[BACK]Return to cyclic.c CVS log [TXT][DIR] Up to [cvs.NetBSD.org] / src / external / cddl / osnet / dev / cyclic

Annotation of src/external/cddl/osnet/dev/cyclic/cyclic.c, Revision 1.2.6.1

1.2.6.1 ! yamt        1: /*     $NetBSD: cyclic.c,v 1.2 2010/02/21 01:46:33 darran Exp $        */
1.2       darran      2:
1.1       darran      3: /*
                      4:  * CDDL HEADER START
                      5:  *
                      6:  * The contents of this file are subject to the terms of the
                      7:  * Common Development and Distribution License, Version 1.0 only
                      8:  * (the "License").  You may not use this file except in compliance
                      9:  * with the License.
                     10:  *
                     11:  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
                     12:  * or http://www.opensolaris.org/os/licensing.
                     13:  * See the License for the specific language governing permissions
                     14:  * and limitations under the License.
                     15:  *
                     16:  * When distributing Covered Code, include this CDDL HEADER in each
                     17:  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
                     18:  * If applicable, add the following below this CDDL HEADER, with the
                     19:  * fields enclosed by brackets "[]" replaced with your own identifying
                     20:  * information: Portions Copyright [yyyy] [name of copyright owner]
                     21:  *
                     22:  * CDDL HEADER END
                     23:  *
                     24:  * Portions Copyright 2008 John Birrell <jb@freebsd.org>
                     25:  *
1.2.6.1 ! yamt       26:  * $FreeBSD$
1.1       darran     27:  *
                     28:  * This is a simplified version of the cyclic timer subsystem from
                     29:  * OpenSolaris. In the FreeBSD version, we don't use interrupt levels.
                     30:  */
                     31:
                     32: /*
                     33:  * Copyright 2004 Sun Microsystems, Inc.  All rights reserved.
                     34:  * Use is subject to license terms.
                     35:  */
                     36:
                     37: /*
                     38:  *  The Cyclic Subsystem
                     39:  *  --------------------
                     40:  *
                     41:  *  Prehistory
                     42:  *
                     43:  *  Historically, most computer architectures have specified interval-based
                     44:  *  timer parts (e.g. SPARCstation's counter/timer; Intel's i8254).  While
                     45:  *  these parts deal in relative (i.e. not absolute) time values, they are
                     46:  *  typically used by the operating system to implement the abstraction of
                     47:  *  absolute time.  As a result, these parts cannot typically be reprogrammed
                     48:  *  without introducing error in the system's notion of time.
                     49:  *
                     50:  *  Starting in about 1994, chip architectures began specifying high resolution
                     51:  *  timestamp registers.  As of this writing (1999), all major chip families
                     52:  *  (UltraSPARC, PentiumPro, MIPS, PowerPC, Alpha) have high resolution
                     53:  *  timestamp registers, and two (UltraSPARC and MIPS) have added the capacity
                     54:  *  to interrupt based on timestamp values.  These timestamp-compare registers
                     55:  *  present a time-based interrupt source which can be reprogrammed arbitrarily
                     56:  *  often without introducing error.  Given the low cost of implementing such a
                     57:  *  timestamp-compare register (and the tangible benefit of eliminating
                     58:  *  discrete timer parts), it is reasonable to expect that future chip
                     59:  *  architectures will adopt this feature.
                     60:  *
                     61:  *  The cyclic subsystem has been designed to take advantage of chip
                     62:  *  architectures with the capacity to interrupt based on absolute, high
                     63:  *  resolution values of time.
                     64:  *
                     65:  *  Subsystem Overview
                     66:  *
                     67:  *  The cyclic subsystem is a low-level kernel subsystem designed to provide
                     68:  *  arbitrarily high resolution, per-CPU interval timers (to avoid colliding
                     69:  *  with existing terms, we dub such an interval timer a "cyclic").
                     70:  *  Alternatively, a cyclic may be specified to be "omnipresent", denoting
                     71:  *  firing on all online CPUs.
