1 /* SPDX-License-Identifier: GPL-2.0 */
2 /* Copyright(c) 2013 - 2018 Intel Corporation. */
9 /* Interrupt Throttling and Rate Limiting Goodies */
10 #define I40E_DEFAULT_IRQ_WORK 256
12 /* The datasheet for the X710 and XL710 indicate that the maximum value for
13 * the ITR is 8160usec which is then called out as 0xFF0 with a 2usec
14 * resolution. 8160 is 0x1FE0 when written out in hex. So instead of storing
15 * the register value which is divided by 2 lets use the actual values and
16 * avoid an excessive amount of translation.
18 #define I40E_ITR_DYNAMIC 0x8000 /* use top bit as a flag */
19 #define I40E_ITR_MASK 0x1FFE /* mask for ITR register value */
20 #define I40E_MIN_ITR 2 /* reg uses 2 usec resolution */
21 #define I40E_ITR_100K 10 /* all values below must be even */
22 #define I40E_ITR_50K 20
23 #define I40E_ITR_20K 50
24 #define I40E_ITR_18K 60
25 #define I40E_ITR_8K 122
26 #define I40E_MAX_ITR 8160 /* maximum value as per datasheet */
27 #define ITR_TO_REG(setting) ((setting) & ~I40E_ITR_DYNAMIC)
28 #define ITR_REG_ALIGN(setting) __ALIGN_MASK(setting, ~I40E_ITR_MASK)
29 #define ITR_IS_DYNAMIC(setting) (!!((setting) & I40E_ITR_DYNAMIC))
31 #define I40E_ITR_RX_DEF (I40E_ITR_20K | I40E_ITR_DYNAMIC)
32 #define I40E_ITR_TX_DEF (I40E_ITR_20K | I40E_ITR_DYNAMIC)
34 /* 0x40 is the enable bit for interrupt rate limiting, and must be set if
35 * the value of the rate limit is non-zero
37 #define INTRL_ENA BIT(6)
38 #define I40E_MAX_INTRL 0x3B /* reg uses 4 usec resolution */
39 #define INTRL_REG_TO_USEC(intrl) ((intrl & ~INTRL_ENA) << 2)
42 * i40e_intrl_usec_to_reg - convert interrupt rate limit to register
43 * @intrl: interrupt rate limit to convert
45 * This function converts a decimal interrupt rate limit to the appropriate
46 * register format expected by the firmware when setting interrupt rate limit.
48 static inline u16 i40e_intrl_usec_to_reg(int intrl)
51 return ((intrl >> 2) | INTRL_ENA);
55 #define I40E_INTRL_8K 125 /* 8000 ints/sec */
56 #define I40E_INTRL_62K 16 /* 62500 ints/sec */
57 #define I40E_INTRL_83K 12 /* 83333 ints/sec */
59 #define I40E_QUEUE_END_OF_LIST 0x7FF
61 /* this enum matches hardware bits and is meant to be used by DYN_CTLN
62 * registers and QINT registers or more generally anywhere in the manual
63 * mentioning ITR_INDX, ITR_NONE cannot be used as an index 'n' into any
64 * register but instead is a special value meaning "don't update" ITR0/1/2.
