xref: /dpdk/drivers/bus/dpaa/base/qbman/qman.c (revision 2d0c29a37a9c080c1cccb1ad7941aba2ccf5437e)
1 /* SPDX-License-Identifier: (BSD-3-Clause OR GPL-2.0)
2  *
3  * Copyright 2008-2016 Freescale Semiconductor Inc.
4  * Copyright 2017,2019 NXP
5  *
6  */
7 
8 #include "qman.h"
9 #include <rte_branch_prediction.h>
10 #include <rte_dpaa_bus.h>
11 #include <rte_eventdev.h>
12 #include <rte_byteorder.h>
13 
14 /* Compilation constants */
15 #define DQRR_MAXFILL	15
16 #define EQCR_ITHRESH	4	/* if EQCR congests, interrupt threshold */
17 #define IRQNAME		"QMan portal %d"
18 #define MAX_IRQNAME	16	/* big enough for "QMan portal %d" */
19 /* maximum number of DQRR entries to process in qman_poll() */
20 #define FSL_QMAN_POLL_LIMIT 8
21 
22 /* Lock/unlock frame queues, subject to the "LOCKED" flag. This is about
23  * inter-processor locking only. Note, FQLOCK() is always called either under a
24  * local_irq_save() or from interrupt context - hence there's no need for irq
25  * protection (and indeed, attempting to nest irq-protection doesn't work, as
26  * the "irq en/disable" machinery isn't recursive...).
27  */
28 #define FQLOCK(fq) \
29 	do { \
30 		struct qman_fq *__fq478 = (fq); \
31 		if (fq_isset(__fq478, QMAN_FQ_FLAG_LOCKED)) \
32 			spin_lock(&__fq478->fqlock); \
33 	} while (0)
34 #define FQUNLOCK(fq) \
35 	do { \
36 		struct qman_fq *__fq478 = (fq); \
37 		if (fq_isset(__fq478, QMAN_FQ_FLAG_LOCKED)) \
38 			spin_unlock(&__fq478->fqlock); \
39 	} while (0)
40 
41 static inline void fq_set(struct qman_fq *fq, u32 mask)
42 {
43 	dpaa_set_bits(mask, &fq->flags);
44 }
45 
46 static inline void fq_clear(struct qman_fq *fq, u32 mask)
47 {
48 	dpaa_clear_bits(mask, &fq->flags);
49 }
50 
51 static inline int fq_isset(struct qman_fq *fq, u32 mask)
52 {
53 	return fq->flags & mask;
54 }
55 
56 static inline int fq_isclear(struct qman_fq *fq, u32 mask)
57 {
58 	return !(fq->flags & mask);
59 }
60 
61 struct qman_portal {
62 	struct qm_portal p;
63 	/* PORTAL_BITS_*** - dynamic, strictly internal */
64 	unsigned long bits;
65 	/* interrupt sources processed by portal_isr(), configurable */
66 	unsigned long irq_sources;
67 	u32 use_eqcr_ci_stashing;
68 	u32 slowpoll;	/* only used when interrupts are off */
69 	/* only 1 volatile dequeue at a time */
70 	struct qman_fq *vdqcr_owned;
71 	u32 sdqcr;
72 	int dqrr_disable_ref;
73 	/* A portal-specific handler for DCP ERNs. If this is NULL, the global
74 	 * handler is called instead.
75 	 */
76 	qman_cb_dc_ern cb_dc_ern;
77 	/* When the cpu-affine portal is activated, this is non-NULL */
78 	const struct qm_portal_config *config;
79 	struct dpa_rbtree retire_table;
80 	char irqname[MAX_IRQNAME];
81 	/* 2-element array. cgrs[0] is mask, cgrs[1] is snapshot. */
82 	struct qman_cgrs *cgrs;
83 	/* linked-list of CSCN handlers. */
84 	struct list_head cgr_cbs;
85 	/* list lock */
86 	spinlock_t cgr_lock;
87 	/* track if memory was allocated by the driver */
88 #if __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__
89 	/* Keep a shadow copy of the DQRR on LE systems as the SW needs to
90 	 * do byte swaps of DQRR read only memory.  First entry must be aligned
91 	 * to 2 ** 10 to ensure DQRR index calculations based shadow copy
92 	 * address (6 bits for address shift + 4 bits for the DQRR size).
93 	 */
94 	struct qm_dqrr_entry shadow_dqrr[QM_DQRR_SIZE]
95 		    __attribute__((aligned(1024)));
96 #endif
97 };
98 
99 /* Global handler for DCP ERNs. Used when the portal receiving the message does
100  * not have a portal-specific handler.
101  */
102 static qman_cb_dc_ern cb_dc_ern;
103 
104 static cpumask_t affine_mask;
105 static DEFINE_SPINLOCK(affine_mask_lock);
106 static u16 affine_channels[NR_CPUS];
107 static RTE_DEFINE_PER_LCORE(struct qman_portal, qman_affine_portal);
108 
109 static inline struct qman_portal *get_affine_portal(void)
110 {
111 	return &RTE_PER_LCORE(qman_affine_portal);
112 }
113 
114 /* This gives a FQID->FQ lookup to cover the fact that we can't directly demux
115  * retirement notifications (the fact they are sometimes h/w-consumed means that
116  * contextB isn't always a s/w demux - and as we can't know which case it is
117  * when looking at the notification, we have to use the slow lookup for all of
118  * them). NB, it's possible to have multiple FQ objects refer to the same FQID
119  * (though at most one of them should be the consumer), so this table isn't for
120  * all FQs - FQs are added when retirement commands are issued, and removed when
121  * they complete, which also massively reduces the size of this table.
122  */
123 IMPLEMENT_DPAA_RBTREE(fqtree, struct qman_fq, node, fqid);
124 /*
125  * This is what everything can wait on, even if it migrates to a different cpu
126  * to the one whose affine portal it is waiting on.
127  */
128 static DECLARE_WAIT_QUEUE_HEAD(affine_queue);
129 
130 static inline int table_push_fq(struct qman_portal *p, struct qman_fq *fq)
131 {
132 	int ret = fqtree_push(&p->retire_table, fq);
133 
134 	if (ret)
135 		pr_err("ERROR: double FQ-retirement %d\n", fq->fqid);
136 	return ret;
137 }
138 
139 static inline void table_del_fq(struct qman_portal *p, struct qman_fq *fq)
140 {
141 	fqtree_del(&p->retire_table, fq);
142 }
143 
144 static inline struct qman_fq *table_find_fq(struct qman_portal *p, u32 fqid)
145 {
146 	return fqtree_find(&p->retire_table, fqid);
147 }
148 
149 #ifdef CONFIG_FSL_QMAN_FQ_LOOKUP
150 static void **qman_fq_lookup_table;
151 static size_t qman_fq_lookup_table_size;
152 
153 int qman_setup_fq_lookup_table(size_t num_entries)
154 {
155 	num_entries++;
156 	/* Allocate 1 more entry since the first entry is not used */
157 	qman_fq_lookup_table = vmalloc((num_entries * sizeof(void *)));
158 	if (!qman_fq_lookup_table) {
159 		pr_err("QMan: Could not allocate fq lookup table\n");
160 		return -ENOMEM;
161 	}
162 	memset(qman_fq_lookup_table, 0, num_entries * sizeof(void *));
163 	qman_fq_lookup_table_size = num_entries;
164 	pr_debug("QMan: Allocated lookup table at %p, entry count %lu\n",
165 		qman_fq_lookup_table,
166 			(unsigned long)qman_fq_lookup_table_size);
167 	return 0;
168 }
169 
170 void qman_set_fq_lookup_table(void **fq_table)
171 {
172 	qman_fq_lookup_table = fq_table;
173 }
174 
175 /* global structure that maintains fq object mapping */
176 static DEFINE_SPINLOCK(fq_hash_table_lock);
177 
178 static int find_empty_fq_table_entry(u32 *entry, struct qman_fq *fq)
179 {
180 	u32 i;
181 
182 	spin_lock(&fq_hash_table_lock);
183 	/* Can't use index zero because this has special meaning
184 	 * in context_b field.
185 	 */
186 	for (i = 1; i < qman_fq_lookup_table_size; i++) {
187 		if (qman_fq_lookup_table[i] == NULL) {
188 			*entry = i;
189 			qman_fq_lookup_table[i] = fq;
190 			spin_unlock(&fq_hash_table_lock);
191 			return 0;
192 		}
193 	}
194 	spin_unlock(&fq_hash_table_lock);
195 	return -ENOMEM;
196 }
197 
198 static void clear_fq_table_entry(u32 entry)
199 {
200 	spin_lock(&fq_hash_table_lock);
201 	DPAA_BUG_ON(entry >= qman_fq_lookup_table_size);
202 	qman_fq_lookup_table[entry] = NULL;
203 	spin_unlock(&fq_hash_table_lock);
204 }
205 
206 static inline struct qman_fq *get_fq_table_entry(u32 entry)
207 {
208 	DPAA_BUG_ON(entry >= qman_fq_lookup_table_size);
209 	return qman_fq_lookup_table[entry];
210 }
211 #endif
212 
213 static inline void cpu_to_hw_fqd(struct qm_fqd *fqd)
214 {
215 	/* Byteswap the FQD to HW format */
216 	fqd->fq_ctrl = cpu_to_be16(fqd->fq_ctrl);
217 	fqd->dest_wq = cpu_to_be16(fqd->dest_wq);
218 	fqd->ics_cred = cpu_to_be16(fqd->ics_cred);
219 	fqd->context_b = cpu_to_be32(fqd->context_b);
220 	fqd->context_a.opaque = cpu_to_be64(fqd->context_a.opaque);
221 	fqd->opaque_td = cpu_to_be16(fqd->opaque_td);
222 }
223 
224 static inline void hw_fqd_to_cpu(struct qm_fqd *fqd)
225 {
226 	/* Byteswap the FQD to CPU format */
227 	fqd->fq_ctrl = be16_to_cpu(fqd->fq_ctrl);
228 	fqd->dest_wq = be16_to_cpu(fqd->dest_wq);
229 	fqd->ics_cred = be16_to_cpu(fqd->ics_cred);
230 	fqd->context_b = be32_to_cpu(fqd->context_b);
231 	fqd->context_a.opaque = be64_to_cpu(fqd->context_a.opaque);
232 }
233 
234 static inline void cpu_to_hw_fd(struct qm_fd *fd)
235 {
236 	fd->addr = cpu_to_be40(fd->addr);
237 	fd->status = cpu_to_be32(fd->status);
238 	fd->opaque = cpu_to_be32(fd->opaque);
239 }
240 
241 static inline void hw_fd_to_cpu(struct qm_fd *fd)
242 {
243 	fd->addr = be40_to_cpu(fd->addr);
244 	fd->status = be32_to_cpu(fd->status);
245 	fd->opaque = be32_to_cpu(fd->opaque);
246 }
247 
248 /* In the case that slow- and fast-path handling are both done by qman_poll()
249  * (ie. because there is no interrupt handling), we ought to balance how often
250  * we do the fast-path poll versus the slow-path poll. We'll use two decrementer
251  * sources, so we call the fast poll 'n' times before calling the slow poll
252  * once. The idle decrementer constant is used when the last slow-poll detected
253  * no work to do, and the busy decrementer constant when the last slow-poll had
254  * work to do.
255  */
256 #define SLOW_POLL_IDLE   1000
257 #define SLOW_POLL_BUSY   10
258 static u32 __poll_portal_slow(struct qman_portal *p, u32 is);
259 static inline unsigned int __poll_portal_fast(struct qman_portal *p,
260 					      unsigned int poll_limit);
261 
262 /* Portal interrupt handler */
263 static irqreturn_t portal_isr(__always_unused int irq, void *ptr)
264 {
265 	struct qman_portal *p = ptr;
266 	/*
267 	 * The CSCI/CCSCI source is cleared inside __poll_portal_slow(), because
268 	 * it could race against a Query Congestion State command also given
269 	 * as part of the handling of this interrupt source. We mustn't
270 	 * clear it a second time in this top-level function.
271 	 */
272 	u32 clear = QM_DQAVAIL_MASK | (p->irq_sources &
273 		~(QM_PIRQ_CSCI | QM_PIRQ_CCSCI));
274 	u32 is = qm_isr_status_read(&p->p) & p->irq_sources;
275 	/* DQRR-handling if it's interrupt-driven */
276 	if (is & QM_PIRQ_DQRI)
277 		__poll_portal_fast(p, FSL_QMAN_POLL_LIMIT);
278 	/* Handling of anything else that's interrupt-driven */
279 	clear |= __poll_portal_slow(p, is);
280 	qm_isr_status_clear(&p->p, clear);
281 	return IRQ_HANDLED;
282 }
283 
284 /* This inner version is used privately by qman_create_affine_portal(), as well
285  * as by the exported qman_stop_dequeues().
286  */
287 static inline void qman_stop_dequeues_ex(struct qman_portal *p)
288 {
289 	if (!(p->dqrr_disable_ref++))
290 		qm_dqrr_set_maxfill(&p->p, 0);
291 }
292 
293 static int drain_mr_fqrni(struct qm_portal *p)
294 {
295 	const struct qm_mr_entry *msg;
296 loop:
297 	msg = qm_mr_current(p);
298 	if (!msg) {
299 		/*
300 		 * if MR was full and h/w had other FQRNI entries to produce, we
301 		 * need to allow it time to produce those entries once the
302 		 * existing entries are consumed. A worst-case situation
303 		 * (fully-loaded system) means h/w sequencers may have to do 3-4
304 		 * other things before servicing the portal's MR pump, each of
305 		 * which (if slow) may take ~50 qman cycles (which is ~200
306 		 * processor cycles). So rounding up and then multiplying this
307 		 * worst-case estimate by a factor of 10, just to be
308 		 * ultra-paranoid, goes as high as 10,000 cycles. NB, we consume
309 		 * one entry at a time, so h/w has an opportunity to produce new
310 		 * entries well before the ring has been fully consumed, so
311 		 * we're being *really* paranoid here.
312 		 */
313 		u64 now, then = mfatb();
314 
315 		do {
316 			now = mfatb();
317 		} while ((then + 10000) > now);
318 		msg = qm_mr_current(p);
319 		if (!msg)
320 			return 0;
321 	}
322 	if ((msg->ern.verb & QM_MR_VERB_TYPE_MASK) != QM_MR_VERB_FQRNI) {
323 		/* We aren't draining anything but FQRNIs */
324 		pr_err("Found verb 0x%x in MR\n", msg->ern.verb);
325 		return -1;
326 	}
327 	qm_mr_next(p);
328 	qm_mr_cci_consume(p, 1);
329 	goto loop;
330 }
331 
332 static inline int qm_eqcr_init(struct qm_portal *portal,
333 			       enum qm_eqcr_pmode pmode,
334 			       unsigned int eq_stash_thresh,
335 			       int eq_stash_prio)
336 {
337 	/* This use of 'register', as well as all other occurrences, is because
338 	 * it has been observed to generate much faster code with gcc than is
339 	 * otherwise the case.
340 	 */
341 	register struct qm_eqcr *eqcr = &portal->eqcr;
342 	u32 cfg;
343 	u8 pi;
344 
345 	eqcr->ring = portal->addr.ce + QM_CL_EQCR;
346 	eqcr->ci = qm_in(EQCR_CI_CINH) & (QM_EQCR_SIZE - 1);
347 	qm_cl_invalidate(EQCR_CI);
348 	pi = qm_in(EQCR_PI_CINH) & (QM_EQCR_SIZE - 1);
349 	eqcr->cursor = eqcr->ring + pi;
350 	eqcr->vbit = (qm_in(EQCR_PI_CINH) & QM_EQCR_SIZE) ?
