Home | History | Annotate | Line # | Download | only in io
      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  * In addition, if you intend to use the FLEX macros, you also need to
     37  * provide the following, before invoking the FLEX macros:
     38  * - size_t
     39  * - memcpy
     40  * - grant_ref_t
     41  * These declarations are provided by string.h of the standard headers,
     42  * and grant_table.h from the Xen public headers.
     43  */
     44 
     45 #include "../xen-compat.h"
     46 
     47 #if __XEN_INTERFACE_VERSION__ < 0x00030208
     48 #define xen_mb()  mb()
     49 #define xen_rmb() rmb()
     50 #define xen_wmb() wmb()
     51 #endif
     52 
     53 typedef unsigned int RING_IDX;
     54 
     55 /* Round a 32-bit unsigned constant down to the nearest power of two. */
     56 #define __RD2(_x)  (((_x) & 0x00000002) ? 0x2                  : ((_x) & 0x1))
     57 #define __RD4(_x)  (((_x) & 0x0000000c) ? __RD2((_x)>>2)<<2    : __RD2(_x))
     58 #define __RD8(_x)  (((_x) & 0x000000f0) ? __RD4((_x)>>4)<<4    : __RD4(_x))
     59 #define __RD16(_x) (((_x) & 0x0000ff00) ? __RD8((_x)>>8)<<8    : __RD8(_x))
     60 #define __RD32(_x) (((_x) & 0xffff0000) ? __RD16((_x)>>16)<<16 : __RD16(_x))
     61 
     62 /*
     63  * Calculate size of a shared ring, given the total available space for the
     64  * ring and indexes (_sz), and the name tag of the request/response structure.
     65  * A ring contains as many entries as will fit, rounded down to the nearest
     66  * power of two (so we can mask with (size-1) to loop around).
     67  */
     68 #define __CONST_RING_SIZE(_s, _sz) \
     69     (__RD32(((_sz) - offsetof(struct _s##_sring, ring)) / \
     70 	    sizeof(((struct _s##_sring *)0)->ring[0])))
     71 /*
     72  * The same for passing in an actual pointer instead of a name tag.
     73  */
     74 #define __RING_SIZE(_s, _sz) \
     75     (__RD32(((_sz) - (long)(_s)->ring + (long)(_s)) / sizeof((_s)->ring[0])))
     76 
     77 /*
     78  * Macros to make the correct C datatypes for a new kind of ring.
     79  *
     80  * To make a new ring datatype, you need to have two message structures,
     81  * let's say request_t, and response_t already defined.
     82  *
     83  * In a header where you want the ring datatype declared, you then do:
     84  *
     85  *     DEFINE_RING_TYPES(mytag, request_t, response_t);
     86  *
     87  * These expand out to give you a set of types, as you can see below.
     88  * The most important of these are:
     89  *
     90  *     mytag_sring_t      - The shared ring.
     91  *     mytag_front_ring_t - The 'front' half of the ring.
     92  *     mytag_back_ring_t  - The 'back' half of the ring.
     93  *
     94  * To initialize a ring in your code you need to know the location and size
     95  * of the shared memory area (PAGE_SIZE, for instance). To initialise
     96  * the front half:
     97  *
     98  *     mytag_front_ring_t front_ring;
     99  *     SHARED_RING_INIT((mytag_sring_t *)shared_page);
    100  *     FRONT_RING_INIT(&front_ring, (mytag_sring_t *)shared_page, PAGE_SIZE);
    101  *
    102  * Initializing the back follows similarly (note that only the front
    103  * initializes the shared ring):
    104  *
    105  *     mytag_back_ring_t back_ring;
    106  *     BACK_RING_INIT(&back_ring, (mytag_sring_t *)shared_page, PAGE_SIZE);
    107  */
    108 
    109 #define DEFINE_RING_TYPES(__name, __req_t, __rsp_t)                     \
    110                                                                         \
    111 /* Shared ring entry */                                                 \
    112 union __name##_sring_entry {                                            \
    113     __req_t req;                                                        \
    114     __rsp_t rsp;                                                        \
    115 };                                                                      \
    116                                                                         \
    117 /* Shared ring page */                                                  \
    118 struct __name##_sring {                                                 \
    119     RING_IDX req_prod, req_event;                                       \
    120     RING_IDX rsp_prod, rsp_event;                                       \
    121     union {                                                             \
    122         struct {                                                        \
    123             uint8_t smartpoll_active;                                   \
    124         } netif;                                                        \
