1/***********************license start***************
2 * Copyright (c) 2003-2010  Cavium Inc. (support@cavium.com). All rights
3 * reserved.
4 *
5 *
6 * Redistribution and use in source and binary forms, with or without
7 * modification, are permitted provided that the following conditions are
8 * met:
9 *
10 *   * Redistributions of source code must retain the above copyright
11 *     notice, this list of conditions and the following disclaimer.
12 *
13 *   * Redistributions in binary form must reproduce the above
14 *     copyright notice, this list of conditions and the following
15 *     disclaimer in the documentation and/or other materials provided
16 *     with the distribution.
17
18 *   * Neither the name of Cavium Inc. nor the names of
19 *     its contributors may be used to endorse or promote products
20 *     derived from this software without specific prior written
21 *     permission.
22
23 * This Software, including technical data, may be subject to U.S. export  control
24 * laws, including the U.S. Export Administration Act and its  associated
25 * regulations, and may be subject to export or import  regulations in other
26 * countries.
27
28 * TO THE MAXIMUM EXTENT PERMITTED BY LAW, THE SOFTWARE IS PROVIDED "AS IS"
29 * AND WITH ALL FAULTS AND CAVIUM INC. MAKES NO PROMISES, REPRESENTATIONS OR
30 * WARRANTIES, EITHER EXPRESS, IMPLIED, STATUTORY, OR OTHERWISE, WITH RESPECT TO
31 * THE SOFTWARE, INCLUDING ITS CONDITION, ITS CONFORMITY TO ANY REPRESENTATION OR
32 * DESCRIPTION, OR THE EXISTENCE OF ANY LATENT OR PATENT DEFECTS, AND CAVIUM
33 * SPECIFICALLY DISCLAIMS ALL IMPLIED (IF ANY) WARRANTIES OF TITLE,
34 * MERCHANTABILITY, NONINFRINGEMENT, FITNESS FOR A PARTICULAR PURPOSE, LACK OF
35 * VIRUSES, ACCURACY OR COMPLETENESS, QUIET ENJOYMENT, QUIET POSSESSION OR
36 * CORRESPONDENCE TO DESCRIPTION. THE ENTIRE  RISK ARISING OUT OF USE OR
37 * PERFORMANCE OF THE SOFTWARE LIES WITH YOU.
38 ***********************license end**************************************/
39
40
41
42
43
44
45
46/**
47 * @file
48 *
49 * Support functions for managing command queues used for
50 * various hardware blocks.
51 *
52 * The common command queue infrastructure abstracts out the
53 * software necessary for adding to Octeon's chained queue
54 * structures. These structures are used for commands to the
55 * PKO, ZIP, DFA, RAID, and DMA engine blocks. Although each
56 * hardware unit takes commands and CSRs of different types,
57 * they all use basic linked command buffers to store the
58 * pending request. In general, users of the CVMX API don't
59 * call cvmx-cmd-queue functions directly. Instead the hardware
60 * unit specific wrapper should be used. The wrappers perform
61 * unit specific validation and CSR writes to submit the
62 * commands.
63 *
64 * Even though most software will never directly interact with
65 * cvmx-cmd-queue, knowledge of its internal workings can help
66 * in diagnosing performance problems and help with debugging.
67 *
68 * Command queue pointers are stored in a global named block
69 * called "cvmx_cmd_queues". Except for the PKO queues, each
70 * hardware queue is stored in its own cache line to reduce SMP
71 * contention on spin locks. The PKO queues are stored such that
72 * every 16th queue is next to each other in memory. This scheme
73 * allows for queues being in separate cache lines when there
74 * are low number of queues per port. With 16 queues per port,
75 * the first queue for each port is in the same cache area. The
76 * second queues for each port are in another area, etc. This
77 * allows software to implement very efficient lockless PKO with
78 * 16 queues per port using a minimum of cache lines per core.
79 * All queues for a given core will be isolated in the same
80 * cache area.
