xref: /openbmc/linux/drivers/infiniband/hw/qib/qib.h (revision 53a2a90d)
1 #ifndef _QIB_KERNEL_H
2 #define _QIB_KERNEL_H
3 /*
4  * Copyright (c) 2012 - 2017 Intel Corporation.  All rights reserved.
5  * Copyright (c) 2006 - 2012 QLogic Corporation. All rights reserved.
6  * Copyright (c) 2003, 2004, 2005, 2006 PathScale, Inc. All rights reserved.
7  *
8  * This software is available to you under a choice of one of two
9  * licenses.  You may choose to be licensed under the terms of the GNU
10  * General Public License (GPL) Version 2, available from the file
11  * COPYING in the main directory of this source tree, or the
12  * OpenIB.org BSD license below:
13  *
14  *     Redistribution and use in source and binary forms, with or
15  *     without modification, are permitted provided that the following
16  *     conditions are met:
17  *
18  *      - Redistributions of source code must retain the above
19  *        copyright notice, this list of conditions and the following
20  *        disclaimer.
21  *
22  *      - Redistributions in binary form must reproduce the above
23  *        copyright notice, this list of conditions and the following
24  *        disclaimer in the documentation and/or other materials
25  *        provided with the distribution.
26  *
27  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
28  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
29  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
30  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
31  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
32  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
33  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
34  * SOFTWARE.
35  */
36 
37 /*
38  * This header file is the base header file for qlogic_ib kernel code
39  * qib_user.h serves a similar purpose for user code.
40  */
41 
42 #include <linux/interrupt.h>
43 #include <linux/pci.h>
44 #include <linux/dma-mapping.h>
45 #include <linux/mutex.h>
46 #include <linux/list.h>
47 #include <linux/scatterlist.h>
48 #include <linux/slab.h>
49 #include <linux/io.h>
50 #include <linux/fs.h>
51 #include <linux/completion.h>
52 #include <linux/kref.h>
53 #include <linux/sched.h>
54 #include <linux/kthread.h>
55 #include <linux/xarray.h>
56 #include <rdma/ib_hdrs.h>
57 #include <rdma/rdma_vt.h>
58 
59 #include "qib_common.h"
60 #include "qib_verbs.h"
61 
62 /* only s/w major version of QLogic_IB we can handle */
63 #define QIB_CHIP_VERS_MAJ 2U
64 
65 /* don't care about this except printing */
66 #define QIB_CHIP_VERS_MIN 0U
67 
68 /* The Organization Unique Identifier (Mfg code), and its position in GUID */
69 #define QIB_OUI 0x001175
70 #define QIB_OUI_LSB 40
71 
72 /*
73  * per driver stats, either not device nor port-specific, or
74  * summed over all of the devices and ports.
75  * They are described by name via ipathfs filesystem, so layout
76  * and number of elements can change without breaking compatibility.
77  * If members are added or deleted qib_statnames[] in qib_fs.c must
78  * change to match.
79  */
80 struct qlogic_ib_stats {
81 	__u64 sps_ints; /* number of interrupts handled */
82 	__u64 sps_errints; /* number of error interrupts */
83 	__u64 sps_txerrs; /* tx-related packet errors */
84 	__u64 sps_rcverrs; /* non-crc rcv packet errors */
85 	__u64 sps_hwerrs; /* hardware errors reported (parity, etc.) */
86 	__u64 sps_nopiobufs; /* no pio bufs avail from kernel */
87 	__u64 sps_ctxts; /* number of contexts currently open */
88 	__u64 sps_lenerrs; /* number of kernel packets where RHF != LRH len */
89 	__u64 sps_buffull;
90 	__u64 sps_hdrfull;
91 };
92 
93 extern struct qlogic_ib_stats qib_stats;
94 extern const struct pci_error_handlers qib_pci_err_handler;
95 
96 #define QIB_CHIP_SWVERSION QIB_CHIP_VERS_MAJ
97 /*
98  * First-cut critierion for "device is active" is
99  * two thousand dwords combined Tx, Rx traffic per
100  * 5-second interval. SMA packets are 64 dwords,
101  * and occur "a few per second", presumably each way.
102  */
103 #define QIB_TRAFFIC_ACTIVE_THRESHOLD (2000)
104 
105 /*
106  * Below contains all data related to a single context (formerly called port).
107  */
108 
109 #ifdef CONFIG_DEBUG_FS
110 struct qib_opcode_stats_perctx;
111 #endif
112 
113 struct qib_ctxtdata {
114 	void **rcvegrbuf;
115 	dma_addr_t *rcvegrbuf_phys;
116 	/* rcvhdrq base, needs mmap before useful */
117 	void *rcvhdrq;
118 	/* kernel virtual address where hdrqtail is updated */
119 	void *rcvhdrtail_kvaddr;
120 	/*
121 	 * temp buffer for expected send setup, allocated at open, instead
122 	 * of each setup call
123 	 */
124 	void *tid_pg_list;
125 	/*
126 	 * Shared page for kernel to signal user processes that send buffers
127 	 * need disarming.  The process should call QIB_CMD_DISARM_BUFS
128 	 * or QIB_CMD_ACK_EVENT with IPATH_EVENT_DISARM_BUFS set.
129 	 */
130 	unsigned long *user_event_mask;
131 	/* when waiting for rcv or pioavail */
132 	wait_queue_head_t wait;
133 	/*
134 	 * rcvegr bufs base, physical, must fit
135 	 * in 44 bits so 32 bit programs mmap64 44 bit works)
136 	 */
137 	dma_addr_t rcvegr_phys;
138 	/* mmap of hdrq, must fit in 44 bits */
139 	dma_addr_t rcvhdrq_phys;
140 	dma_addr_t rcvhdrqtailaddr_phys;
141 
142 	/*
143 	 * number of opens (including slave sub-contexts) on this instance
144 	 * (ignoring forks, dup, etc. for now)
145 	 */
146 	int cnt;
147 	/*
148 	 * how much space to leave at start of eager TID entries for
149 	 * protocol use, on each TID
150 	 */
151 	/* instead of calculating it */
152 	unsigned ctxt;
153 	/* local node of context */
154 	int node_id;
155 	/* non-zero if ctxt is being shared. */
156 	u16 subctxt_cnt;
157 	/* non-zero if ctxt is being shared. */
158 	u16 subctxt_id;
159 	/* number of eager TID entries. */
160 	u16 rcvegrcnt;
161 	/* index of first eager TID entry. */
162 	u16 rcvegr_tid_base;
163 	/* number of pio bufs for this ctxt (all procs, if shared) */
164 	u32 piocnt;
165 	/* first pio buffer for this ctxt */
166 	u32 pio_base;
167 	/* chip offset of PIO buffers for this ctxt */
168 	u32 piobufs;
169 	/* how many alloc_pages() chunks in rcvegrbuf_pages */
170 	u32 rcvegrbuf_chunks;
171 	/* how many egrbufs per chunk */
172 	u16 rcvegrbufs_perchunk;
173 	/* ilog2 of above */
174 	u16 rcvegrbufs_perchunk_shift;
175 	/* order for rcvegrbuf_pages */
176 	size_t rcvegrbuf_size;
177 	/* rcvhdrq size (for freeing) */
178 	size_t rcvhdrq_size;
179 	/* per-context flags for fileops/intr communication */
180 	unsigned long flag;
181 	/* next expected TID to check when looking for free */
182 	u32 tidcursor;
183 	/* WAIT_RCV that timed out, no interrupt */
184 	u32 rcvwait_to;
185 	/* WAIT_PIO that timed out, no interrupt */
186 	u32 piowait_to;
187 	/* WAIT_RCV already happened, no wait */
188 	u32 rcvnowait;
189 	/* WAIT_PIO already happened, no wait */
190 	u32 pionowait;
191 	/* total number of polled urgent packets */
192 	u32 urgent;
193 	/* saved total number of polled urgent packets for poll edge trigger */
194 	u32 urgent_poll;
195 	/* pid of process using this ctxt */
196 	pid_t pid;
197 	pid_t subpid[QLOGIC_IB_MAX_SUBCTXT];
198 	/* same size as task_struct .comm[], command that opened context */
