2 * This file is based on code from OCTEON SDK by Cavium Networks.
4 * Copyright (c) 2003-2010 Cavium Networks
6 * This file is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License, Version 2, as
8 * published by the Free Software Foundation.
11 #include <linux/module.h>
12 #include <linux/kernel.h>
13 #include <linux/cache.h>
14 #include <linux/cpumask.h>
15 #include <linux/netdevice.h>
16 #include <linux/etherdevice.h>
18 #include <linux/string.h>
19 #include <linux/prefetch.h>
20 #include <linux/ratelimit.h>
21 #include <linux/smp.h>
22 #include <linux/interrupt.h>
25 #include <linux/xfrm.h>
27 #endif /* CONFIG_XFRM */
29 #include <asm/octeon/octeon.h>
31 #include "ethernet-defines.h"
32 #include "ethernet-mem.h"
33 #include "ethernet-rx.h"
34 #include "octeon-ethernet.h"
35 #include "ethernet-util.h"
37 #include <asm/octeon/cvmx-helper.h>
38 #include <asm/octeon/cvmx-wqe.h>
39 #include <asm/octeon/cvmx-fau.h>
40 #include <asm/octeon/cvmx-pow.h>
41 #include <asm/octeon/cvmx-pip.h>
42 #include <asm/octeon/cvmx-scratch.h>
44 #include <asm/octeon/cvmx-gmxx-defs.h>
46 static atomic_t oct_rx_ready
= ATOMIC_INIT(0);
48 static struct oct_rx_group
{
51 struct napi_struct napi
;
55 * cvm_oct_do_interrupt - interrupt handler.
56 * @irq: Interrupt number.
57 * @napi_id: Cookie to identify the NAPI instance.
59 * The interrupt occurs whenever the POW has packets in our group.
62 static irqreturn_t
cvm_oct_do_interrupt(int irq
, void *napi_id
)
64 /* Disable the IRQ and start napi_poll. */
65 disable_irq_nosync(irq
);
66 napi_schedule(napi_id
);
72 * cvm_oct_check_rcv_error - process receive errors
73 * @work: Work queue entry pointing to the packet.
75 * Returns Non-zero if the packet can be dropped, zero otherwise.
77 static inline int cvm_oct_check_rcv_error(cvmx_wqe_t
*work
)
81 if (octeon_has_feature(OCTEON_FEATURE_PKND
))
82 port
= work
->word0
.pip
.cn68xx
.pknd
;
84 port
= work
->word1
.cn38xx
.ipprt
;
86 if ((work
->word2
.snoip
.err_code
== 10) && (work
->word1
.len
<= 64)) {
88 * Ignore length errors on min size packets. Some
89 * equipment incorrectly pads packets to 64+4FCS
90 * instead of 60+4FCS. Note these packets still get
91 * counted as frame errors.
93 } else if (work
->word2
.snoip
.err_code
== 5 ||
94 work
->word2
.snoip
.err_code
== 7) {
96 * We received a packet with either an alignment error
97 * or a FCS error. This may be signalling that we are
98 * running 10Mbps with GMXX_RXX_FRM_CTL[PRE_CHK]
99 * off. If this is the case we need to parse the
100 * packet to determine if we can remove a non spec
101 * preamble and generate a correct packet.
