xref: /qemu/include/hw/xen/interface/io/ring.h (revision 138ca49a)
1 /******************************************************************************
2  * ring.h
3  *
4  * Shared producer-consumer ring macros.
5  *
6  * Permission is hereby granted, free of charge, to any person obtaining a copy
7  * of this software and associated documentation files (the "Software"), to
8  * deal in the Software without restriction, including without limitation the
9  * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
10  * sell copies of the Software, and to permit persons to whom the Software is
11  * furnished to do so, subject to the following conditions:
12  *
13  * The above copyright notice and this permission notice shall be included in
14  * all copies or substantial portions of the Software.
15  *
16  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
19  * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
20  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
21  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
22  * DEALINGS IN THE SOFTWARE.
23  *
24  * Tim Deegan and Andrew Warfield November 2004.
25  */
26 
27 #ifndef __XEN_PUBLIC_IO_RING_H__
28 #define __XEN_PUBLIC_IO_RING_H__
29 
30 /*
31  * When #include'ing this header, you need to provide the following
32  * declaration upfront:
33  * - standard integers types (uint8_t, uint16_t, etc)
34  * They are provided by stdint.h of the standard headers.
35  *
36  * Before using the different macros, you need to provide the following
37  * macros:
38  * - xen_mb()  a memory barrier
39  * - xen_rmb() a read memory barrier
40  * - xen_wmb() a write memory barrier
41  * Example of those can be found in xenctrl.h.
42  *
43  * In addition, if you intend to use the FLEX macros, you also need to
44  * provide the following, before invoking the FLEX macros:
45  * - size_t
46  * - memcpy
47  * - grant_ref_t
48  * These declarations are provided by string.h of the standard headers,
49  * and grant_table.h from the Xen public headers.
50  */
51 
52 typedef unsigned int RING_IDX;
53 
54 /* Round a 32-bit unsigned constant down to the nearest power of two. */
55 #define __RD2(_x)  (((_x) & 0x00000002) ? 0x2                  : ((_x) & 0x1))
56 #define __RD4(_x)  (((_x) & 0x0000000c) ? __RD2((_x)>>2)<<2    : __RD2(_x))
57 #define __RD8(_x)  (((_x) & 0x000000f0) ? __RD4((_x)>>4)<<4    : __RD4(_x))
58 #define __RD16(_x) (((_x) & 0x0000ff00) ? __RD8((_x)>>8)<<8    : __RD8(_x))
59 #define __RD32(_x) (((_x) & 0xffff0000) ? __RD16((_x)>>16)<<16 : __RD16(_x))
60 
61 /*
62  * Calculate size of a shared ring, given the total available space for the
63  * ring and indexes (_sz), and the name tag of the request/response structure.
64  * A ring contains as many entries as will fit, rounded down to the nearest
65  * power of two (so we can mask with (size-1) to loop around).
66  */
67 #define __CONST_RING_SIZE(_s, _sz) \
68     (__RD32(((_sz) - offsetof(struct _s##_sring, ring)) / \
69             sizeof_field(struct _s##_sring, ring[0])))
70 /*
71  * The same for passing in an actual pointer instead of a name tag.
72  */
73 #define __RING_SIZE(_s, _sz) \
74     (__RD32(((_sz) - (long)(_s)->ring + (long)(_s)) / sizeof((_s)->ring[0])))
75 
76 /*
77  * Macros to make the correct C datatypes for a new kind of ring.
78  *
79  * To make a new ring datatype, you need to have two message structures,
80  * let's say request_t, and response_t already defined.
81  *
82  * In a header where you want the ring datatype declared, you then do:
83  *
84  *     DEFINE_RING_TYPES(mytag, request_t, response_t);
85  *
86  * These expand out to give you a set of types, as you can see below.
87  * The most important of these are:
88  *
89  *     mytag_sring_t      - The shared ring.
90  *     mytag_front_ring_t - The 'front' half of the ring.
91  *     mytag_back_ring_t  - The 'back' half of the ring.
