1 /* SPDX-License-Identifier: BSD-3-Clause
3 * Copyright (C) 2014-2016 Freescale Semiconductor, Inc.
8 #include <fsl_qbman_portal.h>
10 #define QMAN_REV_4000 0x04000000
11 #define QMAN_REV_4100 0x04010000
12 #define QMAN_REV_4101 0x04010001
14 /* All QBMan command and result structures use this "valid bit" encoding */
15 #define QB_VALID_BIT ((uint32_t)0x80)
17 /* Management command result codes */
18 #define QBMAN_MC_RSLT_OK 0xf0
20 /* QBMan DQRR size is set at runtime in qbman_portal.c */
22 #define QBMAN_EQCR_SIZE 8
24 static inline uint8_t qm_cyc_diff(uint8_t ringsize, uint8_t first,
27 /* 'first' is included, 'last' is excluded */
30 return (2 * ringsize) + last - first;
33 /* --------------------- */
34 /* portal data structure */
35 /* --------------------- */
38 struct qbman_swp_desc desc;
39 /* The qbman_sys (ie. arch/OS-specific) support code can put anything it
42 struct qbman_swp_sys sys;
43 /* Management commands */
47 swp_mc_can_start, /* call __qbman_swp_mc_start() */
48 swp_mc_can_submit, /* call __qbman_swp_mc_submit() */
49 swp_mc_can_poll, /* call __qbman_swp_mc_result() */
52 uint32_t valid_bit; /* 0x00 or 0x80 */
56 /* Volatile dequeues */
58 /* VDQCR supports a "1 deep pipeline", meaning that if you know
59 * the last-submitted command is already executing in the
60 * hardware (as evidenced by at least 1 valid dequeue result),
61 * you can write another dequeue command to the register, the
62 * hardware will start executing it as soon as the
63 * already-executing command terminates. (This minimises latency
64 * and stalls.) With that in mind, this "busy" variable refers
65 * to whether or not a command can be submitted, not whether or
66 * not a previously-submitted command is still executing. In
67 * other words, once proof is seen that the previously-submitted
68 * command is executing, "vdq" is no longer "busy".
71 uint32_t valid_bit; /* 0x00 or 0x80 */
72 /* We need to determine when vdq is no longer busy. This depends
73 * on whether the "busy" (last-submitted) dequeue command is
74 * targeting DQRR or main-memory, and detected is based on the
75 * presence of the dequeue command's "token" showing up in
76 * dequeue entries in DQRR or main-memory (respectively).
78 struct qbman_result *storage; /* NULL if DQRR */
95 /* -------------------------- */
96 /* portal management commands */
97 /* -------------------------- */
99 /* Different management commands all use this common base layer of code to issue
100 * commands and poll for results. The first function returns a pointer to where
101 * the caller should fill in their MC command (though they should ignore the
102 * verb byte), the second function commits merges in the caller-supplied command
103 * verb (which should not include the valid-bit) and submits the command to
104 * hardware, and the third function checks for a completed response (returns
105 * non-NULL if only if the response is complete).
107 void *qbman_swp_mc_start(struct qbman_swp *p);
108 void qbman_swp_mc_submit(struct qbman_swp *p, void *cmd, uint8_t cmd_verb);
109 void *qbman_swp_mc_result(struct qbman_swp *p);
111 /* Wraps up submit + poll-for-result */
112 static inline void *qbman_swp_mc_complete(struct qbman_swp *swp, void *cmd,
117 qbman_swp_mc_submit(swp, cmd, cmd_verb);
119 cmd = qbman_swp_mc_result(swp);
120 } while (!cmd && loopvar--);
121 QBMAN_BUG_ON(!loopvar);
126 /* ---------------------- */
127 /* Descriptors/cachelines */
128 /* ---------------------- */
130 /* To avoid needless dynamic allocation, the driver API often gives the caller
131 * a "descriptor" type that the caller can instantiate however they like.
132 * Ultimately though, it is just a cacheline of binary storage (or something
133 * smaller when it is known that the descriptor doesn't need all 64 bytes) for
134 * holding pre-formatted pieces of hardware commands. The performance-critical
135 * code can then copy these descriptors directly into hardware command
136 * registers more efficiently than trying to construct/format commands
137 * on-the-fly. The API user sees the descriptor as an array of 32-bit words in
138 * order for the compiler to know its size, but the internal details are not
139 * exposed. The following macro is used within the driver for converting *any*
140 * descriptor pointer to a usable array pointer. The use of a macro (instead of
141 * an inline) is necessary to work with different descriptor types and to work
142 * correctly with const and non-const inputs (and similarly-qualified outputs).
144 #define qb_cl(d) (&(d)->donot_manipulate_directly[0])