| Message ID | 20221206055059.1877471-3-edumazet@google.com (mailing list archive) |
|---|---|
| State | Superseded |
| Delegated to: | Netdev Maintainers |
| Headers | show |
| Series | mlx4: better BIG-TCP support | expand |
On 12/6/2022 7:50 AM, Eric Dumazet wrote: > Google production kernel has increased MAX_SKB_FRAGS to 45 > for BIG-TCP rollout. > > Unfortunately mlx4 TX bounce buffer is not big enough whenever > an skb has up to 45 page fragments. > > This can happen often with TCP TX zero copy, as one frag usually > holds 4096 bytes of payload (order-0 page). > > Tested: > Kernel built with MAX_SKB_FRAGS=45 > ip link set dev eth0 gso_max_size 185000 > netperf -t TCP_SENDFILE > > I made sure that "ethtool -G eth0 tx 64" was properly working, > ring->full_size being set to 16. > > Signed-off-by: Eric Dumazet <edumazet@google.com> > Reported-by: Wei Wang <weiwan@google.com> > Cc: Tariq Toukan <tariqt@nvidia.com> > --- > drivers/net/ethernet/mellanox/mlx4/mlx4_en.h | 16 ++++++++++++---- > 1 file changed, 12 insertions(+), 4 deletions(-) > > diff --git a/drivers/net/ethernet/mellanox/mlx4/mlx4_en.h b/drivers/net/ethernet/mellanox/mlx4/mlx4_en.h > index 7cc288db2a64f75ffe64882e3c25b90715e68855..120b8c361e91d443f83f100a1afabcabc776a92a 100644 > --- a/drivers/net/ethernet/mellanox/mlx4/mlx4_en.h > +++ b/drivers/net/ethernet/mellanox/mlx4/mlx4_en.h > @@ -89,8 +89,18 @@ > #define MLX4_EN_FILTER_HASH_SHIFT 4 > #define MLX4_EN_FILTER_EXPIRY_QUOTA 60 > > -/* Typical TSO descriptor with 16 gather entries is 352 bytes... */ > -#define MLX4_TX_BOUNCE_BUFFER_SIZE 512 > +#define CTRL_SIZE sizeof(struct mlx4_wqe_ctrl_seg) > +#define DS_SIZE sizeof(struct mlx4_wqe_data_seg) > + > +/* Maximal size of the bounce buffer: > + * 256 bytes for LSO headers. > + * CTRL_SIZE for control desc. > + * DS_SIZE if skb->head contains some payload. > + * MAX_SKB_FRAGS frags. > + */ > +#define MLX4_TX_BOUNCE_BUFFER_SIZE (256 + CTRL_SIZE + DS_SIZE + \ > + MAX_SKB_FRAGS * DS_SIZE) > + > #define MLX4_MAX_DESC_TXBBS (MLX4_TX_BOUNCE_BUFFER_SIZE / TXBB_SIZE) > Now as MLX4_TX_BOUNCE_BUFFER_SIZE might not be a multiple of TXBB_SIZE, simple integer division won't work to calculate the max num of TXBBs. Roundup is needed. > /* > @@ -217,9 +227,7 @@ struct mlx4_en_tx_info { > > > #define MLX4_EN_BIT_DESC_OWN 0x80000000 > -#define CTRL_SIZE sizeof(struct mlx4_wqe_ctrl_seg) > #define MLX4_EN_MEMTYPE_PAD 0x100 > -#define DS_SIZE sizeof(struct mlx4_wqe_data_seg) > > > struct mlx4_en_tx_desc {
