xref: /linux/include/xen/interface/io/ring.h (revision e58e871becec2d3b04ed91c0c16fe8deac9c9dfa)
1 /******************************************************************************
2  * ring.h
3  *
4  * Shared producer-consumer ring macros.
5  *
6  * Tim Deegan and Andrew Warfield November 2004.
7  */
8 
9 #ifndef __XEN_PUBLIC_IO_RING_H__
10 #define __XEN_PUBLIC_IO_RING_H__
11 
12 typedef unsigned int RING_IDX;
13 
14 /* Round a 32-bit unsigned constant down to the nearest power of two. */
15 #define __RD2(_x)  (((_x) & 0x00000002) ? 0x2		       : ((_x) & 0x1))
16 #define __RD4(_x)  (((_x) & 0x0000000c) ? __RD2((_x)>>2)<<2    : __RD2(_x))
17 #define __RD8(_x)  (((_x) & 0x000000f0) ? __RD4((_x)>>4)<<4    : __RD4(_x))
18 #define __RD16(_x) (((_x) & 0x0000ff00) ? __RD8((_x)>>8)<<8    : __RD8(_x))
19 #define __RD32(_x) (((_x) & 0xffff0000) ? __RD16((_x)>>16)<<16 : __RD16(_x))
20 
21 /*
22  * Calculate size of a shared ring, given the total available space for the
23  * ring and indexes (_sz), and the name tag of the request/response structure.
24  * A ring contains as many entries as will fit, rounded down to the nearest
25  * power of two (so we can mask with (size-1) to loop around).
26  */
27 #define __CONST_RING_SIZE(_s, _sz)				\
28 	(__RD32(((_sz) - offsetof(struct _s##_sring, ring)) /	\
29 		sizeof(((struct _s##_sring *)0)->ring[0])))
30 
31 /*
32  * The same for passing in an actual pointer instead of a name tag.
33  */
34 #define __RING_SIZE(_s, _sz)						\
35 	(__RD32(((_sz) - (long)&(_s)->ring + (long)(_s)) / sizeof((_s)->ring[0])))
36 
37 /*
38  * Macros to make the correct C datatypes for a new kind of ring.
39  *
40  * To make a new ring datatype, you need to have two message structures,
41  * let's say struct request, and struct response already defined.
42  *
43  * In a header where you want the ring datatype declared, you then do:
44  *
45  *     DEFINE_RING_TYPES(mytag, struct request, struct response);
46  *
47  * These expand out to give you a set of types, as you can see below.
48  * The most important of these are:
49  *
50  *     struct mytag_sring      - The shared ring.
51  *     struct mytag_front_ring - The 'front' half of the ring.
52  *     struct mytag_back_ring  - The 'back' half of the ring.
53  *
54  * To initialize a ring in your code you need to know the location and size
55  * of the shared memory area (PAGE_SIZE, for instance). To initialise
56  * the front half:
57  *
58  *     struct mytag_front_ring front_ring;
59  *     SHARED_RING_INIT((struct mytag_sring *)shared_page);
60  *     FRONT_RING_INIT(&front_ring, (struct mytag_sring *)shared_page,
61  *		       PAGE_SIZE);
62  *
63  * Initializing the back follows similarly (note that only the front
64  * initializes the shared ring):
65  *
66  *     struct mytag_back_ring back_ring;
67  *     BACK_RING_INIT(&back_ring, (struct mytag_sring *)shared_page,
68  *		      PAGE_SIZE);
69  */
70 
71 #define DEFINE_RING_TYPES(__name, __req_t, __rsp_t)			\
72 									\
73 /* Shared ring entry */							\
74 union __name##_sring_entry {						\
75     __req_t req;							\
76     __rsp_t rsp;							\
77 };									\
78 									\
79 /* Shared ring page */							\
80 struct __name##_sring {							\
81     RING_IDX req_prod, req_event;					\
82     RING_IDX rsp_prod, rsp_event;					\
83     uint8_t  pad[48];							\
84     union __name##_sring_entry ring[1]; /* variable-length */		\
85 };									\
86 									\
87 /* "Front" end's private variables */					\
88 struct __name##_front_ring {						\
89     RING_IDX req_prod_pvt;						\
90     RING_IDX rsp_cons;							\
91     unsigned int nr_ents;						\
92     struct __name##_sring *sring;					\
93 };									\
94 									\
95 /* "Back" end's private variables */					\
96 struct __name##_back_ring {						\
97     RING_IDX rsp_prod_pvt;						\
98     RING_IDX req_cons;							\
99     unsigned int nr_ents;						\
100     struct __name##_sring *sring;					\
101 };
102 
103 /*
104  * Macros for manipulating rings.