199 	char comm[TASK_COMM_LEN];
200 	/* pkeys set by this use of this ctxt */
201 	u16 pkeys[4];
202 	/* so file ops can get at unit */
203 	struct qib_devdata *dd;
204 	/* so funcs that need physical port can get it easily */
205 	struct qib_pportdata *ppd;
206 	/* A page of memory for rcvhdrhead, rcvegrhead, rcvegrtail * N */
207 	void *subctxt_uregbase;
208 	/* An array of pages for the eager receive buffers * N */
209 	void *subctxt_rcvegrbuf;
210 	/* An array of pages for the eager header queue entries * N */
211 	void *subctxt_rcvhdr_base;
212 	/* The version of the library which opened this ctxt */
213 	u32 userversion;
214 	/* Bitmask of active slaves */
215 	u32 active_slaves;
216 	/* Type of packets or conditions we want to poll for */
217 	u16 poll_type;
218 	/* receive packet sequence counter */
219 	u8 seq_cnt;
220 	u8 redirect_seq_cnt;
221 	/* ctxt rcvhdrq head offset */
222 	u32 head;
223 	/* QPs waiting for context processing */
224 	struct list_head qp_wait_list;
225 #ifdef CONFIG_DEBUG_FS
226 	/* verbs stats per CTX */
227 	struct qib_opcode_stats_perctx *opstats;
228 #endif
229 };
230 
231 struct rvt_sge_state;
232 
233 struct qib_sdma_txreq {
234 	int                 flags;
235 	int                 sg_count;
236 	dma_addr_t          addr;
237 	void              (*callback)(struct qib_sdma_txreq *, int);
238 	u16                 start_idx;  /* sdma private */
239 	u16                 next_descq_idx;  /* sdma private */
240 	struct list_head    list;       /* sdma private */
241 };
242 
243 struct qib_sdma_desc {
244 	__le64 qw[2];
245 };
246 
247 struct qib_verbs_txreq {
248 	struct qib_sdma_txreq   txreq;
249 	struct rvt_qp           *qp;
250 	struct rvt_swqe         *wqe;
251 	u32                     dwords;
252 	u16                     hdr_dwords;
253 	u16                     hdr_inx;
254 	struct qib_pio_header	*align_buf;
255 	struct rvt_mregion	*mr;
256 	struct rvt_sge_state    *ss;
257 };
258 
259 #define QIB_SDMA_TXREQ_F_USELARGEBUF  0x1
260 #define QIB_SDMA_TXREQ_F_HEADTOHOST   0x2
261 #define QIB_SDMA_TXREQ_F_INTREQ       0x4
262 #define QIB_SDMA_TXREQ_F_FREEBUF      0x8
263 #define QIB_SDMA_TXREQ_F_FREEDESC     0x10
264 
265 #define QIB_SDMA_TXREQ_S_OK        0
266 #define QIB_SDMA_TXREQ_S_SENDERROR 1
267 #define QIB_SDMA_TXREQ_S_ABORTED   2
268 #define QIB_SDMA_TXREQ_S_SHUTDOWN  3
269 
270 /*
271  * Get/Set IB link-level config parameters for f_get/set_ib_cfg()
272  * Mostly for MADs that set or query link parameters, also ipath
273  * config interfaces
274  */
275 #define QIB_IB_CFG_LIDLMC 0 /* LID (LS16b) and Mask (MS16b) */
276 #define QIB_IB_CFG_LWID_ENB 2 /* allowed Link-width */
277 #define QIB_IB_CFG_LWID 3 /* currently active Link-width */
278 #define QIB_IB_CFG_SPD_ENB 4 /* allowed Link speeds */
279 #define QIB_IB_CFG_SPD 5 /* current Link spd */
280 #define QIB_IB_CFG_RXPOL_ENB 6 /* Auto-RX-polarity enable */
281 #define QIB_IB_CFG_LREV_ENB 7 /* Auto-Lane-reversal enable */
282 #define QIB_IB_CFG_LINKLATENCY 8 /* Link Latency (IB1.2 only) */
283 #define QIB_IB_CFG_HRTBT 9 /* IB heartbeat off/enable/auto; DDR/QDR only */
284 #define QIB_IB_CFG_OP_VLS 10 /* operational VLs */
285 #define QIB_IB_CFG_VL_HIGH_CAP 11 /* num of VL high priority weights */
286 #define QIB_IB_CFG_VL_LOW_CAP 12 /* num of VL low priority weights */
287 #define QIB_IB_CFG_OVERRUN_THRESH 13 /* IB overrun threshold */
288 #define QIB_IB_CFG_PHYERR_THRESH 14 /* IB PHY error threshold */
289 #define QIB_IB_CFG_LINKDEFAULT 15 /* IB link default (sleep/poll) */
290 #define QIB_IB_CFG_PKEYS 16 /* update partition keys */
291 #define QIB_IB_CFG_MTU 17 /* update MTU in IBC */
292 #define QIB_IB_CFG_LSTATE 18 /* update linkcmd and linkinitcmd in IBC */
293 #define QIB_IB_CFG_VL_HIGH_LIMIT 19
294 #define QIB_IB_CFG_PMA_TICKS 20 /* PMA sample tick resolution */
295 #define QIB_IB_CFG_PORT 21 /* switch port we are connected to */
296 
297 /*
298  * for CFG_LSTATE: LINKCMD in upper 16 bits, LINKINITCMD in lower 16
299  * IB_LINKINITCMD_POLL and SLEEP are also used as set/get values for
300  * QIB_IB_CFG_LINKDEFAULT cmd
301  */
302 #define   IB_LINKCMD_DOWN   (0 << 16)
303 #define   IB_LINKCMD_ARMED  (1 << 16)
304 #define   IB_LINKCMD_ACTIVE (2 << 16)
305 #define   IB_LINKINITCMD_NOP     0
306 #define   IB_LINKINITCMD_POLL    1
307 #define   IB_LINKINITCMD_SLEEP   2
308 #define   IB_LINKINITCMD_DISABLE 3
309 
310 /*
311  * valid states passed to qib_set_linkstate() user call
312  */
313 #define QIB_IB_LINKDOWN         0
314 #define QIB_IB_LINKARM          1
315 #define QIB_IB_LINKACTIVE       2
316 #define QIB_IB_LINKDOWN_ONLY    3
317 #define QIB_IB_LINKDOWN_SLEEP   4
318 #define QIB_IB_LINKDOWN_DISABLE 5
319 
320 /*
321  * These 7 values (SDR, DDR, and QDR may be ORed for auto-speed
322  * negotiation) are used for the 3rd argument to path_f_set_ib_cfg
323  * with cmd QIB_IB_CFG_SPD_ENB, by direct calls or via sysfs.  They
324  * are also the the possible values for qib_link_speed_enabled and active
325  * The values were chosen to match values used within the IB spec.
326  */
327 #define QIB_IB_SDR 1
328 #define QIB_IB_DDR 2
329 #define QIB_IB_QDR 4
330 
331 #define QIB_DEFAULT_MTU 4096
332 
333 /* max number of IB ports supported per HCA */
334 #define QIB_MAX_IB_PORTS 2
335 
336 /*
337  * Possible IB config parameters for f_get/set_ib_table()
338  */
339 #define QIB_IB_TBL_VL_HIGH_ARB 1 /* Get/set VL high priority weights */
340 #define QIB_IB_TBL_VL_LOW_ARB 2 /* Get/set VL low priority weights */
341 
342 /*
343  * Possible "operations" for f_rcvctrl(ppd, op, ctxt)
344  * these are bits so they can be combined, e.g.
345  * QIB_RCVCTRL_INTRAVAIL_ENB | QIB_RCVCTRL_CTXT_ENB
346  */
347 #define QIB_RCVCTRL_TAILUPD_ENB 0x01
348 #define QIB_RCVCTRL_TAILUPD_DIS 0x02
349 #define QIB_RCVCTRL_CTXT_ENB 0x04
350 #define QIB_RCVCTRL_CTXT_DIS 0x08
351 #define QIB_RCVCTRL_INTRAVAIL_ENB 0x10
352 #define QIB_RCVCTRL_INTRAVAIL_DIS 0x20
353 #define QIB_RCVCTRL_PKEY_ENB 0x40  /* Note, default is enabled */
354 #define QIB_RCVCTRL_PKEY_DIS 0x80
355 #define QIB_RCVCTRL_BP_ENB 0x0100
356 #define QIB_RCVCTRL_BP_DIS 0x0200
357 #define QIB_RCVCTRL_TIDFLOW_ENB 0x0400
358 #define QIB_RCVCTRL_TIDFLOW_DIS 0x0800
359 
360 /*
361  * Possible "operations" for f_sendctrl(ppd, op, var)
362  * these are bits so they can be combined, e.g.
363  * QIB_SENDCTRL_BUFAVAIL_ENB | QIB_SENDCTRL_ENB
364  * Some operations (e.g. DISARM, ABORT) are known to
365  * be "one-shot", so do not modify shadow.
366  */
367 #define QIB_SENDCTRL_DISARM       (0x1000)
368 #define QIB_SENDCTRL_DISARM_BUF(bufn) ((bufn) | QIB_SENDCTRL_DISARM)
369 	/* available (0x2000) */
370 #define QIB_SENDCTRL_AVAIL_DIS    (0x4000)
371 #define QIB_SENDCTRL_AVAIL_ENB    (0x8000)
372 #define QIB_SENDCTRL_AVAIL_BLIP  (0x10000)
373 #define QIB_SENDCTRL_SEND_DIS    (0x20000)
374 #define QIB_SENDCTRL_SEND_ENB    (0x40000)
375 #define QIB_SENDCTRL_FLUSH       (0x80000)
376 #define QIB_SENDCTRL_CLEAR      (0x100000)
377 #define QIB_SENDCTRL_DISARM_ALL (0x200000)
378 
379 /*
380  * These are the generic indices for requesting per-port
381  * counter values via the f_portcntr function.  They
382  * are always returned as 64 bit values, although most
383  * are 32 bit counters.
