xref: /dpdk/drivers/bus/dpaa/base/qbman/qman.c (revision d81734caccade4dc17d24d2ffd8b71244d35a69f)
1 /* SPDX-License-Identifier: (BSD-3-Clause OR GPL-2.0)
2  *
3  * Copyright 2008-2016 Freescale Semiconductor Inc.
4  * Copyright 2017 NXP
5  *
6  */
7 
8 #include "qman.h"
9 #include <rte_branch_prediction.h>
10 
11 /* Compilation constants */
12 #define DQRR_MAXFILL	15
13 #define EQCR_ITHRESH	4	/* if EQCR congests, interrupt threshold */
14 #define IRQNAME		"QMan portal %d"
15 #define MAX_IRQNAME	16	/* big enough for "QMan portal %d" */
16 /* maximum number of DQRR entries to process in qman_poll() */
17 #define FSL_QMAN_POLL_LIMIT 8
18 
19 /* Lock/unlock frame queues, subject to the "LOCKED" flag. This is about
20  * inter-processor locking only. Note, FQLOCK() is always called either under a
21  * local_irq_save() or from interrupt context - hence there's no need for irq
22  * protection (and indeed, attempting to nest irq-protection doesn't work, as
23  * the "irq en/disable" machinery isn't recursive...).
24  */
25 #define FQLOCK(fq) \
26 	do { \
27 		struct qman_fq *__fq478 = (fq); \
28 		if (fq_isset(__fq478, QMAN_FQ_FLAG_LOCKED)) \
29 			spin_lock(&__fq478->fqlock); \
30 	} while (0)
31 #define FQUNLOCK(fq) \
32 	do { \
33 		struct qman_fq *__fq478 = (fq); \
34 		if (fq_isset(__fq478, QMAN_FQ_FLAG_LOCKED)) \
35 			spin_unlock(&__fq478->fqlock); \
36 	} while (0)
37 
38 static inline void fq_set(struct qman_fq *fq, u32 mask)
39 {
40 	dpaa_set_bits(mask, &fq->flags);
41 }
42 
43 static inline void fq_clear(struct qman_fq *fq, u32 mask)
44 {
45 	dpaa_clear_bits(mask, &fq->flags);
46 }
47 
48 static inline int fq_isset(struct qman_fq *fq, u32 mask)
49 {
50 	return fq->flags & mask;
51 }
52 
53 static inline int fq_isclear(struct qman_fq *fq, u32 mask)
54 {
55 	return !(fq->flags & mask);
56 }
57 
58 struct qman_portal {
59 	struct qm_portal p;
60 	/* PORTAL_BITS_*** - dynamic, strictly internal */
61 	unsigned long bits;
62 	/* interrupt sources processed by portal_isr(), configurable */
63 	unsigned long irq_sources;
64 	u32 use_eqcr_ci_stashing;
65 	u32 slowpoll;	/* only used when interrupts are off */
66 	/* only 1 volatile dequeue at a time */
67 	struct qman_fq *vdqcr_owned;
68 	u32 sdqcr;
69 	int dqrr_disable_ref;
70 	/* A portal-specific handler for DCP ERNs. If this is NULL, the global
71 	 * handler is called instead.
72 	 */
73 	qman_cb_dc_ern cb_dc_ern;
74 	/* When the cpu-affine portal is activated, this is non-NULL */
75 	const struct qm_portal_config *config;
76 	struct dpa_rbtree retire_table;
77 	char irqname[MAX_IRQNAME];
78 	/* 2-element array. cgrs[0] is mask, cgrs[1] is snapshot. */
79 	struct qman_cgrs *cgrs;
80 	/* linked-list of CSCN handlers. */
81 	struct list_head cgr_cbs;
82 	/* list lock */
83 	spinlock_t cgr_lock;
84 	/* track if memory was allocated by the driver */
85 #if __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__
86 	/* Keep a shadow copy of the DQRR on LE systems as the SW needs to
87 	 * do byte swaps of DQRR read only memory.  First entry must be aligned
88 	 * to 2 ** 10 to ensure DQRR index calculations based shadow copy
89 	 * address (6 bits for address shift + 4 bits for the DQRR size).
90 	 */
91 	struct qm_dqrr_entry shadow_dqrr[QM_DQRR_SIZE]
92 		    __attribute__((aligned(1024)));
93 #endif
94 };
95 
96 /* Global handler for DCP ERNs. Used when the portal receiving the message does
97  * not have a portal-specific handler.
98  */
99 static qman_cb_dc_ern cb_dc_ern;
100 
101 static cpumask_t affine_mask;
102 static DEFINE_SPINLOCK(affine_mask_lock);
103 static u16 affine_channels[NR_CPUS];
104 static RTE_DEFINE_PER_LCORE(struct qman_portal, qman_affine_portal);
105 
106 static inline struct qman_portal *get_affine_portal(void)
107 {
108 	return &RTE_PER_LCORE(qman_affine_portal);
109 }
110 
111 /* This gives a FQID->FQ lookup to cover the fact that we can't directly demux
112  * retirement notifications (the fact they are sometimes h/w-consumed means that
113  * contextB isn't always a s/w demux - and as we can't know which case it is
114  * when looking at the notification, we have to use the slow lookup for all of
115  * them). NB, it's possible to have multiple FQ objects refer to the same FQID
116  * (though at most one of them should be the consumer), so this table isn't for
117  * all FQs - FQs are added when retirement commands are issued, and removed when
118  * they complete, which also massively reduces the size of this table.
119  */
120 IMPLEMENT_DPAA_RBTREE(fqtree, struct qman_fq, node, fqid);
121 /*
122  * This is what everything can wait on, even if it migrates to a different cpu
123  * to the one whose affine portal it is waiting on.
124  */
125 static DECLARE_WAIT_QUEUE_HEAD(affine_queue);
126 
127 static inline int table_push_fq(struct qman_portal *p, struct qman_fq *fq)
128 {
129 	int ret = fqtree_push(&p->retire_table, fq);
130 
131 	if (ret)
132 		pr_err("ERROR: double FQ-retirement %d\n", fq->fqid);
133 	return ret;
134 }
135 
136 static inline void table_del_fq(struct qman_portal *p, struct qman_fq *fq)
137 {
138 	fqtree_del(&p->retire_table, fq);
139 }
140 
141 static inline struct qman_fq *table_find_fq(struct qman_portal *p, u32 fqid)
142 {
143 	return fqtree_find(&p->retire_table, fqid);
144 }
145 
146 #ifdef CONFIG_FSL_QMAN_FQ_LOOKUP
147 static void **qman_fq_lookup_table;
148 static size_t qman_fq_lookup_table_size;
149 
150 int qman_setup_fq_lookup_table(size_t num_entries)
151 {
152 	num_entries++;
153 	/* Allocate 1 more entry since the first entry is not used */
154 	qman_fq_lookup_table = vmalloc((num_entries * sizeof(void *)));
155 	if (!qman_fq_lookup_table) {
156 		pr_err("QMan: Could not allocate fq lookup table\n");
157 		return -ENOMEM;
158 	}
159 	memset(qman_fq_lookup_table, 0, num_entries * sizeof(void *));
160 	qman_fq_lookup_table_size = num_entries;
161 	pr_debug("QMan: Allocated lookup table at %p, entry count %lu\n",
162 		qman_fq_lookup_table,
163 			(unsigned long)qman_fq_lookup_table_size);
164 	return 0;
165 }
166 
167 /* global structure that maintains fq object mapping */
168 static DEFINE_SPINLOCK(fq_hash_table_lock);
169 
170 static int find_empty_fq_table_entry(u32 *entry, struct qman_fq *fq)
171 {
172 	u32 i;
173 
174 	spin_lock(&fq_hash_table_lock);
175 	/* Can't use index zero because this has special meaning
176 	 * in context_b field.
177 	 */
178 	for (i = 1; i < qman_fq_lookup_table_size; i++) {
179 		if (qman_fq_lookup_table[i] == NULL) {
180 			*entry = i;
181 			qman_fq_lookup_table[i] = fq;
182 			spin_unlock(&fq_hash_table_lock);
183 			return 0;
184 		}
185 	}
186 	spin_unlock(&fq_hash_table_lock);
187 	return -ENOMEM;
188 }
189 
190 static void clear_fq_table_entry(u32 entry)
191 {
192 	spin_lock(&fq_hash_table_lock);
193 	DPAA_BUG_ON(entry >= qman_fq_lookup_table_size);
194 	qman_fq_lookup_table[entry] = NULL;
195 	spin_unlock(&fq_hash_table_lock);
196 }
197 
198 static inline struct qman_fq *get_fq_table_entry(u32 entry)
199 {
200 	DPAA_BUG_ON(entry >= qman_fq_lookup_table_size);
201 	return qman_fq_lookup_table[entry];
202 }
203 #endif
204 
205 static inline void cpu_to_hw_fqd(struct qm_fqd *fqd)
206 {
207 	/* Byteswap the FQD to HW format */
208 	fqd->fq_ctrl = cpu_to_be16(fqd->fq_ctrl);
209 	fqd->dest_wq = cpu_to_be16(fqd->dest_wq);
210 	fqd->ics_cred = cpu_to_be16(fqd->ics_cred);
211 	fqd->context_b = cpu_to_be32(fqd->context_b);
212 	fqd->context_a.opaque = cpu_to_be64(fqd->context_a.opaque);
213 	fqd->opaque_td = cpu_to_be16(fqd->opaque_td);
214 }
215 
216 static inline void hw_fqd_to_cpu(struct qm_fqd *fqd)
217 {
218 	/* Byteswap the FQD to CPU format */
219 	fqd->fq_ctrl = be16_to_cpu(fqd->fq_ctrl);
220 	fqd->dest_wq = be16_to_cpu(fqd->dest_wq);
221 	fqd->ics_cred = be16_to_cpu(fqd->ics_cred);
222 	fqd->context_b = be32_to_cpu(fqd->context_b);
223 	fqd->context_a.opaque = be64_to_cpu(fqd->context_a.opaque);
224 }
225 
226 static inline void cpu_to_hw_fd(struct qm_fd *fd)
227 {
228 	fd->addr = cpu_to_be40(fd->addr);
229 	fd->status = cpu_to_be32(fd->status);
230 	fd->opaque = cpu_to_be32(fd->opaque);
231 }
232 
233 static inline void hw_fd_to_cpu(struct qm_fd *fd)
234 {
235 	fd->addr = be40_to_cpu(fd->addr);
236 	fd->status = be32_to_cpu(fd->status);
237 	fd->opaque = be32_to_cpu(fd->opaque);
238 }
239 
240 /* In the case that slow- and fast-path handling are both done by qman_poll()
241  * (ie. because there is no interrupt handling), we ought to balance how often
242  * we do the fast-path poll versus the slow-path poll. We'll use two decrementer
243  * sources, so we call the fast poll 'n' times before calling the slow poll
244  * once. The idle decrementer constant is used when the last slow-poll detected
245  * no work to do, and the busy decrementer constant when the last slow-poll had
246  * work to do.
247  */
248 #define SLOW_POLL_IDLE   1000
249 #define SLOW_POLL_BUSY   10
250 static u32 __poll_portal_slow(struct qman_portal *p, u32 is);
251 static inline unsigned int __poll_portal_fast(struct qman_portal *p,
252 					      unsigned int poll_limit);
253 
254 /* Portal interrupt handler */
255 static irqreturn_t portal_isr(__always_unused int irq, void *ptr)
256 {
257 	struct qman_portal *p = ptr;
258 	/*
259 	 * The CSCI/CCSCI source is cleared inside __poll_portal_slow(), because
260 	 * it could race against a Query Congestion State command also given
261 	 * as part of the handling of this interrupt source. We mustn't
262 	 * clear it a second time in this top-level function.
263 	 */
264 	u32 clear = QM_DQAVAIL_MASK | (p->irq_sources &
265 		~(QM_PIRQ_CSCI | QM_PIRQ_CCSCI));
266 	u32 is = qm_isr_status_read(&p->p) & p->irq_sources;
267 	/* DQRR-handling if it's interrupt-driven */
268 	if (is & QM_PIRQ_DQRI)
269 		__poll_portal_fast(p, FSL_QMAN_POLL_LIMIT);
270 	/* Handling of anything else that's interrupt-driven */
271 	clear |= __poll_portal_slow(p, is);
272 	qm_isr_status_clear(&p->p, clear);
273 	return IRQ_HANDLED;
274 }
275 
276 /* This inner version is used privately by qman_create_affine_portal(), as well
277  * as by the exported qman_stop_dequeues().
278  */
279 static inline void qman_stop_dequeues_ex(struct qman_portal *p)
280 {
281 	if (!(p->dqrr_disable_ref++))
282 		qm_dqrr_set_maxfill(&p->p, 0);
283 }
284 
285 static int drain_mr_fqrni(struct qm_portal *p)
286 {
287 	const struct qm_mr_entry *msg;
288 loop:
289 	msg = qm_mr_current(p);
290 	if (!msg) {
291 		/*
292 		 * if MR was full and h/w had other FQRNI entries to produce, we
293 		 * need to allow it time to produce those entries once the
294 		 * existing entries are consumed. A worst-case situation
295 		 * (fully-loaded system) means h/w sequencers may have to do 3-4
296 		 * other things before servicing the portal's MR pump, each of
297 		 * which (if slow) may take ~50 qman cycles (which is ~200
298 		 * processor cycles). So rounding up and then multiplying this
299 		 * worst-case estimate by a factor of 10, just to be
300 		 * ultra-paranoid, goes as high as 10,000 cycles. NB, we consume
301 		 * one entry at a time, so h/w has an opportunity to produce new
302 		 * entries well before the ring has been fully consumed, so
303 		 * we're being *really* paranoid here.
304 		 */
305 		u64 now, then = mfatb();
306 
307 		do {
308 			now = mfatb();
309 		} while ((then + 10000) > now);
310 		msg = qm_mr_current(p);
311 		if (!msg)
312 			return 0;
313 	}
314 	if ((msg->verb & QM_MR_VERB_TYPE_MASK) != QM_MR_VERB_FQRNI) {
315 		/* We aren't draining anything but FQRNIs */
316 		pr_err("Found verb 0x%x in MR\n", msg->verb);
317 		return -1;
318 	}
319 	qm_mr_next(p);
320 	qm_mr_cci_consume(p, 1);
321 	goto loop;
322 }
323 
324 static inline int qm_eqcr_init(struct qm_portal *portal,
325 			       enum qm_eqcr_pmode pmode,
326 			       unsigned int eq_stash_thresh,
327 			       int eq_stash_prio)
328 {
329 	/* This use of 'register', as well as all other occurrences, is because
330 	 * it has been observed to generate much faster code with gcc than is
331 	 * otherwise the case.
