2f3ff73b1d1ba736a72ccdd14b9c0526a21fb01a
[cascardo/linux.git] / drivers / ntb / ntb_hw.c
1 /*
2  * This file is provided under a dual BSD/GPLv2 license.  When using or
3  *   redistributing this file, you may do so under either license.
4  *
5  *   GPL LICENSE SUMMARY
6  *
7  *   Copyright(c) 2012 Intel Corporation. All rights reserved.
8  *
9  *   This program is free software; you can redistribute it and/or modify
10  *   it under the terms of version 2 of the GNU General Public License as
11  *   published by the Free Software Foundation.
12  *
13  *   BSD LICENSE
14  *
15  *   Copyright(c) 2012 Intel Corporation. All rights reserved.
16  *
17  *   Redistribution and use in source and binary forms, with or without
18  *   modification, are permitted provided that the following conditions
19  *   are met:
20  *
21  *     * Redistributions of source code must retain the above copyright
22  *       notice, this list of conditions and the following disclaimer.
23  *     * Redistributions in binary form must reproduce the above copy
24  *       notice, this list of conditions and the following disclaimer in
25  *       the documentation and/or other materials provided with the
26  *       distribution.
27  *     * Neither the name of Intel Corporation nor the names of its
28  *       contributors may be used to endorse or promote products derived
29  *       from this software without specific prior written permission.
30  *
31  *   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
32  *   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
33  *   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
34  *   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
35  *   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
36  *   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
37  *   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
38  *   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
39  *   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
40  *   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
41  *   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
42  *
43  * Intel PCIe NTB Linux driver
44  *
45  * Contact Information:
46  * Jon Mason <jon.mason@intel.com>
47  */
48 #include <linux/debugfs.h>
49 #include <linux/delay.h>
50 #include <linux/init.h>
51 #include <linux/interrupt.h>
52 #include <linux/module.h>
53 #include <linux/pci.h>
54 #include <linux/random.h>
55 #include <linux/slab.h>
56 #include "ntb_hw.h"
57 #include "ntb_regs.h"
58
59 #define NTB_NAME        "Intel(R) PCI-E Non-Transparent Bridge Driver"
60 #define NTB_VER         "0.25"
61
62 MODULE_DESCRIPTION(NTB_NAME);
63 MODULE_VERSION(NTB_VER);
64 MODULE_LICENSE("Dual BSD/GPL");
65 MODULE_AUTHOR("Intel Corporation");
66
67 static bool xeon_errata_workaround = true;
68 module_param(xeon_errata_workaround, bool, 0644);
69 MODULE_PARM_DESC(xeon_errata_workaround, "Workaround for the Xeon Errata");
70
71 enum {
72         NTB_CONN_TRANSPARENT = 0,
73         NTB_CONN_B2B,
74         NTB_CONN_RP,
75 };
76
77 enum {
78         NTB_DEV_USD = 0,
79         NTB_DEV_DSD,
80 };
81
82 enum {
83         SNB_HW = 0,
84         BWD_HW,
85 };
86
87 static struct dentry *debugfs_dir;
88
89 #define BWD_LINK_RECOVERY_TIME  500
90
91 /* Translate memory window 0,1 to BAR 2,4 */
92 #define MW_TO_BAR(mw)   (mw * NTB_MAX_NUM_MW + 2)
93
94 static DEFINE_PCI_DEVICE_TABLE(ntb_pci_tbl) = {
95         {PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_NTB_B2B_BWD)},
96         {PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_NTB_B2B_JSF)},
97         {PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_NTB_B2B_SNB)},
98         {PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_NTB_B2B_IVT)},
99         {PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_NTB_B2B_HSX)},
100         {PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_NTB_PS_JSF)},
101         {PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_NTB_PS_SNB)},
102         {PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_NTB_PS_IVT)},
103         {PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_NTB_PS_HSX)},
104         {PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_NTB_SS_JSF)},
105         {PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_NTB_SS_SNB)},
106         {PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_NTB_SS_IVT)},
107         {PCI_VDEVICE(INTEL, PCI_DEVICE_ID_INTEL_NTB_SS_HSX)},
108         {0}
109 };
110 MODULE_DEVICE_TABLE(pci, ntb_pci_tbl);
111
112 /**
113  * ntb_register_event_callback() - register event callback
114  * @ndev: pointer to ntb_device instance
115  * @func: callback function to register
116  *
117  * This function registers a callback for any HW driver events such as link
118  * up/down, power management notices and etc.
119  *
120  * RETURNS: An appropriate -ERRNO error value on error, or zero for success.
121  */
122 int ntb_register_event_callback(struct ntb_device *ndev,
123                             void (*func)(void *handle, enum ntb_hw_event event))
124 {
125         if (ndev->event_cb)
126                 return -EINVAL;
127
128         ndev->event_cb = func;
129
130         return 0;
131 }
132
133 /**
134  * ntb_unregister_event_callback() - unregisters the event callback
135  * @ndev: pointer to ntb_device instance
136  *
137  * This function unregisters the existing callback from transport
138  */
139 void ntb_unregister_event_callback(struct ntb_device *ndev)
140 {
141         ndev->event_cb = NULL;
142 }
143
144 /**
145  * ntb_register_db_callback() - register a callback for doorbell interrupt
146  * @ndev: pointer to ntb_device instance
147  * @idx: doorbell index to register callback, zero based
148  * @func: callback function to register
149  *
150  * This function registers a callback function for the doorbell interrupt
151  * on the primary side. The function will unmask the doorbell as well to
152  * allow interrupt.
153  *
154  * RETURNS: An appropriate -ERRNO error value on error, or zero for success.
155  */
156 int ntb_register_db_callback(struct ntb_device *ndev, unsigned int idx,
157                              void *data, void (*func)(void *data, int db_num))
158 {
159         unsigned long mask;
160
161         if (idx >= ndev->max_cbs || ndev->db_cb[idx].callback) {
162                 dev_warn(&ndev->pdev->dev, "Invalid Index.\n");
163                 return -EINVAL;
164         }
165
166         ndev->db_cb[idx].callback = func;
167         ndev->db_cb[idx].data = data;
168
169         /* unmask interrupt */
170         mask = readw(ndev->reg_ofs.ldb_mask);
171         clear_bit(idx * ndev->bits_per_vector, &mask);
172         writew(mask, ndev->reg_ofs.ldb_mask);
173
174         return 0;
175 }
176
177 /**
178  * ntb_unregister_db_callback() - unregister a callback for doorbell interrupt
179  * @ndev: pointer to ntb_device instance
180  * @idx: doorbell index to register callback, zero based
181  *
182  * This function unregisters a callback function for the doorbell interrupt
183  * on the primary side. The function will also mask the said doorbell.
184  */
185 void ntb_unregister_db_callback(struct ntb_device *ndev, unsigned int idx)
186 {
187         unsigned long mask;
188
189         if (idx >= ndev->max_cbs || !ndev->db_cb[idx].callback)
190                 return;
191
192         mask = readw(ndev->reg_ofs.ldb_mask);
193         set_bit(idx * ndev->bits_per_vector, &mask);
194         writew(mask, ndev->reg_ofs.ldb_mask);
195
196         ndev->db_cb[idx].callback = NULL;
197 }
198
199 /**
200  * ntb_find_transport() - find the transport pointer
201  * @transport: pointer to pci device
202  *
203  * Given the pci device pointer, return the transport pointer passed in when
204  * the transport attached when it was inited.
205  *
206  * RETURNS: pointer to transport.
