projects / deliverable / linux.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
blame | history | raw | HEAD
sfc: Add and use MCDI_SET_QWORD() and MCDI_SET_ARRAY_QWORD()
[deliverable/linux.git] / drivers / net / ethernet / sfc / siena_sriov.c
1 /****************************************************************************
2 * Driver for Solarflare Solarstorm network controllers and boards
3 * Copyright 2010-2011 Solarflare Communications Inc.
4 *
5 * This program is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 as published
7 * by the Free Software Foundation, incorporated herein by reference.
8 */
9 #include <linux/pci.h>
10 #include <linux/module.h>
11 #include "net_driver.h"
12 #include "efx.h"
13 #include "nic.h"
14 #include "io.h"
15 #include "mcdi.h"
16 #include "filter.h"
17 #include "mcdi_pcol.h"
18 #include "regs.h"
19 #include "vfdi.h"
20
21 /* Number of longs required to track all the VIs in a VF */
22 #define VI_MASK_LENGTH BITS_TO_LONGS(1 << EFX_VI_SCALE_MAX)
23
24 /* Maximum number of RX queues supported */
25 #define VF_MAX_RX_QUEUES 63
26
27 /**
28 * enum efx_vf_tx_filter_mode - TX MAC filtering behaviour
29 * @VF_TX_FILTER_OFF: Disabled
30 * @VF_TX_FILTER_AUTO: Enabled if MAC address assigned to VF and only
31 * 2 TX queues allowed per VF.
32 * @VF_TX_FILTER_ON: Enabled
33 */
34 enum efx_vf_tx_filter_mode {
35 VF_TX_FILTER_OFF,
36 VF_TX_FILTER_AUTO,
37 VF_TX_FILTER_ON,
38 };
39
40 /**
41 * struct efx_vf - Back-end resource and protocol state for a PCI VF
42 * @efx: The Efx NIC owning this VF
43 * @pci_rid: The PCI requester ID for this VF
44 * @pci_name: The PCI name (formatted address) of this VF
45 * @index: Index of VF within its port and PF.
46 * @req: VFDI incoming request work item. Incoming USR_EV events are received
47 * by the NAPI handler, but must be handled by executing MCDI requests
48 * inside a work item.
49 * @req_addr: VFDI incoming request DMA address (in VF's PCI address space).
50 * @req_type: Expected next incoming (from VF) %VFDI_EV_TYPE member.
51 * @req_seqno: Expected next incoming (from VF) %VFDI_EV_SEQ member.
52 * @msg_seqno: Next %VFDI_EV_SEQ member to reply to VF. Protected by
53 * @status_lock
54 * @busy: VFDI request queued to be processed or being processed. Receiving
55 * a VFDI request when @busy is set is an error condition.
56 * @buf: Incoming VFDI requests are DMA from the VF into this buffer.
57 * @buftbl_base: Buffer table entries for this VF start at this index.
58 * @rx_filtering: Receive filtering has been requested by the VF driver.
59 * @rx_filter_flags: The flags sent in the %VFDI_OP_INSERT_FILTER request.
60 * @rx_filter_qid: VF relative qid for RX filter requested by VF.
61 * @rx_filter_id: Receive MAC filter ID. Only one filter per VF is supported.
62 * @tx_filter_mode: Transmit MAC filtering mode.
63 * @tx_filter_id: Transmit MAC filter ID.
64 * @addr: The MAC address and outer vlan tag of the VF.
65 * @status_addr: VF DMA address of page for &struct vfdi_status updates.
66 * @status_lock: Mutex protecting @msg_seqno, @status_addr, @addr,
67 * @peer_page_addrs and @peer_page_count from simultaneous
68 * updates by the VM and consumption by
69 * efx_sriov_update_vf_addr()
70 * @peer_page_addrs: Pointer to an array of guest pages for local addresses.
71 * @peer_page_count: Number of entries in @peer_page_count.
72 * @evq0_addrs: Array of guest pages backing evq0.
73 * @evq0_count: Number of entries in @evq0_addrs.
74 * @flush_waitq: wait queue used by %VFDI_OP_FINI_ALL_QUEUES handler
75 * to wait for flush completions.
76 * @txq_lock: Mutex for TX queue allocation.
77 * @txq_mask: Mask of initialized transmit queues.
78 * @txq_count: Number of initialized transmit queues.
79 * @rxq_mask: Mask of initialized receive queues.
80 * @rxq_count: Number of initialized receive queues.
81 * @rxq_retry_mask: Mask or receive queues that need to be flushed again
82 * due to flush failure.
83 * @rxq_retry_count: Number of receive queues in @rxq_retry_mask.
84 * @reset_work: Work item to schedule a VF reset.
85 */
86 struct efx_vf {
87 struct efx_nic *efx;
88 unsigned int pci_rid;
89 char pci_name[13]; /* dddd:bb:dd.f */
90 unsigned int index;
91 struct work_struct req;
92 u64 req_addr;
93 int req_type;
94 unsigned req_seqno;
95 unsigned msg_seqno;
96 bool busy;
97 struct efx_buffer buf;
98 unsigned buftbl_base;
99 bool rx_filtering;
100 enum efx_filter_flags rx_filter_flags;
101 unsigned rx_filter_qid;
102 int rx_filter_id;
103 enum efx_vf_tx_filter_mode tx_filter_mode;
104 int tx_filter_id;
105 struct vfdi_endpoint addr;
106 u64 status_addr;
107 struct mutex status_lock;
108 u64 *peer_page_addrs;
109 unsigned peer_page_count;
110 u64 evq0_addrs[EFX_MAX_VF_EVQ_SIZE * sizeof(efx_qword_t) /
111 EFX_BUF_SIZE];
112 unsigned evq0_count;
113 wait_queue_head_t flush_waitq;
114 struct mutex txq_lock;
115 unsigned long txq_mask[VI_MASK_LENGTH];
116 unsigned txq_count;
117 unsigned long rxq_mask[VI_MASK_LENGTH];
118 unsigned rxq_count;
119 unsigned long rxq_retry_mask[VI_MASK_LENGTH];
120 atomic_t rxq_retry_count;
121 struct work_struct reset_work;
122 };
123
124 struct efx_memcpy_req {
125 unsigned int from_rid;
126 void *from_buf;
127 u64 from_addr;
128 unsigned int to_rid;
129 u64 to_addr;
130 unsigned length;
131 };
132
133 /**
134 * struct efx_local_addr - A MAC address on the vswitch without a VF.
135 *
136 * Siena does not have a switch, so VFs can't transmit data to each
137 * other. Instead the VFs must be made aware of the local addresses
138 * on the vswitch, so that they can arrange for an alternative
139 * software datapath to be used.
140 *
141 * @link: List head for insertion into efx->local_addr_list.
142 * @addr: Ethernet address
143 */
144 struct efx_local_addr {
145 struct list_head link;
146 u8 addr[ETH_ALEN];
147 };
148
149 /**
150 * struct efx_endpoint_page - Page of vfdi_endpoint structures
151 *
152 * @link: List head for insertion into efx->local_page_list.
153 * @ptr: Pointer to page.
154 * @addr: DMA address of page.
155 */
156 struct efx_endpoint_page {
157 struct list_head link;
158 void *ptr;
159 dma_addr_t addr;
160 };
161
162 /* Buffer table entries are reserved txq0,rxq0,evq0,txq1,rxq1,evq1 */
163 #define EFX_BUFTBL_TXQ_BASE(_vf, _qid) \
164 ((_vf)->buftbl_base + EFX_VF_BUFTBL_PER_VI * (_qid))
165 #define EFX_BUFTBL_RXQ_BASE(_vf, _qid) \
166 (EFX_BUFTBL_TXQ_BASE(_vf, _qid) + \
167 (EFX_MAX_DMAQ_SIZE * sizeof(efx_qword_t) / EFX_BUF_SIZE))
168 #define EFX_BUFTBL_EVQ_BASE(_vf, _qid) \
169 (EFX_BUFTBL_TXQ_BASE(_vf, _qid) + \
170 (2 * EFX_MAX_DMAQ_SIZE * sizeof(efx_qword_t) / EFX_BUF_SIZE))
171
172 #define EFX_FIELD_MASK(_field) \
173 ((1 << _field ## _WIDTH) - 1)
174
175 /* VFs can only use this many transmit channels */
176 static unsigned int vf_max_tx_channels = 2;
177 module_param(vf_max_tx_channels, uint, 0444);
178 MODULE_PARM_DESC(vf_max_tx_channels,
179 "Limit the number of TX channels VFs can use");
180
181 static int max_vfs = -1;
182 module_param(max_vfs, int, 0444);
183 MODULE_PARM_DESC(max_vfs,
184 "Reduce the number of VFs initialized by the driver");
185
186 /* Workqueue used by VFDI communication. We can't use the global
187 * workqueue because it may be running the VF driver's probe()
188 * routine, which will be blocked there waiting for a VFDI response.