                     72:  *
                     73:  *  Cyclic Subsystem Interface Overview
                     74:  *  -----------------------------------
                     75:  *
                     76:  *  The cyclic subsystem has interfaces with the kernel at-large, with other
                     77:  *  kernel subsystems (e.g. the processor management subsystem, the checkpoint
                     78:  *  resume subsystem) and with the platform (the cyclic backend).  Each
                     79:  *  of these interfaces is given a brief synopsis here, and is described
                     80:  *  in full above the interface's implementation.
                     81:  *
                     82:  *  The following diagram displays the cyclic subsystem's interfaces to
                     83:  *  other kernel components.  The arrows denote a "calls" relationship, with
                     84:  *  the large arrow indicating the cyclic subsystem's consumer interface.
                     85:  *  Each arrow is labeled with the section in which the corresponding
                     86:  *  interface is described.
                     87:  *
                     88:  *           Kernel at-large consumers
                     89:  *           -----------++------------
                     90:  *                      ||
                     91:  *                      ||
                     92:  *                     _||_
                     93:  *                     \  /
                     94:  *                      \/
                     95:  *            +---------------------+
                     96:  *            |                     |
                     97:  *            |  Cyclic subsystem   |<-----------  Other kernel subsystems
                     98:  *            |                     |
                     99:  *            +---------------------+
                    100:  *                   ^       |
                    101:  *                   |       |
                    102:  *                   |       |
                    103:  *                   |       v
                    104:  *            +---------------------+
                    105:  *            |                     |
                    106:  *            |   Cyclic backend    |
                    107:  *            | (platform specific) |
                    108:  *            |                     |
                    109:  *            +---------------------+
                    110:  *
                    111:  *
                    112:  *  Kernel At-Large Interfaces
                    113:  *
                    114:  *      cyclic_add()         <-- Creates a cyclic
                    115:  *      cyclic_add_omni()    <-- Creates an omnipresent cyclic
                    116:  *      cyclic_remove()      <-- Removes a cyclic
                    117:  *
                    118:  *  Backend Interfaces
                    119:  *
                    120:  *      cyclic_init()        <-- Initializes the cyclic subsystem
                    121:  *      cyclic_fire()        <-- Interrupt entry point
                    122:  *
                    123:  *  The backend-supplied interfaces (through the cyc_backend structure) are
                    124:  *  documented in detail in <sys/cyclic_impl.h>
                    125:  *
                    126:  *
                    127:  *  Cyclic Subsystem Implementation Overview
                    128:  *  ----------------------------------------
                    129:  *
                    130:  *  The cyclic subsystem is designed to minimize interference between cyclics
                    131:  *  on different CPUs.  Thus, all of the cyclic subsystem's data structures
                    132:  *  hang off of a per-CPU structure, cyc_cpu.
                    133:  *
                    134:  *  Each cyc_cpu has a power-of-two sized array of cyclic structures (the
                    135:  *  cyp_cyclics member of the cyc_cpu structure).  If cyclic_add() is called
                    136:  *  and there does not exist a free slot in the cyp_cyclics array, the size of
                    137:  *  the array will be doubled.  The array will never shrink.  Cyclics are
                    138:  *  referred to by their index in the cyp_cyclics array, which is of type
                    139:  *  cyc_index_t.
                    140:  *
                    141:  *  The cyclics are kept sorted by expiration time in the cyc_cpu's heap.  The
                    142:  *  heap is keyed by cyclic expiration time, with parents expiring earlier
                    143:  *  than their children.
                    144:  *
                    145:  *  Heap Management
                    146:  *
                    147:  *  The heap is managed primarily by cyclic_fire().  Upon entry, cyclic_fire()
                    148:  *  compares the root cyclic's expiration time to the current time.  If the
                    149:  *  expiration time is in the past, cyclic_expire() is called on the root
                    150:  *  cyclic.  Upon return from cyclic_expire(), the cyclic's new expiration time
                    151:  *  is derived by adding its interval to its old expiration time, and a
                    152:  *  downheap operation is performed.  After the downheap, cyclic_fire()
                    153:  *  examines the (potentially changed) root cyclic, repeating the
                    154:  *  cyclic_expire()/add interval/cyclic_downheap() sequence until the root
                    155:  *  cyclic has an expiration time in the future.  This expiration time
                    156:  *  (guaranteed to be the earliest in the heap) is then communicated to the
                    157:  *  backend via cyb_reprogram.  Optimal backends will next call cyclic_fire()
                    158:  *  shortly after the root cyclic's expiration time.
