1 /*
2  * Copyright (C) 2014 Freescale Semiconductor
3  *
4  * SPDX-License-Identifier:	GPL-2.0+
5  */
6 
7 #include "qbman_private.h"
8 #include <fsl-mc/fsl_qbman_portal.h>
9 #include <fsl-mc/fsl_dpaa_fd.h>
10 
11 /* All QBMan command and result structures use this "valid bit" encoding */
12 #define QB_VALID_BIT ((uint32_t)0x80)
13 
14 /* Management command result codes */
15 #define QBMAN_MC_RSLT_OK      0xf0
16 
17 /* TBD: as of QBMan 4.1, DQRR will be 8 rather than 4! */
18 #define QBMAN_DQRR_SIZE 4
19 
20 
21 /* --------------------- */
22 /* portal data structure */
23 /* --------------------- */
24 
25 struct qbman_swp {
26 	const struct qbman_swp_desc *desc;
27 	/* The qbman_sys (ie. arch/OS-specific) support code can put anything it
28 	 * needs in here. */
29 	struct qbman_swp_sys sys;
30 	/* Management commands */
31 	struct {
32 #ifdef QBMAN_CHECKING
33 		enum swp_mc_check {
34 			swp_mc_can_start, /* call __qbman_swp_mc_start() */
35 			swp_mc_can_submit, /* call __qbman_swp_mc_submit() */
36 			swp_mc_can_poll, /* call __qbman_swp_mc_result() */
37 		} check;
38 #endif
39 		uint32_t valid_bit; /* 0x00 or 0x80 */
40 	} mc;
41 	/* Push dequeues */
42 	uint32_t sdq;
43 	/* Volatile dequeues */
44 	struct {
45 		/* VDQCR supports a "1 deep pipeline", meaning that if you know
46 		 * the last-submitted command is already executing in the
47 		 * hardware (as evidenced by at least 1 valid dequeue result),
48 		 * you can write another dequeue command to the register, the
49 		 * hardware will start executing it as soon as the
50 		 * already-executing command terminates. (This minimises latency
51 		 * and stalls.) With that in mind, this "busy" variable refers
52 		 * to whether or not a command can be submitted, not whether or
53 		 * not a previously-submitted command is still executing. In
54 		 * other words, once proof is seen that the previously-submitted
55 		 * command is executing, "vdq" is no longer "busy".
56 		 */
57 		atomic_t busy;
58 		uint32_t valid_bit; /* 0x00 or 0x80 */
59 		/* We need to determine when vdq is no longer busy. This depends
60 		 * on whether the "busy" (last-submitted) dequeue command is
61 		 * targeting DQRR or main-memory, and detected is based on the
62 		 * presence of the dequeue command's "token" showing up in
63 		 * dequeue entries in DQRR or main-memory (respectively). Debug
64 		 * builds will, when submitting vdq commands, verify that the
65 		 * dequeue result location is not already equal to the command's
66 		 * token value. */
67 		struct ldpaa_dq *storage; /* NULL if DQRR */
68 		uint32_t token;
69 	} vdq;
70 	/* DQRR */
71 	struct {
72 		uint32_t next_idx;
73 		uint32_t valid_bit;
74 	} dqrr;
75 };
76 
77 /* -------------------------- */
78 /* portal management commands */
79 /* -------------------------- */
80 
81 /* Different management commands all use this common base layer of code to issue
82  * commands and poll for results. The first function returns a pointer to where
83  * the caller should fill in their MC command (though they should ignore the
84  * verb byte), the second function commits merges in the caller-supplied command
85  * verb (which should not include the valid-bit) and submits the command to
86  * hardware, and the third function checks for a completed response (returns
87  * non-NULL if only if the response is complete). */
88 void *qbman_swp_mc_start(struct qbman_swp *p);
89 void qbman_swp_mc_submit(struct qbman_swp *p, void *cmd, uint32_t cmd_verb);
90 void *qbman_swp_mc_result(struct qbman_swp *p);
91 
92 /* Wraps up submit + poll-for-result */
93 static inline void *qbman_swp_mc_complete(struct qbman_swp *swp, void *cmd,
94 					  uint32_t cmd_verb)
95 {
96 	int loopvar;
97 
98 	qbman_swp_mc_submit(swp, cmd, cmd_verb);
99 	DBG_POLL_START(loopvar);
100 	do {
101 		DBG_POLL_CHECK(loopvar);
102 		cmd = qbman_swp_mc_result(swp);
103 	} while (!cmd);
104 	return cmd;
105 }
106 
107 /* ------------ */
108 /* qb_attr_code */
109 /* ------------ */
110 
111 /* This struct locates a sub-field within a QBMan portal (CENA) cacheline which
112  * is either serving as a configuration command or a query result. The
113  * representation is inherently little-endian, as the indexing of the words is
114  * itself little-endian in nature and layerscape is little endian for anything
115  * that crosses a word boundary too (64-bit fields are the obvious examples).
116  */
117 struct qb_attr_code {
118 	unsigned int word; /* which uint32_t[] array member encodes the field */
119 	unsigned int lsoffset; /* encoding offset from ls-bit */
120 	unsigned int width; /* encoding width. (bool must be 1.) */
121 };
122 
123 /* Macros to define codes */
124 #define QB_CODE(a, b, c) { a, b, c}
125 
126 /* decode a field from a cacheline */
127 static inline uint32_t qb_attr_code_decode(const struct qb_attr_code *code,
128 				      const uint32_t *cacheline)
129 {
130 	return d32_uint32_t(code->lsoffset, code->width, cacheline[code->word]);
131 }
132 
133 
134 /* encode a field to a cacheline */
135 static inline void qb_attr_code_encode(const struct qb_attr_code *code,
136 				       uint32_t *cacheline, uint32_t val)
137 {
138 	cacheline[code->word] =
139 		r32_uint32_t(code->lsoffset, code->width, cacheline[code->word])
140 		| e32_uint32_t(code->lsoffset, code->width, val);
141 }
142 
143 static inline void qb_attr_code_encode_64(const struct qb_attr_code *code,
144 				       uint64_t *cacheline, uint64_t val)
145 {
146 	cacheline[code->word / 2] = val;
147 }
148 
149 /* ---------------------- */
150 /* Descriptors/cachelines */
151 /* ---------------------- */
152 
153 /* To avoid needless dynamic allocation, the driver API often gives the caller
154  * a "descriptor" type that the caller can instantiate however they like.
155  * Ultimately though, it is just a cacheline of binary storage (or something
156  * smaller when it is known that the descriptor doesn't need all 64 bytes) for
157  * holding pre-formatted pieces of hardware commands. The performance-critical
158  * code can then copy these descriptors directly into hardware command
159  * registers more efficiently than trying to construct/format commands
160  * on-the-fly. The API user sees the descriptor as an array of 32-bit words in
161  * order for the compiler to know its size, but the internal details are not
162  * exposed. The following macro is used within the driver for converting *any*
163  * descriptor pointer to a usable array pointer. The use of a macro (instead of
164  * an inline) is necessary to work with different descriptor types and to work
165  * correctly with const and non-const inputs (and similarly-qualified outputs).
166  */
167 #define qb_cl(d) (&(d)->dont_manipulate_directly[0])
168