81 *
82 * In addition to the memory pointer layout, cvmx-cmd-queue
83 * provides an optimized fair ll/sc locking mechanism for the
84 * queues. The lock uses a "ticket / now serving" model to
85 * maintain fair order on contended locks. In addition, it uses
86 * predicted locking time to limit cache contention. When a core
87 * know it must wait in line for a lock, it spins on the
88 * internal cycle counter to completely eliminate any causes of
89 * bus traffic.
90 *
91 * <hr> $Revision: 70030 $ <hr>
92 */
93
94#ifndef __CVMX_CMD_QUEUE_H__
95#define __CVMX_CMD_QUEUE_H__
96
97#if !defined(CVMX_BUILD_FOR_LINUX_KERNEL) && !defined(CVMX_BUILD_FOR_FREEBSD_KERNEL)
98#include "executive-config.h"
99#include "cvmx-config.h"
100#endif
101
102#include "cvmx-fpa.h"
103
104#ifdef	__cplusplus
105extern "C" {
106#endif
107
108/**
109 * By default we disable the max depth support. Most programs
110 * don't use it and it slows down the command queue processing
111 * significantly.
112 */
113#ifndef CVMX_CMD_QUEUE_ENABLE_MAX_DEPTH
114#define CVMX_CMD_QUEUE_ENABLE_MAX_DEPTH 0
115#endif
116
117/**
118 * Enumeration representing all hardware blocks that use command
119 * queues. Each hardware block has up to 65536 sub identifiers for
120 * multiple command queues. Not all chips support all hardware
121 * units.
122 */
123typedef enum
124{
125    CVMX_CMD_QUEUE_PKO_BASE = 0x00000,
126#define CVMX_CMD_QUEUE_PKO(queue) ((cvmx_cmd_queue_id_t)(CVMX_CMD_QUEUE_PKO_BASE + (0xffff&(queue))))
127    CVMX_CMD_QUEUE_ZIP      = 0x10000,
128#define CVMX_CMD_QUEUE_ZIP_QUE(queue) ((cvmx_cmd_queue_id_t)(CVMX_CMD_QUEUE_ZIP + (0xffff&(queue))))
129    CVMX_CMD_QUEUE_DFA      = 0x20000,
130    CVMX_CMD_QUEUE_RAID     = 0x30000,
131    CVMX_CMD_QUEUE_DMA_BASE = 0x40000,
132#define CVMX_CMD_QUEUE_DMA(queue) ((cvmx_cmd_queue_id_t)(CVMX_CMD_QUEUE_DMA_BASE + (0xffff&(queue))))
133    CVMX_CMD_QUEUE_END      = 0x50000,
134} cvmx_cmd_queue_id_t;
135
136/**
137 * Command write operations can fail if the command queue needs
138 * a new buffer and the associated FPA pool is empty. It can also
139 * fail if the number of queued command words reaches the maximum
140 * set at initialization.
141 */
142typedef enum
143{
144    CVMX_CMD_QUEUE_SUCCESS = 0,
145    CVMX_CMD_QUEUE_NO_MEMORY = -1,
146    CVMX_CMD_QUEUE_FULL = -2,
147    CVMX_CMD_QUEUE_INVALID_PARAM = -3,
148    CVMX_CMD_QUEUE_ALREADY_SETUP = -4,
149} cvmx_cmd_queue_result_t;
150
151typedef struct
152{
153    uint8_t  now_serving;           /**< You have lock when this is your ticket */
154    uint64_t unused1        : 24;
155    uint32_t max_depth;             /**< Maximum outstanding command words */
156    uint64_t fpa_pool       : 3;    /**< FPA pool buffers come from */
157    uint64_t base_ptr_div128: 29;   /**< Top of command buffer pointer shifted 7 */
158    uint64_t unused2        : 6;
159    uint64_t pool_size_m1   : 13;   /**< FPA buffer size in 64bit words minus 1 */
160    uint64_t index          : 13;   /**< Number of commands already used in buffer */
161} __cvmx_cmd_queue_state_t;
162
163/**
164 * This structure contains the global state of all command queues.