384  */
385 /* send-related counters */
386 #define QIBPORTCNTR_PKTSEND         0U
387 #define QIBPORTCNTR_WORDSEND        1U
388 #define QIBPORTCNTR_PSXMITDATA      2U
389 #define QIBPORTCNTR_PSXMITPKTS      3U
390 #define QIBPORTCNTR_PSXMITWAIT      4U
391 #define QIBPORTCNTR_SENDSTALL       5U
392 /* receive-related counters */
393 #define QIBPORTCNTR_PKTRCV          6U
394 #define QIBPORTCNTR_PSRCVDATA       7U
395 #define QIBPORTCNTR_PSRCVPKTS       8U
396 #define QIBPORTCNTR_RCVEBP          9U
397 #define QIBPORTCNTR_RCVOVFL         10U
398 #define QIBPORTCNTR_WORDRCV         11U
399 /* IB link related error counters */
400 #define QIBPORTCNTR_RXLOCALPHYERR   12U
401 #define QIBPORTCNTR_RXVLERR         13U
402 #define QIBPORTCNTR_ERRICRC         14U
403 #define QIBPORTCNTR_ERRVCRC         15U
404 #define QIBPORTCNTR_ERRLPCRC        16U
405 #define QIBPORTCNTR_BADFORMAT       17U
406 #define QIBPORTCNTR_ERR_RLEN        18U
407 #define QIBPORTCNTR_IBSYMBOLERR     19U
408 #define QIBPORTCNTR_INVALIDRLEN     20U
409 #define QIBPORTCNTR_UNSUPVL         21U
410 #define QIBPORTCNTR_EXCESSBUFOVFL   22U
411 #define QIBPORTCNTR_ERRLINK         23U
412 #define QIBPORTCNTR_IBLINKDOWN      24U
413 #define QIBPORTCNTR_IBLINKERRRECOV  25U
414 #define QIBPORTCNTR_LLI             26U
415 /* other error counters */
416 #define QIBPORTCNTR_RXDROPPKT       27U
417 #define QIBPORTCNTR_VL15PKTDROP     28U
418 #define QIBPORTCNTR_ERRPKEY         29U
419 #define QIBPORTCNTR_KHDROVFL        30U
420 /* sampling counters (these are actually control registers) */
421 #define QIBPORTCNTR_PSINTERVAL      31U
422 #define QIBPORTCNTR_PSSTART         32U
423 #define QIBPORTCNTR_PSSTAT          33U
424 
425 /* how often we check for packet activity for "power on hours (in seconds) */
426 #define ACTIVITY_TIMER 5
427 
428 #define MAX_NAME_SIZE 64
429 
430 #ifdef CONFIG_INFINIBAND_QIB_DCA
431 struct qib_irq_notify;
432 #endif
433 
434 struct qib_msix_entry {
435 	void *arg;
436 #ifdef CONFIG_INFINIBAND_QIB_DCA
437 	int dca;
438 	int rcv;
439 	struct qib_irq_notify *notifier;
440 #endif
441 	cpumask_var_t mask;
442 };
443 
444 /* Below is an opaque struct. Each chip (device) can maintain
445  * private data needed for its operation, but not germane to the
446  * rest of the driver.  For convenience, we define another that
447  * is chip-specific, per-port
448  */
449 struct qib_chip_specific;
450 struct qib_chipport_specific;
451 
452 enum qib_sdma_states {
453 	qib_sdma_state_s00_hw_down,
454 	qib_sdma_state_s10_hw_start_up_wait,
455 	qib_sdma_state_s20_idle,
456 	qib_sdma_state_s30_sw_clean_up_wait,
457 	qib_sdma_state_s40_hw_clean_up_wait,
458 	qib_sdma_state_s50_hw_halt_wait,
459 	qib_sdma_state_s99_running,
460 };
461 
462 enum qib_sdma_events {
463 	qib_sdma_event_e00_go_hw_down,
464 	qib_sdma_event_e10_go_hw_start,
465 	qib_sdma_event_e20_hw_started,
466 	qib_sdma_event_e30_go_running,
467 	qib_sdma_event_e40_sw_cleaned,
468 	qib_sdma_event_e50_hw_cleaned,
469 	qib_sdma_event_e60_hw_halted,
470 	qib_sdma_event_e70_go_idle,
471 	qib_sdma_event_e7220_err_halted,
472 	qib_sdma_event_e7322_err_halted,
473 	qib_sdma_event_e90_timer_tick,
474 };
475 
476 struct sdma_set_state_action {
477 	unsigned op_enable:1;
478 	unsigned op_intenable:1;
479 	unsigned op_halt:1;
480 	unsigned op_drain:1;
481 	unsigned go_s99_running_tofalse:1;
482 	unsigned go_s99_running_totrue:1;
483 };
484 
485 struct qib_sdma_state {
486 	struct kref          kref;
487 	struct completion    comp;
488 	enum qib_sdma_states current_state;
489 	struct sdma_set_state_action *set_state_action;
490 	unsigned             current_op;
491 	unsigned             go_s99_running;
492 	unsigned             first_sendbuf;
493 	unsigned             last_sendbuf; /* really last +1 */
494 	/* debugging/devel */
495 	enum qib_sdma_states previous_state;
496 	unsigned             previous_op;
497 	enum qib_sdma_events last_event;
498 };
499 
500 struct xmit_wait {
501 	struct timer_list timer;
502 	u64 counter;
503 	u8 flags;
504 	struct cache {
505 		u64 psxmitdata;
506 		u64 psrcvdata;
507 		u64 psxmitpkts;
508 		u64 psrcvpkts;
509 		u64 psxmitwait;
510 	} counter_cache;
511 };
512 
513 /*
514  * The structure below encapsulates data relevant to a physical IB Port.
515  * Current chips support only one such port, but the separation
516  * clarifies things a bit. Note that to conform to IB conventions,
517  * port-numbers are one-based. The first or only port is port1.
518  */
519 struct qib_pportdata {
520 	struct qib_ibport ibport_data;
521 
522 	struct qib_devdata *dd;
523 	struct qib_chippport_specific *cpspec; /* chip-specific per-port */
524 
525 	/* GUID for this interface, in network order */
526 	__be64 guid;
527 
528 	/* QIB_POLL, etc. link-state specific flags, per port */
529 	u32 lflags;
530 	/* qib_lflags driver is waiting for */
531 	u32 state_wanted;
532 	spinlock_t lflags_lock;
533 
534 	/* ref count for each pkey */
535 	atomic_t pkeyrefs[4];
536 
537 	/*
538 	 * this address is mapped readonly into user processes so they can
539 	 * get status cheaply, whenever they want.  One qword of status per port
540 	 */
541 	u64 *statusp;
542 
543 	/* SendDMA related entries */
544 
545 	/* read mostly */
546 	struct qib_sdma_desc *sdma_descq;
547 	struct workqueue_struct *qib_wq;
548 	struct qib_sdma_state sdma_state;
549 	dma_addr_t       sdma_descq_phys;
550 	volatile __le64 *sdma_head_dma; /* DMA'ed by chip */
551 	dma_addr_t       sdma_head_phys;
552 	u16                   sdma_descq_cnt;
553 
554 	/* read/write using lock */
555 	spinlock_t            sdma_lock ____cacheline_aligned_in_smp;
556 	struct list_head      sdma_activelist;
557 	struct list_head      sdma_userpending;
558 	u64                   sdma_descq_added;
559 	u64                   sdma_descq_removed;
560 	u16                   sdma_descq_tail;
561 	u16                   sdma_descq_head;
562 	u8                    sdma_generation;
563 	u8                    sdma_intrequest;
564 
565 	struct tasklet_struct sdma_sw_clean_up_task
566 		____cacheline_aligned_in_smp;
567 
568 	wait_queue_head_t state_wait; /* for state_wanted */
569 
570 	/* HoL blocking for SMP replies */
571 	unsigned          hol_state;
572 	struct timer_list hol_timer;
573 
574 	/*
575 	 * Shadow copies of registers; size indicates read access size.
576 	 * Most of them are readonly, but some are write-only register,
577 	 * where we manipulate the bits in the shadow copy, and then write
578 	 * the shadow copy to qlogic_ib.
579 	 *
580 	 * We deliberately make most of these 32 bits, since they have
581 	 * restricted range.  For any that we read, we won't to generate 32
582 	 * bit accesses, since Opteron will generate 2 separate 32 bit HT
583 	 * transactions for a 64 bit read, and we want to avoid unnecessary
584 	 * bus transactions.
