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 /* --------------------- */
18 /* portal data structure */
19 /* --------------------- */
20 
21 struct qbman_swp {
22 	const struct qbman_swp_desc *desc;
23 	/* The qbman_sys (ie. arch/OS-specific) support code can put anything it
24 	 * needs in here. */
25 	struct qbman_swp_sys sys;
26 	/* Management commands */
27 	struct {
28 #ifdef QBMAN_CHECKING
29 		enum swp_mc_check {
30 			swp_mc_can_start, /* call __qbman_swp_mc_start() */
31 			swp_mc_can_submit, /* call __qbman_swp_mc_submit() */
32 			swp_mc_can_poll, /* call __qbman_swp_mc_result() */
33 		} check;
34 #endif
35 		uint32_t valid_bit; /* 0x00 or 0x80 */
36 	} mc;
37 	/* Push dequeues */
38 	uint32_t sdq;
39 	/* Volatile dequeues */
40 	struct {
41 		/* VDQCR supports a "1 deep pipeline", meaning that if you know
42 		 * the last-submitted command is already executing in the
43 		 * hardware (as evidenced by at least 1 valid dequeue result),
44 		 * you can write another dequeue command to the register, the
45 		 * hardware will start executing it as soon as the
46 		 * already-executing command terminates. (This minimises latency
47 		 * and stalls.) With that in mind, this "busy" variable refers
48 		 * to whether or not a command can be submitted, not whether or
49 		 * not a previously-submitted command is still executing. In
50 		 * other words, once proof is seen that the previously-submitted
51 		 * command is executing, "vdq" is no longer "busy". TODO:
52 		 * convert this to "atomic_t" so that it is thread-safe (without
53 		 * locking). */
54 		int busy;
55 		uint32_t valid_bit; /* 0x00 or 0x80 */
56 		/* We need to determine when vdq is no longer busy. This depends
57 		 * on whether the "busy" (last-submitted) dequeue command is
58 		 * targetting DQRR or main-memory, and detected is based on the
59 		 * presence of the dequeue command's "token" showing up in
60 		 * dequeue entries in DQRR or main-memory (respectively). Debug
61 		 * builds will, when submitting vdq commands, verify that the
62 		 * dequeue result location is not already equal to the command's
63 		 * token value. */
64 		struct ldpaa_dq *storage; /* NULL if DQRR */
65 		uint32_t token;
66 	} vdq;
67 	/* DQRR */
68 	struct {
69 		uint32_t next_idx;
70 		uint32_t valid_bit;
71 	} dqrr;
72 };
73 
74 /* -------------------------- */
75 /* portal management commands */
76 /* -------------------------- */
77 
78 /* Different management commands all use this common base layer of code to issue
79  * commands and poll for results. The first function returns a pointer to where
80  * the caller should fill in their MC command (though they should ignore the
81  * verb byte), the second function commits merges in the caller-supplied command
82  * verb (which should not include the valid-bit) and submits the command to
83  * hardware, and the third function checks for a completed response (returns
84  * non-NULL if only if the response is complete). */
85 void *qbman_swp_mc_start(struct qbman_swp *p);
86 void qbman_swp_mc_submit(struct qbman_swp *p, void *cmd, uint32_t cmd_verb);
87 void *qbman_swp_mc_result(struct qbman_swp *p);
88 
89 /* Wraps up submit + poll-for-result */
90 static inline void *qbman_swp_mc_complete(struct qbman_swp *swp, void *cmd,
91 					  uint32_t cmd_verb)
92 {
93 	int loopvar;
94 
95 	qbman_swp_mc_submit(swp, cmd, cmd_verb);
96 	DBG_POLL_START(loopvar);
97 	do {
98 		DBG_POLL_CHECK(loopvar);
99 		cmd = qbman_swp_mc_result(swp);
100 	} while (!cmd);
101 	return cmd;
102 }
103 
104 /* ------------ */
105 /* qb_attr_code */
106 /* ------------ */
107 
108 /* This struct locates a sub-field within a QBMan portal (CENA) cacheline which
109  * is either serving as a configuration command or a query result. The
110  * representation is inherently little-endian, as the indexing of the words is
111  * itself little-endian in nature and layerscape is little endian for anything
112  * that crosses a word boundary too (64-bit fields are the obvious examples).
113  */
114 struct qb_attr_code {
115 	unsigned int word; /* which uint32_t[] array member encodes the field */
116 	unsigned int lsoffset; /* encoding offset from ls-bit */
117 	unsigned int width; /* encoding width. (bool must be 1.) */
118 };
119 
120 /* Macros to define codes */
121 #define QB_CODE(a, b, c) { a, b, c}
122 
123 /* decode a field from a cacheline */
124 static inline uint32_t qb_attr_code_decode(const struct qb_attr_code *code,
125 				      const uint32_t *cacheline)
126 {
127 	return d32_uint32_t(code->lsoffset, code->width, cacheline[code->word]);
128 }
129 
130 /* encode a field to a cacheline */
131 static inline void qb_attr_code_encode(const struct qb_attr_code *code,
132 				       uint32_t *cacheline, uint32_t val)
133 {
134 	cacheline[code->word] =
135 		r32_uint32_t(code->lsoffset, code->width, cacheline[code->word])
136 		| e32_uint32_t(code->lsoffset, code->width, val);
137 }
138 
139 /* ---------------------- */
140 /* Descriptors/cachelines */
141 /* ---------------------- */
142 
143 /* To avoid needless dynamic allocation, the driver API often gives the caller
144  * a "descriptor" type that the caller can instantiate however they like.
145  * Ultimately though, it is just a cacheline of binary storage (or something
146  * smaller when it is known that the descriptor doesn't need all 64 bytes) for
147  * holding pre-formatted pieces of harware commands. The performance-critical
148  * code can then copy these descriptors directly into hardware command
149  * registers more efficiently than trying to construct/format commands
150  * on-the-fly. The API user sees the descriptor as an array of 32-bit words in
151  * order for the compiler to know its size, but the internal details are not
152  * exposed. The following macro is used within the driver for converting *any*
153  * descriptor pointer to a usable array pointer. The use of a macro (instead of
154  * an inline) is necessary to work with different descriptor types and to work
155  * correctly with const and non-const inputs (and similarly-qualified outputs).
156  */
157 #define qb_cl(d) (&(d)->dont_manipulate_directly[0])
158