207  */
208 void *ntb_find_transport(struct pci_dev *pdev)
209 {
210         struct ntb_device *ndev = pci_get_drvdata(pdev);
211         return ndev->ntb_transport;
212 }
213
214 /**
215  * ntb_register_transport() - Register NTB transport with NTB HW driver
216  * @transport: transport identifier
217  *
218  * This function allows a transport to reserve the hardware driver for
219  * NTB usage.
220  *
221  * RETURNS: pointer to ntb_device, NULL on error.
222  */
223 struct ntb_device *ntb_register_transport(struct pci_dev *pdev, void *transport)
224 {
225         struct ntb_device *ndev = pci_get_drvdata(pdev);
226
227         if (ndev->ntb_transport)
228                 return NULL;
229
230         ndev->ntb_transport = transport;
231         return ndev;
232 }
233
234 /**
235  * ntb_unregister_transport() - Unregister the transport with the NTB HW driver
236  * @ndev - ntb_device of the transport to be freed
237  *
238  * This function unregisters the transport from the HW driver and performs any
239  * necessary cleanups.
240  */
241 void ntb_unregister_transport(struct ntb_device *ndev)
242 {
243         int i;
244
245         if (!ndev->ntb_transport)
246                 return;
247
248         for (i = 0; i < ndev->max_cbs; i++)
249                 ntb_unregister_db_callback(ndev, i);
250
251         ntb_unregister_event_callback(ndev);
252         ndev->ntb_transport = NULL;
253 }
254
255 /**
256  * ntb_write_local_spad() - write to the secondary scratchpad register
257  * @ndev: pointer to ntb_device instance
258  * @idx: index to the scratchpad register, 0 based
259  * @val: the data value to put into the register
260  *
261  * This function allows writing of a 32bit value to the indexed scratchpad
262  * register. This writes over the data mirrored to the local scratchpad register
263  * by the remote system.
264  *
265  * RETURNS: An appropriate -ERRNO error value on error, or zero for success.
266  */
267 int ntb_write_local_spad(struct ntb_device *ndev, unsigned int idx, u32 val)
268 {
269         if (idx >= ndev->limits.max_spads)
270                 return -EINVAL;
271
272         dev_dbg(&ndev->pdev->dev, "Writing %x to local scratch pad index %d\n",
273                 val, idx);
274         writel(val, ndev->reg_ofs.spad_read + idx * 4);
275
276         return 0;
277 }
278
279 /**
280  * ntb_read_local_spad() - read from the primary scratchpad register
281  * @ndev: pointer to ntb_device instance
282  * @idx: index to scratchpad register, 0 based
283  * @val: pointer to 32bit integer for storing the register value
284  *
285  * This function allows reading of the 32bit scratchpad register on
286  * the primary (internal) side.  This allows the local system to read data
287  * written and mirrored to the scratchpad register by the remote system.
288  *
289  * RETURNS: An appropriate -ERRNO error value on error, or zero for success.
290  */
291 int ntb_read_local_spad(struct ntb_device *ndev, unsigned int idx, u32 *val)
292 {
293         if (idx >= ndev->limits.max_spads)
294                 return -EINVAL;
295
296         *val = readl(ndev->reg_ofs.spad_write + idx * 4);
297         dev_dbg(&ndev->pdev->dev,
298                 "Reading %x from local scratch pad index %d\n", *val, idx);
299
300         return 0;
301 }
302
303 /**
304  * ntb_write_remote_spad() - write to the secondary scratchpad register
305  * @ndev: pointer to ntb_device instance
306  * @idx: index to the scratchpad register, 0 based
307  * @val: the data value to put into the register
308  *
309  * This function allows writing of a 32bit value to the indexed scratchpad
310  * register. The register resides on the secondary (external) side.  This allows
311  * the local system to write data to be mirrored to the remote systems
312  * scratchpad register.
313  *
314  * RETURNS: An appropriate -ERRNO error value on error, or zero for success.
315  */
316 int ntb_write_remote_spad(struct ntb_device *ndev, unsigned int idx, u32 val)
317 {
318         if (idx >= ndev->limits.max_spads)
319                 return -EINVAL;
320
321         dev_dbg(&ndev->pdev->dev, "Writing %x to remote scratch pad index %d\n",
322                 val, idx);
323         writel(val, ndev->reg_ofs.spad_write + idx * 4);
324
325         return 0;
326 }
327
328 /**
329  * ntb_read_remote_spad() - read from the primary scratchpad register
330  * @ndev: pointer to ntb_device instance
331  * @idx: index to scratchpad register, 0 based
332  * @val: pointer to 32bit integer for storing the register value
333  *
334  * This function allows reading of the 32bit scratchpad register on
335  * the primary (internal) side.  This alloows the local system to read the data
336  * it wrote to be mirrored on the remote system.
337  *
338  * RETURNS: An appropriate -ERRNO error value on error, or zero for success.
339  */
340 int ntb_read_remote_spad(struct ntb_device *ndev, unsigned int idx, u32 *val)
341 {
342         if (idx >= ndev->limits.max_spads)
343                 return -EINVAL;
344
345         *val = readl(ndev->reg_ofs.spad_read + idx * 4);
346         dev_dbg(&ndev->pdev->dev,
347                 "Reading %x from remote scratch pad index %d\n", *val, idx);
348
349         return 0;
350 }
351
352 /**
353  * ntb_get_mw_base() - get addr for the NTB memory window
354  * @ndev: pointer to ntb_device instance
355  * @mw: memory window number
356  *
357  * This function provides the base address of the memory window specified.
358  *
359  * RETURNS: address, or NULL on error.
360  */
361 resource_size_t ntb_get_mw_base(struct ntb_device *ndev, unsigned int mw)
362 {
363         if (mw >= ntb_max_mw(ndev))
364                 return 0;
365
366         return pci_resource_start(ndev->pdev, MW_TO_BAR(mw));
367 }
368
369 /**
370  * ntb_get_mw_vbase() - get virtual addr for the NTB memory window
371  * @ndev: pointer to ntb_device instance
372  * @mw: memory window number
373  *
374  * This function provides the base virtual address of the memory window
375  * specified.
376  *
377  * RETURNS: pointer to virtual address, or NULL on error.
378  */
379 void __iomem *ntb_get_mw_vbase(struct ntb_device *ndev, unsigned int mw)
380 {
381         if (mw >= ntb_max_mw(ndev))
382                 return NULL;
383
384         return ndev->mw[mw].vbase;
385 }
386
387 /**
388  * ntb_get_mw_size() - return size of NTB memory window
389  * @ndev: pointer to ntb_device instance
390  * @mw: memory window number
391  *
392  * This function provides the physical size of the memory window specified
393  *
394  * RETURNS: the size of the memory window or zero on error
395  */
396 u64 ntb_get_mw_size(struct ntb_device *ndev, unsigned int mw)
397 {
398         if (mw >= ntb_max_mw(ndev))
399                 return 0;
400
401         return ndev->mw[mw].bar_sz;
402 }
403
404 /**
405  * ntb_set_mw_addr - set the memory window address
406  * @ndev: pointer to ntb_device instance
407  * @mw: memory window number
408  * @addr: base address for data
409  *
410  * This function sets the base physical address of the memory window.  This
411  * memory address is where data from the remote system will be transfered into
412  * or out of depending on how the transport is configured.
