1 | /* SPDX-License-Identifier: GPL-2.0 */ |
2 | /****************************************************************************** |
3 | * ring.h |
4 | * |
5 | * Shared producer-consumer ring macros. |
6 | * |
7 | * Tim Deegan and Andrew Warfield November 2004. |
8 | */ |
9 | |
10 | #ifndef __XEN_PUBLIC_IO_RING_H__ |
11 | #define __XEN_PUBLIC_IO_RING_H__ |
12 | |
13 | #include <xen/interface/grant_table.h> |
14 | |
15 | typedef unsigned int RING_IDX; |
16 | |
17 | /* Round a 32-bit unsigned constant down to the nearest power of two. */ |
18 | #define __RD2(_x) (((_x) & 0x00000002) ? 0x2 : ((_x) & 0x1)) |
19 | #define __RD4(_x) (((_x) & 0x0000000c) ? __RD2((_x)>>2)<<2 : __RD2(_x)) |
20 | #define __RD8(_x) (((_x) & 0x000000f0) ? __RD4((_x)>>4)<<4 : __RD4(_x)) |
21 | #define __RD16(_x) (((_x) & 0x0000ff00) ? __RD8((_x)>>8)<<8 : __RD8(_x)) |
22 | #define __RD32(_x) (((_x) & 0xffff0000) ? __RD16((_x)>>16)<<16 : __RD16(_x)) |
23 | |
24 | /* |
25 | * Calculate size of a shared ring, given the total available space for the |
26 | * ring and indexes (_sz), and the name tag of the request/response structure. |
27 | * A ring contains as many entries as will fit, rounded down to the nearest |
28 | * power of two (so we can mask with (size-1) to loop around). |
29 | */ |
30 | #define __CONST_RING_SIZE(_s, _sz) \ |
31 | (__RD32(((_sz) - offsetof(struct _s##_sring, ring)) / \ |
32 | sizeof(((struct _s##_sring *)0)->ring[0]))) |
33 | |
34 | /* |
35 | * The same for passing in an actual pointer instead of a name tag. |
36 | */ |
37 | #define __RING_SIZE(_s, _sz) \ |
38 | (__RD32(((_sz) - (long)&(_s)->ring + (long)(_s)) / sizeof((_s)->ring[0]))) |
39 | |
40 | /* |
41 | * Macros to make the correct C datatypes for a new kind of ring. |
42 | * |
43 | * To make a new ring datatype, you need to have two message structures, |
44 | * let's say struct request, and struct response already defined. |
45 | * |
46 | * In a header where you want the ring datatype declared, you then do: |
47 | * |
48 | * DEFINE_RING_TYPES(mytag, struct request, struct response); |
49 | * |
50 | * These expand out to give you a set of types, as you can see below. |
51 | * The most important of these are: |
52 | * |
53 | * struct mytag_sring - The shared ring. |
54 | * struct mytag_front_ring - The 'front' half of the ring. |
55 | * struct mytag_back_ring - The 'back' half of the ring. |
56 | * |
57 | * To initialize a ring in your code you need to know the location and size |
58 | * of the shared memory area (PAGE_SIZE, for instance). To initialise |
59 | * the front half: |
60 | * |
61 | * struct mytag_front_ring front_ring; |
62 | * SHARED_RING_INIT((struct mytag_sring *)shared_page); |
63 | * FRONT_RING_INIT(&front_ring, (struct mytag_sring *)shared_page, |
64 | * PAGE_SIZE); |
65 | * |
66 | * Initializing the back follows similarly (note that only the front |
67 | * initializes the shared ring): |
68 | * |
69 | * struct mytag_back_ring back_ring; |
70 | * BACK_RING_INIT(&back_ring, (struct mytag_sring *)shared_page, |
71 | * PAGE_SIZE); |
72 | */ |
73 | |
74 | #define DEFINE_RING_TYPES(__name, __req_t, __rsp_t) \ |
75 | \ |
76 | /* Shared ring entry */ \ |
77 | union __name##_sring_entry { \ |
78 | __req_t req; \ |
79 | __rsp_t rsp; \ |
80 | }; \ |
81 | \ |
82 | /* Shared ring page */ \ |
83 | struct __name##_sring { \ |
84 | RING_IDX req_prod, req_event; \ |
