Message ID | 20201207210649.19194-3-borisp@mellanox.com |
---|---|
State | Superseded |
Headers | show |
Series | nvme-tcp receive offloads | expand |
On 12/7/20 2:06 PM, Boris Pismenny wrote: > This commit introduces direct data placement offload for TCP. > This capability is accompanied by new net_device operations that > configure > hardware contexts. There is a context per socket, and a context per DDP > opreation. Additionally, a resynchronization routine is used to assist > hardware handle TCP OOO, and continue the offload. > Furthermore, we let the offloading driver advertise what is the max hw > sectors/segments. > > Using this interface, the NIC hardware will scatter TCP payload directly > to the BIO pages according to the command_id. > To maintain the correctness of the network stack, the driver is expected > to construct SKBs that point to the BIO pages. > > This, the SKB represents the data on the wire, while it is pointing > to data that is already placed in the destination buffer. > As a result, data from page frags should not be copied out to > the linear part. > > As SKBs that use DDP are already very memory efficient, we modify > skb_condence to avoid copying data from fragments to the linear > part of SKBs that belong to a socket that uses DDP offload. > > A follow-up patch will use this interface for DDP in NVMe-TCP. > You call this Direct Data Placement - which sounds like a marketing name. Fundamentally, this starts with offloading TCP socket buffers for a specific flow, so generically a TCP Rx zerocopy for kernel stack managed sockets (as opposed to AF_XDP's zerocopy). Why is this not building in that level of infrastructure first and adding ULPs like NVME on top?
On 08/12/2020 2:42, David Ahern wrote: > On 12/7/20 2:06 PM, Boris Pismenny wrote: >> This commit introduces direct data placement offload for TCP. >> This capability is accompanied by new net_device operations that >> configure >> hardware contexts. There is a context per socket, and a context per DDP >> opreation. Additionally, a resynchronization routine is used to assist >> hardware handle TCP OOO, and continue the offload. >> Furthermore, we let the offloading driver advertise what is the max hw >> sectors/segments. >> >> Using this interface, the NIC hardware will scatter TCP payload directly >> to the BIO pages according to the command_id. >> To maintain the correctness of the network stack, the driver is expected >> to construct SKBs that point to the BIO pages. >> >> This, the SKB represents the data on the wire, while it is pointing >> to data that is already placed in the destination buffer. >> As a result, data from page frags should not be copied out to >> the linear part. >> >> As SKBs that use DDP are already very memory efficient, we modify >> skb_condence to avoid copying data from fragments to the linear >> part of SKBs that belong to a socket that uses DDP offload. >> >> A follow-up patch will use this interface for DDP in NVMe-TCP. >> > > You call this Direct Data Placement - which sounds like a marketing name. > [Re-sending as the previous one didn't hit the mailing list. Sorry for the spam] Interesting idea. But, unlike SKBTX_DEV_ZEROCOPY this SKB can be inspected/modified by the stack without the need to copy things out. Additionally, the SKB may contain both data that is already placed in its final destination buffer (PDU data) and data that isn't (PDU header); it doesn't matter. Therefore, labeling the entire SKB as zerocopy doesn't convey the desired information. Moreover, skipping copies in the stack to receive zerocopy SKBs will require more invasive changes. Our goal in this approach was to provide the smallest change that enables the desired functionality while preserving the performance of existing flows that do not care for it. An alternative approach, that doesn't affect existing flows at all, which we considered was to make a special version of memcpy_to_page to be used by DDP providers (nvme-tcp). This alternative will require creating corresponding special versions for users of this function such skb_copy_datagram_iter. Thit is more invasive, thus in this patchset we decided to avoid it. > Fundamentally, this starts with offloading TCP socket buffers for a > specific flow, so generically a TCP Rx zerocopy for kernel stack managed > sockets (as opposed to AF_XDP's zerocopy). Why is this not building in > that level of infrastructure first and adding ULPs like NVME on top? > We aren't using AF_XDP or any of the Rx zerocopy infrastructure, because it is unsuitable for data placement for nvme-tcp, which reordes responses relatively to requests for efficiency and requires that data reside in specific destination buffers.
On 12/8/20 7:36 AM, Boris Pismenny wrote: > On 08/12/2020 2:42, David Ahern wrote: >> On 12/7/20 2:06 PM, Boris Pismenny wrote: >>> This commit introduces direct data placement offload for TCP. >>> This capability is accompanied by new net_device operations that >>> configure >>> hardware contexts. There is a context per socket, and a context per DDP >>> opreation. Additionally, a resynchronization routine is used to assist >>> hardware handle TCP OOO, and continue the offload. >>> Furthermore, we let the offloading driver advertise what is the max hw >>> sectors/segments. >>> >>> Using this interface, the NIC hardware will scatter TCP payload directly >>> to the BIO pages according to the command_id. >>> To maintain the correctness of the network stack, the driver is expected >>> to construct SKBs that point to the BIO pages. >>> >>> This, the SKB represents the data on the wire, while it is pointing >>> to data that is already placed in the destination buffer. >>> As a result, data from page frags should not be copied out to >>> the linear part. >>> >>> As SKBs that use DDP are already very memory efficient, we modify >>> skb_condence to avoid copying data from fragments to the linear >>> part of SKBs that belong to a socket that uses DDP offload. >>> >>> A follow-up patch will use this interface for DDP in NVMe-TCP. >>> >> >> You call this Direct Data Placement - which sounds like a marketing name. >> > > [Re-sending as the previous one didn't hit the mailing list. Sorry for the spam] > > Interesting idea. But, unlike SKBTX_DEV_ZEROCOPY this SKB can be inspected/modified by the stack without the need to copy things out. Additionally, the SKB may contain both data that is already placed in its final destination buffer (PDU data) and data that isn't (PDU header); it doesn't matter. Therefore, labeling the entire SKB as zerocopy doesn't convey the desired information. Moreover, skipping copies in the stack to receive zerocopy SKBs will require more invasive changes. > > Our goal in this approach was to provide the smallest change that enables the desired functionality while preserving the performance of existing flows that do not care for it. An alternative approach, that doesn't affect existing flows at all, which we considered was to make a special version of memcpy_to_page to be used by DDP providers (nvme-tcp). This alternative will require creating corresponding special versions for users of this function such skb_copy_datagram_iter. Thit is more invasive, thus in this patchset we decided to avoid it. > >> Fundamentally, this starts with offloading TCP socket buffers for a >> specific flow, so generically a TCP Rx zerocopy for kernel stack managed >> sockets (as opposed to AF_XDP's zerocopy). Why is this not building in >> that level of infrastructure first and adding ULPs like NVME on top? >> > > We aren't using AF_XDP or any of the Rx zerocopy infrastructure, because it is unsuitable for data placement for nvme-tcp, which reordes responses relatively to requests for efficiency and requires that data reside in specific destination buffers. > > The AF_XDP reference was to differentiate one zerocopy use case (all packets go to userspace) from another (kernel managed TCP socket with zerocopy payload). You are focusing on a very narrow use case - kernel based NVMe over TCP - of a more general problem. You have a TCP socket and a design that only works for kernel owned sockets. You have specialized queues in the NIC, a flow rule directing packets to those queues. Presumably some ULP parser in the NIC associated with the queues to process NVMe packets. Rather than copying headers (ethernet/ip/tcp) to one buffer and payload to another (which is similar to what Jonathan Lemon is working on), this design has a ULP processor that just splits out the TCP payload even more making it highly selective about which part of the packet is put into which buffer. Take out the NVMe part, and it is header split with zerocopy for the payload - a generic feature that can have a wider impact with NVMe as a special case.