413  */
414 void ntb_set_mw_addr(struct ntb_device *ndev, unsigned int mw, u64 addr)
415 {
416         if (mw >= ntb_max_mw(ndev))
417                 return;
418
419         dev_dbg(&ndev->pdev->dev, "Writing addr %Lx to BAR %d\n", addr,
420                 MW_TO_BAR(mw));
421
422         ndev->mw[mw].phys_addr = addr;
423
424         switch (MW_TO_BAR(mw)) {
425         case NTB_BAR_23:
426                 writeq(addr, ndev->reg_ofs.bar2_xlat);
427                 break;
428         case NTB_BAR_45:
429                 writeq(addr, ndev->reg_ofs.bar4_xlat);
430                 break;
431         }
432 }
433
434 /**
435  * ntb_ring_doorbell() - Set the doorbell on the secondary/external side
436  * @ndev: pointer to ntb_device instance
437  * @db: doorbell to ring
438  *
439  * This function allows triggering of a doorbell on the secondary/external
440  * side that will initiate an interrupt on the remote host
441  *
442  * RETURNS: An appropriate -ERRNO error value on error, or zero for success.
443  */
444 void ntb_ring_doorbell(struct ntb_device *ndev, unsigned int db)
445 {
446         dev_dbg(&ndev->pdev->dev, "%s: ringing doorbell %d\n", __func__, db);
447
448         if (ndev->hw_type == BWD_HW)
449                 writeq((u64) 1 << db, ndev->reg_ofs.rdb);
450         else
451                 writew(((1 << ndev->bits_per_vector) - 1) <<
452                        (db * ndev->bits_per_vector), ndev->reg_ofs.rdb);
453 }
454
455 static void bwd_recover_link(struct ntb_device *ndev)
456 {
457         u32 status;
458
459         /* Driver resets the NTB ModPhy lanes - magic! */
460         writeb(0xe0, ndev->reg_base + BWD_MODPHY_PCSREG6);
461         writeb(0x40, ndev->reg_base + BWD_MODPHY_PCSREG4);
462         writeb(0x60, ndev->reg_base + BWD_MODPHY_PCSREG4);
463         writeb(0x60, ndev->reg_base + BWD_MODPHY_PCSREG6);
464
465         /* Driver waits 100ms to allow the NTB ModPhy to settle */
466         msleep(100);
467
468         /* Clear AER Errors, write to clear */
469         status = readl(ndev->reg_base + BWD_ERRCORSTS_OFFSET);
470         dev_dbg(&ndev->pdev->dev, "ERRCORSTS = %x\n", status);
471         status &= PCI_ERR_COR_REP_ROLL;
472         writel(status, ndev->reg_base + BWD_ERRCORSTS_OFFSET);
473
474         /* Clear unexpected electrical idle event in LTSSM, write to clear */
475         status = readl(ndev->reg_base + BWD_LTSSMERRSTS0_OFFSET);
476         dev_dbg(&ndev->pdev->dev, "LTSSMERRSTS0 = %x\n", status);
477         status |= BWD_LTSSMERRSTS0_UNEXPECTEDEI;
478         writel(status, ndev->reg_base + BWD_LTSSMERRSTS0_OFFSET);
479
480         /* Clear DeSkew Buffer error, write to clear */
481         status = readl(ndev->reg_base + BWD_DESKEWSTS_OFFSET);
482         dev_dbg(&ndev->pdev->dev, "DESKEWSTS = %x\n", status);
483         status |= BWD_DESKEWSTS_DBERR;
484         writel(status, ndev->reg_base + BWD_DESKEWSTS_OFFSET);
485
486         status = readl(ndev->reg_base + BWD_IBSTERRRCRVSTS0_OFFSET);
487         dev_dbg(&ndev->pdev->dev, "IBSTERRRCRVSTS0 = %x\n", status);
488         status &= BWD_IBIST_ERR_OFLOW;
489         writel(status, ndev->reg_base + BWD_IBSTERRRCRVSTS0_OFFSET);
490
491         /* Releases the NTB state machine to allow the link to retrain */
492         status = readl(ndev->reg_base + BWD_LTSSMSTATEJMP_OFFSET);
493         dev_dbg(&ndev->pdev->dev, "LTSSMSTATEJMP = %x\n", status);
494         status &= ~BWD_LTSSMSTATEJMP_FORCEDETECT;
495         writel(status, ndev->reg_base + BWD_LTSSMSTATEJMP_OFFSET);
496 }
497
498 static void ntb_link_event(struct ntb_device *ndev, int link_state)
499 {
500         unsigned int event;
501
502         if (ndev->link_status == link_state)
503                 return;
504
505         if (link_state == NTB_LINK_UP) {
506                 u16 status;
507
508                 dev_info(&ndev->pdev->dev, "Link Up\n");
509                 ndev->link_status = NTB_LINK_UP;
510                 event = NTB_EVENT_HW_LINK_UP;
511
512                 if (ndev->hw_type == BWD_HW ||
513                     ndev->conn_type == NTB_CONN_TRANSPARENT)
514                         status = readw(ndev->reg_ofs.lnk_stat);
515                 else {
516                         int rc = pci_read_config_word(ndev->pdev,
517                                                       SNB_LINK_STATUS_OFFSET,
518                                                       &status);
519                         if (rc)
520                                 return;
521                 }
522
523                 ndev->link_width = (status & NTB_LINK_WIDTH_MASK) >> 4;
524                 ndev->link_speed = (status & NTB_LINK_SPEED_MASK);
525                 dev_info(&ndev->pdev->dev, "Link Width %d, Link Speed %d\n",
526                          ndev->link_width, ndev->link_speed);
527         } else {
528                 dev_info(&ndev->pdev->dev, "Link Down\n");
529                 ndev->link_status = NTB_LINK_DOWN;
530                 event = NTB_EVENT_HW_LINK_DOWN;
531                 /* Don't modify link width/speed, we need it in link recovery */
532         }
533
534         /* notify the upper layer if we have an event change */
535         if (ndev->event_cb)
536                 ndev->event_cb(ndev->ntb_transport, event);
537 }
538
539 static int ntb_link_status(struct ntb_device *ndev)
540 {
541         int link_state;
542
543         if (ndev->hw_type == BWD_HW) {
544                 u32 ntb_cntl;
545
546                 ntb_cntl = readl(ndev->reg_ofs.lnk_cntl);
547                 if (ntb_cntl & BWD_CNTL_LINK_DOWN)
548                         link_state = NTB_LINK_DOWN;
549                 else
550                         link_state = NTB_LINK_UP;
551         } else {
552                 u16 status;
553                 int rc;
554
555                 rc = pci_read_config_word(ndev->pdev, SNB_LINK_STATUS_OFFSET,
556                                           &status);
557                 if (rc)
558                         return rc;
559
560                 if (status & NTB_LINK_STATUS_ACTIVE)
561                         link_state = NTB_LINK_UP;
562                 else
563                         link_state = NTB_LINK_DOWN;
564         }
565
566         ntb_link_event(ndev, link_state);
567
568         return 0;
569 }
570
571 static void bwd_link_recovery(struct work_struct *work)
572 {
573         struct ntb_device *ndev = container_of(work, struct ntb_device,
574                                                lr_timer.work);
575         u32 status32;
576
577         bwd_recover_link(ndev);
578         /* There is a potential race between the 2 NTB devices recovering at the
579          * same time.  If the times are the same, the link will not recover and
580          * the driver will be stuck in this loop forever.  Add a random interval
581          * to the recovery time to prevent this race.