585 	 */
586 
587 	/* This is the 64 bit group */
588 	/* last ibcstatus.  opaque outside chip-specific code */
589 	u64 lastibcstat;
590 
591 	/* these are the "32 bit" regs */
592 
593 	/*
594 	 * the following two are 32-bit bitmasks, but {test,clear,set}_bit
595 	 * all expect bit fields to be "unsigned long"
596 	 */
597 	unsigned long p_rcvctrl; /* shadow per-port rcvctrl */
598 	unsigned long p_sendctrl; /* shadow per-port sendctrl */
599 
600 	u32 ibmtu; /* The MTU programmed for this unit */
601 	/*
602 	 * Current max size IB packet (in bytes) including IB headers, that
603 	 * we can send. Changes when ibmtu changes.
604 	 */
605 	u32 ibmaxlen;
606 	/*
607 	 * ibmaxlen at init time, limited by chip and by receive buffer
608 	 * size.  Not changed after init.
609 	 */
610 	u32 init_ibmaxlen;
611 	/* LID programmed for this instance */
612 	u16 lid;
613 	/* list of pkeys programmed; 0 if not set */
614 	u16 pkeys[4];
615 	/* LID mask control */
616 	u8 lmc;
617 	u8 link_width_supported;
618 	u16 link_speed_supported;
619 	u8 link_width_enabled;
620 	u16 link_speed_enabled;
621 	u8 link_width_active;
622 	u16 link_speed_active;
623 	u8 vls_supported;
624 	u8 vls_operational;
625 	/* Rx Polarity inversion (compensate for ~tx on partner) */
626 	u8 rx_pol_inv;
627 
628 	u8 hw_pidx;     /* physical port index */
629 	u32 port;        /* IB port number and index into dd->pports - 1 */
630 
631 	u8 delay_mult;
632 
633 	/* used to override LED behavior */
634 	u8 led_override;  /* Substituted for normal value, if non-zero */
635 	u16 led_override_timeoff; /* delta to next timer event */
636 	u8 led_override_vals[2]; /* Alternates per blink-frame */
637 	u8 led_override_phase; /* Just counts, LSB picks from vals[] */
638 	atomic_t led_override_timer_active;
639 	/* Used to flash LEDs in override mode */
640 	struct timer_list led_override_timer;
641 	struct xmit_wait cong_stats;
642 	struct timer_list symerr_clear_timer;
643 
644 	/* Synchronize access between driver writes and sysfs reads */
645 	spinlock_t cc_shadow_lock
646 		____cacheline_aligned_in_smp;
647 
648 	/* Shadow copy of the congestion control table */
649 	struct cc_table_shadow *ccti_entries_shadow;
650 
651 	/* Shadow copy of the congestion control entries */
652 	struct ib_cc_congestion_setting_attr_shadow *congestion_entries_shadow;
653 
654 	/* List of congestion control table entries */
655 	struct ib_cc_table_entry_shadow *ccti_entries;
656 
657 	/* 16 congestion entries with each entry corresponding to a SL */
658 	struct ib_cc_congestion_entry_shadow *congestion_entries;
659 
660 	/* Maximum number of congestion control entries that the agent expects
661 	 * the manager to send.
662 	 */
663 	u16 cc_supported_table_entries;
664 
665 	/* Total number of congestion control table entries */
666 	u16 total_cct_entry;
667 
668 	/* Bit map identifying service level */
669 	u16 cc_sl_control_map;
670 
671 	/* maximum congestion control table index */
672 	u16 ccti_limit;
673 
674 	/* CA's max number of 64 entry units in the congestion control table */
675 	u8 cc_max_table_entries;
676 };
677 
678 /* Observers. Not to be taken lightly, possibly not to ship. */
679 /*
680  * If a diag read or write is to (bottom <= offset <= top),
681  * the "hook" is called, allowing, e.g. shadows to be
682  * updated in sync with the driver. struct diag_observer
683  * is the "visible" part.
684  */
685 struct diag_observer;
686 
687 typedef int (*diag_hook) (struct qib_devdata *dd,
688 	const struct diag_observer *op,
689 	u32 offs, u64 *data, u64 mask, int only_32);
690 
691 struct diag_observer {
692 	diag_hook hook;
693 	u32 bottom;
694 	u32 top;
695 };
696 
697 extern int qib_register_observer(struct qib_devdata *dd,
698 	const struct diag_observer *op);
699 
700 /* Only declared here, not defined. Private to diags */
701 struct diag_observer_list_elt;
702 
703 /* device data struct now contains only "general per-device" info.
704  * fields related to a physical IB port are in a qib_pportdata struct,
705  * described above) while fields only used by a particular chip-type are in
706  * a qib_chipdata struct, whose contents are opaque to this file.
707  */
708 struct qib_devdata {
709 	struct qib_ibdev verbs_dev;     /* must be first */
710 	struct list_head list;
711 	/* pointers to related structs for this device */
712 	/* pci access data structure */
713 	struct pci_dev *pcidev;
714 	struct cdev *user_cdev;
715 	struct cdev *diag_cdev;
716 	struct device *user_device;
717 	struct device *diag_device;
718 
719 	/* mem-mapped pointer to base of chip regs */
720 	u64 __iomem *kregbase;
721 	/* end of mem-mapped chip space excluding sendbuf and user regs */
722 	u64 __iomem *kregend;
723 	/* physical address of chip for io_remap, etc. */
724 	resource_size_t physaddr;
725 	/* qib_cfgctxts pointers */
726 	struct qib_ctxtdata **rcd; /* Receive Context Data */
727 
728 	/* qib_pportdata, points to array of (physical) port-specific
729 	 * data structs, indexed by pidx (0..n-1)
730 	 */
731 	struct qib_pportdata *pport;
732 	struct qib_chip_specific *cspec; /* chip-specific */
733 
734 	/* kvirt address of 1st 2k pio buffer */
735 	void __iomem *pio2kbase;
736 	/* kvirt address of 1st 4k pio buffer */
737 	void __iomem *pio4kbase;
738 	/* mem-mapped pointer to base of PIO buffers (if using WC PAT) */
739 	void __iomem *piobase;
740 	/* mem-mapped pointer to base of user chip regs (if using WC PAT) */
741 	u64 __iomem *userbase;
742 	void __iomem *piovl15base; /* base of VL15 buffers, if not WC */
743 	/*
744 	 * points to area where PIOavail registers will be DMA'ed.
745 	 * Has to be on a page of it's own, because the page will be
746 	 * mapped into user program space.  This copy is *ONLY* ever
747 	 * written by DMA, not by the driver!  Need a copy per device
748 	 * when we get to multiple devices
749 	 */
750 	volatile __le64 *pioavailregs_dma; /* DMA'ed by chip */
751 	/* physical address where updates occur */
752 	dma_addr_t pioavailregs_phys;
753 
754 	/* device-specific implementations of functions needed by
755 	 * common code. Contrary to previous consensus, we can't
756 	 * really just point to a device-specific table, because we
757 	 * may need to "bend", e.g. *_f_put_tid
758 	 */
759 	/* fallback to alternate interrupt type if possible */
760 	int (*f_intr_fallback)(struct qib_devdata *);
761 	/* hard reset chip */
762 	int (*f_reset)(struct qib_devdata *);
763 	void (*f_quiet_serdes)(struct qib_pportdata *);
764 	int (*f_bringup_serdes)(struct qib_pportdata *);
765 	int (*f_early_init)(struct qib_devdata *);
766 	void (*f_clear_tids)(struct qib_devdata *, struct qib_ctxtdata *);
767 	void (*f_put_tid)(struct qib_devdata *, u64 __iomem*,
768 				u32, unsigned long);
769 	void (*f_cleanup)(struct qib_devdata *);
770 	void (*f_setextled)(struct qib_pportdata *, u32);
771 	/* fill out chip-specific fields */
772 	int (*f_get_base_info)(struct qib_ctxtdata *, struct qib_base_info *);
773 	/* free irq */
774 	void (*f_free_irq)(struct qib_devdata *);
775 	struct qib_message_header *(*f_get_msgheader)
776 					(struct qib_devdata *, __le32 *);
777 	void (*f_config_ctxts)(struct qib_devdata *);
778 	int (*f_get_ib_cfg)(struct qib_pportdata *, int);
779 	int (*f_set_ib_cfg)(struct qib_pportdata *, int, u32);
780 	int (*f_set_ib_loopback)(struct qib_pportdata *, const char *);
781 	int (*f_get_ib_table)(struct qib_pportdata *, int, void *);
782 	int (*f_set_ib_table)(struct qib_pportdata *, int, void *);
783 	u32 (*f_iblink_state)(u64);
784 	u8 (*f_ibphys_portstate)(u64);
785 	void (*f_xgxs_reset)(struct qib_pportdata *);
786 	/* per chip actions needed for IB Link up/down changes */
787 	int (*f_ib_updown)(struct qib_pportdata *, int, u64);
788 	u32 __iomem *(*f_getsendbuf)(struct qib_pportdata *, u64, u32 *);
789 	/* Read/modify/write of GPIO pins (potentially chip-specific */
790 	int (*f_gpio_mod)(struct qib_devdata *dd, u32 out, u32 dir,
791 		u32 mask);
792 	/* Enable writes to config EEPROM (if supported) */
793 	int (*f_eeprom_wen)(struct qib_devdata *dd, int wen);
794 	/*
795 	 * modify rcvctrl shadow[s] and write to appropriate chip-regs.