                    159:  *
                    160:  *  To allow efficient, deterministic downheap operations, we implement the
                    161:  *  heap as an array (the cyp_heap member of the cyc_cpu structure), with each
                    162:  *  element containing an index into the CPU's cyp_cyclics array.
                    163:  *
                    164:  *  The heap is laid out in the array according to the following:
                    165:  *
                    166:  *   1.  The root of the heap is always in the 0th element of the heap array
                    167:  *   2.  The left and right children of the nth element are element
                    168:  *       (((n + 1) << 1) - 1) and element ((n + 1) << 1), respectively.
                    169:  *
                    170:  *  This layout is standard (see, e.g., Cormen's "Algorithms"); the proof
                    171:  *  that these constraints correctly lay out a heap (or indeed, any binary
                    172:  *  tree) is trivial and left to the reader.
                    173:  *
                    174:  *  To see the heap by example, assume our cyclics array has the following
                    175:  *  members (at time t):
                    176:  *
                    177:  *            cy_handler                          cy_expire
                    178:  *            ---------------------------------------------
                    179:  *     [ 0]   clock()                            t+10000000
                    180:  *     [ 1]   deadman()                        t+1000000000
                    181:  *     [ 2]   clock_highres_fire()                    t+100
                    182:  *     [ 3]   clock_highres_fire()                   t+1000
                    183:  *     [ 4]   clock_highres_fire()                    t+500
                    184:  *     [ 5]   (free)                                     --
                    185:  *     [ 6]   (free)                                     --
                    186:  *     [ 7]   (free)                                     --
                    187:  *
                    188:  *  The heap array could be:
                    189:  *
                    190:  *                [0]   [1]   [2]   [3]   [4]   [5]   [6]   [7]
                    191:  *              +-----+-----+-----+-----+-----+-----+-----+-----+
                    192:  *              |     |     |     |     |     |     |     |     |
                    193:  *              |  2  |  3  |  4  |  0  |  1  |  x  |  x  |  x  |
                    194:  *              |     |     |     |     |     |     |     |     |
                    195:  *              +-----+-----+-----+-----+-----+-----+-----+-----+
                    196:  *
                    197:  *  Graphically, this array corresponds to the following (excuse the ASCII art):
                    198:  *
                    199:  *                                       2
                    200:  *                                       |
                    201:  *                    +------------------+------------------+
                    202:  *                    3                                     4
                    203:  *                    |
                    204:  *          +---------+--------+
                    205:  *          0                  1
                    206:  *
                    207:  *  Note that the heap is laid out by layer:  all nodes at a given depth are
                    208:  *  stored in consecutive elements of the array.  Moreover, layers of
                    209:  *  consecutive depths are in adjacent element ranges.  This property
                    210:  *  guarantees high locality of reference during downheap operations.
                    211:  *  Specifically, we are guaranteed that we can downheap to a depth of
                    212:  *
                    213:  *      lg (cache_line_size / sizeof (cyc_index_t))
                    214:  *
                    215:  *  nodes with at most one cache miss.  On UltraSPARC (64 byte e-cache line
                    216:  *  size), this corresponds to a depth of four nodes.  Thus, if there are
                    217:  *  fewer than sixteen cyclics in the heap, downheaps on UltraSPARC miss at
                    218:  *  most once in the e-cache.
                    219:  *
                    220:  *  Downheaps are required to compare siblings as they proceed down the
                    221:  *  heap.  For downheaps proceeding beyond the one-cache-miss depth, every
                    222:  *  access to a left child could potentially miss in the cache.  However,
                    223:  *  if we assume
                    224:  *
                    225:  *      (cache_line_size / sizeof (cyc_index_t)) > 2,
                    226:  *
                    227:  *  then all siblings are guaranteed to be on the same cache line.  Thus, the
                    228:  *  miss on the left child will guarantee a hit on the right child; downheaps
                    229:  *  will incur at most one cache miss per layer beyond the one-cache-miss
                    230:  *  depth.  The total number of cache misses for heap management during a
                    231:  *  downheap operation is thus bounded by
                    232:  *
                    233:  *      lg (n) - lg (cache_line_size / sizeof (cyc_index_t))
                    234:  *
                    235:  *  Traditional pointer-based heaps are implemented without regard to
                    236:  *  locality.  Downheaps can thus incur two cache misses per layer (one for
                    237:  *  each child), but at most one cache miss at the root.  This yields a bound
                    238:  *  of
                    239:  *
                    240:  *      2 * lg (n) - 1
                    241:  *
                    242:  *  on the total cache misses.