582          */
583         msleep(BWD_LINK_RECOVERY_TIME + prandom_u32() % BWD_LINK_RECOVERY_TIME);
584
585         status32 = readl(ndev->reg_base + BWD_LTSSMSTATEJMP_OFFSET);
586         if (status32 & BWD_LTSSMSTATEJMP_FORCEDETECT)
587                 goto retry;
588
589         status32 = readl(ndev->reg_base + BWD_IBSTERRRCRVSTS0_OFFSET);
590         if (status32 & BWD_IBIST_ERR_OFLOW)
591                 goto retry;
592
593         status32 = readl(ndev->reg_ofs.lnk_cntl);
594         if (!(status32 & BWD_CNTL_LINK_DOWN)) {
595                 unsigned char speed, width;
596                 u16 status16;
597
598                 status16 = readw(ndev->reg_ofs.lnk_stat);
599                 width = (status16 & NTB_LINK_WIDTH_MASK) >> 4;
600                 speed = (status16 & NTB_LINK_SPEED_MASK);
601                 if (ndev->link_width != width || ndev->link_speed != speed)
602                         goto retry;
603         }
604
605         schedule_delayed_work(&ndev->hb_timer, NTB_HB_TIMEOUT);
606         return;
607
608 retry:
609         schedule_delayed_work(&ndev->lr_timer, NTB_HB_TIMEOUT);
610 }
611
612 /* BWD doesn't have link status interrupt, poll on that platform */
613 static void bwd_link_poll(struct work_struct *work)
614 {
615         struct ntb_device *ndev = container_of(work, struct ntb_device,
616                                                hb_timer.work);
617         unsigned long ts = jiffies;
618
619         /* If we haven't gotten an interrupt in a while, check the BWD link
620          * status bit
621          */
622         if (ts > ndev->last_ts + NTB_HB_TIMEOUT) {
623                 int rc = ntb_link_status(ndev);
624                 if (rc)
625                         dev_err(&ndev->pdev->dev,
626                                 "Error determining link status\n");
627
628                 /* Check to see if a link error is the cause of the link down */
629                 if (ndev->link_status == NTB_LINK_DOWN) {
630                         u32 status32 = readl(ndev->reg_base +
631                                              BWD_LTSSMSTATEJMP_OFFSET);
632                         if (status32 & BWD_LTSSMSTATEJMP_FORCEDETECT) {
633                                 schedule_delayed_work(&ndev->lr_timer, 0);
634                                 return;
635                         }
636                 }
637         }
638
639         schedule_delayed_work(&ndev->hb_timer, NTB_HB_TIMEOUT);
640 }
641
642 static int ntb_xeon_setup(struct ntb_device *ndev)
643 {
644         int rc;
645         u8 val;
646
647         ndev->hw_type = SNB_HW;
648
649         rc = pci_read_config_byte(ndev->pdev, NTB_PPD_OFFSET, &val);
650         if (rc)
651                 return rc;
652
653         if (val & SNB_PPD_DEV_TYPE)
654                 ndev->dev_type = NTB_DEV_USD;
655         else
656                 ndev->dev_type = NTB_DEV_DSD;
657
658         switch (val & SNB_PPD_CONN_TYPE) {
659         case NTB_CONN_B2B:
660                 dev_info(&ndev->pdev->dev, "Conn Type = B2B\n");
661                 ndev->conn_type = NTB_CONN_B2B;
662                 ndev->reg_ofs.ldb = ndev->reg_base + SNB_PDOORBELL_OFFSET;
663                 ndev->reg_ofs.ldb_mask = ndev->reg_base + SNB_PDBMSK_OFFSET;
664                 ndev->reg_ofs.spad_read = ndev->reg_base + SNB_SPAD_OFFSET;
665                 ndev->reg_ofs.bar2_xlat = ndev->reg_base + SNB_SBAR2XLAT_OFFSET;
666                 ndev->reg_ofs.bar4_xlat = ndev->reg_base + SNB_SBAR4XLAT_OFFSET;
667                 ndev->limits.max_spads = SNB_MAX_B2B_SPADS;
668
669                 /* There is a Xeon hardware errata related to writes to
670                  * SDOORBELL or B2BDOORBELL in conjunction with inbound access
671                  * to NTB MMIO Space, which may hang the system.  To workaround
672                  * this use the second memory window to access the interrupt and
673                  * scratch pad registers on the remote system.
674                  */
675                 if (xeon_errata_workaround) {
676                         if (!ndev->mw[1].bar_sz)
677                                 return -EINVAL;
678
679                         ndev->limits.max_mw = SNB_ERRATA_MAX_MW;
680                         ndev->reg_ofs.spad_write = ndev->mw[1].vbase +
681                                                    SNB_SPAD_OFFSET;
682                         ndev->reg_ofs.rdb = ndev->mw[1].vbase +
683                                             SNB_PDOORBELL_OFFSET;
684
685                         /* Set the Limit register to 4k, the minimum size, to
686                          * prevent an illegal access
687                          */
688                         writeq(ndev->mw[1].bar_sz + 0x1000, ndev->reg_base +
689                                SNB_PBAR4LMT_OFFSET);
690                 } else {
691                         ndev->limits.max_mw = SNB_MAX_MW;
692                         ndev->reg_ofs.spad_write = ndev->reg_base +
693                                                    SNB_B2B_SPAD_OFFSET;
694                         ndev->reg_ofs.rdb = ndev->reg_base +
695                                             SNB_B2B_DOORBELL_OFFSET;
696
697                         /* Disable the Limit register, just incase it is set to
698                          * something silly
699                          */
700                         writeq(0, ndev->reg_base + SNB_PBAR4LMT_OFFSET);
701                 }
702
703                 /* The Xeon errata workaround requires setting SBAR Base
704                  * addresses to known values, so that the PBAR XLAT can be
705                  * pointed at SBAR0 of the remote system.
706                  */
707                 if (ndev->dev_type == NTB_DEV_USD) {
708                         writeq(SNB_MBAR23_DSD_ADDR, ndev->reg_base +
709                                SNB_PBAR2XLAT_OFFSET);
710                         if (xeon_errata_workaround)
711                                 writeq(SNB_MBAR01_DSD_ADDR, ndev->reg_base +
712                                        SNB_PBAR4XLAT_OFFSET);
713                         else {
714                                 writeq(SNB_MBAR45_DSD_ADDR, ndev->reg_base +
715                                        SNB_PBAR4XLAT_OFFSET);
716                                 /* B2B_XLAT_OFFSET is a 64bit register, but can
717                                  * only take 32bit writes
718                                  */
719                                 writel(SNB_MBAR01_DSD_ADDR & 0xffffffff,
720                                        ndev->reg_base + SNB_B2B_XLAT_OFFSETL);
721                                 writel(SNB_MBAR01_DSD_ADDR >> 32,
722                                        ndev->reg_base + SNB_B2B_XLAT_OFFSETU);
723                         }
724
725                         writeq(SNB_MBAR01_USD_ADDR, ndev->reg_base +
726                                SNB_SBAR0BASE_OFFSET);
727                         writeq(SNB_MBAR23_USD_ADDR, ndev->reg_base +
728                                SNB_SBAR2BASE_OFFSET);
729                         writeq(SNB_MBAR45_USD_ADDR, ndev->reg_base +
730                                SNB_SBAR4BASE_OFFSET);
731                 } else {
732                         writeq(SNB_MBAR23_USD_ADDR, ndev->reg_base +
733                                SNB_PBAR2XLAT_OFFSET);
734                         if (xeon_errata_workaround)
735                                 writeq(SNB_MBAR01_USD_ADDR, ndev->reg_base +
736                                        SNB_PBAR4XLAT_OFFSET);
737                         else {
738                                 writeq(SNB_MBAR45_USD_ADDR, ndev->reg_base +
739                                        SNB_PBAR4XLAT_OFFSET);
740                                 /* B2B_XLAT_OFFSET is a 64bit register, but can
741                                  * only take 32bit writes
742                                  */
743                                 writel(SNB_MBAR01_DSD_ADDR & 0xffffffff,
744                                        ndev->reg_base + SNB_B2B_XLAT_OFFSETL);
745                                 writel(SNB_MBAR01_USD_ADDR >> 32,
746                                        ndev->reg_base + SNB_B2B_XLAT_OFFSETU);
747                         }
748                         writeq(SNB_MBAR01_DSD_ADDR, ndev->reg_base +
749                                SNB_SBAR0BASE_OFFSET);
750                         writeq(SNB_MBAR23_DSD_ADDR, ndev->reg_base +
751                                SNB_SBAR2BASE_OFFSET);
752                         writeq(SNB_MBAR45_DSD_ADDR, ndev->reg_base +
753                                SNB_SBAR4BASE_OFFSET);
754                 }
755                 break;
756         case NTB_CONN_RP:
757                 dev_info(&ndev->pdev->dev, "Conn Type = RP\n");
758                 ndev->conn_type = NTB_CONN_RP;
759
760                 if (xeon_errata_workaround) {
761                         dev_err(&ndev->pdev->dev, 
762                                 "NTB-RP disabled due to hardware errata.  To disregard this warning and potentially lock-up the system, add the parameter 'xeon_errata_workaround=0'.\n");
763                         return -EINVAL;
764                 }
765
766                 /* Scratch pads need to have exclusive access from the primary
767                  * or secondary side.  Halve the num spads so that each side can
768                  * have an equal amount.