796 	 * see above QIB_RCVCTRL_xxx_ENB/DIS for operations.
797 	 * (ctxt == -1) means "all contexts", only meaningful for
798 	 * clearing. Could remove if chip_spec shutdown properly done.
799 	 */
800 	void (*f_rcvctrl)(struct qib_pportdata *, unsigned int op,
801 		int ctxt);
802 	/* Read/modify/write sendctrl appropriately for op and port. */
803 	void (*f_sendctrl)(struct qib_pportdata *, u32 op);
804 	void (*f_set_intr_state)(struct qib_devdata *, u32);
805 	void (*f_set_armlaunch)(struct qib_devdata *, u32);
806 	void (*f_wantpiobuf_intr)(struct qib_devdata *, u32);
807 	int (*f_late_initreg)(struct qib_devdata *);
808 	int (*f_init_sdma_regs)(struct qib_pportdata *);
809 	u16 (*f_sdma_gethead)(struct qib_pportdata *);
810 	int (*f_sdma_busy)(struct qib_pportdata *);
811 	void (*f_sdma_update_tail)(struct qib_pportdata *, u16);
812 	void (*f_sdma_set_desc_cnt)(struct qib_pportdata *, unsigned);
813 	void (*f_sdma_sendctrl)(struct qib_pportdata *, unsigned);
814 	void (*f_sdma_hw_clean_up)(struct qib_pportdata *);
815 	void (*f_sdma_hw_start_up)(struct qib_pportdata *);
816 	void (*f_sdma_init_early)(struct qib_pportdata *);
817 	void (*f_set_cntr_sample)(struct qib_pportdata *, u32, u32);
818 	void (*f_update_usrhead)(struct qib_ctxtdata *, u64, u32, u32, u32);
819 	u32 (*f_hdrqempty)(struct qib_ctxtdata *);
820 	u64 (*f_portcntr)(struct qib_pportdata *, u32);
821 	u32 (*f_read_cntrs)(struct qib_devdata *, loff_t, char **,
822 		u64 **);
823 	u32 (*f_read_portcntrs)(struct qib_devdata *, loff_t, u32,
824 		char **, u64 **);
825 	u32 (*f_setpbc_control)(struct qib_pportdata *, u32, u8, u8);
826 	void (*f_initvl15_bufs)(struct qib_devdata *);
827 	void (*f_init_ctxt)(struct qib_ctxtdata *);
828 	void (*f_txchk_change)(struct qib_devdata *, u32, u32, u32,
829 		struct qib_ctxtdata *);
830 	void (*f_writescratch)(struct qib_devdata *, u32);
831 	int (*f_tempsense_rd)(struct qib_devdata *, int regnum);
832 #ifdef CONFIG_INFINIBAND_QIB_DCA
833 	int (*f_notify_dca)(struct qib_devdata *, unsigned long event);
834 #endif
835 
836 	char *boardname; /* human readable board info */
837 
838 	/* template for writing TIDs  */
839 	u64 tidtemplate;
840 	/* value to write to free TIDs */
841 	u64 tidinvalid;
842 
843 	/* number of registers used for pioavail */
844 	u32 pioavregs;
845 	/* device (not port) flags, basically device capabilities */
846 	u32 flags;
847 	/* last buffer for user use */
848 	u32 lastctxt_piobuf;
849 
850 	/* reset value */
851 	u64 z_int_counter;
852 	/* percpu intcounter */
853 	u64 __percpu *int_counter;
854 
855 	/* pio bufs allocated per ctxt */
856 	u32 pbufsctxt;
857 	/* if remainder on bufs/ctxt, ctxts < extrabuf get 1 extra */
858 	u32 ctxts_extrabuf;
859 	/*
860 	 * number of ctxts configured as max; zero is set to number chip
861 	 * supports, less gives more pio bufs/ctxt, etc.
862 	 */
863 	u32 cfgctxts;
864 	/*
865 	 * number of ctxts available for PSM open
866 	 */
867 	u32 freectxts;
868 
869 	/*
870 	 * hint that we should update pioavailshadow before
871 	 * looking for a PIO buffer
872 	 */
873 	u32 upd_pio_shadow;
874 
875 	/* internal debugging stats */
876 	u32 maxpkts_call;
877 	u32 avgpkts_call;
878 	u64 nopiobufs;
879 
880 	/* PCI Vendor ID (here for NodeInfo) */
881 	u16 vendorid;
882 	/* PCI Device ID (here for NodeInfo) */
883 	u16 deviceid;
884 	/* for write combining settings */
885 	int wc_cookie;
886 	unsigned long wc_base;
887 	unsigned long wc_len;
888 
889 	/* shadow copy of struct page *'s for exp tid pages */
890 	struct page **pageshadow;
891 	/* shadow copy of dma handles for exp tid pages */
892 	dma_addr_t *physshadow;
893 	u64 __iomem *egrtidbase;
894 	spinlock_t sendctrl_lock; /* protect changes to sendctrl shadow */
895 	/* around rcd and (user ctxts) ctxt_cnt use (intr vs free) */
896 	spinlock_t uctxt_lock; /* rcd and user context changes */
897 	/*
898 	 * per unit status, see also portdata statusp
899 	 * mapped readonly into user processes so they can get unit and
900 	 * IB link status cheaply
901 	 */
902 	u64 *devstatusp;
903 	char *freezemsg; /* freeze msg if hw error put chip in freeze */
904 	u32 freezelen; /* max length of freezemsg */
905 	/* timer used to prevent stats overflow, error throttling, etc. */
906 	struct timer_list stats_timer;
907 
908 	/* timer to verify interrupts work, and fallback if possible */
909 	struct timer_list intrchk_timer;
910 	unsigned long ureg_align; /* user register alignment */
911 
912 	/*
913 	 * Protects pioavailshadow, pioavailkernel, pio_need_disarm, and
914 	 * pio_writing.
915 	 */
916 	spinlock_t pioavail_lock;
917 	/*
918 	 * index of last buffer to optimize search for next
919 	 */
920 	u32 last_pio;
921 	/*
922 	 * min kernel pio buffer to optimize search
923 	 */
924 	u32 min_kernel_pio;
925 	/*
926 	 * Shadow copies of registers; size indicates read access size.
927 	 * Most of them are readonly, but some are write-only register,
928 	 * where we manipulate the bits in the shadow copy, and then write
929 	 * the shadow copy to qlogic_ib.
930 	 *
931 	 * We deliberately make most of these 32 bits, since they have
932 	 * restricted range.  For any that we read, we won't to generate 32
933 	 * bit accesses, since Opteron will generate 2 separate 32 bit HT
934 	 * transactions for a 64 bit read, and we want to avoid unnecessary
935 	 * bus transactions.