92  *
93  * To initialize a ring in your code you need to know the location and size
94  * of the shared memory area (PAGE_SIZE, for instance). To initialise
95  * the front half:
96  *
97  *     mytag_front_ring_t front_ring;
98  *     SHARED_RING_INIT((mytag_sring_t *)shared_page);
99  *     FRONT_RING_INIT(&front_ring, (mytag_sring_t *)shared_page, PAGE_SIZE);
100  *
101  * Initializing the back follows similarly (note that only the front
102  * initializes the shared ring):
103  *
104  *     mytag_back_ring_t back_ring;
105  *     BACK_RING_INIT(&back_ring, (mytag_sring_t *)shared_page, PAGE_SIZE);
106  */
107 
108 #define DEFINE_RING_TYPES(__name, __req_t, __rsp_t)                     \
109                                                                         \
110 /* Shared ring entry */                                                 \
111 union __name##_sring_entry {                                            \
112     __req_t req;                                                        \
113     __rsp_t rsp;                                                        \
114 };                                                                      \
115                                                                         \
116 /* Shared ring page */                                                  \
117 struct __name##_sring {                                                 \
118     RING_IDX req_prod, req_event;                                       \
119     RING_IDX rsp_prod, rsp_event;                                       \
120     union {                                                             \
121         struct {                                                        \
122             uint8_t smartpoll_active;                                   \
123         } netif;                                                        \
124         struct {                                                        \
125             uint8_t msg;                                                \
126         } tapif_user;                                                   \
127         uint8_t pvt_pad[4];                                             \
128     } pvt;                                                              \
129     uint8_t __pad[44];                                                  \
130     union __name##_sring_entry ring[1]; /* variable-length */           \
131 };                                                                      \
132                                                                         \
133 /* "Front" end's private variables */                                   \
134 struct __name##_front_ring {                                            \
135     RING_IDX req_prod_pvt;                                              \
136     RING_IDX rsp_cons;                                                  \
137     unsigned int nr_ents;                                               \
138     struct __name##_sring *sring;                                       \
139 };                                                                      \
140                                                                         \
141 /* "Back" end's private variables */                                    \
142 struct __name##_back_ring {                                             \
143     RING_IDX rsp_prod_pvt;                                              \
144     RING_IDX req_cons;                                                  \
145     unsigned int nr_ents;                                               \
146     struct __name##_sring *sring;                                       \
147 };                                                                      \
148                                                                         \
149 /* Syntactic sugar */                                                   \
150 typedef struct __name##_sring __name##_sring_t;                         \
151 typedef struct __name##_front_ring __name##_front_ring_t;               \
152 typedef struct __name##_back_ring __name##_back_ring_t
153 
154 /*
155  * Macros for manipulating rings.
156  *
157  * FRONT_RING_whatever works on the "front end" of a ring: here
158  * requests are pushed on to the ring and responses taken off it.
159  *
160  * BACK_RING_whatever works on the "back end" of a ring: here
161  * requests are taken off the ring and responses put on.
162  *
163  * N.B. these macros do NO INTERLOCKS OR FLOW CONTROL.
164  * This is OK in 1-for-1 request-response situations where the
165  * requestor (front end) never has more than RING_SIZE()-1
166  * outstanding requests.
167  */
168 
169 /* Initialising empty rings */
170 #define SHARED_RING_INIT(_s) do {                                       \
171     (_s)->req_prod  = (_s)->rsp_prod  = 0;                              \
172     (_s)->req_event = (_s)->rsp_event = 1;                              \
173     (void)memset((_s)->pvt.pvt_pad, 0, sizeof((_s)->pvt.pvt_pad));      \
174     (void)memset((_s)->__pad, 0, sizeof((_s)->__pad));                  \
175 } while(0)
176 
177 #define FRONT_RING_INIT(_r, _s, __size) do {                            \
178     (_r)->req_prod_pvt = 0;                                             \
179     (_r)->rsp_cons = 0;                                                 \
180     (_r)->nr_ents = __RING_SIZE(_s, __size);                            \
181     (_r)->sring = (_s);                                                 \
182 } while (0)
183 
184 #define BACK_RING_INIT(_r, _s, __size) do {                             \
185     (_r)->rsp_prod_pvt = 0;                                             \
186     (_r)->req_cons = 0;                                                 \
187     (_r)->nr_ents = __RING_SIZE(_s, __size);                            \
188     (_r)->sring = (_s);                                                 \
189 } while (0)
190 
191 /* How big is this ring? */
192 #define RING_SIZE(_r)                                                   \
193     ((_r)->nr_ents)
194 
195 /* Number of free requests (for use on front side only). */
196 #define RING_FREE_REQUESTS(_r)                                          \
197     (RING_SIZE(_r) - ((_r)->req_prod_pvt - (_r)->rsp_cons))
198 
199 /* Test if there is an empty slot available on the front ring.
