory.

Also, as populate_virt() will skip page boundaries, keeping
a multiple of min_chunk_size may not make sense in that case.


> 
> > +		} while (max_alloc_size >= min_chunk_size);
> > +
> >  		if (mz == NULL) {
> >  			ret = -rte_errno;
> >  			goto fail;
> >  		}
> >  
> > -		if (mz->len < min_chunk_size) {
> > -			rte_memzone_free(mz);
> > -			ret = -ENOMEM;
> > -			goto fail;
> > -		}
> > -
> >  		if (need_iova_contig_obj)
> >  			iova = mz->iova;
> >  		else
> > 
>
  

Patch

diff --git a/lib/librte_mempool/rte_mempool.c b/lib/librte_mempool/rte_mempool.c
index bda361ce6..03c8d984c 100644
--- a/lib/librte_mempool/rte_mempool.c
+++ b/lib/librte_mempool/rte_mempool.c
@@ -481,6 +481,7 @@  rte_mempool_populate_default(struct rte_mempool *mp)
 	unsigned mz_id, n;
 	int ret;
 	bool need_iova_contig_obj;
+	size_t max_alloc_size = SIZE_MAX;
 
 	ret = mempool_ops_alloc_once(mp);
 	if (ret != 0)
@@ -560,30 +561,24 @@  rte_mempool_populate_default(struct rte_mempool *mp)
 		if (min_chunk_size == (size_t)mem_size)
 			mz_flags |= RTE_MEMZONE_IOVA_CONTIG;
 
-		mz = rte_memzone_reserve_aligned(mz_name, mem_size,
+		/* Allocate a memzone, retrying with a smaller area on ENOMEM */
+		do {
+			mz = rte_memzone_reserve_aligned(mz_name,
+				RTE_MIN((size_t)mem_size, max_alloc_size),
 				mp->socket_id, mz_flags, align);
 
-		/* don't try reserving with 0 size if we were asked to reserve
-		 * IOVA-contiguous memory.
-		 */
-		if (min_chunk_size < (size_t)mem_size && mz == NULL) {
-			/* not enough memory, retry with the biggest zone we
-			 * have
-			 */
-			mz = rte_memzone_reserve_aligned(mz_name, 0,
-					mp->socket_id, mz_flags, align);
-		}
+			if (mz == NULL && rte_errno != ENOMEM)
+				break;
+
+			max_alloc_size = RTE_MIN(max_alloc_size,
+						(size_t)mem_size) / 2;
+		} while (max_alloc_size >= min_chunk_size);
+
 		if (mz == NULL) {
 			ret = -rte_errno;
 			goto fail;
 		}
 
-		if (mz->len < min_chunk_size) {
-			rte_memzone_free(mz);
-			ret = -ENOMEM;
-			goto fail;
-		}
-
 		if (need_iova_contig_obj)
 			iova = mz->iova;
 		else