332 	 */
333 	register struct qm_eqcr *eqcr = &portal->eqcr;
334 	u32 cfg;
335 	u8 pi;
336 
337 	eqcr->ring = portal->addr.ce + QM_CL_EQCR;
338 	eqcr->ci = qm_in(EQCR_CI_CINH) & (QM_EQCR_SIZE - 1);
339 	qm_cl_invalidate(EQCR_CI);
340 	pi = qm_in(EQCR_PI_CINH) & (QM_EQCR_SIZE - 1);
341 	eqcr->cursor = eqcr->ring + pi;
342 	eqcr->vbit = (qm_in(EQCR_PI_CINH) & QM_EQCR_SIZE) ?
343 			QM_EQCR_VERB_VBIT : 0;
344 	eqcr->available = QM_EQCR_SIZE - 1 -
345 			qm_cyc_diff(QM_EQCR_SIZE, eqcr->ci, pi);
346 	eqcr->ithresh = qm_in(EQCR_ITR);
347 #ifdef RTE_LIBRTE_DPAA_HWDEBUG
348 	eqcr->busy = 0;
349 	eqcr->pmode = pmode;
350 #endif
351 	cfg = (qm_in(CFG) & 0x00ffffff) |
352 		(eq_stash_thresh << 28) | /* QCSP_CFG: EST */
353 		(eq_stash_prio << 26)	| /* QCSP_CFG: EP */
354 		((pmode & 0x3) << 24);	/* QCSP_CFG::EPM */
355 	qm_out(CFG, cfg);
356 	return 0;
357 }
358 
359 static inline void qm_eqcr_finish(struct qm_portal *portal)
360 {
361 	register struct qm_eqcr *eqcr = &portal->eqcr;
362 	u8 pi, ci;
363 	u32 cfg;
364 
365 	/*
366 	 * Disable EQCI stashing because the QMan only
367 	 * presents the value it previously stashed to
368 	 * maintain coherency.  Setting the stash threshold
369 	 * to 1 then 0 ensures that QMan has resyncronized
370 	 * its internal copy so that the portal is clean
371 	 * when it is reinitialized in the future
372 	 */
373 	cfg = (qm_in(CFG) & 0x0fffffff) |
374 		(1 << 28); /* QCSP_CFG: EST */
375 	qm_out(CFG, cfg);
376 	cfg &= 0x0fffffff; /* stash threshold = 0 */
377 	qm_out(CFG, cfg);
378 
379 	pi = qm_in(EQCR_PI_CINH) & (QM_EQCR_SIZE - 1);
380 	ci = qm_in(EQCR_CI_CINH) & (QM_EQCR_SIZE - 1);
381 
382 	/* Refresh EQCR CI cache value */
383 	qm_cl_invalidate(EQCR_CI);
384 	eqcr->ci = qm_cl_in(EQCR_CI) & (QM_EQCR_SIZE - 1);
385 
386 #ifdef RTE_LIBRTE_DPAA_HWDEBUG
387 	DPAA_ASSERT(!eqcr->busy);
388 #endif
389 	if (pi != EQCR_PTR2IDX(eqcr->cursor))
390 		pr_crit("losing uncommitted EQCR entries\n");
391 	if (ci != eqcr->ci)
392 		pr_crit("missing existing EQCR completions\n");
393 	if (eqcr->ci != EQCR_PTR2IDX(eqcr->cursor))
394 		pr_crit("EQCR destroyed unquiesced\n");
395 }
396 
397 static inline int qm_dqrr_init(struct qm_portal *portal,
398 			__maybe_unused const struct qm_portal_config *config,
399 			enum qm_dqrr_dmode dmode,
400 			__maybe_unused enum qm_dqrr_pmode pmode,
401 			enum qm_dqrr_cmode cmode, u8 max_fill)
402 {
403 	register struct qm_dqrr *dqrr = &portal->dqrr;
404 	u32 cfg;
405 
406 	/* Make sure the DQRR will be idle when we enable */
407 	qm_out(DQRR_SDQCR, 0);
408 	qm_out(DQRR_VDQCR, 0);
409 	qm_out(DQRR_PDQCR, 0);
410 	dqrr->ring = portal->addr.ce + QM_CL_DQRR;
411 	dqrr->pi = qm_in(DQRR_PI_CINH) & (QM_DQRR_SIZE - 1);
412 	dqrr->ci = qm_in(DQRR_CI_CINH) & (QM_DQRR_SIZE - 1);
413 	dqrr->cursor = dqrr->ring + dqrr->ci;
414 	dqrr->fill = qm_cyc_diff(QM_DQRR_SIZE, dqrr->ci, dqrr->pi);
415 	dqrr->vbit = (qm_in(DQRR_PI_CINH) & QM_DQRR_SIZE) ?
416 			QM_DQRR_VERB_VBIT : 0;
417 	dqrr->ithresh = qm_in(DQRR_ITR);
418 #ifdef RTE_LIBRTE_DPAA_HWDEBUG
419 	dqrr->dmode = dmode;
420 	dqrr->pmode = pmode;
421 	dqrr->cmode = cmode;
422 #endif
423 	/* Invalidate every ring entry before beginning */
424 	for (cfg = 0; cfg < QM_DQRR_SIZE; cfg++)
425 		dccivac(qm_cl(dqrr->ring, cfg));
426 	cfg = (qm_in(CFG) & 0xff000f00) |
427 		((max_fill & (QM_DQRR_SIZE - 1)) << 20) | /* DQRR_MF */
428 		((dmode & 1) << 18) |			/* DP */
429 		((cmode & 3) << 16) |			/* DCM */
430 		0xa0 |					/* RE+SE */
431 		(0 ? 0x40 : 0) |			/* Ignore RP */
432 		(0 ? 0x10 : 0);				/* Ignore SP */
433 	qm_out(CFG, cfg);
434 	qm_dqrr_set_maxfill(portal, max_fill);
435 	return 0;
436 }
437 
438 static inline void qm_dqrr_finish(struct qm_portal *portal)
439 {
440 	__maybe_unused register struct qm_dqrr *dqrr = &portal->dqrr;
441 #ifdef RTE_LIBRTE_DPAA_HWDEBUG
442 	if ((dqrr->cmode != qm_dqrr_cdc) &&
443 	    (dqrr->ci != DQRR_PTR2IDX(dqrr->cursor)))
444 		pr_crit("Ignoring completed DQRR entries\n");
445 #endif
446 }
447 
448 static inline int qm_mr_init(struct qm_portal *portal,
449 			     __maybe_unused enum qm_mr_pmode pmode,
450 			     enum qm_mr_cmode cmode)
451 {
452 	register struct qm_mr *mr = &portal->mr;
453 	u32 cfg;
454 
455 	mr->ring = portal->addr.ce + QM_CL_MR;
456 	mr->pi = qm_in(MR_PI_CINH) & (QM_MR_SIZE - 1);
457 	mr->ci = qm_in(MR_CI_CINH) & (QM_MR_SIZE - 1);
458 	mr->cursor = mr->ring + mr->ci;
459 	mr->fill = qm_cyc_diff(QM_MR_SIZE, mr->ci, mr->pi);
460 	mr->vbit = (qm_in(MR_PI_CINH) & QM_MR_SIZE) ? QM_MR_VERB_VBIT : 0;
461 	mr->ithresh = qm_in(MR_ITR);
462 #ifdef RTE_LIBRTE_DPAA_HWDEBUG
463 	mr->pmode = pmode;
464 	mr->cmode = cmode;
465 #endif
466 	cfg = (qm_in(CFG) & 0xfffff0ff) |
467 		((cmode & 1) << 8);		/* QCSP_CFG:MM */
468 	qm_out(CFG, cfg);
469 	return 0;
470 }
471 
472 static inline void qm_mr_pvb_update(struct qm_portal *portal)
473 {
474 	register struct qm_mr *mr = &portal->mr;
475 	const struct qm_mr_entry *res = qm_cl(mr->ring, mr->pi);
476 
477 #ifdef RTE_LIBRTE_DPAA_HWDEBUG
478 	DPAA_ASSERT(mr->pmode == qm_mr_pvb);
479 #endif
480 	/* when accessing 'verb', use __raw_readb() to ensure that compiler
481 	 * inlining doesn't try to optimise out "excess reads".
482 	 */
483 	if ((__raw_readb(&res->verb) & QM_MR_VERB_VBIT) == mr->vbit) {
484 		mr->pi = (mr->pi + 1) & (QM_MR_SIZE - 1);
485 		if (!mr->pi)
486 			mr->vbit ^= QM_MR_VERB_VBIT;
487 		mr->fill++;
488 		res = MR_INC(res);
489 	}
490 	dcbit_ro(res);
491 }
492 
493 static inline
494 struct qman_portal *qman_create_portal(
495 			struct qman_portal *portal,
496 			      const struct qm_portal_config *c,
497 			      const struct qman_cgrs *cgrs)
498 {
499 	struct qm_portal *p;
500 	char buf[16];
501 	int ret;
502 	u32 isdr;
503 
504 	p = &portal->p;
505 
506 	portal->use_eqcr_ci_stashing = ((qman_ip_rev >= QMAN_REV30) ? 1 : 0);
507 	/*
508 	 * prep the low-level portal struct with the mapped addresses from the
509 	 * config, everything that follows depends on it and "config" is more
510 	 * for (de)reference
511 	 */
512 	p->addr.ce = c->addr_virt[DPAA_PORTAL_CE];
513 	p->addr.ci = c->addr_virt[DPAA_PORTAL_CI];
514 	/*
515 	 * If CI-stashing is used, the current defaults use a threshold of 3,
516 	 * and stash with high-than-DQRR priority.
517 	 */
518 	if (qm_eqcr_init(p, qm_eqcr_pvb,
519 			 portal->use_eqcr_ci_stashing ? 3 : 0, 1)) {
520 		pr_err("Qman EQCR initialisation failed\n");
521 		goto fail_eqcr;
522 	}
523 	if (qm_dqrr_init(p, c, qm_dqrr_dpush, qm_dqrr_pvb,
524 			 qm_dqrr_cdc, DQRR_MAXFILL)) {
525 		pr_err("Qman DQRR initialisation failed\n");
526 		goto fail_dqrr;
527 	}
528 	if (qm_mr_init(p, qm_mr_pvb, qm_mr_cci)) {
529 		pr_err("Qman MR initialisation failed\n");
530 		goto fail_mr;
531 	}
532 	if (qm_mc_init(p)) {
533 		pr_err("Qman MC initialisation failed\n");
534 		goto fail_mc;
535 	}
536 
537 	/* static interrupt-gating controls */
538 	qm_dqrr_set_ithresh(p, 0);
539 	qm_mr_set_ithresh(p, 0);
540 	qm_isr_set_iperiod(p, 0);
541 	portal->cgrs = kmalloc(2 * sizeof(*cgrs), GFP_KERNEL);
542 	if (!portal->cgrs)
543 		goto fail_cgrs;
544 	/* initial snapshot is no-depletion */
545 	qman_cgrs_init(&portal->cgrs[1]);
546 	if (cgrs)
547 		portal->cgrs[0] = *cgrs;
548 	else
549 		/* if the given mask is NULL, assume all CGRs can be seen */
550 		qman_cgrs_fill(&portal->cgrs[0]);
551 	INIT_LIST_HEAD(&portal->cgr_cbs);
552 	spin_lock_init(&portal->cgr_lock);
553 	portal->bits = 0;
554 	portal->slowpoll = 0;
555 	portal->sdqcr = QM_SDQCR_SOURCE_CHANNELS | QM_SDQCR_COUNT_UPTO3 |
556 			QM_SDQCR_DEDICATED_PRECEDENCE | QM_SDQCR_TYPE_PRIO_QOS |
557 			QM_SDQCR_TOKEN_SET(0xab) | QM_SDQCR_CHANNELS_DEDICATED;
558 	portal->dqrr_disable_ref = 0;
559 	portal->cb_dc_ern = NULL;
560 	sprintf(buf, "qportal-%d", c->channel);
561 	dpa_rbtree_init(&portal->retire_table);
562 	isdr = 0xffffffff;
563 	qm_isr_disable_write(p, isdr);
564 	portal->irq_sources = 0;
565 	qm_isr_enable_write(p, portal->irq_sources);
566 	qm_isr_status_clear(p, 0xffffffff);
567 	snprintf(portal->irqname, MAX_IRQNAME, IRQNAME, c->cpu);
568 	if (request_irq(c->irq, portal_isr, 0, portal->irqname,
569 			portal)) {
570 		pr_err("request_irq() failed\n");
571 		goto fail_irq;
572 	}
573 
574 	/* Need EQCR to be empty before continuing */
575 	isdr &= ~QM_PIRQ_EQCI;
576 	qm_isr_disable_write(p, isdr);
577 	ret = qm_eqcr_get_fill(p);
578 	if (ret) {
579 		pr_err("Qman EQCR unclean\n");
580 		goto fail_eqcr_empty;
581 	}
582 	isdr &= ~(QM_PIRQ_DQRI | QM_PIRQ_MRI);
583 	qm_isr_disable_write(p, isdr);
584 	if (qm_dqrr_current(p)) {
585 		pr_err("Qman DQRR unclean\n");
586 		qm_dqrr_cdc_consume_n(p, 0xffff);
587 	}
588 	if (qm_mr_current(p) && drain_mr_fqrni(p)) {
589 		/* special handling, drain just in case it's a few FQRNIs */
590 		if (drain_mr_fqrni(p))
591 			goto fail_dqrr_mr_empty;
592 	}
593 	/* Success */
594 	portal->config = c;
595 	qm_isr_disable_write(p, 0);
596 	qm_isr_uninhibit(p);
597 	/* Write a sane SDQCR */
598 	qm_dqrr_sdqcr_set(p, portal->sdqcr);
599 	return portal;
600 fail_dqrr_mr_empty:
601 fail_eqcr_empty:
602 	free_irq(c->irq, portal);
603 fail_irq:
604 	kfree(portal->cgrs);
605 	spin_lock_destroy(&portal->cgr_lock);
606 fail_cgrs:
607 	qm_mc_finish(p);
608 fail_mc:
609 	qm_mr_finish(p);
610 fail_mr:
611 	qm_dqrr_finish(p);
612 fail_dqrr:
613 	qm_eqcr_finish(p);
614 fail_eqcr:
615 	return NULL;
616 }
617 
618 struct qman_portal *qman_create_affine_portal(const struct qm_portal_config *c,
619 					      const struct qman_cgrs *cgrs)
620 {
621 	struct qman_portal *res;
622 	struct qman_portal *portal = get_affine_portal();
623 	/* A criteria for calling this function (from qman_driver.c) is that
624 	 * we're already affine to the cpu and won't schedule onto another cpu.