200  * (This is only meaningful from the front. )
201  */
202 #define RING_FULL(_r)                                                   \
203     (RING_FREE_REQUESTS(_r) == 0)
204 
205 /* Test if there are outstanding messages to be processed on a ring. */
206 #define RING_HAS_UNCONSUMED_RESPONSES(_r)                               \
207     ((_r)->sring->rsp_prod - (_r)->rsp_cons)
208 
209 #define RING_HAS_UNCONSUMED_REQUESTS(_r) ({                             \
210     unsigned int req = (_r)->sring->req_prod - (_r)->req_cons;          \
211     unsigned int rsp = RING_SIZE(_r) -                                  \
212         ((_r)->req_cons - (_r)->rsp_prod_pvt);                          \
213     req < rsp ? req : rsp;                                              \
214 })
215 
216 /* Direct access to individual ring elements, by index. */
217 #define RING_GET_REQUEST(_r, _idx)                                      \
218     (&((_r)->sring->ring[((_idx) & (RING_SIZE(_r) - 1))].req))
219 
220 /*
221  * Get a local copy of a request.
222  *
223  * Use this in preference to RING_GET_REQUEST() so all processing is
224  * done on a local copy that cannot be modified by the other end.
225  *
226  * Note that https://gcc.gnu.org/bugzilla/show_bug.cgi?id=58145 may cause this
227  * to be ineffective where _req is a struct which consists of only bitfields.
228  */
229 #define RING_COPY_REQUEST(_r, _idx, _req) do {				\
230         /* Use volatile to force the copy into _req. */			\
231         *(_req) = *(volatile typeof(_req))RING_GET_REQUEST(_r, _idx);	\
232 } while (0)
233 
234 #define RING_GET_RESPONSE(_r, _idx)                                     \
235     (&((_r)->sring->ring[((_idx) & (RING_SIZE(_r) - 1))].rsp))
236 
237 /* Loop termination condition: Would the specified index overflow the ring? */
238 #define RING_REQUEST_CONS_OVERFLOW(_r, _cons)                           \
239     (((_cons) - (_r)->rsp_prod_pvt) >= RING_SIZE(_r))
240 
241 /* Ill-behaved frontend determination: Can there be this many requests? */
242 #define RING_REQUEST_PROD_OVERFLOW(_r, _prod)                           \
243     (((_prod) - (_r)->rsp_prod_pvt) > RING_SIZE(_r))
244 
245 #define RING_PUSH_REQUESTS(_r) do {                                     \
246     xen_wmb(); /* back sees requests /before/ updated producer index */ \
247     (_r)->sring->req_prod = (_r)->req_prod_pvt;                         \
248 } while (0)
249 
250 #define RING_PUSH_RESPONSES(_r) do {                                    \
251     xen_wmb(); /* front sees resps /before/ updated producer index */   \
252     (_r)->sring->rsp_prod = (_r)->rsp_prod_pvt;                         \
253 } while (0)
254 
255 /*
256  * Notification hold-off (req_event and rsp_event):
257  *
258  * When queueing requests or responses on a shared ring, it may not always be
259  * necessary to notify the remote end. For example, if requests are in flight
260  * in a backend, the front may be able to queue further requests without
261  * notifying the back (if the back checks for new requests when it queues
262  * responses).
263  *
264  * When enqueuing requests or responses:
265  *
266  *  Use RING_PUSH_{REQUESTS,RESPONSES}_AND_CHECK_NOTIFY(). The second argument
267  *  is a boolean return value. True indicates that the receiver requires an
268  *  asynchronous notification.
269  *
270  * After dequeuing requests or responses (before sleeping the connection):
271  *
272  *  Use RING_FINAL_CHECK_FOR_REQUESTS() or RING_FINAL_CHECK_FOR_RESPONSES().
273  *  The second argument is a boolean return value. True indicates that there
274  *  are pending messages on the ring (i.e., the connection should not be put
275  *  to sleep).
276  *
277  *  These macros will set the req_event/rsp_event field to trigger a
278  *  notification on the very next message that is enqueued. If you want to
279  *  create batches of work (i.e., only receive a notification after several
280  *  messages have been enqueued) then you will need to create a customised
281  *  version of the FINAL_CHECK macro in your own code, which sets the event
282  *  field appropriately.