105  *
106  * FRONT_RING_whatever works on the "front end" of a ring: here
107  * requests are pushed on to the ring and responses taken off it.
108  *
109  * BACK_RING_whatever works on the "back end" of a ring: here
110  * requests are taken off the ring and responses put on.
111  *
112  * N.B. these macros do NO INTERLOCKS OR FLOW CONTROL.
113  * This is OK in 1-for-1 request-response situations where the
114  * requestor (front end) never has more than RING_SIZE()-1
115  * outstanding requests.
116  */
117 
118 /* Initialising empty rings */
119 #define SHARED_RING_INIT(_s) do {					\
120     (_s)->req_prod  = (_s)->rsp_prod  = 0;				\
121     (_s)->req_event = (_s)->rsp_event = 1;				\
122     memset((_s)->pad, 0, sizeof((_s)->pad));				\
123 } while(0)
124 
125 #define FRONT_RING_INIT(_r, _s, __size) do {				\
126     (_r)->req_prod_pvt = 0;						\
127     (_r)->rsp_cons = 0;							\
128     (_r)->nr_ents = __RING_SIZE(_s, __size);				\
129     (_r)->sring = (_s);							\
130 } while (0)
131 
132 #define BACK_RING_INIT(_r, _s, __size) do {				\
133     (_r)->rsp_prod_pvt = 0;						\
134     (_r)->req_cons = 0;							\
135     (_r)->nr_ents = __RING_SIZE(_s, __size);				\
136     (_r)->sring = (_s);							\
137 } while (0)
138 
139 /* Initialize to existing shared indexes -- for recovery */
140 #define FRONT_RING_ATTACH(_r, _s, __size) do {				\
141     (_r)->sring = (_s);							\
142     (_r)->req_prod_pvt = (_s)->req_prod;				\
143     (_r)->rsp_cons = (_s)->rsp_prod;					\
144     (_r)->nr_ents = __RING_SIZE(_s, __size);				\
145 } while (0)
146 
147 #define BACK_RING_ATTACH(_r, _s, __size) do {				\
148     (_r)->sring = (_s);							\
149     (_r)->rsp_prod_pvt = (_s)->rsp_prod;				\
150     (_r)->req_cons = (_s)->req_prod;					\
151     (_r)->nr_ents = __RING_SIZE(_s, __size);				\
152 } while (0)
153 
154 /* How big is this ring? */
155 #define RING_SIZE(_r)							\
156     ((_r)->nr_ents)
157 
158 /* Number of free requests (for use on front side only). */
159 #define RING_FREE_REQUESTS(_r)						\
160     (RING_SIZE(_r) - ((_r)->req_prod_pvt - (_r)->rsp_cons))
161 
162 /* Test if there is an empty slot available on the front ring.
163  * (This is only meaningful from the front. )
164  */
165 #define RING_FULL(_r)							\
166     (RING_FREE_REQUESTS(_r) == 0)
167 
168 /* Test if there are outstanding messages to be processed on a ring. */
169 #define RING_HAS_UNCONSUMED_RESPONSES(_r)				\
170     ((_r)->sring->rsp_prod - (_r)->rsp_cons)
171 
172 #define RING_HAS_UNCONSUMED_REQUESTS(_r)				\
173     ({									\
174 	unsigned int req = (_r)->sring->req_prod - (_r)->req_cons;	\
175 	unsigned int rsp = RING_SIZE(_r) -				\
176 			   ((_r)->req_cons - (_r)->rsp_prod_pvt);	\
177 	req < rsp ? req : rsp;						\
178     })
179 
180 /* Direct access to individual ring elements, by index. */
181 #define RING_GET_REQUEST(_r, _idx)					\
182     (&((_r)->sring->ring[((_idx) & (RING_SIZE(_r) - 1))].req))
183 
184 /*
185  * Get a local copy of a request.