351 			QM_EQCR_VERB_VBIT : 0;
352 	eqcr->available = QM_EQCR_SIZE - 1 -
353 			qm_cyc_diff(QM_EQCR_SIZE, eqcr->ci, pi);
354 	eqcr->ithresh = qm_in(EQCR_ITR);
355 #ifdef RTE_LIBRTE_DPAA_HWDEBUG
356 	eqcr->busy = 0;
357 	eqcr->pmode = pmode;
358 #endif
359 	cfg = (qm_in(CFG) & 0x00ffffff) |
360 		(eq_stash_thresh << 28) | /* QCSP_CFG: EST */
361 		(eq_stash_prio << 26)	| /* QCSP_CFG: EP */
362 		((pmode & 0x3) << 24);	/* QCSP_CFG::EPM */
363 	qm_out(CFG, cfg);
364 	return 0;
365 }
366 
367 static inline void qm_eqcr_finish(struct qm_portal *portal)
368 {
369 	register struct qm_eqcr *eqcr = &portal->eqcr;
370 	u8 pi, ci;
371 	u32 cfg;
372 
373 	/*
374 	 * Disable EQCI stashing because the QMan only
375 	 * presents the value it previously stashed to
376 	 * maintain coherency.  Setting the stash threshold
377 	 * to 1 then 0 ensures that QMan has resyncronized
378 	 * its internal copy so that the portal is clean
379 	 * when it is reinitialized in the future
380 	 */
381 	cfg = (qm_in(CFG) & 0x0fffffff) |
382 		(1 << 28); /* QCSP_CFG: EST */
383 	qm_out(CFG, cfg);
384 	cfg &= 0x0fffffff; /* stash threshold = 0 */
385 	qm_out(CFG, cfg);
386 
387 	pi = qm_in(EQCR_PI_CINH) & (QM_EQCR_SIZE - 1);
388 	ci = qm_in(EQCR_CI_CINH) & (QM_EQCR_SIZE - 1);
389 
390 	/* Refresh EQCR CI cache value */
391 	qm_cl_invalidate(EQCR_CI);
392 	eqcr->ci = qm_cl_in(EQCR_CI) & (QM_EQCR_SIZE - 1);
393 
394 #ifdef RTE_LIBRTE_DPAA_HWDEBUG
395 	DPAA_ASSERT(!eqcr->busy);
396 #endif
397 	if (pi != EQCR_PTR2IDX(eqcr->cursor))
398 		pr_crit("losing uncommitted EQCR entries\n");
399 	if (ci != eqcr->ci)
400 		pr_crit("missing existing EQCR completions\n");
401 	if (eqcr->ci != EQCR_PTR2IDX(eqcr->cursor))
402 		pr_crit("EQCR destroyed unquiesced\n");
403 }
404 
405 static inline int qm_dqrr_init(struct qm_portal *portal,
406 			__maybe_unused const struct qm_portal_config *config,
407 			enum qm_dqrr_dmode dmode,
408 			__maybe_unused enum qm_dqrr_pmode pmode,
409 			enum qm_dqrr_cmode cmode, u8 max_fill)
410 {
411 	register struct qm_dqrr *dqrr = &portal->dqrr;
412 	u32 cfg;
413 
414 	/* Make sure the DQRR will be idle when we enable */
415 	qm_out(DQRR_SDQCR, 0);
416 	qm_out(DQRR_VDQCR, 0);
417 	qm_out(DQRR_PDQCR, 0);
418 	dqrr->ring = portal->addr.ce + QM_CL_DQRR;
419 	dqrr->pi = qm_in(DQRR_PI_CINH) & (QM_DQRR_SIZE - 1);
420 	dqrr->ci = qm_in(DQRR_CI_CINH) & (QM_DQRR_SIZE - 1);
421 	dqrr->cursor = dqrr->ring + dqrr->ci;
422 	dqrr->fill = qm_cyc_diff(QM_DQRR_SIZE, dqrr->ci, dqrr->pi);
423 	dqrr->vbit = (qm_in(DQRR_PI_CINH) & QM_DQRR_SIZE) ?
424 			QM_DQRR_VERB_VBIT : 0;
425 	dqrr->ithresh = qm_in(DQRR_ITR);
426 #ifdef RTE_LIBRTE_DPAA_HWDEBUG
427 	dqrr->dmode = dmode;
428 	dqrr->pmode = pmode;
429 	dqrr->cmode = cmode;
430 #endif
431 	/* Invalidate every ring entry before beginning */
432 	for (cfg = 0; cfg < QM_DQRR_SIZE; cfg++)
433 		dccivac(qm_cl(dqrr->ring, cfg));
434 	cfg = (qm_in(CFG) & 0xff000f00) |
435 		((max_fill & (QM_DQRR_SIZE - 1)) << 20) | /* DQRR_MF */
436 		((dmode & 1) << 18) |			/* DP */
437 		((cmode & 3) << 16) |			/* DCM */
438 		0xa0 |					/* RE+SE */
439 		(0 ? 0x40 : 0) |			/* Ignore RP */
440 		(0 ? 0x10 : 0);				/* Ignore SP */
441 	qm_out(CFG, cfg);
442 	qm_dqrr_set_maxfill(portal, max_fill);
443 	return 0;
444 }
445 
446 static inline void qm_dqrr_finish(struct qm_portal *portal)
447 {
448 	__maybe_unused register struct qm_dqrr *dqrr = &portal->dqrr;
449 #ifdef RTE_LIBRTE_DPAA_HWDEBUG
450 	if ((dqrr->cmode != qm_dqrr_cdc) &&
451 	    (dqrr->ci != DQRR_PTR2IDX(dqrr->cursor)))
452 		pr_crit("Ignoring completed DQRR entries\n");
453 #endif
454 }
455 
456 static inline int qm_mr_init(struct qm_portal *portal,
457 			     __maybe_unused enum qm_mr_pmode pmode,
458 			     enum qm_mr_cmode cmode)
459 {
460 	register struct qm_mr *mr = &portal->mr;
461 	u32 cfg;
462 
463 	mr->ring = portal->addr.ce + QM_CL_MR;
464 	mr->pi = qm_in(MR_PI_CINH) & (QM_MR_SIZE - 1);
465 	mr->ci = qm_in(MR_CI_CINH) & (QM_MR_SIZE - 1);
466 	mr->cursor = mr->ring + mr->ci;
467 	mr->fill = qm_cyc_diff(QM_MR_SIZE, mr->ci, mr->pi);
468 	mr->vbit = (qm_in(MR_PI_CINH) & QM_MR_SIZE) ? QM_MR_VERB_VBIT : 0;
469 	mr->ithresh = qm_in(MR_ITR);
470 #ifdef RTE_LIBRTE_DPAA_HWDEBUG
471 	mr->pmode = pmode;
472 	mr->cmode = cmode;
473 #endif
474 	cfg = (qm_in(CFG) & 0xfffff0ff) |
475 		((cmode & 1) << 8);		/* QCSP_CFG:MM */
476 	qm_out(CFG, cfg);
477 	return 0;
478 }
479 
480 static inline void qm_mr_pvb_update(struct qm_portal *portal)
481 {
482 	register struct qm_mr *mr = &portal->mr;
483 	const struct qm_mr_entry *res = qm_cl(mr->ring, mr->pi);
484 
485 #ifdef RTE_LIBRTE_DPAA_HWDEBUG
486 	DPAA_ASSERT(mr->pmode == qm_mr_pvb);
487 #endif
488 	/* when accessing 'verb', use __raw_readb() to ensure that compiler
489 	 * inlining doesn't try to optimise out "excess reads".
490 	 */
491 	if ((__raw_readb(&res->ern.verb) & QM_MR_VERB_VBIT) == mr->vbit) {
492 		mr->pi = (mr->pi + 1) & (QM_MR_SIZE - 1);
493 		if (!mr->pi)
494 			mr->vbit ^= QM_MR_VERB_VBIT;
495 		mr->fill++;
496 		res = MR_INC(res);
497 	}
498 	dcbit_ro(res);
499 }
500 
501 static inline
502 struct qman_portal *qman_create_portal(
503 			struct qman_portal *portal,
504 			      const struct qm_portal_config *c,
505 			      const struct qman_cgrs *cgrs)
506 {
507 	struct qm_portal *p;
508 	char buf[16];
509 	int ret;
510 	u32 isdr;
511 
512 	p = &portal->p;
513 
514 	if (dpaa_svr_family == SVR_LS1043A_FAMILY)
515 		portal->use_eqcr_ci_stashing = 3;
516 	else
517 		portal->use_eqcr_ci_stashing =
518 					((qman_ip_rev >= QMAN_REV30) ? 1 : 0);
519 
520 	/*
521 	 * prep the low-level portal struct with the mapped addresses from the
522 	 * config, everything that follows depends on it and "config" is more
523 	 * for (de)reference
524 	 */
525 	p->addr.ce = c->addr_virt[DPAA_PORTAL_CE];
526 	p->addr.ci = c->addr_virt[DPAA_PORTAL_CI];
527 	/*
528 	 * If CI-stashing is used, the current defaults use a threshold of 3,
529 	 * and stash with high-than-DQRR priority.
530 	 */
531 	if (qm_eqcr_init(p, qm_eqcr_pvb,
532 			 portal->use_eqcr_ci_stashing, 1)) {
533 		pr_err("Qman EQCR initialisation failed\n");
534 		goto fail_eqcr;
535 	}
536 	if (qm_dqrr_init(p, c, qm_dqrr_dpush, qm_dqrr_pvb,
537 			 qm_dqrr_cdc, DQRR_MAXFILL)) {
538 		pr_err("Qman DQRR initialisation failed\n");
539 		goto fail_dqrr;
540 	}
541 	if (qm_mr_init(p, qm_mr_pvb, qm_mr_cci)) {
542 		pr_err("Qman MR initialisation failed\n");
543 		goto fail_mr;
544 	}
545 	if (qm_mc_init(p)) {
546 		pr_err("Qman MC initialisation failed\n");
547 		goto fail_mc;
548 	}
549 
550 	/* static interrupt-gating controls */
551 	qm_dqrr_set_ithresh(p, 0);
552 	qm_mr_set_ithresh(p, 0);
553 	qm_isr_set_iperiod(p, 0);
554 	portal->cgrs = kmalloc(2 * sizeof(*cgrs), GFP_KERNEL);
555 	if (!portal->cgrs)
556 		goto fail_cgrs;
557 	/* initial snapshot is no-depletion */
558 	qman_cgrs_init(&portal->cgrs[1]);
559 	if (cgrs)
560 		portal->cgrs[0] = *cgrs;
561 	else
562 		/* if the given mask is NULL, assume all CGRs can be seen */
563 		qman_cgrs_fill(&portal->cgrs[0]);
564 	INIT_LIST_HEAD(&portal->cgr_cbs);
565 	spin_lock_init(&portal->cgr_lock);
566 	portal->bits = 0;
567 	portal->slowpoll = 0;
568 	portal->sdqcr = QM_SDQCR_SOURCE_CHANNELS | QM_SDQCR_COUNT_UPTO3 |
569 			QM_SDQCR_DEDICATED_PRECEDENCE | QM_SDQCR_TYPE_PRIO_QOS |
570 			QM_SDQCR_TOKEN_SET(0xab) | QM_SDQCR_CHANNELS_DEDICATED;
571 	portal->dqrr_disable_ref = 0;
572 	portal->cb_dc_ern = NULL;
573 	sprintf(buf, "qportal-%d", c->channel);
574 	dpa_rbtree_init(&portal->retire_table);
575 	isdr = 0xffffffff;
576 	qm_isr_disable_write(p, isdr);
577 	portal->irq_sources = 0;
578 	qm_isr_enable_write(p, portal->irq_sources);
579 	qm_isr_status_clear(p, 0xffffffff);
580 	snprintf(portal->irqname, MAX_IRQNAME, IRQNAME, c->cpu);
581 	if (request_irq(c->irq, portal_isr, 0, portal->irqname,
582 			portal)) {
583 		pr_err("request_irq() failed\n");
584 		goto fail_irq;
585 	}
586 
587 	/* Need EQCR to be empty before continuing */
588 	isdr &= ~QM_PIRQ_EQCI;
589 	qm_isr_disable_write(p, isdr);
590 	ret = qm_eqcr_get_fill(p);
591 	if (ret) {
592 		pr_err("Qman EQCR unclean\n");
593 		goto fail_eqcr_empty;
594 	}
595 	isdr &= ~(QM_PIRQ_DQRI | QM_PIRQ_MRI);
596 	qm_isr_disable_write(p, isdr);
597 	if (qm_dqrr_current(p)) {
598 		pr_err("Qman DQRR unclean\n");
599 		qm_dqrr_cdc_consume_n(p, 0xffff);
600 	}
601 	if (qm_mr_current(p) && drain_mr_fqrni(p)) {
602 		/* special handling, drain just in case it's a few FQRNIs */
603 		if (drain_mr_fqrni(p))
604 			goto fail_dqrr_mr_empty;
605 	}
606 	/* Success */
607 	portal->config = c;
608 	qm_isr_disable_write(p, 0);
609 	qm_isr_uninhibit(p);
610 	/* Write a sane SDQCR */
611 	qm_dqrr_sdqcr_set(p, portal->sdqcr);
612 	return portal;
613 fail_dqrr_mr_empty:
614 fail_eqcr_empty:
615 	free_irq(c->irq, portal);
616 fail_irq:
617 	kfree(portal->cgrs);
618 	spin_lock_destroy(&portal->cgr_lock);
619 fail_cgrs:
620 	qm_mc_finish(p);
621 fail_mc:
622 	qm_mr_finish(p);
623 fail_mr:
624 	qm_dqrr_finish(p);
625 fail_dqrr:
626 	qm_eqcr_finish(p);
627 fail_eqcr:
628 	return NULL;
629 }
630 
631 #define MAX_GLOBAL_PORTALS 8
632 static struct qman_portal global_portals[MAX_GLOBAL_PORTALS];
633 static rte_atomic16_t global_portals_used[MAX_GLOBAL_PORTALS];
634 
635 static struct qman_portal *
636 qman_alloc_global_portal(void)
637 {
638 	unsigned int i;
639 
640 	for (i = 0; i < MAX_GLOBAL_PORTALS; i++) {
641 		if (rte_atomic16_test_and_set(&global_portals_used[i]))
642 			return &global_portals[i];
643 	}
644 	pr_err("No portal available (%x)\n", MAX_GLOBAL_PORTALS);
645 
646 	return NULL;
647 }
648 
649 static int
650 qman_free_global_portal(struct qman_portal *portal)
651 {
652 	unsigned int i;
653 
654 	for (i = 0; i < MAX_GLOBAL_PORTALS; i++) {
655 		if (&global_portals[i] == portal) {
656 			rte_atomic16_clear(&global_portals_used[i]);
657 			return 0;
658 		}
659 	}
660 	return -1;
661 }
662 
663 struct qman_portal *qman_create_affine_portal(const struct qm_portal_config *c,
664 					      const struct qman_cgrs *cgrs,
665 					      int alloc)
666 {
667 	struct qman_portal *res;
668 	struct qman_portal *portal;
669 
670 	if (alloc)
671 		portal = qman_alloc_global_portal();
672 	else
673 		portal = get_affine_portal();
674 
675 	/* A criteria for calling this function (from qman_driver.c) is that
676 	 * we're already affine to the cpu and won't schedule onto another cpu.
677 	 */
678 
679 	res = qman_create_portal(portal, c, cgrs);
680 	if (res) {
681 		spin_lock(&affine_mask_lock);
682 		CPU_SET(c->cpu, &affine_mask);
683 		affine_channels[c->cpu] =
684 			c->channel;
685 		spin_unlock(&affine_mask_lock);
686 	}
687 	return res;
688 }
689 
690 static inline
691 void qman_destroy_portal(struct qman_portal *qm)
692 {
693 	const struct qm_portal_config *pcfg;
694 
695 	/* Stop dequeues on the portal */
696 	qm_dqrr_sdqcr_set(&qm->p, 0);
697 
698 	/*
699 	 * NB we do this to "quiesce" EQCR. If we add enqueue-completions or
700 	 * something related to QM_PIRQ_EQCI, this may need fixing.
701 	 * Also, due to the prefetching model used for CI updates in the enqueue
702 	 * path, this update will only invalidate the CI cacheline *after*
703 	 * working on it, so we need to call this twice to ensure a full update
704 	 * irrespective of where the enqueue processing was at when the teardown
705 	 * began.