189 */
190 static struct workqueue_struct *vfdi_workqueue;
191
192 static unsigned abs_index(struct efx_vf *vf, unsigned index)
193 {
194 return EFX_VI_BASE + vf->index * efx_vf_size(vf->efx) + index;
195 }
196
197 static int efx_sriov_cmd(struct efx_nic *efx, bool enable,
198 unsigned *vi_scale_out, unsigned *vf_total_out)
199 {
200 MCDI_DECLARE_BUF(inbuf, MC_CMD_SRIOV_IN_LEN);
201 MCDI_DECLARE_BUF(outbuf, MC_CMD_SRIOV_OUT_LEN);
202 unsigned vi_scale, vf_total;
203 size_t outlen;
204 int rc;
205
206 MCDI_SET_DWORD(inbuf, SRIOV_IN_ENABLE, enable ? 1 : 0);
207 MCDI_SET_DWORD(inbuf, SRIOV_IN_VI_BASE, EFX_VI_BASE);
208 MCDI_SET_DWORD(inbuf, SRIOV_IN_VF_COUNT, efx->vf_count);
209
210 rc = efx_mcdi_rpc(efx, MC_CMD_SRIOV, inbuf, MC_CMD_SRIOV_IN_LEN,
211 outbuf, MC_CMD_SRIOV_OUT_LEN, &outlen);
212 if (rc)
213 return rc;
214 if (outlen < MC_CMD_SRIOV_OUT_LEN)
215 return -EIO;
216
217 vf_total = MCDI_DWORD(outbuf, SRIOV_OUT_VF_TOTAL);
218 vi_scale = MCDI_DWORD(outbuf, SRIOV_OUT_VI_SCALE);
219 if (vi_scale > EFX_VI_SCALE_MAX)
220 return -EOPNOTSUPP;
221
222 if (vi_scale_out)
223 *vi_scale_out = vi_scale;
224 if (vf_total_out)
225 *vf_total_out = vf_total;
226
227 return 0;
228 }
229
230 static void efx_sriov_usrev(struct efx_nic *efx, bool enabled)
231 {
232 efx_oword_t reg;
233
234 EFX_POPULATE_OWORD_2(reg,
235 FRF_CZ_USREV_DIS, enabled ? 0 : 1,
236 FRF_CZ_DFLT_EVQ, efx->vfdi_channel->channel);
237 efx_writeo(efx, &reg, FR_CZ_USR_EV_CFG);
238 }
239
240 static int efx_sriov_memcpy(struct efx_nic *efx, struct efx_memcpy_req *req,
241 unsigned int count)
242 {
243 MCDI_DECLARE_BUF(inbuf, MCDI_CTL_SDU_LEN_MAX_V1);
244 MCDI_DECLARE_STRUCT_PTR(record);
245 unsigned int index, used;
246 u64 from_addr;
247 u32 from_rid;
248 int rc;
249
250 mb(); /* Finish writing source/reading dest before DMA starts */
251
252 if (WARN_ON(count > MC_CMD_MEMCPY_IN_RECORD_MAXNUM))
253 return -ENOBUFS;
254 used = MC_CMD_MEMCPY_IN_LEN(count);
255
256 for (index = 0; index < count; index++) {
257 record = MCDI_ARRAY_STRUCT_PTR(inbuf, MEMCPY_IN_RECORD, index);
258 MCDI_SET_DWORD(record, MEMCPY_RECORD_TYPEDEF_NUM_RECORDS,
259 count);
260 MCDI_SET_DWORD(record, MEMCPY_RECORD_TYPEDEF_TO_RID,
261 req->to_rid);
262 MCDI_SET_QWORD(record, MEMCPY_RECORD_TYPEDEF_TO_ADDR,
263 req->to_addr);
264 if (req->from_buf == NULL) {
265 from_rid = req->from_rid;
266 from_addr = req->from_addr;
267 } else {
268 if (WARN_ON(used + req->length >
269 MCDI_CTL_SDU_LEN_MAX_V1)) {
270 rc = -ENOBUFS;
271 goto out;
272 }
273
274 from_rid = MC_CMD_MEMCPY_RECORD_TYPEDEF_RID_INLINE;
275 from_addr = used;
276 memcpy(_MCDI_PTR(inbuf, used), req->from_buf,
277 req->length);
278 used += req->length;
279 }
280
281 MCDI_SET_DWORD(record, MEMCPY_RECORD_TYPEDEF_FROM_RID, from_rid);
282 MCDI_SET_QWORD(record, MEMCPY_RECORD_TYPEDEF_FROM_ADDR,
283 from_addr);
284 MCDI_SET_DWORD(record, MEMCPY_RECORD_TYPEDEF_LENGTH,
285 req->length);
286
287 ++req;
288 }
289
290 rc = efx_mcdi_rpc(efx, MC_CMD_MEMCPY, inbuf, used, NULL, 0, NULL);
291 out:
292 mb(); /* Don't write source/read dest before DMA is complete */
293
294 return rc;
295 }
296
297 /* The TX filter is entirely controlled by this driver, and is modified
298 * underneath the feet of the VF
299 */
300 static void efx_sriov_reset_tx_filter(struct efx_vf *vf)
301 {
302 struct efx_nic *efx = vf->efx;
303 struct efx_filter_spec filter;
304 u16 vlan;
305 int rc;
306
307 if (vf->tx_filter_id != -1) {
308 efx_filter_remove_id_safe(efx, EFX_FILTER_PRI_REQUIRED,
309 vf->tx_filter_id);
310 netif_dbg(efx, hw, efx->net_dev, "Removed vf %s tx filter %d\n",
311 vf->pci_name, vf->tx_filter_id);
312 vf->tx_filter_id = -1;
313 }
314
315 if (is_zero_ether_addr(vf->addr.mac_addr))
316 return;
317
318 /* Turn on TX filtering automatically if not explicitly
319 * enabled or disabled.
320 */
321 if (vf->tx_filter_mode == VF_TX_FILTER_AUTO && vf_max_tx_channels <= 2)
322 vf->tx_filter_mode = VF_TX_FILTER_ON;
323
324 vlan = ntohs(vf->addr.tci) & VLAN_VID_MASK;
325 efx_filter_init_tx(&filter, abs_index(vf, 0));
326 rc = efx_filter_set_eth_local(&filter,
327 vlan ? vlan : EFX_FILTER_VID_UNSPEC,
328 vf->addr.mac_addr);
329 BUG_ON(rc);
330
331 rc = efx_filter_insert_filter(efx, &filter, true);
332 if (rc < 0) {
333 netif_warn(efx, hw, efx->net_dev,
334 "Unable to migrate tx filter for vf %s\n",
335 vf->pci_name);
336 } else {
337 netif_dbg(efx, hw, efx->net_dev, "Inserted vf %s tx filter %d\n",
338 vf->pci_name, rc);
339 vf->tx_filter_id = rc;
340 }
341 }
342
343 /* The RX filter is managed here on behalf of the VF driver */
344 static void efx_sriov_reset_rx_filter(struct efx_vf *vf)
345 {
346 struct efx_nic *efx = vf->efx;
347 struct efx_filter_spec filter;
348 u16 vlan;
349 int rc;
350
351 if (vf->rx_filter_id != -1) {
352 efx_filter_remove_id_safe(efx, EFX_FILTER_PRI_REQUIRED,
353 vf->rx_filter_id);
354 netif_dbg(efx, hw, efx->net_dev, "Removed vf %s rx filter %d\n",
355 vf->pci_name, vf->rx_filter_id);
356 vf->rx_filter_id = -1;
357 }
358
359 if (!vf->rx_filtering || is_zero_ether_addr(vf->addr.mac_addr))
360 return;
361
362 vlan = ntohs(vf->addr.tci) & VLAN_VID_MASK;
363 efx_filter_init_rx(&filter, EFX_FILTER_PRI_REQUIRED,
364 vf->rx_filter_flags,
365 abs_index(vf, vf->rx_filter_qid));
366 rc = efx_filter_set_eth_local(&filter,
367 vlan ? vlan : EFX_FILTER_VID_UNSPEC,
368 vf->addr.mac_addr);
369 BUG_ON(rc);
370
371 rc = efx_filter_insert_filter(efx, &filter, true);
372 if (rc < 0) {
373 netif_warn(efx, hw, efx->net_dev,
374 "Unable to insert rx filter for vf %s\n",
375 vf->pci_name);
376 } else {
377 netif_dbg(efx, hw, efx->net_dev, "Inserted vf %s rx filter %d\n",
378 vf->pci_name, rc);
379 vf->rx_filter_id = rc;
380 }
381 }
382
383 static void __efx_sriov_update_vf_addr(struct efx_vf *vf)
384 {
385 efx_sriov_reset_tx_filter(vf);
386 efx_sriov_reset_rx_filter(vf);
387 queue_work(vfdi_workqueue, &vf->efx->peer_work);
388 }
389
390 /* Push the peer list to this VF. The caller must hold status_lock to interlock
391 * with VFDI requests, and they must be serialised against manipulation of
392 * local_page_list, either by acquiring local_lock or by running from
393 * efx_sriov_peer_work()
394 */
395 static void __efx_sriov_push_vf_status(struct efx_vf *vf)
396 {
397 struct efx_nic *efx = vf->efx;
398 struct vfdi_status *status = efx->vfdi_status.addr;
399 struct efx_memcpy_req copy[4];
400 struct efx_endpoint_page *epp;
401 unsigned int pos, count;
402 unsigned data_offset;
403 efx_qword_t event;
404
405 WARN_ON(!mutex_is_locked(&vf->status_lock));
406 WARN_ON(!vf->status_addr);
407
408 status->local = vf->addr;
409 status->generation_end = ++status->generation_start;
410
411 memset(copy, '\0', sizeof(copy));
412 /* Write generation_start */
413 copy[0].from_buf = &status->generation_start;
414 copy[0].to_rid = vf->pci_rid;
415 copy[0].to_addr = vf->status_addr + offsetof(struct vfdi_status,
416 generation_start);
417 copy[0].length = sizeof(status->generation_start);
418 /* DMA the rest of the structure (excluding the generations). This
419 * assumes that the non-generation portion of vfdi_status is in
420 * one chunk starting at the version member.