165 * It is stored in a bootmem named block and shared by all
166 * applications running on Octeon. Tickets are stored in a different
167 * cache line that queue information to reduce the contention on the
168 * ll/sc used to get a ticket. If this is not the case, the update
169 * of queue state causes the ll/sc to fail quite often.
170 */
171typedef struct
172{
173    uint64_t                 ticket[(CVMX_CMD_QUEUE_END>>16) * 256];
174    __cvmx_cmd_queue_state_t state[(CVMX_CMD_QUEUE_END>>16) * 256];
175} __cvmx_cmd_queue_all_state_t;
176
177extern CVMX_SHARED __cvmx_cmd_queue_all_state_t *__cvmx_cmd_queue_state_ptr;
178
179/**
180 * Initialize a command queue for use. The initial FPA buffer is
181 * allocated and the hardware unit is configured to point to the
182 * new command queue.
183 *
184 * @param queue_id  Hardware command queue to initialize.
185 * @param max_depth Maximum outstanding commands that can be queued.
186 * @param fpa_pool  FPA pool the command queues should come from.
187 * @param pool_size Size of each buffer in the FPA pool (bytes)
188 *
189 * @return CVMX_CMD_QUEUE_SUCCESS or a failure code
190 */
191cvmx_cmd_queue_result_t cvmx_cmd_queue_initialize(cvmx_cmd_queue_id_t queue_id, int max_depth, int fpa_pool, int pool_size);
192
193/**
194 * Shutdown a queue a free it's command buffers to the FPA. The
195 * hardware connected to the queue must be stopped before this
196 * function is called.
197 *
198 * @param queue_id Queue to shutdown
199 *
200 * @return CVMX_CMD_QUEUE_SUCCESS or a failure code
201 */
202cvmx_cmd_queue_result_t cvmx_cmd_queue_shutdown(cvmx_cmd_queue_id_t queue_id);
203
204/**
205 * Return the number of command words pending in the queue. This
206 * function may be relatively slow for some hardware units.
207 *
208 * @param queue_id Hardware command queue to query
209 *
210 * @return Number of outstanding commands
211 */
212int cvmx_cmd_queue_length(cvmx_cmd_queue_id_t queue_id);
213
214/**
215 * Return the command buffer to be written to. The purpose of this
216 * function is to allow CVMX routine access to the low level buffer
217 * for initial hardware setup. User applications should not call this
218 * function directly.
219 *
220 * @param queue_id Command queue to query
221 *
222 * @return Command buffer or NULL on failure
223 */
224void *cvmx_cmd_queue_buffer(cvmx_cmd_queue_id_t queue_id);
225
226/**
227 * @INTERNAL
228 * Get the index into the state arrays for the supplied queue id.
229 *
230 * @param queue_id Queue ID to get an index for
231 *
232 * @return Index into the state arrays
233 */
234static inline int __cvmx_cmd_queue_get_index(cvmx_cmd_queue_id_t queue_id)
235{
236    /* Warning: This code currently only works with devices that have 256 queues
237        or less. Devices with more than 16 queues are laid out in memory to allow
238        cores quick access to every 16th queue. This reduces cache thrashing
239        when you are running 16 queues per port to support lockless operation */
240    int unit = queue_id>>16;
241    int q = (queue_id >> 4) & 0xf;
242    int core = queue_id & 0xf;
243    return unit*256 + core*16 + q;
244}
245
246
247/**
248 * @INTERNAL
249 * Lock the supplied queue so nobody else is updating it at the same
250 * time as us.