On Wed, Dec 7, 2022 at 1:40 PM Tariq Toukan <ttoukan.linux@gmail.com> wrote: > > > > On 12/6/2022 7:50 AM, Eric Dumazet wrote: > > Google production kernel has increased MAX_SKB_FRAGS to 45 > > for BIG-TCP rollout. > > > > Unfortunately mlx4 TX bounce buffer is not big enough whenever > > an skb has up to 45 page fragments. > > > > This can happen often with TCP TX zero copy, as one frag usually > > holds 4096 bytes of payload (order-0 page). > > > > Tested: > > Kernel built with MAX_SKB_FRAGS=45 > > ip link set dev eth0 gso_max_size 185000 > > netperf -t TCP_SENDFILE > > > > I made sure that "ethtool -G eth0 tx 64" was properly working, > > ring->full_size being set to 16. > > > > Signed-off-by: Eric Dumazet <edumazet@google.com> > > Reported-by: Wei Wang <weiwan@google.com> > > Cc: Tariq Toukan <tariqt@nvidia.com> > > --- > > drivers/net/ethernet/mellanox/mlx4/mlx4_en.h | 16 ++++++++++++---- > > 1 file changed, 12 insertions(+), 4 deletions(-) > > > > diff --git a/drivers/net/ethernet/mellanox/mlx4/mlx4_en.h b/drivers/net/ethernet/mellanox/mlx4/mlx4_en.h > > index 7cc288db2a64f75ffe64882e3c25b90715e68855..120b8c361e91d443f83f100a1afabcabc776a92a 100644 > > --- a/drivers/net/ethernet/mellanox/mlx4/mlx4_en.h > > +++ b/drivers/net/ethernet/mellanox/mlx4/mlx4_en.h > > @@ -89,8 +89,18 @@ > > #define MLX4_EN_FILTER_HASH_SHIFT 4 > > #define MLX4_EN_FILTER_EXPIRY_QUOTA 60 > > > > -/* Typical TSO descriptor with 16 gather entries is 352 bytes... */ > > -#define MLX4_TX_BOUNCE_BUFFER_SIZE 512 > > +#define CTRL_SIZE sizeof(struct mlx4_wqe_ctrl_seg) > > +#define DS_SIZE sizeof(struct mlx4_wqe_data_seg) > > + > > +/* Maximal size of the bounce buffer: > > + * 256 bytes for LSO headers. > > + * CTRL_SIZE for control desc. > > + * DS_SIZE if skb->head contains some payload. > > + * MAX_SKB_FRAGS frags. > > + */ > > +#define MLX4_TX_BOUNCE_BUFFER_SIZE (256 + CTRL_SIZE + DS_SIZE + \ > > + MAX_SKB_FRAGS * DS_SIZE) > > + > > #define MLX4_MAX_DESC_TXBBS (MLX4_TX_BOUNCE_BUFFER_SIZE / TXBB_SIZE) > > > > Now as MLX4_TX_BOUNCE_BUFFER_SIZE might not be a multiple of TXBB_SIZE, > simple integer division won't work to calculate the max num of TXBBs. > Roundup is needed. I do not see why a roundup is needed. This seems like obfuscation to me. A divide by TXBB_SIZE always "works". A round up is already done in mlx4_en_xmit() /* Align descriptor to TXBB size */ desc_size = ALIGN(real_size, TXBB_SIZE); nr_txbb = desc_size >> LOG_TXBB_SIZE; Then the check is : if (unlikely(nr_txbb > MLX4_MAX_DESC_TXBBS)) { if (netif_msg_tx_err(priv)) en_warn(priv, "Oversized header or SG list\n"); goto tx_drop_count; } If we allocate X extra bytes (in case MLX4_TX_BOUNCE_BUFFER_SIZE % TXBB_SIZE == X), we are not going to use them anyway.