936 	 */
937 
938 	/* This is the 64 bit group */
939 
940 	unsigned long pioavailshadow[6];
941 	/* bitmap of send buffers available for the kernel to use with PIO. */
942 	unsigned long pioavailkernel[6];
943 	/* bitmap of send buffers which need to be disarmed. */
944 	unsigned long pio_need_disarm[3];
945 	/* bitmap of send buffers which are being written to. */
946 	unsigned long pio_writing[3];
947 	/* kr_revision shadow */
948 	u64 revision;
949 	/* Base GUID for device (from eeprom, network order) */
950 	__be64 base_guid;
951 
952 	/*
953 	 * kr_sendpiobufbase value (chip offset of pio buffers), and the
954 	 * base of the 2KB buffer s(user processes only use 2K)
955 	 */
956 	u64 piobufbase;
957 	u32 pio2k_bufbase;
958 
959 	/* these are the "32 bit" regs */
960 
961 	/* number of GUIDs in the flash for this interface */
962 	u32 nguid;
963 	/*
964 	 * the following two are 32-bit bitmasks, but {test,clear,set}_bit
965 	 * all expect bit fields to be "unsigned long"
966 	 */
967 	unsigned long rcvctrl; /* shadow per device rcvctrl */
968 	unsigned long sendctrl; /* shadow per device sendctrl */
969 
970 	/* value we put in kr_rcvhdrcnt */
971 	u32 rcvhdrcnt;
972 	/* value we put in kr_rcvhdrsize */
973 	u32 rcvhdrsize;
974 	/* value we put in kr_rcvhdrentsize */
975 	u32 rcvhdrentsize;
976 	/* kr_ctxtcnt value */
977 	u32 ctxtcnt;
978 	/* kr_pagealign value */
979 	u32 palign;
980 	/* number of "2KB" PIO buffers */
981 	u32 piobcnt2k;
982 	/* size in bytes of "2KB" PIO buffers */
983 	u32 piosize2k;
984 	/* max usable size in dwords of a "2KB" PIO buffer before going "4KB" */
985 	u32 piosize2kmax_dwords;
986 	/* number of "4KB" PIO buffers */
987 	u32 piobcnt4k;
988 	/* size in bytes of "4KB" PIO buffers */
989 	u32 piosize4k;
990 	/* kr_rcvegrbase value */
991 	u32 rcvegrbase;
992 	/* kr_rcvtidbase value */
993 	u32 rcvtidbase;
994 	/* kr_rcvtidcnt value */
995 	u32 rcvtidcnt;
996 	/* kr_userregbase */
997 	u32 uregbase;
998 	/* shadow the control register contents */
999 	u32 control;
1000 
1001 	/* chip address space used by 4k pio buffers */
1002 	u32 align4k;
1003 	/* size of each rcvegrbuffer */
1004 	u16 rcvegrbufsize;
1005 	/* log2 of above */
1006 	u16 rcvegrbufsize_shift;
1007 	/* localbus width (1, 2,4,8,16,32) from config space  */
1008 	u32 lbus_width;
1009 	/* localbus speed in MHz */
1010 	u32 lbus_speed;
1011 	int unit; /* unit # of this chip */
1012 
1013 	/* start of CHIP_SPEC move to chipspec, but need code changes */
1014 	/* low and high portions of MSI capability/vector */
1015 	u32 msi_lo;
1016 	/* saved after PCIe init for restore after reset */
1017 	u32 msi_hi;
1018 	/* MSI data (vector) saved for restore */
1019 	u16 msi_data;
1020 	/* so we can rewrite it after a chip reset */
1021 	u32 pcibar0;
1022 	/* so we can rewrite it after a chip reset */
1023 	u32 pcibar1;
1024 	u64 rhdrhead_intr_off;
1025 
1026 	/*
1027 	 * ASCII serial number, from flash, large enough for original
1028 	 * all digit strings, and longer QLogic serial number format
1029 	 */
1030 	u8 serial[16];
1031 	/* human readable board version */
1032 	u8 boardversion[96];
1033 	u8 lbus_info[32]; /* human readable localbus info */
1034 	/* chip major rev, from qib_revision */
1035 	u8 majrev;
1036 	/* chip minor rev, from qib_revision */
1037 	u8 minrev;
1038 
1039 	/* Misc small ints */
1040 	/* Number of physical ports available */
1041 	u8 num_pports;
1042 	/* Lowest context number which can be used by user processes */
1043 	u8 first_user_ctxt;
1044 	u8 n_krcv_queues;
1045 	u8 qpn_mask;
1046 	u8 skip_kctxt_mask;
1047 
1048 	u16 rhf_offset; /* offset of RHF within receive header entry */
1049 
1050 	/*
1051 	 * GPIO pins for twsi-connected devices, and device code for eeprom
1052 	 */
1053 	u8 gpio_sda_num;
1054 	u8 gpio_scl_num;
1055 	u8 twsi_eeprom_dev;
1056 	u8 board_atten;
1057 
1058 	/* Support (including locks) for EEPROM logging of errors and time */
1059 	/* control access to actual counters, timer */
1060 	spinlock_t eep_st_lock;
1061 	/* control high-level access to EEPROM */
1062 	struct mutex eep_lock;
1063 	uint64_t traffic_wds;
1064 	struct qib_diag_client *diag_client;
1065 	spinlock_t qib_diag_trans_lock; /* protect diag observer ops */
1066 	struct diag_observer_list_elt *diag_observer_list;
1067 
1068 	u8 psxmitwait_supported;
1069 	/* cycle length of PS* counters in HW (in picoseconds) */
1070 	u16 psxmitwait_check_rate;
1071 	/* high volume overflow errors defered to tasklet */
1072 	struct tasklet_struct error_tasklet;
1073 
1074 	int assigned_node_id; /* NUMA node closest to HCA */
1075 };
1076 
1077 /* hol_state values */
1078 #define QIB_HOL_UP       0
1079 #define QIB_HOL_INIT     1
1080 
1081 #define QIB_SDMA_SENDCTRL_OP_ENABLE    (1U << 0)
1082 #define QIB_SDMA_SENDCTRL_OP_INTENABLE (1U << 1)
1083 #define QIB_SDMA_SENDCTRL_OP_HALT      (1U << 2)
1084 #define QIB_SDMA_SENDCTRL_OP_CLEANUP   (1U << 3)
1085 #define QIB_SDMA_SENDCTRL_OP_DRAIN     (1U << 4)
1086 
1087 /* operation types for f_txchk_change() */
1088 #define TXCHK_CHG_TYPE_DIS1  3
1089 #define TXCHK_CHG_TYPE_ENAB1 2
1090 #define TXCHK_CHG_TYPE_KERN  1
1091 #define TXCHK_CHG_TYPE_USER  0
1092 
1093 #define QIB_CHASE_TIME msecs_to_jiffies(145)
1094 #define QIB_CHASE_DIS_TIME msecs_to_jiffies(160)
1095 
1096 /* Private data for file operations */
1097 struct qib_filedata {
1098 	struct qib_ctxtdata *rcd;
1099 	unsigned subctxt;
1100 	unsigned tidcursor;
1101 	struct qib_user_sdma_queue *pq;
1102 	int rec_cpu_num; /* for cpu affinity; -1 if none */
1103 };
1104 
1105 extern struct xarray qib_dev_table;
1106 extern struct qib_devdata *qib_lookup(int unit);
1107 extern u32 qib_cpulist_count;
1108 extern unsigned long *qib_cpulist;
1109 extern unsigned qib_cc_table_size;
1110 
1111 int qib_init(struct qib_devdata *, int);
1112 int init_chip_wc_pat(struct qib_devdata *dd, u32);
1113 int qib_enable_wc(struct qib_devdata *dd);
1114 void qib_disable_wc(struct qib_devdata *dd);
1115 int qib_count_units(int *npresentp, int *nupp);
1116 int qib_count_active_units(void);
1117 
1118 int qib_cdev_init(int minor, const char *name,
1119 		  const struct file_operations *fops,
1120 		  struct cdev **cdevp, struct device **devp);
1121 void qib_cdev_cleanup(struct cdev **cdevp, struct device **devp);
1122 int qib_dev_init(void);
1123 void qib_dev_cleanup(void);
1124 
1125 int qib_diag_add(struct qib_devdata *);
1126 void qib_diag_remove(struct qib_devdata *);
1127 void qib_handle_e_ibstatuschanged(struct qib_pportdata *, u64);
1128 void qib_sdma_update_tail(struct qib_pportdata *, u16); /* hold sdma_lock */
1129 
1130 int qib_decode_err(struct qib_devdata *dd, char *buf, size_t blen, u64 err);
1131 void qib_bad_intrstatus(struct qib_devdata *);
1132 void qib_handle_urcv(struct qib_devdata *, u64);
1133 
1134 /* clean up any per-chip chip-specific stuff */
1135 void qib_chip_cleanup(struct qib_devdata *);
1136 /* clean up any chip type-specific stuff */
1137 void qib_chip_done(void);
1138 
1139 /* check to see if we have to force ordering for write combining */
1140 int qib_unordered_wc(void);
1141 void qib_pio_copy(void __iomem *to, const void *from, size_t count);
1142 
1143 void qib_disarm_piobufs(struct qib_devdata *, unsigned, unsigned);
1144 int qib_disarm_piobufs_ifneeded(struct qib_ctxtdata *);
1145 void qib_disarm_piobufs_set(struct qib_devdata *, unsigned long *, unsigned);
1146 void qib_cancel_sends(struct qib_pportdata *);
1147 
1148 int qib_create_rcvhdrq(struct qib_devdata *, struct qib_ctxtdata *);
1149 int qib_setup_eagerbufs(struct qib_ctxtdata *);
1150 void qib_set_ctxtcnt(struct qib_devdata *);
1151 int qib_create_ctxts(struct qib_devdata *dd);
1152 struct qib_ctxtdata *qib_create_ctxtdata(struct qib_pportdata *, u32, int);