251 *
252 * @param queue_id Queue ID to lock
253 * @param qptr     Pointer to the queue's global state
254 */
255static inline void __cvmx_cmd_queue_lock(cvmx_cmd_queue_id_t queue_id, __cvmx_cmd_queue_state_t *qptr)
256{
257    int tmp;
258    int my_ticket;
259    CVMX_PREFETCH(qptr, 0);
260    asm volatile (
261        ".set push\n"
262        ".set noreorder\n"
263        "1:\n"
264        "ll     %[my_ticket], %[ticket_ptr]\n"          /* Atomic add one to ticket_ptr */
265        "li     %[ticket], 1\n"                         /*    and store the original value */
266        "baddu  %[ticket], %[my_ticket]\n"              /*    in my_ticket */
267        "sc     %[ticket], %[ticket_ptr]\n"
268        "beqz   %[ticket], 1b\n"
269        " nop\n"
270        "lbu    %[ticket], %[now_serving]\n"            /* Load the current now_serving ticket */
271        "2:\n"
272        "beq    %[ticket], %[my_ticket], 4f\n"          /* Jump out if now_serving == my_ticket */
273        " subu   %[ticket], %[my_ticket], %[ticket]\n"  /* Find out how many tickets are in front of me */
274        "subu  %[ticket], 1\n"                          /* Use tickets in front of me minus one to delay */
275        "cins   %[ticket], %[ticket], 5, 7\n"           /* Delay will be ((tickets in front)-1)*32 loops */
276        "3:\n"
277        "bnez   %[ticket], 3b\n"                        /* Loop here until our ticket might be up */
278        " subu  %[ticket], 1\n"
279        "b      2b\n"                                   /* Jump back up to check out ticket again */
280        " lbu   %[ticket], %[now_serving]\n"            /* Load the current now_serving ticket */
281        "4:\n"
282        ".set pop\n"
283        : [ticket_ptr] "=m" (__cvmx_cmd_queue_state_ptr->ticket[__cvmx_cmd_queue_get_index(queue_id)]),
284          [now_serving] "=m" (qptr->now_serving),
285          [ticket] "=r" (tmp),
286          [my_ticket] "=r" (my_ticket)
287    );
288}
289
290
291/**
292 * @INTERNAL
293 * Unlock the queue, flushing all writes.
294 *
295 * @param qptr   Queue to unlock
296 */
297static inline void __cvmx_cmd_queue_unlock(__cvmx_cmd_queue_state_t *qptr)
298{
299    uint8_t ns;
300
301    ns = qptr->now_serving + 1;
302    CVMX_SYNCWS; /* Order queue manipulation with respect to the unlock.  */
303    qptr->now_serving = ns;
304    CVMX_SYNCWS; /* nudge out the unlock. */
305}
306
307
308/**
309 * @INTERNAL
310 * Get the queue state structure for the given queue id
311 *
312 * @param queue_id Queue id to get
313 *
314 * @return Queue structure or NULL on failure
315 */
316static inline __cvmx_cmd_queue_state_t *__cvmx_cmd_queue_get_state(cvmx_cmd_queue_id_t queue_id)
317{
318    if (CVMX_ENABLE_PARAMETER_CHECKING)
319    {
320        if (cvmx_unlikely(queue_id >= CVMX_CMD_QUEUE_END))
321            return NULL;
322        if (cvmx_unlikely((queue_id & 0xffff) >= 256))
323            return NULL;
324    }
325    return &__cvmx_cmd_queue_state_ptr->state[__cvmx_cmd_queue_get_index(queue_id)];
326}
327
328
329/**
330 * Write an arbitrary number of command words to a command queue.
331 * This is a generic function; the fixed number of command word
332 * functions yield higher performance.
333 *
334 * @param queue_id  Hardware command queue to write to
335 * @param use_locking
336 *                  Use internal locking to ensure exclusive access for queue
337 *                  updates. If you don't use this locking you must ensure
338 *                  exclusivity some other way. Locking is strongly recommended.