625 	 */
626 
627 	res = qman_create_portal(portal, c, cgrs);
628 	if (res) {
629 		spin_lock(&affine_mask_lock);
630 		CPU_SET(c->cpu, &affine_mask);
631 		affine_channels[c->cpu] =
632 			c->channel;
633 		spin_unlock(&affine_mask_lock);
634 	}
635 	return res;
636 }
637 
638 static inline
639 void qman_destroy_portal(struct qman_portal *qm)
640 {
641 	const struct qm_portal_config *pcfg;
642 
643 	/* Stop dequeues on the portal */
644 	qm_dqrr_sdqcr_set(&qm->p, 0);
645 
646 	/*
647 	 * NB we do this to "quiesce" EQCR. If we add enqueue-completions or
648 	 * something related to QM_PIRQ_EQCI, this may need fixing.
649 	 * Also, due to the prefetching model used for CI updates in the enqueue
650 	 * path, this update will only invalidate the CI cacheline *after*
651 	 * working on it, so we need to call this twice to ensure a full update
652 	 * irrespective of where the enqueue processing was at when the teardown
653 	 * began.
654 	 */
655 	qm_eqcr_cce_update(&qm->p);
656 	qm_eqcr_cce_update(&qm->p);
657 	pcfg = qm->config;
658 
659 	free_irq(pcfg->irq, qm);
660 
661 	kfree(qm->cgrs);
662 	qm_mc_finish(&qm->p);
663 	qm_mr_finish(&qm->p);
664 	qm_dqrr_finish(&qm->p);
665 	qm_eqcr_finish(&qm->p);
666 
667 	qm->config = NULL;
668 
669 	spin_lock_destroy(&qm->cgr_lock);
670 }
671 
672 const struct qm_portal_config *qman_destroy_affine_portal(void)
673 {
674 	/* We don't want to redirect if we're a slave, use "raw" */
675 	struct qman_portal *qm = get_affine_portal();
676 	const struct qm_portal_config *pcfg;
677 	int cpu;
678 
679 	pcfg = qm->config;
680 	cpu = pcfg->cpu;
681 
682 	qman_destroy_portal(qm);
683 
684 	spin_lock(&affine_mask_lock);
685 	CPU_CLR(cpu, &affine_mask);
686 	spin_unlock(&affine_mask_lock);
687 	return pcfg;
688 }
689 
690 int qman_get_portal_index(void)
691 {
692 	struct qman_portal *p = get_affine_portal();
693 	return p->config->index;
694 }
695 
696 /* Inline helper to reduce nesting in __poll_portal_slow() */
697 static inline void fq_state_change(struct qman_portal *p, struct qman_fq *fq,
698 				   const struct qm_mr_entry *msg, u8 verb)
699 {
700 	FQLOCK(fq);
701 	switch (verb) {
702 	case QM_MR_VERB_FQRL:
703 		DPAA_ASSERT(fq_isset(fq, QMAN_FQ_STATE_ORL));
704 		fq_clear(fq, QMAN_FQ_STATE_ORL);
705 		table_del_fq(p, fq);
706 		break;
707 	case QM_MR_VERB_FQRN:
708 		DPAA_ASSERT((fq->state == qman_fq_state_parked) ||
709 			    (fq->state == qman_fq_state_sched));
710 		DPAA_ASSERT(fq_isset(fq, QMAN_FQ_STATE_CHANGING));
711 		fq_clear(fq, QMAN_FQ_STATE_CHANGING);
712 		if (msg->fq.fqs & QM_MR_FQS_NOTEMPTY)
713 			fq_set(fq, QMAN_FQ_STATE_NE);
714 		if (msg->fq.fqs & QM_MR_FQS_ORLPRESENT)
715 			fq_set(fq, QMAN_FQ_STATE_ORL);
716 		else
717 			table_del_fq(p, fq);
718 		fq->state = qman_fq_state_retired;
719 		break;
720 	case QM_MR_VERB_FQPN:
721 		DPAA_ASSERT(fq->state == qman_fq_state_sched);
722 		DPAA_ASSERT(fq_isclear(fq, QMAN_FQ_STATE_CHANGING));
723 		fq->state = qman_fq_state_parked;
724 	}
725 	FQUNLOCK(fq);
726 }
727 
728 static u32 __poll_portal_slow(struct qman_portal *p, u32 is)
729 {
730 	const struct qm_mr_entry *msg;
731 	struct qm_mr_entry swapped_msg;
732 
733 	if (is & QM_PIRQ_CSCI) {
734 		struct qman_cgrs rr, c;
735 		struct qm_mc_result *mcr;
736 		struct qman_cgr *cgr;
737 
738 		spin_lock(&p->cgr_lock);
739 		/*
740 		 * The CSCI bit must be cleared _before_ issuing the
741 		 * Query Congestion State command, to ensure that a long
742 		 * CGR State Change callback cannot miss an intervening
743 		 * state change.
744 		 */
745 		qm_isr_status_clear(&p->p, QM_PIRQ_CSCI);
746 		qm_mc_start(&p->p);
747 		qm_mc_commit(&p->p, QM_MCC_VERB_QUERYCONGESTION);
748 		while (!(mcr = qm_mc_result(&p->p)))
749 			cpu_relax();
750 		/* mask out the ones I'm not interested in */
751 		qman_cgrs_and(&rr, (const struct qman_cgrs *)
752 			&mcr->querycongestion.state, &p->cgrs[0]);
753 		/* check previous snapshot for delta, enter/exit congestion */
754 		qman_cgrs_xor(&c, &rr, &p->cgrs[1]);
755 		/* update snapshot */
756 		qman_cgrs_cp(&p->cgrs[1], &rr);
757 		/* Invoke callback */
758 		list_for_each_entry(cgr, &p->cgr_cbs, node)
759 			if (cgr->cb && qman_cgrs_get(&c, cgr->cgrid))
760 				cgr->cb(p, cgr, qman_cgrs_get(&rr, cgr->cgrid));
761 		spin_unlock(&p->cgr_lock);
762 	}
763 
764 	if (is & QM_PIRQ_EQRI) {
765 		qm_eqcr_cce_update(&p->p);
766 		qm_eqcr_set_ithresh(&p->p, 0);
767 		wake_up(&affine_queue);
768 	}
769 
770 	if (is & QM_PIRQ_MRI) {
771 		struct qman_fq *fq;
772 		u8 verb, num = 0;
773 mr_loop:
774 		qm_mr_pvb_update(&p->p);
775 		msg = qm_mr_current(&p->p);
776 		if (!msg)
777 			goto mr_done;
778 		swapped_msg = *msg;
779 		hw_fd_to_cpu(&swapped_msg.ern.fd);
780 		verb = msg->verb & QM_MR_VERB_TYPE_MASK;
781 		/* The message is a software ERN iff the 0x20 bit is set */
782 		if (verb & 0x20) {
783 			switch (verb) {
784 			case QM_MR_VERB_FQRNI:
785 				/* nada, we drop FQRNIs on the floor */
786 				break;
787 			case QM_MR_VERB_FQRN:
788 			case QM_MR_VERB_FQRL:
789 				/* Lookup in the retirement table */
790 				fq = table_find_fq(p,
791 						   be32_to_cpu(msg->fq.fqid));
792 				DPAA_BUG_ON(!fq);
793 				fq_state_change(p, fq, &swapped_msg, verb);
794 				if (fq->cb.fqs)
795 					fq->cb.fqs(p, fq, &swapped_msg);
796 				break;
797 			case QM_MR_VERB_FQPN:
798 				/* Parked */
799 #ifdef CONFIG_FSL_QMAN_FQ_LOOKUP
800 				fq = get_fq_table_entry(
801 					be32_to_cpu(msg->fq.contextB));
802 #else
803 				fq = (void *)(uintptr_t)
804 					be32_to_cpu(msg->fq.contextB);
805 #endif
806 				fq_state_change(p, fq, msg, verb);
807 				if (fq->cb.fqs)
808 					fq->cb.fqs(p, fq, &swapped_msg);
809 				break;
810 			case QM_MR_VERB_DC_ERN:
811 				/* DCP ERN */
812 				if (p->cb_dc_ern)
813 					p->cb_dc_ern(p, msg);
814 				else if (cb_dc_ern)
815 					cb_dc_ern(p, msg);
816 				else {
817 					static int warn_once;
818 
819 					if (!warn_once) {
820 						pr_crit("Leaking DCP ERNs!\n");
821 						warn_once = 1;
822 					}
823 				}
824 				break;
825 			default:
826 				pr_crit("Invalid MR verb 0x%02x\n", verb);
827 			}
828 		} else {
829 			/* Its a software ERN */
830 #ifdef CONFIG_FSL_QMAN_FQ_LOOKUP
831 			fq = get_fq_table_entry(be32_to_cpu(msg->ern.tag));
832 #else
833 			fq = (void *)(uintptr_t)be32_to_cpu(msg->ern.tag);
834 #endif
835 			fq->cb.ern(p, fq, &swapped_msg);
836 		}
837 		num++;
838 		qm_mr_next(&p->p);
839 		goto mr_loop;
840 mr_done:
841 		qm_mr_cci_consume(&p->p, num);
842 	}
843 	/*
844 	 * QM_PIRQ_CSCI/CCSCI has already been cleared, as part of its specific
845 	 * processing. If that interrupt source has meanwhile been re-asserted,
846 	 * we mustn't clear it here (or in the top-level interrupt handler).
847 	 */
848 	return is & (QM_PIRQ_EQCI | QM_PIRQ_EQRI | QM_PIRQ_MRI);
849 }
850 
851 /*
852  * remove some slowish-path stuff from the "fast path" and make sure it isn't
853  * inlined.
854  */
855 static noinline void clear_vdqcr(struct qman_portal *p, struct qman_fq *fq)
856 {
857 	p->vdqcr_owned = NULL;
858 	FQLOCK(fq);
859 	fq_clear(fq, QMAN_FQ_STATE_VDQCR);
860 	FQUNLOCK(fq);
861 	wake_up(&affine_queue);
862 }
863 
864 /*
865  * The only states that would conflict with other things if they ran at the
866  * same time on the same cpu are:
867  *
868  *   (i) setting/clearing vdqcr_owned, and
869  *  (ii) clearing the NE (Not Empty) flag.
870  *
871  * Both are safe. Because;
872  *
873  *   (i) this clearing can only occur after qman_set_vdq() has set the
874  *	 vdqcr_owned field (which it does before setting VDQCR), and
875  *	 qman_volatile_dequeue() blocks interrupts and preemption while this is
876  *	 done so that we can't interfere.
877  *  (ii) the NE flag is only cleared after qman_retire_fq() has set it, and as
878  *	 with (i) that API prevents us from interfering until it's safe.
879  *
880  * The good thing is that qman_set_vdq() and qman_retire_fq() run far
881  * less frequently (ie. per-FQ) than __poll_portal_fast() does, so the nett
882  * advantage comes from this function not having to "lock" anything at all.
883  *
884  * Note also that the callbacks are invoked at points which are safe against the
885  * above potential conflicts, but that this function itself is not re-entrant
886  * (this is because the function tracks one end of each FIFO in the portal and
887  * we do *not* want to lock that). So the consequence is that it is safe for
888  * user callbacks to call into any QMan API.
889  */
890 static inline unsigned int __poll_portal_fast(struct qman_portal *p,
891 					      unsigned int poll_limit)
892 {
893 	const struct qm_dqrr_entry *dq;
894 	struct qman_fq *fq;
895 	enum qman_cb_dqrr_result res;
896 	unsigned int limit = 0;
897 #if __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__
898 	struct qm_dqrr_entry *shadow;
899 #endif
900 	do {
901 		qm_dqrr_pvb_update(&p->p);
902 		dq = qm_dqrr_current(&p->p);
903 		if (!dq)
904 			break;
905 #if __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__
906 	/* If running on an LE system the fields of the
907 	 * dequeue entry must be swapper.  Because the
908 	 * QMan HW will ignore writes the DQRR entry is
909 	 * copied and the index stored within the copy
910 	 */
911 		shadow = &p->shadow_dqrr[DQRR_PTR2IDX(dq)];
912 		*shadow = *dq;
913 		dq = shadow;
914 		shadow->fqid = be32_to_cpu(shadow->fqid);
915 		shadow->contextB = be32_to_cpu(shadow->contextB);
916 		shadow->seqnum = be16_to_cpu(shadow->seqnum);
917 		hw_fd_to_cpu(&shadow->fd);
918 #endif
919 
920 		if (dq->stat & QM_DQRR_STAT_UNSCHEDULED) {
921 			/*
922 			 * VDQCR: don't trust context_b as the FQ may have
923 			 * been configured for h/w consumption and we're
924 			 * draining it post-retirement.
925 			 */
926 			fq = p->vdqcr_owned;
927 			/*
928 			 * We only set QMAN_FQ_STATE_NE when retiring, so we
929 			 * only need to check for clearing it when doing
930 			 * volatile dequeues.  It's one less thing to check
931 			 * in the critical path (SDQCR).
932 			 */
933 			if (dq->stat & QM_DQRR_STAT_FQ_EMPTY)
934 				fq_clear(fq, QMAN_FQ_STATE_NE);
935 			/*
936 			 * This is duplicated from the SDQCR code, but we
937 			 * have stuff to do before *and* after this callback,
938 			 * and we don't want multiple if()s in the critical
939 			 * path (SDQCR).