                    243:  *
                    244:  *  This difference may seem theoretically trivial (the difference is, after
                    245:  *  all, constant), but can become substantial in practice -- especially for
                    246:  *  caches with very large cache lines and high miss penalties (e.g. TLBs).
                    247:  *
                    248:  *  Heaps must always be full, balanced trees.  Heap management must therefore
                    249:  *  track the next point-of-insertion into the heap.  In pointer-based heaps,
                    250:  *  recomputing this point takes O(lg (n)).  Given the layout of the
                    251:  *  array-based implementation, however, the next point-of-insertion is
                    252:  *  always:
                    253:  *
                    254:  *      heap[number_of_elements]
                    255:  *
                    256:  *  We exploit this property by implementing the free-list in the usused
                    257:  *  heap elements.  Heap insertion, therefore, consists only of filling in
                    258:  *  the cyclic at cyp_cyclics[cyp_heap[number_of_elements]], incrementing
                    259:  *  the number of elements, and performing an upheap.  Heap deletion consists
                    260:  *  of decrementing the number of elements, swapping the to-be-deleted element
                    261:  *  with the element at cyp_heap[number_of_elements], and downheaping.
                    262:  *
                    263:  *  Filling in more details in our earlier example:
                    264:  *
                    265:  *                                               +--- free list head
                    266:  *                                               |
                    267:  *                                               V
                    268:  *
                    269:  *                [0]   [1]   [2]   [3]   [4]   [5]   [6]   [7]
                    270:  *              +-----+-----+-----+-----+-----+-----+-----+-----+
                    271:  *              |     |     |     |     |     |     |     |     |
                    272:  *              |  2  |  3  |  4  |  0  |  1  |  5  |  6  |  7  |
                    273:  *              |     |     |     |     |     |     |     |     |
                    274:  *              +-----+-----+-----+-----+-----+-----+-----+-----+
                    275:  *
                    276:  *  To insert into this heap, we would just need to fill in the cyclic at
                    277:  *  cyp_cyclics[5], bump the number of elements (from 5 to 6) and perform
                    278:  *  an upheap.
                    279:  *
                    280:  *  If we wanted to remove, say, cyp_cyclics[3], we would first scan for it
                    281:  *  in the cyp_heap, and discover it at cyp_heap[1].  We would then decrement
                    282:  *  the number of elements (from 5 to 4), swap cyp_heap[1] with cyp_heap[4],
                    283:  *  and perform a downheap from cyp_heap[1].  The linear scan is required
                    284:  *  because the cyclic does not keep a backpointer into the heap.  This makes
                    285:  *  heap manipulation (e.g. downheaps) faster at the expense of removal
                    286:  *  operations.
                    287:  *
                    288:  *  Expiry processing
                    289:  *
                    290:  *  As alluded to above, cyclic_expire() is called by cyclic_fire() to expire
                    291:  *  a cyclic.  Cyclic subsystem consumers are guaranteed that for an arbitrary
                    292:  *  time t in the future, their cyclic handler will have been called
                    293:  *  (t - cyt_when) / cyt_interval times. cyclic_expire() simply needs to call
                    294:  *  the handler.
                    295:  *
                    296:  *  Resizing
                    297:  *
                    298:  *  All of the discussion thus far has assumed a static number of cyclics.
                    299:  *  Obviously, static limitations are not practical; we need the capacity
                    300:  *  to resize our data structures dynamically.
                    301:  *
                    302:  *  We resize our data structures lazily, and only on a per-CPU basis.