    125         struct {                                                        \
    126             uint8_t msg;                                                \
    127         } tapif_user;                                                   \
    128         uint8_t pvt_pad[4];                                             \
    129     } pvt;                                                              \
    130     uint8_t __pad[44];                                                  \
    131     union __name##_sring_entry ring[1]; /* variable-length */           \
    132 };                                                                      \
    133                                                                         \
    134 /* "Front" end's private variables */                                   \
    135 struct __name##_front_ring {                                            \
    136     RING_IDX req_prod_pvt;                                              \
    137     RING_IDX rsp_cons;                                                  \
    138     unsigned int nr_ents;                                               \
    139     struct __name##_sring *sring;                                       \
    140 };                                                                      \
    141                                                                         \
    142 /* "Back" end's private variables */                                    \
    143 struct __name##_back_ring {                                             \
    144     RING_IDX rsp_prod_pvt;                                              \
    145     RING_IDX req_cons;                                                  \
    146     unsigned int nr_ents;                                               \
    147     struct __name##_sring *sring;                                       \
    148 };                                                                      \
    149                                                                         \
    150 /* Syntactic sugar */                                                   \
    151 typedef struct __name##_sring __name##_sring_t;                         \
    152 typedef struct __name##_front_ring __name##_front_ring_t;               \
    153 typedef struct __name##_back_ring __name##_back_ring_t
    154 
    155 /*
    156  * Macros for manipulating rings.
    157  *
    158  * FRONT_RING_whatever works on the "front end" of a ring: here
    159  * requests are pushed on to the ring and responses taken off it.
    160  *
    161  * BACK_RING_whatever works on the "back end" of a ring: here
    162  * requests are taken off the ring and responses put on.
    163  *
    164  * N.B. these macros do NO INTERLOCKS OR FLOW CONTROL.
    165  * This is OK in 1-for-1 request-response situations where the
    166  * requestor (front end) never has more than RING_SIZE()-1
    167  * outstanding requests.
    168  */
    169 
    170 /* Initialising empty rings */
    171 #define SHARED_RING_INIT(_s) do {                                       \
    172     (_s)->req_prod  = (_s)->rsp_prod  = 0;                              \
    173     (_s)->req_event = (_s)->rsp_event = 1;                              \
    174     (void)memset((_s)->pvt.pvt_pad, 0, sizeof((_s)->pvt.pvt_pad));      \
    175     (void)memset((_s)->__pad, 0, sizeof((_s)->__pad));                  \
    176 } while(0)
    177 
    178 #define FRONT_RING_INIT(_r, _s, __size) do {                            \
    179     (_r)->req_prod_pvt = 0;                                             \
    180     (_r)->rsp_cons = 0;                                                 \
    181     (_r)->nr_ents = __RING_SIZE(_s, __size);                            \
    182     (_r)->sring = (_s);                                                 \
    183 } while (0)
    184 
    185 #define BACK_RING_INIT(_r, _s, __size) do {                             \
    186     (_r)->rsp_prod_pvt = 0;                                             \
    187     (_r)->req_cons = 0;                                                 \
    188     (_r)->nr_ents = __RING_SIZE(_s, __size);                            \
    189     (_r)->sring = (_s);                                                 \
    190 } while (0)
    191 
    192 /* How big is this ring? */
    193 #define RING_SIZE(_r)                                                   \
    194     ((_r)->nr_ents)
    195 
    196 /* Number of free requests (for use on front side only). */
    197 #define RING_FREE_REQUESTS(_r)                                          \
    198     (RING_SIZE(_r) - ((_r)->req_prod_pvt - (_r)->rsp_cons))
    199 
    200 /* Test if there is an empty slot available on the front ring.