103 int interface
= cvmx_helper_get_interface_num(port
);
104 int index
= cvmx_helper_get_interface_index_num(port
);
105 union cvmx_gmxx_rxx_frm_ctl gmxx_rxx_frm_ctl
;
107 gmxx_rxx_frm_ctl
.u64
=
108 cvmx_read_csr(CVMX_GMXX_RXX_FRM_CTL(index
, interface
));
109 if (gmxx_rxx_frm_ctl
.s
.pre_chk
== 0) {
111 cvmx_phys_to_ptr(work
->packet_ptr
.s
.addr
);
114 while (i
< work
->word1
.len
- 1) {
122 /* Port received 0xd5 preamble */
123 work
->packet_ptr
.s
.addr
+= i
+ 1;
124 work
->word1
.len
-= i
+ 5;
125 } else if ((*ptr
& 0xf) == 0xd) {
126 /* Port received 0xd preamble */
127 work
->packet_ptr
.s
.addr
+= i
;
128 work
->word1
.len
-= i
+ 4;
129 for (i
= 0; i
< work
->word1
.len
; i
++) {
131 ((*ptr
& 0xf0) >> 4) |
132 ((*(ptr
+ 1) & 0xf) << 4);
136 printk_ratelimited("Port %d unknown preamble, packet dropped\n",
138 cvm_oct_free_work(work
);
143 printk_ratelimited("Port %d receive error code %d, packet dropped\n",
144 port
, work
->word2
.snoip
.err_code
);
145 cvm_oct_free_work(work
);
152 static int cvm_oct_poll(struct oct_rx_group
*rx_group
, int budget
)
154 const int coreid
= cvmx_get_core_num();
158 int did_work_request
= 0;
159 int packet_not_copied
;
161 /* Prefetch cvm_oct_device since we know we need it soon */
162 prefetch(cvm_oct_device
);
164 if (USE_ASYNC_IOBDMA
) {
165 /* Save scratch in case userspace is using it */
167 old_scratch
= cvmx_scratch_read64(CVMX_SCR_SCRATCH
);
170 /* Only allow work for our group (and preserve priorities) */
171 if (OCTEON_IS_MODEL(OCTEON_CN68XX
)) {
172 old_group_mask
= cvmx_read_csr(CVMX_SSO_PPX_GRP_MSK(coreid
));
173 cvmx_write_csr(CVMX_SSO_PPX_GRP_MSK(coreid
),
174 BIT(rx_group
->group
));
175 cvmx_read_csr(CVMX_SSO_PPX_GRP_MSK(coreid
)); /* Flush */
177 old_group_mask
= cvmx_read_csr(CVMX_POW_PP_GRP_MSKX(coreid
));
178 cvmx_write_csr(CVMX_POW_PP_GRP_MSKX(coreid
),
179 (old_group_mask
& ~0xFFFFull
) |
180 BIT(rx_group
->group
));
183 if (USE_ASYNC_IOBDMA
) {
184 cvmx_pow_work_request_async(CVMX_SCR_SCRATCH
, CVMX_POW_NO_WAIT
);
185 did_work_request
= 1;
188 while (rx_count
< budget
) {
189 struct sk_buff
*skb
= NULL
;
190 struct sk_buff
**pskb
= NULL
;
195 if (USE_ASYNC_IOBDMA
&& did_work_request
)
196 work
= cvmx_pow_work_response_async(CVMX_SCR_SCRATCH
);
198 work
= cvmx_pow_work_request_sync(CVMX_POW_NO_WAIT
);
201 did_work_request
= 0;
203 if (OCTEON_IS_MODEL(OCTEON_CN68XX
)) {
204 cvmx_write_csr(CVMX_SSO_WQ_IQ_DIS
,
205 BIT(rx_group
->group
));
206 cvmx_write_csr(CVMX_SSO_WQ_INT
,
207 BIT(rx_group
->group
));
209 union cvmx_pow_wq_int wq_int
;
212 wq_int
.s
.iq_dis
= BIT(rx_group
->group
);
213 wq_int
.s
.wq_int
= BIT(rx_group
->group
);
214 cvmx_write_csr(CVMX_POW_WQ_INT
, wq_int
.u64
);
218 pskb
= (struct sk_buff
**)
219 (cvm_oct_get_buffer_ptr(work
->packet_ptr
) -
223 if (USE_ASYNC_IOBDMA
&& rx_count
< (budget
- 1)) {
224 cvmx_pow_work_request_async_nocheck(CVMX_SCR_SCRATCH
,
226 did_work_request
= 1;
230 skb_in_hw
= work
->word2
.s
.bufs
== 1;
231 if (likely(skb_in_hw
)) {
233 prefetch(&skb
->head
);
237 if (octeon_has_feature(OCTEON_FEATURE_PKND
))
238 port
= work
->word0
.pip
.cn68xx
.pknd
;
240 port
= work
->word1
.cn38xx
.ipprt
;
242 prefetch(cvm_oct_device
[port
]);
244 /* Immediately throw away all packets with receive errors */
245 if (unlikely(work
->word2
.snoip
.rcv_error
)) {
246 if (cvm_oct_check_rcv_error(work
))
251 * We can only use the zero copy path if skbuffs are
252 * in the FPA pool and the packet fits in a single
255 if (likely(skb_in_hw
)) {
256 skb
->data
= skb
->head
+ work
->packet_ptr
.s
.addr
-
257 cvmx_ptr_to_phys(skb
->head
);
259 skb
->len
= work
->word1
.len
;
260 skb_set_tail_pointer(skb
, skb
->len
);
261 packet_not_copied
= 1;
264 * We have to copy the packet. First allocate
267 skb
= dev_alloc_skb(work
->word1
.len
);
269 cvm_oct_free_work(work
);
274 * Check if we've received a packet that was
275 * entirely stored in the work entry.