                    303:  *  The size of the data structures always doubles and never shrinks.  We
                    304:  *  serialize adds (and thus resizes) on cpu_lock; we never need to deal
                    305:  *  with concurrent resizes.  Resizes should be rare; they may induce jitter
                    306:  *  on the CPU being resized, but should not affect cyclic operation on other
                    307:  *  CPUs.
                    308:  *
                    309:  *  Three key cyc_cpu data structures need to be resized:  the cyclics array,
                    310:  *  nad the heap array.  Resizing is relatively straightforward:
                    311:  *
                    312:  *    1.  The new, larger arrays are allocated in cyclic_expand() (called
                    313:  *        from cyclic_add()).
                    314:  *    2.  The contents of the old arrays are copied into the new arrays.
                    315:  *    3.  The old cyclics array is bzero()'d
                    316:  *    4.  The pointers are updated.
                    317:  *
                    318:  *  Removals
                    319:  *
                    320:  *  Cyclic removals should be rare.  To simplify the implementation (and to
                    321:  *  allow optimization for the cyclic_fire()/cyclic_expire()
                    322:  *  path), we force removals and adds to serialize on cpu_lock.
                    323:  *
                    324:  */
                    325: #include <sys/cdefs.h>
                    326: #include <sys/param.h>
                    327: #include <sys/conf.h>
                    328: #include <sys/kernel.h>
1.2.6.1 ! yamt      329: #ifdef __FreeBSD___
1.1       darran    330: #include <sys/lock.h>
                    331: #include <sys/sx.h>
1.2.6.1 ! yamt      332: #endif
1.1       darran    333: #include <sys/cyclic_impl.h>
                    334: #include <sys/module.h>
                    335: #include <sys/systm.h>
                    336: #include <sys/atomic.h>
                    337: #include <sys/kmem.h>
                    338: #include <sys/cmn_err.h>
                    339: #include <sys/dtrace_bsd.h>
1.2.6.1 ! yamt      340: #ifdef __FreeBSD__
1.1       darran    341: #include <machine/cpu.h>
1.2.6.1 ! yamt      342: #endif
        !           343:
        !           344: #ifdef __NetBSD__
        !           345: #include <sys/cpu.h>
        !           346: #include <sys/malloc.h>
        !           347: #include <sys/xcall.h>
        !           348:
        !           349: #undef mutex_init
        !           350: #define mtx_init(m, d, p, f) mutex_init(m, MUTEX_DEFAULT, IPL_CLOCK)
        !           351: #define mtx_lock_spin(x) mutex_spin_enter(x)
        !           352: #define mtx_unlock_spin(x) mutex_spin_exit(x)
        !           353: #define mtx_destroy(x) mutex_destroy(x)
        !           354:
        !           355: #define ASSERT(x) KASSERT(x)
        !           356: #define SYSINIT(a1, a2, a3, a4, a5)
        !           357: #define SYSUNINIT(a1, a2, a3, a4, a5)
        !           358: #define CPU_FOREACH(var) \
        !           359:        CPU_INFO_ITERATOR cii; \
        !           360:        struct cpu_info *ci; \
        !           361:        for (CPU_INFO_FOREACH(cii, ci))
        !           362: #define MAXCPU MAXCPUS
        !           363: #define TRAPF_USERMODE(x) CLKF_USERMODE(x)
        !           364: #define TRAPF_PC(x) CLKF_PC(x)
        !           365: #endif
1.1       darran    366:
                    367: static kmem_cache_t *cyclic_id_cache;
                    368: static cyc_id_t *cyclic_id_head;
                    369: static cyc_backend_t cyclic_backend;
                    370:
                    371: MALLOC_DEFINE(M_CYCLIC, "cyclic", "Cyclic timer subsystem");
                    372:
                    373: /*
                    374:  * Returns 1 if the upheap propagated to the root, 0 if it did not.  This
                    375:  * allows the caller to reprogram the backend only when the root has been
                    376:  * modified.