706 	 */
707 	qm_eqcr_cce_update(&qm->p);
708 	qm_eqcr_cce_update(&qm->p);
709 	pcfg = qm->config;
710 
711 	free_irq(pcfg->irq, qm);
712 
713 	kfree(qm->cgrs);
714 	qm_mc_finish(&qm->p);
715 	qm_mr_finish(&qm->p);
716 	qm_dqrr_finish(&qm->p);
717 	qm_eqcr_finish(&qm->p);
718 
719 	qm->config = NULL;
720 
721 	spin_lock_destroy(&qm->cgr_lock);
722 }
723 
724 const struct qm_portal_config *
725 qman_destroy_affine_portal(struct qman_portal *qp)
726 {
727 	/* We don't want to redirect if we're a slave, use "raw" */
728 	struct qman_portal *qm;
729 	const struct qm_portal_config *pcfg;
730 	int cpu;
731 
732 	if (qp == NULL)
733 		qm = get_affine_portal();
734 	else
735 		qm = qp;
736 	pcfg = qm->config;
737 	cpu = pcfg->cpu;
738 
739 	qman_destroy_portal(qm);
740 
741 	spin_lock(&affine_mask_lock);
742 	CPU_CLR(cpu, &affine_mask);
743 	spin_unlock(&affine_mask_lock);
744 
745 	qman_free_global_portal(qm);
746 
747 	return pcfg;
748 }
749 
750 int qman_get_portal_index(void)
751 {
752 	struct qman_portal *p = get_affine_portal();
753 	return p->config->index;
754 }
755 
756 /* Inline helper to reduce nesting in __poll_portal_slow() */
757 static inline void fq_state_change(struct qman_portal *p, struct qman_fq *fq,
758 				   const struct qm_mr_entry *msg, u8 verb)
759 {
760 	FQLOCK(fq);
761 	switch (verb) {
762 	case QM_MR_VERB_FQRL:
763 		DPAA_ASSERT(fq_isset(fq, QMAN_FQ_STATE_ORL));
764 		fq_clear(fq, QMAN_FQ_STATE_ORL);
765 		table_del_fq(p, fq);
766 		break;
767 	case QM_MR_VERB_FQRN:
768 		DPAA_ASSERT((fq->state == qman_fq_state_parked) ||
769 			    (fq->state == qman_fq_state_sched));
770 		DPAA_ASSERT(fq_isset(fq, QMAN_FQ_STATE_CHANGING));
771 		fq_clear(fq, QMAN_FQ_STATE_CHANGING);
772 		if (msg->fq.fqs & QM_MR_FQS_NOTEMPTY)
773 			fq_set(fq, QMAN_FQ_STATE_NE);
774 		if (msg->fq.fqs & QM_MR_FQS_ORLPRESENT)
775 			fq_set(fq, QMAN_FQ_STATE_ORL);
776 		else
777 			table_del_fq(p, fq);
778 		fq->state = qman_fq_state_retired;
779 		break;
780 	case QM_MR_VERB_FQPN:
781 		DPAA_ASSERT(fq->state == qman_fq_state_sched);
782 		DPAA_ASSERT(fq_isclear(fq, QMAN_FQ_STATE_CHANGING));
783 		fq->state = qman_fq_state_parked;
784 	}
785 	FQUNLOCK(fq);
786 }
787 
788 static u32 __poll_portal_slow(struct qman_portal *p, u32 is)
789 {
790 	const struct qm_mr_entry *msg;
791 	struct qm_mr_entry swapped_msg;
792 
793 	if (is & QM_PIRQ_CSCI) {
794 		struct qman_cgrs rr, c;
795 		struct qm_mc_result *mcr;
796 		struct qman_cgr *cgr;
797 
798 		spin_lock(&p->cgr_lock);
799 		/*
800 		 * The CSCI bit must be cleared _before_ issuing the
801 		 * Query Congestion State command, to ensure that a long
802 		 * CGR State Change callback cannot miss an intervening
803 		 * state change.
804 		 */
805 		qm_isr_status_clear(&p->p, QM_PIRQ_CSCI);
806 		qm_mc_start(&p->p);
807 		qm_mc_commit(&p->p, QM_MCC_VERB_QUERYCONGESTION);
808 		while (!(mcr = qm_mc_result(&p->p)))
809 			cpu_relax();
810 		/* mask out the ones I'm not interested in */
811 		qman_cgrs_and(&rr, (const struct qman_cgrs *)
812 			&mcr->querycongestion.state, &p->cgrs[0]);
813 		/* check previous snapshot for delta, enter/exit congestion */
814 		qman_cgrs_xor(&c, &rr, &p->cgrs[1]);
815 		/* update snapshot */
816 		qman_cgrs_cp(&p->cgrs[1], &rr);
817 		/* Invoke callback */
818 		list_for_each_entry(cgr, &p->cgr_cbs, node)
819 			if (cgr->cb && qman_cgrs_get(&c, cgr->cgrid))
820 				cgr->cb(p, cgr, qman_cgrs_get(&rr, cgr->cgrid));
821 		spin_unlock(&p->cgr_lock);
822 	}
823 
824 	if (is & QM_PIRQ_EQRI) {
825 		qm_eqcr_cce_update(&p->p);
826 		qm_eqcr_set_ithresh(&p->p, 0);
827 		wake_up(&affine_queue);
828 	}
829 
830 	if (is & QM_PIRQ_MRI) {
831 		struct qman_fq *fq;
832 		u8 verb, num = 0;
833 mr_loop:
834 		qm_mr_pvb_update(&p->p);
835 		msg = qm_mr_current(&p->p);
836 		if (!msg)
837 			goto mr_done;
838 		swapped_msg = *msg;
839 		hw_fd_to_cpu(&swapped_msg.ern.fd);
840 		verb = msg->ern.verb & QM_MR_VERB_TYPE_MASK;
841 		/* The message is a software ERN iff the 0x20 bit is set */
842 		if (verb & 0x20) {
843 			switch (verb) {
844 			case QM_MR_VERB_FQRNI:
845 				/* nada, we drop FQRNIs on the floor */
846 				break;
847 			case QM_MR_VERB_FQRN:
848 			case QM_MR_VERB_FQRL:
849 				/* Lookup in the retirement table */
850 				fq = table_find_fq(p,
851 						   be32_to_cpu(msg->fq.fqid));
852 				DPAA_BUG_ON(!fq);
853 				fq_state_change(p, fq, &swapped_msg, verb);
854 				if (fq->cb.fqs)
855 					fq->cb.fqs(p, fq, &swapped_msg);
856 				break;
857 			case QM_MR_VERB_FQPN:
858 				/* Parked */
859 #ifdef CONFIG_FSL_QMAN_FQ_LOOKUP
860 				fq = get_fq_table_entry(msg->fq.contextB);
861 #else
862 				fq = (void *)(uintptr_t)msg->fq.contextB;
863 #endif
864 				fq_state_change(p, fq, msg, verb);
865 				if (fq->cb.fqs)
866 					fq->cb.fqs(p, fq, &swapped_msg);
867 				break;
868 			case QM_MR_VERB_DC_ERN:
869 				/* DCP ERN */
870 				if (p->cb_dc_ern)
871 					p->cb_dc_ern(p, msg);
872 				else if (cb_dc_ern)
873 					cb_dc_ern(p, msg);
874 				else {
875 					static int warn_once;
876 
877 					if (!warn_once) {
878 						pr_crit("Leaking DCP ERNs!\n");
879 						warn_once = 1;
880 					}
881 				}
882 				break;
883 			default:
884 				pr_crit("Invalid MR verb 0x%02x\n", verb);
885 			}
886 		} else {
887 			/* Its a software ERN */
888 #ifdef CONFIG_FSL_QMAN_FQ_LOOKUP
889 			fq = get_fq_table_entry(be32_to_cpu(msg->ern.tag));
890 #else
891 			fq = (void *)(uintptr_t)be32_to_cpu(msg->ern.tag);
892 #endif
893 			fq->cb.ern(p, fq, &swapped_msg);
894 		}
895 		num++;
896 		qm_mr_next(&p->p);
897 		goto mr_loop;
898 mr_done:
899 		qm_mr_cci_consume(&p->p, num);
900 	}
901 	/*
902 	 * QM_PIRQ_CSCI/CCSCI has already been cleared, as part of its specific
903 	 * processing. If that interrupt source has meanwhile been re-asserted,
904 	 * we mustn't clear it here (or in the top-level interrupt handler).
905 	 */
906 	return is & (QM_PIRQ_EQCI | QM_PIRQ_EQRI | QM_PIRQ_MRI);
907 }
908 
909 /*
910  * remove some slowish-path stuff from the "fast path" and make sure it isn't
911  * inlined.
912  */
913 static noinline void clear_vdqcr(struct qman_portal *p, struct qman_fq *fq)
914 {
915 	p->vdqcr_owned = NULL;
916 	FQLOCK(fq);
917 	fq_clear(fq, QMAN_FQ_STATE_VDQCR);
918 	FQUNLOCK(fq);
919 	wake_up(&affine_queue);
920 }
921 
922 /*
923  * The only states that would conflict with other things if they ran at the
924  * same time on the same cpu are:
925  *
926  *   (i) setting/clearing vdqcr_owned, and
927  *  (ii) clearing the NE (Not Empty) flag.
928  *
929  * Both are safe. Because;
930  *
931  *   (i) this clearing can only occur after qman_set_vdq() has set the
932  *	 vdqcr_owned field (which it does before setting VDQCR), and
933  *	 qman_volatile_dequeue() blocks interrupts and preemption while this is
934  *	 done so that we can't interfere.
935  *  (ii) the NE flag is only cleared after qman_retire_fq() has set it, and as
936  *	 with (i) that API prevents us from interfering until it's safe.
937  *
938  * The good thing is that qman_set_vdq() and qman_retire_fq() run far
939  * less frequently (ie. per-FQ) than __poll_portal_fast() does, so the nett
940  * advantage comes from this function not having to "lock" anything at all.
941  *
942  * Note also that the callbacks are invoked at points which are safe against the
943  * above potential conflicts, but that this function itself is not re-entrant
944  * (this is because the function tracks one end of each FIFO in the portal and
945  * we do *not* want to lock that). So the consequence is that it is safe for
946  * user callbacks to call into any QMan API.
947  */
948 static inline unsigned int __poll_portal_fast(struct qman_portal *p,
949 					      unsigned int poll_limit)
950 {
951 	const struct qm_dqrr_entry *dq;
952 	struct qman_fq *fq;
953 	enum qman_cb_dqrr_result res;
954 	unsigned int limit = 0;
955 #if __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__
956 	struct qm_dqrr_entry *shadow;
957 #endif
958 	do {
959 		qm_dqrr_pvb_update(&p->p);
960 		dq = qm_dqrr_current(&p->p);
961 		if (unlikely(!dq))
962 			break;
963 #if __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__
964 	/* If running on an LE system the fields of the
965 	 * dequeue entry must be swapper.  Because the
966 	 * QMan HW will ignore writes the DQRR entry is
967 	 * copied and the index stored within the copy
968 	 */
969 		shadow = &p->shadow_dqrr[DQRR_PTR2IDX(dq)];
970 		*shadow = *dq;
971 		dq = shadow;
972 		shadow->fqid = be32_to_cpu(shadow->fqid);
973 		shadow->seqnum = be16_to_cpu(shadow->seqnum);
974 		hw_fd_to_cpu(&shadow->fd);
975 #endif
976 
977 		if (dq->stat & QM_DQRR_STAT_UNSCHEDULED) {
978 			/*
979 			 * VDQCR: don't trust context_b as the FQ may have
980 			 * been configured for h/w consumption and we're
981 			 * draining it post-retirement.
982 			 */
983 			fq = p->vdqcr_owned;
984 			/*
985 			 * We only set QMAN_FQ_STATE_NE when retiring, so we
986 			 * only need to check for clearing it when doing
987 			 * volatile dequeues.  It's one less thing to check
988 			 * in the critical path (SDQCR).
989 			 */
990 			if (dq->stat & QM_DQRR_STAT_FQ_EMPTY)
991 				fq_clear(fq, QMAN_FQ_STATE_NE);
992 			/*
993 			 * This is duplicated from the SDQCR code, but we
994 			 * have stuff to do before *and* after this callback,
995 			 * and we don't want multiple if()s in the critical
996 			 * path (SDQCR).
997 			 */
998 			res = fq->cb.dqrr(p, fq, dq);
999 			if (res == qman_cb_dqrr_stop)
1000 				break;
1001 			/* Check for VDQCR completion */
1002 			if (dq->stat & QM_DQRR_STAT_DQCR_EXPIRED)
1003 				clear_vdqcr(p, fq);
1004 		} else {
1005 			/* SDQCR: context_b points to the FQ */
1006 #ifdef CONFIG_FSL_QMAN_FQ_LOOKUP
1007 			fq = get_fq_table_entry(dq->contextB);
1008 #else
1009 			fq = (void *)(uintptr_t)dq->contextB;
1010 #endif
1011 			/* Now let the callback do its stuff */
1012 			res = fq->cb.dqrr(p, fq, dq);
1013 			/*
1014 			 * The callback can request that we exit without
1015 			 * consuming this entry nor advancing;
1016 			 */
1017 			if (res == qman_cb_dqrr_stop)
1018 				break;
1019 		}
1020 		/* Interpret 'dq' from a driver perspective. */
1021 		/*
1022 		 * Parking isn't possible unless HELDACTIVE was set. NB,
1023 		 * FORCEELIGIBLE implies HELDACTIVE, so we only need to
1024 		 * check for HELDACTIVE to cover both.
1025 		 */
1026 		DPAA_ASSERT((dq->stat & QM_DQRR_STAT_FQ_HELDACTIVE) ||
1027 			    (res != qman_cb_dqrr_park));
1028 		/* just means "skip it, I'll consume it myself later on" */
1029 		if (res != qman_cb_dqrr_defer)
1030 			qm_dqrr_cdc_consume_1ptr(&p->p, dq,
1031 						 res == qman_cb_dqrr_park);
1032 		/* Move forward */
1033 		qm_dqrr_next(&p->p);
1034 		/*
1035 		 * Entry processed and consumed, increment our counter.  The
1036 		 * callback can request that we exit after consuming the
1037 		 * entry, and we also exit if we reach our processing limit,
1038 		 * so loop back only if neither of these conditions is met.
1039 		 */
1040 	} while (++limit < poll_limit && res != qman_cb_dqrr_consume_stop);
1041 
1042 	return limit;
1043 }
1044 
1045 int qman_irqsource_add(u32 bits)
1046 {
1047 	struct qman_portal *p = get_affine_portal();
1048 
1049 	bits = bits & QM_PIRQ_VISIBLE;
1050 
1051 	/* Clear any previously remaining interrupt conditions in
1052 	 * QCSP_ISR. This prevents raising a false interrupt when
1053 	 * interrupt conditions are enabled in QCSP_IER.
1054 	 */
1055 	qm_isr_status_clear(&p->p, bits);
1056 	dpaa_set_bits(bits, &p->irq_sources);
1057 	qm_isr_enable_write(&p->p, p->irq_sources);
1058 
1059 
1060 	return 0;
1061 }
1062 
1063 int qman_irqsource_remove(u32 bits)
1064 {
1065 	struct qman_portal *p = get_affine_portal();
1066 	u32 ier;
1067 
1068 	/* Our interrupt handler only processes+clears status register bits that
1069 	 * are in p->irq_sources. As we're trimming that mask, if one of them
1070 	 * were to assert in the status register just before we remove it from
1071 	 * the enable register, there would be an interrupt-storm when we
1072 	 * release the IRQ lock. So we wait for the enable register update to
1073 	 * take effect in h/w (by reading it back) and then clear all other bits
1074 	 * in the status register. Ie. we clear them from ISR once it's certain
1075 	 * IER won't allow them to reassert.
1076 	 */
1077 
1078 	bits &= QM_PIRQ_VISIBLE;
1079 	dpaa_clear_bits(bits, &p->irq_sources);
1080 	qm_isr_enable_write(&p->p, p->irq_sources);
1081 	ier = qm_isr_enable_read(&p->p);
1082 	/* Using "~ier" (rather than "bits" or "~p->irq_sources") creates a
1083 	 * data-dependency, ie. to protect against re-ordering.