421 */
422 data_offset = offsetof(struct vfdi_status, version);
423 copy[1].from_rid = efx->pci_dev->devfn;
424 copy[1].from_addr = efx->vfdi_status.dma_addr + data_offset;
425 copy[1].to_rid = vf->pci_rid;
426 copy[1].to_addr = vf->status_addr + data_offset;
427 copy[1].length = status->length - data_offset;
428
429 /* Copy the peer pages */
430 pos = 2;
431 count = 0;
432 list_for_each_entry(epp, &efx->local_page_list, link) {
433 if (count == vf->peer_page_count) {
434 /* The VF driver will know they need to provide more
435 * pages because peer_addr_count is too large.
436 */
437 break;
438 }
439 copy[pos].from_buf = NULL;
440 copy[pos].from_rid = efx->pci_dev->devfn;
441 copy[pos].from_addr = epp->addr;
442 copy[pos].to_rid = vf->pci_rid;
443 copy[pos].to_addr = vf->peer_page_addrs[count];
444 copy[pos].length = EFX_PAGE_SIZE;
445
446 if (++pos == ARRAY_SIZE(copy)) {
447 efx_sriov_memcpy(efx, copy, ARRAY_SIZE(copy));
448 pos = 0;
449 }
450 ++count;
451 }
452
453 /* Write generation_end */
454 copy[pos].from_buf = &status->generation_end;
455 copy[pos].to_rid = vf->pci_rid;
456 copy[pos].to_addr = vf->status_addr + offsetof(struct vfdi_status,
457 generation_end);
458 copy[pos].length = sizeof(status->generation_end);
459 efx_sriov_memcpy(efx, copy, pos + 1);
460
461 /* Notify the guest */
462 EFX_POPULATE_QWORD_3(event,
463 FSF_AZ_EV_CODE, FSE_CZ_EV_CODE_USER_EV,
464 VFDI_EV_SEQ, (vf->msg_seqno & 0xff),
465 VFDI_EV_TYPE, VFDI_EV_TYPE_STATUS);
466 ++vf->msg_seqno;
467 efx_generate_event(efx, EFX_VI_BASE + vf->index * efx_vf_size(efx),
468 &event);
469 }
470
471 static void efx_sriov_bufs(struct efx_nic *efx, unsigned offset,
472 u64 *addr, unsigned count)
473 {
474 efx_qword_t buf;
475 unsigned pos;
476
477 for (pos = 0; pos < count; ++pos) {
478 EFX_POPULATE_QWORD_3(buf,
479 FRF_AZ_BUF_ADR_REGION, 0,
480 FRF_AZ_BUF_ADR_FBUF,
481 addr ? addr[pos] >> 12 : 0,
482 FRF_AZ_BUF_OWNER_ID_FBUF, 0);
483 efx_sram_writeq(efx, efx->membase + FR_BZ_BUF_FULL_TBL,
484 &buf, offset + pos);
485 }
486 }
487
488 static bool bad_vf_index(struct efx_nic *efx, unsigned index)
489 {
490 return index >= efx_vf_size(efx);
491 }
492
493 static bool bad_buf_count(unsigned buf_count, unsigned max_entry_count)
494 {
495 unsigned max_buf_count = max_entry_count *
496 sizeof(efx_qword_t) / EFX_BUF_SIZE;
497
498 return ((buf_count & (buf_count - 1)) || buf_count > max_buf_count);
499 }
500
501 /* Check that VI specified by per-port index belongs to a VF.
502 * Optionally set VF index and VI index within the VF.
503 */
504 static bool map_vi_index(struct efx_nic *efx, unsigned abs_index,
505 struct efx_vf **vf_out, unsigned *rel_index_out)
506 {
507 unsigned vf_i;
508
509 if (abs_index < EFX_VI_BASE)
510 return true;
511 vf_i = (abs_index - EFX_VI_BASE) / efx_vf_size(efx);
512 if (vf_i >= efx->vf_init_count)
513 return true;
514
515 if (vf_out)
516 *vf_out = efx->vf + vf_i;
517 if (rel_index_out)
518 *rel_index_out = abs_index % efx_vf_size(efx);
519 return false;
520 }
521
522 static int efx_vfdi_init_evq(struct efx_vf *vf)
523 {
524 struct efx_nic *efx = vf->efx;
525 struct vfdi_req *req = vf->buf.addr;
526 unsigned vf_evq = req->u.init_evq.index;
527 unsigned buf_count = req->u.init_evq.buf_count;
528 unsigned abs_evq = abs_index(vf, vf_evq);
529 unsigned buftbl = EFX_BUFTBL_EVQ_BASE(vf, vf_evq);
530 efx_oword_t reg;
531
532 if (bad_vf_index(efx, vf_evq) ||
533 bad_buf_count(buf_count, EFX_MAX_VF_EVQ_SIZE)) {
534 if (net_ratelimit())
535 netif_err(efx, hw, efx->net_dev,
536 "ERROR: Invalid INIT_EVQ from %s: evq %d bufs %d\n",
537 vf->pci_name, vf_evq, buf_count);
538 return VFDI_RC_EINVAL;
539 }
540
541 efx_sriov_bufs(efx, buftbl, req->u.init_evq.addr, buf_count);
542
543 EFX_POPULATE_OWORD_3(reg,
544 FRF_CZ_TIMER_Q_EN, 1,
545 FRF_CZ_HOST_NOTIFY_MODE, 0,
546 FRF_CZ_TIMER_MODE, FFE_CZ_TIMER_MODE_DIS);
547 efx_writeo_table(efx, &reg, FR_BZ_TIMER_TBL, abs_evq);
548 EFX_POPULATE_OWORD_3(reg,
549 FRF_AZ_EVQ_EN, 1,
550 FRF_AZ_EVQ_SIZE, __ffs(buf_count),
551 FRF_AZ_EVQ_BUF_BASE_ID, buftbl);
552 efx_writeo_table(efx, &reg, FR_BZ_EVQ_PTR_TBL, abs_evq);
553
554 if (vf_evq == 0) {
555 memcpy(vf->evq0_addrs, req->u.init_evq.addr,
556 buf_count * sizeof(u64));
557 vf->evq0_count = buf_count;
558 }
559
560 return VFDI_RC_SUCCESS;
561 }
562
563 static int efx_vfdi_init_rxq(struct efx_vf *vf)
564 {
565 struct efx_nic *efx = vf->efx;
566 struct vfdi_req *req = vf->buf.addr;
567 unsigned vf_rxq = req->u.init_rxq.index;
568 unsigned vf_evq = req->u.init_rxq.evq;
569 unsigned buf_count = req->u.init_rxq.buf_count;
570 unsigned buftbl = EFX_BUFTBL_RXQ_BASE(vf, vf_rxq);
571 unsigned label;
572 efx_oword_t reg;
573
574 if (bad_vf_index(efx, vf_evq) || bad_vf_index(efx, vf_rxq) ||
575 vf_rxq >= VF_MAX_RX_QUEUES ||
576 bad_buf_count(buf_count, EFX_MAX_DMAQ_SIZE)) {
577 if (net_ratelimit())
578 netif_err(efx, hw, efx->net_dev,
579 "ERROR: Invalid INIT_RXQ from %s: rxq %d evq %d "
580 "buf_count %d\n", vf->pci_name, vf_rxq,
581 vf_evq, buf_count);
582 return VFDI_RC_EINVAL;
583 }
584 if (__test_and_set_bit(req->u.init_rxq.index, vf->rxq_mask))
585 ++vf->rxq_count;
586 efx_sriov_bufs(efx, buftbl, req->u.init_rxq.addr, buf_count);
587
588 label = req->u.init_rxq.label & EFX_FIELD_MASK(FRF_AZ_RX_DESCQ_LABEL);