277 if (unlikely(work
->word2
.s
.bufs
== 0)) {
278 u8
*ptr
= work
->packet_data
;
280 if (likely(!work
->word2
.s
.not_IP
)) {
282 * The beginning of the packet
283 * moves for IP packets.
285 if (work
->word2
.s
.is_v6
)
290 memcpy(skb_put(skb
, work
->word1
.len
), ptr
,
292 /* No packet buffers to free */
294 int segments
= work
->word2
.s
.bufs
;
295 union cvmx_buf_ptr segment_ptr
=
297 int len
= work
->word1
.len
;
300 union cvmx_buf_ptr next_ptr
=
301 *(union cvmx_buf_ptr
*)
303 segment_ptr
.s
.addr
- 8);
306 * Octeon Errata PKI-100: The segment size is
307 * wrong. Until it is fixed, calculate the
308 * segment size based on the packet pool
309 * buffer size. When it is fixed, the
310 * following line should be replaced with this
311 * one: int segment_size =
312 * segment_ptr.s.size;
315 CVMX_FPA_PACKET_POOL_SIZE
-
316 (segment_ptr
.s
.addr
-
317 (((segment_ptr
.s
.addr
>> 7) -
318 segment_ptr
.s
.back
) << 7));
320 * Don't copy more than what
321 * is left in the packet.
323 if (segment_size
> len
)
325 /* Copy the data into the packet */
326 memcpy(skb_put(skb
, segment_size
),
331 segment_ptr
= next_ptr
;
334 packet_not_copied
= 0;
336 if (likely((port
< TOTAL_NUMBER_OF_PORTS
) &&
337 cvm_oct_device
[port
])) {
338 struct net_device
*dev
= cvm_oct_device
[port
];
341 * Only accept packets for devices that are
344 if (likely(dev
->flags
& IFF_UP
)) {
345 skb
->protocol
= eth_type_trans(skb
, dev
);
348 if (unlikely(work
->word2
.s
.not_IP
||
349 work
->word2
.s
.IP_exc
||
350 work
->word2
.s
.L4_error
||
351 !work
->word2
.s
.tcp_or_udp
))
352 skb
->ip_summed
= CHECKSUM_NONE
;
354 skb
->ip_summed
= CHECKSUM_UNNECESSARY
;
356 /* Increment RX stats for virtual ports */
357 if (port
>= CVMX_PIP_NUM_INPUT_PORTS
) {
358 dev
->stats
.rx_packets
++;
359 dev
->stats
.rx_bytes
+= skb
->len
;
361 netif_receive_skb(skb
);
364 * Drop any packet received for a device that
367 dev
->stats
.rx_dropped
++;
368 dev_kfree_skb_irq(skb
);
372 * Drop any packet received for a device that
375 printk_ratelimited("Port %d not controlled by Linux, packet dropped\n",
377 dev_kfree_skb_irq(skb
);
380 * Check to see if the skbuff and work share the same
383 if (likely(packet_not_copied
)) {
385 * This buffer needs to be replaced, increment
386 * the number of buffers we need to free by
389 cvmx_fau_atomic_add32(FAU_NUM_PACKET_BUFFERS_TO_FREE
,
392 cvmx_fpa_free(work
, CVMX_FPA_WQE_POOL
, 1);
394 cvm_oct_free_work(work
);
397 /* Restore the original POW group mask */
398 if (OCTEON_IS_MODEL(OCTEON_CN68XX
)) {
399 cvmx_write_csr(CVMX_SSO_PPX_GRP_MSK(coreid
), old_group_mask
);
400 cvmx_read_csr(CVMX_SSO_PPX_GRP_MSK(coreid
)); /* Flush */
402 cvmx_write_csr(CVMX_POW_PP_GRP_MSKX(coreid
), old_group_mask
);
405 if (USE_ASYNC_IOBDMA
) {
406 /* Restore the scratch area */
407 cvmx_scratch_write64(CVMX_SCR_SCRATCH
, old_scratch
);
409 cvm_oct_rx_refill_pool(0);
415 * cvm_oct_napi_poll - the NAPI poll function.
