xref: /linux/drivers/net/ethernet/intel/libeth/rx.c (revision cdd30ebb1b9f36159d66f088b61aee264e649d7a)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /* Copyright (C) 2024 Intel Corporation */
3 
4 #include <net/libeth/rx.h>
5 
6 /* Rx buffer management */
7 
8 /**
9  * libeth_rx_hw_len_mtu - get the actual buffer size to be passed to HW
10  * @pp: &page_pool_params of the netdev to calculate the size for
11  * @max_len: maximum buffer size for a single descriptor
12  *
13  * Return: HW-writeable length per one buffer to pass it to the HW accounting:
14  * MTU the @dev has, HW required alignment, minimum and maximum allowed values,
15  * and system's page size.
16  */
libeth_rx_hw_len_mtu(const struct page_pool_params * pp,u32 max_len)17 static u32 libeth_rx_hw_len_mtu(const struct page_pool_params *pp, u32 max_len)
18 {
19 	u32 len;
20 
21 	len = READ_ONCE(pp->netdev->mtu) + LIBETH_RX_LL_LEN;
22 	len = ALIGN(len, LIBETH_RX_BUF_STRIDE);
23 	len = min3(len, ALIGN_DOWN(max_len ? : U32_MAX, LIBETH_RX_BUF_STRIDE),
24 		   pp->max_len);
25 
26 	return len;
27 }
28 
29 /**
30  * libeth_rx_hw_len_truesize - get the short buffer size to be passed to HW
31  * @pp: &page_pool_params of the netdev to calculate the size for
32  * @max_len: maximum buffer size for a single descriptor
33  * @truesize: desired truesize for the buffers
34  *
35  * Return: HW-writeable length per one buffer to pass it to the HW ignoring the
36  * MTU and closest to the passed truesize. Can be used for "short" buffer
37  * queues to fragment pages more efficiently.
38  */
libeth_rx_hw_len_truesize(const struct page_pool_params * pp,u32 max_len,u32 truesize)39 static u32 libeth_rx_hw_len_truesize(const struct page_pool_params *pp,
40 				     u32 max_len, u32 truesize)
41 {
42 	u32 min, len;
43 
44 	min = SKB_HEAD_ALIGN(pp->offset + LIBETH_RX_BUF_STRIDE);
45 	truesize = clamp(roundup_pow_of_two(truesize), roundup_pow_of_two(min),
46 			 PAGE_SIZE << LIBETH_RX_PAGE_ORDER);
47 
48 	len = SKB_WITH_OVERHEAD(truesize - pp->offset);
49 	len = ALIGN_DOWN(len, LIBETH_RX_BUF_STRIDE) ? : LIBETH_RX_BUF_STRIDE;
50 	len = min3(len, ALIGN_DOWN(max_len ? : U32_MAX, LIBETH_RX_BUF_STRIDE),
51 		   pp->max_len);
52 
53 	return len;
54 }
55 
56 /**
57  * libeth_rx_page_pool_params - calculate params with the stack overhead
58  * @fq: buffer queue to calculate the size for
59  * @pp: &page_pool_params of the netdev
60  *
61  * Set the PP params to will all needed stack overhead (headroom, tailroom) and
62  * both the HW buffer length and the truesize for all types of buffers. For
63  * "short" buffers, truesize never exceeds the "wanted" one; for the rest,
64  * it can be up to the page size.
65  *
66  * Return: true on success, false on invalid input params.