1153 int qib_init_pportdata(struct qib_pportdata *, struct qib_devdata *, u8, u8);
1154 void qib_free_ctxtdata(struct qib_devdata *, struct qib_ctxtdata *);
1155 
1156 u32 qib_kreceive(struct qib_ctxtdata *, u32 *, u32 *);
1157 int qib_reset_device(int);
1158 int qib_wait_linkstate(struct qib_pportdata *, u32, int);
1159 int qib_set_linkstate(struct qib_pportdata *, u8);
1160 int qib_set_mtu(struct qib_pportdata *, u16);
1161 int qib_set_lid(struct qib_pportdata *, u32, u8);
1162 void qib_hol_down(struct qib_pportdata *);
1163 void qib_hol_init(struct qib_pportdata *);
1164 void qib_hol_up(struct qib_pportdata *);
1165 void qib_hol_event(struct timer_list *);
1166 void qib_disable_after_error(struct qib_devdata *);
1167 int qib_set_uevent_bits(struct qib_pportdata *, const int);
1168 
1169 /* for use in system calls, where we want to know device type, etc. */
1170 #define ctxt_fp(fp) \
1171 	(((struct qib_filedata *)(fp)->private_data)->rcd)
1172 #define subctxt_fp(fp) \
1173 	(((struct qib_filedata *)(fp)->private_data)->subctxt)
1174 #define tidcursor_fp(fp) \
1175 	(((struct qib_filedata *)(fp)->private_data)->tidcursor)
1176 #define user_sdma_queue_fp(fp) \
1177 	(((struct qib_filedata *)(fp)->private_data)->pq)
1178 
1179 static inline struct qib_devdata *dd_from_ppd(struct qib_pportdata *ppd)
1180 {
1181 	return ppd->dd;
1182 }
1183 
1184 static inline struct qib_devdata *dd_from_dev(struct qib_ibdev *dev)
1185 {
1186 	return container_of(dev, struct qib_devdata, verbs_dev);
1187 }
1188 
1189 static inline struct qib_devdata *dd_from_ibdev(struct ib_device *ibdev)
1190 {
1191 	return dd_from_dev(to_idev(ibdev));
1192 }
1193 
1194 static inline struct qib_pportdata *ppd_from_ibp(struct qib_ibport *ibp)
1195 {
1196 	return container_of(ibp, struct qib_pportdata, ibport_data);
1197 }
1198 
1199 static inline struct qib_ibport *to_iport(struct ib_device *ibdev, u32 port)
1200 {
1201 	struct qib_devdata *dd = dd_from_ibdev(ibdev);
1202 	u32 pidx = port - 1; /* IB number port from 1, hdw from 0 */
1203 
1204 	WARN_ON(pidx >= dd->num_pports);
1205 	return &dd->pport[pidx].ibport_data;
1206 }
1207 
1208 /*
1209  * values for dd->flags (_device_ related flags) and
1210  */
1211 #define QIB_HAS_LINK_LATENCY  0x1 /* supports link latency (IB 1.2) */
1212 #define QIB_INITTED           0x2 /* chip and driver up and initted */
1213 #define QIB_DOING_RESET       0x4  /* in the middle of doing chip reset */
1214 #define QIB_PRESENT           0x8  /* chip accesses can be done */
1215 #define QIB_PIO_FLUSH_WC      0x10 /* Needs Write combining flush for PIO */
1216 #define QIB_HAS_THRESH_UPDATE 0x40
1217 #define QIB_HAS_SDMA_TIMEOUT  0x80
1218 #define QIB_USE_SPCL_TRIG     0x100 /* SpecialTrigger launch enabled */
1219 #define QIB_NODMA_RTAIL       0x200 /* rcvhdrtail register DMA enabled */
1220 #define QIB_HAS_INTX          0x800 /* Supports INTx interrupts */
1221 #define QIB_HAS_SEND_DMA      0x1000 /* Supports Send DMA */
1222 #define QIB_HAS_VLSUPP        0x2000 /* Supports multiple VLs; PBC different */
1223 #define QIB_HAS_HDRSUPP       0x4000 /* Supports header suppression */
1224 #define QIB_BADINTR           0x8000 /* severe interrupt problems */
1225 #define QIB_DCA_ENABLED       0x10000 /* Direct Cache Access enabled */
1226 #define QIB_HAS_QSFP          0x20000 /* device (card instance) has QSFP */
1227 #define QIB_SHUTDOWN          0x40000 /* device is shutting down */
1228 
1229 /*
1230  * values for ppd->lflags (_ib_port_ related flags)
1231  */
1232 #define QIBL_LINKV             0x1 /* IB link state valid */
1233 #define QIBL_LINKDOWN          0x8 /* IB link is down */
1234 #define QIBL_LINKINIT          0x10 /* IB link level is up */
1235 #define QIBL_LINKARMED         0x20 /* IB link is ARMED */
1236 #define QIBL_LINKACTIVE        0x40 /* IB link is ACTIVE */
1237 /* leave a gap for more IB-link state */
1238 #define QIBL_IB_AUTONEG_INPROG 0x1000 /* non-IBTA DDR/QDR neg active */
1239 #define QIBL_IB_AUTONEG_FAILED 0x2000 /* non-IBTA DDR/QDR neg failed */
1240 #define QIBL_IB_LINK_DISABLED  0x4000 /* Linkdown-disable forced,
1241 				       * Do not try to bring up */
1242 #define QIBL_IB_FORCE_NOTIFY   0x8000 /* force notify on next ib change */
1243 
1244 /* IB dword length mask in PBC (lower 11 bits); same for all chips */
1245 #define QIB_PBC_LENGTH_MASK                     ((1 << 11) - 1)
1246 
1247 
1248 /* ctxt_flag bit offsets */
1249 		/* waiting for a packet to arrive */
1250 #define QIB_CTXT_WAITING_RCV   2
1251 		/* master has not finished initializing */
1252 #define QIB_CTXT_MASTER_UNINIT 4
1253 		/* waiting for an urgent packet to arrive */
1254 #define QIB_CTXT_WAITING_URG 5
1255 
1256 /* free up any allocated data at closes */
1257 void qib_free_data(struct qib_ctxtdata *dd);
1258 void qib_chg_pioavailkernel(struct qib_devdata *, unsigned, unsigned,
1259 			    u32, struct qib_ctxtdata *);
1260 struct qib_devdata *qib_init_iba7322_funcs(struct pci_dev *,
1261 					   const struct pci_device_id *);
1262 struct qib_devdata *qib_init_iba7220_funcs(struct pci_dev *,
1263 					   const struct pci_device_id *);
1264 struct qib_devdata *qib_init_iba6120_funcs(struct pci_dev *,
1265 					   const struct pci_device_id *);
1266 void qib_free_devdata(struct qib_devdata *);
1267 struct qib_devdata *qib_alloc_devdata(struct pci_dev *pdev, size_t extra);
1268 
1269 #define QIB_TWSI_NO_DEV 0xFF
1270 /* Below qib_twsi_ functions must be called with eep_lock held */
1271 int qib_twsi_reset(struct qib_devdata *dd);
1272 int qib_twsi_blk_rd(struct qib_devdata *dd, int dev, int addr, void *buffer,
1273 		    int len);
1274 int qib_twsi_blk_wr(struct qib_devdata *dd, int dev, int addr,
1275 		    const void *buffer, int len);
1276 void qib_get_eeprom_info(struct qib_devdata *);
1277 void qib_dump_lookup_output_queue(struct qib_devdata *);
1278 void qib_force_pio_avail_update(struct qib_devdata *);
1279 void qib_clear_symerror_on_linkup(struct timer_list *t);
1280 
1281 /*
1282  * Set LED override, only the two LSBs have "public" meaning, but
1283  * any non-zero value substitutes them for the Link and LinkTrain
1284  * LED states.
1285  */
1286 #define QIB_LED_PHYS 1 /* Physical (linktraining) GREEN LED */
1287 #define QIB_LED_LOG 2  /* Logical (link) YELLOW LED */
1288 void qib_set_led_override(struct qib_pportdata *ppd, unsigned int val);
1289 
1290 /* send dma routines */
1291 int qib_setup_sdma(struct qib_pportdata *);
1292 void qib_teardown_sdma(struct qib_pportdata *);
1293 void __qib_sdma_intr(struct qib_pportdata *);
1294 void qib_sdma_intr(struct qib_pportdata *);
1295 void qib_user_sdma_send_desc(struct qib_pportdata *dd,
1296 			struct list_head *pktlist);
1297 int qib_sdma_verbs_send(struct qib_pportdata *, struct rvt_sge_state *,
1298 			u32, struct qib_verbs_txreq *);
1299 /* ppd->sdma_lock should be locked before calling this. */
1300 int qib_sdma_make_progress(struct qib_pportdata *dd);
1301 
1302 /* must be called under qib_sdma_lock */
1303 static inline u16 qib_sdma_descq_freecnt(const struct qib_pportdata *ppd)
1304 {
1305 	return ppd->sdma_descq_cnt -
1306 		(ppd->sdma_descq_added - ppd->sdma_descq_removed) - 1;
1307 }
1308 
1309 static inline int __qib_sdma_running(struct qib_pportdata *ppd)
1310 {
1311 	return ppd->sdma_state.current_state == qib_sdma_state_s99_running;
1312 }
1313 int qib_sdma_running(struct qib_pportdata *);
1314 void dump_sdma_state(struct qib_pportdata *ppd);
1315 void __qib_sdma_process_event(struct qib_pportdata *, enum qib_sdma_events);
1316 void qib_sdma_process_event(struct qib_pportdata *, enum qib_sdma_events);
1317 
1318 /*
1319  * number of words used for protocol header if not set by qib_userinit();
1320  */
1321 #define QIB_DFLT_RCVHDRSIZE 9
1322 
1323 /*
1324  * We need to be able to handle an IB header of at least 24 dwords.