339 * @param cmd_count Number of command words to write
340 * @param cmds      Array of commands to write
341 *
342 * @return CVMX_CMD_QUEUE_SUCCESS or a failure code
343 */
344static inline cvmx_cmd_queue_result_t cvmx_cmd_queue_write(cvmx_cmd_queue_id_t queue_id, int use_locking, int cmd_count, uint64_t *cmds)
345{
346    __cvmx_cmd_queue_state_t *qptr = __cvmx_cmd_queue_get_state(queue_id);
347
348    if (CVMX_ENABLE_PARAMETER_CHECKING)
349    {
350        if (cvmx_unlikely(qptr == NULL))
351            return CVMX_CMD_QUEUE_INVALID_PARAM;
352        if (cvmx_unlikely((cmd_count < 1) || (cmd_count > 32)))
353            return CVMX_CMD_QUEUE_INVALID_PARAM;
354        if (cvmx_unlikely(cmds == NULL))
355            return CVMX_CMD_QUEUE_INVALID_PARAM;
356    }
357
358    /* Make sure nobody else is updating the same queue */
359    if (cvmx_likely(use_locking))
360        __cvmx_cmd_queue_lock(queue_id, qptr);
361
362    /* If a max queue length was specified then make sure we don't
363        exceed it. If any part of the command would be below the limit
364        we allow it */
365    if (CVMX_CMD_QUEUE_ENABLE_MAX_DEPTH && cvmx_unlikely(qptr->max_depth))
366    {
367        if (cvmx_unlikely(cvmx_cmd_queue_length(queue_id) > (int)qptr->max_depth))
368        {
369            if (cvmx_likely(use_locking))
370                __cvmx_cmd_queue_unlock(qptr);
371            return CVMX_CMD_QUEUE_FULL;
372        }
373    }
374
375    /* Normally there is plenty of room in the current buffer for the command */
376    if (cvmx_likely(qptr->index + cmd_count < qptr->pool_size_m1))
377    {
378        uint64_t *ptr = (uint64_t *)cvmx_phys_to_ptr((uint64_t)qptr->base_ptr_div128<<7);
379        ptr += qptr->index;
380        qptr->index += cmd_count;
381        while (cmd_count--)
382            *ptr++ = *cmds++;
383    }
384    else
385    {
386        uint64_t *ptr;
387        int count;
388        /* We need a new command buffer. Fail if there isn't one available */
389        uint64_t *new_buffer = (uint64_t *)cvmx_fpa_alloc(qptr->fpa_pool);
390        if (cvmx_unlikely(new_buffer == NULL))
391        {
392            if (cvmx_likely(use_locking))
393                __cvmx_cmd_queue_unlock(qptr);
394            return CVMX_CMD_QUEUE_NO_MEMORY;
395        }
396        ptr = (uint64_t *)cvmx_phys_to_ptr((uint64_t)qptr->base_ptr_div128<<7);
397        /* Figure out how many command words will fit in this buffer. One
398            location will be needed for the next buffer pointer */
399        count = qptr->pool_size_m1 - qptr->index;
400        ptr += qptr->index;
401        cmd_count-=count;
402        while (count--)
403            *ptr++ = *cmds++;
404        *ptr = cvmx_ptr_to_phys(new_buffer);
405        /* The current buffer is full and has a link to the next buffer. Time
406            to write the rest of the commands into the new buffer */
407        qptr->base_ptr_div128 = *ptr >> 7;
408        qptr->index = cmd_count;
409        ptr = new_buffer;
410        while (cmd_count--)
411            *ptr++ = *cmds++;
412    }
413
414    /* All updates are complete. Release the lock and return */
415    if (cvmx_likely(use_locking))
416        __cvmx_cmd_queue_unlock(qptr);
417    return CVMX_CMD_QUEUE_SUCCESS;
418}
419
420
421/**
422 * Simple function to write two command words to a command
423 * queue.
424 *
425 * @param queue_id Hardware command queue to write to
426 * @param use_locking
427 *                 Use internal locking to ensure exclusive access for queue
428 *                 updates. If you don't use this locking you must ensure
429 *                 exclusivity some other way. Locking is strongly recommended.