70 I40E_ITR_NONE = 3 /* ITR_NONE must not be used as an index */
73 /* these are indexes into ITRN registers */
74 #define I40E_RX_ITR I40E_IDX_ITR0
75 #define I40E_TX_ITR I40E_IDX_ITR1
76 #define I40E_PE_ITR I40E_IDX_ITR2
78 /* Supported RSS offloads */
79 #define I40E_DEFAULT_RSS_HENA ( \
80 BIT_ULL(I40E_FILTER_PCTYPE_NONF_IPV4_UDP) | \
81 BIT_ULL(I40E_FILTER_PCTYPE_NONF_IPV4_SCTP) | \
82 BIT_ULL(I40E_FILTER_PCTYPE_NONF_IPV4_TCP) | \
83 BIT_ULL(I40E_FILTER_PCTYPE_NONF_IPV4_OTHER) | \
84 BIT_ULL(I40E_FILTER_PCTYPE_FRAG_IPV4) | \
85 BIT_ULL(I40E_FILTER_PCTYPE_NONF_IPV6_UDP) | \
86 BIT_ULL(I40E_FILTER_PCTYPE_NONF_IPV6_TCP) | \
87 BIT_ULL(I40E_FILTER_PCTYPE_NONF_IPV6_SCTP) | \
88 BIT_ULL(I40E_FILTER_PCTYPE_NONF_IPV6_OTHER) | \
89 BIT_ULL(I40E_FILTER_PCTYPE_FRAG_IPV6) | \
90 BIT_ULL(I40E_FILTER_PCTYPE_L2_PAYLOAD))
92 #define I40E_DEFAULT_RSS_HENA_EXPANDED (I40E_DEFAULT_RSS_HENA | \
93 BIT_ULL(I40E_FILTER_PCTYPE_NONF_IPV4_TCP_SYN_NO_ACK) | \
94 BIT_ULL(I40E_FILTER_PCTYPE_NONF_UNICAST_IPV4_UDP) | \
95 BIT_ULL(I40E_FILTER_PCTYPE_NONF_MULTICAST_IPV4_UDP) | \
96 BIT_ULL(I40E_FILTER_PCTYPE_NONF_IPV6_TCP_SYN_NO_ACK) | \
97 BIT_ULL(I40E_FILTER_PCTYPE_NONF_UNICAST_IPV6_UDP) | \
98 BIT_ULL(I40E_FILTER_PCTYPE_NONF_MULTICAST_IPV6_UDP))
100 #define i40e_pf_get_default_rss_hena(pf) \
101 (((pf)->hw_features & I40E_HW_MULTIPLE_TCP_UDP_RSS_PCTYPE) ? \
102 I40E_DEFAULT_RSS_HENA_EXPANDED : I40E_DEFAULT_RSS_HENA)
104 /* Supported Rx Buffer Sizes (a multiple of 128) */
105 #define I40E_RXBUFFER_256 256
106 #define I40E_RXBUFFER_1536 1536 /* 128B aligned standard Ethernet frame */
107 #define I40E_RXBUFFER_2048 2048
108 #define I40E_RXBUFFER_3072 3072 /* Used for large frames w/ padding */
109 #define I40E_MAX_RXBUFFER 9728 /* largest size for single descriptor */
111 /* NOTE: netdev_alloc_skb reserves up to 64 bytes, NET_IP_ALIGN means we
112 * reserve 2 more, and skb_shared_info adds an additional 384 bytes more,
113 * this adds up to 512 bytes of extra data meaning the smallest allocation
114 * we could have is 1K.
115 * i.e. RXBUFFER_256 --> 960 byte skb (size-1024 slab)
116 * i.e. RXBUFFER_512 --> 1216 byte skb (size-2048 slab)
118 #define I40E_RX_HDR_SIZE I40E_RXBUFFER_256
119 #define I40E_PACKET_HDR_PAD (ETH_HLEN + ETH_FCS_LEN + (VLAN_HLEN * 2))
120 #define i40e_rx_desc i40e_32byte_rx_desc
122 #define I40E_RX_DMA_ATTR \
123 (DMA_ATTR_SKIP_CPU_SYNC | DMA_ATTR_WEAK_ORDERING)
125 /* Attempt to maximize the headroom available for incoming frames. We
126 * use a 2K buffer for receives and need 1536/1534 to store the data for
127 * the frame. This leaves us with 512 bytes of room. From that we need
128 * to deduct the space needed for the shared info and the padding needed
129 * to IP align the frame.