67  */
libeth_rx_page_pool_params(struct libeth_fq * fq,struct page_pool_params * pp)68 static bool libeth_rx_page_pool_params(struct libeth_fq *fq,
69 				       struct page_pool_params *pp)
70 {
71 	pp->offset = LIBETH_SKB_HEADROOM;
72 	/* HW-writeable / syncable length per one page */
73 	pp->max_len = LIBETH_RX_PAGE_LEN(pp->offset);
74 
75 	/* HW-writeable length per buffer */
76 	switch (fq->type) {
77 	case LIBETH_FQE_MTU:
78 		fq->buf_len = libeth_rx_hw_len_mtu(pp, fq->buf_len);
79 		break;
80 	case LIBETH_FQE_SHORT:
81 		fq->buf_len = libeth_rx_hw_len_truesize(pp, fq->buf_len,
82 							fq->truesize);
83 		break;
84 	case LIBETH_FQE_HDR:
85 		fq->buf_len = ALIGN(LIBETH_MAX_HEAD, LIBETH_RX_BUF_STRIDE);
86 		break;
87 	default:
88 		return false;
89 	}
90 
91 	/* Buffer size to allocate */
92 	fq->truesize = roundup_pow_of_two(SKB_HEAD_ALIGN(pp->offset +
93 							 fq->buf_len));
94 
95 	return true;
96 }
97 
98 /**
99  * libeth_rx_page_pool_params_zc - calculate params without the stack overhead
100  * @fq: buffer queue to calculate the size for
101  * @pp: &page_pool_params of the netdev
102  *
103  * Set the PP params to exclude the stack overhead and both the buffer length
104  * and the truesize, which are equal for the data buffers. Note that this
105  * requires separate header buffers to be always active and account the
106  * overhead.
107  * With the MTU == ``PAGE_SIZE``, this allows the kernel to enable the zerocopy
108  * mode.
109  *
110  * Return: true on success, false on invalid input params.
111  */
libeth_rx_page_pool_params_zc(struct libeth_fq * fq,struct page_pool_params * pp)112 static bool libeth_rx_page_pool_params_zc(struct libeth_fq *fq,
113 					  struct page_pool_params *pp)
114 {
115 	u32 mtu, max;
116 
117 	pp->offset = 0;
118 	pp->max_len = PAGE_SIZE << LIBETH_RX_PAGE_ORDER;
119 
120 	switch (fq->type) {
121 	case LIBETH_FQE_MTU:
122 		mtu = READ_ONCE(pp->netdev->mtu);
123 		break;
124 	case LIBETH_FQE_SHORT:
125 		mtu = fq->truesize;
126 		break;
127 	default:
128 		return false;
129 	}
130 
131 	mtu = roundup_pow_of_two(mtu);
132 	max = min(rounddown_pow_of_two(fq->buf_len ? : U32_MAX),
133 		  pp->max_len);
134 
135 	fq->buf_len = clamp(mtu, LIBETH_RX_BUF_STRIDE, max);
136 	fq->truesize = fq->buf_len;
137 
138 	return true;
139 }
140 
141 /**
142  * libeth_rx_fq_create - create a PP with the default libeth settings
143  * @fq: buffer queue struct to fill
144  * @napi: &napi_struct covering this PP (no usage outside its poll loops)
145  *
146  * Return: %0 on success, -%errno on failure.
147  */
libeth_rx_fq_create(struct libeth_fq * fq,struct napi_struct * napi)148 int libeth_rx_fq_create(struct libeth_fq *fq, struct napi_struct *napi)
149 {
150 	struct page_pool_params pp = {
151 		.flags		= PP_FLAG_DMA_MAP | PP_FLAG_DMA_SYNC_DEV,
152 		.order		= LIBETH_RX_PAGE_ORDER,
153 		.pool_size	= fq->count,
154 		.nid		= fq->nid,
155 		.dev		= napi->dev->dev.parent,
156 		.netdev		= napi->dev,
157 		.napi		= napi,
158 		.dma_dir	= DMA_FROM_DEVICE,
159 	};
160 	struct libeth_fqe *fqes;
161 	struct page_pool *pool;
162 	bool ret;
163 
164 	if (!fq->hsplit)
165 		ret = libeth_rx_page_pool_params(fq, &pp);
166 	else
167 		ret = libeth_rx_page_pool_params_zc(fq, &pp);
168 	if (!ret)
169 		return -EINVAL;
170 
171 	pool = page_pool_create(&pp);
172 	if (IS_ERR(pool))
173 		return PTR_ERR(pool);
174 
175 	fqes = kvcalloc_node(fq->count, sizeof(*fqes), GFP_KERNEL, fq->nid);
176 	if (!fqes)
177 		goto err_buf;
178 
179 	fq->fqes = fqes;
180 	fq->pp = pool;
181 
182 	return 0;
183 
184 err_buf:
185 	page_pool_destroy(pool);
186 
187 	return -ENOMEM;
188 }
189 EXPORT_SYMBOL_NS_GPL(libeth_rx_fq_create, "LIBETH");
190 
191 /**
192  * libeth_rx_fq_destroy - destroy a &page_pool created by libeth
193  * @fq: buffer queue to process
194  */
libeth_rx_fq_destroy(struct libeth_fq * fq)195 void libeth_rx_fq_destroy(struct libeth_fq *fq)
196 {
197 	kvfree(fq->fqes);
198 	page_pool_destroy(fq->pp);
199 }
200 EXPORT_SYMBOL_NS_GPL(libeth_rx_fq_destroy, "LIBETH");
201 
202 /**
203  * libeth_rx_recycle_slow - recycle a libeth page from the NAPI context
204  * @page: page to recycle
205  *
206  * To be used on exceptions or rare cases not requiring fast inline recycling.
