xref: /linux/drivers/net/ethernet/fungible/funeth/funeth_txrx.h (revision f9bff0e31881d03badf191d3b0005839391f5f2b)
1 /* SPDX-License-Identifier: (GPL-2.0-only OR BSD-3-Clause) */
2 
3 #ifndef _FUNETH_TXRX_H
4 #define _FUNETH_TXRX_H
5 
6 #include <linux/netdevice.h>
7 #include <linux/u64_stats_sync.h>
8 
9 /* Tx descriptor size */
10 #define FUNETH_SQE_SIZE 64U
11 
12 /* Size of device headers per Tx packet */
13 #define FUNETH_FUNOS_HDR_SZ (sizeof(struct fun_eth_tx_req))
14 
15 /* Number of gather list entries per Tx descriptor */
16 #define FUNETH_GLE_PER_DESC (FUNETH_SQE_SIZE / sizeof(struct fun_dataop_gl))
17 
18 /* Max gather list size in bytes for an sk_buff. */
19 #define FUNETH_MAX_GL_SZ ((MAX_SKB_FRAGS + 1) * sizeof(struct fun_dataop_gl))
20 
21 #if IS_ENABLED(CONFIG_TLS_DEVICE)
22 # define FUNETH_TLS_SZ sizeof(struct fun_eth_tls)
23 #else
24 # define FUNETH_TLS_SZ 0
25 #endif
26 
27 /* Max number of Tx descriptors for an sk_buff using a gather list. */
28 #define FUNETH_MAX_GL_DESC \
29 	DIV_ROUND_UP((FUNETH_FUNOS_HDR_SZ + FUNETH_MAX_GL_SZ + FUNETH_TLS_SZ), \
30 		     FUNETH_SQE_SIZE)
31 
32 /* Max number of Tx descriptors for any packet. */
33 #define FUNETH_MAX_PKT_DESC FUNETH_MAX_GL_DESC
34 
35 /* Rx CQ descriptor size. */
36 #define FUNETH_CQE_SIZE 64U
37 
38 /* Offset of cqe_info within a CQE. */
39 #define FUNETH_CQE_INFO_OFFSET (FUNETH_CQE_SIZE - sizeof(struct fun_cqe_info))
40 
41 /* Construct the IRQ portion of a CQ doorbell. The resulting value arms the
42  * interrupt with the supplied time delay and packet count moderation settings.
43  */
44 #define FUN_IRQ_CQ_DB(usec, pkts) \
45 	(FUN_DB_IRQ_ARM_F | ((usec) << FUN_DB_INTCOAL_USEC_S) | \
46 	 ((pkts) << FUN_DB_INTCOAL_ENTRIES_S))
47 
48 /* As above for SQ doorbells. */
49 #define FUN_IRQ_SQ_DB(usec, pkts) \
50 	(FUN_DB_IRQ_ARM_F | \
51 	 ((usec) << FUN_DB_INTCOAL_USEC_S) | \
52 	 ((pkts) << FUN_DB_INTCOAL_ENTRIES_S))
53 
54 /* Per packet tailroom. Present only for 1-frag packets. */
55 #define FUN_RX_TAILROOM SKB_DATA_ALIGN(sizeof(struct skb_shared_info))
56 
57 /* Per packet headroom for XDP. Preferred over XDP_PACKET_HEADROOM to
58  * accommodate two packets per buffer for 4K pages and 1500B MTUs.
59  */
60 #define FUN_XDP_HEADROOM 192
61 
62 /* Initialization state of a queue. */
63 enum {
64 	FUN_QSTATE_DESTROYED, /* what queue? */
65 	FUN_QSTATE_INIT_SW,   /* exists in SW, not on the device */
66 	FUN_QSTATE_INIT_FULL, /* exists both in SW and on device */
67 };
68 
69 /* Initialization state of an interrupt. */
70 enum {
71 	FUN_IRQ_INIT,      /* initialized and in the XArray but inactive */
72 	FUN_IRQ_REQUESTED, /* request_irq() done */
73 	FUN_IRQ_ENABLED,   /* processing enabled */
74 	FUN_IRQ_DISABLED,  /* processing disabled */
75 };
76 
77 struct bpf_prog;
78 
79 struct funeth_txq_stats {  /* per Tx queue SW counters */
80 	u64 tx_pkts;       /* # of Tx packets */
81 	u64 tx_bytes;      /* total bytes of Tx packets */
82 	u64 tx_cso;        /* # of packets with checksum offload */
83 	u64 tx_tso;        /* # of non-encapsulated TSO super-packets */
84 	u64 tx_encap_tso;  /* # of encapsulated TSO super-packets */
85 	u64 tx_uso;        /* # of non-encapsulated UDP LSO super-packets */
86 	u64 tx_more;       /* # of DBs elided due to xmit_more */
87 	u64 tx_nstops;     /* # of times the queue has stopped */
88 	u64 tx_nrestarts;  /* # of times the queue has restarted */
89 	u64 tx_map_err;    /* # of packets dropped due to DMA mapping errors */
90 	u64 tx_xdp_full;   /* # of XDP packets that could not be enqueued */
91 	u64 tx_tls_pkts;   /* # of Tx TLS packets offloaded to HW */
92 	u64 tx_tls_bytes;  /* Tx bytes of HW-handled TLS payload */
93 	u64 tx_tls_fallback; /* attempted Tx TLS offloads punted to SW */
94 	u64 tx_tls_drops;  /* attempted Tx TLS offloads dropped */