                    377:  */
                    378: static int
                    379: cyclic_upheap(cyc_cpu_t *cpu, cyc_index_t ndx)
                    380: {
                    381:        cyclic_t *cyclics;
                    382:        cyc_index_t *heap;
                    383:        cyc_index_t heap_parent, heap_current = ndx;
                    384:        cyc_index_t parent, current;
                    385:
                    386:        if (heap_current == 0)
                    387:                return (1);
                    388:
                    389:        heap = cpu->cyp_heap;
                    390:        cyclics = cpu->cyp_cyclics;
                    391:        heap_parent = CYC_HEAP_PARENT(heap_current);
                    392:
                    393:        for (;;) {
                    394:                current = heap[heap_current];
                    395:                parent = heap[heap_parent];
                    396:
                    397:                /*
                    398:                 * We have an expiration time later than our parent; we're
                    399:                 * done.
                    400:                 */
                    401:                if (cyclics[current].cy_expire >= cyclics[parent].cy_expire)
                    402:                        return (0);
                    403:
                    404:                /*
                    405:                 * We need to swap with our parent, and continue up the heap.
                    406:                 */
                    407:                heap[heap_parent] = current;
                    408:                heap[heap_current] = parent;
                    409:
                    410:                /*
                    411:                 * If we just reached the root, we're done.
                    412:                 */
                    413:                if (heap_parent == 0)
                    414:                        return (1);
                    415:
                    416:                heap_current = heap_parent;
                    417:                heap_parent = CYC_HEAP_PARENT(heap_current);
                    418:        }
                    419: }
                    420:
                    421: static void
                    422: cyclic_downheap(cyc_cpu_t *cpu, cyc_index_t ndx)
                    423: {
                    424:        cyclic_t *cyclics = cpu->cyp_cyclics;
                    425:        cyc_index_t *heap = cpu->cyp_heap;
                    426:
                    427:        cyc_index_t heap_left, heap_right, heap_me = ndx;
                    428:        cyc_index_t left, right, me;
                    429:        cyc_index_t nelems = cpu->cyp_nelems;
                    430:
                    431:        for (;;) {
                    432:                /*
                    433:                 * If we don't have a left child (i.e., we're a leaf), we're
                    434:                 * done.
                    435:                 */
                    436:                if ((heap_left = CYC_HEAP_LEFT(heap_me)) >= nelems)
                    437:                        return;
                    438:
                    439:                left = heap[heap_left];
                    440:                me = heap[heap_me];
                    441:
                    442:                heap_right = CYC_HEAP_RIGHT(heap_me);
                    443:
                    444:                /*
                    445:                 * Even if we don't have a right child, we still need to compare
                    446:                 * our expiration time against that of our left child.
                    447:                 */
                    448:                if (heap_right >= nelems)
                    449:                        goto comp_left;
                    450:
                    451:                right = heap[heap_right];
                    452:
                    453:                /*
                    454:                 * We have both a left and a right child.  We need to compare
                    455:                 * the expiration times of the children to determine which
                    456:                 * expires earlier.
                    457:                 */
                    458:                if (cyclics[right].cy_expire < cyclics[left].cy_expire) {
                    459:                        /*
                    460:                         * Our right child is the earlier of our children.
                    461:                         * We'll now compare our expiration time to its; if
                    462:                         * ours is the earlier, we're done.
                    463:                         */
                    464:                        if (cyclics[me].cy_expire <= cyclics[right].cy_expire)
                    465:                                return;
                    466:
                    467:                        /*
                    468:                         * Our right child expires earlier than we do; swap
                    469:                         * with our right child, and descend right.
                    470:                         */
                    471:                        heap[heap_right] = me;
                    472:                        heap[heap_me] = right;
                    473:                        heap_me = heap_right;
                    474:                        continue;
                    475:                }
                    476:
                    477: comp_left:
                    478:                /*
                    479:                 * Our left child is the earlier of our children (or we have
                    480:                 * no right child).  We'll now compare our expiration time
                    481:                 * to its; if ours is the earlier, we're done.
                    482:                 */
                    483:                if (cyclics[me].cy_expire <= cyclics[left].cy_expire)
                    484:                        return;
                    485:
                    486:                /*
                    487:                 * Our left child expires earlier than we do; swap with our
                    488:                 * left child, and descend left.