85 | RING_IDX rsp_prod, rsp_event; \ |
86 | uint8_t pad[48]; \ |
87 | union __name##_sring_entry ring[1]; /* variable-length */ \ |
88 | }; \ |
89 | \ |
90 | /* "Front" end's private variables */ \ |
91 | struct __name##_front_ring { \ |
92 | RING_IDX req_prod_pvt; \ |
93 | RING_IDX rsp_cons; \ |
94 | unsigned int nr_ents; \ |
95 | struct __name##_sring *sring; \ |
96 | }; \ |
97 | \ |
98 | /* "Back" end's private variables */ \ |
99 | struct __name##_back_ring { \ |
100 | RING_IDX rsp_prod_pvt; \ |
101 | RING_IDX req_cons; \ |
102 | unsigned int nr_ents; \ |
103 | struct __name##_sring *sring; \ |
104 | }; |
105 | |
106 | /* |
107 | * Macros for manipulating rings. |
108 | * |
109 | * FRONT_RING_whatever works on the "front end" of a ring: here |
110 | * requests are pushed on to the ring and responses taken off it. |
111 | * |
112 | * BACK_RING_whatever works on the "back end" of a ring: here |
113 | * requests are taken off the ring and responses put on. |
114 | * |
115 | * N.B. these macros do NO INTERLOCKS OR FLOW CONTROL. |
116 | * This is OK in 1-for-1 request-response situations where the |
117 | * requestor (front end) never has more than RING_SIZE()-1 |
118 | * outstanding requests. |
119 | */ |
120 | |
121 | /* Initialising empty rings */ |
122 | #define SHARED_RING_INIT(_s) do { \ |
123 | (_s)->req_prod = (_s)->rsp_prod = 0; \ |
124 | (_s)->req_event = (_s)->rsp_event = 1; \ |
125 | memset((_s)->pad, 0, sizeof((_s)->pad)); \ |
126 | } while(0) |
127 | |
128 | #define FRONT_RING_INIT(_r, _s, __size) do { \ |
129 | (_r)->req_prod_pvt = 0; \ |
130 | (_r)->rsp_cons = 0; \ |
131 | (_r)->nr_ents = __RING_SIZE(_s, __size); \ |
132 | (_r)->sring = (_s); \ |
133 | } while (0) |
134 | |
135 | #define BACK_RING_INIT(_r, _s, __size) do { \ |
136 | (_r)->rsp_prod_pvt = 0; \ |
137 | (_r)->req_cons = 0; \ |
138 | (_r)->nr_ents = __RING_SIZE(_s, __size); \ |
139 | (_r)->sring = (_s); \ |
140 | } while (0) |
141 | |
142 | /* Initialize to existing shared indexes -- for recovery */ |
143 | #define FRONT_RING_ATTACH(_r, _s, __size) do { \ |
144 | (_r)->sring = (_s); \ |
145 | (_r)->req_prod_pvt = (_s)->req_prod; \ |
146 | (_r)->rsp_cons = (_s)->rsp_prod; \ |
147 | (_r)->nr_ents = __RING_SIZE(_s, __size); \ |
148 | } while (0) |
149 | |
150 | #define BACK_RING_ATTACH(_r, _s, __size) do { \ |
151 | (_r)->sring = (_s); \ |
152 | (_r)->rsp_prod_pvt = (_s)->rsp_prod; \ |
153 | (_r)->req_cons = (_s)->req_prod; \ |
154 | (_r)->nr_ents = __RING_SIZE(_s, __size); \ |
155 | } while (0) |
156 | |
157 | /* How big is this ring? */ |
158 | #define RING_SIZE(_r) \ |
159 | ((_r)->nr_ents) |
160 | |
161 | /* Number of free requests (for use on front side only). */ |
162 | #define RING_FREE_REQUESTS(_r) \ |
163 | (RING_SIZE(_r) - ((_r)->req_prod_pvt - (_r)->rsp_cons)) |
164 | |
165 | /* Test if there is an empty slot available on the front ring. |
166 | * (This is only meaningful from the front. ) |
167 | */ |
168 | #define RING_FULL(_r) \ |
169 | (RING_FREE_REQUESTS(_r) == 0) |
170 | |
171 | /* Test if there are outstanding messages to be processed on a ring. */ |
172 | #define RING_HAS_UNCONSUMED_RESPONSES(_r) \ |
173 | ((_r)->sring->rsp_prod - (_r)->rsp_cons) |
174 | |
175 | #define RING_HAS_UNCONSUMED_REQUESTS(_r) \ |
176 | ({ \ |
177 | unsigned int req = (_r)->sring->req_prod - (_r)->req_cons; \ |