    201  * (This is only meaningful from the front. )
    202  */
    203 #define RING_FULL(_r)                                                   \
    204     (RING_FREE_REQUESTS(_r) == 0)
    205 
    206 /* Test if there are outstanding messages to be processed on a ring. */
    207 #define RING_HAS_UNCONSUMED_RESPONSES(_r)                               \
    208     ((_r)->sring->rsp_prod - (_r)->rsp_cons)
    209 
    210 #ifdef __GNUC__
    211 #define RING_HAS_UNCONSUMED_REQUESTS(_r) ({                             \
    212     unsigned int req = (_r)->sring->req_prod - (_r)->req_cons;          \
    213     unsigned int rsp = RING_SIZE(_r) -                                  \
    214         ((_r)->req_cons - (_r)->rsp_prod_pvt);                          \
    215     req < rsp ? req : rsp;                                              \
    216 })
    217 #else
    218 /* Same as above, but without the nice GCC ({ ... }) syntax. */
    219 #define RING_HAS_UNCONSUMED_REQUESTS(_r)                                \
    220     ((((_r)->sring->req_prod - (_r)->req_cons) <                        \
    221       (RING_SIZE(_r) - ((_r)->req_cons - (_r)->rsp_prod_pvt))) ?        \
    222      ((_r)->sring->req_prod - (_r)->req_cons) :                         \
    223      (RING_SIZE(_r) - ((_r)->req_cons - (_r)->rsp_prod_pvt)))
    224 #endif
    225 
    226 /* Direct access to individual ring elements, by index. */
    227 #define RING_GET_REQUEST(_r, _idx)                                      \
    228     (&((_r)->sring->ring[((_idx) & (RING_SIZE(_r) - 1))].req))
    229 
    230 /*
    231  * Get a local copy of a request.
    232  *
    233  * Use this in preference to RING_GET_REQUEST() so all processing is
    234  * done on a local copy that cannot be modified by the other end.
    235  *
    236  * Note that https://gcc.gnu.org/bugzilla/show_bug.cgi?id=58145 may cause this
    237  * to be ineffective where _req is a struct which consists of only bitfields.
    238  */
    239 #define RING_COPY_REQUEST(_r, _idx, _req) do {				\
    240 	/* Use volatile to force the copy into _req. */			\
    241 	*(_req) = *(volatile typeof(_req))RING_GET_REQUEST(_r, _idx);	\
    242 } while (0)
    243 
    244 #define RING_GET_RESPONSE(_r, _idx)                                     \
    245     (&((_r)->sring->ring[((_idx) & (RING_SIZE(_r) - 1))].rsp))
    246 
    247 /* Loop termination condition: Would the specified index overflow the ring? */
    248 #define RING_REQUEST_CONS_OVERFLOW(_r, _cons)                           \
    249     (((_cons) - (_r)->rsp_prod_pvt) >= RING_SIZE(_r))
    250 
    251 /* Ill-behaved frontend determination: Can there be this many requests? */
    252 #define RING_REQUEST_PROD_OVERFLOW(_r, _prod)                           \
    253     (((_prod) - (_r)->rsp_prod_pvt) > RING_SIZE(_r))
    254 
    255 #define RING_PUSH_REQUESTS(_r) do {                                     \
    256     xen_wmb(); /* back sees requests /before/ updated producer index */ \
    257     (_r)->sring->req_prod = (_r)->req_prod_pvt;                         \
    258 } while (0)
    259 
    260 #define RING_PUSH_RESPONSES(_r) do {                                    \
    261     xen_wmb(); /* front sees resps /before/ updated producer index */   \
    262     (_r)->sring->rsp_prod = (_r)->rsp_prod_pvt;                         \
    263 } while (0)
    264 
    265 /*
    266  * Notification hold-off (req_event and rsp_event):
    267  *
    268  * When queueing requests or responses on a shared ring, it may not always be
    269  * necessary to notify the remote end. For example, if requests are in flight
    270  * in a backend, the front may be able to queue further requests without
    271  * notifying the back (if the back checks for new requests when it queues
    272  * responses).
    273  *
    274  * When enqueuing requests or responses:
    275  *
    276  *  Use RING_PUSH_{REQUESTS,RESPONSES}_AND_CHECK_NOTIFY(). The second argument
    277  *  is a boolean return value. True indicates that the receiver requires an
    278  *  asynchronous notification.
    279  *
    280  * After dequeuing requests or responses (before sleeping the connection):
    281  *
    282  *  Use RING_FINAL_CHECK_FOR_REQUESTS() or RING_FINAL_CHECK_FOR_RESPONSES().
    283  *  The second argument is a boolean return value. True indicates that there
    284  *  are pending messages on the ring (i.e., the connection should not be put
    285  *  to sleep).
    286  *
    287  *  These macros will set the req_event/rsp_event field to trigger a
    288  *  notification on the very next message that is enqueued. If you want to
    289  *  create batches of work (i.e., only receive a notification after several
    290  *  messages have been enqueued) then you will need to create a customised
    291  *  version of the FINAL_CHECK macro in your own code, which sets the event
    292  *  field appropriately.