430 * @param cmd1     Command
431 * @param cmd2     Command
432 *
433 * @return CVMX_CMD_QUEUE_SUCCESS or a failure code
434 */
435static inline cvmx_cmd_queue_result_t cvmx_cmd_queue_write2(cvmx_cmd_queue_id_t queue_id, int use_locking, uint64_t cmd1, uint64_t cmd2)
436{
437    __cvmx_cmd_queue_state_t *qptr = __cvmx_cmd_queue_get_state(queue_id);
438
439    if (CVMX_ENABLE_PARAMETER_CHECKING)
440    {
441        if (cvmx_unlikely(qptr == NULL))
442            return CVMX_CMD_QUEUE_INVALID_PARAM;
443    }
444
445    /* Make sure nobody else is updating the same queue */
446    if (cvmx_likely(use_locking))
447        __cvmx_cmd_queue_lock(queue_id, qptr);
448
449    /* If a max queue length was specified then make sure we don't
450        exceed it. If any part of the command would be below the limit
451        we allow it */
452    if (CVMX_CMD_QUEUE_ENABLE_MAX_DEPTH && cvmx_unlikely(qptr->max_depth))
453    {
454        if (cvmx_unlikely(cvmx_cmd_queue_length(queue_id) > (int)qptr->max_depth))
455        {
456            if (cvmx_likely(use_locking))
457                __cvmx_cmd_queue_unlock(qptr);
458            return CVMX_CMD_QUEUE_FULL;
459        }
460    }
461
462    /* Normally there is plenty of room in the current buffer for the command */
463    if (cvmx_likely(qptr->index + 2 < qptr->pool_size_m1))
464    {
465        uint64_t *ptr = (uint64_t *)cvmx_phys_to_ptr((uint64_t)qptr->base_ptr_div128<<7);
466        ptr += qptr->index;
467        qptr->index += 2;
468        ptr[0] = cmd1;
469        ptr[1] = cmd2;
470    }
471    else
472    {
473        uint64_t *ptr;
474        /* Figure out how many command words will fit in this buffer. One
475            location will be needed for the next buffer pointer */
476        int count = qptr->pool_size_m1 - qptr->index;
477        /* We need a new command buffer. Fail if there isn't one available */
478        uint64_t *new_buffer = (uint64_t *)cvmx_fpa_alloc(qptr->fpa_pool);
479        if (cvmx_unlikely(new_buffer == NULL))
480        {
481            if (cvmx_likely(use_locking))
482                __cvmx_cmd_queue_unlock(qptr);
483            return CVMX_CMD_QUEUE_NO_MEMORY;
484        }
485        count--;
486        ptr = (uint64_t *)cvmx_phys_to_ptr((uint64_t)qptr->base_ptr_div128<<7);
487        ptr += qptr->index;
488        *ptr++ = cmd1;
489        if (cvmx_likely(count))
490            *ptr++ = cmd2;
491        *ptr = cvmx_ptr_to_phys(new_buffer);
492        /* The current buffer is full and has a link to the next buffer. Time
493            to write the rest of the commands into the new buffer */
494        qptr->base_ptr_div128 = *ptr >> 7;
495        qptr->index = 0;
496        if (cvmx_unlikely(count == 0))
497        {
498            qptr->index = 1;
499            new_buffer[0] = cmd2;
500        }
501    }
502
503    /* All updates are complete. Release the lock and return */
504    if (cvmx_likely(use_locking))
505        __cvmx_cmd_queue_unlock(qptr);
506    return CVMX_CMD_QUEUE_SUCCESS;
507}
508
509
510/**
511 * Simple function to write three command words to a command
512 * queue.
513 *
514 * @param queue_id Hardware command queue to write to
515 * @param use_locking
516 *                 Use internal locking to ensure exclusive access for queue
517 *                 updates. If you don't use this locking you must ensure
518 *                 exclusivity some other way. Locking is strongly recommended.