178 | unsigned int rsp = RING_SIZE(_r) - \ |
179 | ((_r)->req_cons - (_r)->rsp_prod_pvt); \ |
180 | req < rsp ? req : rsp; \ |
181 | }) |
182 | |
183 | /* Direct access to individual ring elements, by index. */ |
184 | #define RING_GET_REQUEST(_r, _idx) \ |
185 | (&((_r)->sring->ring[((_idx) & (RING_SIZE(_r) - 1))].req)) |
186 | |
187 | /* |
188 | * Get a local copy of a request. |
189 | * |
190 | * Use this in preference to RING_GET_REQUEST() so all processing is |
191 | * done on a local copy that cannot be modified by the other end. |
192 | * |
193 | * Note that https://gcc.gnu.org/bugzilla/show_bug.cgi?id=58145 may cause this |
194 | * to be ineffective where _req is a struct which consists of only bitfields. |
195 | */ |
196 | #define RING_COPY_REQUEST(_r, _idx, _req) do { \ |
197 | /* Use volatile to force the copy into _req. */ \ |
198 | *(_req) = *(volatile typeof(_req))RING_GET_REQUEST(_r, _idx); \ |
199 | } while (0) |
200 | |
201 | #define RING_GET_RESPONSE(_r, _idx) \ |
202 | (&((_r)->sring->ring[((_idx) & (RING_SIZE(_r) - 1))].rsp)) |
203 | |
204 | /* Loop termination condition: Would the specified index overflow the ring? */ |
205 | #define RING_REQUEST_CONS_OVERFLOW(_r, _cons) \ |
206 | (((_cons) - (_r)->rsp_prod_pvt) >= RING_SIZE(_r)) |
207 | |
208 | /* Ill-behaved frontend determination: Can there be this many requests? */ |
209 | #define RING_REQUEST_PROD_OVERFLOW(_r, _prod) \ |
210 | (((_prod) - (_r)->rsp_prod_pvt) > RING_SIZE(_r)) |
211 | |
212 | |
213 | #define RING_PUSH_REQUESTS(_r) do { \ |
214 | virt_wmb(); /* back sees requests /before/ updated producer index */ \ |
215 | (_r)->sring->req_prod = (_r)->req_prod_pvt; \ |
216 | } while (0) |
217 | |
218 | #define RING_PUSH_RESPONSES(_r) do { \ |
219 | virt_wmb(); /* front sees responses /before/ updated producer index */ \ |
220 | (_r)->sring->rsp_prod = (_r)->rsp_prod_pvt; \ |
221 | } while (0) |
222 | |
223 | /* |
224 | * Notification hold-off (req_event and rsp_event): |
225 | * |
226 | * When queueing requests or responses on a shared ring, it may not always be |
227 | * necessary to notify the remote end. For example, if requests are in flight |
228 | * in a backend, the front may be able to queue further requests without |
229 | * notifying the back (if the back checks for new requests when it queues |
230 | * responses). |
231 | * |
232 | * When enqueuing requests or responses: |
233 | * |
234 | * Use RING_PUSH_{REQUESTS,RESPONSES}_AND_CHECK_NOTIFY(). The second argument |
235 | * is a boolean return value. True indicates that the receiver requires an |
236 | * asynchronous notification. |
237 | * |
238 | * After dequeuing requests or responses (before sleeping the connection): |
239 | * |
240 | * Use RING_FINAL_CHECK_FOR_REQUESTS() or RING_FINAL_CHECK_FOR_RESPONSES(). |
241 | * The second argument is a boolean return value. True indicates that there |
242 | * are pending messages on the ring (i.e., the connection should not be put |
243 | * to sleep). |
244 | * |
245 | * These macros will set the req_event/rsp_event field to trigger a |
246 | * notification on the very next message that is enqueued. If you want to |
247 | * create batches of work (i.e., only receive a notification after several |
248 | * messages have been enqueued) then you will need to create a customised |
249 | * version of the FINAL_CHECK macro in your own code, which sets the event |