589 EFX_POPULATE_OWORD_6(reg,
590 FRF_AZ_RX_DESCQ_BUF_BASE_ID, buftbl,
591 FRF_AZ_RX_DESCQ_EVQ_ID, abs_index(vf, vf_evq),
592 FRF_AZ_RX_DESCQ_LABEL, label,
593 FRF_AZ_RX_DESCQ_SIZE, __ffs(buf_count),
594 FRF_AZ_RX_DESCQ_JUMBO,
595 !!(req->u.init_rxq.flags &
596 VFDI_RXQ_FLAG_SCATTER_EN),
597 FRF_AZ_RX_DESCQ_EN, 1);
598 efx_writeo_table(efx, &reg, FR_BZ_RX_DESC_PTR_TBL,
599 abs_index(vf, vf_rxq));
600
601 return VFDI_RC_SUCCESS;
602 }
603
604 static int efx_vfdi_init_txq(struct efx_vf *vf)
605 {
606 struct efx_nic *efx = vf->efx;
607 struct vfdi_req *req = vf->buf.addr;
608 unsigned vf_txq = req->u.init_txq.index;
609 unsigned vf_evq = req->u.init_txq.evq;
610 unsigned buf_count = req->u.init_txq.buf_count;
611 unsigned buftbl = EFX_BUFTBL_TXQ_BASE(vf, vf_txq);
612 unsigned label, eth_filt_en;
613 efx_oword_t reg;
614
615 if (bad_vf_index(efx, vf_evq) || bad_vf_index(efx, vf_txq) ||
616 vf_txq >= vf_max_tx_channels ||
617 bad_buf_count(buf_count, EFX_MAX_DMAQ_SIZE)) {
618 if (net_ratelimit())
619 netif_err(efx, hw, efx->net_dev,
620 "ERROR: Invalid INIT_TXQ from %s: txq %d evq %d "
621 "buf_count %d\n", vf->pci_name, vf_txq,
622 vf_evq, buf_count);
623 return VFDI_RC_EINVAL;
624 }
625
626 mutex_lock(&vf->txq_lock);
627 if (__test_and_set_bit(req->u.init_txq.index, vf->txq_mask))
628 ++vf->txq_count;
629 mutex_unlock(&vf->txq_lock);
630 efx_sriov_bufs(efx, buftbl, req->u.init_txq.addr, buf_count);
631
632 eth_filt_en = vf->tx_filter_mode == VF_TX_FILTER_ON;
633
634 label = req->u.init_txq.label & EFX_FIELD_MASK(FRF_AZ_TX_DESCQ_LABEL);
635 EFX_POPULATE_OWORD_8(reg,
636 FRF_CZ_TX_DPT_Q_MASK_WIDTH, min(efx->vi_scale, 1U),
637 FRF_CZ_TX_DPT_ETH_FILT_EN, eth_filt_en,
638 FRF_AZ_TX_DESCQ_EN, 1,
639 FRF_AZ_TX_DESCQ_BUF_BASE_ID, buftbl,
640 FRF_AZ_TX_DESCQ_EVQ_ID, abs_index(vf, vf_evq),
641 FRF_AZ_TX_DESCQ_LABEL, label,
642 FRF_AZ_TX_DESCQ_SIZE, __ffs(buf_count),
643 FRF_BZ_TX_NON_IP_DROP_DIS, 1);
644 efx_writeo_table(efx, &reg, FR_BZ_TX_DESC_PTR_TBL,
645 abs_index(vf, vf_txq));
646
647 return VFDI_RC_SUCCESS;
648 }
649
650 /* Returns true when efx_vfdi_fini_all_queues should wake */
651 static bool efx_vfdi_flush_wake(struct efx_vf *vf)
652 {
653 /* Ensure that all updates are visible to efx_vfdi_fini_all_queues() */
654 smp_mb();
655
656 return (!vf->txq_count && !vf->rxq_count) ||
657 atomic_read(&vf->rxq_retry_count);
658 }
659
660 static void efx_vfdi_flush_clear(struct efx_vf *vf)
661 {
662 memset(vf->txq_mask, 0, sizeof(vf->txq_mask));
663 vf->txq_count = 0;
664 memset(vf->rxq_mask, 0, sizeof(vf->rxq_mask));
665 vf->rxq_count = 0;
666 memset(vf->rxq_retry_mask, 0, sizeof(vf->rxq_retry_mask));
667 atomic_set(&vf->rxq_retry_count, 0);
668 }
669
670 static int efx_vfdi_fini_all_queues(struct efx_vf *vf)
671 {
672 struct efx_nic *efx = vf->efx;
673 efx_oword_t reg;
674 unsigned count = efx_vf_size(efx);
675 unsigned vf_offset = EFX_VI_BASE + vf->index * efx_vf_size(efx);
676 unsigned timeout = HZ;
677 unsigned index, rxqs_count;
678 MCDI_DECLARE_BUF(inbuf, MC_CMD_FLUSH_RX_QUEUES_IN_LENMAX);
679 int rc;
680
681 BUILD_BUG_ON(VF_MAX_RX_QUEUES >
682 MC_CMD_FLUSH_RX_QUEUES_IN_QID_OFST_MAXNUM);
683
684 rtnl_lock();
685 siena_prepare_flush(efx);
686 rtnl_unlock();
687
688 /* Flush all the initialized queues */
689 rxqs_count = 0;
690 for (index = 0; index < count; ++index) {
691 if (test_bit(index, vf->txq_mask)) {
692 EFX_POPULATE_OWORD_2(reg,
693 FRF_AZ_TX_FLUSH_DESCQ_CMD, 1,
694 FRF_AZ_TX_FLUSH_DESCQ,
695 vf_offset + index);
696 efx_writeo(efx, &reg, FR_AZ_TX_FLUSH_DESCQ);
697 }
698 if (test_bit(index, vf->rxq_mask)) {
699 MCDI_SET_ARRAY_DWORD(
700 inbuf, FLUSH_RX_QUEUES_IN_QID_OFST,
701 rxqs_count, vf_offset + index);
702 rxqs_count++;
703 }
704 }
705
706 atomic_set(&vf->rxq_retry_count, 0);
707 while (timeout && (vf->rxq_count || vf->txq_count)) {
708 rc = efx_mcdi_rpc(efx, MC_CMD_FLUSH_RX_QUEUES, inbuf,
709 MC_CMD_FLUSH_RX_QUEUES_IN_LEN(rxqs_count),
710 NULL, 0, NULL);
711 WARN_ON(rc < 0);
712
713 timeout = wait_event_timeout(vf->flush_waitq,
714 efx_vfdi_flush_wake(vf),
715 timeout);
716 rxqs_count = 0;
717 for (index = 0; index < count; ++index) {
718 if (test_and_clear_bit(index, vf->rxq_retry_mask)) {
719 atomic_dec(&vf->rxq_retry_count);
720 MCDI_SET_ARRAY_DWORD(
721 inbuf, FLUSH_RX_QUEUES_IN_QID_OFST,
722 rxqs_count, vf_offset + index);
723 rxqs_count++;
724 }
725 }
726 }
727
728 rtnl_lock();
729 siena_finish_flush(efx);
730 rtnl_unlock();
731
732 /* Irrespective of success/failure, fini the queues */
733 EFX_ZERO_OWORD(reg);
734 for (index = 0; index < count; ++index) {
735 efx_writeo_table(efx, &reg, FR_BZ_RX_DESC_PTR_TBL,
736 vf_offset + index);
737 efx_writeo_table(efx, &reg, FR_BZ_TX_DESC_PTR_TBL,
738 vf_offset + index);
739 efx_writeo_table(efx, &reg, FR_BZ_EVQ_PTR_TBL,
740 vf_offset + index);
741 efx_writeo_table(efx, &reg, FR_BZ_TIMER_TBL,
742 vf_offset + index);
743 }
744 efx_sriov_bufs(efx, vf->buftbl_base, NULL,
745 EFX_VF_BUFTBL_PER_VI * efx_vf_size(efx));
746 efx_vfdi_flush_clear(vf);
747
748 vf->evq0_count = 0;
749
750 return timeout ? 0 : VFDI_RC_ETIMEDOUT;
751 }
752
753 static int efx_vfdi_insert_filter(struct efx_vf *vf)
754 {
755 struct efx_nic *efx = vf->efx;
756 struct vfdi_req *req = vf->buf.addr;