95 };
96 
97 struct funeth_tx_info {      /* per Tx descriptor state */
98 	union {
99 		struct sk_buff *skb;    /* associated packet (sk_buff path) */
100 		struct xdp_frame *xdpf; /* associated XDP frame (XDP path) */
101 	};
102 };
103 
104 struct funeth_txq {
105 	/* RO cacheline of frequently accessed data */
106 	u32 mask;               /* queue depth - 1 */
107 	u32 hw_qid;             /* device ID of the queue */
108 	void *desc;             /* base address of descriptor ring */
109 	struct funeth_tx_info *info;
110 	struct device *dma_dev; /* device for DMA mappings */
111 	volatile __be64 *hw_wb; /* HW write-back location */
112 	u32 __iomem *db;        /* SQ doorbell register address */
113 	struct netdev_queue *ndq;
114 	dma_addr_t dma_addr;    /* DMA address of descriptor ring */
115 	/* producer R/W cacheline */
116 	u16 qidx;               /* queue index within net_device */
117 	u16 ethid;
118 	u32 prod_cnt;           /* producer counter */
119 	struct funeth_txq_stats stats;
120 	/* shared R/W cacheline, primarily accessed by consumer */
121 	u32 irq_db_val;         /* value written to IRQ doorbell */
122 	u32 cons_cnt;           /* consumer (cleanup) counter */
123 	struct net_device *netdev;
124 	struct fun_irq *irq;
125 	int numa_node;
126 	u8 init_state;          /* queue initialization state */
127 	struct u64_stats_sync syncp;
128 };
129 
130 struct funeth_rxq_stats {  /* per Rx queue SW counters */
131 	u64 rx_pkts;       /* # of received packets, including SW drops */
132 	u64 rx_bytes;      /* total size of received packets */
133 	u64 rx_cso;        /* # of packets with checksum offload */
134 	u64 rx_bufs;       /* total # of Rx buffers provided to device */
135 	u64 gro_pkts;      /* # of GRO superpackets */
136 	u64 gro_merged;    /* # of pkts merged into existing GRO superpackets */
137 	u64 rx_page_alloc; /* # of page allocations for Rx buffers */
138 	u64 rx_budget;     /* NAPI iterations that exhausted their budget */
139 	u64 rx_mem_drops;  /* # of packets dropped due to memory shortage */
140 	u64 rx_map_err;    /* # of page DMA mapping errors */
141 	u64 xdp_drops;     /* XDP_DROPped packets */
142 	u64 xdp_tx;        /* successful XDP transmits */
143 	u64 xdp_redir;     /* successful XDP redirects */
144 	u64 xdp_err;       /* packets dropped due to XDP errors */
145 };
146 
147 struct funeth_rxbuf {          /* per Rx buffer state */
148 	struct page *page;     /* associated page */
149 	dma_addr_t dma_addr;   /* DMA address of page start */
150 	int pg_refs;           /* page refs held by driver */
151 	int node;              /* page node, or -1 if it is PF_MEMALLOC */
152 };
153 
154 struct funeth_rx_cache {       /* cache of DMA-mapped previously used buffers */
155 	struct funeth_rxbuf *bufs; /* base of Rx buffer state ring */
156 	unsigned int prod_cnt;     /* producer counter */
157 	unsigned int cons_cnt;     /* consumer counter */
158 	unsigned int mask;         /* depth - 1 */
159 };
160 
161 /* An Rx queue consists of a CQ and an SQ used to provide Rx buffers. */
162 struct funeth_rxq {
163 	struct net_device *netdev;
164 	struct napi_struct *napi;
165 	struct device *dma_dev;    /* device for DMA mappings */
166 	void *cqes;                /* base of CQ descriptor ring */
167 	const void *next_cqe_info; /* fun_cqe_info of next CQE */
168 	u32 __iomem *cq_db;        /* CQ doorbell register address */
169 	unsigned int cq_head;      /* CQ head index */
170 	unsigned int cq_mask;      /* CQ depth - 1 */
171 	u16 phase;                 /* CQ phase tag */
172 	u16 qidx;                  /* queue index within net_device */
173 	unsigned int irq_db_val;   /* IRQ info for CQ doorbell */
174 	struct fun_eprq_rqbuf *rqes; /* base of RQ descriptor ring */
175 	struct funeth_rxbuf *bufs; /* base of Rx buffer state ring */