131 * Note: For cache line sizes 256 or larger this value is going to end
132 * up negative. In these cases we should fall back to the legacy
135 #if (PAGE_SIZE < 8192)
136 #define I40E_2K_TOO_SMALL_WITH_PADDING \
137 ((NET_SKB_PAD + I40E_RXBUFFER_1536) > SKB_WITH_OVERHEAD(I40E_RXBUFFER_2048))
139 static inline int i40e_compute_pad(int rx_buf_len)
141 int page_size, pad_size;
143 page_size = ALIGN(rx_buf_len, PAGE_SIZE / 2);
144 pad_size = SKB_WITH_OVERHEAD(page_size) - rx_buf_len;
149 static inline int i40e_skb_pad(void)
153 /* If a 2K buffer cannot handle a standard Ethernet frame then
154 * optimize padding for a 3K buffer instead of a 1.5K buffer.
156 * For a 3K buffer we need to add enough padding to allow for
157 * tailroom due to NET_IP_ALIGN possibly shifting us out of
158 * cache-line alignment.
160 if (I40E_2K_TOO_SMALL_WITH_PADDING)
161 rx_buf_len = I40E_RXBUFFER_3072 + SKB_DATA_ALIGN(NET_IP_ALIGN);
163 rx_buf_len = I40E_RXBUFFER_1536;
165 /* if needed make room for NET_IP_ALIGN */
166 rx_buf_len -= NET_IP_ALIGN;
168 return i40e_compute_pad(rx_buf_len);
171 #define I40E_SKB_PAD i40e_skb_pad()
173 #define I40E_2K_TOO_SMALL_WITH_PADDING false
174 #define I40E_SKB_PAD (NET_SKB_PAD + NET_IP_ALIGN)
178 * i40e_test_staterr - tests bits in Rx descriptor status and error fields
179 * @rx_desc: pointer to receive descriptor (in le64 format)
180 * @stat_err_bits: value to mask
182 * This function does some fast chicanery in order to return the
183 * value of the mask which is really only used for boolean tests.
184 * The status_error_len doesn't need to be shifted because it begins
187 static inline bool i40e_test_staterr(union i40e_rx_desc *rx_desc,
188 const u64 stat_err_bits)
190 return !!(rx_desc->wb.qword1.status_error_len &
191 cpu_to_le64(stat_err_bits));
194 /* How many Rx Buffers do we bundle into one write to the hardware ? */
195 #define I40E_RX_BUFFER_WRITE 32 /* Must be power of 2 */
196 #define I40E_RX_INCREMENT(r, i) \
199 if ((i) == (r)->count) \
201 r->next_to_clean = i; \
204 #define I40E_RX_NEXT_DESC(r, i, n) \
207 if ((i) == (r)->count) \
209 (n) = I40E_RX_DESC((r), (i)); \
212 #define I40E_RX_NEXT_DESC_PREFETCH(r, i, n) \
214 I40E_RX_NEXT_DESC((r), (i), (n)); \
218 #define I40E_MAX_BUFFER_TXD 8
219 #define I40E_MIN_TX_LEN 17
221 /* The size limit for a transmit buffer in a descriptor is (16K - 1).
222 * In order to align with the read requests we will align the value to
223 * the nearest 4K which represents our maximum read request size.
225 #define I40E_MAX_READ_REQ_SIZE 4096
226 #define I40E_MAX_DATA_PER_TXD (16 * 1024 - 1)
227 #define I40E_MAX_DATA_PER_TXD_ALIGNED \
228 (I40E_MAX_DATA_PER_TXD & ~(I40E_MAX_READ_REQ_SIZE - 1))
231 * i40e_txd_use_count - estimate the number of descriptors needed for Tx
232 * @size: transmit request size in bytes
234 * Due to hardware alignment restrictions (4K alignment), we need to
235 * assume that we can have no more than 12K of data per descriptor, even
236 * though each descriptor can take up to 16K - 1 bytes of aligned memory.
237 * Thus, we need to divide by 12K. But division is slow! Instead,
238 * we decompose the operation into shifts and one relatively cheap
239 * multiply operation.