1084 	 */
1085 	qm_isr_status_clear(&p->p, ~ier);
1086 	return 0;
1087 }
1088 
1089 u16 qman_affine_channel(int cpu)
1090 {
1091 	if (cpu < 0) {
1092 		struct qman_portal *portal = get_affine_portal();
1093 
1094 		cpu = portal->config->cpu;
1095 	}
1096 	DPAA_BUG_ON(!CPU_ISSET(cpu, &affine_mask));
1097 	return affine_channels[cpu];
1098 }
1099 
1100 unsigned int qman_portal_poll_rx(unsigned int poll_limit,
1101 				 void **bufs,
1102 				 struct qman_portal *p)
1103 {
1104 	struct qm_portal *portal = &p->p;
1105 	register struct qm_dqrr *dqrr = &portal->dqrr;
1106 	struct qm_dqrr_entry *dq[QM_DQRR_SIZE], *shadow[QM_DQRR_SIZE];
1107 	struct qman_fq *fq;
1108 	unsigned int limit = 0, rx_number = 0;
1109 	uint32_t consume = 0;
1110 
1111 	do {
1112 		qm_dqrr_pvb_update(&p->p);
1113 		if (!dqrr->fill)
1114 			break;
1115 
1116 		dq[rx_number] = dqrr->cursor;
1117 		dqrr->cursor = DQRR_CARRYCLEAR(dqrr->cursor + 1);
1118 		/* Prefetch the next DQRR entry */
1119 		rte_prefetch0(dqrr->cursor);
1120 
1121 #if __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__
1122 		/* If running on an LE system the fields of the
1123 		 * dequeue entry must be swapper.  Because the
1124 		 * QMan HW will ignore writes the DQRR entry is
1125 		 * copied and the index stored within the copy
1126 		 */
1127 		shadow[rx_number] =
1128 			&p->shadow_dqrr[DQRR_PTR2IDX(dq[rx_number])];
1129 		shadow[rx_number]->fd.opaque_addr =
1130 			dq[rx_number]->fd.opaque_addr;
1131 		shadow[rx_number]->fd.addr =
1132 			be40_to_cpu(dq[rx_number]->fd.addr);
1133 		shadow[rx_number]->fd.opaque =
1134 			be32_to_cpu(dq[rx_number]->fd.opaque);
1135 #else
1136 		shadow[rx_number] = dq[rx_number];
1137 #endif
1138 
1139 		/* SDQCR: context_b points to the FQ */
1140 #ifdef CONFIG_FSL_QMAN_FQ_LOOKUP
1141 		fq = qman_fq_lookup_table[dq[rx_number]->contextB];
1142 #else
1143 		fq = (void *)dq[rx_number]->contextB;
1144 #endif
1145 		if (fq->cb.dqrr_prepare)
1146 			fq->cb.dqrr_prepare(shadow[rx_number],
1147 					    &bufs[rx_number]);
1148 
1149 		consume |= (1 << (31 - DQRR_PTR2IDX(shadow[rx_number])));
1150 		rx_number++;
1151 		--dqrr->fill;
1152 	} while (++limit < poll_limit);
1153 
1154 	if (rx_number)
1155 		fq->cb.dqrr_dpdk_pull_cb(&fq, shadow, bufs, rx_number);
1156 
1157 	/* Consume all the DQRR enries together */
1158 	qm_out(DQRR_DCAP, (1 << 8) | consume);
1159 
1160 	return rx_number;
1161 }
1162 
1163 void qman_clear_irq(void)
1164 {
1165 	struct qman_portal *p = get_affine_portal();
1166 	u32 clear = QM_DQAVAIL_MASK | (p->irq_sources &
1167 		~(QM_PIRQ_CSCI | QM_PIRQ_CCSCI));
1168 	qm_isr_status_clear(&p->p, clear);
1169 }
1170 
1171 u32 qman_portal_dequeue(struct rte_event ev[], unsigned int poll_limit,
1172 			void **bufs)
1173 {
1174 	const struct qm_dqrr_entry *dq;
1175 	struct qman_fq *fq;
1176 	enum qman_cb_dqrr_result res;
1177 	unsigned int limit = 0;
1178 	struct qman_portal *p = get_affine_portal();
1179 #if RTE_BYTE_ORDER == RTE_LITTLE_ENDIAN
1180 	struct qm_dqrr_entry *shadow;
1181 #endif
1182 	unsigned int rx_number = 0;
1183 
1184 	do {
1185 		qm_dqrr_pvb_update(&p->p);
1186 		dq = qm_dqrr_current(&p->p);
1187 		if (!dq)
1188 			break;
1189 #if RTE_BYTE_ORDER == RTE_LITTLE_ENDIAN
1190 		/*
1191 		 * If running on an LE system the fields of the
1192 		 * dequeue entry must be swapper.  Because the
1193 		 * QMan HW will ignore writes the DQRR entry is
1194 		 * copied and the index stored within the copy
1195 		 */
1196 		shadow = &p->shadow_dqrr[DQRR_PTR2IDX(dq)];
1197 		*shadow = *dq;
1198 		dq = shadow;
1199 		shadow->fqid = be32_to_cpu(shadow->fqid);
1200 		shadow->seqnum = be16_to_cpu(shadow->seqnum);
1201 		hw_fd_to_cpu(&shadow->fd);
1202 #endif
1203 
1204 	       /* SDQCR: context_b points to the FQ */
1205 #ifdef CONFIG_FSL_QMAN_FQ_LOOKUP
1206 		fq = get_fq_table_entry(dq->contextB);
1207 #else
1208 		fq = (void *)(uintptr_t)dq->contextB;
1209 #endif
1210 		/* Now let the callback do its stuff */
1211 		res = fq->cb.dqrr_dpdk_cb(&ev[rx_number], p, fq,
1212 					 dq, &bufs[rx_number]);
1213 		rx_number++;
1214 		/* Interpret 'dq' from a driver perspective. */
1215 		/*
1216 		 * Parking isn't possible unless HELDACTIVE was set. NB,
1217 		 * FORCEELIGIBLE implies HELDACTIVE, so we only need to
1218 		 * check for HELDACTIVE to cover both.
1219 		 */
1220 		DPAA_ASSERT((dq->stat & QM_DQRR_STAT_FQ_HELDACTIVE) ||
1221 			    (res != qman_cb_dqrr_park));
1222 		if (res != qman_cb_dqrr_defer)
1223 			qm_dqrr_cdc_consume_1ptr(&p->p, dq,
1224 						 res == qman_cb_dqrr_park);
1225 		/* Move forward */
1226 		qm_dqrr_next(&p->p);
1227 		/*
1228 		 * Entry processed and consumed, increment our counter.  The
1229 		 * callback can request that we exit after consuming the
1230 		 * entry, and we also exit if we reach our processing limit,
1231 		 * so loop back only if neither of these conditions is met.
1232 		 */
1233 	} while (++limit < poll_limit);
1234 
1235 	return limit;
1236 }
1237 
1238 struct qm_dqrr_entry *qman_dequeue(struct qman_fq *fq)
1239 {
1240 	struct qman_portal *p = get_affine_portal();
1241 	const struct qm_dqrr_entry *dq;
1242 #if __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__
1243 	struct qm_dqrr_entry *shadow;
1244 #endif
1245 
1246 	qm_dqrr_pvb_update(&p->p);
1247 	dq = qm_dqrr_current(&p->p);
1248 	if (!dq)
1249 		return NULL;
1250 
1251 	if (!(dq->stat & QM_DQRR_STAT_FD_VALID)) {
1252 		/* Invalid DQRR - put the portal and consume the DQRR.
1253 		 * Return NULL to user as no packet is seen.
1254 		 */
1255 		qman_dqrr_consume(fq, (struct qm_dqrr_entry *)dq);
1256 		return NULL;
1257 	}
1258 
1259 #if __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__
1260 	shadow = &p->shadow_dqrr[DQRR_PTR2IDX(dq)];
1261 	*shadow = *dq;
1262 	dq = shadow;
1263 	shadow->fqid = be32_to_cpu(shadow->fqid);
1264 	shadow->seqnum = be16_to_cpu(shadow->seqnum);
1265 	hw_fd_to_cpu(&shadow->fd);
1266 #endif
1267 
1268 	if (dq->stat & QM_DQRR_STAT_FQ_EMPTY)
1269 		fq_clear(fq, QMAN_FQ_STATE_NE);
1270 
1271 	return (struct qm_dqrr_entry *)dq;
1272 }
1273 
1274 void qman_dqrr_consume(struct qman_fq *fq,
1275 		       struct qm_dqrr_entry *dq)
1276 {
1277 	struct qman_portal *p = get_affine_portal();
1278 
1279 	if (dq->stat & QM_DQRR_STAT_DQCR_EXPIRED)
1280 		clear_vdqcr(p, fq);
1281 
1282 	qm_dqrr_cdc_consume_1ptr(&p->p, dq, 0);
1283 	qm_dqrr_next(&p->p);
1284 }
1285 
1286 int qman_poll_dqrr(unsigned int limit)
1287 {
1288 	struct qman_portal *p = get_affine_portal();
1289 	int ret;
1290 
1291 	ret = __poll_portal_fast(p, limit);
1292 	return ret;
1293 }
1294 
1295 void qman_poll(void)
1296 {
1297 	struct qman_portal *p = get_affine_portal();
1298 
1299 	if ((~p->irq_sources) & QM_PIRQ_SLOW) {
1300 		if (!(p->slowpoll--)) {
1301 			u32 is = qm_isr_status_read(&p->p) & ~p->irq_sources;
1302 			u32 active = __poll_portal_slow(p, is);
1303 
1304 			if (active) {
1305 				qm_isr_status_clear(&p->p, active);
1306 				p->slowpoll = SLOW_POLL_BUSY;
1307 			} else
1308 				p->slowpoll = SLOW_POLL_IDLE;
1309 		}
1310 	}
1311 	if ((~p->irq_sources) & QM_PIRQ_DQRI)
1312 		__poll_portal_fast(p, FSL_QMAN_POLL_LIMIT);
1313 }
1314 
1315 void qman_stop_dequeues(void)
1316 {
1317 	struct qman_portal *p = get_affine_portal();
1318 
1319 	qman_stop_dequeues_ex(p);
1320 }
1321 
1322 void qman_start_dequeues(void)
1323 {
1324 	struct qman_portal *p = get_affine_portal();
1325 
1326 	DPAA_ASSERT(p->dqrr_disable_ref > 0);
1327 	if (!(--p->dqrr_disable_ref))
1328 		qm_dqrr_set_maxfill(&p->p, DQRR_MAXFILL);
1329 }
1330 
1331 void qman_static_dequeue_add(u32 pools, struct qman_portal *qp)
1332 {
1333 	struct qman_portal *p = qp ? qp : get_affine_portal();
1334 
1335 	pools &= p->config->pools;
1336 	p->sdqcr |= pools;
1337 	qm_dqrr_sdqcr_set(&p->p, p->sdqcr);
1338 }
1339 
1340 void qman_static_dequeue_del(u32 pools, struct qman_portal *qp)
1341 {
1342 	struct qman_portal *p = qp ? qp : get_affine_portal();
1343 
1344 	pools &= p->config->pools;
1345 	p->sdqcr &= ~pools;
1346 	qm_dqrr_sdqcr_set(&p->p, p->sdqcr);
1347 }
1348 
1349 u32 qman_static_dequeue_get(struct qman_portal *qp)
1350 {
1351 	struct qman_portal *p = qp ? qp : get_affine_portal();
1352 	return p->sdqcr;
1353 }
1354 
1355 void qman_dca(const struct qm_dqrr_entry *dq, int park_request)
1356 {
1357 	struct qman_portal *p = get_affine_portal();
1358 
1359 	qm_dqrr_cdc_consume_1ptr(&p->p, dq, park_request);
1360 }
1361 
1362 void qman_dca_index(u8 index, int park_request)
1363 {
1364 	struct qman_portal *p = get_affine_portal();
1365 
1366 	qm_dqrr_cdc_consume_1(&p->p, index, park_request);
1367 }
1368 
1369 /* Frame queue API */
1370 static const char *mcr_result_str(u8 result)
1371 {
1372 	switch (result) {
1373 	case QM_MCR_RESULT_NULL:
1374 		return "QM_MCR_RESULT_NULL";
1375 	case QM_MCR_RESULT_OK:
1376 		return "QM_MCR_RESULT_OK";
1377 	case QM_MCR_RESULT_ERR_FQID:
1378 		return "QM_MCR_RESULT_ERR_FQID";
1379 	case QM_MCR_RESULT_ERR_FQSTATE:
1380 		return "QM_MCR_RESULT_ERR_FQSTATE";
1381 	case QM_MCR_RESULT_ERR_NOTEMPTY:
1382 		return "QM_MCR_RESULT_ERR_NOTEMPTY";
1383 	case QM_MCR_RESULT_PENDING:
1384 		return "QM_MCR_RESULT_PENDING";
1385 	case QM_MCR_RESULT_ERR_BADCOMMAND:
1386 		return "QM_MCR_RESULT_ERR_BADCOMMAND";
1387 	}
1388 	return "<unknown MCR result>";
1389 }
1390 
1391 int qman_create_fq(u32 fqid, u32 flags, struct qman_fq *fq)
1392 {
1393 	struct qm_fqd fqd;
1394 	struct qm_mcr_queryfq_np np;
1395 	struct qm_mc_command *mcc;
1396 	struct qm_mc_result *mcr;
1397 	struct qman_portal *p;
1398 
1399 	if (flags & QMAN_FQ_FLAG_DYNAMIC_FQID) {
1400 		int ret = qman_alloc_fqid(&fqid);
1401 
1402 		if (ret)
1403 			return ret;
1404 	}
1405 	spin_lock_init(&fq->fqlock);
1406 	fq->fqid = fqid;
1407 	fq->fqid_le = cpu_to_be32(fqid);
1408 	fq->flags = flags;
1409 	fq->state = qman_fq_state_oos;
1410 	fq->cgr_groupid = 0;
1411 #ifdef CONFIG_FSL_QMAN_FQ_LOOKUP
1412 	if (unlikely(find_empty_fq_table_entry(&fq->key, fq))) {
1413 		pr_info("Find empty table entry failed\n");
1414 		return -ENOMEM;
1415 	}
1416 	fq->qman_fq_lookup_table = qman_fq_lookup_table;
1417 #endif
1418 	if (!(flags & QMAN_FQ_FLAG_AS_IS) || (flags & QMAN_FQ_FLAG_NO_MODIFY))
1419 		return 0;
1420 	/* Everything else is AS_IS support */
1421 	p = get_affine_portal();
1422 	mcc = qm_mc_start(&p->p);
1423 	mcc->queryfq.fqid = cpu_to_be32(fqid);
1424 	qm_mc_commit(&p->p, QM_MCC_VERB_QUERYFQ);
1425 	while (!(mcr = qm_mc_result(&p->p)))
1426 		cpu_relax();
1427 	DPAA_ASSERT((mcr->verb & QM_MCR_VERB_MASK) == QM_MCC_VERB_QUERYFQ);
1428 	if (mcr->result != QM_MCR_RESULT_OK) {
1429 		pr_err("QUERYFQ failed: %s\n", mcr_result_str(mcr->result));
1430 		goto err;
1431 	}
1432 	fqd = mcr->queryfq.fqd;
1433 	hw_fqd_to_cpu(&fqd);
1434 	mcc = qm_mc_start(&p->p);
1435 	mcc->queryfq_np.fqid = cpu_to_be32(fqid);
1436 	qm_mc_commit(&p->p, QM_MCC_VERB_QUERYFQ_NP);
1437 	while (!(mcr = qm_mc_result(&p->p)))
1438 		cpu_relax();
1439 	DPAA_ASSERT((mcr->verb & QM_MCR_VERB_MASK) == QM_MCC_VERB_QUERYFQ_NP);
1440 	if (mcr->result != QM_MCR_RESULT_OK) {
1441 		pr_err("QUERYFQ_NP failed: %s\n", mcr_result_str(mcr->result));
1442 		goto err;
1443 	}
1444 	np = mcr->queryfq_np;
1445 	/* Phew, have queryfq and queryfq_np results, stitch together
1446 	 * the FQ object from those.