757 unsigned vf_rxq = req->u.mac_filter.rxq;
758 unsigned flags;
759
760 if (bad_vf_index(efx, vf_rxq) || vf->rx_filtering) {
761 if (net_ratelimit())
762 netif_err(efx, hw, efx->net_dev,
763 "ERROR: Invalid INSERT_FILTER from %s: rxq %d "
764 "flags 0x%x\n", vf->pci_name, vf_rxq,
765 req->u.mac_filter.flags);
766 return VFDI_RC_EINVAL;
767 }
768
769 flags = 0;
770 if (req->u.mac_filter.flags & VFDI_MAC_FILTER_FLAG_RSS)
771 flags |= EFX_FILTER_FLAG_RX_RSS;
772 if (req->u.mac_filter.flags & VFDI_MAC_FILTER_FLAG_SCATTER)
773 flags |= EFX_FILTER_FLAG_RX_SCATTER;
774 vf->rx_filter_flags = flags;
775 vf->rx_filter_qid = vf_rxq;
776 vf->rx_filtering = true;
777
778 efx_sriov_reset_rx_filter(vf);
779 queue_work(vfdi_workqueue, &efx->peer_work);
780
781 return VFDI_RC_SUCCESS;
782 }
783
784 static int efx_vfdi_remove_all_filters(struct efx_vf *vf)
785 {
786 vf->rx_filtering = false;
787 efx_sriov_reset_rx_filter(vf);
788 queue_work(vfdi_workqueue, &vf->efx->peer_work);
789
790 return VFDI_RC_SUCCESS;
791 }
792
793 static int efx_vfdi_set_status_page(struct efx_vf *vf)
794 {
795 struct efx_nic *efx = vf->efx;
796 struct vfdi_req *req = vf->buf.addr;
797 u64 page_count = req->u.set_status_page.peer_page_count;
798 u64 max_page_count =
799 (EFX_PAGE_SIZE -
800 offsetof(struct vfdi_req, u.set_status_page.peer_page_addr[0]))
801 / sizeof(req->u.set_status_page.peer_page_addr[0]);
802
803 if (!req->u.set_status_page.dma_addr || page_count > max_page_count) {
804 if (net_ratelimit())
805 netif_err(efx, hw, efx->net_dev,
806 "ERROR: Invalid SET_STATUS_PAGE from %s\n",
807 vf->pci_name);
808 return VFDI_RC_EINVAL;
809 }
810
811 mutex_lock(&efx->local_lock);
812 mutex_lock(&vf->status_lock);
813 vf->status_addr = req->u.set_status_page.dma_addr;
814
815 kfree(vf->peer_page_addrs);
816 vf->peer_page_addrs = NULL;
817 vf->peer_page_count = 0;
818
819 if (page_count) {
820 vf->peer_page_addrs = kcalloc(page_count, sizeof(u64),
821 GFP_KERNEL);
822 if (vf->peer_page_addrs) {
823 memcpy(vf->peer_page_addrs,
824 req->u.set_status_page.peer_page_addr,
825 page_count * sizeof(u64));
826 vf->peer_page_count = page_count;
827 }
828 }
829
830 __efx_sriov_push_vf_status(vf);
831 mutex_unlock(&vf->status_lock);
832 mutex_unlock(&efx->local_lock);
833
834 return VFDI_RC_SUCCESS;
835 }
836
837 static int efx_vfdi_clear_status_page(struct efx_vf *vf)
838 {
839 mutex_lock(&vf->status_lock);
840 vf->status_addr = 0;
841 mutex_unlock(&vf->status_lock);
842
843 return VFDI_RC_SUCCESS;
844 }
845
846 typedef int (*efx_vfdi_op_t)(struct efx_vf *vf);
847
848 static const efx_vfdi_op_t vfdi_ops[VFDI_OP_LIMIT] = {
849 [VFDI_OP_INIT_EVQ] = efx_vfdi_init_evq,
850 [VFDI_OP_INIT_TXQ] = efx_vfdi_init_txq,
851 [VFDI_OP_INIT_RXQ] = efx_vfdi_init_rxq,
852 [VFDI_OP_FINI_ALL_QUEUES] = efx_vfdi_fini_all_queues,
853 [VFDI_OP_INSERT_FILTER] = efx_vfdi_insert_filter,
854 [VFDI_OP_REMOVE_ALL_FILTERS] = efx_vfdi_remove_all_filters,
855 [VFDI_OP_SET_STATUS_PAGE] = efx_vfdi_set_status_page,
856 [VFDI_OP_CLEAR_STATUS_PAGE] = efx_vfdi_clear_status_page,
857 };
858
859 static void efx_sriov_vfdi(struct work_struct *work)
860 {
861 struct efx_vf *vf = container_of(work, struct efx_vf, req);
862 struct efx_nic *efx = vf->efx;
863 struct vfdi_req *req = vf->buf.addr;
864 struct efx_memcpy_req copy[2];
865 int rc;
866
867 /* Copy this page into the local address space */
868 memset(copy, '\0', sizeof(copy));
869 copy[0].from_rid = vf->pci_rid;
870 copy[0].from_addr = vf->req_addr;
871 copy[0].to_rid = efx->pci_dev->devfn;
872 copy[0].to_addr = vf->buf.dma_addr;
873 copy[0].length = EFX_PAGE_SIZE;
874 rc = efx_sriov_memcpy(efx, copy, 1);
875 if (rc) {
876 /* If we can't get the request, we can't reply to the caller */
877 if (net_ratelimit())
878 netif_err(efx, hw, efx->net_dev,
879 "ERROR: Unable to fetch VFDI request from %s rc %d\n",
880 vf->pci_name, -rc);
881 vf->busy = false;
882 return;
883 }
884
885 if (req->op < VFDI_OP_LIMIT && vfdi_ops[req->op] != NULL) {
886 rc = vfdi_ops[req->op](vf);
887 if (rc == 0) {
888 netif_dbg(efx, hw, efx->net_dev,
889 "vfdi request %d from %s ok\n",
890 req->op, vf->pci_name);
891 }
892 } else {
893 netif_dbg(efx, hw, efx->net_dev,
894 "ERROR: Unrecognised request %d from VF %s addr "
895 "%llx\n", req->op, vf->pci_name,
896 (unsigned long long)vf->req_addr);
897 rc = VFDI_RC_EOPNOTSUPP;
898 }
899
900 /* Allow subsequent VF requests */
901 vf->busy = false;
902 smp_wmb();
903
904 /* Respond to the request */
905 req->rc = rc;
906 req->op = VFDI_OP_RESPONSE;
907
908 memset(copy, '\0', sizeof(copy));
909 copy[0].from_buf = &req->rc;
910 copy[0].to_rid = vf->pci_rid;
911 copy[0].to_addr = vf->req_addr + offsetof(struct vfdi_req, rc);
912 copy[0].length = sizeof(req->rc);
913 copy[1].from_buf = &req->op;
914 copy[1].to_rid = vf->pci_rid;
915 copy[1].to_addr = vf->req_addr + offsetof(struct vfdi_req, op);
916 copy[1].length = sizeof(req->op);
917
918 (void) efx_sriov_memcpy(efx, copy, ARRAY_SIZE(copy));
919 }
920
921
922
923 /* After a reset the event queues inside the guests no longer exist. Fill the
924 * event ring in guest memory with VFDI reset events, then (re-initialise) the
925 * event queue to raise an interrupt. The guest driver will then recover.