241 * To divide by 12K, we first divide by 4K, then divide by 3:
242 * To divide by 4K, shift right by 12 bits
243 * To divide by 3, multiply by 85, then divide by 256
244 * (Divide by 256 is done by shifting right by 8 bits)
245 * Finally, we add one to round up. Because 256 isn't an exact multiple of
246 * 3, we'll underestimate near each multiple of 12K. This is actually more
247 * accurate as we have 4K - 1 of wiggle room that we can fit into the last
248 * segment. For our purposes this is accurate out to 1M which is orders of
249 * magnitude greater than our largest possible GSO size.
251 * This would then be implemented as:
252 * return (((size >> 12) * 85) >> 8) + 1;
254 * Since multiplication and division are commutative, we can reorder
256 * return ((size * 85) >> 20) + 1;
258 static inline unsigned int i40e_txd_use_count(unsigned int size)
260 return ((size * 85) >> 20) + 1;
263 /* Tx Descriptors needed, worst case */
264 #define DESC_NEEDED (MAX_SKB_FRAGS + 6)
265 #define I40E_MIN_DESC_PENDING 4
267 #define I40E_TX_FLAGS_HW_VLAN BIT(1)
268 #define I40E_TX_FLAGS_SW_VLAN BIT(2)
269 #define I40E_TX_FLAGS_TSO BIT(3)
270 #define I40E_TX_FLAGS_IPV4 BIT(4)
271 #define I40E_TX_FLAGS_IPV6 BIT(5)
272 #define I40E_TX_FLAGS_FCCRC BIT(6)
273 #define I40E_TX_FLAGS_FSO BIT(7)
274 #define I40E_TX_FLAGS_TSYN BIT(8)
275 #define I40E_TX_FLAGS_FD_SB BIT(9)
276 #define I40E_TX_FLAGS_UDP_TUNNEL BIT(10)
277 #define I40E_TX_FLAGS_VLAN_MASK 0xffff0000
278 #define I40E_TX_FLAGS_VLAN_PRIO_MASK 0xe0000000
279 #define I40E_TX_FLAGS_VLAN_PRIO_SHIFT 29
280 #define I40E_TX_FLAGS_VLAN_SHIFT 16
282 struct i40e_tx_buffer {
283 struct i40e_tx_desc *next_to_watch;
285 struct xdp_frame *xdpf;
289 unsigned int bytecount;
290 unsigned short gso_segs;
292 DEFINE_DMA_UNMAP_ADDR(dma);
293 DEFINE_DMA_UNMAP_LEN(len);
297 struct i40e_rx_buffer {
312 struct i40e_queue_stats {
317 struct i40e_tx_queue_stats {
326 struct i40e_rx_queue_stats {
328 u64 alloc_page_failed;
329 u64 alloc_buff_failed;
330 u64 page_reuse_count;
334 enum i40e_ring_state_t {
335 __I40E_TX_FDIR_INIT_DONE,
336 __I40E_TX_XPS_INIT_DONE,
337 __I40E_RING_STATE_NBITS /* must be last */
340 /* some useful defines for virtchannel interface, which
341 * is the only remaining user of header split
343 #define I40E_RX_DTYPE_NO_SPLIT 0
344 #define I40E_RX_DTYPE_HEADER_SPLIT 1
345 #define I40E_RX_DTYPE_SPLIT_ALWAYS 2
346 #define I40E_RX_SPLIT_L2 0x1
347 #define I40E_RX_SPLIT_IP 0x2
348 #define I40E_RX_SPLIT_TCP_UDP 0x4
349 #define I40E_RX_SPLIT_SCTP 0x8
351 /* struct that defines a descriptor ring, associated with a VSI */
353 struct i40e_ring *next; /* pointer to next ring in q_vector */
354 void *desc; /* Descriptor ring memory */
355 struct device *dev; /* Used for DMA mapping */
356 struct net_device *netdev; /* netdev ring maps to */
357 struct bpf_prog *xdp_prog;
359 struct i40e_tx_buffer *tx_bi;
360 struct i40e_rx_buffer *rx_bi;
362 DECLARE_BITMAP(state, __I40E_RING_STATE_NBITS);
363 u16 queue_index; /* Queue number of ring */
364 u8 dcb_tc; /* Traffic class of ring */
367 /* high bit set means dynamic, use accessor routines to read/write.