1447 	 */
1448 	fq->cgr_groupid = fqd.cgid;
1449 	switch (np.state & QM_MCR_NP_STATE_MASK) {
1450 	case QM_MCR_NP_STATE_OOS:
1451 		break;
1452 	case QM_MCR_NP_STATE_RETIRED:
1453 		fq->state = qman_fq_state_retired;
1454 		if (np.frm_cnt)
1455 			fq_set(fq, QMAN_FQ_STATE_NE);
1456 		break;
1457 	case QM_MCR_NP_STATE_TEN_SCHED:
1458 	case QM_MCR_NP_STATE_TRU_SCHED:
1459 	case QM_MCR_NP_STATE_ACTIVE:
1460 		fq->state = qman_fq_state_sched;
1461 		if (np.state & QM_MCR_NP_STATE_R)
1462 			fq_set(fq, QMAN_FQ_STATE_CHANGING);
1463 		break;
1464 	case QM_MCR_NP_STATE_PARKED:
1465 		fq->state = qman_fq_state_parked;
1466 		break;
1467 	default:
1468 		DPAA_ASSERT(NULL == "invalid FQ state");
1469 	}
1470 	if (fqd.fq_ctrl & QM_FQCTRL_CGE)
1471 		fq->state |= QMAN_FQ_STATE_CGR_EN;
1472 	return 0;
1473 err:
1474 	if (flags & QMAN_FQ_FLAG_DYNAMIC_FQID)
1475 		qman_release_fqid(fqid);
1476 	return -EIO;
1477 }
1478 
1479 void qman_destroy_fq(struct qman_fq *fq, u32 flags __maybe_unused)
1480 {
1481 	/*
1482 	 * We don't need to lock the FQ as it is a pre-condition that the FQ be
1483 	 * quiesced. Instead, run some checks.
1484 	 */
1485 	switch (fq->state) {
1486 	case qman_fq_state_parked:
1487 		DPAA_ASSERT(flags & QMAN_FQ_DESTROY_PARKED);
1488 		/* Fallthrough */
1489 	case qman_fq_state_oos:
1490 		if (fq_isset(fq, QMAN_FQ_FLAG_DYNAMIC_FQID))
1491 			qman_release_fqid(fq->fqid);
1492 #ifdef CONFIG_FSL_QMAN_FQ_LOOKUP
1493 		clear_fq_table_entry(fq->key);
1494 #endif
1495 		return;
1496 	default:
1497 		break;
1498 	}
1499 	DPAA_ASSERT(NULL == "qman_free_fq() on unquiesced FQ!");
1500 }
1501 
1502 u32 qman_fq_fqid(struct qman_fq *fq)
1503 {
1504 	return fq->fqid;
1505 }
1506 
1507 void qman_fq_state(struct qman_fq *fq, enum qman_fq_state *state, u32 *flags)
1508 {
1509 	if (state)
1510 		*state = fq->state;
1511 	if (flags)
1512 		*flags = fq->flags;
1513 }
1514 
1515 int qman_init_fq(struct qman_fq *fq, u32 flags, struct qm_mcc_initfq *opts)
1516 {
1517 	struct qm_mc_command *mcc;
1518 	struct qm_mc_result *mcr;
1519 	struct qman_portal *p;
1520 
1521 	u8 res, myverb = (flags & QMAN_INITFQ_FLAG_SCHED) ?
1522 		QM_MCC_VERB_INITFQ_SCHED : QM_MCC_VERB_INITFQ_PARKED;
1523 
1524 	if ((fq->state != qman_fq_state_oos) &&
1525 	    (fq->state != qman_fq_state_parked))
1526 		return -EINVAL;
1527 #ifdef RTE_LIBRTE_DPAA_HWDEBUG
1528 	if (unlikely(fq_isset(fq, QMAN_FQ_FLAG_NO_MODIFY)))
1529 		return -EINVAL;
1530 #endif
1531 	if (opts && (opts->we_mask & QM_INITFQ_WE_OAC)) {
1532 		/* And can't be set at the same time as TDTHRESH */
1533 		if (opts->we_mask & QM_INITFQ_WE_TDTHRESH)
1534 			return -EINVAL;
1535 	}
1536 	/* Issue an INITFQ_[PARKED|SCHED] management command */
1537 	p = get_affine_portal();
1538 	FQLOCK(fq);
1539 	if (unlikely((fq_isset(fq, QMAN_FQ_STATE_CHANGING)) ||
1540 		     ((fq->state != qman_fq_state_oos) &&
1541 				(fq->state != qman_fq_state_parked)))) {
1542 		FQUNLOCK(fq);
1543 		return -EBUSY;
1544 	}
1545 	mcc = qm_mc_start(&p->p);
1546 	if (opts)
1547 		mcc->initfq = *opts;
1548 	mcc->initfq.fqid = cpu_to_be32(fq->fqid);
1549 	mcc->initfq.count = 0;
1550 	/*
1551 	 * If the FQ does *not* have the TO_DCPORTAL flag, context_b is set as a
1552 	 * demux pointer. Otherwise, the caller-provided value is allowed to
1553 	 * stand, don't overwrite it.
1554 	 */
1555 	if (fq_isclear(fq, QMAN_FQ_FLAG_TO_DCPORTAL)) {
1556 		dma_addr_t phys_fq;
1557 
1558 		mcc->initfq.we_mask |= QM_INITFQ_WE_CONTEXTB;
1559 #ifdef CONFIG_FSL_QMAN_FQ_LOOKUP
1560 		mcc->initfq.fqd.context_b = cpu_to_be32(fq->key);
1561 #else
1562 		mcc->initfq.fqd.context_b = (u32)(uintptr_t)fq;
1563 #endif
1564 		/*
1565 		 *  and the physical address - NB, if the user wasn't trying to
1566 		 * set CONTEXTA, clear the stashing settings.
1567 		 */
1568 		if (!(mcc->initfq.we_mask & QM_INITFQ_WE_CONTEXTA)) {
1569 			mcc->initfq.we_mask |= QM_INITFQ_WE_CONTEXTA;
1570 			memset(&mcc->initfq.fqd.context_a, 0,
1571 			       sizeof(mcc->initfq.fqd.context_a));
1572 		} else {
1573 			phys_fq = rte_mem_virt2iova(fq);
1574 			qm_fqd_stashing_set64(&mcc->initfq.fqd, phys_fq);
1575 		}
1576 	}
1577 	if (flags & QMAN_INITFQ_FLAG_LOCAL) {
1578 		mcc->initfq.fqd.dest.channel = p->config->channel;
1579 		if (!(mcc->initfq.we_mask & QM_INITFQ_WE_DESTWQ)) {
1580 			mcc->initfq.we_mask |= QM_INITFQ_WE_DESTWQ;
1581 			mcc->initfq.fqd.dest.wq = 4;
1582 		}
1583 	}
1584 	mcc->initfq.we_mask = cpu_to_be16(mcc->initfq.we_mask);
1585 	cpu_to_hw_fqd(&mcc->initfq.fqd);
1586 	qm_mc_commit(&p->p, myverb);
1587 	while (!(mcr = qm_mc_result(&p->p)))
1588 		cpu_relax();
1589 	DPAA_ASSERT((mcr->verb & QM_MCR_VERB_MASK) == myverb);
1590 	res = mcr->result;
1591 	if (res != QM_MCR_RESULT_OK) {
1592 		FQUNLOCK(fq);
1593 		return -EIO;
1594 	}
1595 	if (opts) {
1596 		if (opts->we_mask & QM_INITFQ_WE_FQCTRL) {
1597 			if (opts->fqd.fq_ctrl & QM_FQCTRL_CGE)
1598 				fq_set(fq, QMAN_FQ_STATE_CGR_EN);
1599 			else
1600 				fq_clear(fq, QMAN_FQ_STATE_CGR_EN);
1601 		}
1602 		if (opts->we_mask & QM_INITFQ_WE_CGID)
1603 			fq->cgr_groupid = opts->fqd.cgid;
1604 	}
1605 	fq->state = (flags & QMAN_INITFQ_FLAG_SCHED) ?
1606 		qman_fq_state_sched : qman_fq_state_parked;
1607 	FQUNLOCK(fq);
1608 	return 0;
1609 }
1610 
1611 int qman_schedule_fq(struct qman_fq *fq)
1612 {
1613 	struct qm_mc_command *mcc;
1614 	struct qm_mc_result *mcr;
1615 	struct qman_portal *p;
1616 
1617 	int ret = 0;
1618 	u8 res;
1619 
1620 	if (fq->state != qman_fq_state_parked)
1621 		return -EINVAL;
1622 #ifdef RTE_LIBRTE_DPAA_HWDEBUG
1623 	if (unlikely(fq_isset(fq, QMAN_FQ_FLAG_NO_MODIFY)))
1624 		return -EINVAL;
1625 #endif
1626 	/* Issue a ALTERFQ_SCHED management command */
1627 	p = get_affine_portal();
1628 
1629 	FQLOCK(fq);
1630 	if (unlikely((fq_isset(fq, QMAN_FQ_STATE_CHANGING)) ||
1631 		     (fq->state != qman_fq_state_parked))) {
1632 		ret = -EBUSY;
1633 		goto out;
1634 	}
1635 	mcc = qm_mc_start(&p->p);
1636 	mcc->alterfq.fqid = cpu_to_be32(fq->fqid);
1637 	qm_mc_commit(&p->p, QM_MCC_VERB_ALTER_SCHED);
1638 	while (!(mcr = qm_mc_result(&p->p)))
1639 		cpu_relax();
1640 	DPAA_ASSERT((mcr->verb & QM_MCR_VERB_MASK) == QM_MCR_VERB_ALTER_SCHED);
1641 	res = mcr->result;
1642 	if (res != QM_MCR_RESULT_OK) {
1643 		ret = -EIO;
1644 		goto out;
1645 	}
1646 	fq->state = qman_fq_state_sched;
1647 out:
1648 	FQUNLOCK(fq);
1649 
1650 	return ret;
1651 }
1652 
1653 int qman_retire_fq(struct qman_fq *fq, u32 *flags)
1654 {
1655 	struct qm_mc_command *mcc;
1656 	struct qm_mc_result *mcr;
1657 	struct qman_portal *p;
1658 
1659 	int rval;
1660 	u8 res;
1661 
1662 	if ((fq->state != qman_fq_state_parked) &&
1663 	    (fq->state != qman_fq_state_sched))
1664 		return -EINVAL;
1665 #ifdef RTE_LIBRTE_DPAA_HWDEBUG
1666 	if (unlikely(fq_isset(fq, QMAN_FQ_FLAG_NO_MODIFY)))
1667 		return -EINVAL;
1668 #endif
1669 	p = get_affine_portal();
1670 
1671 	FQLOCK(fq);
1672 	if (unlikely((fq_isset(fq, QMAN_FQ_STATE_CHANGING)) ||
1673 		     (fq->state == qman_fq_state_retired) ||
1674 				(fq->state == qman_fq_state_oos))) {
1675 		rval = -EBUSY;
1676 		goto out;
1677 	}
1678 	rval = table_push_fq(p, fq);
1679 	if (rval)
1680 		goto out;
1681 	mcc = qm_mc_start(&p->p);
1682 	mcc->alterfq.fqid = cpu_to_be32(fq->fqid);
1683 	qm_mc_commit(&p->p, QM_MCC_VERB_ALTER_RETIRE);
1684 	while (!(mcr = qm_mc_result(&p->p)))
1685 		cpu_relax();
1686 	DPAA_ASSERT((mcr->verb & QM_MCR_VERB_MASK) == QM_MCR_VERB_ALTER_RETIRE);
1687 	res = mcr->result;
1688 	/*
1689 	 * "Elegant" would be to treat OK/PENDING the same way; set CHANGING,
1690 	 * and defer the flags until FQRNI or FQRN (respectively) show up. But
1691 	 * "Friendly" is to process OK immediately, and not set CHANGING. We do
1692 	 * friendly, otherwise the caller doesn't necessarily have a fully
1693 	 * "retired" FQ on return even if the retirement was immediate. However
1694 	 * this does mean some code duplication between here and
1695 	 * fq_state_change().
1696 	 */
1697 	if (likely(res == QM_MCR_RESULT_OK)) {
1698 		rval = 0;
1699 		/* Process 'fq' right away, we'll ignore FQRNI */
1700 		if (mcr->alterfq.fqs & QM_MCR_FQS_NOTEMPTY)
1701 			fq_set(fq, QMAN_FQ_STATE_NE);
1702 		if (mcr->alterfq.fqs & QM_MCR_FQS_ORLPRESENT)
1703 			fq_set(fq, QMAN_FQ_STATE_ORL);
1704 		else
1705 			table_del_fq(p, fq);
1706 		if (flags)
1707 			*flags = fq->flags;
1708 		fq->state = qman_fq_state_retired;
1709 		if (fq->cb.fqs) {
1710 			/*
1711 			 * Another issue with supporting "immediate" retirement
1712 			 * is that we're forced to drop FQRNIs, because by the
1713 			 * time they're seen it may already be "too late" (the
1714 			 * fq may have been OOS'd and free()'d already). But if
1715 			 * the upper layer wants a callback whether it's
1716 			 * immediate or not, we have to fake a "MR" entry to
1717 			 * look like an FQRNI...