176 	struct funeth_rxbuf *cur_buf; /* currently active buffer */
177 	u32 __iomem *rq_db;        /* RQ doorbell register address */
178 	unsigned int rq_cons;      /* RQ consumer counter */
179 	unsigned int rq_mask;      /* RQ depth - 1 */
180 	unsigned int buf_offset;   /* offset of next pkt in head buffer */
181 	u8 xdp_flush;              /* XDP flush types needed at NAPI end */
182 	u8 init_state;             /* queue initialization state */
183 	u16 headroom;              /* per packet headroom */
184 	unsigned int rq_cons_db;   /* value of rq_cons at last RQ db */
185 	unsigned int rq_db_thres;  /* # of new buffers needed to write RQ db */
186 	struct funeth_rxbuf spare_buf; /* spare for next buffer replacement */
187 	struct funeth_rx_cache cache; /* used buffer cache */
188 	struct bpf_prog *xdp_prog; /* optional XDP BPF program */
189 	struct funeth_rxq_stats stats;
190 	dma_addr_t cq_dma_addr;    /* DMA address of CQE ring */
191 	dma_addr_t rq_dma_addr;    /* DMA address of RQE ring */
192 	u16 irq_cnt;
193 	u32 hw_cqid;               /* device ID of the queue's CQ */
194 	u32 hw_sqid;               /* device ID of the queue's SQ */
195 	int numa_node;
196 	struct u64_stats_sync syncp;
197 	struct xdp_rxq_info xdp_rxq;
198 };
199 
200 #define FUN_QSTAT_INC(q, counter) \
201 	do { \
202 		u64_stats_update_begin(&(q)->syncp); \
203 		(q)->stats.counter++; \
204 		u64_stats_update_end(&(q)->syncp); \
205 	} while (0)
206 
207 #define FUN_QSTAT_READ(q, seq, stats_copy) \
208 	do { \
209 		seq = u64_stats_fetch_begin(&(q)->syncp); \
210 		stats_copy = (q)->stats; \
211 	} while (u64_stats_fetch_retry(&(q)->syncp, (seq)))
212 
213 #define FUN_INT_NAME_LEN (IFNAMSIZ + 16)
214 
215 struct fun_irq {
216 	struct napi_struct napi;
217 	struct funeth_txq *txq;
218 	struct funeth_rxq *rxq;
219 	u8 state;
220 	u16 irq_idx;              /* index of MSI-X interrupt */
221 	int irq;                  /* Linux IRQ vector */
222 	cpumask_t affinity_mask;  /* IRQ affinity */
223 	struct irq_affinity_notify aff_notify;
224 	char name[FUN_INT_NAME_LEN];
225 } ____cacheline_internodealigned_in_smp;
226 
227 /* Return the start address of the idx-th Tx descriptor. */
228 static inline void *fun_tx_desc_addr(const struct funeth_txq *q,
229 				     unsigned int idx)
230 {
231 	return q->desc + idx * FUNETH_SQE_SIZE;
232 }
233 
234 static inline void fun_txq_wr_db(const struct funeth_txq *q)
235 {
236 	unsigned int tail = q->prod_cnt & q->mask;
237 
238 	writel(tail, q->db);
239 }
240 
241 static inline int fun_irq_node(const struct fun_irq *p)
242 {
243 	return cpu_to_mem(cpumask_first(&p->affinity_mask));
244 }
245 
246 int fun_rxq_napi_poll(struct napi_struct *napi, int budget);
247 int fun_txq_napi_poll(struct napi_struct *napi, int budget);
248 netdev_tx_t fun_start_xmit(struct sk_buff *skb, struct net_device *netdev);
249 bool fun_xdp_tx(struct funeth_txq *q, struct xdp_frame *xdpf);
250 int fun_xdp_xmit_frames(struct net_device *dev, int n,
251 			struct xdp_frame **frames, u32 flags);
252 
253 int funeth_txq_create(struct net_device *dev, unsigned int qidx,
254 		      unsigned int ndesc, struct fun_irq *irq, int state,
255 		      struct funeth_txq **qp);
256 int fun_txq_create_dev(struct funeth_txq *q, struct fun_irq *irq);
257 struct funeth_txq *funeth_txq_free(struct funeth_txq *q, int state);
258 int funeth_rxq_create(struct net_device *dev, unsigned int qidx,
259 		      unsigned int ncqe, unsigned int nrqe, struct fun_irq *irq,
260 		      int state, struct funeth_rxq **qp);
261 int fun_rxq_create_dev(struct funeth_rxq *q, struct fun_irq *irq);
262 struct funeth_rxq *funeth_rxq_free(struct funeth_rxq *q, int state);
263 int fun_rxq_set_bpf(struct funeth_rxq *q, struct bpf_prog *prog);
264 
265 #endif /* _FUNETH_TXRX_H */
266