940 			 */
941 			res = fq->cb.dqrr(p, fq, dq);
942 			if (res == qman_cb_dqrr_stop)
943 				break;
944 			/* Check for VDQCR completion */
945 			if (dq->stat & QM_DQRR_STAT_DQCR_EXPIRED)
946 				clear_vdqcr(p, fq);
947 		} else {
948 			/* SDQCR: context_b points to the FQ */
949 #ifdef CONFIG_FSL_QMAN_FQ_LOOKUP
950 			fq = get_fq_table_entry(dq->contextB);
951 #else
952 			fq = (void *)(uintptr_t)dq->contextB;
953 #endif
954 			/* Now let the callback do its stuff */
955 			res = fq->cb.dqrr(p, fq, dq);
956 			/*
957 			 * The callback can request that we exit without
958 			 * consuming this entry nor advancing;
959 			 */
960 			if (res == qman_cb_dqrr_stop)
961 				break;
962 		}
963 		/* Interpret 'dq' from a driver perspective. */
964 		/*
965 		 * Parking isn't possible unless HELDACTIVE was set. NB,
966 		 * FORCEELIGIBLE implies HELDACTIVE, so we only need to
967 		 * check for HELDACTIVE to cover both.
968 		 */
969 		DPAA_ASSERT((dq->stat & QM_DQRR_STAT_FQ_HELDACTIVE) ||
970 			    (res != qman_cb_dqrr_park));
971 		/* just means "skip it, I'll consume it myself later on" */
972 		if (res != qman_cb_dqrr_defer)
973 			qm_dqrr_cdc_consume_1ptr(&p->p, dq,
974 						 res == qman_cb_dqrr_park);
975 		/* Move forward */
976 		qm_dqrr_next(&p->p);
977 		/*
978 		 * Entry processed and consumed, increment our counter.  The
979 		 * callback can request that we exit after consuming the
980 		 * entry, and we also exit if we reach our processing limit,
981 		 * so loop back only if neither of these conditions is met.
982 		 */
983 	} while (++limit < poll_limit && res != qman_cb_dqrr_consume_stop);
984 
985 	return limit;
986 }
987 
988 u16 qman_affine_channel(int cpu)
989 {
990 	if (cpu < 0) {
991 		struct qman_portal *portal = get_affine_portal();
992 
993 		cpu = portal->config->cpu;
994 	}
995 	DPAA_BUG_ON(!CPU_ISSET(cpu, &affine_mask));
996 	return affine_channels[cpu];
997 }
998 
999 struct qm_dqrr_entry *qman_dequeue(struct qman_fq *fq)
1000 {
1001 	struct qman_portal *p = get_affine_portal();
1002 	const struct qm_dqrr_entry *dq;
1003 #if __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__
1004 	struct qm_dqrr_entry *shadow;
1005 #endif
1006 
1007 	qm_dqrr_pvb_update(&p->p);
1008 	dq = qm_dqrr_current(&p->p);
1009 	if (!dq)
1010 		return NULL;
1011 
1012 	if (!(dq->stat & QM_DQRR_STAT_FD_VALID)) {
1013 		/* Invalid DQRR - put the portal and consume the DQRR.
1014 		 * Return NULL to user as no packet is seen.
1015 		 */
1016 		qman_dqrr_consume(fq, (struct qm_dqrr_entry *)dq);
1017 		return NULL;
1018 	}
1019 
1020 #if __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__
1021 	shadow = &p->shadow_dqrr[DQRR_PTR2IDX(dq)];
1022 	*shadow = *dq;
1023 	dq = shadow;
1024 	shadow->fqid = be32_to_cpu(shadow->fqid);
1025 	shadow->contextB = be32_to_cpu(shadow->contextB);
1026 	shadow->seqnum = be16_to_cpu(shadow->seqnum);
1027 	hw_fd_to_cpu(&shadow->fd);
1028 #endif
1029 
1030 	if (dq->stat & QM_DQRR_STAT_FQ_EMPTY)
1031 		fq_clear(fq, QMAN_FQ_STATE_NE);
1032 
1033 	return (struct qm_dqrr_entry *)dq;
1034 }
1035 
1036 void qman_dqrr_consume(struct qman_fq *fq,
1037 		       struct qm_dqrr_entry *dq)
1038 {
1039 	struct qman_portal *p = get_affine_portal();
1040 
1041 	if (dq->stat & QM_DQRR_STAT_DQCR_EXPIRED)
1042 		clear_vdqcr(p, fq);
1043 
1044 	qm_dqrr_cdc_consume_1ptr(&p->p, dq, 0);
1045 	qm_dqrr_next(&p->p);
1046 }
1047 
1048 int qman_poll_dqrr(unsigned int limit)
1049 {
1050 	struct qman_portal *p = get_affine_portal();
1051 	int ret;
1052 
1053 	ret = __poll_portal_fast(p, limit);
1054 	return ret;
1055 }
1056 
1057 void qman_poll(void)
1058 {
1059 	struct qman_portal *p = get_affine_portal();
1060 
1061 	if ((~p->irq_sources) & QM_PIRQ_SLOW) {
1062 		if (!(p->slowpoll--)) {
1063 			u32 is = qm_isr_status_read(&p->p) & ~p->irq_sources;
1064 			u32 active = __poll_portal_slow(p, is);
1065 
1066 			if (active) {
1067 				qm_isr_status_clear(&p->p, active);
1068 				p->slowpoll = SLOW_POLL_BUSY;
1069 			} else
1070 				p->slowpoll = SLOW_POLL_IDLE;
1071 		}
1072 	}
1073 	if ((~p->irq_sources) & QM_PIRQ_DQRI)
1074 		__poll_portal_fast(p, FSL_QMAN_POLL_LIMIT);
1075 }
1076 
1077 void qman_stop_dequeues(void)
1078 {
1079 	struct qman_portal *p = get_affine_portal();
1080 
1081 	qman_stop_dequeues_ex(p);
1082 }
1083 
1084 void qman_start_dequeues(void)
1085 {
1086 	struct qman_portal *p = get_affine_portal();
1087 
1088 	DPAA_ASSERT(p->dqrr_disable_ref > 0);
1089 	if (!(--p->dqrr_disable_ref))
1090 		qm_dqrr_set_maxfill(&p->p, DQRR_MAXFILL);
1091 }
1092 
1093 void qman_static_dequeue_add(u32 pools)
1094 {
1095 	struct qman_portal *p = get_affine_portal();
1096 
1097 	pools &= p->config->pools;
1098 	p->sdqcr |= pools;
1099 	qm_dqrr_sdqcr_set(&p->p, p->sdqcr);
1100 }
1101 
1102 void qman_static_dequeue_del(u32 pools)
1103 {
1104 	struct qman_portal *p = get_affine_portal();
1105 
1106 	pools &= p->config->pools;
1107 	p->sdqcr &= ~pools;
1108 	qm_dqrr_sdqcr_set(&p->p, p->sdqcr);
1109 }
1110 
1111 u32 qman_static_dequeue_get(void)
1112 {
1113 	struct qman_portal *p = get_affine_portal();
1114 	return p->sdqcr;
1115 }
1116 
1117 void qman_dca(struct qm_dqrr_entry *dq, int park_request)
1118 {
1119 	struct qman_portal *p = get_affine_portal();
1120 
1121 	qm_dqrr_cdc_consume_1ptr(&p->p, dq, park_request);
1122 }
1123 
1124 /* Frame queue API */
1125 static const char *mcr_result_str(u8 result)
1126 {
1127 	switch (result) {
1128 	case QM_MCR_RESULT_NULL:
1129 		return "QM_MCR_RESULT_NULL";
1130 	case QM_MCR_RESULT_OK:
1131 		return "QM_MCR_RESULT_OK";
1132 	case QM_MCR_RESULT_ERR_FQID:
1133 		return "QM_MCR_RESULT_ERR_FQID";
1134 	case QM_MCR_RESULT_ERR_FQSTATE:
1135 		return "QM_MCR_RESULT_ERR_FQSTATE";
1136 	case QM_MCR_RESULT_ERR_NOTEMPTY:
1137 		return "QM_MCR_RESULT_ERR_NOTEMPTY";
1138 	case QM_MCR_RESULT_PENDING:
1139 		return "QM_MCR_RESULT_PENDING";
1140 	case QM_MCR_RESULT_ERR_BADCOMMAND:
1141 		return "QM_MCR_RESULT_ERR_BADCOMMAND";
1142 	}
1143 	return "<unknown MCR result>";
1144 }
1145 
1146 int qman_create_fq(u32 fqid, u32 flags, struct qman_fq *fq)
1147 {
1148 	struct qm_fqd fqd;
1149 	struct qm_mcr_queryfq_np np;
1150 	struct qm_mc_command *mcc;
1151 	struct qm_mc_result *mcr;
1152 	struct qman_portal *p;
1153 
1154 	if (flags & QMAN_FQ_FLAG_DYNAMIC_FQID) {
1155 		int ret = qman_alloc_fqid(&fqid);
1156 
1157 		if (ret)
1158 			return ret;
1159 	}
1160 	spin_lock_init(&fq->fqlock);
1161 	fq->fqid = fqid;
1162 	fq->flags = flags;
1163 	fq->state = qman_fq_state_oos;
1164 	fq->cgr_groupid = 0;
1165 #ifdef CONFIG_FSL_QMAN_FQ_LOOKUP
1166 	if (unlikely(find_empty_fq_table_entry(&fq->key, fq))) {
1167 		pr_info("Find empty table entry failed\n");
1168 		return -ENOMEM;
1169 	}
1170 #endif
1171 	if (!(flags & QMAN_FQ_FLAG_AS_IS) || (flags & QMAN_FQ_FLAG_NO_MODIFY))
1172 		return 0;
1173 	/* Everything else is AS_IS support */
1174 	p = get_affine_portal();
1175 	mcc = qm_mc_start(&p->p);
1176 	mcc->queryfq.fqid = cpu_to_be32(fqid);
1177 	qm_mc_commit(&p->p, QM_MCC_VERB_QUERYFQ);
1178 	while (!(mcr = qm_mc_result(&p->p)))
1179 		cpu_relax();
1180 	DPAA_ASSERT((mcr->verb & QM_MCR_VERB_MASK) == QM_MCC_VERB_QUERYFQ);
1181 	if (mcr->result != QM_MCR_RESULT_OK) {
1182 		pr_err("QUERYFQ failed: %s\n", mcr_result_str(mcr->result));
1183 		goto err;
1184 	}
1185 	fqd = mcr->queryfq.fqd;
1186 	hw_fqd_to_cpu(&fqd);
1187 	mcc = qm_mc_start(&p->p);
1188 	mcc->queryfq_np.fqid = cpu_to_be32(fqid);
1189 	qm_mc_commit(&p->p, QM_MCC_VERB_QUERYFQ_NP);
1190 	while (!(mcr = qm_mc_result(&p->p)))
1191 		cpu_relax();
1192 	DPAA_ASSERT((mcr->verb & QM_MCR_VERB_MASK) == QM_MCC_VERB_QUERYFQ_NP);
1193 	if (mcr->result != QM_MCR_RESULT_OK) {
1194 		pr_err("QUERYFQ_NP failed: %s\n", mcr_result_str(mcr->result));
1195 		goto err;
1196 	}
1197 	np = mcr->queryfq_np;
1198 	/* Phew, have queryfq and queryfq_np results, stitch together
1199 	 * the FQ object from those.
1200 	 */
1201 	fq->cgr_groupid = fqd.cgid;
1202 	switch (np.state & QM_MCR_NP_STATE_MASK) {
1203 	case QM_MCR_NP_STATE_OOS:
1204 		break;
1205 	case QM_MCR_NP_STATE_RETIRED:
1206 		fq->state = qman_fq_state_retired;
1207 		if (np.frm_cnt)
1208 			fq_set(fq, QMAN_FQ_STATE_NE);
1209 		break;
1210 	case QM_MCR_NP_STATE_TEN_SCHED:
1211 	case QM_MCR_NP_STATE_TRU_SCHED:
1212 	case QM_MCR_NP_STATE_ACTIVE:
1213 		fq->state = qman_fq_state_sched;
1214 		if (np.state & QM_MCR_NP_STATE_R)
1215 			fq_set(fq, QMAN_FQ_STATE_CHANGING);
1216 		break;
1217 	case QM_MCR_NP_STATE_PARKED:
1218 		fq->state = qman_fq_state_parked;
1219 		break;
1220 	default:
1221 		DPAA_ASSERT(NULL == "invalid FQ state");
1222 	}
1223 	if (fqd.fq_ctrl & QM_FQCTRL_CGE)
1224 		fq->state |= QMAN_FQ_STATE_CGR_EN;
1225 	return 0;
1226 err:
1227 	if (flags & QMAN_FQ_FLAG_DYNAMIC_FQID)
1228 		qman_release_fqid(fqid);
1229 	return -EIO;
1230 }
1231 
1232 void qman_destroy_fq(struct qman_fq *fq, u32 flags __maybe_unused)
1233 {
1234 	/*
1235 	 * We don't need to lock the FQ as it is a pre-condition that the FQ be
1236 	 * quiesced. Instead, run some checks.
1237 	 */
1238 	switch (fq->state) {
1239 	case qman_fq_state_parked:
1240 		DPAA_ASSERT(flags & QMAN_FQ_DESTROY_PARKED);
1241 		/* Fallthrough */
1242 	case qman_fq_state_oos:
1243 		if (fq_isset(fq, QMAN_FQ_FLAG_DYNAMIC_FQID))
1244 			qman_release_fqid(fq->fqid);
1245 #ifdef CONFIG_FSL_QMAN_FQ_LOOKUP
1246 		clear_fq_table_entry(fq->key);
1247 #endif
1248 		return;
1249 	default:
1250 		break;
1251 	}
1252 	DPAA_ASSERT(NULL == "qman_free_fq() on unquiesced FQ!");
1253 }
1254 
1255 u32 qman_fq_fqid(struct qman_fq *fq)
1256 {
1257 	return fq->fqid;
1258 }
1259 
1260 void qman_fq_state(struct qman_fq *fq, enum qman_fq_state *state, u32 *flags)
1261 {
1262 	if (state)
1263 		*state = fq->state;
1264 	if (flags)
1265 		*flags = fq->flags;
1266 }
1267 
1268 int qman_init_fq(struct qman_fq *fq, u32 flags, struct qm_mcc_initfq *opts)
1269 {
1270 	struct qm_mc_command *mcc;
1271 	struct qm_mc_result *mcr;
1272 	struct qman_portal *p;
1273 
1274 	u8 res, myverb = (flags & QMAN_INITFQ_FLAG_SCHED) ?