On 12/7/20 2:06 PM, Boris Pismenny wrote: > diff --git a/include/linux/netdev_features.h b/include/linux/netdev_features.h > index 934de56644e7..fb35dcac03d2 100644 > --- a/include/linux/netdev_features.h > +++ b/include/linux/netdev_features.h > @@ -84,6 +84,7 @@ enum { > NETIF_F_GRO_FRAGLIST_BIT, /* Fraglist GRO */ > > NETIF_F_HW_MACSEC_BIT, /* Offload MACsec operations */ > + NETIF_F_HW_TCP_DDP_BIT, /* TCP direct data placement offload */ > > /* > * Add your fresh new feature above and remember to update > @@ -157,6 +158,7 @@ enum { > #define NETIF_F_GRO_FRAGLIST __NETIF_F(GRO_FRAGLIST) > #define NETIF_F_GSO_FRAGLIST __NETIF_F(GSO_FRAGLIST) > #define NETIF_F_HW_MACSEC __NETIF_F(HW_MACSEC) > +#define NETIF_F_HW_TCP_DDP __NETIF_F(HW_TCP_DDP) All of the DDP naming seems wrong to me. I realize the specific use case is targeted payloads of a ULP, but it is still S/W handing H/W specific buffers for a payload of a flow. > > /* Finds the next feature with the highest number of the range of start till 0. > */ > diff --git a/include/linux/netdevice.h b/include/linux/netdevice.h > index a07c8e431f45..755766976408 100644 > --- a/include/linux/netdevice.h > +++ b/include/linux/netdevice.h > @@ -934,6 +934,7 @@ struct dev_ifalias { > > struct devlink; > struct tlsdev_ops; > +struct tcp_ddp_dev_ops; > > struct netdev_name_node { > struct hlist_node hlist; > @@ -1930,6 +1931,10 @@ struct net_device { > const struct tlsdev_ops *tlsdev_ops; > #endif > > +#ifdef CONFIG_TCP_DDP > + const struct tcp_ddp_dev_ops *tcp_ddp_ops; > +#endif > + > const struct header_ops *header_ops; > > unsigned int flags; > diff --git a/include/net/inet_connection_sock.h b/include/net/inet_connection_sock.h > index 7338b3865a2a..a08b85b53aa8 100644 > --- a/include/net/inet_connection_sock.h > +++ b/include/net/inet_connection_sock.h > @@ -66,6 +66,8 @@ struct inet_connection_sock_af_ops { > * @icsk_ulp_ops Pluggable ULP control hook > * @icsk_ulp_data ULP private data > * @icsk_clean_acked Clean acked data hook > + * @icsk_ulp_ddp_ops Pluggable ULP direct data placement control hook > + * @icsk_ulp_ddp_data ULP direct data placement private data Neither of these socket layer intrusions are needed. All references but 1 -- the skbuff check -- are in the mlx5 driver. Any skb check that is needed can be handled with a different setting. > * @icsk_listen_portaddr_node hash to the portaddr listener hashtable > * @icsk_ca_state: Congestion control state > * @icsk_retransmits: Number of unrecovered [RTO] timeouts > @@ -94,6 +96,8 @@ struct inet_connection_sock { > const struct tcp_ulp_ops *icsk_ulp_ops; > void __rcu *icsk_ulp_data; > void (*icsk_clean_acked)(struct sock *sk, u32 acked_seq); > + const struct tcp_ddp_ulp_ops *icsk_ulp_ddp_ops; > + void __rcu *icsk_ulp_ddp_data; > struct hlist_node icsk_listen_portaddr_node; > unsigned int (*icsk_sync_mss)(struct sock *sk, u32 pmtu); > __u8 icsk_ca_state:5, > diff --git a/include/net/tcp_ddp.h b/include/net/tcp_ddp.h > new file mode 100644 > index 000000000000..df3264be4600 > --- /dev/null > +++ b/include/net/tcp_ddp.h > @@ -0,0 +1,129 @@ > +/* SPDX-License-Identifier: GPL-2.0 > + * > + * tcp_ddp.h > + * Author: Boris Pismenny <borisp@mellanox.com> > + * Copyright (C) 2020 Mellanox Technologies. > + */ > +#ifndef _TCP_DDP_H > +#define _TCP_DDP_H > + > +#include <linux/netdevice.h> > +#include <net/inet_connection_sock.h> > +#include <net/sock.h> > + > +/* limits returned by the offload driver, zero means don't care */ > +struct tcp_ddp_limits { > + int max_ddp_sgl_len; > +}; > + > +enum tcp_ddp_type { > + TCP_DDP_NVME = 1, > +}; > + > +/** > + * struct tcp_ddp_config - Generic tcp ddp configuration: tcp ddp IO queue > + * config implementations must use this as the first member. > + * Add new instances of tcp_ddp_config below (nvme-tcp, etc.). > + */ > +struct tcp_ddp_config { > + enum tcp_ddp_type type; > + unsigned char buf[]; you have this variable length buf, but it is not used (as far as I can tell). But then ... > +}; > + > +/** > + * struct nvme_tcp_ddp_config - nvme tcp ddp configuration for an IO queue > + * > + * @pfv: pdu