368 * hardware only supports 2us resolution for the ITR registers.
369 * these values always store the USER setting, and must be converted
370 * before programming to a register.
374 u16 count; /* Number of descriptors */
375 u16 reg_idx; /* HW register index of the ring */
378 /* used in interrupt processing */
385 bool ring_active; /* is ring online or not */
386 bool arm_wb; /* do something to arm write back */
390 #define I40E_TXR_FLAGS_WB_ON_ITR BIT(0)
391 #define I40E_RXR_FLAGS_BUILD_SKB_ENABLED BIT(1)
392 #define I40E_TXR_FLAGS_XDP BIT(2)
395 struct i40e_queue_stats stats;
396 struct u64_stats_sync syncp;
398 struct i40e_tx_queue_stats tx_stats;
399 struct i40e_rx_queue_stats rx_stats;
402 unsigned int size; /* length of descriptor ring in bytes */
403 dma_addr_t dma; /* physical address of ring */
405 struct i40e_vsi *vsi; /* Backreference to associated VSI */
406 struct i40e_q_vector *q_vector; /* Backreference to associated vector */
408 struct rcu_head rcu; /* to avoid race on free */
410 struct sk_buff *skb; /* When i40e_clean_rx_ring_irq() must
411 * return before it sees the EOP for
412 * the current packet, we save that skb
413 * here and resume receiving this
414 * packet the next time
415 * i40e_clean_rx_ring_irq() is called
419 struct i40e_channel *ch;
420 struct xdp_rxq_info xdp_rxq;
421 struct xdp_umem *xsk_umem;
422 struct zero_copy_allocator zca; /* ZC allocator anchor */
423 } ____cacheline_internodealigned_in_smp;
425 static inline bool ring_uses_build_skb(struct i40e_ring *ring)
427 return !!(ring->flags & I40E_RXR_FLAGS_BUILD_SKB_ENABLED);
430 static inline void set_ring_build_skb_enabled(struct i40e_ring *ring)
432 ring->flags |= I40E_RXR_FLAGS_BUILD_SKB_ENABLED;
435 static inline void clear_ring_build_skb_enabled(struct i40e_ring *ring)
437 ring->flags &= ~I40E_RXR_FLAGS_BUILD_SKB_ENABLED;
440 static inline bool ring_is_xdp(struct i40e_ring *ring)
442 return !!(ring->flags & I40E_TXR_FLAGS_XDP);
445 static inline void set_ring_xdp(struct i40e_ring *ring)
447 ring->flags |= I40E_TXR_FLAGS_XDP;
450 #define I40E_ITR_ADAPTIVE_MIN_INC 0x0002
451 #define I40E_ITR_ADAPTIVE_MIN_USECS 0x0002
452 #define I40E_ITR_ADAPTIVE_MAX_USECS 0x007e
453 #define I40E_ITR_ADAPTIVE_LATENCY 0x8000
454 #define I40E_ITR_ADAPTIVE_BULK 0x0000
455 #define ITR_IS_BULK(x) (!((x) & I40E_ITR_ADAPTIVE_LATENCY))
457 struct i40e_ring_container {
458 struct i40e_ring *ring; /* pointer to linked list of ring(s) */
459 unsigned long next_update; /* jiffies value of next update */
460 unsigned int total_bytes; /* total bytes processed this int */
461 unsigned int total_packets; /* total packets processed this int */
463 u16 target_itr; /* target ITR setting for ring(s) */
464 u16 current_itr; /* current ITR setting for ring(s) */
467 /* iterator for handling rings in ring container */
468 #define i40e_for_each_ring(pos, head) \
469 for (pos = (head).ring; pos != NULL; pos = pos->next)
471 static inline unsigned int i40e_rx_pg_order(struct i40e_ring *ring)
473 #if (PAGE_SIZE < 8192)
474 if (ring->rx_buf_len > (PAGE_SIZE / 2))
480 #define i40e_rx_pg_size(_ring) (PAGE_SIZE << i40e_rx_pg_order(_ring))
482 bool i40e_alloc_rx_buffers(struct i40e_ring *rxr, u16 cleaned_count);