1275 		QM_MCC_VERB_INITFQ_SCHED : QM_MCC_VERB_INITFQ_PARKED;
1276 
1277 	if ((fq->state != qman_fq_state_oos) &&
1278 	    (fq->state != qman_fq_state_parked))
1279 		return -EINVAL;
1280 #ifdef RTE_LIBRTE_DPAA_HWDEBUG
1281 	if (unlikely(fq_isset(fq, QMAN_FQ_FLAG_NO_MODIFY)))
1282 		return -EINVAL;
1283 #endif
1284 	if (opts && (opts->we_mask & QM_INITFQ_WE_OAC)) {
1285 		/* And can't be set at the same time as TDTHRESH */
1286 		if (opts->we_mask & QM_INITFQ_WE_TDTHRESH)
1287 			return -EINVAL;
1288 	}
1289 	/* Issue an INITFQ_[PARKED|SCHED] management command */
1290 	p = get_affine_portal();
1291 	FQLOCK(fq);
1292 	if (unlikely((fq_isset(fq, QMAN_FQ_STATE_CHANGING)) ||
1293 		     ((fq->state != qman_fq_state_oos) &&
1294 				(fq->state != qman_fq_state_parked)))) {
1295 		FQUNLOCK(fq);
1296 		return -EBUSY;
1297 	}
1298 	mcc = qm_mc_start(&p->p);
1299 	if (opts)
1300 		mcc->initfq = *opts;
1301 	mcc->initfq.fqid = cpu_to_be32(fq->fqid);
1302 	mcc->initfq.count = 0;
1303 	/*
1304 	 * If the FQ does *not* have the TO_DCPORTAL flag, context_b is set as a
1305 	 * demux pointer. Otherwise, the caller-provided value is allowed to
1306 	 * stand, don't overwrite it.
1307 	 */
1308 	if (fq_isclear(fq, QMAN_FQ_FLAG_TO_DCPORTAL)) {
1309 		dma_addr_t phys_fq;
1310 
1311 		mcc->initfq.we_mask |= QM_INITFQ_WE_CONTEXTB;
1312 #ifdef CONFIG_FSL_QMAN_FQ_LOOKUP
1313 		mcc->initfq.fqd.context_b = fq->key;
1314 #else
1315 		mcc->initfq.fqd.context_b = (u32)(uintptr_t)fq;
1316 #endif
1317 		/*
1318 		 *  and the physical address - NB, if the user wasn't trying to
1319 		 * set CONTEXTA, clear the stashing settings.
1320 		 */
1321 		if (!(mcc->initfq.we_mask & QM_INITFQ_WE_CONTEXTA)) {
1322 			mcc->initfq.we_mask |= QM_INITFQ_WE_CONTEXTA;
1323 			memset(&mcc->initfq.fqd.context_a, 0,
1324 			       sizeof(mcc->initfq.fqd.context_a));
1325 		} else {
1326 			phys_fq = rte_mem_virt2iova(fq);
1327 			qm_fqd_stashing_set64(&mcc->initfq.fqd, phys_fq);
1328 		}
1329 	}
1330 	if (flags & QMAN_INITFQ_FLAG_LOCAL) {
1331 		mcc->initfq.fqd.dest.channel = p->config->channel;
1332 		if (!(mcc->initfq.we_mask & QM_INITFQ_WE_DESTWQ)) {
1333 			mcc->initfq.we_mask |= QM_INITFQ_WE_DESTWQ;
1334 			mcc->initfq.fqd.dest.wq = 4;
1335 		}
1336 	}
1337 	mcc->initfq.we_mask = cpu_to_be16(mcc->initfq.we_mask);
1338 	cpu_to_hw_fqd(&mcc->initfq.fqd);
1339 	qm_mc_commit(&p->p, myverb);
1340 	while (!(mcr = qm_mc_result(&p->p)))
1341 		cpu_relax();
1342 	DPAA_ASSERT((mcr->verb & QM_MCR_VERB_MASK) == myverb);
1343 	res = mcr->result;
1344 	if (res != QM_MCR_RESULT_OK) {
1345 		FQUNLOCK(fq);
1346 		return -EIO;
1347 	}
1348 	if (opts) {
1349 		if (opts->we_mask & QM_INITFQ_WE_FQCTRL) {
1350 			if (opts->fqd.fq_ctrl & QM_FQCTRL_CGE)
1351 				fq_set(fq, QMAN_FQ_STATE_CGR_EN);
1352 			else
1353 				fq_clear(fq, QMAN_FQ_STATE_CGR_EN);
1354 		}
1355 		if (opts->we_mask & QM_INITFQ_WE_CGID)
1356 			fq->cgr_groupid = opts->fqd.cgid;
1357 	}
1358 	fq->state = (flags & QMAN_INITFQ_FLAG_SCHED) ?
1359 		qman_fq_state_sched : qman_fq_state_parked;
1360 	FQUNLOCK(fq);
1361 	return 0;
1362 }
1363 
1364 int qman_schedule_fq(struct qman_fq *fq)
1365 {
1366 	struct qm_mc_command *mcc;
1367 	struct qm_mc_result *mcr;
1368 	struct qman_portal *p;
1369 
1370 	int ret = 0;
1371 	u8 res;
1372 
1373 	if (fq->state != qman_fq_state_parked)
1374 		return -EINVAL;
1375 #ifdef RTE_LIBRTE_DPAA_HWDEBUG
1376 	if (unlikely(fq_isset(fq, QMAN_FQ_FLAG_NO_MODIFY)))
1377 		return -EINVAL;
1378 #endif
1379 	/* Issue a ALTERFQ_SCHED management command */
1380 	p = get_affine_portal();
1381 
1382 	FQLOCK(fq);
1383 	if (unlikely((fq_isset(fq, QMAN_FQ_STATE_CHANGING)) ||
1384 		     (fq->state != qman_fq_state_parked))) {
1385 		ret = -EBUSY;
1386 		goto out;
1387 	}
1388 	mcc = qm_mc_start(&p->p);
1389 	mcc->alterfq.fqid = cpu_to_be32(fq->fqid);
1390 	qm_mc_commit(&p->p, QM_MCC_VERB_ALTER_SCHED);
1391 	while (!(mcr = qm_mc_result(&p->p)))
1392 		cpu_relax();
1393 	DPAA_ASSERT((mcr->verb & QM_MCR_VERB_MASK) == QM_MCR_VERB_ALTER_SCHED);
1394 	res = mcr->result;
1395 	if (res != QM_MCR_RESULT_OK) {
1396 		ret = -EIO;
1397 		goto out;
1398 	}
1399 	fq->state = qman_fq_state_sched;
1400 out:
1401 	FQUNLOCK(fq);
1402 
1403 	return ret;
1404 }
1405 
1406 int qman_retire_fq(struct qman_fq *fq, u32 *flags)
1407 {
1408 	struct qm_mc_command *mcc;
1409 	struct qm_mc_result *mcr;
1410 	struct qman_portal *p;
1411 
1412 	int rval;
1413 	u8 res;
1414 
1415 	if ((fq->state != qman_fq_state_parked) &&
1416 	    (fq->state != qman_fq_state_sched))
1417 		return -EINVAL;
1418 #ifdef RTE_LIBRTE_DPAA_HWDEBUG
1419 	if (unlikely(fq_isset(fq, QMAN_FQ_FLAG_NO_MODIFY)))
1420 		return -EINVAL;
1421 #endif
1422 	p = get_affine_portal();
1423 
1424 	FQLOCK(fq);
1425 	if (unlikely((fq_isset(fq, QMAN_FQ_STATE_CHANGING)) ||
1426 		     (fq->state == qman_fq_state_retired) ||
1427 				(fq->state == qman_fq_state_oos))) {
1428 		rval = -EBUSY;
1429 		goto out;
1430 	}
1431 	rval = table_push_fq(p, fq);
1432 	if (rval)
1433 		goto out;
1434 	mcc = qm_mc_start(&p->p);
1435 	mcc->alterfq.fqid = cpu_to_be32(fq->fqid);
1436 	qm_mc_commit(&p->p, QM_MCC_VERB_ALTER_RETIRE);
1437 	while (!(mcr = qm_mc_result(&p->p)))
1438 		cpu_relax();
1439 	DPAA_ASSERT((mcr->verb & QM_MCR_VERB_MASK) == QM_MCR_VERB_ALTER_RETIRE);
1440 	res = mcr->result;
1441 	/*
1442 	 * "Elegant" would be to treat OK/PENDING the same way; set CHANGING,
1443 	 * and defer the flags until FQRNI or FQRN (respectively) show up. But
1444 	 * "Friendly" is to process OK immediately, and not set CHANGING. We do
1445 	 * friendly, otherwise the caller doesn't necessarily have a fully
1446 	 * "retired" FQ on return even if the retirement was immediate. However
1447 	 * this does mean some code duplication between here and
1448 	 * fq_state_change().
1449 	 */
1450 	if (likely(res == QM_MCR_RESULT_OK)) {
1451 		rval = 0;
1452 		/* Process 'fq' right away, we'll ignore FQRNI */
1453 		if (mcr->alterfq.fqs & QM_MCR_FQS_NOTEMPTY)
1454 			fq_set(fq, QMAN_FQ_STATE_NE);
1455 		if (mcr->alterfq.fqs & QM_MCR_FQS_ORLPRESENT)
1456 			fq_set(fq, QMAN_FQ_STATE_ORL);
1457 		else
1458 			table_del_fq(p, fq);
1459 		if (flags)
1460 			*flags = fq->flags;
1461 		fq->state = qman_fq_state_retired;
1462 		if (fq->cb.fqs) {
1463 			/*
1464 			 * Another issue with supporting "immediate" retirement
1465 			 * is that we're forced to drop FQRNIs, because by the
1466 			 * time they're seen it may already be "too late" (the
1467 			 * fq may have been OOS'd and free()'d already). But if
1468 			 * the upper layer wants a callback whether it's
1469 			 * immediate or not, we have to fake a "MR" entry to
1470 			 * look like an FQRNI...
1471 			 */
1472 			struct qm_mr_entry msg;
1473 
1474 			msg.verb = QM_MR_VERB_FQRNI;
1475 			msg.fq.fqs = mcr->alterfq.fqs;
1476 			msg.fq.fqid = fq->fqid;
1477 #ifdef CONFIG_FSL_QMAN_FQ_LOOKUP
1478 			msg.fq.contextB = fq->key;
1479 #else
1480 			msg.fq.contextB = (u32)(uintptr_t)fq;
1481 #endif
1482 			fq->cb.fqs(p, fq, &msg);
1483 		}
1484 	} else if (res == QM_MCR_RESULT_PENDING) {
1485 		rval = 1;
1486 		fq_set(fq, QMAN_FQ_STATE_CHANGING);
1487 	} else {
1488 		rval = -EIO;
1489 		table_del_fq(p, fq);
1490 	}
1491 out:
1492 	FQUNLOCK(fq);
1493 	return rval;
1494 }
1495 
1496 int qman_oos_fq(struct qman_fq *fq)
1497 {
1498 	struct qm_mc_command *mcc;
1499 	struct qm_mc_result *mcr;
1500 	struct qman_portal *p;
1501 
1502 	int ret = 0;
1503 	u8 res;
1504 
1505 	if (fq->state != qman_fq_state_retired)
1506 		return -EINVAL;
1507 #ifdef RTE_LIBRTE_DPAA_HWDEBUG
1508 	if (unlikely(fq_isset(fq, QMAN_FQ_FLAG_NO_MODIFY)))
1509 		return -EINVAL;
1510 #endif
1511 	p = get_affine_portal();
1512 	FQLOCK(fq);
1513 	if (unlikely((fq_isset(fq, QMAN_FQ_STATE_BLOCKOOS)) ||
1514 		     (fq->state != qman_fq_state_retired))) {
1515 		ret = -EBUSY;
1516 		goto out;
1517 	}
1518 	mcc = qm_mc_start(&p->p);
1519 	mcc->alterfq.fqid = cpu_to_be32(fq->fqid);
1520 	qm_mc_commit(&p->p, QM_MCC_VERB_ALTER_OOS);
1521 	while (!(mcr = qm_mc_result(&p->p)))
1522 		cpu_relax();
1523 	DPAA_ASSERT((mcr->verb & QM_MCR_VERB_MASK) == QM_MCR_VERB_ALTER_OOS);
1524 	res = mcr->result;
1525 	if (res != QM_MCR_RESULT_OK) {
1526 		ret = -EIO;
1527 		goto out;
1528 	}
1529 	fq->state = qman_fq_state_oos;
1530 out:
1531 	FQUNLOCK(fq);
1532 	return ret;
1533 }
1534 
1535 int qman_fq_flow_control(struct qman_fq *fq, int xon)
1536 {
1537 	struct qm_mc_command *mcc;
1538 	struct qm_mc_result *mcr;
1539 	struct qman_portal *p;
1540 
1541 	int ret = 0;
1542 	u8 res;
1543 	u8 myverb;
1544 
1545 	if ((fq->state == qman_fq_state_oos) ||
1546 	    (fq->state == qman_fq_state_retired) ||
1547 		(fq->state == qman_fq_state_parked))
1548 		return -EINVAL;
1549 
1550 #ifdef RTE_LIBRTE_DPAA_HWDEBUG
1551 	if (unlikely(fq_isset(fq, QMAN_FQ_FLAG_NO_MODIFY)))
1552 		return -EINVAL;
1553 #endif
1554 	/* Issue a ALTER_FQXON or ALTER_FQXOFF management command */
1555 	p = get_affine_portal();
1556 	FQLOCK(fq);
1557 	if (unlikely((fq_isset(fq, QMAN_FQ_STATE_CHANGING)) ||
1558 		     (fq->state == qman_fq_state_parked) ||
1559 			(fq->state == qman_fq_state_oos) ||
1560 			(fq->state == qman_fq_state_retired))) {
1561 		ret = -EBUSY;
1562 		goto out;
1563 	}
1564 	mcc = qm_mc_start(&p->p);
1565 	mcc->alterfq.fqid = fq->fqid;
1566 	mcc->alterfq.count = 0;
1567 	myverb = xon ? QM_MCC_VERB_ALTER_FQXON : QM_MCC_VERB_ALTER_FQXOFF;
1568 
1569 	qm_mc_commit(&p->p, myverb);
1570 	while (!(mcr = qm_mc_result(&p->p)))
1571 		cpu_relax();
1572 	DPAA_ASSERT((mcr->verb & QM_MCR_VERB_MASK) == myverb);
1573 
1574 	res = mcr->result;
1575 	if (res != QM_MCR_RESULT_OK) {
1576 		ret = -EIO;
1577 		goto out;
1578 	}
1579 out:
1580 	FQUNLOCK(fq);
1581 	return ret;