186  *
187  * Use this in preference to RING_GET_REQUEST() so all processing is
188  * done on a local copy that cannot be modified by the other end.
189  *
190  * Note that https://gcc.gnu.org/bugzilla/show_bug.cgi?id=58145 may cause this
191  * to be ineffective where _req is a struct which consists of only bitfields.
192  */
193 #define RING_COPY_REQUEST(_r, _idx, _req) do {				\
194 	/* Use volatile to force the copy into _req. */			\
195 	*(_req) = *(volatile typeof(_req))RING_GET_REQUEST(_r, _idx);	\
196 } while (0)
197 
198 #define RING_GET_RESPONSE(_r, _idx)					\
199     (&((_r)->sring->ring[((_idx) & (RING_SIZE(_r) - 1))].rsp))
200 
201 /* Loop termination condition: Would the specified index overflow the ring? */
202 #define RING_REQUEST_CONS_OVERFLOW(_r, _cons)				\
203     (((_cons) - (_r)->rsp_prod_pvt) >= RING_SIZE(_r))
204 
205 /* Ill-behaved frontend determination: Can there be this many requests? */
206 #define RING_REQUEST_PROD_OVERFLOW(_r, _prod)               \
207     (((_prod) - (_r)->rsp_prod_pvt) > RING_SIZE(_r))
208 
209 
210 #define RING_PUSH_REQUESTS(_r) do {					\
211     virt_wmb(); /* back sees requests /before/ updated producer index */	\
212     (_r)->sring->req_prod = (_r)->req_prod_pvt;				\
213 } while (0)
214 
215 #define RING_PUSH_RESPONSES(_r) do {					\
216     virt_wmb(); /* front sees responses /before/ updated producer index */	\
217     (_r)->sring->rsp_prod = (_r)->rsp_prod_pvt;				\
218 } while (0)
219 
220 /*
221  * Notification hold-off (req_event and rsp_event):
222  *
223  * When queueing requests or responses on a shared ring, it may not always be
224  * necessary to notify the remote end. For example, if requests are in flight
225  * in a backend, the front may be able to queue further requests without
226  * notifying the back (if the back checks for new requests when it queues
227  * responses).
228  *
229  * When enqueuing requests or responses:
230  *
231  *  Use RING_PUSH_{REQUESTS,RESPONSES}_AND_CHECK_NOTIFY(). The second argument
232  *  is a boolean return value. True indicates that the receiver requires an
233  *  asynchronous notification.
234  *
235  * After dequeuing requests or responses (before sleeping the connection):
236  *
237  *  Use RING_FINAL_CHECK_FOR_REQUESTS() or RING_FINAL_CHECK_FOR_RESPONSES().
238  *  The second argument is a boolean return value. True indicates that there
239  *  are pending messages on the ring (i.e., the connection should not be put
240  *  to sleep).
241  *
242  *  These macros will set the req_event/rsp_event field to trigger a
243  *  notification on the very next message that is enqueued. If you want to
244  *  create batches of work (i.e., only receive a notification after several
245  *  messages have been enqueued) then you will need to create a customised
246  *  version of the FINAL_CHECK macro in your own code, which sets the event
247  *  field appropriately.