1718 			 */
1719 			struct qm_mr_entry msg;
1720 
1721 			msg.ern.verb = QM_MR_VERB_FQRNI;
1722 			msg.fq.fqs = mcr->alterfq.fqs;
1723 			msg.fq.fqid = fq->fqid;
1724 #ifdef CONFIG_FSL_QMAN_FQ_LOOKUP
1725 			msg.fq.contextB = fq->key;
1726 #else
1727 			msg.fq.contextB = (u32)(uintptr_t)fq;
1728 #endif
1729 			fq->cb.fqs(p, fq, &msg);
1730 		}
1731 	} else if (res == QM_MCR_RESULT_PENDING) {
1732 		rval = 1;
1733 		fq_set(fq, QMAN_FQ_STATE_CHANGING);
1734 	} else {
1735 		rval = -EIO;
1736 		table_del_fq(p, fq);
1737 	}
1738 out:
1739 	FQUNLOCK(fq);
1740 	return rval;
1741 }
1742 
1743 int qman_oos_fq(struct qman_fq *fq)
1744 {
1745 	struct qm_mc_command *mcc;
1746 	struct qm_mc_result *mcr;
1747 	struct qman_portal *p;
1748 
1749 	int ret = 0;
1750 	u8 res;
1751 
1752 	if (fq->state != qman_fq_state_retired)
1753 		return -EINVAL;
1754 #ifdef RTE_LIBRTE_DPAA_HWDEBUG
1755 	if (unlikely(fq_isset(fq, QMAN_FQ_FLAG_NO_MODIFY)))
1756 		return -EINVAL;
1757 #endif
1758 	p = get_affine_portal();
1759 	FQLOCK(fq);
1760 	if (unlikely((fq_isset(fq, QMAN_FQ_STATE_BLOCKOOS)) ||
1761 		     (fq->state != qman_fq_state_retired))) {
1762 		ret = -EBUSY;
1763 		goto out;
1764 	}
1765 	mcc = qm_mc_start(&p->p);
1766 	mcc->alterfq.fqid = cpu_to_be32(fq->fqid);
1767 	qm_mc_commit(&p->p, QM_MCC_VERB_ALTER_OOS);
1768 	while (!(mcr = qm_mc_result(&p->p)))
1769 		cpu_relax();
1770 	DPAA_ASSERT((mcr->verb & QM_MCR_VERB_MASK) == QM_MCR_VERB_ALTER_OOS);
1771 	res = mcr->result;
1772 	if (res != QM_MCR_RESULT_OK) {
1773 		ret = -EIO;
1774 		goto out;
1775 	}
1776 	fq->state = qman_fq_state_oos;
1777 out:
1778 	FQUNLOCK(fq);
1779 	return ret;
1780 }
1781 
1782 int qman_fq_flow_control(struct qman_fq *fq, int xon)
1783 {
1784 	struct qm_mc_command *mcc;
1785 	struct qm_mc_result *mcr;
1786 	struct qman_portal *p;
1787 
1788 	int ret = 0;
1789 	u8 res;
1790 	u8 myverb;
1791 
1792 	if ((fq->state == qman_fq_state_oos) ||
1793 	    (fq->state == qman_fq_state_retired) ||
1794 		(fq->state == qman_fq_state_parked))
1795 		return -EINVAL;
1796 
1797 #ifdef RTE_LIBRTE_DPAA_HWDEBUG
1798 	if (unlikely(fq_isset(fq, QMAN_FQ_FLAG_NO_MODIFY)))
1799 		return -EINVAL;
1800 #endif
1801 	/* Issue a ALTER_FQXON or ALTER_FQXOFF management command */
1802 	p = get_affine_portal();
1803 	FQLOCK(fq);
1804 	if (unlikely((fq_isset(fq, QMAN_FQ_STATE_CHANGING)) ||
1805 		     (fq->state == qman_fq_state_parked) ||
1806 			(fq->state == qman_fq_state_oos) ||
1807 			(fq->state == qman_fq_state_retired))) {
1808 		ret = -EBUSY;
1809 		goto out;
1810 	}
1811 	mcc = qm_mc_start(&p->p);
1812 	mcc->alterfq.fqid = fq->fqid;
1813 	mcc->alterfq.count = 0;
1814 	myverb = xon ? QM_MCC_VERB_ALTER_FQXON : QM_MCC_VERB_ALTER_FQXOFF;
1815 
1816 	qm_mc_commit(&p->p, myverb);
1817 	while (!(mcr = qm_mc_result(&p->p)))
1818 		cpu_relax();
1819 	DPAA_ASSERT((mcr->verb & QM_MCR_VERB_MASK) == myverb);
1820 
1821 	res = mcr->result;
1822 	if (res != QM_MCR_RESULT_OK) {
1823 		ret = -EIO;
1824 		goto out;
1825 	}
1826 out:
1827 	FQUNLOCK(fq);
1828 	return ret;
1829 }
1830 
1831 int qman_query_fq(struct qman_fq *fq, struct qm_fqd *fqd)
1832 {
1833 	struct qm_mc_command *mcc;
1834 	struct qm_mc_result *mcr;
1835 	struct qman_portal *p = get_affine_portal();
1836 
1837 	u8 res;
1838 
1839 	mcc = qm_mc_start(&p->p);
1840 	mcc->queryfq.fqid = cpu_to_be32(fq->fqid);
1841 	qm_mc_commit(&p->p, QM_MCC_VERB_QUERYFQ);
1842 	while (!(mcr = qm_mc_result(&p->p)))
1843 		cpu_relax();
1844 	DPAA_ASSERT((mcr->verb & QM_MCR_VERB_MASK) == QM_MCR_VERB_QUERYFQ);
1845 	res = mcr->result;
1846 	if (res == QM_MCR_RESULT_OK)
1847 		*fqd = mcr->queryfq.fqd;
1848 	hw_fqd_to_cpu(fqd);
1849 	if (res != QM_MCR_RESULT_OK)
1850 		return -EIO;
1851 	return 0;
1852 }
1853 
1854 int qman_query_fq_has_pkts(struct qman_fq *fq)
1855 {
1856 	struct qm_mc_command *mcc;
1857 	struct qm_mc_result *mcr;
1858 	struct qman_portal *p = get_affine_portal();
1859 
1860 	int ret = 0;
1861 	u8 res;
1862 
1863 	mcc = qm_mc_start(&p->p);
1864 	mcc->queryfq.fqid = cpu_to_be32(fq->fqid);
1865 	qm_mc_commit(&p->p, QM_MCC_VERB_QUERYFQ_NP);
1866 	while (!(mcr = qm_mc_result(&p->p)))
1867 		cpu_relax();
1868 	res = mcr->result;
1869 	if (res == QM_MCR_RESULT_OK)
1870 		ret = !!mcr->queryfq_np.frm_cnt;
1871 	return ret;
1872 }
1873 
1874 int qman_query_fq_np(struct qman_fq *fq, struct qm_mcr_queryfq_np *np)
1875 {
1876 	struct qm_mc_command *mcc;
1877 	struct qm_mc_result *mcr;
1878 	struct qman_portal *p = get_affine_portal();
1879 
1880 	u8 res;
1881 
1882 	mcc = qm_mc_start(&p->p);
1883 	mcc->queryfq.fqid = cpu_to_be32(fq->fqid);
1884 	qm_mc_commit(&p->p, QM_MCC_VERB_QUERYFQ_NP);
1885 	while (!(mcr = qm_mc_result(&p->p)))
1886 		cpu_relax();
1887 	DPAA_ASSERT((mcr->verb & QM_MCR_VERB_MASK) == QM_MCR_VERB_QUERYFQ_NP);
1888 	res = mcr->result;
1889 	if (res == QM_MCR_RESULT_OK) {
1890 		*np = mcr->queryfq_np;
1891 		np->fqd_link = be24_to_cpu(np->fqd_link);
1892 		np->odp_seq = be16_to_cpu(np->odp_seq);
1893 		np->orp_nesn = be16_to_cpu(np->orp_nesn);
1894 		np->orp_ea_hseq  = be16_to_cpu(np->orp_ea_hseq);
1895 		np->orp_ea_tseq  = be16_to_cpu(np->orp_ea_tseq);
1896 		np->orp_ea_hptr = be24_to_cpu(np->orp_ea_hptr);
1897 		np->orp_ea_tptr = be24_to_cpu(np->orp_ea_tptr);
1898 		np->pfdr_hptr = be24_to_cpu(np->pfdr_hptr);
1899 		np->pfdr_tptr = be24_to_cpu(np->pfdr_tptr);
1900 		np->ics_surp = be16_to_cpu(np->ics_surp);
1901 		np->byte_cnt = be32_to_cpu(np->byte_cnt);
1902 		np->frm_cnt = be24_to_cpu(np->frm_cnt);
1903 		np->ra1_sfdr = be16_to_cpu(np->ra1_sfdr);
1904 		np->ra2_sfdr = be16_to_cpu(np->ra2_sfdr);
1905 		np->od1_sfdr = be16_to_cpu(np->od1_sfdr);
1906 		np->od2_sfdr = be16_to_cpu(np->od2_sfdr);
1907 		np->od3_sfdr = be16_to_cpu(np->od3_sfdr);
1908 	}
1909 	if (res == QM_MCR_RESULT_ERR_FQID)
1910 		return -ERANGE;
1911 	else if (res != QM_MCR_RESULT_OK)
1912 		return -EIO;
1913 	return 0;
1914 }
1915 
1916 int qman_query_fq_frm_cnt(struct qman_fq *fq, u32 *frm_cnt)
1917 {
1918 	struct qm_mc_command *mcc;
1919 	struct qm_mc_result *mcr;
1920 	struct qman_portal *p = get_affine_portal();
1921 
1922 	mcc = qm_mc_start(&p->p);
1923 	mcc->queryfq.fqid = cpu_to_be32(fq->fqid);
1924 	qm_mc_commit(&p->p, QM_MCC_VERB_QUERYFQ_NP);
1925 	while (!(mcr = qm_mc_result(&p->p)))
1926 		cpu_relax();
1927 	DPAA_ASSERT((mcr->verb & QM_MCR_VERB_MASK) == QM_MCR_VERB_QUERYFQ_NP);
1928 
1929 	if (mcr->result == QM_MCR_RESULT_OK)
1930 		*frm_cnt = be24_to_cpu(mcr->queryfq_np.frm_cnt);
1931 	else if (mcr->result == QM_MCR_RESULT_ERR_FQID)
1932 		return -ERANGE;
1933 	else if (mcr->result != QM_MCR_RESULT_OK)
1934 		return -EIO;
1935 	return 0;
1936 }
1937 
1938 int qman_query_wq(u8 query_dedicated, struct qm_mcr_querywq *wq)
1939 {
1940 	struct qm_mc_command *mcc;
1941 	struct qm_mc_result *mcr;
1942 	struct qman_portal *p = get_affine_portal();
1943 
1944 	u8 res, myverb;
1945 
1946 	myverb = (query_dedicated) ? QM_MCR_VERB_QUERYWQ_DEDICATED :
1947 				 QM_MCR_VERB_QUERYWQ;
1948 	mcc = qm_mc_start(&p->p);
1949 	mcc->querywq.channel.id = cpu_to_be16(wq->channel.id);
1950 	qm_mc_commit(&p->p, myverb);
1951 	while (!(mcr = qm_mc_result(&p->p)))
1952 		cpu_relax();
1953 	DPAA_ASSERT((mcr->verb & QM_MCR_VERB_MASK) == myverb);
1954 	res = mcr->result;
1955 	if (res == QM_MCR_RESULT_OK) {
1956 		int i, array_len;
1957 
1958 		wq->channel.id = be16_to_cpu(mcr->querywq.channel.id);
1959 		array_len = ARRAY_SIZE(mcr->querywq.wq_len);
1960 		for (i = 0; i < array_len; i++)
1961 			wq->wq_len[i] = be32_to_cpu(mcr->querywq.wq_len[i]);
1962 	}
1963 	if (res != QM_MCR_RESULT_OK) {
1964 		pr_err("QUERYWQ failed: %s\n", mcr_result_str(res));
1965 		return -EIO;
1966 	}
1967 	return 0;
1968 }
1969 
1970 int qman_testwrite_cgr(struct qman_cgr *cgr, u64 i_bcnt,
1971 		       struct qm_mcr_cgrtestwrite *result)
1972 {
1973 	struct qm_mc_command *mcc;
1974 	struct qm_mc_result *mcr;
1975 	struct qman_portal *p = get_affine_portal();
1976 
1977 	u8 res;
1978 
1979 	mcc = qm_mc_start(&p->p);
1980 	mcc->cgrtestwrite.cgid = cgr->cgrid;
1981 	mcc->cgrtestwrite.i_bcnt_hi = (u8)(i_bcnt >> 32);
1982 	mcc->cgrtestwrite.i_bcnt_lo = (u32)i_bcnt;
1983 	qm_mc_commit(&p->p, QM_MCC_VERB_CGRTESTWRITE);
1984 	while (!(mcr = qm_mc_result(&p->p)))
1985 		cpu_relax();
1986 	DPAA_ASSERT((mcr->verb & QM_MCR_VERB_MASK) == QM_MCC_VERB_CGRTESTWRITE);
1987 	res = mcr->result;
1988 	if (res == QM_MCR_RESULT_OK)
1989 		*result = mcr->cgrtestwrite;
1990 	if (res != QM_MCR_RESULT_OK) {
1991 		pr_err("CGR TEST WRITE failed: %s\n", mcr_result_str(res));
1992 		return -EIO;
1993 	}
1994 	return 0;
1995 }
1996 
1997 int qman_query_cgr(struct qman_cgr *cgr, struct qm_mcr_querycgr *cgrd)
1998 {
1999 	struct qm_mc_command *mcc;
2000 	struct qm_mc_result *mcr;
2001 	struct qman_portal *p = get_affine_portal();
2002 	u8 res;
2003 	unsigned int i;
2004 
2005 	mcc = qm_mc_start(&p->p);
2006 	mcc->querycgr.cgid = cgr->cgrid;
2007 	qm_mc_commit(&p->p, QM_MCC_VERB_QUERYCGR);
2008 	while (!(mcr = qm_mc_result(&p->p)))
2009 		cpu_relax();
2010 	DPAA_ASSERT((mcr->verb & QM_MCR_VERB_MASK) == QM_MCC_VERB_QUERYCGR);
2011 	res = mcr->result;
2012 	if (res == QM_MCR_RESULT_OK)
2013 		*cgrd = mcr->querycgr;
2014 	if (res != QM_MCR_RESULT_OK) {
2015 		pr_err("QUERY_CGR failed: %s\n", mcr_result_str(res));
2016 		return -EIO;
2017 	}
2018 	cgrd->cgr.wr_parm_g.word =
2019 		be32_to_cpu(cgrd->cgr.wr_parm_g.word);
2020 	cgrd->cgr.wr_parm_y.word =
2021 		be32_to_cpu(cgrd->cgr.wr_parm_y.word);
2022 	cgrd->cgr.wr_parm_r.word =
2023 		be32_to_cpu(cgrd->cgr.wr_parm_r.word);
2024 	cgrd->cgr.cscn_targ =  be32_to_cpu(cgrd->cgr.cscn_targ);
2025 	cgrd->cgr.__cs_thres = be16_to_cpu(cgrd->cgr.__cs_thres);
2026 	for (i = 0; i < ARRAY_SIZE(cgrd->cscn_targ_swp); i++)
2027 		cgrd->cscn_targ_swp[i] =
2028 			be32_to_cpu(cgrd->cscn_targ_swp[i]);
2029 	return 0;
2030 }
2031 
2032 int qman_query_congestion(struct qm_mcr_querycongestion *congestion)
2033 {
2034 	struct qm_mc_result *mcr;
2035 	struct qman_portal *p = get_affine_portal();
2036 	u8 res;
2037 	unsigned int i;
2038 
2039 	qm_mc_start(&p->p);
2040 	qm_mc_commit(&p->p, QM_MCC_VERB_QUERYCONGESTION);
2041 	while (!(mcr = qm_mc_result(&p->p)))
2042 		cpu_relax();
2043 	DPAA_ASSERT((mcr->verb & QM_MCR_VERB_MASK) ==
2044 			QM_MCC_VERB_QUERYCONGESTION);
2045 	res = mcr->result;
2046 	if (res == QM_MCR_RESULT_OK)
2047 		*congestion = mcr->querycongestion;
2048 	if (res != QM_MCR_RESULT_OK) {
2049 		pr_err("QUERY_CONGESTION failed: %s\n", mcr_result_str(res));
2050 		return -EIO;
2051 	}
2052 	for (i = 0; i < ARRAY_SIZE(congestion->state.state); i++)
2053 		congestion->state.state[i] =
2054 			be32_to_cpu(congestion->state.state[i]);
2055 	return 0;
2056 }
2057 
2058 int qman_set_vdq(struct qman_fq *fq, u16 num, uint32_t vdqcr_flags)
2059 {
2060 	struct qman_portal *p = get_affine_portal();
2061 	uint32_t vdqcr;
2062 	int ret = -EBUSY;
2063 
2064 	vdqcr = vdqcr_flags;
2065 	vdqcr |= QM_VDQCR_NUMFRAMES_SET(num);
2066 
2067 	if ((fq->state != qman_fq_state_parked) &&
2068 	    (fq->state != qman_fq_state_retired)) {
2069 		ret = -EINVAL;
2070 		goto out;
2071 	}
2072 	if (fq_isset(fq, QMAN_FQ_STATE_VDQCR)) {
2073 		ret = -EBUSY;
2074 		goto out;
2075 	}
2076 	vdqcr = (vdqcr & ~QM_VDQCR_FQID_MASK) | fq->fqid;
2077 
2078 	if (!p->vdqcr_owned) {
2079 		FQLOCK(fq);
2080 		if (fq_isset(fq, QMAN_FQ_STATE_VDQCR))
2081 			goto escape;
2082 		fq_set(fq, QMAN_FQ_STATE_VDQCR);
2083 		FQUNLOCK(fq);
2084 		p->vdqcr_owned = fq;
2085 		ret = 0;
2086 	}
2087 escape:
2088 	if (!ret)
2089 		qm_dqrr_vdqcr_set(&p->p, vdqcr);
2090 
2091 out:
2092 	return ret;
2093 }
2094 
2095 int qman_volatile_dequeue(struct qman_fq *fq, u32 flags __maybe_unused,
2096 			  u32 vdqcr)
2097 {
2098 	struct qman_portal *p;
2099 	int ret = -EBUSY;
2100 
2101 	if ((fq->state != qman_fq_state_parked) &&
2102 	    (fq->state != qman_fq_state_retired))
2103 		return -EINVAL;
2104 	if (vdqcr & QM_VDQCR_FQID_MASK)
2105 		return -EINVAL;
2106 	if (fq_isset(fq, QMAN_FQ_STATE_VDQCR))
2107 		return -EBUSY;
2108 	vdqcr = (vdqcr & ~QM_VDQCR_FQID_MASK) | fq->fqid;
2109 
2110 	p = get_affine_portal();
2111 
2112 	if (!p->vdqcr_owned) {
2113 		FQLOCK(fq);
2114 		if (fq_isset(fq, QMAN_FQ_STATE_VDQCR))
2115 			goto escape;
2116 		fq_set(fq, QMAN_FQ_STATE_VDQCR);
2117 		FQUNLOCK(fq);
2118 		p->vdqcr_owned = fq;
2119 		ret = 0;
2120 	}
2121 escape:
2122 	if (ret)
2123 		return ret;
2124 
2125 	/* VDQCR is set */
2126 	qm_dqrr_vdqcr_set(&p->p, vdqcr);
2127 	return 0;
2128 }
2129 
2130 static noinline void update_eqcr_ci(struct qman_portal *p, u8 avail)
2131 {
2132 	if (avail)
2133 		qm_eqcr_cce_prefetch(&p->p);
2134 	else
2135 		qm_eqcr_cce_update(&p->p);
2136 }
2137 
2138 int qman_eqcr_is_empty(void)
2139 {
2140 	struct qman_portal *p = get_affine_portal();
2141 	u8 avail;
2142 
2143 	update_eqcr_ci(p, 0);
2144 	avail = qm_eqcr_get_fill(&p->p);
2145 	return (avail == 0);
2146 }
2147 
2148 void qman_set_dc_ern(qman_cb_dc_ern handler, int affine)
2149 {
2150 	if (affine) {
2151 		struct qman_portal *p = get_affine_portal();
2152 
2153 		p->cb_dc_ern = handler;
2154 	} else
2155 		cb_dc_ern = handler;
2156 }
2157 
2158 static inline struct qm_eqcr_entry *try_p_eq_start(struct qman_portal *p,
2159 					struct qman_fq *fq,
2160 					const struct qm_fd *fd,
2161 					u32 flags)
2162 {
2163 	struct qm_eqcr_entry *eq;
2164 	u8 avail;
2165 
2166 	if (p->use_eqcr_ci_stashing) {
2167 		/*
2168 		 * The stashing case is easy, only update if we need to in
2169 		 * order to try and liberate ring entries.