1582 }
1583 
1584 int qman_query_fq(struct qman_fq *fq, struct qm_fqd *fqd)
1585 {
1586 	struct qm_mc_command *mcc;
1587 	struct qm_mc_result *mcr;
1588 	struct qman_portal *p = get_affine_portal();
1589 
1590 	u8 res;
1591 
1592 	mcc = qm_mc_start(&p->p);
1593 	mcc->queryfq.fqid = cpu_to_be32(fq->fqid);
1594 	qm_mc_commit(&p->p, QM_MCC_VERB_QUERYFQ);
1595 	while (!(mcr = qm_mc_result(&p->p)))
1596 		cpu_relax();
1597 	DPAA_ASSERT((mcr->verb & QM_MCR_VERB_MASK) == QM_MCR_VERB_QUERYFQ);
1598 	res = mcr->result;
1599 	if (res == QM_MCR_RESULT_OK)
1600 		*fqd = mcr->queryfq.fqd;
1601 	hw_fqd_to_cpu(fqd);
1602 	if (res != QM_MCR_RESULT_OK)
1603 		return -EIO;
1604 	return 0;
1605 }
1606 
1607 int qman_query_fq_has_pkts(struct qman_fq *fq)
1608 {
1609 	struct qm_mc_command *mcc;
1610 	struct qm_mc_result *mcr;
1611 	struct qman_portal *p = get_affine_portal();
1612 
1613 	int ret = 0;
1614 	u8 res;
1615 
1616 	mcc = qm_mc_start(&p->p);
1617 	mcc->queryfq.fqid = cpu_to_be32(fq->fqid);
1618 	qm_mc_commit(&p->p, QM_MCC_VERB_QUERYFQ_NP);
1619 	while (!(mcr = qm_mc_result(&p->p)))
1620 		cpu_relax();
1621 	res = mcr->result;
1622 	if (res == QM_MCR_RESULT_OK)
1623 		ret = !!mcr->queryfq_np.frm_cnt;
1624 	return ret;
1625 }
1626 
1627 int qman_query_fq_np(struct qman_fq *fq, struct qm_mcr_queryfq_np *np)
1628 {
1629 	struct qm_mc_command *mcc;
1630 	struct qm_mc_result *mcr;
1631 	struct qman_portal *p = get_affine_portal();
1632 
1633 	u8 res;
1634 
1635 	mcc = qm_mc_start(&p->p);
1636 	mcc->queryfq.fqid = cpu_to_be32(fq->fqid);
1637 	qm_mc_commit(&p->p, QM_MCC_VERB_QUERYFQ_NP);
1638 	while (!(mcr = qm_mc_result(&p->p)))
1639 		cpu_relax();
1640 	DPAA_ASSERT((mcr->verb & QM_MCR_VERB_MASK) == QM_MCR_VERB_QUERYFQ_NP);
1641 	res = mcr->result;
1642 	if (res == QM_MCR_RESULT_OK) {
1643 		*np = mcr->queryfq_np;
1644 		np->fqd_link = be24_to_cpu(np->fqd_link);
1645 		np->odp_seq = be16_to_cpu(np->odp_seq);
1646 		np->orp_nesn = be16_to_cpu(np->orp_nesn);
1647 		np->orp_ea_hseq  = be16_to_cpu(np->orp_ea_hseq);
1648 		np->orp_ea_tseq  = be16_to_cpu(np->orp_ea_tseq);
1649 		np->orp_ea_hptr = be24_to_cpu(np->orp_ea_hptr);
1650 		np->orp_ea_tptr = be24_to_cpu(np->orp_ea_tptr);
1651 		np->pfdr_hptr = be24_to_cpu(np->pfdr_hptr);
1652 		np->pfdr_tptr = be24_to_cpu(np->pfdr_tptr);
1653 		np->ics_surp = be16_to_cpu(np->ics_surp);
1654 		np->byte_cnt = be32_to_cpu(np->byte_cnt);
1655 		np->frm_cnt = be24_to_cpu(np->frm_cnt);
1656 		np->ra1_sfdr = be16_to_cpu(np->ra1_sfdr);
1657 		np->ra2_sfdr = be16_to_cpu(np->ra2_sfdr);
1658 		np->od1_sfdr = be16_to_cpu(np->od1_sfdr);
1659 		np->od2_sfdr = be16_to_cpu(np->od2_sfdr);
1660 		np->od3_sfdr = be16_to_cpu(np->od3_sfdr);
1661 	}
1662 	if (res == QM_MCR_RESULT_ERR_FQID)
1663 		return -ERANGE;
1664 	else if (res != QM_MCR_RESULT_OK)
1665 		return -EIO;
1666 	return 0;
1667 }
1668 
1669 int qman_query_wq(u8 query_dedicated, struct qm_mcr_querywq *wq)
1670 {
1671 	struct qm_mc_command *mcc;
1672 	struct qm_mc_result *mcr;
1673 	struct qman_portal *p = get_affine_portal();
1674 
1675 	u8 res, myverb;
1676 
1677 	myverb = (query_dedicated) ? QM_MCR_VERB_QUERYWQ_DEDICATED :
1678 				 QM_MCR_VERB_QUERYWQ;
1679 	mcc = qm_mc_start(&p->p);
1680 	mcc->querywq.channel.id = cpu_to_be16(wq->channel.id);
1681 	qm_mc_commit(&p->p, myverb);
1682 	while (!(mcr = qm_mc_result(&p->p)))
1683 		cpu_relax();
1684 	DPAA_ASSERT((mcr->verb & QM_MCR_VERB_MASK) == myverb);
1685 	res = mcr->result;
1686 	if (res == QM_MCR_RESULT_OK) {
1687 		int i, array_len;
1688 
1689 		wq->channel.id = be16_to_cpu(mcr->querywq.channel.id);
1690 		array_len = ARRAY_SIZE(mcr->querywq.wq_len);
1691 		for (i = 0; i < array_len; i++)
1692 			wq->wq_len[i] = be32_to_cpu(mcr->querywq.wq_len[i]);
1693 	}
1694 	if (res != QM_MCR_RESULT_OK) {
1695 		pr_err("QUERYWQ failed: %s\n", mcr_result_str(res));
1696 		return -EIO;
1697 	}
1698 	return 0;
1699 }
1700 
1701 int qman_testwrite_cgr(struct qman_cgr *cgr, u64 i_bcnt,
1702 		       struct qm_mcr_cgrtestwrite *result)
1703 {
1704 	struct qm_mc_command *mcc;
1705 	struct qm_mc_result *mcr;
1706 	struct qman_portal *p = get_affine_portal();
1707 
1708 	u8 res;
1709 
1710 	mcc = qm_mc_start(&p->p);
1711 	mcc->cgrtestwrite.cgid = cgr->cgrid;
1712 	mcc->cgrtestwrite.i_bcnt_hi = (u8)(i_bcnt >> 32);
1713 	mcc->cgrtestwrite.i_bcnt_lo = (u32)i_bcnt;
1714 	qm_mc_commit(&p->p, QM_MCC_VERB_CGRTESTWRITE);
1715 	while (!(mcr = qm_mc_result(&p->p)))
1716 		cpu_relax();
1717 	DPAA_ASSERT((mcr->verb & QM_MCR_VERB_MASK) == QM_MCC_VERB_CGRTESTWRITE);
1718 	res = mcr->result;
1719 	if (res == QM_MCR_RESULT_OK)
1720 		*result = mcr->cgrtestwrite;
1721 	if (res != QM_MCR_RESULT_OK) {
1722 		pr_err("CGR TEST WRITE failed: %s\n", mcr_result_str(res));
1723 		return -EIO;
1724 	}
1725 	return 0;
1726 }
1727 
1728 int qman_query_cgr(struct qman_cgr *cgr, struct qm_mcr_querycgr *cgrd)
1729 {
1730 	struct qm_mc_command *mcc;
1731 	struct qm_mc_result *mcr;
1732 	struct qman_portal *p = get_affine_portal();
1733 	u8 res;
1734 	unsigned int i;
1735 
1736 	mcc = qm_mc_start(&p->p);
1737 	mcc->querycgr.cgid = cgr->cgrid;
1738 	qm_mc_commit(&p->p, QM_MCC_VERB_QUERYCGR);
1739 	while (!(mcr = qm_mc_result(&p->p)))
1740 		cpu_relax();
1741 	DPAA_ASSERT((mcr->verb & QM_MCR_VERB_MASK) == QM_MCC_VERB_QUERYCGR);
1742 	res = mcr->result;
1743 	if (res == QM_MCR_RESULT_OK)
1744 		*cgrd = mcr->querycgr;
1745 	if (res != QM_MCR_RESULT_OK) {
1746 		pr_err("QUERY_CGR failed: %s\n", mcr_result_str(res));
1747 		return -EIO;
1748 	}
1749 	cgrd->cgr.wr_parm_g.word =
1750 		be32_to_cpu(cgrd->cgr.wr_parm_g.word);
1751 	cgrd->cgr.wr_parm_y.word =
1752 		be32_to_cpu(cgrd->cgr.wr_parm_y.word);
1753 	cgrd->cgr.wr_parm_r.word =
1754 		be32_to_cpu(cgrd->cgr.wr_parm_r.word);
1755 	cgrd->cgr.cscn_targ =  be32_to_cpu(cgrd->cgr.cscn_targ);
1756 	cgrd->cgr.__cs_thres = be16_to_cpu(cgrd->cgr.__cs_thres);
1757 	for (i = 0; i < ARRAY_SIZE(cgrd->cscn_targ_swp); i++)
1758 		cgrd->cscn_targ_swp[i] =
1759 			be32_to_cpu(cgrd->cscn_targ_swp[i]);
1760 	return 0;
1761 }
1762 
1763 int qman_query_congestion(struct qm_mcr_querycongestion *congestion)
1764 {
1765 	struct qm_mc_result *mcr;
1766 	struct qman_portal *p = get_affine_portal();
1767 	u8 res;
1768 	unsigned int i;
1769 
1770 	qm_mc_start(&p->p);
1771 	qm_mc_commit(&p->p, QM_MCC_VERB_QUERYCONGESTION);
1772 	while (!(mcr = qm_mc_result(&p->p)))
1773 		cpu_relax();
1774 	DPAA_ASSERT((mcr->verb & QM_MCR_VERB_MASK) ==
1775 			QM_MCC_VERB_QUERYCONGESTION);
1776 	res = mcr->result;
1777 	if (res == QM_MCR_RESULT_OK)
1778 		*congestion = mcr->querycongestion;
1779 	if (res != QM_MCR_RESULT_OK) {
1780 		pr_err("QUERY_CONGESTION failed: %s\n", mcr_result_str(res));
1781 		return -EIO;
1782 	}
1783 	for (i = 0; i < ARRAY_SIZE(congestion->state.state); i++)
1784 		congestion->state.state[i] =
1785 			be32_to_cpu(congestion->state.state[i]);
1786 	return 0;
1787 }
1788 
1789 int qman_set_vdq(struct qman_fq *fq, u16 num)
1790 {
1791 	struct qman_portal *p = get_affine_portal();
1792 	uint32_t vdqcr;
1793 	int ret = -EBUSY;
1794 
1795 	vdqcr = QM_VDQCR_EXACT;
1796 	vdqcr |= QM_VDQCR_NUMFRAMES_SET(num);
1797 
1798 	if ((fq->state != qman_fq_state_parked) &&
1799 	    (fq->state != qman_fq_state_retired)) {
1800 		ret = -EINVAL;
1801 		goto out;
1802 	}
1803 	if (fq_isset(fq, QMAN_FQ_STATE_VDQCR)) {
1804 		ret = -EBUSY;
1805 		goto out;
1806 	}
1807 	vdqcr = (vdqcr & ~QM_VDQCR_FQID_MASK) | fq->fqid;
1808 
1809 	if (!p->vdqcr_owned) {
1810 		FQLOCK(fq);
1811 		if (fq_isset(fq, QMAN_FQ_STATE_VDQCR))
1812 			goto escape;
1813 		fq_set(fq, QMAN_FQ_STATE_VDQCR);
1814 		FQUNLOCK(fq);
1815 		p->vdqcr_owned = fq;
1816 		ret = 0;
1817 	}
1818 escape:
1819 	if (!ret)
1820 		qm_dqrr_vdqcr_set(&p->p, vdqcr);
1821 
1822 out:
1823 	return ret;
1824 }
1825 
1826 int qman_volatile_dequeue(struct qman_fq *fq, u32 flags __maybe_unused,
1827 			  u32 vdqcr)
1828 {
1829 	struct qman_portal *p;
1830 	int ret = -EBUSY;
1831 
1832 	if ((fq->state != qman_fq_state_parked) &&
1833 	    (fq->state != qman_fq_state_retired))
1834 		return -EINVAL;
1835 	if (vdqcr & QM_VDQCR_FQID_MASK)
1836 		return -EINVAL;
1837 	if (fq_isset(fq, QMAN_FQ_STATE_VDQCR))
1838 		return -EBUSY;
1839 	vdqcr = (vdqcr & ~QM_VDQCR_FQID_MASK) | fq->fqid;
1840 
1841 	p = get_affine_portal();
1842 
1843 	if (!p->vdqcr_owned) {
1844 		FQLOCK(fq);
1845 		if (fq_isset(fq, QMAN_FQ_STATE_VDQCR))
1846 			goto escape;
1847 		fq_set(fq, QMAN_FQ_STATE_VDQCR);
1848 		FQUNLOCK(fq);
1849 		p->vdqcr_owned = fq;
1850 		ret = 0;
1851 	}
1852 escape:
1853 	if (ret)
1854 		return ret;
1855 
1856 	/* VDQCR is set */
1857 	qm_dqrr_vdqcr_set(&p->p, vdqcr);
1858 	return 0;
1859 }
1860 
1861 static noinline void update_eqcr_ci(struct qman_portal *p, u8 avail)
1862 {
1863 	if (avail)
1864 		qm_eqcr_cce_prefetch(&p->p);
1865 	else
1866 		qm_eqcr_cce_update(&p->p);
1867 }
1868 
1869 int qman_eqcr_is_empty(void)
1870 {
1871 	struct qman_portal *p = get_affine_portal();
1872 	u8 avail;
1873 
1874 	update_eqcr_ci(p, 0);
1875 	avail = qm_eqcr_get_fill(&p->p);
1876 	return (avail == 0);
1877 }
1878 
1879 void qman_set_dc_ern(qman_cb_dc_ern handler, int affine)
1880 {
1881 	if (affine) {
1882 		struct qman_portal *p = get_affine_portal();
1883 
1884 		p->cb_dc_ern = handler;
1885 	} else
1886 		cb_dc_ern = handler;
1887 }
1888 
1889 static inline struct qm_eqcr_entry *try_p_eq_start(struct qman_portal *p,
1890 					struct qman_fq *fq,
1891 					const struct qm_fd *fd,
1892 					u32 flags)
1893 {
1894 	struct qm_eqcr_entry *eq;
1895 	u8 avail;
1896 
1897 	if (p->use_eqcr_ci_stashing) {
1898 		/*
1899 		 * The stashing case is easy, only update if we need to in
1900 		 * order to try and liberate ring entries.