416 * @napi: The NAPI instance.
417 * @budget: Maximum number of packets to receive.
419 * Returns the number of packets processed.
421 static int cvm_oct_napi_poll(struct napi_struct
*napi
, int budget
)
423 struct oct_rx_group
*rx_group
= container_of(napi
, struct oct_rx_group
,
427 rx_count
= cvm_oct_poll(rx_group
, budget
);
429 if (rx_count
< budget
) {
431 napi_complete_done(napi
, rx_count
);
432 enable_irq(rx_group
->irq
);
437 #ifdef CONFIG_NET_POLL_CONTROLLER
439 * cvm_oct_poll_controller - poll for receive packets
442 * @dev: Device to poll. Unused
444 void cvm_oct_poll_controller(struct net_device
*dev
)
448 if (!atomic_read(&oct_rx_ready
))
451 for (i
= 0; i
< ARRAY_SIZE(oct_rx_group
); i
++) {
452 if (!(pow_receive_groups
& BIT(i
)))
455 cvm_oct_poll(&oct_rx_group
[i
], 16);
460 void cvm_oct_rx_initialize(void)
463 struct net_device
*dev_for_napi
= NULL
;
465 for (i
= 0; i
< TOTAL_NUMBER_OF_PORTS
; i
++) {
466 if (cvm_oct_device
[i
]) {
467 dev_for_napi
= cvm_oct_device
[i
];
473 panic("No net_devices were allocated.");
475 for (i
= 0; i
< ARRAY_SIZE(oct_rx_group
); i
++) {
478 if (!(pow_receive_groups
& BIT(i
)))
481 netif_napi_add(dev_for_napi
, &oct_rx_group
[i
].napi
,
482 cvm_oct_napi_poll
, rx_napi_weight
);
483 napi_enable(&oct_rx_group
[i
].napi
);
485 oct_rx_group
[i
].irq
= OCTEON_IRQ_WORKQ0
+ i
;
486 oct_rx_group
[i
].group
= i
;
488 /* Register an IRQ handler to receive POW interrupts */
489 ret
= request_irq(oct_rx_group
[i
].irq
, cvm_oct_do_interrupt
, 0,
490 "Ethernet", &oct_rx_group
[i
].napi
);
492 panic("Could not acquire Ethernet IRQ %d\n",
493 oct_rx_group
[i
].irq
);
495 disable_irq_nosync(oct_rx_group
[i
].irq
);
497 /* Enable POW interrupt when our port has at least one packet */
498 if (OCTEON_IS_MODEL(OCTEON_CN68XX
)) {
499 union cvmx_sso_wq_int_thrx int_thr
;
500 union cvmx_pow_wq_int_pc int_pc
;
504 int_thr
.s
.tc_thr
= 1;
505 cvmx_write_csr(CVMX_SSO_WQ_INT_THRX(i
), int_thr
.u64
);
509 cvmx_write_csr(CVMX_SSO_WQ_INT_PC
, int_pc
.u64
);
511 union cvmx_pow_wq_int_thrx int_thr
;
512 union cvmx_pow_wq_int_pc int_pc
;
516 int_thr
.s
.tc_thr
= 1;
517 cvmx_write_csr(CVMX_POW_WQ_INT_THRX(i
), int_thr
.u64
);
521 cvmx_write_csr(CVMX_POW_WQ_INT_PC
, int_pc
.u64
);
524 /* Schedule NAPI now. This will indirectly enable the
527 napi_schedule(&oct_rx_group
[i
].napi
);
529 atomic_inc(&oct_rx_ready
);
532 void cvm_oct_rx_shutdown(void)
536 for (i
= 0; i
< ARRAY_SIZE(oct_rx_group
); i
++) {
537 if (!(pow_receive_groups
& BIT(i
)))
540 /* Disable POW interrupt */
541 if (OCTEON_IS_MODEL(OCTEON_CN68XX
))
542 cvmx_write_csr(CVMX_SSO_WQ_INT_THRX(i
), 0);
544 cvmx_write_csr(CVMX_POW_WQ_INT_THRX(i
), 0);
546 /* Free the interrupt handler */
547 free_irq(oct_rx_group
[i
].irq
, cvm_oct_device
);
549 netif_napi_del(&oct_rx_group
[i
].napi
);