    293  */
    294 
    295 #define RING_PUSH_REQUESTS_AND_CHECK_NOTIFY(_r, _notify) do {           \
    296     RING_IDX __old = (_r)->sring->req_prod;                             \
    297     RING_IDX __new = (_r)->req_prod_pvt;                                \
    298     xen_wmb(); /* back sees requests /before/ updated producer index */ \
    299     (_r)->sring->req_prod = __new;                                      \
    300     xen_mb(); /* back sees new requests /before/ we check req_event */  \
    301     (_notify) = ((RING_IDX)(__new - (_r)->sring->req_event) <           \
    302                  (RING_IDX)(__new - __old));                            \
    303 } while (0)
    304 
    305 #define RING_PUSH_RESPONSES_AND_CHECK_NOTIFY(_r, _notify) do {          \
    306     RING_IDX __old = (_r)->sring->rsp_prod;                             \
    307     RING_IDX __new = (_r)->rsp_prod_pvt;                                \
    308     xen_wmb(); /* front sees resps /before/ updated producer index */   \
    309     (_r)->sring->rsp_prod = __new;                                      \
    310     xen_mb(); /* front sees new resps /before/ we check rsp_event */    \
    311     (_notify) = ((RING_IDX)(__new - (_r)->sring->rsp_event) <           \
    312                  (RING_IDX)(__new - __old));                            \
    313 } while (0)
    314 
    315 #define RING_FINAL_CHECK_FOR_REQUESTS(_r, _work_to_do) do {             \
    316     (_work_to_do) = RING_HAS_UNCONSUMED_REQUESTS(_r);                   \
    317     if (_work_to_do) break;                                             \
    318     (_r)->sring->req_event = (_r)->req_cons + 1;                        \
    319     xen_mb();                                                           \
    320     (_work_to_do) = RING_HAS_UNCONSUMED_REQUESTS(_r);                   \
    321 } while (0)
    322 
    323 #define RING_FINAL_CHECK_FOR_RESPONSES(_r, _work_to_do) do {            \
    324     (_work_to_do) = RING_HAS_UNCONSUMED_RESPONSES(_r);                  \
    325     if (_work_to_do) break;                                             \
    326     (_r)->sring->rsp_event = (_r)->rsp_cons + 1;                        \
    327     xen_mb();                                                           \
    328     (_work_to_do) = RING_HAS_UNCONSUMED_RESPONSES(_r);                  \
    329 } while (0)
    330 
    331 
    332 /*
    333  * DEFINE_XEN_FLEX_RING_AND_INTF defines two monodirectional rings and
    334  * functions to check if there is data on the ring, and to read and
    335  * write to them.
    336  *
    337  * DEFINE_XEN_FLEX_RING is similar to DEFINE_XEN_FLEX_RING_AND_INTF, but
    338  * does not define the indexes page. As different protocols can have
    339  * extensions to the basic format, this macro allow them to define their
    340  * own struct.
    341  *
    342  * XEN_FLEX_RING_SIZE
    343  *   Convenience macro to calculate the size of one of the two rings
    344  *   from the overall order.
    345  *
    346  * $NAME_mask
    347  *   Function to apply the size mask to an index, to reduce the index
    348  *   within the range [0-size].
    349  *
    350  * $NAME_read_packet
    351  *   Function to read data from the ring. The amount of data to read is
    352  *   specified by the "size" argument.
    353  *
    354  * $NAME_write_packet
    355  *   Function to write data to the ring. The amount of data to write is
    356  *   specified by the "size" argument.
    357  *
    358  * $NAME_get_ring_ptr
    359  *   Convenience function that returns a pointer to read/write to the
    360  *   ring at the right location.
    361  *
    362  * $NAME_data_intf
    363  *   Indexes page, shared between frontend and backend. It also
    364  *   contains the array of grant refs.
    365  *
    366  * $NAME_queued
    367  *   Function to calculate how many bytes are currently on the ring,
    368  *   ready to be read. It can also be used to calculate how much free
    369  *   space is currently on the ring (XEN_FLEX_RING_SIZE() -
    370  *   $NAME_queued()).
    371  */
    372 
    373 #ifndef XEN_PAGE_SHIFT
    374 /* The PAGE_SIZE for ring protocols and hypercall interfaces is always
    375  * 4K, regardless of the architecture, and page granularity chosen by
    376  * operating systems.