283  */
284 
285 #define RING_PUSH_REQUESTS_AND_CHECK_NOTIFY(_r, _notify) do {           \
286     RING_IDX __old = (_r)->sring->req_prod;                             \
287     RING_IDX __new = (_r)->req_prod_pvt;                                \
288     xen_wmb(); /* back sees requests /before/ updated producer index */ \
289     (_r)->sring->req_prod = __new;                                      \
290     xen_mb(); /* back sees new requests /before/ we check req_event */  \
291     (_notify) = ((RING_IDX)(__new - (_r)->sring->req_event) <           \
292                  (RING_IDX)(__new - __old));                            \
293 } while (0)
294 
295 #define RING_PUSH_RESPONSES_AND_CHECK_NOTIFY(_r, _notify) do {          \
296     RING_IDX __old = (_r)->sring->rsp_prod;                             \
297     RING_IDX __new = (_r)->rsp_prod_pvt;                                \
298     xen_wmb(); /* front sees resps /before/ updated producer index */   \
299     (_r)->sring->rsp_prod = __new;                                      \
300     xen_mb(); /* front sees new resps /before/ we check rsp_event */    \
301     (_notify) = ((RING_IDX)(__new - (_r)->sring->rsp_event) <           \
302                  (RING_IDX)(__new - __old));                            \
303 } while (0)
304 
305 #define RING_FINAL_CHECK_FOR_REQUESTS(_r, _work_to_do) do {             \
306     (_work_to_do) = RING_HAS_UNCONSUMED_REQUESTS(_r);                   \
307     if (_work_to_do) break;                                             \
308     (_r)->sring->req_event = (_r)->req_cons + 1;                        \
309     xen_mb();                                                           \
310     (_work_to_do) = RING_HAS_UNCONSUMED_REQUESTS(_r);                   \
311 } while (0)
312 
313 #define RING_FINAL_CHECK_FOR_RESPONSES(_r, _work_to_do) do {            \
314     (_work_to_do) = RING_HAS_UNCONSUMED_RESPONSES(_r);                  \
315     if (_work_to_do) break;                                             \
316     (_r)->sring->rsp_event = (_r)->rsp_cons + 1;                        \
317     xen_mb();                                                           \
318     (_work_to_do) = RING_HAS_UNCONSUMED_RESPONSES(_r);                  \
319 } while (0)
320 
321 
322 /*
323  * DEFINE_XEN_FLEX_RING_AND_INTF defines two monodirectional rings and
324  * functions to check if there is data on the ring, and to read and
325  * write to them.
326  *
327  * DEFINE_XEN_FLEX_RING is similar to DEFINE_XEN_FLEX_RING_AND_INTF, but
328  * does not define the indexes page. As different protocols can have
329  * extensions to the basic format, this macro allow them to define their
330  * own struct.
331  *
332  * XEN_FLEX_RING_SIZE
333  *   Convenience macro to calculate the size of one of the two rings
334  *   from the overall order.
335  *
336  * $NAME_mask
337  *   Function to apply the size mask to an index, to reduce the index
338  *   within the range [0-size].
339  *
340  * $NAME_read_packet
341  *   Function to read data from the ring. The amount of data to read is
342  *   specified by the "size" argument.
343  *
344  * $NAME_write_packet
345  *   Function to write data to the ring. The amount of data to write is
346  *   specified by the "size" argument.
347  *
348  * $NAME_get_ring_ptr
349  *   Convenience function that returns a pointer to read/write to the
350  *   ring at the right location.
351  *
352  * $NAME_data_intf
353  *   Indexes page, shared between frontend and backend. It also
354  *   contains the array of grant refs.
355  *
356  * $NAME_queued
357  *   Function to calculate how many bytes are currently on the ring,
358  *   ready to be read. It can also be used to calculate how much free
359  *   space is currently on the ring (XEN_FLEX_RING_SIZE() -
360  *   $NAME_queued()).
361  */
362 
363 #ifndef XEN_PAGE_SHIFT
364 /* The PAGE_SIZE for ring protocols and hypercall interfaces is always
365  * 4K, regardless of the architecture, and page granularity chosen by
366  * operating systems.