769                  */
770                 ndev->limits.max_spads = SNB_MAX_COMPAT_SPADS / 2;
771                 /* Note: The SDOORBELL is the cause of the errata.  You REALLY
772                  * don't want to touch it.
773                  */
774                 ndev->reg_ofs.rdb = ndev->reg_base + SNB_SDOORBELL_OFFSET;
775                 ndev->reg_ofs.ldb = ndev->reg_base + SNB_PDOORBELL_OFFSET;
776                 ndev->reg_ofs.ldb_mask = ndev->reg_base + SNB_PDBMSK_OFFSET;
777                 /* Offset the start of the spads to correspond to whether it is
778                  * primary or secondary
779                  */
780                 ndev->reg_ofs.spad_write = ndev->reg_base + SNB_SPAD_OFFSET +
781                                            ndev->limits.max_spads * 4;
782                 ndev->reg_ofs.spad_read = ndev->reg_base + SNB_SPAD_OFFSET;
783                 ndev->reg_ofs.bar2_xlat = ndev->reg_base + SNB_SBAR2XLAT_OFFSET;
784                 ndev->reg_ofs.bar4_xlat = ndev->reg_base + SNB_SBAR4XLAT_OFFSET;
785                 ndev->limits.max_mw = SNB_MAX_MW;
786                 break;
787         case NTB_CONN_TRANSPARENT:
788                 dev_info(&ndev->pdev->dev, "Conn Type = TRANSPARENT\n");
789                 ndev->conn_type = NTB_CONN_TRANSPARENT;
790                 /* Scratch pads need to have exclusive access from the primary
791                  * or secondary side.  Halve the num spads so that each side can
792                  * have an equal amount.
793                  */
794                 ndev->limits.max_spads = SNB_MAX_COMPAT_SPADS / 2;
795                 ndev->reg_ofs.rdb = ndev->reg_base + SNB_PDOORBELL_OFFSET;
796                 ndev->reg_ofs.ldb = ndev->reg_base + SNB_SDOORBELL_OFFSET;
797                 ndev->reg_ofs.ldb_mask = ndev->reg_base + SNB_SDBMSK_OFFSET;
798                 ndev->reg_ofs.spad_write = ndev->reg_base + SNB_SPAD_OFFSET;
799                 /* Offset the start of the spads to correspond to whether it is
800                  * primary or secondary
801                  */
802                 ndev->reg_ofs.spad_read = ndev->reg_base + SNB_SPAD_OFFSET +
803                                           ndev->limits.max_spads * 4;
804                 ndev->reg_ofs.bar2_xlat = ndev->reg_base + SNB_PBAR2XLAT_OFFSET;
805                 ndev->reg_ofs.bar4_xlat = ndev->reg_base + SNB_PBAR4XLAT_OFFSET;
806
807                 ndev->limits.max_mw = SNB_MAX_MW;
808                 break;
809         default:
810                 /* Most likely caused by the remote NTB-RP device not being
811                  * configured
812                  */
813                 dev_err(&ndev->pdev->dev, "Unknown PPD %x\n", val);
814                 return -EINVAL;
815         }
816
817         ndev->reg_ofs.lnk_cntl = ndev->reg_base + SNB_NTBCNTL_OFFSET;
818         ndev->reg_ofs.lnk_stat = ndev->reg_base + SNB_SLINK_STATUS_OFFSET;
819         ndev->reg_ofs.spci_cmd = ndev->reg_base + SNB_PCICMD_OFFSET;
820
821         ndev->limits.max_db_bits = SNB_MAX_DB_BITS;
822         ndev->limits.msix_cnt = SNB_MSIX_CNT;
823         ndev->bits_per_vector = SNB_DB_BITS_PER_VEC;
824
825         return 0;
826 }
827
828 static int ntb_bwd_setup(struct ntb_device *ndev)
829 {
830         int rc;
831         u32 val;
832
833         ndev->hw_type = BWD_HW;
834
835         rc = pci_read_config_dword(ndev->pdev, NTB_PPD_OFFSET, &val);
836         if (rc)
837                 return rc;
838
839         switch ((val & BWD_PPD_CONN_TYPE) >> 8) {
840         case NTB_CONN_B2B:
841                 ndev->conn_type = NTB_CONN_B2B;
842                 break;
843         case NTB_CONN_RP:
844         default:
845                 dev_err(&ndev->pdev->dev, "Only B2B supported at this time\n");
846                 return -EINVAL;
847         }
848
849         if (val & BWD_PPD_DEV_TYPE)
850                 ndev->dev_type = NTB_DEV_DSD;
851         else
852                 ndev->dev_type = NTB_DEV_USD;
853
854         /* Initiate PCI-E link training */
855         rc = pci_write_config_dword(ndev->pdev, NTB_PPD_OFFSET,
856                                     val | BWD_PPD_INIT_LINK);
857         if (rc)
858                 return rc;
859
860         ndev->reg_ofs.ldb = ndev->reg_base + BWD_PDOORBELL_OFFSET;
861         ndev->reg_ofs.ldb_mask = ndev->reg_base + BWD_PDBMSK_OFFSET;
862         ndev->reg_ofs.bar2_xlat = ndev->reg_base + BWD_SBAR2XLAT_OFFSET;
863         ndev->reg_ofs.bar4_xlat = ndev->reg_base + BWD_SBAR4XLAT_OFFSET;
864         ndev->reg_ofs.lnk_cntl = ndev->reg_base + BWD_NTBCNTL_OFFSET;
865         ndev->reg_ofs.lnk_stat = ndev->reg_base + BWD_LINK_STATUS_OFFSET;
866         ndev->reg_ofs.spad_read = ndev->reg_base + BWD_SPAD_OFFSET;
867         ndev->reg_ofs.spci_cmd = ndev->reg_base + BWD_PCICMD_OFFSET;
868
869         if (ndev->conn_type == NTB_CONN_B2B) {
870                 ndev->reg_ofs.rdb = ndev->reg_base + BWD_B2B_DOORBELL_OFFSET;
871                 ndev->reg_ofs.spad_write = ndev->reg_base + BWD_B2B_SPAD_OFFSET;
872                 ndev->limits.max_spads = BWD_MAX_SPADS;
873         } else {
874                 ndev->reg_ofs.rdb = ndev->reg_base + BWD_PDOORBELL_OFFSET;
875                 ndev->reg_ofs.spad_write = ndev->reg_base + BWD_SPAD_OFFSET;
876                 ndev->limits.max_spads = BWD_MAX_COMPAT_SPADS;
877         }
878
879         ndev->limits.max_mw = BWD_MAX_MW;
880         ndev->limits.max_db_bits = BWD_MAX_DB_BITS;
881         ndev->limits.msix_cnt = BWD_MSIX_CNT;
882         ndev->bits_per_vector = BWD_DB_BITS_PER_VEC;
883
884         /* Since bwd doesn't have a link interrupt, setup a poll timer */
885         INIT_DELAYED_WORK(&ndev->hb_timer, bwd_link_poll);
886         INIT_DELAYED_WORK(&ndev->lr_timer, bwd_link_recovery);
887         schedule_delayed_work(&ndev->hb_timer, NTB_HB_TIMEOUT);
888
889         return 0;
890 }
891
892 static int ntb_device_setup(struct ntb_device *ndev)
893 {
894         int rc;
895
896         switch (ndev->pdev->device) {
897         case PCI_DEVICE_ID_INTEL_NTB_SS_JSF:
898         case PCI_DEVICE_ID_INTEL_NTB_SS_SNB:
899         case PCI_DEVICE_ID_INTEL_NTB_SS_IVT:
900         case PCI_DEVICE_ID_INTEL_NTB_SS_HSX:
901         case PCI_DEVICE_ID_INTEL_NTB_PS_JSF:
902         case PCI_DEVICE_ID_INTEL_NTB_PS_SNB:
903         case PCI_DEVICE_ID_INTEL_NTB_PS_IVT:
904         case PCI_DEVICE_ID_INTEL_NTB_PS_HSX:
905         case PCI_DEVICE_ID_INTEL_NTB_B2B_JSF:
906         case PCI_DEVICE_ID_INTEL_NTB_B2B_SNB:
907         case PCI_DEVICE_ID_INTEL_NTB_B2B_IVT:
908         case PCI_DEVICE_ID_INTEL_NTB_B2B_HSX:
909                 rc = ntb_xeon_setup(ndev);
910                 break;
911         case PCI_DEVICE_ID_INTEL_NTB_B2B_BWD:
912                 rc = ntb_bwd_setup(ndev);
913                 break;
914         default:
915                 rc = -ENODEV;
916         }
917
918         if (rc)
919                 return rc;
920
921         dev_info(&ndev->pdev->dev, "Device Type = %s\n",
922                  ndev->dev_type == NTB_DEV_USD ? "USD/DSP" : "DSD/USP");
923
924         if (ndev->conn_type == NTB_CONN_B2B)
925                 /* Enable Bus Master and Memory Space on the secondary side */
926                 writew(PCI_COMMAND_MEMORY | PCI_COMMAND_MASTER,
927                        ndev->reg_ofs.spci_cmd);
928
929         return 0;
930 }
931
932 static void ntb_device_free(struct ntb_device *ndev)
933 {
934         if (ndev->hw_type == BWD_HW) {
935                 cancel_delayed_work_sync(&ndev->hb_timer);
936                 cancel_delayed_work_sync(&ndev->lr_timer);
937         }
938 }
939
940 static irqreturn_t bwd_callback_msix_irq(int irq, void *data)
941 {
942         struct ntb_db_cb *db_cb = data;
943         struct ntb_device *ndev = db_cb->ndev;
944
945         dev_dbg(&ndev->pdev->dev, "MSI-X irq %d received for DB %d\n", irq,
946                 db_cb->db_num);
947
948         if (db_cb->callback)
949                 db_cb->callback(db_cb->data, db_cb->db_num);
950
951         /* No need to check for the specific HB irq, any interrupt means
952          * we're connected.
953          */
954         ndev->last_ts = jiffies;
955
956         writeq((u64) 1 << db_cb->db_num, ndev->reg_ofs.ldb);
957
958         return IRQ_HANDLED;
959 }
960
961 static irqreturn_t xeon_callback_msix_irq(int irq, void *data)
962 {
963         struct ntb_db_cb *db_cb = data;
964         struct ntb_device *ndev = db_cb->ndev;
965
966         dev_dbg(&ndev->pdev->dev, "MSI-X irq %d received for DB %d\n", irq,
967                 db_cb->db_num);
968
969         if (db_cb->callback)
970                 db_cb->callback(db_cb->data, db_cb->db_num);
971
972         /* On Sandybridge, there are 16 bits in the interrupt register
973          * but only 4 vectors.  So, 5 bits are assigned to the first 3
974          * vectors, with the 4th having a single bit for link
975          * interrupts.
976          */
977         writew(((1 << ndev->bits_per_vector) - 1) <<
978                (db_cb->db_num * ndev->bits_per_vector), ndev->reg_ofs.ldb);
979
980         return IRQ_HANDLED;
981 }
982
983 /* Since we do not have a HW doorbell in BWD, this is only used in JF/JT */
984 static irqreturn_t xeon_event_msix_irq(int irq, void *dev)
985 {
986         struct ntb_device *ndev = dev;
987         int rc;
988
989         dev_dbg(&ndev->pdev->dev, "MSI-X irq %d received for Events\n", irq);
990
991         rc = ntb_link_status(ndev);
992         if (rc)
993                 dev_err(&ndev->pdev->dev, "Error determining link status\n");
994
995         /* bit 15 is always the link bit */
996         writew(1 << ndev->limits.max_db_bits, ndev->reg_ofs.ldb);
997
998         return IRQ_HANDLED;
999 }
1000
1001 static irqreturn_t ntb_interrupt(int irq, void *dev)
1002 {
1003         struct ntb_device *ndev = dev;
1004         unsigned int i = 0;
1005
1006         if (ndev->hw_type == BWD_HW) {
1007                 u64 ldb = readq(ndev->reg_ofs.ldb);
1008
1009                 dev_dbg(&ndev->pdev->dev, "irq %d - ldb = %Lx\n", irq, ldb);
1010
1011                 while (ldb) {
1012                         i = __ffs(ldb);
1013                         ldb &= ldb - 1;
1014                         bwd_callback_msix_irq(irq, &ndev->db_cb[i]);
1015                 }
1016         } else {
1017                 u16 ldb = readw(ndev->reg_ofs.ldb);
1018
1019                 dev_dbg(&ndev->pdev->dev, "irq %d - ldb = %x\n", irq, ldb);
1020
1021                 if (ldb & SNB_DB_HW_LINK) {
1022                         xeon_event_msix_irq(irq, dev);
1023                         ldb &= ~SNB_DB_HW_LINK;
1024                 }
1025
1026                 while (ldb) {
1027                         i = __ffs(ldb);
1028                         ldb &= ldb - 1;
1029                         xeon_callback_msix_irq(irq, &ndev->db_cb[i]);
1030                 }
1031         }
1032
1033         return IRQ_HANDLED;
1034 }
1035
1036 static int ntb_setup_msix(struct ntb_device *ndev)
1037 {
1038         struct pci_dev *pdev = ndev->pdev;
1039         struct msix_entry *msix;
1040         int msix_entries;
1041         int rc, i, pos;
1042         u16 val;
1043
1044         pos = pci_find_capability(pdev, PCI_CAP_ID_MSIX);
1045         if (!pos) {
1046                 rc = -EIO;
1047                 goto err;
1048         }
1049
1050         rc = pci_read_config_word(pdev, pos + PCI_MSIX_FLAGS, &val);
1051         if (rc)
1052                 goto err;
1053
1054         msix_entries = msix_table_size(val);
1055         if (msix_entries > ndev->limits.msix_cnt) {
1056                 rc = -EINVAL;
1057                 goto err;
1058         }
1059
1060         ndev->msix_entries = kmalloc(sizeof(struct msix_entry) * msix_entries,
1061                                      GFP_KERNEL);
1062         if (!ndev->msix_entries) {
1063                 rc = -ENOMEM;
1064                 goto err;
1065         }
1066
1067         for (i = 0; i < msix_entries; i++)
1068                 ndev->msix_entries[i].entry = i;
1069
1070         rc = pci_enable_msix(pdev, ndev->msix_entries, msix_entries);
1071         if (rc < 0)
1072                 goto err1;
1073         if (rc > 0) {
1074                 /* On SNB, the link interrupt is always tied to 4th vector.  If
1075                  * we can't get all 4, then we can't use MSI-X.