    377  */
    378 #define XEN_PAGE_SHIFT 12
    379 #endif
    380 #define XEN_FLEX_RING_SIZE(order)                                             \
    381     (1UL << ((order) + XEN_PAGE_SHIFT - 1))
    382 
    383 #define DEFINE_XEN_FLEX_RING(name)                                            \
    384 static inline RING_IDX name##_mask(RING_IDX idx, RING_IDX ring_size)          \
    385 {                                                                             \
    386     return idx & (ring_size - 1);                                             \
    387 }                                                                             \
    388                                                                               \
    389 static inline unsigned char *name##_get_ring_ptr(unsigned char *buf,          \
    390                                                  RING_IDX idx,                \
    391                                                  RING_IDX ring_size)          \
    392 {                                                                             \
    393     return buf + name##_mask(idx, ring_size);                                 \
    394 }                                                                             \
    395                                                                               \
    396 static inline void name##_read_packet(void *opaque,                           \
    397                                       const unsigned char *buf,               \
    398                                       size_t size,                            \
    399                                       RING_IDX masked_prod,                   \
    400                                       RING_IDX *masked_cons,                  \
    401                                       RING_IDX ring_size)                     \
    402 {                                                                             \
    403     if (*masked_cons < masked_prod ||                                         \
    404         size <= ring_size - *masked_cons) {                                   \
    405         memcpy(opaque, buf + *masked_cons, size);                             \
    406     } else {                                                                  \
    407         memcpy(opaque, buf + *masked_cons, ring_size - *masked_cons);         \
    408         memcpy((unsigned char *)opaque + ring_size - *masked_cons, buf,       \
    409                size - (ring_size - *masked_cons));                            \
    410     }                                                                         \
    411     *masked_cons = name##_mask(*masked_cons + size, ring_size);               \
    412 }                                                                             \
    413                                                                               \
    414 static inline void name##_write_packet(unsigned char *buf,                    \
    415                                        const void *opaque,                    \
    416                                        size_t size,                           \
    417                                        RING_IDX *masked_prod,                 \
    418                                        RING_IDX masked_cons,                  \
    419                                        RING_IDX ring_size)                    \
    420 {                                                                             \
    421     if (*masked_prod < masked_cons ||                                         \
    422         size <= ring_size - *masked_prod) {                                   \
    423         memcpy(buf + *masked_prod, opaque, size);                             \
    424     } else {                                                                  \
    425         memcpy(buf + *masked_prod, opaque, ring_size - *masked_prod);         \
    426         memcpy(buf, (unsigned char *)opaque + (ring_size - *masked_prod),     \
    427                size - (ring_size - *masked_prod));                            \
    428     }                                                                         \
    429     *masked_prod = name##_mask(*masked_prod + size, ring_size);               \
    430 }                                                                             \
    431                                                                               \
    432 static inline RING_IDX name##_queued(RING_IDX prod,                           \
    433                                      RING_IDX cons,                           \
    434                                      RING_IDX ring_size)                      \
    435 {                                                                             \
    436     RING_IDX size;                                                            \
    437                                                                               \
    438     if (prod == cons)                                                         \
    439         return 0;                                                             \
    440                                                                               \
    441     prod = name##_mask(prod, ring_size);                                      \
    442     cons = name##_mask(cons, ring_size);                                      \
    443                                                                               \
    444     if (prod == cons)                                                         \
    445         return ring_size;                                                     \
    446                                                                               \
    447     if (prod > cons)                                                          \
    448         size = prod - cons;                                                   \
    449     else                                                                      \
    450         size = ring_size - (cons - prod);                                     \
    451     return size;                                                              \
    452 }                                                                             \
    453                                                                               \
    454 struct name##_data {                                                          \
    455     unsigned char *in; /* half of the allocation */                           \
    456     unsigned char *out; /* half of the allocation */                          \
    457 }
    458 
    459 #define DEFINE_XEN_FLEX_RING_AND_INTF(name)                                   \
    460 struct name##_data_intf {                                                     \
    461     RING_IDX in_cons, in_prod;                                                \
    462                                                                               \
    463     uint8_t pad1[56];                                                         \
    464                                                                               \
    465     RING_IDX out_cons, out_prod;                                              \
    466                                                                               \
    467     uint8_t pad2[56];                                                         \
    468                                                                               \
    469     RING_IDX ring_order;                                                      \
    470     grant_ref_t ref[];                                                        \
    471 };                                                                            \
    472 DEFINE_XEN_FLEX_RING(name)
    473 
    474 #endif /* __XEN_PUBLIC_IO_RING_H__ */
    475 
    476 /*
    477  * Local variables:
    478  * mode: C
    479  * c-file-style: "BSD"
    480  * c-basic-offset: 4
    481  * tab-width: 4
    482  * indent-tabs-mode: nil
    483  * End:
    484  */
    485