250 | * field appropriately. |
251 | */ |
252 | |
253 | #define RING_PUSH_REQUESTS_AND_CHECK_NOTIFY(_r, _notify) do { \ |
254 | RING_IDX __old = (_r)->sring->req_prod; \ |
255 | RING_IDX __new = (_r)->req_prod_pvt; \ |
256 | virt_wmb(); /* back sees requests /before/ updated producer index */ \ |
257 | (_r)->sring->req_prod = __new; \ |
258 | virt_mb(); /* back sees new requests /before/ we check req_event */ \ |
259 | (_notify) = ((RING_IDX)(__new - (_r)->sring->req_event) < \ |
260 | (RING_IDX)(__new - __old)); \ |
261 | } while (0) |
262 | |
263 | #define RING_PUSH_RESPONSES_AND_CHECK_NOTIFY(_r, _notify) do { \ |
264 | RING_IDX __old = (_r)->sring->rsp_prod; \ |
265 | RING_IDX __new = (_r)->rsp_prod_pvt; \ |
266 | virt_wmb(); /* front sees responses /before/ updated producer index */ \ |
267 | (_r)->sring->rsp_prod = __new; \ |
268 | virt_mb(); /* front sees new responses /before/ we check rsp_event */ \ |
269 | (_notify) = ((RING_IDX)(__new - (_r)->sring->rsp_event) < \ |
270 | (RING_IDX)(__new - __old)); \ |
271 | } while (0) |
272 | |
273 | #define RING_FINAL_CHECK_FOR_REQUESTS(_r, _work_to_do) do { \ |
274 | (_work_to_do) = RING_HAS_UNCONSUMED_REQUESTS(_r); \ |
275 | if (_work_to_do) break; \ |
276 | (_r)->sring->req_event = (_r)->req_cons + 1; \ |
277 | virt_mb(); \ |
278 | (_work_to_do) = RING_HAS_UNCONSUMED_REQUESTS(_r); \ |
279 | } while (0) |
280 | |
281 | #define RING_FINAL_CHECK_FOR_RESPONSES(_r, _work_to_do) do { \ |
282 | (_work_to_do) = RING_HAS_UNCONSUMED_RESPONSES(_r); \ |
283 | if (_work_to_do) break; \ |
284 | (_r)->sring->rsp_event = (_r)->rsp_cons + 1; \ |
285 | virt_mb(); \ |
286 | (_work_to_do) = RING_HAS_UNCONSUMED_RESPONSES(_r); \ |
287 | } while (0) |
288 | |
289 | |
290 | /* |
291 | * DEFINE_XEN_FLEX_RING_AND_INTF defines two monodirectional rings and |
292 | * functions to check if there is data on the ring, and to read and |
293 | * write to them. |
294 | * |
295 | * DEFINE_XEN_FLEX_RING is similar to DEFINE_XEN_FLEX_RING_AND_INTF, but |
296 | * does not define the indexes page. As different protocols can have |
297 | * extensions to the basic format, this macro allow them to define their |
298 | * own struct. |
299 | * |
300 | * XEN_FLEX_RING_SIZE |
301 | * Convenience macro to calculate the size of one of the two rings |
302 | * from the overall order. |
303 | * |
304 | * $NAME_mask |
305 | * Function to apply the size mask to an index, to reduce the index |
306 | * within the range [0-size]. |
307 | * |
308 | * $NAME_read_packet |
309 | * Function to read data from the ring. The amount of data to read is |
310 | * specified by the "size" argument. |
311 | * |
312 | * $NAME_write_packet |
313 | * Function to write data to the ring. The amount of data to write is |
314 | * specified by the "size" argument. |
315 | * |
316 | * $NAME_get_ring_ptr |
317 | * Convenience function that returns a pointer to read/write to the |
318 | * ring at the right location. |
319 | * |
320 | * $NAME_data_intf |
321 | * Indexes page, shared between frontend and backend. It also |
322 | * contains the array of grant refs. |
323 | * |
324 | * $NAME_queued |
325 | * Function to calculate how many bytes are currently on the ring, |
326 | * ready to be read. It can also be used to calculate how much free |
327 | * space is currently on the ring (XEN_FLEX_RING_SIZE() - |