1901 		 */
1902 		eq = qm_eqcr_start_stash(&p->p);
1903 	} else {
1904 		/*
1905 		 * The non-stashing case is harder, need to prefetch ahead of
1906 		 * time.
1907 		 */
1908 		avail = qm_eqcr_get_avail(&p->p);
1909 		if (avail < 2)
1910 			update_eqcr_ci(p, avail);
1911 		eq = qm_eqcr_start_no_stash(&p->p);
1912 	}
1913 
1914 	if (unlikely(!eq))
1915 		return NULL;
1916 
1917 	if (flags & QMAN_ENQUEUE_FLAG_DCA)
1918 		eq->dca = QM_EQCR_DCA_ENABLE |
1919 			((flags & QMAN_ENQUEUE_FLAG_DCA_PARK) ?
1920 					QM_EQCR_DCA_PARK : 0) |
1921 			((flags >> 8) & QM_EQCR_DCA_IDXMASK);
1922 	eq->fqid = cpu_to_be32(fq->fqid);
1923 #ifdef CONFIG_FSL_QMAN_FQ_LOOKUP
1924 	eq->tag = cpu_to_be32(fq->key);
1925 #else
1926 	eq->tag = cpu_to_be32((u32)(uintptr_t)fq);
1927 #endif
1928 	eq->fd = *fd;
1929 	cpu_to_hw_fd(&eq->fd);
1930 	return eq;
1931 }
1932 
1933 int qman_enqueue(struct qman_fq *fq, const struct qm_fd *fd, u32 flags)
1934 {
1935 	struct qman_portal *p = get_affine_portal();
1936 	struct qm_eqcr_entry *eq;
1937 
1938 	eq = try_p_eq_start(p, fq, fd, flags);
1939 	if (!eq)
1940 		return -EBUSY;
1941 	/* Note: QM_EQCR_VERB_INTERRUPT == QMAN_ENQUEUE_FLAG_WAIT_SYNC */
1942 	qm_eqcr_pvb_commit(&p->p, QM_EQCR_VERB_CMD_ENQUEUE |
1943 		(flags & (QM_EQCR_VERB_COLOUR_MASK | QM_EQCR_VERB_INTERRUPT)));
1944 	/* Factor the below out, it's used from qman_enqueue_orp() too */
1945 	return 0;
1946 }
1947 
1948 int qman_enqueue_multi(struct qman_fq *fq,
1949 		       const struct qm_fd *fd,
1950 		int frames_to_send)
1951 {
1952 	struct qman_portal *p = get_affine_portal();
1953 	struct qm_portal *portal = &p->p;
1954 
1955 	register struct qm_eqcr *eqcr = &portal->eqcr;
1956 	struct qm_eqcr_entry *eq = eqcr->cursor, *prev_eq;
1957 
1958 	u8 i, diff, old_ci, sent = 0;
1959 
1960 	/* Update the available entries if no entry is free */
1961 	if (!eqcr->available) {
1962 		old_ci = eqcr->ci;
1963 		eqcr->ci = qm_cl_in(EQCR_CI) & (QM_EQCR_SIZE - 1);
1964 		diff = qm_cyc_diff(QM_EQCR_SIZE, old_ci, eqcr->ci);
1965 		eqcr->available += diff;
1966 		if (!diff)
1967 			return 0;
1968 	}
1969 
1970 	/* try to send as many frames as possible */
1971 	while (eqcr->available && frames_to_send--) {
1972 		eq->fqid = cpu_to_be32(fq->fqid);
1973 #ifdef CONFIG_FSL_QMAN_FQ_LOOKUP
1974 		eq->tag = cpu_to_be32(fq->key);
1975 #else
1976 		eq->tag = cpu_to_be32((u32)(uintptr_t)fq);
1977 #endif
1978 		eq->fd.opaque_addr = fd->opaque_addr;
1979 		eq->fd.addr = cpu_to_be40(fd->addr);
1980 		eq->fd.status = cpu_to_be32(fd->status);
1981 		eq->fd.opaque = cpu_to_be32(fd->opaque);
1982 
1983 		eq = (void *)((unsigned long)(eq + 1) &
1984 			(~(unsigned long)(QM_EQCR_SIZE << 6)));
1985 		eqcr->available--;
1986 		sent++;
1987 		fd++;
1988 	}
1989 	lwsync();
1990 
1991 	/* In order for flushes to complete faster, all lines are recorded in
1992 	 * 32 bit word.
1993 	 */
1994 	eq = eqcr->cursor;
1995 	for (i = 0; i < sent; i++) {
1996 		eq->__dont_write_directly__verb =
1997 			QM_EQCR_VERB_CMD_ENQUEUE | eqcr->vbit;
1998 		prev_eq = eq;
1999 		eq = (void *)((unsigned long)(eq + 1) &
2000 			(~(unsigned long)(QM_EQCR_SIZE << 6)));
2001 		if (unlikely((prev_eq + 1) != eq))
2002 			eqcr->vbit ^= QM_EQCR_VERB_VBIT;
2003 	}
2004 
2005 	/* We need  to flush all the lines but without load/store operations
2006 	 * between them
2007 	 */
2008 	eq = eqcr->cursor;
2009 	for (i = 0; i < sent; i++) {
2010 		dcbf(eq);
2011 		eq = (void *)((unsigned long)(eq + 1) &
2012 			(~(unsigned long)(QM_EQCR_SIZE << 6)));
2013 	}
2014 	/* Update cursor for the next call */
2015 	eqcr->cursor = eq;
2016 	return sent;
2017 }
2018 
2019 int qman_enqueue_orp(struct qman_fq *fq, const struct qm_fd *fd, u32 flags,
2020 		     struct qman_fq *orp, u16 orp_seqnum)
2021 {
2022 	struct qman_portal *p  = get_affine_portal();
2023 	struct qm_eqcr_entry *eq;
2024 
2025 	eq = try_p_eq_start(p, fq, fd, flags);
2026 	if (!eq)
2027 		return -EBUSY;
2028 	/* Process ORP-specifics here */
2029 	if (flags & QMAN_ENQUEUE_FLAG_NLIS)
2030 		orp_seqnum |= QM_EQCR_SEQNUM_NLIS;
2031 	else {
2032 		orp_seqnum &= ~QM_EQCR_SEQNUM_NLIS;
2033 		if (flags & QMAN_ENQUEUE_FLAG_NESN)
2034 			orp_seqnum |= QM_EQCR_SEQNUM_NESN;
2035 		else
2036 			/* No need to check 4 QMAN_ENQUEUE_FLAG_HOLE */
2037 			orp_seqnum &= ~QM_EQCR_SEQNUM_NESN;
2038 	}
2039 	eq->seqnum = cpu_to_be16(orp_seqnum);
2040 	eq->orp = cpu_to_be32(orp->fqid);
2041 	/* Note: QM_EQCR_VERB_INTERRUPT == QMAN_ENQUEUE_FLAG_WAIT_SYNC */
2042 	qm_eqcr_pvb_commit(&p->p, QM_EQCR_VERB_ORP |
2043 		((flags & (QMAN_ENQUEUE_FLAG_HOLE | QMAN_ENQUEUE_FLAG_NESN)) ?
2044 				0 : QM_EQCR_VERB_CMD_ENQUEUE) |
2045 		(flags & (QM_EQCR_VERB_COLOUR_MASK | QM_EQCR_VERB_INTERRUPT)));
2046 
2047 	return 0;
2048 }
2049 
2050 int qman_modify_cgr(struct qman_cgr *cgr, u32 flags,
2051 		    struct qm_mcc_initcgr *opts)
2052 {
2053 	struct qm_mc_command *mcc;
2054 	struct qm_mc_result *mcr;
2055 	struct qman_portal *p = get_affine_portal();
2056 
2057 	u8 res;
2058 	u8 verb = QM_MCC_VERB_MODIFYCGR;
2059 
2060 	mcc = qm_mc_start(&p->p);
2061 	if (opts)
2062 		mcc->initcgr = *opts;
2063 	mcc->initcgr.we_mask = cpu_to_be16(mcc->initcgr.we_mask);
2064 	mcc->initcgr.cgr.wr_parm_g.word =
2065 		cpu_to_be32(mcc->initcgr.cgr.wr_parm_g.word);
2066 	mcc->initcgr.cgr.wr_parm_y.word =
2067 		cpu_to_be32(mcc->initcgr.cgr.wr_parm_y.word);
2068 	mcc->initcgr.cgr.wr_parm_r.word =
2069 		cpu_to_be32(mcc->initcgr.cgr.wr_parm_r.word);
2070 	mcc->initcgr.cgr.cscn_targ =  cpu_to_be32(mcc->initcgr.cgr.cscn_targ);
2071 	mcc->initcgr.cgr.__cs_thres = cpu_to_be16(mcc->initcgr.cgr.__cs_thres);
2072 
2073 	mcc->initcgr.cgid = cgr->cgrid;
2074 	if (flags & QMAN_CGR_FLAG_USE_INIT)
2075 		verb = QM_MCC_VERB_INITCGR;
2076 	qm_mc_commit(&p->p, verb);
2077 	while (!(mcr = qm_mc_result(&p->p)))
2078 		cpu_relax();
2079 
2080 	DPAA_ASSERT((mcr->verb & QM_MCR_VERB_MASK) == verb);
2081 	res = mcr->result;
2082 	return (res == QM_MCR_RESULT_OK) ? 0 : -EIO;
2083 }
2084 
2085 #define TARG_MASK(n) (0x80000000 >> (n->config->channel - \
2086 					QM_CHANNEL_SWPORTAL0))
2087 #define TARG_DCP_MASK(n) (0x80000000 >> (10 + n))
2088 #define PORTAL_IDX(n) (n->config->channel - QM_CHANNEL_SWPORTAL0)
2089 
2090 int qman_create_cgr(struct qman_cgr *cgr, u32 flags,
2091 		    struct qm_mcc_initcgr *opts)
2092 {
2093 	struct qm_mcr_querycgr cgr_state;
2094 	struct qm_mcc_initcgr local_opts;
2095 	int ret;
2096 	struct qman_portal *p;
2097 
2098 	/* We have to check that the provided CGRID is within the limits of the
2099 	 * data-structures, for obvious reasons. However we'll let h/w take
2100 	 * care of determining whether it's within the limits of what exists on
2101 	 * the SoC.
2102 	 */
2103 	if (cgr->cgrid >= __CGR_NUM)
2104 		return -EINVAL;
2105 
2106 	p = get_affine_portal();
2107 
2108 	memset(&local_opts, 0, sizeof(struct qm_mcc_initcgr));
2109 	cgr->chan = p->config->channel;
2110 	spin_lock(&p->cgr_lock);
2111 
2112 	/* if no opts specified, just add it to the list */
2113 	if (!opts)
2114 		goto add_list;
2115 
2116 	ret = qman_query_cgr(cgr, &cgr_state);
2117 	if (ret)
2118 		goto release_lock;
2119 	if (opts)
2120 		local_opts = *opts;
2121 	if ((qman_ip_rev & 0xFF00) >= QMAN_REV30)
2122 		local_opts.cgr.cscn_targ_upd_ctrl =
2123 			QM_CGR_TARG_UDP_CTRL_WRITE_BIT | PORTAL_IDX(p);
2124 	else
2125 		/* Overwrite TARG */
2126 		local_opts.cgr.cscn_targ = cgr_state.cgr.cscn_targ |
2127 							TARG_MASK(p);
2128 	local_opts.we_mask |= QM_CGR_WE_CSCN_TARG;
2129 
2130 	/* send init if flags indicate so */
2131 	if (opts && (flags & QMAN_CGR_FLAG_USE_INIT))
2132 		ret = qman_modify_cgr(cgr, QMAN_CGR_FLAG_USE_INIT, &local_opts);
2133 	else
2134 		ret = qman_modify_cgr(cgr, 0, &local_opts);
2135 	if (ret)
2136 		goto release_lock;
2137 add_list:
2138 	list_add(&cgr->node, &p->cgr_cbs);
2139 
2140 	/* Determine if newly added object requires its callback to be called */
2141 	ret = qman_query_cgr(cgr, &cgr_state);
2142 	if (ret) {
2143 		/* we can't go back, so proceed and return success, but screen
2144 		 * and wail to the log file.
2145 		 */
2146 		pr_crit("CGR HW state partially modified\n");
2147 		ret = 0;
2148 		goto release_lock;
2149 	}
2150 	if (cgr->cb && cgr_state.cgr.cscn_en && qman_cgrs_get(&p->cgrs[1],
2151 							      cgr->cgrid))
2152 		cgr->cb(p, cgr, 1);
2153 release_lock:
2154 	spin_unlock(&p->cgr_lock);
2155 	return ret;
2156 }
2157 
2158 int qman_create_cgr_to_dcp(struct qman_cgr *cgr, u32 flags, u16 dcp_portal,
2159 			   struct qm_mcc_initcgr *opts)
2160 {
2161 	struct qm_mcc_initcgr local_opts;
2162 	struct qm_mcr_querycgr cgr_state;
2163 	int ret;
2164 
2165 	if ((qman_ip_rev & 0xFF00) < QMAN_REV30) {
2166 		pr_warn("QMan version doesn't support CSCN => DCP portal\n");
2167 		return -EINVAL;
2168 	}
2169 	/* We have to check that the provided CGRID is within the limits of the
2170 	 * data-structures, for obvious reasons. However we'll let h/w take
2171 	 * care of determining whether it's within the limits of what exists on
2172 	 * the SoC.
