• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /*
2  * Copyright 2012 Tilera Corporation. All Rights Reserved.
3  *
4  *   This program is free software; you can redistribute it and/or
5  *   modify it under the terms of the GNU General Public License
6  *   as published by the Free Software Foundation, version 2.
7  *
8  *   This program is distributed in the hope that it will be useful, but
9  *   WITHOUT ANY WARRANTY; without even the implied warranty of
10  *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
11  *   NON INFRINGEMENT.  See the GNU General Public License for
12  *   more details.
13  */
14 
15 /*
16  *
17  * An API for allocating, configuring, and manipulating TRIO hardware
18  * resources
19  */
20 
21 /*
22  *
23  * The TILE-Gx TRIO shim provides connections to external devices via
24  * PCIe or other transaction IO standards.  The gxio_trio_ API,
25  * declared in <gxio/trio.h>, allows applications to allocate and
26  * configure TRIO IO resources like DMA command rings, memory map
27  * windows, and device interrupts.  The following sections introduce
28  * the various components of the API.  We strongly recommend reading
29  * the TRIO section of the IO Device Guide (UG404) before working with
30  * this API.
31  *
32  * @section trio__ingress TRIO Ingress Hardware Resources
33  *
34  * The TRIO ingress hardware is responsible for examining incoming
35  * PCIe or StreamIO packets and choosing a processing mechanism based
36  * on the packets' bus address.  The gxio_trio_ API can be used to
37  * configure different handlers for different ranges of bus address
38  * space.  The user can configure "mapped memory" and "scatter queue"
39  * regions to match incoming packets within 4kB-aligned ranges of bus
40  * addresses.  Each range specifies a different set of mapping
41  * parameters to be applied when handling the ingress packet.  The
42  * following sections describe how to work with MapMem and scatter
43  * queue regions.
44  *
45  * @subsection trio__mapmem TRIO MapMem Regions
46  *
47  * TRIO mapped memory (or MapMem) regions allow the user to map
48  * incoming read and write requests directly to the application's
49  * memory space.  MapMem regions are allocated via
50  * gxio_trio_alloc_memory_maps().  Given an integer MapMem number,
51  * applications can use gxio_trio_init_memory_map() to specify the
52  * range of bus addresses that will match the region and the range of
53  * virtual addresses to which those packets will be applied.
54  *
55  * As with many other gxio APIs, the programmer must be sure to
56  * register memory pages that will be used with MapMem regions.  Pages
57  * can be registered with TRIO by allocating an ASID (address space
58  * identifier) and then using gxio_trio_register_page() to register up to
59  * 16 pages with the hardware.  The initialization functions for
60  * resources that require registered memory (MapMem, scatter queues,
61  * push DMA, and pull DMA) then take an 'asid' parameter in order to
62  * configure which set of registered pages is used by each resource.
63  *
64  * @subsection trio__scatter_queue TRIO Scatter Queues
65  *
66  * The TRIO shim's scatter queue regions allow users to dynamically
67  * map buffers from a large address space into a small range of bus
68  * addresses.  This is particularly helpful for PCIe endpoint devices,
69  * where the host generally limits the size of BARs to tens of
70  * megabytes.
71  *
72  * Each scatter queue consists of a memory map region, a queue of
73  * tile-side buffer VAs to be mapped to that region, and a bus-mapped
74  * "doorbell" register that the remote endpoint can write to trigger a
75  * dequeue of the current buffer VA, thus swapping in a new buffer.
76  * The VAs pushed onto a scatter queue must be 4kB aligned, so
77  * applications may need to use higher-level protocols to inform
78  * remote entities that they should apply some additional, sub-4kB
79  * offset when reading or writing the scatter queue region.  For more
80  * information, see the IO Device Guide (UG404).
81  *
82  * @section trio__egress TRIO Egress Hardware Resources
83  *
84  * The TRIO shim supports two mechanisms for egress packet generation:
85  * programmed IO (PIO) and push/pull DMA.  PIO allows applications to
86  * create MMIO mappings for PCIe or StreamIO address space, such that
87  * the application can generate word-sized read or write transactions
88  * by issuing load or store instructions.  Push and pull DMA are tuned
89  * for larger transactions; they use specialized hardware engines to
90  * transfer large blocks of data at line rate.
91  *
92  * @subsection trio__pio TRIO Programmed IO
93  *
94  * Programmed IO allows applications to create MMIO mappings for PCIe
95  * or StreamIO address space.  The hardware PIO regions support access
96  * to PCIe configuration, IO, and memory space, but the gxio_trio API
97  * only supports memory space accesses.  PIO regions are allocated
98  * with gxio_trio_alloc_pio_regions() and initialized via
99  * gxio_trio_init_pio_region().  Once a region is bound to a range of
100  * bus address via the initialization function, the application can
101  * use gxio_trio_map_pio_region() to create MMIO mappings from its VA
102  * space onto the range of bus addresses supported by the PIO region.