483 netdev_tx_t i40e_lan_xmit_frame(struct sk_buff *skb, struct net_device *netdev);
484 void i40e_clean_tx_ring(struct i40e_ring *tx_ring);
485 void i40e_clean_rx_ring(struct i40e_ring *rx_ring);
486 int i40e_setup_tx_descriptors(struct i40e_ring *tx_ring);
487 int i40e_setup_rx_descriptors(struct i40e_ring *rx_ring);
488 void i40e_free_tx_resources(struct i40e_ring *tx_ring);
489 void i40e_free_rx_resources(struct i40e_ring *rx_ring);
490 int i40e_napi_poll(struct napi_struct *napi, int budget);
491 void i40e_force_wb(struct i40e_vsi *vsi, struct i40e_q_vector *q_vector);
492 u32 i40e_get_tx_pending(struct i40e_ring *ring, bool in_sw);
493 void i40e_detect_recover_hung(struct i40e_vsi *vsi);
494 int __i40e_maybe_stop_tx(struct i40e_ring *tx_ring, int size);
495 bool __i40e_chk_linearize(struct sk_buff *skb);
496 int i40e_xdp_xmit(struct net_device *dev, int n, struct xdp_frame **frames,
500 * i40e_get_head - Retrieve head from head writeback
501 * @tx_ring: tx ring to fetch head of
503 * Returns value of Tx ring head based on value stored
504 * in head write-back location
506 static inline u32 i40e_get_head(struct i40e_ring *tx_ring)
508 void *head = (struct i40e_tx_desc *)tx_ring->desc + tx_ring->count;
510 return le32_to_cpu(*(volatile __le32 *)head);
514 * i40e_xmit_descriptor_count - calculate number of Tx descriptors needed
516 * @tx_ring: ring to send buffer on
518 * Returns number of data descriptors needed for this skb. Returns 0 to indicate
519 * there is not enough descriptors available in this ring since we need at least
522 static inline int i40e_xmit_descriptor_count(struct sk_buff *skb)
524 const skb_frag_t *frag = &skb_shinfo(skb)->frags[0];
525 unsigned int nr_frags = skb_shinfo(skb)->nr_frags;
526 int count = 0, size = skb_headlen(skb);
529 count += i40e_txd_use_count(size);
534 size = skb_frag_size(frag++);
541 * i40e_maybe_stop_tx - 1st level check for Tx stop conditions
542 * @tx_ring: the ring to be checked
543 * @size: the size buffer we want to assure is available
545 * Returns 0 if stop is not needed
547 static inline int i40e_maybe_stop_tx(struct i40e_ring *tx_ring, int size)
549 if (likely(I40E_DESC_UNUSED(tx_ring) >= size))
551 return __i40e_maybe_stop_tx(tx_ring, size);
555 * i40e_chk_linearize - Check if there are more than 8 fragments per packet
557 * @count: number of buffers used
559 * Note: Our HW can't scatter-gather more than 8 fragments to build
560 * a packet on the wire and so we need to figure out the cases where we
561 * need to linearize the skb.
563 static inline bool i40e_chk_linearize(struct sk_buff *skb, int count)
565 /* Both TSO and single send will work if count is less than 8 */
566 if (likely(count < I40E_MAX_BUFFER_TXD))
570 return __i40e_chk_linearize(skb);
572 /* we can support up to 8 data buffers for a single send */
573 return count != I40E_MAX_BUFFER_TXD;
577 * txring_txq - Find the netdev Tx ring based on the i40e Tx ring
578 * @ring: Tx ring to find the netdev equivalent of
580 static inline struct netdev_queue *txring_txq(const struct i40e_ring *ring)
582 return netdev_get_tx_queue(ring->netdev, ring->queue_index);
584 #endif /* _I40E_TXRX_H_ */