207  */
libeth_rx_recycle_slow(struct page * page)208 void libeth_rx_recycle_slow(struct page *page)
209 {
210 	page_pool_recycle_direct(page->pp, page);
211 }
212 EXPORT_SYMBOL_NS_GPL(libeth_rx_recycle_slow, "LIBETH");
213 
214 /* Converting abstract packet type numbers into a software structure with
215  * the packet parameters to do O(1) lookup on Rx.
216  */
217 
218 static const u16 libeth_rx_pt_xdp_oip[] = {
219 	[LIBETH_RX_PT_OUTER_L2]		= XDP_RSS_TYPE_NONE,
220 	[LIBETH_RX_PT_OUTER_IPV4]	= XDP_RSS_L3_IPV4,
221 	[LIBETH_RX_PT_OUTER_IPV6]	= XDP_RSS_L3_IPV6,
222 };
223 
224 static const u16 libeth_rx_pt_xdp_iprot[] = {
225 	[LIBETH_RX_PT_INNER_NONE]	= XDP_RSS_TYPE_NONE,
226 	[LIBETH_RX_PT_INNER_UDP]	= XDP_RSS_L4_UDP,
227 	[LIBETH_RX_PT_INNER_TCP]	= XDP_RSS_L4_TCP,
228 	[LIBETH_RX_PT_INNER_SCTP]	= XDP_RSS_L4_SCTP,
229 	[LIBETH_RX_PT_INNER_ICMP]	= XDP_RSS_L4_ICMP,
230 	[LIBETH_RX_PT_INNER_TIMESYNC]	= XDP_RSS_TYPE_NONE,
231 };
232 
233 static const u16 libeth_rx_pt_xdp_pl[] = {
234 	[LIBETH_RX_PT_PAYLOAD_NONE]	= XDP_RSS_TYPE_NONE,
235 	[LIBETH_RX_PT_PAYLOAD_L2]	= XDP_RSS_TYPE_NONE,
236 	[LIBETH_RX_PT_PAYLOAD_L3]	= XDP_RSS_TYPE_NONE,
237 	[LIBETH_RX_PT_PAYLOAD_L4]	= XDP_RSS_L4,
238 };
239 
240 /**
241  * libeth_rx_pt_gen_hash_type - generate an XDP RSS hash type for a PT
242  * @pt: PT structure to evaluate
243  *
244  * Generates ```hash_type``` field with XDP RSS type values from the parsed
245  * packet parameters if they're obtained dynamically at runtime.
246  */
libeth_rx_pt_gen_hash_type(struct libeth_rx_pt * pt)247 void libeth_rx_pt_gen_hash_type(struct libeth_rx_pt *pt)
248 {
249 	pt->hash_type = 0;
250 	pt->hash_type |= libeth_rx_pt_xdp_oip[pt->outer_ip];
251 	pt->hash_type |= libeth_rx_pt_xdp_iprot[pt->inner_prot];
252 	pt->hash_type |= libeth_rx_pt_xdp_pl[pt->payload_layer];
253 }
254 EXPORT_SYMBOL_NS_GPL(libeth_rx_pt_gen_hash_type, "LIBETH");
255 
256 /* Module */
257 
258 MODULE_DESCRIPTION("Common Ethernet library");
259 MODULE_LICENSE("GPL");
260