103  *
104  * @subsection trio_dma TRIO Push and Pull DMA
105  *
106  * The TRIO push and pull DMA engines allow users to copy blocks of
107  * data between application memory and the bus.  Push DMA generates
108  * write packets that copy from application memory to the bus and pull
109  * DMA generates read packets that copy from the bus into application
110  * memory.  The DMA engines are managed via an API that is very
111  * similar to the mPIPE eDMA interface.  For a detailed explanation of
112  * the eDMA queue API, see @ref gxio_mpipe_wrappers.
113  *
114  * Push and pull DMA queues are allocated via
115  * gxio_trio_alloc_push_dma_ring() / gxio_trio_alloc_pull_dma_ring().
116  * Once allocated, users generally use a ::gxio_trio_dma_queue_t
117  * object to manage the queue, providing easy wrappers for reserving
118  * command slots in the DMA command ring, filling those slots, and
119  * waiting for commands to complete.  DMA queues can be initialized
120  * via gxio_trio_init_push_dma_queue() or
121  * gxio_trio_init_pull_dma_queue().
122  *
123  * See @ref trio/push_dma/app.c for an example of how to use push DMA.
124  *
125  * @section trio_shortcomings Plans for Future API Revisions
126  *
127  * The simulation framework is incomplete.  Future features include:
128  *
129  * - Support for reset and deallocation of resources.
130  *
131  * - Support for pull DMA.
132  *
133  * - Support for interrupt regions and user-space interrupt delivery.
134  *
135  * - Support for getting BAR mappings and reserving regions of BAR
136  *   address space.
137  */
138 #ifndef _GXIO_TRIO_H_
139 #define _GXIO_TRIO_H_
140 
141 #include <linux/types.h>
142 
143 #include <gxio/common.h>
144 #include <gxio/dma_queue.h>
145 
146 #include <arch/trio_constants.h>
147 #include <arch/trio.h>
148 #include <arch/trio_pcie_intfc.h>
149 #include <arch/trio_pcie_rc.h>
150 #include <arch/trio_shm.h>
151 #include <hv/drv_trio_intf.h>
152 #include <hv/iorpc.h>
153 
154 /* A context object used to manage TRIO hardware resources. */
155 typedef struct {
156 
157 	/* File descriptor for calling up to Linux (and thus the HV). */
158 	int fd;
159 
160 	/* The VA at which the MAC MMIO registers are mapped. */
161 	char *mmio_base_mac;
162 
163 	/* The VA at which the PIO config space are mapped for each PCIe MAC.
164 	   Gx36 has max 3 PCIe MACs per TRIO shim. */
165 	char *mmio_base_pio_cfg[TILEGX_TRIO_PCIES];
166 
167 #ifdef USE_SHARED_PCIE_CONFIG_REGION
168 	/* Index of the shared PIO region for PCI config access. */
169 	int pio_cfg_index;
170 #else
171 	/* Index of the PIO region for PCI config access per MAC. */
172 	int pio_cfg_index[TILEGX_TRIO_PCIES];
173 #endif
174 
175 	/*  The VA at which the push DMA MMIO registers are mapped. */
176 	char *mmio_push_dma[TRIO_NUM_PUSH_DMA_RINGS];
177 
178 	/*  The VA at which the pull DMA MMIO registers are mapped. */
179 	char *mmio_pull_dma[TRIO_NUM_PUSH_DMA_RINGS];
180 
181 	/* Application space ID. */
182 	unsigned int asid;
183 
184 } gxio_trio_context_t;
185 
186 /* Command descriptor for push or pull DMA. */
187 typedef TRIO_DMA_DESC_t gxio_trio_dma_desc_t;
188 
189 /* A convenient, thread-safe interface to an eDMA ring. */
190 typedef struct {
191 
192 	/* State object for tracking head and tail pointers. */
193 	__gxio_dma_queue_t dma_queue;
194 
195 	/* The ring entries. */
196 	gxio_trio_dma_desc_t *dma_descs;
197 
198 	/* The number of entries minus one. */
199 	unsigned long mask_num_entries;
200 
201 	/* The log2() of the number of entries. */
202 	unsigned int log2_num_entries;
203 
204 } gxio_trio_dma_queue_t;
205 
206 /* Initialize a TRIO context.
207  *
208  * This function allocates a TRIO "service domain" and maps the MMIO
209  * registers into the the caller's VA space.
210  *
211  * @param trio_index Which TRIO shim; Gx36 must pass 0.