On Wed, Dec 7, 2022 at 1:53 PM Eric Dumazet <edumazet@google.com> wrote: > > On Wed, Dec 7, 2022 at 1:40 PM Tariq Toukan <ttoukan.linux@gmail.com> wrote: > > > > > > > > On 12/6/2022 7:50 AM, Eric Dumazet wrote: > > > Google production kernel has increased MAX_SKB_FRAGS to 45 > > > for BIG-TCP rollout. > > > > > > Unfortunately mlx4 TX bounce buffer is not big enough whenever > > > an skb has up to 45 page fragments. > > > > > > This can happen often with TCP TX zero copy, as one frag usually > > > holds 4096 bytes of payload (order-0 page). > > > > > > Tested: > > > Kernel built with MAX_SKB_FRAGS=45 > > > ip link set dev eth0 gso_max_size 185000 > > > netperf -t TCP_SENDFILE > > > > > > I made sure that "ethtool -G eth0 tx 64" was properly working, > > > ring->full_size being set to 16. > > > > > > Signed-off-by: Eric Dumazet <edumazet@google.com> > > > Reported-by: Wei Wang <weiwan@google.com> > > > Cc: Tariq Toukan <tariqt@nvidia.com> > > > --- > > > drivers/net/ethernet/mellanox/mlx4/mlx4_en.h | 16 ++++++++++++---- > > > 1 file changed, 12 insertions(+), 4 deletions(-) > > > > > > diff --git a/drivers/net/ethernet/mellanox/mlx4/mlx4_en.h b/drivers/net/ethernet/mellanox/mlx4/mlx4_en.h > > > index 7cc288db2a64f75ffe64882e3c25b90715e68855..120b8c361e91d443f83f100a1afabcabc776a92a 100644 > > > --- a/drivers/net/ethernet/mellanox/mlx4/mlx4_en.h > > > +++ b/drivers/net/ethernet/mellanox/mlx4/mlx4_en.h > > > @@ -89,8 +89,18 @@ > > > #define MLX4_EN_FILTER_HASH_SHIFT 4 > > > #define MLX4_EN_FILTER_EXPIRY_QUOTA 60 > > > > > > -/* Typical TSO descriptor with 16 gather entries is 352 bytes... */ > > > -#define MLX4_TX_BOUNCE_BUFFER_SIZE 512 > > > +#define CTRL_SIZE sizeof(struct mlx4_wqe_ctrl_seg) > > > +#define DS_SIZE sizeof(struct mlx4_wqe_data_seg) > > > + > > > +/* Maximal size of the bounce buffer: > > > + * 256 bytes for LSO headers. > > > + * CTRL_SIZE for control desc. > > > + * DS_SIZE if skb->head contains some payload. > > > + * MAX_SKB_FRAGS frags. > > > + */ > > > +#define MLX4_TX_BOUNCE_BUFFER_SIZE (256 + CTRL_SIZE + DS_SIZE + \ > > > + MAX_SKB_FRAGS * DS_SIZE) > > > + > > > #define MLX4_MAX_DESC_TXBBS (MLX4_TX_BOUNCE_BUFFER_SIZE / TXBB_SIZE) > > > > > > > Now as MLX4_TX_BOUNCE_BUFFER_SIZE might not be a multiple of TXBB_SIZE, > > simple integer division won't work to calculate the max num of TXBBs. > > Roundup is needed. > > I do not see why a roundup is needed. This seems like obfuscation to me. > > A divide by TXBB_SIZE always "works". > > A round up is already done in mlx4_en_xmit() > > /* Align descriptor to TXBB size */ > desc_size = ALIGN(real_size, TXBB_SIZE); > nr_txbb = desc_size >> LOG_TXBB_SIZE; > > Then the check is : > > if (unlikely(nr_txbb > MLX4_MAX_DESC_TXBBS)) { > if (netif_msg_tx_err(priv)) > en_warn(priv, "Oversized header or SG list\n"); > goto tx_drop_count; > } > > If we allocate X extra bytes (in case MLX4_TX_BOUNCE_BUFFER_SIZE % > TXBB_SIZE == X), > we are not going to use them anyway. I guess you are worried about not having exactly 256 bytes for the headers ? Currently, the amount of space for headers is 208 bytes. If MAX_SKB_FRAGS is 17, MLX4_TX_BOUNCE_BUFFER_SIZE would be 0x230 after my patch, so the same usable space as before the patch.