version (e.g., NVME_TCP_PFV_1_0) > + * @cpda: controller pdu data alignmend (dwords, 0's based) > + * @dgst: digest types enabled. > + * The netdev will offload crc if ddp_crc is supported. > + * @queue_size: number of nvme-tcp IO queue elements > + * @queue_id: queue identifier > + * @cpu_io: cpu core running the IO thread for this queue > + */ > +struct nvme_tcp_ddp_config { > + struct tcp_ddp_config cfg; ... how would you use it within another struct like this? > + > + u16 pfv; > + u8 cpda; > + u8 dgst; > + int queue_size; > + int queue_id; > + int io_cpu; > +}; > + > +/** > + * struct tcp_ddp_io - tcp ddp configuration for an IO request. > + * > + * @command_id: identifier on the wire associated with these buffers > + * @nents: number of entries in the sg_table > + * @sg_table: describing the buffers for this IO request > + * @first_sgl: first SGL in sg_table > + */ > +struct tcp_ddp_io { > + u32 command_id; > + int nents; > + struct sg_table sg_table; > + struct scatterlist first_sgl[SG_CHUNK_SIZE]; > +}; > + > +/* struct tcp_ddp_dev_ops - operations used by an upper layer protocol to configure ddp offload > + * > + * @tcp_ddp_limits: limit the number of scatter gather entries per IO. > + * the device driver can use this to limit the resources allocated per queue. > + * @tcp_ddp_sk_add: add offload for the queue represennted by the socket+config pair. > + * this function is used to configure either copy, crc or both offloads. > + * @tcp_ddp_sk_del: remove offload from the socket, and release any device related resources. > + * @tcp_ddp_setup: request copy offload for buffers associated with a command_id in tcp_ddp_io. > + * @tcp_ddp_teardown: release offload resources association between buffers and command_id in > + * tcp_ddp_io. > + * @tcp_ddp_resync: respond to the driver's resync_request. Called only if resync is successful. > + */ > +struct tcp_ddp_dev_ops { > + int (*tcp_ddp_limits)(struct net_device *netdev, > + struct tcp_ddp_limits *limits); > + int (*tcp_ddp_sk_add)(struct net_device *netdev, > + struct sock *sk, > + struct tcp_ddp_config *config); > + void (*tcp_ddp_sk_del)(struct net_device *netdev, > + struct sock *sk); > + int (*tcp_ddp_setup)(struct net_device *netdev, > + struct sock *sk, > + struct tcp_ddp_io *io); > + int (*tcp_ddp_teardown)(struct net_device *netdev, > + struct sock *sk, > + struct tcp_ddp_io *io, > + void *ddp_ctx); > + void (*tcp_ddp_resync)(struct net_device *netdev, > + struct sock *sk, u32 seq); > +}; > + > +#define TCP_DDP_RESYNC_REQ (1 << 0) > + > +/** > + * struct tcp_ddp_ulp_ops - Interface to register uppper layer Direct Data Placement (DDP) TCP offload > + */ > +struct tcp_ddp_ulp_ops { > + /* NIC requests ulp to indicate if @seq is the start of a message */ > + bool (*resync_request)(struct sock *sk, u32 seq, u32 flags); > + /* NIC driver informs the ulp that ddp teardown is done - used for async completions*/ > + void (*ddp_teardown_done)(void *ddp_ctx); > +}; > + > +/** > + * struct tcp_ddp_ctx - Generic tcp ddp context: device driver per queue contexts must > + * use this as the first member. > + */ > +struct tcp_ddp_ctx { > + enum tcp_ddp_type type; > + unsigned char buf[]; similar to my comment above, I did not see any uses of the buf element.
On 12/8/20 5:57 PM, David Ahern wrote: >> diff --git a/include/net/inet_connection_sock.h b/include/net/inet_connection_sock.h >> index 7338b3865a2a..a08b85b53aa8 100644 >> --- a/include/net/inet_connection_sock.h >> +++ b/include/net/inet_connection_sock.h >> @@ -66,6 +66,8 @@ struct inet_connection_sock_af_ops { >> * @icsk_ulp_ops Pluggable ULP control hook >> * @icsk_ulp_data ULP private data >> * @icsk_clean_acked Clean acked data hook >> + * @icsk_ulp_ddp_ops Pluggable ULP direct data placement control hook >> + * @icsk_ulp_ddp_data ULP direct data placement private data > > Neither of these socket layer intrusions are needed. All references but > 1 -- the skbuff check -- are in the mlx5 driver. Any skb check that is > needed can be handled with a different setting. missed the nvme ops for the driver to callback to the socket owner.