926 */
927 static void efx_sriov_reset_vf(struct efx_vf *vf, struct efx_buffer *buffer)
928 {
929 struct efx_nic *efx = vf->efx;
930 struct efx_memcpy_req copy_req[4];
931 efx_qword_t event;
932 unsigned int pos, count, k, buftbl, abs_evq;
933 efx_oword_t reg;
934 efx_dword_t ptr;
935 int rc;
936
937 BUG_ON(buffer->len != EFX_PAGE_SIZE);
938
939 if (!vf->evq0_count)
940 return;
941 BUG_ON(vf->evq0_count & (vf->evq0_count - 1));
942
943 mutex_lock(&vf->status_lock);
944 EFX_POPULATE_QWORD_3(event,
945 FSF_AZ_EV_CODE, FSE_CZ_EV_CODE_USER_EV,
946 VFDI_EV_SEQ, vf->msg_seqno,
947 VFDI_EV_TYPE, VFDI_EV_TYPE_RESET);
948 vf->msg_seqno++;
949 for (pos = 0; pos < EFX_PAGE_SIZE; pos += sizeof(event))
950 memcpy(buffer->addr + pos, &event, sizeof(event));
951
952 for (pos = 0; pos < vf->evq0_count; pos += count) {
953 count = min_t(unsigned, vf->evq0_count - pos,
954 ARRAY_SIZE(copy_req));
955 for (k = 0; k < count; k++) {
956 copy_req[k].from_buf = NULL;
957 copy_req[k].from_rid = efx->pci_dev->devfn;
958 copy_req[k].from_addr = buffer->dma_addr;
959 copy_req[k].to_rid = vf->pci_rid;
960 copy_req[k].to_addr = vf->evq0_addrs[pos + k];
961 copy_req[k].length = EFX_PAGE_SIZE;
962 }
963 rc = efx_sriov_memcpy(efx, copy_req, count);
964 if (rc) {
965 if (net_ratelimit())
966 netif_err(efx, hw, efx->net_dev,
967 "ERROR: Unable to notify %s of reset"
968 ": %d\n", vf->pci_name, -rc);
969 break;
970 }
971 }
972
973 /* Reinitialise, arm and trigger evq0 */
974 abs_evq = abs_index(vf, 0);
975 buftbl = EFX_BUFTBL_EVQ_BASE(vf, 0);
976 efx_sriov_bufs(efx, buftbl, vf->evq0_addrs, vf->evq0_count);
977
978 EFX_POPULATE_OWORD_3(reg,
979 FRF_CZ_TIMER_Q_EN, 1,
980 FRF_CZ_HOST_NOTIFY_MODE, 0,
981 FRF_CZ_TIMER_MODE, FFE_CZ_TIMER_MODE_DIS);
982 efx_writeo_table(efx, &reg, FR_BZ_TIMER_TBL, abs_evq);
983 EFX_POPULATE_OWORD_3(reg,
984 FRF_AZ_EVQ_EN, 1,
985 FRF_AZ_EVQ_SIZE, __ffs(vf->evq0_count),
986 FRF_AZ_EVQ_BUF_BASE_ID, buftbl);
987 efx_writeo_table(efx, &reg, FR_BZ_EVQ_PTR_TBL, abs_evq);
988 EFX_POPULATE_DWORD_1(ptr, FRF_AZ_EVQ_RPTR, 0);
989 efx_writed(efx, &ptr, FR_BZ_EVQ_RPTR + FR_BZ_EVQ_RPTR_STEP * abs_evq);
990
991 mutex_unlock(&vf->status_lock);
992 }
993
994 static void efx_sriov_reset_vf_work(struct work_struct *work)
995 {
996 struct efx_vf *vf = container_of(work, struct efx_vf, req);
997 struct efx_nic *efx = vf->efx;
998 struct efx_buffer buf;
999
1000 if (!efx_nic_alloc_buffer(efx, &buf, EFX_PAGE_SIZE)) {
1001 efx_sriov_reset_vf(vf, &buf);
1002 efx_nic_free_buffer(efx, &buf);
1003 }
1004 }
1005
1006 static void efx_sriov_handle_no_channel(struct efx_nic *efx)
1007 {
1008 netif_err(efx, drv, efx->net_dev,
1009 "ERROR: IOV requires MSI-X and 1 additional interrupt"
1010 "vector. IOV disabled\n");
1011 efx->vf_count = 0;
1012 }
1013
1014 static int efx_sriov_probe_channel(struct efx_channel *channel)
1015 {
1016 channel->efx->vfdi_channel = channel;
1017 return 0;
1018 }
1019
1020 static void
1021 efx_sriov_get_channel_name(struct efx_channel *channel, char *buf, size_t len)
1022 {
1023 snprintf(buf, len, "%s-iov", channel->efx->name);
1024 }
1025
1026 static const struct efx_channel_type efx_sriov_channel_type = {
1027 .handle_no_channel = efx_sriov_handle_no_channel,
1028 .pre_probe = efx_sriov_probe_channel,
1029 .post_remove = efx_channel_dummy_op_void,
1030 .get_name = efx_sriov_get_channel_name,
1031 /* no copy operation; channel must not be reallocated */
1032 .keep_eventq = true,
1033 };
1034
1035 void efx_sriov_probe(struct efx_nic *efx)
1036 {
1037 unsigned count;
1038
1039 if (!max_vfs)
1040 return;
1041
1042 if (efx_sriov_cmd(efx, false, &efx->vi_scale, &count))
1043 return;
1044 if (count > 0 && count > max_vfs)
1045 count = max_vfs;
1046
1047 /* efx_nic_dimension_resources() will reduce vf_count as appopriate */
1048 efx->vf_count = count;
1049
1050 efx->extra_channel_type[EFX_EXTRA_CHANNEL_IOV] = &efx_sriov_channel_type;
1051 }
1052
1053 /* Copy the list of individual addresses into the vfdi_status.peers
1054 * array and auxillary pages, protected by %local_lock. Drop that lock
1055 * and then broadcast the address list to every VF.
1056 */
1057 static void efx_sriov_peer_work(struct work_struct *data)
1058 {
1059 struct efx_nic *efx = container_of(data, struct efx_nic, peer_work);
1060 struct vfdi_status *vfdi_status = efx->vfdi_status.addr;
1061 struct efx_vf *vf;
1062 struct efx_local_addr *local_addr;
1063 struct vfdi_endpoint *peer;
1064 struct efx_endpoint_page *epp;
1065 struct list_head pages;
1066 unsigned int peer_space;
1067 unsigned int peer_count;
1068 unsigned int pos;
1069
1070 mutex_lock(&efx->local_lock);
1071
1072 /* Move the existing peer pages off %local_page_list */
1073 INIT_LIST_HEAD(&pages);
1074 list_splice_tail_init(&efx->local_page_list, &pages);
1075
1076 /* Populate the VF addresses starting from entry 1 (entry 0 is
1077 * the PF address)
1078 */
1079 peer = vfdi_status->peers + 1;
1080 peer_space = ARRAY_SIZE(vfdi_status->peers) - 1;
1081 peer_count = 1;
1082 for (pos = 0; pos < efx->vf_count; ++pos) {
1083 vf = efx->vf + pos;
1084
1085 mutex_lock(&vf->status_lock);
1086 if (vf->rx_filtering && !is_zero_ether_addr(vf->addr.mac_addr)) {
1087 *peer++ = vf->addr;
1088 ++peer_count;
1089 --peer_space;
1090 BUG_ON(peer_space == 0);
1091 }
1092 mutex_unlock(&vf->status_lock);
1093 }
1094
1095 /* Fill the remaining addresses */
1096 list_for_each_entry(local_addr, &efx->local_addr_list, link) {
1097 memcpy(peer->mac_addr, local_addr->addr, ETH_ALEN);
1098 peer->tci = 0;
1099 ++peer;
1100 ++peer_count;
1101 if (--peer_space == 0) {
1102 if (list_empty(&pages)) {
1103 epp = kmalloc(sizeof(*epp), GFP_KERNEL);
1104 if (!epp)
1105 break;
1106 epp->ptr = dma_alloc_coherent(
1107 &efx->pci_dev->dev, EFX_PAGE_SIZE,
1108 &epp->addr, GFP_KERNEL);
1109 if (!epp->ptr) {
1110 kfree(epp);
1111 break;
1112 }
1113 } else {
1114 epp = list_first_entry(
1115 &pages, struct efx_endpoint_page, link);
1116 list_del(&epp->link);
1117 }
1118
1119 list_add_tail(&epp->link, &efx->local_page_list);
1120 peer = (struct vfdi_endpoint *)epp->ptr;
1121 peer_space = EFX_PAGE_SIZE / sizeof(struct vfdi_endpoint);
1122 }
1123 }
1124 vfdi_status->peer_count = peer_count;
1125 mutex_unlock(&efx->local_lock);
1126
1127 /* Free any now unused endpoint pages */
1128 while (!list_empty(&pages)) {
1129 epp = list_first_entry(
1130 &pages, struct efx_endpoint_page, link);
1131 list_del(&epp->link);
1132 dma_free_coherent(&efx->pci_dev->dev, EFX_PAGE_SIZE,
1133 epp->ptr, epp->addr);
1134 kfree(epp);
1135 }
1136
1137 /* Finally, push the pages */
1138 for (pos = 0; pos < efx->vf_count; ++pos) {
1139 vf = efx->vf + pos;
1140
1141 mutex_lock(&vf->status_lock);
1142 if (vf->status_addr)
1143 __efx_sriov_push_vf_status(vf);
1144 mutex_unlock(&vf->status_lock);
1145 }
1146 }