2170 		 */
2171 		eq = qm_eqcr_start_stash(&p->p);
2172 	} else {
2173 		/*
2174 		 * The non-stashing case is harder, need to prefetch ahead of
2175 		 * time.
2176 		 */
2177 		avail = qm_eqcr_get_avail(&p->p);
2178 		if (avail < 2)
2179 			update_eqcr_ci(p, avail);
2180 		eq = qm_eqcr_start_no_stash(&p->p);
2181 	}
2182 
2183 	if (unlikely(!eq))
2184 		return NULL;
2185 
2186 	if (flags & QMAN_ENQUEUE_FLAG_DCA)
2187 		eq->dca = QM_EQCR_DCA_ENABLE |
2188 			((flags & QMAN_ENQUEUE_FLAG_DCA_PARK) ?
2189 					QM_EQCR_DCA_PARK : 0) |
2190 			((flags >> 8) & QM_EQCR_DCA_IDXMASK);
2191 	eq->fqid = cpu_to_be32(fq->fqid);
2192 #ifdef CONFIG_FSL_QMAN_FQ_LOOKUP
2193 	eq->tag = cpu_to_be32(fq->key);
2194 #else
2195 	eq->tag = cpu_to_be32((u32)(uintptr_t)fq);
2196 #endif
2197 	eq->fd = *fd;
2198 	cpu_to_hw_fd(&eq->fd);
2199 	return eq;
2200 }
2201 
2202 int qman_enqueue(struct qman_fq *fq, const struct qm_fd *fd, u32 flags)
2203 {
2204 	struct qman_portal *p = get_affine_portal();
2205 	struct qm_eqcr_entry *eq;
2206 
2207 	eq = try_p_eq_start(p, fq, fd, flags);
2208 	if (!eq)
2209 		return -EBUSY;
2210 	/* Note: QM_EQCR_VERB_INTERRUPT == QMAN_ENQUEUE_FLAG_WAIT_SYNC */
2211 	qm_eqcr_pvb_commit(&p->p, QM_EQCR_VERB_CMD_ENQUEUE |
2212 		(flags & (QM_EQCR_VERB_COLOUR_MASK | QM_EQCR_VERB_INTERRUPT)));
2213 	/* Factor the below out, it's used from qman_enqueue_orp() too */
2214 	return 0;
2215 }
2216 
2217 int qman_enqueue_multi(struct qman_fq *fq,
2218 		       const struct qm_fd *fd, u32 *flags,
2219 		int frames_to_send)
2220 {
2221 	struct qman_portal *p = get_affine_portal();
2222 	struct qm_portal *portal = &p->p;
2223 
2224 	register struct qm_eqcr *eqcr = &portal->eqcr;
2225 	struct qm_eqcr_entry *eq = eqcr->cursor, *prev_eq;
2226 
2227 	u8 i = 0, diff, old_ci, sent = 0;
2228 
2229 	/* Update the available entries if no entry is free */
2230 	if (!eqcr->available) {
2231 		old_ci = eqcr->ci;
2232 		eqcr->ci = qm_cl_in(EQCR_CI) & (QM_EQCR_SIZE - 1);
2233 		diff = qm_cyc_diff(QM_EQCR_SIZE, old_ci, eqcr->ci);
2234 		eqcr->available += diff;
2235 		if (!diff)
2236 			return 0;
2237 	}
2238 
2239 	/* try to send as many frames as possible */
2240 	while (eqcr->available && frames_to_send--) {
2241 		eq->fqid = fq->fqid_le;
2242 		eq->fd.opaque_addr = fd->opaque_addr;
2243 		eq->fd.addr = cpu_to_be40(fd->addr);
2244 		eq->fd.status = cpu_to_be32(fd->status);
2245 		eq->fd.opaque = cpu_to_be32(fd->opaque);
2246 		if (flags && (flags[i] & QMAN_ENQUEUE_FLAG_DCA)) {
2247 			eq->dca = QM_EQCR_DCA_ENABLE |
2248 				((flags[i] >> 8) & QM_EQCR_DCA_IDXMASK);
2249 		}
2250 		i++;
2251 		eq = (void *)((unsigned long)(eq + 1) &
2252 			(~(unsigned long)(QM_EQCR_SIZE << 6)));
2253 		eqcr->available--;
2254 		sent++;
2255 		fd++;
2256 	}
2257 	lwsync();
2258 
2259 	/* In order for flushes to complete faster, all lines are recorded in
2260 	 * 32 bit word.
2261 	 */
2262 	eq = eqcr->cursor;
2263 	for (i = 0; i < sent; i++) {
2264 		eq->__dont_write_directly__verb =
2265 			QM_EQCR_VERB_CMD_ENQUEUE | eqcr->vbit;
2266 		prev_eq = eq;
2267 		eq = (void *)((unsigned long)(eq + 1) &
2268 			(~(unsigned long)(QM_EQCR_SIZE << 6)));
2269 		if (unlikely((prev_eq + 1) != eq))
2270 			eqcr->vbit ^= QM_EQCR_VERB_VBIT;
2271 	}
2272 
2273 	/* We need  to flush all the lines but without load/store operations
2274 	 * between them
2275 	 */
2276 	eq = eqcr->cursor;
2277 	for (i = 0; i < sent; i++) {
2278 		dcbf(eq);
2279 		eq = (void *)((unsigned long)(eq + 1) &
2280 			(~(unsigned long)(QM_EQCR_SIZE << 6)));
2281 	}
2282 	/* Update cursor for the next call */
2283 	eqcr->cursor = eq;
2284 	return sent;
2285 }
2286 
2287 int
2288 qman_enqueue_multi_fq(struct qman_fq *fq[], const struct qm_fd *fd,
2289 		      int frames_to_send)
2290 {
2291 	struct qman_portal *p = get_affine_portal();
2292 	struct qm_portal *portal = &p->p;
2293 
2294 	register struct qm_eqcr *eqcr = &portal->eqcr;
2295 	struct qm_eqcr_entry *eq = eqcr->cursor, *prev_eq;
2296 
2297 	u8 i, diff, old_ci, sent = 0;
2298 
2299 	/* Update the available entries if no entry is free */
2300 	if (!eqcr->available) {
2301 		old_ci = eqcr->ci;
2302 		eqcr->ci = qm_cl_in(EQCR_CI) & (QM_EQCR_SIZE - 1);
2303 		diff = qm_cyc_diff(QM_EQCR_SIZE, old_ci, eqcr->ci);
2304 		eqcr->available += diff;
2305 		if (!diff)
2306 			return 0;
2307 	}
2308 
2309 	/* try to send as many frames as possible */
2310 	while (eqcr->available && frames_to_send--) {
2311 		eq->fqid = fq[sent]->fqid_le;
2312 		eq->fd.opaque_addr = fd->opaque_addr;
2313 		eq->fd.addr = cpu_to_be40(fd->addr);
2314 		eq->fd.status = cpu_to_be32(fd->status);
2315 		eq->fd.opaque = cpu_to_be32(fd->opaque);
2316 
2317 		eq = (void *)((unsigned long)(eq + 1) &
2318 			(~(unsigned long)(QM_EQCR_SIZE << 6)));
2319 		eqcr->available--;
2320 		sent++;
2321 		fd++;
2322 	}
2323 	lwsync();
2324 
2325 	/* In order for flushes to complete faster, all lines are recorded in
2326 	 * 32 bit word.
2327 	 */
2328 	eq = eqcr->cursor;
2329 	for (i = 0; i < sent; i++) {
2330 		eq->__dont_write_directly__verb =
2331 			QM_EQCR_VERB_CMD_ENQUEUE | eqcr->vbit;
2332 		prev_eq = eq;
2333 		eq = (void *)((unsigned long)(eq + 1) &
2334 			(~(unsigned long)(QM_EQCR_SIZE << 6)));
2335 		if (unlikely((prev_eq + 1) != eq))
2336 			eqcr->vbit ^= QM_EQCR_VERB_VBIT;
2337 	}
2338 
2339 	/* We need  to flush all the lines but without load/store operations
2340 	 * between them
2341 	 */
2342 	eq = eqcr->cursor;
2343 	for (i = 0; i < sent; i++) {
2344 		dcbf(eq);
2345 		eq = (void *)((unsigned long)(eq + 1) &
2346 			(~(unsigned long)(QM_EQCR_SIZE << 6)));
2347 	}
2348 	/* Update cursor for the next call */
2349 	eqcr->cursor = eq;
2350 	return sent;
2351 }
2352 
2353 int qman_enqueue_orp(struct qman_fq *fq, const struct qm_fd *fd, u32 flags,
2354 		     struct qman_fq *orp, u16 orp_seqnum)
2355 {
2356 	struct qman_portal *p  = get_affine_portal();
2357 	struct qm_eqcr_entry *eq;
2358 
2359 	eq = try_p_eq_start(p, fq, fd, flags);
2360 	if (!eq)
2361 		return -EBUSY;
2362 	/* Process ORP-specifics here */
2363 	if (flags & QMAN_ENQUEUE_FLAG_NLIS)
2364 		orp_seqnum |= QM_EQCR_SEQNUM_NLIS;
2365 	else {
2366 		orp_seqnum &= ~QM_EQCR_SEQNUM_NLIS;
2367 		if (flags & QMAN_ENQUEUE_FLAG_NESN)
2368 			orp_seqnum |= QM_EQCR_SEQNUM_NESN;
2369 		else
2370 			/* No need to check 4 QMAN_ENQUEUE_FLAG_HOLE */
2371 			orp_seqnum &= ~QM_EQCR_SEQNUM_NESN;
2372 	}
2373 	eq->seqnum = cpu_to_be16(orp_seqnum);
2374 	eq->orp = cpu_to_be32(orp->fqid);
2375 	/* Note: QM_EQCR_VERB_INTERRUPT == QMAN_ENQUEUE_FLAG_WAIT_SYNC */
2376 	qm_eqcr_pvb_commit(&p->p, QM_EQCR_VERB_ORP |
2377 		((flags & (QMAN_ENQUEUE_FLAG_HOLE | QMAN_ENQUEUE_FLAG_NESN)) ?
2378 				0 : QM_EQCR_VERB_CMD_ENQUEUE) |
2379 		(flags & (QM_EQCR_VERB_COLOUR_MASK | QM_EQCR_VERB_INTERRUPT)));
2380 
2381 	return 0;
2382 }
2383 
2384 int qman_modify_cgr(struct qman_cgr *cgr, u32 flags,
2385 		    struct qm_mcc_initcgr *opts)
2386 {
2387 	struct qm_mc_command *mcc;
2388 	struct qm_mc_result *mcr;
2389 	struct qman_portal *p = get_affine_portal();
2390 
2391 	u8 res;
2392 	u8 verb = QM_MCC_VERB_MODIFYCGR;
2393 
2394 	mcc = qm_mc_start(&p->p);
2395 	if (opts)
2396 		mcc->initcgr = *opts;
2397 	mcc->initcgr.we_mask = cpu_to_be16(mcc->initcgr.we_mask);
2398 	mcc->initcgr.cgr.wr_parm_g.word =
2399 		cpu_to_be32(mcc->initcgr.cgr.wr_parm_g.word);
2400 	mcc->initcgr.cgr.wr_parm_y.word =
2401 		cpu_to_be32(mcc->initcgr.cgr.wr_parm_y.word);
2402 	mcc->initcgr.cgr.wr_parm_r.word =
2403 		cpu_to_be32(mcc->initcgr.cgr.wr_parm_r.word);
2404 	mcc->initcgr.cgr.cscn_targ =  cpu_to_be32(mcc->initcgr.cgr.cscn_targ);
2405 	mcc->initcgr.cgr.__cs_thres = cpu_to_be16(mcc->initcgr.cgr.__cs_thres);
2406 
2407 	mcc->initcgr.cgid = cgr->cgrid;
2408 	if (flags & QMAN_CGR_FLAG_USE_INIT)
2409 		verb = QM_MCC_VERB_INITCGR;
2410 	qm_mc_commit(&p->p, verb);
2411 	while (!(mcr = qm_mc_result(&p->p)))
2412 		cpu_relax();
2413 
2414 	DPAA_ASSERT((mcr->verb & QM_MCR_VERB_MASK) == verb);
2415 	res = mcr->result;
2416 	return (res == QM_MCR_RESULT_OK) ? 0 : -EIO;
2417 }
2418 
2419 #define TARG_MASK(n) (0x80000000 >> (n->config->channel - \
2420 					QM_CHANNEL_SWPORTAL0))
2421 #define TARG_DCP_MASK(n) (0x80000000 >> (10 + n))
2422 #define PORTAL_IDX(n) (n->config->channel - QM_CHANNEL_SWPORTAL0)
2423 
2424 int qman_create_cgr(struct qman_cgr *cgr, u32 flags,
2425 		    struct qm_mcc_initcgr *opts)
2426 {
2427 	struct qm_mcr_querycgr cgr_state;
2428 	struct qm_mcc_initcgr local_opts;
2429 	int ret;
2430 	struct qman_portal *p;
2431 
2432 	/* We have to check that the provided CGRID is within the limits of the
2433 	 * data-structures, for obvious reasons. However we'll let h/w take
2434 	 * care of determining whether it's within the limits of what exists on
2435 	 * the SoC.
2436 	 */
2437 	if (cgr->cgrid >= __CGR_NUM)
2438 		return -EINVAL;
2439 
2440 	p = get_affine_portal();
2441 
2442 	memset(&local_opts, 0, sizeof(struct qm_mcc_initcgr));
2443 	cgr->chan = p->config->channel;
2444 	spin_lock(&p->cgr_lock);
2445 
2446 	/* if no opts specified, just add it to the list */
2447 	if (!opts)
2448 		goto add_list;
2449 
2450 	ret = qman_query_cgr(cgr, &cgr_state);
2451 	if (ret)
2452 		goto release_lock;
2453 	if (opts)
2454 		local_opts = *opts;
2455 	if ((qman_ip_rev & 0xFF00) >= QMAN_REV30)
2456 		local_opts.cgr.cscn_targ_upd_ctrl =
2457 			QM_CGR_TARG_UDP_CTRL_WRITE_BIT | PORTAL_IDX(p);
2458 	else
2459 		/* Overwrite TARG */
2460 		local_opts.cgr.cscn_targ = cgr_state.cgr.cscn_targ |
2461 							TARG_MASK(p);
2462 	local_opts.we_mask |= QM_CGR_WE_CSCN_TARG;
2463 
2464 	/* send init if flags indicate so */
2465 	if (opts && (flags & QMAN_CGR_FLAG_USE_INIT))
2466 		ret = qman_modify_cgr(cgr, QMAN_CGR_FLAG_USE_INIT, &local_opts);
2467 	else
2468 		ret = qman_modify_cgr(cgr, 0, &local_opts);
2469 	if (ret)
2470 		goto release_lock;
2471 add_list:
2472 	list_add(&cgr->node, &p->cgr_cbs);
2473 
2474 	/* Determine if newly added object requires its callback to be called */
2475 	ret = qman_query_cgr(cgr, &cgr_state);
2476 	if (ret) {
2477 		/* we can't go back, so proceed and return success, but screen
2478 		 * and wail to the log file.