212  * @param context Context object to be initialized.
213  */
214 extern int gxio_trio_init(gxio_trio_context_t *context,
215 			  unsigned int trio_index);
216 
217 /* This indicates that an ASID hasn't been allocated. */
218 #define GXIO_ASID_NULL -1
219 
220 /* Ordering modes for map memory regions and scatter queue regions. */
221 typedef enum gxio_trio_order_mode_e {
222 	/* Writes are not ordered.  Reads always wait for previous writes. */
223 	GXIO_TRIO_ORDER_MODE_UNORDERED =
224 		TRIO_MAP_MEM_SETUP__ORDER_MODE_VAL_UNORDERED,
225 	/* Both writes and reads wait for previous transactions to complete. */
226 	GXIO_TRIO_ORDER_MODE_STRICT =
227 		TRIO_MAP_MEM_SETUP__ORDER_MODE_VAL_STRICT,
228 	/* Writes are ordered unless the incoming packet has the
229 	   relaxed-ordering attributes set. */
230 	GXIO_TRIO_ORDER_MODE_OBEY_PACKET =
231 		TRIO_MAP_MEM_SETUP__ORDER_MODE_VAL_REL_ORD
232 } gxio_trio_order_mode_t;
233 
234 /* Initialize a memory mapping region.
235  *
236  * @param context An initialized TRIO context.
237  * @param map A Memory map region allocated by gxio_trio_alloc_memory_map().
238  * @param target_mem VA of backing memory, should be registered via
239  *   gxio_trio_register_page() and aligned to 4kB.
240  * @param target_size Length of the memory mapping, must be a multiple
241  * of 4kB.
242  * @param asid ASID to be used for Tile-side address translation.
243  * @param mac MAC number.
244  * @param bus_address Bus address at which the mapping starts.
245  * @param order_mode Memory ordering mode for this mapping.
246  * @return Zero on success, else ::GXIO_TRIO_ERR_BAD_MEMORY_MAP,
247  * GXIO_TRIO_ERR_BAD_ASID, or ::GXIO_TRIO_ERR_BAD_BUS_RANGE.
248  */
249 extern int gxio_trio_init_memory_map(gxio_trio_context_t *context,
250 				     unsigned int map, void *target_mem,
251 				     size_t target_size, unsigned int asid,
252 				     unsigned int mac, uint64_t bus_address,
253 				     gxio_trio_order_mode_t order_mode);
254 
255 /* Flags that can be passed to resource allocation functions. */
256 enum gxio_trio_alloc_flags_e {
257 	GXIO_TRIO_ALLOC_FIXED = HV_TRIO_ALLOC_FIXED,
258 };
259 
260 /* Flags that can be passed to memory registration functions. */
261 enum gxio_trio_mem_flags_e {
262 	/* Do not fill L3 when writing, and invalidate lines upon egress. */
263 	GXIO_TRIO_MEM_FLAG_NT_HINT = IORPC_MEM_BUFFER_FLAG_NT_HINT,
264 
265 	/* L3 cache fills should only populate IO cache ways. */
266 	GXIO_TRIO_MEM_FLAG_IO_PIN = IORPC_MEM_BUFFER_FLAG_IO_PIN,
267 };
268 
269 /* Flag indicating a request generator uses a special traffic
270     class. */
271 #define GXIO_TRIO_FLAG_TRAFFIC_CLASS(N) HV_TRIO_FLAG_TC(N)
272 
273 /* Flag indicating a request generator uses a virtual function
274     number. */
275 #define GXIO_TRIO_FLAG_VFUNC(N) HV_TRIO_FLAG_VFUNC(N)
276 
277 /*****************************************************************
278  *                       Memory Registration                      *
279  ******************************************************************/
280 
281 /* Allocate Application Space Identifiers (ASIDs).  Each ASID can
282  * register up to 16 page translations.  ASIDs are used by memory map
283  * regions, scatter queues, and DMA queues to translate application
284  * VAs into memory system PAs.
285  *
286  * @param context An initialized TRIO context.
287  * @param count Number of ASIDs required.
288  * @param first Index of first ASID if ::GXIO_TRIO_ALLOC_FIXED flag
289  *   is set, otherwise ignored.
290  * @param flags Flag bits, including bits from ::gxio_trio_alloc_flags_e.
291  * @return Index of first ASID, or ::GXIO_TRIO_ERR_NO_ASID if allocation
292  *   failed.
293  */
294 extern int gxio_trio_alloc_asids(gxio_trio_context_t *context,
295 				 unsigned int count, unsigned int first,
296 				 unsigned int flags);
297 
298 #endif /* ! _GXIO_TRIO_H_ */
299