367  */
368 #define XEN_PAGE_SHIFT 12
369 #endif
370 #define XEN_FLEX_RING_SIZE(order)                                             \
371     (1UL << ((order) + XEN_PAGE_SHIFT - 1))
372 
373 #define DEFINE_XEN_FLEX_RING(name)                                            \
374 static inline RING_IDX name##_mask(RING_IDX idx, RING_IDX ring_size)          \
375 {                                                                             \
376     return idx & (ring_size - 1);                                             \
377 }                                                                             \
378                                                                               \
379 static inline unsigned char *name##_get_ring_ptr(unsigned char *buf,          \
380                                                  RING_IDX idx,                \
381                                                  RING_IDX ring_size)          \
382 {                                                                             \
383     return buf + name##_mask(idx, ring_size);                                 \
384 }                                                                             \
385                                                                               \
386 static inline void name##_read_packet(void *opaque,                           \
387                                       const unsigned char *buf,               \
388                                       size_t size,                            \
389                                       RING_IDX masked_prod,                   \
390                                       RING_IDX *masked_cons,                  \
391                                       RING_IDX ring_size)                     \
392 {                                                                             \
393     if (*masked_cons < masked_prod ||                                         \
394         size <= ring_size - *masked_cons) {                                   \
395         memcpy(opaque, buf + *masked_cons, size);                             \
396     } else {                                                                  \
397         memcpy(opaque, buf + *masked_cons, ring_size - *masked_cons);         \
398         memcpy((unsigned char *)opaque + ring_size - *masked_cons, buf,       \
399                size - (ring_size - *masked_cons));                            \
400     }                                                                         \
401     *masked_cons = name##_mask(*masked_cons + size, ring_size);               \
402 }                                                                             \
403                                                                               \
404 static inline void name##_write_packet(unsigned char *buf,                    \
405                                        const void *opaque,                    \
406                                        size_t size,                           \
407                                        RING_IDX *masked_prod,                 \
408                                        RING_IDX masked_cons,                  \
409                                        RING_IDX ring_size)                    \
410 {                                                                             \
411     if (*masked_prod < masked_cons ||                                         \
412         size <= ring_size - *masked_prod) {                                   \
413         memcpy(buf + *masked_prod, opaque, size);                             \
414     } else {                                                                  \
415         memcpy(buf + *masked_prod, opaque, ring_size - *masked_prod);         \
416         memcpy(buf, (unsigned char *)opaque + (ring_size - *masked_prod),     \
417                size - (ring_size - *masked_prod));                            \
418     }                                                                         \
419     *masked_prod = name##_mask(*masked_prod + size, ring_size);               \
420 }                                                                             \
421                                                                               \
422 static inline RING_IDX name##_queued(RING_IDX prod,                           \
423                                      RING_IDX cons,                           \
424                                      RING_IDX ring_size)                      \
425 {                                                                             \
426     RING_IDX size;                                                            \
427                                                                               \
428     if (prod == cons)                                                         \
429         return 0;                                                             \
430                                                                               \
431     prod = name##_mask(prod, ring_size);                                      \
432     cons = name##_mask(cons, ring_size);                                      \
433                                                                               \
434     if (prod == cons)                                                         \
435         return ring_size;                                                     \
436                                                                               \
437     if (prod > cons)                                                          \
438         size = prod - cons;                                                   \
439     else                                                                      \
440         size = ring_size - (cons - prod);                                     \
441     return size;                                                              \
442 }                                                                             \
443                                                                               \
444 struct name##_data {                                                          \
445     unsigned char *in; /* half of the allocation */                           \
446     unsigned char *out; /* half of the allocation */                          \
447 }
448 
449 #define DEFINE_XEN_FLEX_RING_AND_INTF(name)                                   \
450 struct name##_data_intf {                                                     \
451     RING_IDX in_cons, in_prod;                                                \
452                                                                               \
453     uint8_t pad1[56];                                                         \
454                                                                               \
455     RING_IDX out_cons, out_prod;                                              \
456                                                                               \
457     uint8_t pad2[56];                                                         \
458                                                                               \
459     RING_IDX ring_order;                                                      \
460     grant_ref_t ref[];                                                        \
461 };                                                                            \
462 DEFINE_XEN_FLEX_RING(name)
463 
464 #endif /* __XEN_PUBLIC_IO_RING_H__ */
465 
466 /*
467  * Local variables:
468  * mode: C
469  * c-file-style: "BSD"
470  * c-basic-offset: 4
471  * tab-width: 4
472  * indent-tabs-mode: nil
473  * End:
474  */
475