2479 		 */
2480 		pr_crit("CGR HW state partially modified\n");
2481 		ret = 0;
2482 		goto release_lock;
2483 	}
2484 	if (cgr->cb && cgr_state.cgr.cscn_en && qman_cgrs_get(&p->cgrs[1],
2485 							      cgr->cgrid))
2486 		cgr->cb(p, cgr, 1);
2487 release_lock:
2488 	spin_unlock(&p->cgr_lock);
2489 	return ret;
2490 }
2491 
2492 int qman_create_cgr_to_dcp(struct qman_cgr *cgr, u32 flags, u16 dcp_portal,
2493 			   struct qm_mcc_initcgr *opts)
2494 {
2495 	struct qm_mcc_initcgr local_opts;
2496 	struct qm_mcr_querycgr cgr_state;
2497 	int ret;
2498 
2499 	if ((qman_ip_rev & 0xFF00) < QMAN_REV30) {
2500 		pr_warn("QMan version doesn't support CSCN => DCP portal\n");
2501 		return -EINVAL;
2502 	}
2503 	/* We have to check that the provided CGRID is within the limits of the
2504 	 * data-structures, for obvious reasons. However we'll let h/w take
2505 	 * care of determining whether it's within the limits of what exists on
2506 	 * the SoC.
2507 	 */
2508 	if (cgr->cgrid >= __CGR_NUM)
2509 		return -EINVAL;
2510 
2511 	ret = qman_query_cgr(cgr, &cgr_state);
2512 	if (ret)
2513 		return ret;
2514 
2515 	memset(&local_opts, 0, sizeof(struct qm_mcc_initcgr));
2516 	if (opts)
2517 		local_opts = *opts;
2518 
2519 	if ((qman_ip_rev & 0xFF00) >= QMAN_REV30)
2520 		local_opts.cgr.cscn_targ_upd_ctrl =
2521 				QM_CGR_TARG_UDP_CTRL_WRITE_BIT |
2522 				QM_CGR_TARG_UDP_CTRL_DCP | dcp_portal;
2523 	else
2524 		local_opts.cgr.cscn_targ = cgr_state.cgr.cscn_targ |
2525 					TARG_DCP_MASK(dcp_portal);
2526 	local_opts.we_mask |= QM_CGR_WE_CSCN_TARG;
2527 
2528 	/* send init if flags indicate so */
2529 	if (opts && (flags & QMAN_CGR_FLAG_USE_INIT))
2530 		ret = qman_modify_cgr(cgr, QMAN_CGR_FLAG_USE_INIT,
2531 				      &local_opts);
2532 	else
2533 		ret = qman_modify_cgr(cgr, 0, &local_opts);
2534 
2535 	return ret;
2536 }
2537 
2538 int qman_delete_cgr(struct qman_cgr *cgr)
2539 {
2540 	struct qm_mcr_querycgr cgr_state;
2541 	struct qm_mcc_initcgr local_opts;
2542 	int ret = 0;
2543 	struct qman_cgr *i;
2544 	struct qman_portal *p = get_affine_portal();
2545 
2546 	if (cgr->chan != p->config->channel) {
2547 		pr_crit("Attempting to delete cgr from different portal than"
2548 			" it was create: create 0x%x, delete 0x%x\n",
2549 			cgr->chan, p->config->channel);
2550 		ret = -EINVAL;
2551 		goto put_portal;
2552 	}
2553 	memset(&local_opts, 0, sizeof(struct qm_mcc_initcgr));
2554 	spin_lock(&p->cgr_lock);
2555 	list_del(&cgr->node);
2556 	/*
2557 	 * If there are no other CGR objects for this CGRID in the list,
2558 	 * update CSCN_TARG accordingly
2559 	 */
2560 	list_for_each_entry(i, &p->cgr_cbs, node)
2561 		if ((i->cgrid == cgr->cgrid) && i->cb)
2562 			goto release_lock;
2563 	ret = qman_query_cgr(cgr, &cgr_state);
2564 	if (ret)  {
2565 		/* add back to the list */
2566 		list_add(&cgr->node, &p->cgr_cbs);
2567 		goto release_lock;
2568 	}
2569 	/* Overwrite TARG */
2570 	local_opts.we_mask = QM_CGR_WE_CSCN_TARG;
2571 	if ((qman_ip_rev & 0xFF00) >= QMAN_REV30)
2572 		local_opts.cgr.cscn_targ_upd_ctrl = PORTAL_IDX(p);
2573 	else
2574 		local_opts.cgr.cscn_targ = cgr_state.cgr.cscn_targ &
2575 							 ~(TARG_MASK(p));
2576 	ret = qman_modify_cgr(cgr, 0, &local_opts);
2577 	if (ret)
2578 		/* add back to the list */
2579 		list_add(&cgr->node, &p->cgr_cbs);
2580 release_lock:
2581 	spin_unlock(&p->cgr_lock);
2582 put_portal:
2583 	return ret;
2584 }
2585 
2586 int qman_shutdown_fq(u32 fqid)
2587 {
2588 	struct qman_portal *p;
2589 	struct qm_portal *low_p;
2590 	struct qm_mc_command *mcc;
2591 	struct qm_mc_result *mcr;
2592 	u8 state;
2593 	int orl_empty, fq_empty, drain = 0;
2594 	u32 result;
2595 	u32 channel, wq;
2596 	u16 dest_wq;
2597 
2598 	p = get_affine_portal();
2599 	low_p = &p->p;
2600 
2601 	/* Determine the state of the FQID */
2602 	mcc = qm_mc_start(low_p);
2603 	mcc->queryfq_np.fqid = cpu_to_be32(fqid);
2604 	qm_mc_commit(low_p, QM_MCC_VERB_QUERYFQ_NP);
2605 	while (!(mcr = qm_mc_result(low_p)))
2606 		cpu_relax();
2607 	DPAA_ASSERT((mcr->verb & QM_MCR_VERB_MASK) == QM_MCR_VERB_QUERYFQ_NP);
2608 	state = mcr->queryfq_np.state & QM_MCR_NP_STATE_MASK;
2609 	if (state == QM_MCR_NP_STATE_OOS)
2610 		return 0; /* Already OOS, no need to do anymore checks */
2611 
2612 	/* Query which channel the FQ is using */
2613 	mcc = qm_mc_start(low_p);
2614 	mcc->queryfq.fqid = cpu_to_be32(fqid);
2615 	qm_mc_commit(low_p, QM_MCC_VERB_QUERYFQ);
2616 	while (!(mcr = qm_mc_result(low_p)))
2617 		cpu_relax();
2618 	DPAA_ASSERT((mcr->verb & QM_MCR_VERB_MASK) == QM_MCR_VERB_QUERYFQ);
2619 
2620 	/* Need to store these since the MCR gets reused */
2621 	dest_wq = be16_to_cpu(mcr->queryfq.fqd.dest_wq);
2622 	channel = dest_wq & 0x7;
2623 	wq = dest_wq >> 3;
2624 
2625 	switch (state) {
2626 	case QM_MCR_NP_STATE_TEN_SCHED:
2627 	case QM_MCR_NP_STATE_TRU_SCHED:
2628 	case QM_MCR_NP_STATE_ACTIVE:
2629 	case QM_MCR_NP_STATE_PARKED:
2630 		orl_empty = 0;
2631 		mcc = qm_mc_start(low_p);
2632 		mcc->alterfq.fqid = cpu_to_be32(fqid);
2633 		qm_mc_commit(low_p, QM_MCC_VERB_ALTER_RETIRE);
2634 		while (!(mcr = qm_mc_result(low_p)))
2635 			cpu_relax();
2636 		DPAA_ASSERT((mcr->verb & QM_MCR_VERB_MASK) ==
2637 			   QM_MCR_VERB_ALTER_RETIRE);
2638 		result = mcr->result; /* Make a copy as we reuse MCR below */
2639 
2640 		if (result == QM_MCR_RESULT_PENDING) {
2641 			/* Need to wait for the FQRN in the message ring, which
2642 			 * will only occur once the FQ has been drained.  In
2643 			 * order for the FQ to drain the portal needs to be set
2644 			 * to dequeue from the channel the FQ is scheduled on
2645 			 */
2646 			const struct qm_mr_entry *msg;
2647 			const struct qm_dqrr_entry *dqrr = NULL;
2648 			int found_fqrn = 0;
2649 			__maybe_unused u16 dequeue_wq = 0;
2650 
2651 			/* Flag that we need to drain FQ */
2652 			drain = 1;
2653 
2654 			if (channel >= qm_channel_pool1 &&
2655 			    channel < (u16)(qm_channel_pool1 + 15)) {
2656 				/* Pool channel, enable the bit in the portal */
2657 				dequeue_wq = (channel -
2658 					      qm_channel_pool1 + 1) << 4 | wq;
2659 			} else if (channel < qm_channel_pool1) {
2660 				/* Dedicated channel */
2661 				dequeue_wq = wq;
2662 			} else {
2663 				pr_info("Cannot recover FQ 0x%x,"
2664 					" it is scheduled on channel 0x%x",
2665 					fqid, channel);
2666 				return -EBUSY;
2667 			}
2668 			/* Set the sdqcr to drain this channel */
2669 			if (channel < qm_channel_pool1)
2670 				qm_dqrr_sdqcr_set(low_p,
2671 						  QM_SDQCR_TYPE_ACTIVE |
2672 					  QM_SDQCR_CHANNELS_DEDICATED);
2673 			else
2674 				qm_dqrr_sdqcr_set(low_p,
2675 						  QM_SDQCR_TYPE_ACTIVE |
2676 						  QM_SDQCR_CHANNELS_POOL_CONV
2677 						  (channel));
2678 			while (!found_fqrn) {
2679 				/* Keep draining DQRR while checking the MR*/
2680 				qm_dqrr_pvb_update(low_p);
2681 				dqrr = qm_dqrr_current(low_p);
2682 				while (dqrr) {
2683 					qm_dqrr_cdc_consume_1ptr(
2684 						low_p, dqrr, 0);
2685 					qm_dqrr_pvb_update(low_p);
2686 					qm_dqrr_next(low_p);
2687 					dqrr = qm_dqrr_current(low_p);
2688 				}
2689 				/* Process message ring too */
2690 				qm_mr_pvb_update(low_p);
2691 				msg = qm_mr_current(low_p);
2692 				while (msg) {
2693 					if ((msg->ern.verb &
2694 					     QM_MR_VERB_TYPE_MASK)
2695 					    == QM_MR_VERB_FQRN)
2696 						found_fqrn = 1;
2697 					qm_mr_next(low_p);
2698 					qm_mr_cci_consume_to_current(low_p);
2699 					qm_mr_pvb_update(low_p);
2700 					msg = qm_mr_current(low_p);
2701 				}
2702 				cpu_relax();
2703 			}
2704 		}
2705 		if (result != QM_MCR_RESULT_OK &&
2706 		    result !=  QM_MCR_RESULT_PENDING) {
2707 			/* error */
2708 			pr_err("qman_retire_fq failed on FQ 0x%x,"
2709 			       " result=0x%x\n", fqid, result);
2710 			return -1;
2711 		}
2712 		if (!(mcr->alterfq.fqs & QM_MCR_FQS_ORLPRESENT)) {
2713 			/* ORL had no entries, no need to wait until the
2714 			 * ERNs come in.
2715 			 */
2716 			orl_empty = 1;
2717 		}
2718 		/* Retirement succeeded, check to see if FQ needs
2719 		 * to be drained.
2720 		 */
2721 		if (drain || mcr->alterfq.fqs & QM_MCR_FQS_NOTEMPTY) {
2722 			/* FQ is Not Empty, drain using volatile DQ commands */
2723 			fq_empty = 0;
2724 			do {
2725 				const struct qm_dqrr_entry *dqrr = NULL;
2726 				u32 vdqcr = fqid | QM_VDQCR_NUMFRAMES_SET(3);
2727 
2728 				qm_dqrr_vdqcr_set(low_p, vdqcr);
2729 
2730 				/* Wait for a dequeue to occur */
2731 				while (dqrr == NULL) {
2732 					qm_dqrr_pvb_update(low_p);
2733 					dqrr = qm_dqrr_current(low_p);
2734 					if (!dqrr)
2735 						cpu_relax();
2736 				}
2737 				/* Process the dequeues, making sure to
2738 				 * empty the ring completely.
2739 				 */
2740 				while (dqrr) {
2741 					if (dqrr->fqid == fqid &&
2742 					    dqrr->stat & QM_DQRR_STAT_FQ_EMPTY)
2743 						fq_empty = 1;
2744 					qm_dqrr_cdc_consume_1ptr(low_p,
2745 								 dqrr, 0);
2746 					qm_dqrr_pvb_update(low_p);
2747 					qm_dqrr_next(low_p);
2748 					dqrr = qm_dqrr_current(low_p);
2749 				}
2750 			} while (fq_empty == 0);
2751 		}
2752 		qm_dqrr_sdqcr_set(low_p, 0);
2753 
2754 		/* Wait for the ORL to have been completely drained */
2755 		while (orl_empty == 0) {
2756 			const struct qm_mr_entry *msg;
2757 
2758 			qm_mr_pvb_update(low_p);
2759 			msg = qm_mr_current(low_p);
2760 			while (msg) {
2761 				if ((msg->ern.verb & QM_MR_VERB_TYPE_MASK) ==
2762 				    QM_MR_VERB_FQRL)
2763 					orl_empty = 1;
2764 				qm_mr_next(low_p);
2765 				qm_mr_cci_consume_to_current(low_p);
2766 				qm_mr_pvb_update(low_p);
2767 				msg = qm_mr_current(low_p);
2768 			}
2769 			cpu_relax();
2770 		}
2771 		mcc = qm_mc_start(low_p);
2772 		mcc->alterfq.fqid = cpu_to_be32(fqid);
2773 		qm_mc_commit(low_p, QM_MCC_VERB_ALTER_OOS);
2774 		while (!(mcr = qm_mc_result(low_p)))
2775 			cpu_relax();
2776 		DPAA_ASSERT((mcr->verb & QM_MCR_VERB_MASK) ==
2777 			   QM_MCR_VERB_ALTER_OOS);
2778 		if (mcr->result != QM_MCR_RESULT_OK) {
2779 			pr_err(
2780 			"OOS after drain Failed on FQID 0x%x, result 0x%x\n",
2781 			       fqid, mcr->result);
2782 			return -1;
2783 		}
2784 		return 0;
2785 
2786 	case QM_MCR_NP_STATE_RETIRED:
2787 		/* Send OOS Command */
2788 		mcc = qm_mc_start(low_p);
2789 		mcc->alterfq.fqid = cpu_to_be32(fqid);
2790 		qm_mc_commit(low_p, QM_MCC_VERB_ALTER_OOS);
2791 		while (!(mcr = qm_mc_result(low_p)))
2792 			cpu_relax();
2793 		DPAA_ASSERT((mcr->verb & QM_MCR_VERB_MASK) ==
2794 			   QM_MCR_VERB_ALTER_OOS);
2795 		if (mcr->result) {
2796 			pr_err("OOS Failed on FQID 0x%x\n", fqid);
2797 			return -1;
2798 		}
2799 		return 0;
2800 
2801 	}
2802 	return -1;
2803 }
2804