On 09/12/2020 2:38, David Ahern wrote: > > The AF_XDP reference was to differentiate one zerocopy use case (all > packets go to userspace) from another (kernel managed TCP socket with > zerocopy payload). You are focusing on a very narrow use case - kernel > based NVMe over TCP - of a more general problem. > Please note that although our framework implements support for nvme-tcp, we designed it to fit iscsi as well, and hopefully future protocols too, as general as we could. For why this could not be generalized further see below. > You have a TCP socket and a design that only works for kernel owned > sockets. You have specialized queues in the NIC, a flow rule directing > packets to those queues. Presumably some ULP parser in the NIC > associated with the queues to process NVMe packets. Rather than copying > headers (ethernet/ip/tcp) to one buffer and payload to another (which is > similar to what Jonathan Lemon is working on), this design has a ULP > processor that just splits out the TCP payload even more making it > highly selective about which part of the packet is put into which > buffer. Take out the NVMe part, and it is header split with zerocopy for > the payload - a generic feature that can have a wider impact with NVMe > as a special case. > There is more to this than TCP zerocopy that exists in userspace or inside the kernel. First, please note that the patches include support for CRC offload as well as data placement. Second, data-placement is not the same as zerocopy for the following reasons: (1) The former places buffers *exactly* where the user requests regardless of the order of response arrivals, while the latter places packets in anonymous buffers according to packet arrival order. Therefore, zerocopy can be implemented using data placement, but not vice versa. (2) Data-placement supports sub-page zerocopy, unlike page-flipping techniques (i.e., TCP_ZEROCOPY). (3) Page-flipping can't work for any storage initiator because the destination buffer is owned by some user pagecache or process using O_DIRECT. (4) Storage over TCP PDUs are not necessarily aligned to TCP packets, i.e., the PDU header can be in the middle of a packet, so header-data split alone isn't enough. I wish we could do the same using some simpler zerocopy mechanism, it would indeed simplify things. But, unfortunately this would severely restrict generality, no sub-page support and alignment between PDUs and packets, and performance (ordering of PDUs).
On 09/12/2020 2:57, David Ahern wrote: > On 12/7/20 2:06 PM, Boris Pismenny wrote: >> diff --git a/include/linux/netdev_features.h b/include/linux/netdev_features.h >> index 934de56644e7..fb35dcac03d2 100644 >> --- a/include/linux/netdev_features.h >> +++ b/include/linux/netdev_features.h >> @@ -84,6 +84,7 @@ enum { >> NETIF_F_GRO_FRAGLIST_BIT, /* Fraglist GRO */ >> >> NETIF_F_HW_MACSEC_BIT, /* Offload MACsec operations */ >> + NETIF_F_HW_TCP_DDP_BIT, /* TCP direct data placement offload */ >> >> /* >> * Add your fresh new feature above and remember to update >> @@ -157,6 +158,7 @@ enum { >> #define NETIF_F_GRO_FRAGLIST __NETIF_F(GRO_FRAGLIST) >> #define NETIF_F_GSO_FRAGLIST __NETIF_F(GSO_FRAGLIST) >> #define NETIF_F_HW_MACSEC __NETIF_F(HW_MACSEC) >> +#define NETIF_F_HW_TCP_DDP __NETIF_F(HW_TCP_DDP) > > All of the DDP naming seems wrong to me. I realize the specific use case > is targeted payloads of a ULP, but it is still S/W handing H/W specific > buffers for a payload of a flow. > > This is intended to be used strictly by ULPs. DDP is how such things were called in the past. It is more than zerocopy as explained before, so naming it zerocopy will be misleading at best. I can propose another name. How about: "Autonomous copy offload (ACO)" and "Autonomous crc offload (ACRC)"? This will indicate that it is independent of other offloads, i.e. autonomous, while also informing users of the functionality (copy/crc). Other names are welcome too. >> * @icsk_listen_portaddr_node hash to the portaddr listener hashtable >> * @icsk_ca_state: Congestion control state >> * @icsk_retransmits: Number of unrecovered [RTO] timeouts >> @@ -94,6 +96,8 @@ struct inet_connection_sock { >> const struct tcp_ulp_ops *icsk_ulp_ops; >> void __rcu *icsk_ulp_data; >> void (*icsk_clean_acked)(struct sock *sk, u32 acked_seq); >> + const struct tcp_ddp_ulp_ops *icsk_ulp_ddp_ops; >> + void __rcu *icsk_ulp_ddp_data; >> struct hlist_node icsk_listen_portaddr_node; >> unsigned int (*icsk_sync_mss)(struct sock *sk, u32 pmtu); >> __u8 icsk_ca_state:5, >> diff --git a/include/net/tcp_ddp.h b/include/net/tcp_ddp.h >> new file mode 100644 >> index 000000000000..df3264be4600 >> --- /dev/null >> +++ b/include/net/tcp_ddp.h >> @@ -0,0 +1,129 @@ >> +/* SPDX-License-Identifier: GPL-2.0 >> + * >> + * tcp_ddp.h >> + * Author: Boris Pismenny <borisp@mellanox.com> >> + * Copyright (C) 2020 Mellanox Technologies. >> + */ >> +#ifndef _TCP_DDP_H >> +#define _TCP_DDP_H >> + >> +#include <linux/netdevice.h> >> +#include <net/inet_connection_sock.h> >> +#include <net/sock.h> >> + >> +/* limits returned by the offload driver, zero means don't care */ >> +struct tcp_ddp_limits { >> + int max_ddp_sgl_len; >> +}; >> + >> +enum