248  */
249 
250 #define RING_PUSH_REQUESTS_AND_CHECK_NOTIFY(_r, _notify) do {		\
251     RING_IDX __old = (_r)->sring->req_prod;				\
252     RING_IDX __new = (_r)->req_prod_pvt;				\
253     virt_wmb(); /* back sees requests /before/ updated producer index */	\
254     (_r)->sring->req_prod = __new;					\
255     virt_mb(); /* back sees new requests /before/ we check req_event */	\
256     (_notify) = ((RING_IDX)(__new - (_r)->sring->req_event) <		\
257 		 (RING_IDX)(__new - __old));				\
258 } while (0)
259 
260 #define RING_PUSH_RESPONSES_AND_CHECK_NOTIFY(_r, _notify) do {		\
261     RING_IDX __old = (_r)->sring->rsp_prod;				\
262     RING_IDX __new = (_r)->rsp_prod_pvt;				\
263     virt_wmb(); /* front sees responses /before/ updated producer index */	\
264     (_r)->sring->rsp_prod = __new;					\
265     virt_mb(); /* front sees new responses /before/ we check rsp_event */	\
266     (_notify) = ((RING_IDX)(__new - (_r)->sring->rsp_event) <		\
267 		 (RING_IDX)(__new - __old));				\
268 } while (0)
269 
270 #define RING_FINAL_CHECK_FOR_REQUESTS(_r, _work_to_do) do {		\
271     (_work_to_do) = RING_HAS_UNCONSUMED_REQUESTS(_r);			\
272     if (_work_to_do) break;						\
273     (_r)->sring->req_event = (_r)->req_cons + 1;			\
274     virt_mb();								\
275     (_work_to_do) = RING_HAS_UNCONSUMED_REQUESTS(_r);			\
276 } while (0)
277 
278 #define RING_FINAL_CHECK_FOR_RESPONSES(_r, _work_to_do) do {		\
279     (_work_to_do) = RING_HAS_UNCONSUMED_RESPONSES(_r);			\
280     if (_work_to_do) break;						\
281     (_r)->sring->rsp_event = (_r)->rsp_cons + 1;			\
282     virt_mb();								\
283     (_work_to_do) = RING_HAS_UNCONSUMED_RESPONSES(_r);			\
284 } while (0)
285 
286 
287 /*
288  * DEFINE_XEN_FLEX_RING_AND_INTF defines two monodirectional rings and
289  * functions to check if there is data on the ring, and to read and
290  * write to them.
291  *
292  * DEFINE_XEN_FLEX_RING is similar to DEFINE_XEN_FLEX_RING_AND_INTF, but
293  * does not define the indexes page. As different protocols can have
294  * extensions to the basic format, this macro allow them to define their
295  * own struct.
296  *
297  * XEN_FLEX_RING_SIZE
298  *   Convenience macro to calculate the size of one of the two rings
299  *   from the overall order.
300  *
301  * $NAME_mask
302  *   Function to apply the size mask to an index, to reduce the index
303  *   within the range [0-size].
304  *
305  * $NAME_read_packet
306  *   Function to read data from the ring. The amount of data to read is
307  *   specified by the "size" argument.
308  *
309  * $NAME_write_packet
310  *   Function to write data to the ring. The amount of data to write is
311  *   specified by the "size" argument.
312  *
313  * $NAME_get_ring_ptr
314  *   Convenience function that returns a pointer to read/write to the
315  *   ring at the right location.
316  *
317  * $NAME_data_intf
318  *   Indexes page, shared between frontend and backend. It also
319  *   contains the array of grant refs.
320  *
321  * $NAME_queued
322  *   Function to calculate how many bytes are currently on the ring,
323  *   ready to be read. It can also be used to calculate how much free
324  *   space is currently on the ring (XEN_FLEX_RING_SIZE() -
325  *   $NAME_queued()).
326  */
327 
328 #ifndef XEN_PAGE_SHIFT
329 /* The PAGE_SIZE for ring protocols and hypercall interfaces is always
330  * 4K, regardless of the architecture, and page granularity chosen by
331  * operating systems.