On 12/7/2022 3:06 PM, Eric Dumazet wrote: > On Wed, Dec 7, 2022 at 1:53 PM Eric Dumazet <edumazet@google.com> wrote: >> >> On Wed, Dec 7, 2022 at 1:40 PM Tariq Toukan <ttoukan.linux@gmail.com> wrote: >>> >>> >>> >>> On 12/6/2022 7:50 AM, Eric Dumazet wrote: >>>> Google production kernel has increased MAX_SKB_FRAGS to 45 >>>> for BIG-TCP rollout. >>>> >>>> Unfortunately mlx4 TX bounce buffer is not big enough whenever >>>> an skb has up to 45 page fragments. >>>> >>>> This can happen often with TCP TX zero copy, as one frag usually >>>> holds 4096 bytes of payload (order-0 page). >>>> >>>> Tested: >>>> Kernel built with MAX_SKB_FRAGS=45 >>>> ip link set dev eth0 gso_max_size 185000 >>>> netperf -t TCP_SENDFILE >>>> >>>> I made sure that "ethtool -G eth0 tx 64" was properly working, >>>> ring->full_size being set to 16. >>>> >>>> Signed-off-by: Eric Dumazet <edumazet@google.com> >>>> Reported-by: Wei Wang <weiwan@google.com> >>>> Cc: Tariq Toukan <tariqt@nvidia.com> >>>> --- >>>> drivers/net/ethernet/mellanox/mlx4/mlx4_en.h | 16 ++++++++++++---- >>>> 1 file changed, 12 insertions(+), 4 deletions(-) >>>> >>>> diff --git a/drivers/net/ethernet/mellanox/mlx4/mlx4_en.h b/drivers/net/ethernet/mellanox/mlx4/mlx4_en.h >>>> index 7cc288db2a64f75ffe64882e3c25b90715e68855..120b8c361e91d443f83f100a1afabcabc776a92a 100644 >>>> --- a/drivers/net/ethernet/mellanox/mlx4/mlx4_en.h >>>> +++ b/drivers/net/ethernet/mellanox/mlx4/mlx4_en.h >>>> @@ -89,8 +89,18 @@ >>>> #define MLX4_EN_FILTER_HASH_SHIFT 4 >>>> #define MLX4_EN_FILTER_EXPIRY_QUOTA 60 >>>> >>>> -/* Typical TSO descriptor with 16 gather entries is 352 bytes... */ >>>> -#define MLX4_TX_BOUNCE_BUFFER_SIZE 512 >>>> +#define CTRL_SIZE sizeof(struct mlx4_wqe_ctrl_seg) >>>> +#define DS_SIZE sizeof(struct mlx4_wqe_data_seg) >>>> + >>>> +/* Maximal size of the bounce buffer: >>>> + * 256 bytes for LSO headers. >>>> + * CTRL_SIZE for control desc. >>>> + * DS_SIZE if skb->head contains some payload. >>>> + * MAX_SKB_FRAGS frags. >>>> + */ >>>> +#define MLX4_TX_BOUNCE_BUFFER_SIZE (256 + CTRL_SIZE + DS_SIZE + \ >>>> + MAX_SKB_FRAGS * DS_SIZE) >>>> + >>>> #define MLX4_MAX_DESC_TXBBS (MLX4_TX_BOUNCE_BUFFER_SIZE / TXBB_SIZE) >>>> >>> >>> Now as MLX4_TX_BOUNCE_BUFFER_SIZE might not be a multiple of TXBB_SIZE, >>> simple integer division won't work to calculate the max num of TXBBs. >>> Roundup is needed. >> >> I do not see why a roundup is needed. This seems like obfuscation to me. >> >> A divide by TXBB_SIZE always "works". >> >> A round up is already done in mlx4_en_xmit() >> >> /* Align descriptor to TXBB size */ >> desc_size = ALIGN(real_size, TXBB_SIZE); >> nr_txbb = desc_size >> LOG_TXBB_SIZE; >> >> Then the check is : >> >> if (unlikely(nr_txbb > MLX4_MAX_DESC_TXBBS)) { >> if (netif_msg_tx_err(priv)) >> en_warn(priv, "Oversized header or SG list\n"); >> goto tx_drop_count; >> } >> >> If we allocate X extra bytes (in case MLX4_TX_BOUNCE_BUFFER_SIZE % >> TXBB_SIZE == X), >> we are not going to use them anyway. Now the MLX4_MAX_DESC_TXBBS gives a stricter limit than the allocated size MLX4_TX_BOUNCE_BUFFER_SIZE. > > I guess you are worried about not having exactly 256 bytes for the headers ? > > Currently, the amount of space for headers is 208 bytes. > > If MAX_SKB_FRAGS is 17, MLX4_TX_BOUNCE_BUFFER_SIZE would be 0x230 > after my patch, > so the same usable space as before the patch. So what you're saying is, if all the elements of MLX4_TX_BOUNCE_BUFFER_SIZE co-exist together for a TX descriptor, then the actual "headers" part can go only up to 208 (similar to