1147
1148 static void efx_sriov_free_local(struct efx_nic *efx)
1149 {
1150 struct efx_local_addr *local_addr;
1151 struct efx_endpoint_page *epp;
1152
1153 while (!list_empty(&efx->local_addr_list)) {
1154 local_addr = list_first_entry(&efx->local_addr_list,
1155 struct efx_local_addr, link);
1156 list_del(&local_addr->link);
1157 kfree(local_addr);
1158 }
1159
1160 while (!list_empty(&efx->local_page_list)) {
1161 epp = list_first_entry(&efx->local_page_list,
1162 struct efx_endpoint_page, link);
1163 list_del(&epp->link);
1164 dma_free_coherent(&efx->pci_dev->dev, EFX_PAGE_SIZE,
1165 epp->ptr, epp->addr);
1166 kfree(epp);
1167 }
1168 }
1169
1170 static int efx_sriov_vf_alloc(struct efx_nic *efx)
1171 {
1172 unsigned index;
1173 struct efx_vf *vf;
1174
1175 efx->vf = kzalloc(sizeof(struct efx_vf) * efx->vf_count, GFP_KERNEL);
1176 if (!efx->vf)
1177 return -ENOMEM;
1178
1179 for (index = 0; index < efx->vf_count; ++index) {
1180 vf = efx->vf + index;
1181
1182 vf->efx = efx;
1183 vf->index = index;
1184 vf->rx_filter_id = -1;
1185 vf->tx_filter_mode = VF_TX_FILTER_AUTO;
1186 vf->tx_filter_id = -1;
1187 INIT_WORK(&vf->req, efx_sriov_vfdi);
1188 INIT_WORK(&vf->reset_work, efx_sriov_reset_vf_work);
1189 init_waitqueue_head(&vf->flush_waitq);
1190 mutex_init(&vf->status_lock);
1191 mutex_init(&vf->txq_lock);
1192 }
1193
1194 return 0;
1195 }
1196
1197 static void efx_sriov_vfs_fini(struct efx_nic *efx)
1198 {
1199 struct efx_vf *vf;
1200 unsigned int pos;
1201
1202 for (pos = 0; pos < efx->vf_count; ++pos) {
1203 vf = efx->vf + pos;
1204
1205 efx_nic_free_buffer(efx, &vf->buf);
1206 kfree(vf->peer_page_addrs);
1207 vf->peer_page_addrs = NULL;
1208 vf->peer_page_count = 0;
1209
1210 vf->evq0_count = 0;
1211 }
1212 }
1213
1214 static int efx_sriov_vfs_init(struct efx_nic *efx)
1215 {
1216 struct pci_dev *pci_dev = efx->pci_dev;
1217 unsigned index, devfn, sriov, buftbl_base;
1218 u16 offset, stride;
1219 struct efx_vf *vf;
1220 int rc;
1221
1222 sriov = pci_find_ext_capability(pci_dev, PCI_EXT_CAP_ID_SRIOV);
1223 if (!sriov)
1224 return -ENOENT;
1225
1226 pci_read_config_word(pci_dev, sriov + PCI_SRIOV_VF_OFFSET, &offset);
1227 pci_read_config_word(pci_dev, sriov + PCI_SRIOV_VF_STRIDE, &stride);
1228
1229 buftbl_base = efx->vf_buftbl_base;
1230 devfn = pci_dev->devfn + offset;
1231 for (index = 0; index < efx->vf_count; ++index) {
1232 vf = efx->vf + index;
1233
1234 /* Reserve buffer entries */
1235 vf->buftbl_base = buftbl_base;
1236 buftbl_base += EFX_VF_BUFTBL_PER_VI * efx_vf_size(efx);
1237
1238 vf->pci_rid = devfn;
1239 snprintf(vf->pci_name, sizeof(vf->pci_name),
1240 "%04x:%02x:%02x.%d",
1241 pci_domain_nr(pci_dev->bus), pci_dev->bus->number,
1242 PCI_SLOT(devfn), PCI_FUNC(devfn));
1243
1244 rc = efx_nic_alloc_buffer(efx, &vf->buf, EFX_PAGE_SIZE);
1245 if (rc)
1246 goto fail;
1247
1248 devfn += stride;
1249 }
1250
1251 return 0;
1252
1253 fail:
1254 efx_sriov_vfs_fini(efx);
1255 return rc;
1256 }
1257
1258 int efx_sriov_init(struct efx_nic *efx)
1259 {
1260 struct net_device *net_dev = efx->net_dev;
1261 struct vfdi_status *vfdi_status;
1262 int rc;
1263
1264 /* Ensure there's room for vf_channel */
1265 BUILD_BUG_ON(EFX_MAX_CHANNELS + 1 >= EFX_VI_BASE);
1266 /* Ensure that VI_BASE is aligned on VI_SCALE */
1267 BUILD_BUG_ON(EFX_VI_BASE & ((1 << EFX_VI_SCALE_MAX) - 1));
1268
1269 if (efx->vf_count == 0)
1270 return 0;
1271
1272 rc = efx_sriov_cmd(efx, true, NULL, NULL);
1273 if (rc)
1274 goto fail_cmd;
1275
1276 rc = efx_nic_alloc_buffer(efx, &efx->vfdi_status, sizeof(*vfdi_status));
1277 if (rc)
1278 goto fail_status;
1279 vfdi_status = efx->vfdi_status.addr;
1280 memset(vfdi_status, 0, sizeof(*vfdi_status));
1281 vfdi_status->version = 1;
1282 vfdi_status->length = sizeof(*vfdi_status);
1283 vfdi_status->max_tx_channels = vf_max_tx_channels;
1284 vfdi_status->vi_scale = efx->vi_scale;
1285 vfdi_status->rss_rxq_count = efx->rss_spread;
1286 vfdi_status->peer_count = 1 + efx->vf_count;
1287 vfdi_status->timer_quantum_ns = efx->timer_quantum_ns;
1288
1289 rc = efx_sriov_vf_alloc(efx);
1290 if (rc)
1291 goto fail_alloc;
1292
1293 mutex_init(&efx->local_lock);
1294 INIT_WORK(&efx->peer_work, efx_sriov_peer_work);
1295 INIT_LIST_HEAD(&efx->local_addr_list);
1296 INIT_LIST_HEAD(&efx->local_page_list);
1297
1298 rc = efx_sriov_vfs_init(efx);
1299 if (rc)
1300 goto fail_vfs;
1301
1302 rtnl_lock();
1303 memcpy(vfdi_status->peers[0].mac_addr,
1304 net_dev->dev_addr, ETH_ALEN);
1305 efx->vf_init_count = efx->vf_count;
1306 rtnl_unlock();
1307
1308 efx_sriov_usrev(efx, true);
1309
1310 /* At this point we must be ready to accept VFDI requests */
1311
1312 rc = pci_enable_sriov(efx->pci_dev, efx->vf_count);
1313 if (rc)
1314 goto fail_pci;
1315
1316 netif_info(efx, probe, net_dev,
1317 "enabled SR-IOV for %d VFs, %d VI per VF\n",
1318 efx->vf_count, efx_vf_size(efx));
1319 return 0;
1320
1321 fail_pci:
1322 efx_sriov_usrev(efx, false);
1323 rtnl_lock();
1324 efx->vf_init_count = 0;
1325 rtnl_unlock();
1326 efx_sriov_vfs_fini(efx);
1327 fail_vfs:
1328 cancel_work_sync(&efx->peer_work);
1329 efx_sriov_free_local(efx);
1330 kfree(efx->vf);
1331 fail_alloc:
1332 efx_nic_free_buffer(efx, &efx->vfdi_status);
1333 fail_status:
1334 efx_sriov_cmd(efx, false, NULL, NULL);
1335 fail_cmd:
1336 return rc;
1337 }
1338
1339 void efx_sriov_fini(struct efx_nic *efx)
1340 {
1341 struct efx_vf *vf;
1342 unsigned int pos;
1343
1344 if (efx->vf_init_count == 0)
1345 return;
1346
1347 /* Disable all interfaces to reconfiguration */
1348 BUG_ON(efx->vfdi_channel->enabled);
1349 efx_sriov_usrev(efx, false);
1350 rtnl_lock();
1351 efx->vf_init_count = 0;
1352 rtnl_unlock();
1353
1354 /* Flush all reconfiguration work */
1355 for (pos = 0; pos < efx->vf_count; ++pos) {
1356 vf = efx->vf + pos;
1357 cancel_work_sync(&vf->req);
1358 cancel_work_sync(&vf->reset_work);
1359 }
1360 cancel_work_sync(&efx->peer_work);
1361
1362 pci_disable_sriov(efx->pci_dev);
1363
1364 /* Tear down back-end state */
1365 efx_sriov_vfs_fini(efx);
1366 efx_sriov_free_local(efx);
1367 kfree(efx->vf);
1368 efx_nic_free_buffer(efx, &efx->vfdi_status);
1369 efx_sriov_cmd(efx, false, NULL, NULL);
1370 }
1371
1372 void efx_sriov_event(struct efx_channel *channel, efx_qword_t *event)
1373 {
1374 struct efx_nic *efx = channel->efx;
1375 struct efx_vf *vf;
1376 unsigned qid, seq, type, data;
1377
1378 qid = EFX_QWORD_FIELD(*event, FSF_CZ_USER_QID);
1379
1380 /* USR_EV_REG_VALUE is dword0, so access the VFDI_EV fields directly */
1381 BUILD_BUG_ON(FSF_CZ_USER_EV_REG_VALUE_LBN != 0);
1382 seq = EFX_QWORD_FIELD(*event, VFDI_EV_SEQ);
1383 type = EFX_QWORD_FIELD(*event, VFDI_EV_TYPE);
1384 data = EFX_QWORD_FIELD(*event, VFDI_EV_DATA);
1385
1386 netif_vdbg(efx, hw, efx->net_dev,