tcp_ddp_type { >> + TCP_DDP_NVME = 1, >> +}; >> + >> +/** >> + * struct tcp_ddp_config - Generic tcp ddp configuration: tcp ddp IO queue >> + * config implementations must use this as the first member. >> + * Add new instances of tcp_ddp_config below (nvme-tcp, etc.). >> + */ >> +struct tcp_ddp_config { >> + enum tcp_ddp_type type; >> + unsigned char buf[]; > > you have this variable length buf, but it is not used (as far as I can > tell). But then ... > > True. This buf[] is here to indicate that users are expected to extend it with the ULP specific data as in nvme_tcp_config. We can remove it and leave a comment if you prefer that. >> +}; >> + >> +/** >> + * struct nvme_tcp_ddp_config - nvme tcp ddp configuration for an IO queue >> + * >> + * @pfv: pdu version (e.g., NVME_TCP_PFV_1_0) >> + * @cpda: controller pdu data alignmend (dwords, 0's based) >> + * @dgst: digest types enabled. >> + * The netdev will offload crc if ddp_crc is supported. >> + * @queue_size: number of nvme-tcp IO queue elements >> + * @queue_id: queue identifier >> + * @cpu_io: cpu core running the IO thread for this queue >> + */ >> +struct nvme_tcp_ddp_config { >> + struct tcp_ddp_config cfg; > > ... how would you use it within another struct like this? > You don't. >> + >> + u16 pfv; >> + u8 cpda; >> + u8 dgst; >> + int queue_size; >> + int queue_id; >> + int io_cpu; >> +}; >> + >> +/** >> + * struct tcp_ddp_io - tcp ddp configuration for an IO request. >> + * >> + * @command_id: identifier on the wire associated with these buffers >> + * @nents: number of entries in the sg_table >> + * @sg_table: describing the buffers for this IO request >> + * @first_sgl: first SGL in sg_table >> + */ >> +struct tcp_ddp_io { >> + u32 command_id; >> + int nents; >> + struct sg_table sg_table; >> + struct scatterlist first_sgl[SG_CHUNK_SIZE]; >> +}; >> + >> +/* struct tcp_ddp_dev_ops - operations used by an upper layer protocol to configure ddp offload >> + * >> + * @tcp_ddp_limits: limit the number of scatter gather entries per IO. >> + * the device driver can use this to limit the resources allocated per queue. >> + * @tcp_ddp_sk_add: add offload for the queue represennted by the socket+config pair. >> + * this function is used to configure either copy, crc or both offloads. >> + * @tcp_ddp_sk_del: remove offload from the socket, and release any device related resources. >> + * @tcp_ddp_setup: request copy offload for buffers associated with a command_id in tcp_ddp_io. >> + * @tcp_ddp_teardown: release offload resources association between buffers and command_id in >> + * tcp_ddp_io. >> + * @tcp_ddp_resync: respond to the driver's resync_request. Called only if resync is successful. >> + */ >> +struct tcp_ddp_dev_ops { >> + int (*tcp_ddp_limits)(struct net_device *netdev, >> + struct tcp_ddp_limits *limits); >> + int (*tcp_ddp_sk_add)(struct net_device *netdev, >> + struct sock *sk, >> + struct tcp_ddp_config *config); >> + void (*tcp_ddp_sk_del)(struct net_device *netdev, >> + struct sock *sk); >> + int (*tcp_ddp_setup)(struct net_device *netdev, >> + struct sock *sk, >> + struct tcp_ddp_io *io); >> + int (*tcp_ddp_teardown)(struct net_device *netdev, >> + struct sock *sk, >> + struct tcp_ddp_io *io, >> + void *ddp_ctx); >> + void (*tcp_ddp_resync)(struct net_device *netdev, >> + struct sock *sk, u32 seq); >> +}; >> + >> +#define TCP_DDP_RESYNC_REQ (1 << 0) >> + >> +/** >> + * struct tcp_ddp_ulp_ops - Interface to register uppper layer Direct Data Placement (DDP) TCP offload >> + */ >> +struct tcp_ddp_ulp_ops { >> + /* NIC requests ulp to indicate if @seq is the start of a message */ >> + bool (*resync_request)(struct sock *sk, u32 seq, u32 flags); >> + /* NIC driver informs the ulp that ddp teardown is done - used for async completions*/ >> + void (*ddp_teardown_done)(void *ddp_ctx); >> +}; >> + >> +/** >> + * struct tcp_ddp_ctx - Generic tcp ddp context: device driver per queue contexts must >> + * use this as the first member. >> + */ >> +struct tcp_ddp_ctx { >> + enum tcp_ddp_type type; >> + unsigned char buf[]; > > similar to my comment above, I did not see any uses of the buf element. > Same idea, we will remove it an leave a comment.
On 09/12/2020 3:11, David Ahern wrote: > On 12/8/20 5:57 PM, David Ahern wrote: >>> diff --git a/include/net/inet_connection_sock.h b/include/net/inet_connection_sock.h >>> index 7338b3865a2a..a08b85b53aa8 100644 >>> --- a/include/net/inet_connection_sock.h >>> +++ b/include/net/inet_connection_sock.h >>> @@ -66,6 +66,8 @@ struct inet_connection_sock_af_ops { >>> * @icsk_ulp_ops Pluggable ULP control hook >>> * @icsk_ulp_data ULP private data >>> * @icsk_clean_acked Clean acked data hook >>> + * @icsk_ulp_ddp_ops Pluggable ULP direct data placement control hook >>> + * @icsk_ulp_ddp_data ULP direct data placement private data >> >> Neither of these socket layer intrusions are needed. All references but >> 1 -- the skbuff check -- are in the mlx5 driver. Any skb check that is >> needed can be handled with a different setting. > > missed the nvme ops for the driver to callback to the socket owner. > Hopefully it is clear that these are needed, and indeed we use them in both driver and nvme-tcp layers.