2173 	 */
2174 	if (cgr->cgrid >= __CGR_NUM)
2175 		return -EINVAL;
2176 
2177 	ret = qman_query_cgr(cgr, &cgr_state);
2178 	if (ret)
2179 		return ret;
2180 
2181 	memset(&local_opts, 0, sizeof(struct qm_mcc_initcgr));
2182 	if (opts)
2183 		local_opts = *opts;
2184 
2185 	if ((qman_ip_rev & 0xFF00) >= QMAN_REV30)
2186 		local_opts.cgr.cscn_targ_upd_ctrl =
2187 				QM_CGR_TARG_UDP_CTRL_WRITE_BIT |
2188 				QM_CGR_TARG_UDP_CTRL_DCP | dcp_portal;
2189 	else
2190 		local_opts.cgr.cscn_targ = cgr_state.cgr.cscn_targ |
2191 					TARG_DCP_MASK(dcp_portal);
2192 	local_opts.we_mask |= QM_CGR_WE_CSCN_TARG;
2193 
2194 	/* send init if flags indicate so */
2195 	if (opts && (flags & QMAN_CGR_FLAG_USE_INIT))
2196 		ret = qman_modify_cgr(cgr, QMAN_CGR_FLAG_USE_INIT,
2197 				      &local_opts);
2198 	else
2199 		ret = qman_modify_cgr(cgr, 0, &local_opts);
2200 
2201 	return ret;
2202 }
2203 
2204 int qman_delete_cgr(struct qman_cgr *cgr)
2205 {
2206 	struct qm_mcr_querycgr cgr_state;
2207 	struct qm_mcc_initcgr local_opts;
2208 	int ret = 0;
2209 	struct qman_cgr *i;
2210 	struct qman_portal *p = get_affine_portal();
2211 
2212 	if (cgr->chan != p->config->channel) {
2213 		pr_crit("Attempting to delete cgr from different portal than"
2214 			" it was create: create 0x%x, delete 0x%x\n",
2215 			cgr->chan, p->config->channel);
2216 		ret = -EINVAL;
2217 		goto put_portal;
2218 	}
2219 	memset(&local_opts, 0, sizeof(struct qm_mcc_initcgr));
2220 	spin_lock(&p->cgr_lock);
2221 	list_del(&cgr->node);
2222 	/*
2223 	 * If there are no other CGR objects for this CGRID in the list,
2224 	 * update CSCN_TARG accordingly
2225 	 */
2226 	list_for_each_entry(i, &p->cgr_cbs, node)
2227 		if ((i->cgrid == cgr->cgrid) && i->cb)
2228 			goto release_lock;
2229 	ret = qman_query_cgr(cgr, &cgr_state);
2230 	if (ret)  {
2231 		/* add back to the list */
2232 		list_add(&cgr->node, &p->cgr_cbs);
2233 		goto release_lock;
2234 	}
2235 	/* Overwrite TARG */
2236 	local_opts.we_mask = QM_CGR_WE_CSCN_TARG;
2237 	if ((qman_ip_rev & 0xFF00) >= QMAN_REV30)
2238 		local_opts.cgr.cscn_targ_upd_ctrl = PORTAL_IDX(p);
2239 	else
2240 		local_opts.cgr.cscn_targ = cgr_state.cgr.cscn_targ &
2241 							 ~(TARG_MASK(p));
2242 	ret = qman_modify_cgr(cgr, 0, &local_opts);
2243 	if (ret)
2244 		/* add back to the list */
2245 		list_add(&cgr->node, &p->cgr_cbs);
2246 release_lock:
2247 	spin_unlock(&p->cgr_lock);
2248 put_portal:
2249 	return ret;
2250 }
2251 
2252 int qman_shutdown_fq(u32 fqid)
2253 {
2254 	struct qman_portal *p;
2255 	struct qm_portal *low_p;
2256 	struct qm_mc_command *mcc;
2257 	struct qm_mc_result *mcr;
2258 	u8 state;
2259 	int orl_empty, fq_empty, drain = 0;
2260 	u32 result;
2261 	u32 channel, wq;
2262 	u16 dest_wq;
2263 
2264 	p = get_affine_portal();
2265 	low_p = &p->p;
2266 
2267 	/* Determine the state of the FQID */
2268 	mcc = qm_mc_start(low_p);
2269 	mcc->queryfq_np.fqid = cpu_to_be32(fqid);
2270 	qm_mc_commit(low_p, QM_MCC_VERB_QUERYFQ_NP);
2271 	while (!(mcr = qm_mc_result(low_p)))
2272 		cpu_relax();
2273 	DPAA_ASSERT((mcr->verb & QM_MCR_VERB_MASK) == QM_MCR_VERB_QUERYFQ_NP);
2274 	state = mcr->queryfq_np.state & QM_MCR_NP_STATE_MASK;
2275 	if (state == QM_MCR_NP_STATE_OOS)
2276 		return 0; /* Already OOS, no need to do anymore checks */
2277 
2278 	/* Query which channel the FQ is using */
2279 	mcc = qm_mc_start(low_p);
2280 	mcc->queryfq.fqid = cpu_to_be32(fqid);
2281 	qm_mc_commit(low_p, QM_MCC_VERB_QUERYFQ);
2282 	while (!(mcr = qm_mc_result(low_p)))
2283 		cpu_relax();
2284 	DPAA_ASSERT((mcr->verb & QM_MCR_VERB_MASK) == QM_MCR_VERB_QUERYFQ);
2285 
2286 	/* Need to store these since the MCR gets reused */
2287 	dest_wq = be16_to_cpu(mcr->queryfq.fqd.dest_wq);
2288 	channel = dest_wq & 0x7;
2289 	wq = dest_wq >> 3;
2290 
2291 	switch (state) {
2292 	case QM_MCR_NP_STATE_TEN_SCHED:
2293 	case QM_MCR_NP_STATE_TRU_SCHED:
2294 	case QM_MCR_NP_STATE_ACTIVE:
2295 	case QM_MCR_NP_STATE_PARKED:
2296 		orl_empty = 0;
2297 		mcc = qm_mc_start(low_p);
2298 		mcc->alterfq.fqid = cpu_to_be32(fqid);
2299 		qm_mc_commit(low_p, QM_MCC_VERB_ALTER_RETIRE);
2300 		while (!(mcr = qm_mc_result(low_p)))
2301 			cpu_relax();
2302 		DPAA_ASSERT((mcr->verb & QM_MCR_VERB_MASK) ==
2303 			   QM_MCR_VERB_ALTER_RETIRE);
2304 		result = mcr->result; /* Make a copy as we reuse MCR below */
2305 
2306 		if (result == QM_MCR_RESULT_PENDING) {
2307 			/* Need to wait for the FQRN in the message ring, which
2308 			 * will only occur once the FQ has been drained.  In
2309 			 * order for the FQ to drain the portal needs to be set
2310 			 * to dequeue from the channel the FQ is scheduled on
2311 			 */
2312 			const struct qm_mr_entry *msg;
2313 			const struct qm_dqrr_entry *dqrr = NULL;
2314 			int found_fqrn = 0;
2315 			__maybe_unused u16 dequeue_wq = 0;
2316 
2317 			/* Flag that we need to drain FQ */
2318 			drain = 1;
2319 
2320 			if (channel >= qm_channel_pool1 &&
2321 			    channel < (u16)(qm_channel_pool1 + 15)) {
2322 				/* Pool channel, enable the bit in the portal */
2323 				dequeue_wq = (channel -
2324 					      qm_channel_pool1 + 1) << 4 | wq;
2325 			} else if (channel < qm_channel_pool1) {
2326 				/* Dedicated channel */
2327 				dequeue_wq = wq;
2328 			} else {
2329 				pr_info("Cannot recover FQ 0x%x,"
2330 					" it is scheduled on channel 0x%x",
2331 					fqid, channel);
2332 				return -EBUSY;
2333 			}
2334 			/* Set the sdqcr to drain this channel */
2335 			if (channel < qm_channel_pool1)
2336 				qm_dqrr_sdqcr_set(low_p,
2337 						  QM_SDQCR_TYPE_ACTIVE |
2338 					  QM_SDQCR_CHANNELS_DEDICATED);
2339 			else
2340 				qm_dqrr_sdqcr_set(low_p,
2341 						  QM_SDQCR_TYPE_ACTIVE |
2342 						  QM_SDQCR_CHANNELS_POOL_CONV
2343 						  (channel));
2344 			while (!found_fqrn) {
2345 				/* Keep draining DQRR while checking the MR*/
2346 				qm_dqrr_pvb_update(low_p);
2347 				dqrr = qm_dqrr_current(low_p);
2348 				while (dqrr) {
2349 					qm_dqrr_cdc_consume_1ptr(
2350 						low_p, dqrr, 0);
2351 					qm_dqrr_pvb_update(low_p);
2352 					qm_dqrr_next(low_p);
2353 					dqrr = qm_dqrr_current(low_p);
2354 				}
2355 				/* Process message ring too */
2356 				qm_mr_pvb_update(low_p);
2357 				msg = qm_mr_current(low_p);
2358 				while (msg) {
2359 					if ((msg->verb &
2360 					     QM_MR_VERB_TYPE_MASK)
2361 					    == QM_MR_VERB_FQRN)
2362 						found_fqrn = 1;
2363 					qm_mr_next(low_p);
2364 					qm_mr_cci_consume_to_current(low_p);
2365 					qm_mr_pvb_update(low_p);
2366 					msg = qm_mr_current(low_p);
2367 				}
2368 				cpu_relax();
2369 			}
2370 		}
2371 		if (result != QM_MCR_RESULT_OK &&
2372 		    result !=  QM_MCR_RESULT_PENDING) {
2373 			/* error */
2374 			pr_err("qman_retire_fq failed on FQ 0x%x,"
2375 			       " result=0x%x\n", fqid, result);
2376 			return -1;
2377 		}
2378 		if (!(mcr->alterfq.fqs & QM_MCR_FQS_ORLPRESENT)) {
2379 			/* ORL had no entries, no need to wait until the
2380 			 * ERNs come in.
2381 			 */
2382 			orl_empty = 1;
2383 		}
2384 		/* Retirement succeeded, check to see if FQ needs
2385 		 * to be drained.
2386 		 */
2387 		if (drain || mcr->alterfq.fqs & QM_MCR_FQS_NOTEMPTY) {
2388 			/* FQ is Not Empty, drain using volatile DQ commands */
2389 			fq_empty = 0;
2390 			do {
2391 				const struct qm_dqrr_entry *dqrr = NULL;
2392 				u32 vdqcr = fqid | QM_VDQCR_NUMFRAMES_SET(3);
2393 
2394 				qm_dqrr_vdqcr_set(low_p, vdqcr);
2395 
2396 				/* Wait for a dequeue to occur */
2397 				while (dqrr == NULL) {
2398 					qm_dqrr_pvb_update(low_p);
2399 					dqrr = qm_dqrr_current(low_p);
2400 					if (!dqrr)
2401 						cpu_relax();
2402 				}
2403 				/* Process the dequeues, making sure to
2404 				 * empty the ring completely.
2405 				 */
2406 				while (dqrr) {
2407 					if (dqrr->fqid == fqid &&
2408 					    dqrr->stat & QM_DQRR_STAT_FQ_EMPTY)
2409 						fq_empty = 1;
2410 					qm_dqrr_cdc_consume_1ptr(low_p,
2411 								 dqrr, 0);
2412 					qm_dqrr_pvb_update(low_p);
2413 					qm_dqrr_next(low_p);
2414 					dqrr = qm_dqrr_current(low_p);
2415 				}
2416 			} while (fq_empty == 0);
2417 		}
2418 		qm_dqrr_sdqcr_set(low_p, 0);
2419 
2420 		/* Wait for the ORL to have been completely drained */
2421 		while (orl_empty == 0) {
2422 			const struct qm_mr_entry *msg;
2423 
2424 			qm_mr_pvb_update(low_p);
2425 			msg = qm_mr_current(low_p);
2426 			while (msg) {
2427 				if ((msg->verb & QM_MR_VERB_TYPE_MASK) ==
2428 				    QM_MR_VERB_FQRL)
2429 					orl_empty = 1;
2430 				qm_mr_next(low_p);
2431 				qm_mr_cci_consume_to_current(low_p);
2432 				qm_mr_pvb_update(low_p);
2433 				msg = qm_mr_current(low_p);
2434 			}
2435 			cpu_relax();
2436 		}
2437 		mcc = qm_mc_start(low_p);
2438 		mcc->alterfq.fqid = cpu_to_be32(fqid);
2439 		qm_mc_commit(low_p, QM_MCC_VERB_ALTER_OOS);
2440 		while (!(mcr = qm_mc_result(low_p)))
2441 			cpu_relax();
2442 		DPAA_ASSERT((mcr->verb & QM_MCR_VERB_MASK) ==
2443 			   QM_MCR_VERB_ALTER_OOS);
2444 		if (mcr->result != QM_MCR_RESULT_OK) {
2445 			pr_err(
2446 			"OOS after drain Failed on FQID 0x%x, result 0x%x\n",
2447 			       fqid, mcr->result);
2448 			return -1;
2449 		}
2450 		return 0;
2451 
2452 	case QM_MCR_NP_STATE_RETIRED:
2453 		/* Send OOS Command */
2454 		mcc = qm_mc_start(low_p);
2455 		mcc->alterfq.fqid = cpu_to_be32(fqid);
2456 		qm_mc_commit(low_p, QM_MCC_VERB_ALTER_OOS);
2457 		while (!(mcr = qm_mc_result(low_p)))
2458 			cpu_relax();
2459 		DPAA_ASSERT((mcr->verb & QM_MCR_VERB_MASK) ==
2460 			   QM_MCR_VERB_ALTER_OOS);
2461 		if (mcr->result) {
2462 			pr_err("OOS Failed on FQID 0x%x\n", fqid);
2463 			return -1;
2464 		}
2465 		return 0;
2466 
2467 	}
2468 	return -1;
2469 }
2470