today), not the whole 256 (as the new define documentation says). This keeps the current behavior, but makes the code a bit more confusing. IMO it is cleaner to have MLX4_TX_BOUNCE_BUFFER_SIZE explicitly defined as a multiple of TXBB_SIZE in the first place. This way, both the allocation size and the desc size limit will be in perfect sync, without having assumptions on the amount X lost in the division. How about the below, to keep today's values for the defines? #define MLX4_TX_BOUNCE_BUFFER_SIZE \ ALIGN(208 + CTRL_SIZE + DS_SIZE + \ MAX_SKB_FRAGS * DS_SIZE, TXBB_SIZE)
On Wed, Dec 7, 2022 at 4:14 PM Tariq Toukan <ttoukan.linux@gmail.com> wrote: > > So what you're saying is, if all the elements of > MLX4_TX_BOUNCE_BUFFER_SIZE co-exist together for a TX descriptor, then > the actual "headers" part can go only up to 208 (similar to today), not > the whole 256 (as the new define documentation says). > > This keeps the current behavior, but makes the code a bit more confusing. > > IMO it is cleaner to have MLX4_TX_BOUNCE_BUFFER_SIZE explicitly defined > as a multiple of TXBB_SIZE in the first place. This way, both the > allocation size and the desc size limit will be in perfect sync, without > having assumptions on the amount X lost in the division. > > How about the below, to keep today's values for the defines? > > #define MLX4_TX_BOUNCE_BUFFER_SIZE \ > ALIGN(208 + CTRL_SIZE + DS_SIZE + \ > MAX_SKB_FRAGS * DS_SIZE, TXBB_SIZE) I already sent a v2, with: +#define MLX4_TX_BOUNCE_BUFFER_SIZE \ + ALIGN(256 + CTRL_SIZE + DS_SIZE + MAX_SKB_FRAGS * DS_SIZE, TXBB_SIZE) + Please take a look, thanks.
diff --git a/drivers/net/ethernet/mellanox/mlx4/mlx4_en.h b/drivers/net/ethernet/mellanox/mlx4/mlx4_en.h index 7cc288db2a64f75ffe64882e3c25b90715e68855..120b8c361e91d443f83f100a1afabcabc776a92a 100644 --- a/drivers/net/ethernet/mellanox/mlx4/mlx4_en.h +++ b/drivers/net/ethernet/mellanox/mlx4/mlx4_en.h @@ -89,8 +89,18 @@ #define MLX4_EN_FILTER_HASH_SHIFT 4 #define MLX4_EN_FILTER_EXPIRY_QUOTA 60 -/* Typical TSO descriptor with 16 gather entries is 352 bytes... */ -#define MLX4_TX_BOUNCE_BUFFER_SIZE 512 +#define CTRL_SIZE sizeof(struct mlx4_wqe_ctrl_seg) +#define DS_SIZE sizeof(struct mlx4_wqe_data_seg) + +/* Maximal size of the bounce buffer: + * 256 bytes for LSO headers. + * CTRL_SIZE for control desc. + * DS_SIZE if skb->head contains some payload. + * MAX_SKB_FRAGS frags. + */ +#define MLX4_TX_BOUNCE_BUFFER_SIZE (256 + CTRL_SIZE + DS_SIZE + \ + MAX_SKB_FRAGS * DS_SIZE) + #define MLX4_MAX_DESC_TXBBS (MLX4_TX_BOUNCE_BUFFER_SIZE / TXBB_SIZE) /* @@ -217,9 +227,7 @@ struct mlx4_en_tx_info { #define MLX4_EN_BIT_DESC_OWN 0x80000000 -#define CTRL_SIZE sizeof(struct mlx4_wqe_ctrl_seg) #define MLX4_EN_MEMTYPE_PAD 0x100 -#define DS_SIZE sizeof(struct mlx4_wqe_data_seg) struct mlx4_en_tx_desc {
Google production kernel has increased MAX_SKB_FRAGS to 45 for BIG-TCP rollout. Unfortunately mlx4 TX bounce buffer is not big enough whenever an skb has up to 45 page fragments. This can happen often with TCP TX zero copy, as one frag usually holds 4096 bytes of payload (order-0 page). Tested: Kernel built with MAX_SKB_FRAGS=45 ip link set dev eth0 gso_max_size 185000 netperf -t TCP_SENDFILE I made sure that "ethtool -G eth0 tx 64" was properly working, ring->full_size being set to 16. Signed-off-by: Eric Dumazet <edumazet@google.com> Reported-by: Wei Wang <weiwan@google.com> Cc: Tariq Toukan <tariqt@nvidia.com> --- drivers/net/ethernet/mellanox/mlx4/mlx4_en.h | 16 ++++++++++++---- 1 file changed, 12 insertions(+), 4 deletions(-)