1387 "USR_EV event from qid %d seq 0x%x type %d data 0x%x\n",
1388 qid, seq, type, data);
1389
1390 if (map_vi_index(efx, qid, &vf, NULL))
1391 return;
1392 if (vf->busy)
1393 goto error;
1394
1395 if (type == VFDI_EV_TYPE_REQ_WORD0) {
1396 /* Resynchronise */
1397 vf->req_type = VFDI_EV_TYPE_REQ_WORD0;
1398 vf->req_seqno = seq + 1;
1399 vf->req_addr = 0;
1400 } else if (seq != (vf->req_seqno++ & 0xff) || type != vf->req_type)
1401 goto error;
1402
1403 switch (vf->req_type) {
1404 case VFDI_EV_TYPE_REQ_WORD0:
1405 case VFDI_EV_TYPE_REQ_WORD1:
1406 case VFDI_EV_TYPE_REQ_WORD2:
1407 vf->req_addr |= (u64)data << (vf->req_type << 4);
1408 ++vf->req_type;
1409 return;
1410
1411 case VFDI_EV_TYPE_REQ_WORD3:
1412 vf->req_addr |= (u64)data << 48;
1413 vf->req_type = VFDI_EV_TYPE_REQ_WORD0;
1414 vf->busy = true;
1415 queue_work(vfdi_workqueue, &vf->req);
1416 return;
1417 }
1418
1419 error:
1420 if (net_ratelimit())
1421 netif_err(efx, hw, efx->net_dev,
1422 "ERROR: Screaming VFDI request from %s\n",
1423 vf->pci_name);
1424 /* Reset the request and sequence number */
1425 vf->req_type = VFDI_EV_TYPE_REQ_WORD0;
1426 vf->req_seqno = seq + 1;
1427 }
1428
1429 void efx_sriov_flr(struct efx_nic *efx, unsigned vf_i)
1430 {
1431 struct efx_vf *vf;
1432
1433 if (vf_i > efx->vf_init_count)
1434 return;
1435 vf = efx->vf + vf_i;
1436 netif_info(efx, hw, efx->net_dev,
1437 "FLR on VF %s\n", vf->pci_name);
1438
1439 vf->status_addr = 0;
1440 efx_vfdi_remove_all_filters(vf);
1441 efx_vfdi_flush_clear(vf);
1442
1443 vf->evq0_count = 0;
1444 }
1445
1446 void efx_sriov_mac_address_changed(struct efx_nic *efx)
1447 {
1448 struct vfdi_status *vfdi_status = efx->vfdi_status.addr;
1449
1450 if (!efx->vf_init_count)
1451 return;
1452 memcpy(vfdi_status->peers[0].mac_addr,
1453 efx->net_dev->dev_addr, ETH_ALEN);
1454 queue_work(vfdi_workqueue, &efx->peer_work);
1455 }
1456
1457 void efx_sriov_tx_flush_done(struct efx_nic *efx, efx_qword_t *event)
1458 {
1459 struct efx_vf *vf;
1460 unsigned queue, qid;
1461
1462 queue = EFX_QWORD_FIELD(*event, FSF_AZ_DRIVER_EV_SUBDATA);
1463 if (map_vi_index(efx, queue, &vf, &qid))
1464 return;
1465 /* Ignore flush completions triggered by an FLR */
1466 if (!test_bit(qid, vf->txq_mask))
1467 return;
1468
1469 __clear_bit(qid, vf->txq_mask);
1470 --vf->txq_count;
1471
1472 if (efx_vfdi_flush_wake(vf))
1473 wake_up(&vf->flush_waitq);
1474 }
1475
1476 void efx_sriov_rx_flush_done(struct efx_nic *efx, efx_qword_t *event)
1477 {
1478 struct efx_vf *vf;
1479 unsigned ev_failed, queue, qid;
1480
1481 queue = EFX_QWORD_FIELD(*event, FSF_AZ_DRIVER_EV_RX_DESCQ_ID);
1482 ev_failed = EFX_QWORD_FIELD(*event,
1483 FSF_AZ_DRIVER_EV_RX_FLUSH_FAIL);
1484 if (map_vi_index(efx, queue, &vf, &qid))
1485 return;
1486 if (!test_bit(qid, vf->rxq_mask))
1487 return;
1488
1489 if (ev_failed) {
1490 set_bit(qid, vf->rxq_retry_mask);
1491 atomic_inc(&vf->rxq_retry_count);
1492 } else {
1493 __clear_bit(qid, vf->rxq_mask);
1494 --vf->rxq_count;
1495 }
1496 if (efx_vfdi_flush_wake(vf))
1497 wake_up(&vf->flush_waitq);
1498 }
1499
1500 /* Called from napi. Schedule the reset work item */
1501 void efx_sriov_desc_fetch_err(struct efx_nic *efx, unsigned dmaq)
1502 {
1503 struct efx_vf *vf;
1504 unsigned int rel;
1505
1506 if (map_vi_index(efx, dmaq, &vf, &rel))
1507 return;
1508
1509 if (net_ratelimit())
1510 netif_err(efx, hw, efx->net_dev,
1511 "VF %d DMA Q %d reports descriptor fetch error.\n",
1512 vf->index, rel);
1513 queue_work(vfdi_workqueue, &vf->reset_work);
1514 }
1515
1516 /* Reset all VFs */
1517 void efx_sriov_reset(struct efx_nic *efx)
1518 {
1519 unsigned int vf_i;
1520 struct efx_buffer buf;
1521 struct efx_vf *vf;
1522
1523 ASSERT_RTNL();
1524
1525 if (efx->vf_init_count == 0)
1526 return;
1527
1528 efx_sriov_usrev(efx, true);
1529 (void)efx_sriov_cmd(efx, true, NULL, NULL);
1530
1531 if (efx_nic_alloc_buffer(efx, &buf, EFX_PAGE_SIZE))
1532 return;
1533
1534 for (vf_i = 0; vf_i < efx->vf_init_count; ++vf_i) {
1535 vf = efx->vf + vf_i;
1536 efx_sriov_reset_vf(vf, &buf);
1537 }
1538
1539 efx_nic_free_buffer(efx, &buf);
1540 }
1541
1542 int efx_init_sriov(void)
1543 {
1544 /* A single threaded workqueue is sufficient. efx_sriov_vfdi() and
1545 * efx_sriov_peer_work() spend almost all their time sleeping for
1546 * MCDI to complete anyway
1547 */
1548 vfdi_workqueue = create_singlethread_workqueue("sfc_vfdi");
1549 if (!vfdi_workqueue)
1550 return -ENOMEM;
1551
1552 return 0;
1553 }
1554
1555 void efx_fini_sriov(void)
1556 {
1557 destroy_workqueue(vfdi_workqueue);
1558 }
1559
1560 int efx_sriov_set_vf_mac(struct net_device *net_dev, int vf_i, u8 *mac)
1561 {
1562 struct efx_nic *efx = netdev_priv(net_dev);
1563 struct efx_vf *vf;
1564
1565 if (vf_i >= efx->vf_init_count)
1566 return -EINVAL;
1567 vf = efx->vf + vf_i;
1568
1569 mutex_lock(&vf->status_lock);
1570 memcpy(vf->addr.mac_addr, mac, ETH_ALEN);
1571 __efx_sriov_update_vf_addr(vf);
1572 mutex_unlock(&vf->status_lock);
1573
1574 return 0;
1575 }
1576
1577 int efx_sriov_set_vf_vlan(struct net_device *net_dev, int vf_i,
1578 u16 vlan, u8 qos)
1579 {
1580 struct efx_nic *efx = netdev_priv(net_dev);
1581 struct efx_vf *vf;
1582 u16 tci;
1583
1584 if (vf_i >= efx->vf_init_count)
1585 return -EINVAL;
1586 vf = efx->vf + vf_i;
1587
1588 mutex_lock(&vf->status_lock);
1589 tci = (vlan & VLAN_VID_MASK) | ((qos & 0x7) << VLAN_PRIO_SHIFT);
1590 vf->addr.tci = htons(tci);
1591 __efx_sriov_update_vf_addr(vf);
1592 mutex_unlock(&vf->status_lock);
1593
1594 return 0;
1595 }
1596
1597 int efx_sriov_set_vf_spoofchk(struct net_device *net_dev, int vf_i,
1598 bool spoofchk)
1599 {
1600 struct efx_nic *efx = netdev_priv(net_dev);
1601 struct efx_vf *vf;
1602 int rc;
1603
1604 if (vf_i >= efx->vf_init_count)
1605 return -EINVAL;
1606 vf = efx->vf + vf_i;
1607
1608 mutex_lock(&vf->txq_lock);
1609 if (vf->txq_count == 0) {
1610 vf->tx_filter_mode =
1611 spoofchk ? VF_TX_FILTER_ON : VF_TX_FILTER_OFF;
1612 rc = 0;
1613 } else {
1614 /* This cannot be changed while TX queues are running */
1615 rc = -EBUSY;
1616 }
1617 mutex_unlock(&vf->txq_lock);
1618 return rc;
1619 }
1620
1621 int efx_sriov_get_vf_config(struct net_device *net_dev, int vf_i,
1622 struct ifla_vf_info *ivi)
1623 {
1624 struct efx_nic *efx = netdev_priv(net_dev);
1625 struct efx_vf *vf;
1626 u16 tci;
1627
1628 if (vf_i >= efx->vf_init_count)
1629 return -EINVAL;
1630 vf = efx->vf + vf_i;
1631
1632 ivi->vf = vf_i;
1633 memcpy(ivi->mac, vf->addr.mac_addr, ETH_ALEN);
1634 ivi->tx_rate = 0;
1635 tci = ntohs(vf->addr.tci);
1636 ivi->vlan = tci & VLAN_VID_MASK;
1637 ivi->qos = (tci >> VLAN_PRIO_SHIFT) & 0x7;
1638 ivi->spoofchk = vf->tx_filter_mode == VF_TX_FILTER_ON;
1639
1640 return 0;
1641 }
1642
Linux kernel - Deliverables
This page took 0.065624 seconds and 6 git commands to generate.