328 | * $NAME_queued()). |
329 | */ |
330 | |
331 | #ifndef XEN_PAGE_SHIFT |
332 | /* The PAGE_SIZE for ring protocols and hypercall interfaces is always |
333 | * 4K, regardless of the architecture, and page granularity chosen by |
334 | * operating systems. |
335 | */ |
336 | #define XEN_PAGE_SHIFT 12 |
337 | #endif |
338 | #define XEN_FLEX_RING_SIZE(order) \ |
339 | (1UL << ((order) + XEN_PAGE_SHIFT - 1)) |
340 | |
341 | #define DEFINE_XEN_FLEX_RING(name) \ |
342 | static inline RING_IDX name##_mask(RING_IDX idx, RING_IDX ring_size) \ |
343 | { \ |
344 | return idx & (ring_size - 1); \ |
345 | } \ |
346 | \ |
347 | static inline unsigned char *name##_get_ring_ptr(unsigned char *buf, \ |
348 | RING_IDX idx, \ |
349 | RING_IDX ring_size) \ |
350 | { \ |
351 | return buf + name##_mask(idx, ring_size); \ |
352 | } \ |
353 | \ |
354 | static inline void name##_read_packet(void *opaque, \ |
355 | const unsigned char *buf, \ |
356 | size_t size, \ |
357 | RING_IDX masked_prod, \ |
358 | RING_IDX *masked_cons, \ |
359 | RING_IDX ring_size) \ |
360 | { \ |
361 | if (*masked_cons < masked_prod || \ |
362 | size <= ring_size - *masked_cons) { \ |
363 | memcpy(opaque, buf + *masked_cons, size); \ |
364 | } else { \ |
365 | memcpy(opaque, buf + *masked_cons, ring_size - *masked_cons); \ |
366 | memcpy((unsigned char *)opaque + ring_size - *masked_cons, buf, \ |
367 | size - (ring_size - *masked_cons)); \ |
368 | } \ |
369 | *masked_cons = name##_mask(*masked_cons + size, ring_size); \ |
370 | } \ |
371 | \ |
372 | static inline void name##_write_packet(unsigned char *buf, \ |
373 | const void *opaque, \ |
374 | size_t size, \ |
375 | RING_IDX *masked_prod, \ |
376 | RING_IDX masked_cons, \ |
377 | RING_IDX ring_size) \ |
378 | { \ |
379 | if (*masked_prod < masked_cons || \ |
380 | size <= ring_size - *masked_prod) { \ |
381 | memcpy(buf + *masked_prod, opaque, size); \ |
382 | } else { \ |
383 | memcpy(buf + *masked_prod, opaque, ring_size - *masked_prod); \ |
384 | memcpy(buf, (unsigned char *)opaque + (ring_size - *masked_prod), \ |
385 | size - (ring_size - *masked_prod)); \ |
386 | } \ |
387 | *masked_prod = name##_mask(*masked_prod + size, ring_size); \ |
388 | } \ |
389 | \ |
390 | static inline RING_IDX name##_queued(RING_IDX prod, \ |
391 | RING_IDX cons, \ |
392 | RING_IDX ring_size) \ |
393 | { \ |
394 | RING_IDX size; \ |
395 | \ |
396 | if (prod == cons) \ |
397 | return 0; \ |
398 | \ |
399 | prod = name##_mask(prod, ring_size); \ |
400 | cons = name##_mask(cons, ring_size); \ |
401 | \ |
402 | if (prod == cons) \ |
403 | return ring_size; \ |
404 | \ |
405 | if (prod > cons) \ |
406 | size = prod - cons; \ |
407 | else \ |
408 | size = ring_size - (cons - prod); \ |
409 | return size; \ |
410 | } \ |
411 | \ |
412 | struct name##_data { \ |
413 | unsigned char *in; /* half of the allocation */ \ |
414 | unsigned char *out; /* half of the allocation */ \ |
415 | } |
416 | |
417 | #define DEFINE_XEN_FLEX_RING_AND_INTF(name) \ |
418 | struct name##_data_intf { \ |
419 | RING_IDX in_cons, in_prod; \ |
420 | \ |
421 | uint8_t pad1[56]; \ |
422 | \ |
423 | RING_IDX out_cons, out_prod; \ |
424 | \ |
425 | uint8_t pad2[56]; \ |
426 | \ |
427 | RING_IDX ring_order; \ |
428 | grant_ref_t ref[]; \ |
429 | }; \ |
430 | DEFINE_XEN_FLEX_RING(name) |
431 | |
432 | #endif /* __XEN_PUBLIC_IO_RING_H__ */ |
433 | |