519 * @param cmd1     Command
520 * @param cmd2     Command
521 * @param cmd3     Command
522 *
523 * @return CVMX_CMD_QUEUE_SUCCESS or a failure code
524 */
525static inline cvmx_cmd_queue_result_t cvmx_cmd_queue_write3(cvmx_cmd_queue_id_t queue_id, int use_locking, uint64_t cmd1, uint64_t cmd2, uint64_t cmd3)
526{
527    __cvmx_cmd_queue_state_t *qptr = __cvmx_cmd_queue_get_state(queue_id);
528
529    if (CVMX_ENABLE_PARAMETER_CHECKING)
530    {
531        if (cvmx_unlikely(qptr == NULL))
532            return CVMX_CMD_QUEUE_INVALID_PARAM;
533    }
534
535    /* Make sure nobody else is updating the same queue */
536    if (cvmx_likely(use_locking))
537        __cvmx_cmd_queue_lock(queue_id, qptr);
538
539    /* If a max queue length was specified then make sure we don't
540        exceed it. If any part of the command would be below the limit
541        we allow it */
542    if (CVMX_CMD_QUEUE_ENABLE_MAX_DEPTH && cvmx_unlikely(qptr->max_depth))
543    {
544        if (cvmx_unlikely(cvmx_cmd_queue_length(queue_id) > (int)qptr->max_depth))
545        {
546            if (cvmx_likely(use_locking))
547                __cvmx_cmd_queue_unlock(qptr);
548            return CVMX_CMD_QUEUE_FULL;
549        }
550    }
551
552    /* Normally there is plenty of room in the current buffer for the command */
553    if (cvmx_likely(qptr->index + 3 < qptr->pool_size_m1))
554    {
555        uint64_t *ptr = (uint64_t *)cvmx_phys_to_ptr((uint64_t)qptr->base_ptr_div128<<7);
556        ptr += qptr->index;
557        qptr->index += 3;
558        ptr[0] = cmd1;
559        ptr[1] = cmd2;
560        ptr[2] = cmd3;
561    }
562    else
563    {
564        uint64_t *ptr;
565        /* Figure out how many command words will fit in this buffer. One
566            location will be needed for the next buffer pointer */
567        int count = qptr->pool_size_m1 - qptr->index;
568        /* We need a new command buffer. Fail if there isn't one available */
569        uint64_t *new_buffer = (uint64_t *)cvmx_fpa_alloc(qptr->fpa_pool);
570        if (cvmx_unlikely(new_buffer == NULL))
571        {
572            if (cvmx_likely(use_locking))
573                __cvmx_cmd_queue_unlock(qptr);
574            return CVMX_CMD_QUEUE_NO_MEMORY;
575        }
576        count--;
577        ptr = (uint64_t *)cvmx_phys_to_ptr((uint64_t)qptr->base_ptr_div128<<7);
578        ptr += qptr->index;
579        *ptr++ = cmd1;
580        if (count)
581        {
582            *ptr++ = cmd2;
583            if (count > 1)
584                *ptr++ = cmd3;
585        }
586        *ptr = cvmx_ptr_to_phys(new_buffer);
587        /* The current buffer is full and has a link to the next buffer. Time
588            to write the rest of the commands into the new buffer */
589        qptr->base_ptr_div128 = *ptr >> 7;
590        qptr->index = 0;
591        ptr = new_buffer;
592        if (count == 0)
593        {
594            *ptr++ = cmd2;
595            qptr->index++;
596        }
597        if (count < 2)
598        {
599            *ptr++ = cmd3;
600            qptr->index++;
601        }
602    }
603
604    /* All updates are complete. Release the lock and return */
605    if (cvmx_likely(use_locking))
606        __cvmx_cmd_queue_unlock(qptr);
607    return CVMX_CMD_QUEUE_SUCCESS;
608}
609
610#ifdef	__cplusplus
611}
612#endif
613
614#endif /* __CVMX_CMD_QUEUE_H__ */
615