1325  * We need the rcvhdrq large enough to handle largest IB header, but
1326  * still have room for a 2KB MTU standard IB packet.
1327  * Additionally, some processor/memory controller combinations
1328  * benefit quite strongly from having the DMA'ed data be cacheline
1329  * aligned and a cacheline multiple, so we set the size to 32 dwords
1330  * (2 64-byte primary cachelines for pretty much all processors of
1331  * interest).  The alignment hurts nothing, other than using somewhat
1332  * more memory.
1333  */
1334 #define QIB_RCVHDR_ENTSIZE 32
1335 
1336 int qib_get_user_pages(unsigned long, size_t, struct page **);
1337 void qib_release_user_pages(struct page **, size_t);
1338 int qib_eeprom_read(struct qib_devdata *, u8, void *, int);
1339 int qib_eeprom_write(struct qib_devdata *, u8, const void *, int);
1340 u32 __iomem *qib_getsendbuf_range(struct qib_devdata *, u32 *, u32, u32);
1341 void qib_sendbuf_done(struct qib_devdata *, unsigned);
1342 
1343 static inline void qib_clear_rcvhdrtail(const struct qib_ctxtdata *rcd)
1344 {
1345 	*((u64 *) rcd->rcvhdrtail_kvaddr) = 0ULL;
1346 }
1347 
1348 static inline u32 qib_get_rcvhdrtail(const struct qib_ctxtdata *rcd)
1349 {
1350 	/*
1351 	 * volatile because it's a DMA target from the chip, routine is
1352 	 * inlined, and don't want register caching or reordering.
1353 	 */
1354 	return (u32) le64_to_cpu(
1355 		*((volatile __le64 *)rcd->rcvhdrtail_kvaddr)); /* DMA'ed */
1356 }
1357 
1358 /*
1359  * sysfs interface.
1360  */
1361 
1362 extern const char ib_qib_version[];
1363 extern const struct attribute_group qib_attr_group;
1364 extern const struct attribute_group *qib_attr_port_groups[];
1365 
1366 int qib_device_create(struct qib_devdata *);
1367 void qib_device_remove(struct qib_devdata *);
1368 
1369 /* Hook for sysfs read of QSFP */
1370 extern int qib_qsfp_dump(struct qib_pportdata *ppd, char *buf, int len);
1371 
1372 int __init qib_init_qibfs(void);
1373 int __exit qib_exit_qibfs(void);
1374 
1375 int qibfs_add(struct qib_devdata *);
1376 int qibfs_remove(struct qib_devdata *);
1377 
1378 int qib_pcie_init(struct pci_dev *, const struct pci_device_id *);
1379 int qib_pcie_ddinit(struct qib_devdata *, struct pci_dev *,
1380 		    const struct pci_device_id *);
1381 void qib_pcie_ddcleanup(struct qib_devdata *);
1382 int qib_pcie_params(struct qib_devdata *dd, u32 minw, u32 *nent);
1383 void qib_free_irq(struct qib_devdata *dd);
1384 int qib_reinit_intr(struct qib_devdata *dd);
1385 void qib_pcie_getcmd(struct qib_devdata *, u16 *, u8 *, u8 *);
1386 void qib_pcie_reenable(struct qib_devdata *, u16, u8, u8);
1387 /* interrupts for device */
1388 u64 qib_int_counter(struct qib_devdata *);
1389 /* interrupt for all devices */
1390 u64 qib_sps_ints(void);
1391 
1392 /*
1393  * dma_addr wrappers - all 0's invalid for hw
1394  */
1395 int qib_map_page(struct pci_dev *d, struct page *p, dma_addr_t *daddr);
1396 struct pci_dev *qib_get_pci_dev(struct rvt_dev_info *rdi);
1397 
1398 /*
1399  * Flush write combining store buffers (if present) and perform a write
1400  * barrier.
1401  */
1402 static inline void qib_flush_wc(void)
1403 {
1404 #if defined(CONFIG_X86_64)
1405 	asm volatile("sfence" : : : "memory");
1406 #else
1407 	wmb(); /* no reorder around wc flush */
1408 #endif
1409 }
1410 
1411 /* global module parameter variables */
1412 extern unsigned qib_ibmtu;
1413 extern ushort qib_cfgctxts;
1414 extern ushort qib_num_cfg_vls;
1415 extern ushort qib_mini_init; /* If set, do few (ideally 0) writes to chip */
1416 extern unsigned qib_n_krcv_queues;
1417 extern unsigned qib_sdma_fetch_arb;
1418 extern unsigned qib_compat_ddr_negotiate;
1419 extern int qib_special_trigger;
1420 extern unsigned qib_numa_aware;
1421 
1422 extern struct mutex qib_mutex;
1423 
1424 /* Number of seconds before our card status check...  */
1425 #define STATUS_TIMEOUT 60
1426 
1427 #define QIB_DRV_NAME            "ib_qib"
1428 #define QIB_USER_MINOR_BASE     0
1429 #define QIB_TRACE_MINOR         127
1430 #define QIB_DIAGPKT_MINOR       128
1431 #define QIB_DIAG_MINOR_BASE     129
1432 #define QIB_NMINORS             255
1433 
1434 #define PCI_VENDOR_ID_PATHSCALE 0x1fc1
1435 #define PCI_VENDOR_ID_QLOGIC 0x1077
1436 #define PCI_DEVICE_ID_QLOGIC_IB_6120 0x10
1437 #define PCI_DEVICE_ID_QLOGIC_IB_7220 0x7220
1438 #define PCI_DEVICE_ID_QLOGIC_IB_7322 0x7322
1439 
1440 /*
1441  * qib_early_err is used (only!) to print early errors before devdata is
1442  * allocated, or when dd->pcidev may not be valid, and at the tail end of
1443  * cleanup when devdata may have been freed, etc.  qib_dev_porterr is
1444  * the same as qib_dev_err, but is used when the message really needs
1445  * the IB port# to be definitive as to what's happening..
1446  * All of these go to the trace log, and the trace log entry is done
1447  * first to avoid possible serial port delays from printk.
1448  */
1449 #define qib_early_err(dev, fmt, ...) \
1450 	dev_err(dev, fmt, ##__VA_ARGS__)
1451 
1452 #define qib_dev_err(dd, fmt, ...) \
1453 	dev_err(&(dd)->pcidev->dev, "%s: " fmt, \
1454 		rvt_get_ibdev_name(&(dd)->verbs_dev.rdi), ##__VA_ARGS__)
1455 
1456 #define qib_dev_warn(dd, fmt, ...) \
1457 	dev_warn(&(dd)->pcidev->dev, "%s: " fmt, \
1458 		 rvt_get_ibdev_name(&(dd)->verbs_dev.rdi), ##__VA_ARGS__)
1459 
1460 #define qib_dev_porterr(dd, port, fmt, ...) \
1461 	dev_err(&(dd)->pcidev->dev, "%s: IB%u:%u " fmt, \
1462 		rvt_get_ibdev_name(&(dd)->verbs_dev.rdi), (dd)->unit, (port), \
1463 		##__VA_ARGS__)
1464 
1465 #define qib_devinfo(pcidev, fmt, ...) \
1466 	dev_info(&(pcidev)->dev, fmt, ##__VA_ARGS__)
1467 
1468 /*
1469  * this is used for formatting hw error messages...
1470  */
1471 struct qib_hwerror_msgs {
1472 	u64 mask;
1473 	const char *msg;
1474 	size_t sz;
1475 };
1476 
1477 #define QLOGIC_IB_HWE_MSG(a, b) { .mask = a, .msg = b }
1478 
1479 /* in qib_intr.c... */
1480 void qib_format_hwerrors(u64 hwerrs,
1481 			 const struct qib_hwerror_msgs *hwerrmsgs,
1482 			 size_t nhwerrmsgs, char *msg, size_t lmsg);
1483 
1484 void qib_stop_send_queue(struct rvt_qp *qp);
1485 void qib_quiesce_qp(struct rvt_qp *qp);
1486 void qib_flush_qp_waiters(struct rvt_qp *qp);
1487 int qib_mtu_to_path_mtu(u32 mtu);
1488 u32 qib_mtu_from_qp(struct rvt_dev_info *rdi, struct rvt_qp *qp, u32 pmtu);
1489 void qib_notify_error_qp(struct rvt_qp *qp);
1490 int qib_get_pmtu_from_attr(struct rvt_dev_info *rdi, struct rvt_qp *qp,
1491 			   struct ib_qp_attr *attr);
1492 
1493 #endif                          /* _QIB_KERNEL_H */
1494