1076                  */
1077                 if (ndev->hw_type != BWD_HW) {
1078                         rc = -EIO;
1079                         goto err1;
1080                 }
1081
1082                 dev_warn(&pdev->dev,
1083                          "Only %d MSI-X vectors.  Limiting the number of queues to that number.\n",
1084                          rc);
1085                 msix_entries = rc;
1086         }
1087
1088         for (i = 0; i < msix_entries; i++) {
1089                 msix = &ndev->msix_entries[i];
1090                 WARN_ON(!msix->vector);
1091
1092                 /* Use the last MSI-X vector for Link status */
1093                 if (ndev->hw_type == BWD_HW) {
1094                         rc = request_irq(msix->vector, bwd_callback_msix_irq, 0,
1095                                          "ntb-callback-msix", &ndev->db_cb[i]);
1096                         if (rc)
1097                                 goto err2;
1098                 } else {
1099                         if (i == msix_entries - 1) {
1100                                 rc = request_irq(msix->vector,
1101                                                  xeon_event_msix_irq, 0,
1102                                                  "ntb-event-msix", ndev);
1103                                 if (rc)
1104                                         goto err2;
1105                         } else {
1106                                 rc = request_irq(msix->vector,
1107                                                  xeon_callback_msix_irq, 0,
1108                                                  "ntb-callback-msix",
1109                                                  &ndev->db_cb[i]);
1110                                 if (rc)
1111                                         goto err2;
1112                         }
1113                 }
1114         }
1115
1116         ndev->num_msix = msix_entries;
1117         if (ndev->hw_type == BWD_HW)
1118                 ndev->max_cbs = msix_entries;
1119         else
1120                 ndev->max_cbs = msix_entries - 1;
1121
1122         return 0;
1123
1124 err2:
1125         while (--i >= 0) {
1126                 msix = &ndev->msix_entries[i];
1127                 if (ndev->hw_type != BWD_HW && i == ndev->num_msix - 1)
1128                         free_irq(msix->vector, ndev);
1129                 else
1130                         free_irq(msix->vector, &ndev->db_cb[i]);
1131         }
1132         pci_disable_msix(pdev);
1133 err1:
1134         kfree(ndev->msix_entries);
1135         dev_err(&pdev->dev, "Error allocating MSI-X interrupt\n");
1136 err:
1137         ndev->num_msix = 0;
1138         return rc;
1139 }
1140
1141 static int ntb_setup_msi(struct ntb_device *ndev)
1142 {
1143         struct pci_dev *pdev = ndev->pdev;
1144         int rc;
1145
1146         rc = pci_enable_msi(pdev);
1147         if (rc)
1148                 return rc;
1149
1150         rc = request_irq(pdev->irq, ntb_interrupt, 0, "ntb-msi", ndev);
1151         if (rc) {
1152                 pci_disable_msi(pdev);
1153                 dev_err(&pdev->dev, "Error allocating MSI interrupt\n");
1154                 return rc;
1155         }
1156
1157         return 0;
1158 }
1159
1160 static int ntb_setup_intx(struct ntb_device *ndev)
1161 {
1162         struct pci_dev *pdev = ndev->pdev;
1163         int rc;
1164
1165         pci_msi_off(pdev);
1166
1167         /* Verify intx is enabled */
1168         pci_intx(pdev, 1);
1169
1170         rc = request_irq(pdev->irq, ntb_interrupt, IRQF_SHARED, "ntb-intx",
1171                          ndev);
1172         if (rc)
1173                 return rc;
1174
1175         return 0;
1176 }
1177
1178 static int ntb_setup_interrupts(struct ntb_device *ndev)
1179 {
1180         int rc;
1181
1182         /* On BWD, disable all interrupts.  On SNB, disable all but Link
1183          * Interrupt.  The rest will be unmasked as callbacks are registered.
1184          */
1185         if (ndev->hw_type == BWD_HW)
1186                 writeq(~0, ndev->reg_ofs.ldb_mask);
1187         else
1188                 writew(~(1 << ndev->limits.max_db_bits),
1189                        ndev->reg_ofs.ldb_mask);
1190
1191         rc = ntb_setup_msix(ndev);
1192         if (!rc)
1193                 goto done;
1194
1195         ndev->bits_per_vector = 1;
1196         ndev->max_cbs = ndev->limits.max_db_bits;
1197
1198         rc = ntb_setup_msi(ndev);
1199         if (!rc)
1200                 goto done;
1201
1202         rc = ntb_setup_intx(ndev);
1203         if (rc) {
1204                 dev_err(&ndev->pdev->dev, "no usable interrupts\n");
1205                 return rc;
1206         }
1207
1208 done:
1209         return 0;
1210 }
1211
1212 static void ntb_free_interrupts(struct ntb_device *ndev)
1213 {
1214         struct pci_dev *pdev = ndev->pdev;
1215
1216         /* mask interrupts */
1217         if (ndev->hw_type == BWD_HW)
1218                 writeq(~0, ndev->reg_ofs.ldb_mask);
1219         else
1220                 writew(~0, ndev->reg_ofs.ldb_mask);
1221
1222         if (ndev->num_msix) {
1223                 struct msix_entry *msix;
1224                 u32 i;
1225
1226                 for (i = 0; i < ndev->num_msix; i++) {
1227                         msix = &ndev->msix_entries[i];
1228                         if (ndev->hw_type != BWD_HW && i == ndev->num_msix - 1)
1229                                 free_irq(msix->vector, ndev);
1230                         else
1231                                 free_irq(msix->vector, &ndev->db_cb[i]);
1232                 }
1233                 pci_disable_msix(pdev);
1234         } else {
1235                 free_irq(pdev->irq, ndev);
1236
1237                 if (pci_dev_msi_enabled(pdev))
1238                         pci_disable_msi(pdev);
1239         }
1240 }
1241
1242 static int ntb_create_callbacks(struct ntb_device *ndev)
1243 {
1244         int i;
1245
1246         /* Checken-egg issue.  We won't know how many callbacks are necessary
1247          * until we see how many MSI-X vectors we get, but these pointers need
1248          * to be passed into the MSI-X register fucntion.  So, we allocate the
1249          * max, knowing that they might not all be used, to work around this.
