4 * Copyright (C) 2014-2016 Freescale Semiconductor, Inc.
6 * Redistribution and use in source and binary forms, with or without
7 * modification, are permitted provided that the following conditions are met:
8 * * Redistributions of source code must retain the above copyright
9 * notice, this list of conditions and the following disclaimer.
10 * * Redistributions in binary form must reproduce the above copyright
11 * notice, this list of conditions and the following disclaimer in the
12 * documentation and/or other materials provided with the distribution.
13 * * Neither the name of Freescale Semiconductor nor the
14 * names of its contributors may be used to endorse or promote products
15 * derived from this software without specific prior written permission.
17 * THIS SOFTWARE IS PROVIDED BY Freescale Semiconductor ``AS IS'' AND ANY
18 * EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
19 * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
20 * DISCLAIMED. IN NO EVENT SHALL Freescale Semiconductor BE LIABLE FOR ANY
21 * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
22 * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
23 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND
24 * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
25 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
26 * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
29 #include "qbman_private.h"
30 #include <fsl_qbman_portal.h>
32 uint32_t qman_version;
33 /* All QBMan command and result structures use this "valid bit" encoding */
34 #define QB_VALID_BIT ((uint32_t)0x80)
36 /* Management command result codes */
37 #define QBMAN_MC_RSLT_OK 0xf0
39 /* QBMan DQRR size is set at runtime in qbman_portal.c */
41 #define QBMAN_EQCR_SIZE 8
43 static inline u8 qm_cyc_diff(u8 ringsize, u8 first, u8 last)
45 /* 'first' is included, 'last' is excluded */
48 return (2 * ringsize) + last - first;
51 /* --------------------- */
52 /* portal data structure */
53 /* --------------------- */
56 struct qbman_swp_desc desc;
57 /* The qbman_sys (ie. arch/OS-specific) support code can put anything it
60 struct qbman_swp_sys sys;
61 /* Management commands */
65 swp_mc_can_start, /* call __qbman_swp_mc_start() */
66 swp_mc_can_submit, /* call __qbman_swp_mc_submit() */
67 swp_mc_can_poll, /* call __qbman_swp_mc_result() */
70 uint32_t valid_bit; /* 0x00 or 0x80 */
74 /* Volatile dequeues */
76 /* VDQCR supports a "1 deep pipeline", meaning that if you know
77 * the last-submitted command is already executing in the
78 * hardware (as evidenced by at least 1 valid dequeue result),
79 * you can write another dequeue command to the register, the
80 * hardware will start executing it as soon as the
81 * already-executing command terminates. (This minimises latency
82 * and stalls.) With that in mind, this "busy" variable refers
83 * to whether or not a command can be submitted, not whether or
84 * not a previously-submitted command is still executing. In
85 * other words, once proof is seen that the previously-submitted
86 * command is executing, "vdq" is no longer "busy".
89 uint32_t valid_bit; /* 0x00 or 0x80 */
90 /* We need to determine when vdq is no longer busy. This depends
91 * on whether the "busy" (last-submitted) dequeue command is
92 * targeting DQRR or main-memory, and detected is based on the
93 * presence of the dequeue command's "token" showing up in
94 * dequeue entries in DQRR or main-memory (respectively).
96 struct qbman_result *storage; /* NULL if DQRR */
113 /* -------------------------- */
114 /* portal management commands */
115 /* -------------------------- */
117 /* Different management commands all use this common base layer of code to issue
118 * commands and poll for results. The first function returns a pointer to where
119 * the caller should fill in their MC command (though they should ignore the
120 * verb byte), the second function commits merges in the caller-supplied command
121 * verb (which should not include the valid-bit) and submits the command to
122 * hardware, and the third function checks for a completed response (returns
123 * non-NULL if only if the response is complete).
125 void *qbman_swp_mc_start(struct qbman_swp *p);
126 void qbman_swp_mc_submit(struct qbman_swp *p, void *cmd, uint8_t cmd_verb);
127 void *qbman_swp_mc_result(struct qbman_swp *p);
129 /* Wraps up submit + poll-for-result */
130 static inline void *qbman_swp_mc_complete(struct qbman_swp *swp, void *cmd,
135 qbman_swp_mc_submit(swp, cmd, cmd_verb);
136 DBG_POLL_START(loopvar);
138 DBG_POLL_CHECK(loopvar);
139 cmd = qbman_swp_mc_result(swp);
148 /* This struct locates a sub-field within a QBMan portal (CENA) cacheline which
149 * is either serving as a configuration command or a query result. The
150 * representation is inherently little-endian, as the indexing of the words is
151 * itself little-endian in nature and DPAA2 QBMan is little endian for anything
152 * that crosses a word boundary too (64-bit fields are the obvious examples).