On 12/9/20 1:15 AM, Boris Pismenny wrote: > On 09/12/2020 2:38, David Ahern wrote: >> >> The AF_XDP reference was to differentiate one zerocopy use case (all >> packets go to userspace) from another (kernel managed TCP socket with >> zerocopy payload). You are focusing on a very narrow use case - kernel >> based NVMe over TCP - of a more general problem. >> > > Please note that although our framework implements support for nvme-tcp, > we designed it to fit iscsi as well, and hopefully future protocols too, > as general as we could. For why this could not be generalized further > see below. > >> You have a TCP socket and a design that only works for kernel owned >> sockets. You have specialized queues in the NIC, a flow rule directing >> packets to those queues. Presumably some ULP parser in the NIC >> associated with the queues to process NVMe packets. Rather than copying >> headers (ethernet/ip/tcp) to one buffer and payload to another (which is >> similar to what Jonathan Lemon is working on), this design has a ULP >> processor that just splits out the TCP payload even more making it >> highly selective about which part of the packet is put into which >> buffer. Take out the NVMe part, and it is header split with zerocopy for >> the payload - a generic feature that can have a wider impact with NVMe >> as a special case. >> > > There is more to this than TCP zerocopy that exists in userspace or > inside the kernel. First, please note that the patches include support for > CRC offload as well as data placement. Second, data-placement is not the same Yes, the CRC offload is different, but I think it is orthogonal to the 'where does h/w put the data' problem. > as zerocopy for the following reasons: > (1) The former places buffers *exactly* where the user requests > regardless of the order of response arrivals, while the latter places packets > in anonymous buffers according to packet arrival order. Therefore, zerocopy > can be implemented using data placement, but not vice versa. Fundamentally, it is an SGL and a TCP sequence number. There is a starting point where seq N == sgl element 0, position 0. Presumably there is a hardware cursor to track where you are in filling the SGL as packets are processed. You abort on OOO, so it seems like a fairly straightfoward problem. > (2) Data-placement supports sub-page zerocopy, unlike page-flipping > techniques (i.e., TCP_ZEROCOPY). I am not pushing for or suggesting any page-flipping. I understand the limitations of that approach. > (3) Page-flipping can't work for any storage initiator because the > destination buffer is owned by some user pagecache or process using O_DIRECT. > (4) Storage over TCP PDUs are not necessarily aligned to TCP packets, > i.e., the PDU header can be in the middle of a packet, so header-data split > alone isn't enough. yes, TCP is a byte stream and you have to have a cursor marking last written spot in the SGL. More below. > > I wish we could do the same using some simpler zerocopy mechanism, > it would indeed simplify things. But, unfortunately this would severely > restrict generality, no sub-page support and alignment between PDUs > and packets, and performance (ordering of PDUs). > My biggest concern is that you are adding checks in the fast path for a very specific use case. If / when Rx zerocopy happens (and I suspect it has to happen soon to handle the ever increasing speeds), nothing about this patch set is reusable and worse more checks are needed in the fast path. I think it is best if you make this more generic — at least anything touching core code. For example, you have an iov static key hook managed by a driver for generic code. There are a few ways around that. One is by adding skb details to the nvme code — ie., walking the skb fragments, seeing that a given frag is in your allocated memory and skipping the copy. This would offer best performance since it skips all unnecessary checks. Another option is to export __skb_datagram_iter, use it and define your own copy handler that does the address compare and skips the copy. Key point - only your code path is affected. Similarly for the NVMe SGLs and DDP offload - a more generic solution allows other use cases to build on this as opposed to the checks you want for a special case. For example, a split at the protocol headers / payload boundaries would be a generic solution where kernel managed protocols get data in one buffer and socket data is put into a given SGL. I am guessing that you have to be already doing this to put PDU payloads into an SGL and other headers into other memory to make a complete packet, so this is not too far off from what you are already doing. Let me walk through an example with assumptions about your hardware's capabilities, and you correct me where I am wrong. Assume you have a 'full' command response of this form: +------------- ... ----------------+---------+---------+--------+-----+ | big data segment | PDU hdr | TCP hdr | IP hdr | eth | +------------- ... ----------------+---------+---------+--------+-----+ but it shows up to the host in 3 packets like this (ideal case): +-------------------------+---------+---------+--------+-----+ | data - seg 1 | PDU hdr | TCP hdr | IP hdr | eth | +-------------------------+---------+---------+--------+-----+ +-----------------------------------+---------+--------+-----+ | data - seg 2 | TCP hdr | IP hdr | eth | +-----------------------------------+---------+--------+-----+ +-----------------+---------+--------+-----+ | payload - seg 3 | TCP hdr | IP hdr | eth | +-----------------+---------+--------+-----+ The hardware splits the eth/IP/tcp headers from payload like this (again, your hardware has to know these boundaries to accomplish what you want): +-------------------------+---------+ +---------+--------+-----+ | data - seg 1 | PDU hdr | | TCP hdr | IP hdr | eth | +-------------------------+---------+ +---------+--------+-----+ +-----------------------------------+ +---------+--------+-----+ | data - seg 2 | | TCP hdr | IP hdr | eth | +-----------------------------------+ +---------+--------+-----+ +-----------------+ +---------+--------+-----+ | payload - seg 3 | | TCP hdr | IP hdr | eth | +-----------------+ +---------+--------+-----+ Left side goes into the SGLs posted for this socket / flow; the right side goes into some other memory resource made available for headers. This is very close to what you are doing now - with the exception of the PDU header being put to the right side. NVMe code then just needs to set the iov offset (or adjust the base_addr) to skip over the PDU header - standard options for an iov. Yes, TCP is a byte stream, so the packets could very well show up like this: +--------------+---------+-----------+---------+--------+-----+ | data - seg 1 | PDU hdr | prev data | TCP hdr | IP hdr | eth | +--------------+---------+-----------+---------+--------+-----+ +-----------------------------------+---------+--------+-----+ | payload - seg 2 | TCP hdr | IP hdr | eth | +-----------------------------------+---------+--------+-----+ +-------- +-------------------------+---------+--------+-----+ | PDU hdr | payload - seg 3 | TCP hdr | IP hdr | eth | +---------+-------------------------+---------+--------+-----+ If your hardware can extract the NVMe payload into a targeted SGL like you want in this set, then it has some logic for parsing headers and "snapping" an SGL to a new element. ie., it already knows 'prev data' goes with the in-progress PDU, sees more data, recognizes a new PDU header and a new payload. That means it already has to handle a 'snap-to-PDU' style argument where the end of the payload closes out an SGL element and the next PDU hdr starts in a new SGL element (ie., 'prev data' closes out sgl[i], and the next PDU hdr starts sgl[i+1]). So in this case, you want 'snap-to-PDU' but that could just as easily be 'no snap at all', just a byte stream and filling an SGL after the protocol headers. Key point here is that this is the start of a generic header / data split that could work for other applications - not just NVMe. eth/IP/TCP headers are consumed by the Linux networking stack; data is in application owned, socket based SGLs to avoid copies. ### A dump of other comments about this patch set: - there are a LOT of unnecessary typecasts around tcp_ddp_ctx that can be avoided by using container_of. - you have an accessor tcp_ddp_get_ctx but no setter; all uses of tcp_ddp_get_ctx are within mlx5. why open code the set but use the accessor for the get? Worse, mlx5e_nvmeotcp_queue_teardown actually has both — uses the accessor and open codes setting icsk_ulp_ddp_data. - the driver is storing private data on the socket. Nothing about the socket layer cares and the mlx5 driver is already tracking that data in priv->nvmeotcp->queue_hash. As I mentioned in a previous response, I understand the socket ops are needed for the driver level to call into the socket layer, but the data part does not seem to be needed. - nvme_tcp_offload_socket and nvme_tcp_offload_limits both return int yet the value is ignored - the build robot found a number of problems (it pulls my github tree and I pushed this set to it to move across computers). I think the patch set would be easier to follow if you restructured the patches to 1 thing only per patch -- e.g., split patch 2 into netdev bits and socket bits. Add the netdev feature bit and operations in 1 patch and add the socket ops in a second patch with better commit logs about why each is needed and what is done.