1250          */
1251         ndev->db_cb = kcalloc(ndev->limits.max_db_bits,
1252                               sizeof(struct ntb_db_cb),
1253                               GFP_KERNEL);
1254         if (!ndev->db_cb)
1255                 return -ENOMEM;
1256
1257         for (i = 0; i < ndev->limits.max_db_bits; i++) {
1258                 ndev->db_cb[i].db_num = i;
1259                 ndev->db_cb[i].ndev = ndev;
1260         }
1261
1262         return 0;
1263 }
1264
1265 static void ntb_free_callbacks(struct ntb_device *ndev)
1266 {
1267         int i;
1268
1269         for (i = 0; i < ndev->limits.max_db_bits; i++)
1270                 ntb_unregister_db_callback(ndev, i);
1271
1272         kfree(ndev->db_cb);
1273 }
1274
1275 static void ntb_setup_debugfs(struct ntb_device *ndev)
1276 {
1277         if (!debugfs_initialized())
1278                 return;
1279
1280         if (!debugfs_dir)
1281                 debugfs_dir = debugfs_create_dir(KBUILD_MODNAME, NULL);
1282
1283         ndev->debugfs_dir = debugfs_create_dir(pci_name(ndev->pdev),
1284                                                debugfs_dir);
1285 }
1286
1287 static void ntb_free_debugfs(struct ntb_device *ndev)
1288 {
1289         debugfs_remove_recursive(ndev->debugfs_dir);
1290
1291         if (debugfs_dir && simple_empty(debugfs_dir)) {
1292                 debugfs_remove_recursive(debugfs_dir);
1293                 debugfs_dir = NULL;
1294         }
1295 }
1296
1297 static int ntb_pci_probe(struct pci_dev *pdev, const struct pci_device_id *id)
1298 {
1299         struct ntb_device *ndev;
1300         int rc, i;
1301
1302         ndev = kzalloc(sizeof(struct ntb_device), GFP_KERNEL);
1303         if (!ndev)
1304                 return -ENOMEM;
1305
1306         ndev->pdev = pdev;
1307         ndev->link_status = NTB_LINK_DOWN;
1308         pci_set_drvdata(pdev, ndev);
1309         ntb_setup_debugfs(ndev);
1310
1311         rc = pci_enable_device(pdev);
1312         if (rc)
1313                 goto err;
1314
1315         pci_set_master(ndev->pdev);
1316
1317         rc = pci_request_selected_regions(pdev, NTB_BAR_MASK, KBUILD_MODNAME);
1318         if (rc)
1319                 goto err1;
1320
1321         ndev->reg_base = pci_ioremap_bar(pdev, NTB_BAR_MMIO);
1322         if (!ndev->reg_base) {
1323                 dev_warn(&pdev->dev, "Cannot remap BAR 0\n");
1324                 rc = -EIO;
1325                 goto err2;
1326         }
1327
1328         for (i = 0; i < NTB_MAX_NUM_MW; i++) {
1329                 ndev->mw[i].bar_sz = pci_resource_len(pdev, MW_TO_BAR(i));
1330                 ndev->mw[i].vbase =
1331                     ioremap_wc(pci_resource_start(pdev, MW_TO_BAR(i)),
1332                                ndev->mw[i].bar_sz);
1333                 dev_info(&pdev->dev, "MW %d size %llu\n", i,
1334                          (unsigned long long) ndev->mw[i].bar_sz);
1335                 if (!ndev->mw[i].vbase) {
1336                         dev_warn(&pdev->dev, "Cannot remap BAR %d\n",
1337                                  MW_TO_BAR(i));
1338                         rc = -EIO;
1339                         goto err3;
1340                 }
1341         }
1342
1343         rc = pci_set_dma_mask(pdev, DMA_BIT_MASK(64));
1344         if (rc) {
1345                 rc = pci_set_dma_mask(pdev, DMA_BIT_MASK(32));
1346                 if (rc)
1347                         goto err3;
1348
1349                 dev_warn(&pdev->dev, "Cannot DMA highmem\n");
1350         }
1351
1352         rc = pci_set_consistent_dma_mask(pdev, DMA_BIT_MASK(64));
1353         if (rc) {
1354                 rc = pci_set_consistent_dma_mask(pdev, DMA_BIT_MASK(32));
1355                 if (rc)
1356                         goto err3;
1357
1358                 dev_warn(&pdev->dev, "Cannot DMA consistent highmem\n");
1359         }
1360
1361         rc = ntb_device_setup(ndev);
1362         if (rc)
1363                 goto err3;
1364
1365         rc = ntb_create_callbacks(ndev);
1366         if (rc)
1367                 goto err4;
1368
1369         rc = ntb_setup_interrupts(ndev);
1370         if (rc)
1371                 goto err5;
1372
1373         /* The scratchpad registers keep the values between rmmod/insmod,
1374          * blast them now
1375          */
1376         for (i = 0; i < ndev->limits.max_spads; i++) {
1377                 ntb_write_local_spad(ndev, i, 0);
1378                 ntb_write_remote_spad(ndev, i, 0);
1379         }
1380
1381         rc = ntb_transport_init(pdev);
1382         if (rc)
1383                 goto err6;
1384
1385         /* Let's bring the NTB link up */
1386         writel(NTB_CNTL_BAR23_SNOOP | NTB_CNTL_BAR45_SNOOP,
1387                ndev->reg_ofs.lnk_cntl);
1388
1389         return 0;
1390
1391 err6:
1392         ntb_free_interrupts(ndev);
1393 err5:
1394         ntb_free_callbacks(ndev);
1395 err4:
1396         ntb_device_free(ndev);
1397 err3:
1398         for (i--; i >= 0; i--)
1399                 iounmap(ndev->mw[i].vbase);
1400         iounmap(ndev->reg_base);
1401 err2:
1402         pci_release_selected_regions(pdev, NTB_BAR_MASK);
1403 err1:
1404         pci_disable_device(pdev);
1405 err:
1406         ntb_free_debugfs(ndev);
1407         kfree(ndev);
1408
1409         dev_err(&pdev->dev, "Error loading %s module\n", KBUILD_MODNAME);
1410         return rc;
1411 }
1412
1413 static void ntb_pci_remove(struct pci_dev *pdev)
1414 {
1415         struct ntb_device *ndev = pci_get_drvdata(pdev);
1416         int i;
1417         u32 ntb_cntl;
1418
1419         /* Bring NTB link down */
1420         ntb_cntl = readl(ndev->reg_ofs.lnk_cntl);
1421         ntb_cntl |= NTB_CNTL_LINK_DISABLE;
1422         writel(ntb_cntl, ndev->reg_ofs.lnk_cntl);
1423
1424         ntb_transport_free(ndev->ntb_transport);
1425
1426         ntb_free_interrupts(ndev);
1427         ntb_free_callbacks(ndev);
1428         ntb_device_free(ndev);
1429
1430         for (i = 0; i < NTB_MAX_NUM_MW; i++)
1431                 iounmap(ndev->mw[i].vbase);
1432
1433         iounmap(ndev->reg_base);
1434         pci_release_selected_regions(pdev, NTB_BAR_MASK);
1435         pci_disable_device(pdev);
1436         ntb_free_debugfs(ndev);
1437         kfree(ndev);
1438 }
1439
1440 static struct pci_driver ntb_pci_driver = {
1441         .name = KBUILD_MODNAME,
1442         .id_table = ntb_pci_tbl,
1443         .probe = ntb_pci_probe,
1444         .remove = ntb_pci_remove,
1445 };
1446 module_pci_driver(ntb_pci_driver);