154 struct qb_attr_code {
155 unsigned int word; /* which uint32_t[] array member encodes the field */
156 unsigned int lsoffset; /* encoding offset from ls-bit */
157 unsigned int width; /* encoding width. (bool must be 1.) */
160 /* Some pre-defined codes */
161 extern struct qb_attr_code code_generic_verb;
162 extern struct qb_attr_code code_generic_rslt;
164 /* Macros to define codes */
165 #define QB_CODE(a, b, c) { a, b, c}
166 #define QB_CODE_NULL \
167 QB_CODE((unsigned int)-1, (unsigned int)-1, (unsigned int)-1)
169 /* Rotate a code "ms", meaning that it moves from less-significant bytes to
170 * more-significant, from less-significant words to more-significant, etc. The
171 * "ls" version does the inverse, from more-significant towards
174 static inline void qb_attr_code_rotate_ms(struct qb_attr_code *code,
177 code->lsoffset += bits;
178 while (code->lsoffset > 31) {
180 code->lsoffset -= 32;
184 static inline void qb_attr_code_rotate_ls(struct qb_attr_code *code,
187 /* Don't be fooled, this trick should work because the types are
188 * unsigned. So the case that interests the while loop (the rotate has
189 * gone too far and the word count needs to compensate for it), is
190 * manifested when lsoffset is negative. But that equates to a really
191 * large unsigned value, starting with lots of "F"s. As such, we can
192 * continue adding 32 back to it until it wraps back round above zero,
193 * to a value of 31 or less...
195 code->lsoffset -= bits;
196 while (code->lsoffset > 31) {
198 code->lsoffset += 32;
202 /* Implement a loop of code rotations until 'expr' evaluates to FALSE (0). */
203 #define qb_attr_code_for_ms(code, bits, expr) \
204 for (; expr; qb_attr_code_rotate_ms(code, bits))
205 #define qb_attr_code_for_ls(code, bits, expr) \
206 for (; expr; qb_attr_code_rotate_ls(code, bits))
208 /* decode a field from a cacheline */
209 static inline uint32_t qb_attr_code_decode(const struct qb_attr_code *code,
210 const uint32_t *cacheline)
212 return d32_uint32_t(code->lsoffset, code->width, cacheline[code->word]);
215 static inline uint64_t qb_attr_code_decode_64(const struct qb_attr_code *code,
216 const uint64_t *cacheline)
218 return cacheline[code->word / 2];
221 /* encode a field to a cacheline */
222 static inline void qb_attr_code_encode(const struct qb_attr_code *code,
223 uint32_t *cacheline, uint32_t val)
225 cacheline[code->word] =
226 r32_uint32_t(code->lsoffset, code->width, cacheline[code->word])
227 | e32_uint32_t(code->lsoffset, code->width, val);
230 static inline void qb_attr_code_encode_64(const struct qb_attr_code *code,
231 uint64_t *cacheline, uint64_t val)
233 cacheline[code->word / 2] = val;
236 /* Small-width signed values (two's-complement) will decode into medium-width
237 * positives. (Eg. for an 8-bit signed field, which stores values from -128 to
238 * +127, a setting of -7 would appear to decode to the 32-bit unsigned value
239 * 249. Likewise -120 would decode as 136.) This function allows the caller to
240 * "re-sign" such fields to 32-bit signed. (Eg. -7, which was 249 with an 8-bit
241 * encoding, will become 0xfffffff9 if you cast the return value to uint32_t).
243 static inline int32_t qb_attr_code_makesigned(const struct qb_attr_code *code,
246 QBMAN_BUG_ON(val >= (1u << code->width));
247 /* code->width should never exceed the width of val. If it does then a
248 * different function with larger val size must be used to translate
249 * from unsigned to signed
251 QBMAN_BUG_ON(code->width > sizeof(val) * CHAR_BIT);
252 /* If the high bit was set, it was encoding a negative */
253 if (val >= 1u << (code->width - 1))
254 return (int32_t)0 - (int32_t)(((uint32_t)1 << code->width) -
256 /* Otherwise, it was encoding a positive */
260 /* ---------------------- */
261 /* Descriptors/cachelines */
262 /* ---------------------- */
264 /* To avoid needless dynamic allocation, the driver API often gives the caller
265 * a "descriptor" type that the caller can instantiate however they like.
266 * Ultimately though, it is just a cacheline of binary storage (or something
267 * smaller when it is known that the descriptor doesn't need all 64 bytes) for
268 * holding pre-formatted pieces of hardware commands. The performance-critical
269 * code can then copy these descriptors directly into hardware command
270 * registers more efficiently than trying to construct/format commands
271 * on-the-fly. The API user sees the descriptor as an array of 32-bit words in
272 * order for the compiler to know its size, but the internal details are not
273 * exposed. The following macro is used within the driver for converting *any*
274 * descriptor pointer to a usable array pointer. The use of a macro (instead of
275 * an inline) is necessary to work with different descriptor types and to work
276 * correctly with const and non-const inputs (and similarly-qualified outputs).
278 #define qb_cl(d) (&(d)->donot_manipulate_directly[0])