diff --git a/include/linux/netdev_features.h b/include/linux/netdev_features.h index 934de56644e7..fb35dcac03d2 100644 --- a/include/linux/netdev_features.h +++ b/include/linux/netdev_features.h @@ -84,6 +84,7 @@ enum { NETIF_F_GRO_FRAGLIST_BIT, /* Fraglist GRO */ NETIF_F_HW_MACSEC_BIT, /* Offload MACsec operations */ + NETIF_F_HW_TCP_DDP_BIT, /* TCP direct data placement offload */ /* * Add your fresh new feature above and remember to update @@ -157,6 +158,7 @@ enum { #define NETIF_F_GRO_FRAGLIST __NETIF_F(GRO_FRAGLIST) #define NETIF_F_GSO_FRAGLIST __NETIF_F(GSO_FRAGLIST) #define NETIF_F_HW_MACSEC __NETIF_F(HW_MACSEC) +#define NETIF_F_HW_TCP_DDP __NETIF_F(HW_TCP_DDP) /* Finds the next feature with the highest number of the range of start till 0. */ diff --git a/include/linux/netdevice.h b/include/linux/netdevice.h index a07c8e431f45..755766976408 100644 --- a/include/linux/netdevice.h +++ b/include/linux/netdevice.h @@ -934,6 +934,7 @@ struct dev_ifalias { struct devlink; struct tlsdev_ops; +struct tcp_ddp_dev_ops; struct netdev_name_node { struct hlist_node hlist; @@ -1930,6 +1931,10 @@ struct net_device { const struct tlsdev_ops *tlsdev_ops; #endif +#ifdef CONFIG_TCP_DDP + const struct tcp_ddp_dev_ops *tcp_ddp_ops; +#endif + const struct header_ops *header_ops; unsigned int flags; diff --git a/include/net/inet_connection_sock.h b/include/net/inet_connection_sock.h index 7338b3865a2a..a08b85b53aa8 100644 --- a/include/net/inet_connection_sock.h +++ b/include/net/inet_connection_sock.h @@ -66,6 +66,8 @@ struct inet_connection_sock_af_ops { * @icsk_ulp_ops Pluggable ULP control hook * @icsk_ulp_data ULP private data * @icsk_clean_acked Clean acked data hook + * @icsk_ulp_ddp_ops Pluggable ULP direct data placement control hook + * @icsk_ulp_ddp_data ULP direct data placement private data * @icsk_listen_portaddr_node hash to the portaddr listener hashtable * @icsk_ca_state: Congestion control state * @icsk_retransmits: Number of unrecovered [RTO] timeouts @@ -94,6 +96,8 @@ struct inet_connection_sock { const struct tcp_ulp_ops *icsk_ulp_ops; void __rcu *icsk_ulp_data; void (*icsk_clean_acked)(struct sock *sk, u32 acked_seq); + const struct tcp_ddp_ulp_ops *icsk_ulp_ddp_ops; + void __rcu *icsk_ulp_ddp_data; struct hlist_node icsk_listen_portaddr_node; unsigned int (*icsk_sync_mss)(struct sock *sk, u32 pmtu); __u8 icsk_ca_state:5, diff --git a/include/net/tcp_ddp.h b/include/net/tcp_ddp.h new file mode 100644 index 000000000000..df3264be4600 --- /dev/null +++ b/include/net/tcp_ddp.h @@ -0,0 +1,129 @@ +/* SPDX-License-Identifier: GPL-2.0 + * + * tcp_ddp.h + * Author: Boris Pismenny <borisp@mellanox.com> + * Copyright (C) 2020 Mellanox Technologies. + */ +#ifndef _TCP_DDP_H +#define _TCP_DDP_H + +#include <linux/netdevice.h> +#include <net/inet_connection_sock.h> +#include <net/sock.h> + +/* limits returned by the offload driver, zero means don't care */ +struct tcp_ddp_limits { + int max_ddp_sgl_len; +}; + +enum tcp_ddp_type { + TCP_DDP_NVME = 1, +}; + +/** + * struct tcp_ddp_config - Generic tcp ddp configuration: tcp ddp IO queue + * config implementations must use this as the first member. + * Add new instances of tcp_ddp_config below (nvme-tcp, etc.). + */ +struct tcp_ddp_config { + enum tcp_ddp_type type; + unsigned char buf[]; +}; + +/** + * struct nvme_tcp_ddp_config - nvme tcp ddp configuration for an IO queue + * + * @pfv: pdu version (e.g., NVME_TCP_PFV_1_0) + * @cpda: controller pdu data alignmend (dwords, 0's based) + * @dgst: digest types enabled. + * The netdev will offload crc if ddp_crc is supported. + * @queue_size: number of nvme-tcp IO queue elements + * @queue_id: queue identifier + * @cpu_io: cpu core running the IO thread for this queue + */ +struct nvme_tcp_ddp_config { + struct tcp_ddp_config cfg; + + u16 pfv; + u8 cpda; + u8 dgst; + int queue_size; + int queue_id; + int io_cpu; +}; + +/** + * struct tcp_ddp_io - tcp ddp configuration for an IO request. + * + * @command_id: identifier on the wire associated with these buffers + * @nents: number of entries in the sg_table + * @sg_table: describing the buffers for this IO request + * @first_sgl: first SGL in sg_table + */ +struct