332  */
333 #define XEN_PAGE_SHIFT 12
334 #endif
335 #define XEN_FLEX_RING_SIZE(order)                                             \
336     (1UL << ((order) + XEN_PAGE_SHIFT - 1))
337 
338 #define DEFINE_XEN_FLEX_RING(name)                                            \
339 static inline RING_IDX name##_mask(RING_IDX idx, RING_IDX ring_size)          \
340 {                                                                             \
341     return idx & (ring_size - 1);                                             \
342 }                                                                             \
343                                                                               \
344 static inline unsigned char *name##_get_ring_ptr(unsigned char *buf,          \
345                                                  RING_IDX idx,                \
346                                                  RING_IDX ring_size)          \
347 {                                                                             \
348     return buf + name##_mask(idx, ring_size);                                 \
349 }                                                                             \
350                                                                               \
351 static inline void name##_read_packet(void *opaque,                           \
352                                       const unsigned char *buf,               \
353                                       size_t size,                            \
354                                       RING_IDX masked_prod,                   \
355                                       RING_IDX *masked_cons,                  \
356                                       RING_IDX ring_size)                     \
357 {                                                                             \
358     if (*masked_cons < masked_prod ||                                         \
359         size <= ring_size - *masked_cons) {                                   \
360         memcpy(opaque, buf + *masked_cons, size);                             \
361     } else {                                                                  \
362         memcpy(opaque, buf + *masked_cons, ring_size - *masked_cons);         \
363         memcpy((unsigned char *)opaque + ring_size - *masked_cons, buf,       \
364                size - (ring_size - *masked_cons));                            \
365     }                                                                         \
366     *masked_cons = name##_mask(*masked_cons + size, ring_size);               \
367 }                                                                             \
368                                                                               \
369 static inline void name##_write_packet(unsigned char *buf,                    \
370                                        const void *opaque,                    \
371                                        size_t size,                           \
372                                        RING_IDX *masked_prod,                 \
373                                        RING_IDX masked_cons,                  \
374                                        RING_IDX ring_size)                    \
375 {                                                                             \
376     if (*masked_prod < masked_cons ||                                         \
377         size <= ring_size - *masked_prod) {                                   \
378         memcpy(buf + *masked_prod, opaque, size);                             \
379     } else {                                                                  \
380         memcpy(buf + *masked_prod, opaque, ring_size - *masked_prod);         \
381         memcpy(buf, (unsigned char *)opaque + (ring_size - *masked_prod),     \
382                size - (ring_size - *masked_prod));                            \
383     }                                                                         \
384     *masked_prod = name##_mask(*masked_prod + size, ring_size);               \
385 }                                                                             \
386                                                                               \
387 static inline RING_IDX name##_queued(RING_IDX prod,                           \
388                                      RING_IDX cons,                           \
389                                      RING_IDX ring_size)                      \
390 {                                                                             \
391     RING_IDX size;                                                            \
392                                                                               \
393     if (prod == cons)                                                         \
394         return 0;                                                             \
395                                                                               \
396     prod = name##_mask(prod, ring_size);                                      \
397     cons = name##_mask(cons, ring_size);                                      \
398                                                                               \
399     if (prod == cons)                                                         \
400         return ring_size;                                                     \
401                                                                               \
402     if (prod > cons)                                                          \
403         size = prod - cons;                                                   \
404     else                                                                      \
405         size = ring_size - (cons - prod);                                     \
406     return size;                                                              \
407 }                                                                             \
408                                                                               \
409 struct name##_data {                                                          \
410     unsigned char *in; /* half of the allocation */                           \
411     unsigned char *out; /* half of the allocation */                          \
412 }
413 
414 #define DEFINE_XEN_FLEX_RING_AND_INTF(name)                                   \
415 struct name##_data_intf {                                                     \
416     RING_IDX in_cons, in_prod;                                                \
417                                                                               \
418     uint8_t pad1[56];                                                         \
419                                                                               \
420     RING_IDX out_cons, out_prod;                                              \
421                                                                               \
422     uint8_t pad2[56];                                                         \
423                                                                               \
424     RING_IDX ring_order;                                                      \
425     grant_ref_t ref[];                                                        \
426 };                                                                            \
427 DEFINE_XEN_FLEX_RING(name)
428 
429 #endif /* __XEN_PUBLIC_IO_RING_H__ */
430