tcp_ddp_io { + u32 command_id; + int nents; + struct sg_table sg_table; + struct scatterlist first_sgl[SG_CHUNK_SIZE]; +}; + +/* struct tcp_ddp_dev_ops - operations used by an upper layer protocol to configure ddp offload + * + * @tcp_ddp_limits: limit the number of scatter gather entries per IO. + * the device driver can use this to limit the resources allocated per queue. + * @tcp_ddp_sk_add: add offload for the queue represennted by the socket+config pair. + * this function is used to configure either copy, crc or both offloads. + * @tcp_ddp_sk_del: remove offload from the socket, and release any device related resources. + * @tcp_ddp_setup: request copy offload for buffers associated with a command_id in tcp_ddp_io. + * @tcp_ddp_teardown: release offload resources association between buffers and command_id in + * tcp_ddp_io. + * @tcp_ddp_resync: respond to the driver's resync_request. Called only if resync is successful. + */ +struct tcp_ddp_dev_ops { + int (*tcp_ddp_limits)(struct net_device *netdev, + struct tcp_ddp_limits *limits); + int (*tcp_ddp_sk_add)(struct net_device *netdev, + struct sock *sk, + struct tcp_ddp_config *config); + void (*tcp_ddp_sk_del)(struct net_device *netdev, + struct sock *sk); + int (*tcp_ddp_setup)(struct net_device *netdev, + struct sock *sk, + struct tcp_ddp_io *io); + int (*tcp_ddp_teardown)(struct net_device *netdev, + struct sock *sk, + struct tcp_ddp_io *io, + void *ddp_ctx); + void (*tcp_ddp_resync)(struct net_device *netdev, + struct sock *sk, u32 seq); +}; + +#define TCP_DDP_RESYNC_REQ (1 << 0) + +/** + * struct tcp_ddp_ulp_ops - Interface to register uppper layer Direct Data Placement (DDP) TCP offload + */ +struct tcp_ddp_ulp_ops { + /* NIC requests ulp to indicate if @seq is the start of a message */ + bool (*resync_request)(struct sock *sk, u32 seq, u32 flags); + /* NIC driver informs the ulp that ddp teardown is done - used for async completions*/ + void (*ddp_teardown_done)(void *ddp_ctx); +}; + +/** + * struct tcp_ddp_ctx - Generic tcp ddp context: device driver per queue contexts must + * use this as the first member. + */ +struct tcp_ddp_ctx { + enum tcp_ddp_type type; + unsigned char buf[]; +}; + +static inline struct tcp_ddp_ctx *tcp_ddp_get_ctx(const struct sock *sk) +{ + struct inet_connection_sock *icsk = inet_csk(sk); + + return (__force struct tcp_ddp_ctx *)icsk->icsk_ulp_ddp_data; +} + +#endif //_TCP_DDP_H diff --git a/net/Kconfig b/net/Kconfig index f4c32d982af6..3876861cdc90 100644 --- a/net/Kconfig +++ b/net/Kconfig @@ -457,6 +457,15 @@ config ETHTOOL_NETLINK netlink. It provides better extensibility and some new features, e.g. notification messages. +config TCP_DDP + bool "TCP direct data placement offload" + default n + help + Direct Data Placement (DDP) offload for TCP enables ULP, such as + NVMe-TCP/iSCSI, to request the NIC to place TCP payload data + of a command response directly into kernel pages. + + endif # if NET # Used by archs to tell that they support BPF JIT compiler plus which flavour. diff --git a/net/core/skbuff.c b/net/core/skbuff.c index ed61eed1195d..75354fb8fe94 100644 --- a/net/core/skbuff.c +++ b/net/core/skbuff.c @@ -69,6 +69,7 @@ #include <net/xfrm.h> #include <net/mpls.h> #include <net/mptcp.h> +#include <net/tcp_ddp.h> #include <linux/uaccess.h> #include <trace/events/skb.h> @@ -6135,9 +6136,15 @@ EXPORT_SYMBOL(pskb_extract); */ void skb_condense(struct sk_buff *skb) { + bool is_ddp = false; + +#ifdef CONFIG_TCP_DDP + is_ddp = skb->sk && inet_csk(skb->sk) && + inet_csk(skb->sk)->icsk_ulp_ddp_data; +#endif if (skb->data_len) { if (skb->data_len > skb->end - skb->tail || - skb_cloned(skb)) + skb_cloned(skb) || is_ddp) return; /* Nice, we can free page frag(s) right now */ diff --git a/net/ethtool/common.c b/net/ethtool/common.c index 24036e3055a1..a2ff7a4a6bbf 100644 --- a/net/ethtool/common.c +++ b/net/ethtool/common.c @@ -68,6 +68,7 @@ const char netdev_features_strings[NETDEV_FEATURE_COUNT][ETH_GSTRING_LEN] = { [NETIF_F_HW_TLS_RX_BIT] = "tls-hw-rx-offload", [NETIF_F_GRO_FRAGLIST_BIT] = "rx-gro-list", [NETIF_F_HW_MACSEC_BIT] = "macsec-hw-offload", + [NETIF_F_HW_TCP_DDP_BIT] = "tcp-ddp-offload", }; const char