                    489:                 */
                    490:                heap[heap_left] = me;
                    491:                heap[heap_me] = left;
                    492:                heap_me = heap_left;
                    493:        }
                    494: }
                    495:
                    496: static void
                    497: cyclic_expire(cyc_cpu_t *cpu, cyc_index_t ndx, cyclic_t *cyclic)
                    498: {
                    499:        cyc_func_t handler = cyclic->cy_handler;
                    500:        void *arg = cyclic->cy_arg;
                    501:
                    502:        (*handler)(arg);
                    503: }
                    504:
                    505: /*
                    506:  *  cyclic_fire(cpu_t *)
                    507:  *
                    508:  *  Overview
                    509:  *
                    510:  *    cyclic_fire() is the cyclic subsystem's interrupt handler.
                    511:  *    Called by the cyclic backend.
                    512:  *
                    513:  *  Arguments and notes
                    514:  *
                    515:  *    The only argument is the CPU on which the interrupt is executing;
                    516:  *    backends must call into cyclic_fire() on the specified CPU.
                    517:  *
                    518:  *    cyclic_fire() may be called spuriously without ill effect.  Optimal
                    519:  *    backends will call into cyclic_fire() at or shortly after the time
                    520:  *    requested via cyb_reprogram().  However, calling cyclic_fire()
                    521:  *    arbitrarily late will only manifest latency bubbles; the correctness
                    522:  *    of the cyclic subsystem does not rely on the timeliness of the backend.
                    523:  *
                    524:  *    cyclic_fire() is wait-free; it will not block or spin.
                    525:  *
                    526:  *  Return values
                    527:  *
                    528:  *    None.
                    529:  *
                    530:  */
                    531: static void
                    532: cyclic_fire(cpu_t *c)
                    533: {
                    534:        cyc_cpu_t *cpu = c->cpu_cyclic;
1.2.6.1 ! yamt      535:        cyc_backend_t *be = cpu->cyp_backend;
1.1       darran    536:        cyc_index_t *heap = cpu->cyp_heap;
                    537:        cyclic_t *cyclic, *cyclics = cpu->cyp_cyclics;
1.2.6.1 ! yamt      538:        void *arg = be->cyb_arg;
1.1       darran    539:        hrtime_t now = gethrtime();
                    540:        hrtime_t exp;
                    541:
                    542:        if (cpu->cyp_nelems == 0) {
                    543:                /* This is a spurious fire. */
                    544:                return;
                    545:        }
                    546:
                    547:        for (;;) {
                    548:                cyc_index_t ndx = heap[0];
                    549:
                    550:                cyclic = &cyclics[ndx];
                    551:
                    552:                ASSERT(!(cyclic->cy_flags & CYF_FREE));
                    553:
                    554:                if ((exp = cyclic->cy_expire) > now)
                    555:                        break;
                    556:
                    557:                cyclic_expire(cpu, ndx, cyclic);
                    558:
                    559:                /*
                    560:                 * If this cyclic will be set to next expire in the distant
                    561:                 * past, we have one of two situations:
                    562:                 *
                    563:                 *   a) This is the first firing of a cyclic which had
                    564:                 *      cy_expire set to 0.
                    565:                 *
                    566:                 *   b) We are tragically late for a cyclic -- most likely
                    567:                 *      due to being in the debugger.
                    568:                 *
                    569:                 * In either case, we set the new expiration time to be the
                    570:                 * the next interval boundary.  This assures that the
                    571:                 * expiration time modulo the interval is invariant.
                    572:                 *
                    573:                 * We arbitrarily define "distant" to be one second (one second
                    574:                 * is chosen because it's shorter than any foray to the
                    575:                 * debugger while still being longer than any legitimate
                    576:                 * stretch).
                    577:                 */
                    578:                exp += cyclic->cy_interval;
                    579:
                    580:                if (now - exp > NANOSEC) {
                    581:                        hrtime_t interval = cyclic->cy_interval;
                    582:
                    583:                        exp += ((now - exp) / interval + 1) * interval;
                    584:                }
                    585:
                    586:                cyclic->cy_expire = exp;
                    587:                cyclic_downheap(cpu, 0);
                    588:        }
                    589:
                    590:        /*
                    591:         * Now we have a cyclic in the root slot which isn't in the past;
                    592:         * reprogram the interrupt source.
                    593:         */
1.2.6.1 ! yamt      594:        be->cyb_reprogram(arg, exp);
        !           595: }
1.1       darran    596:
1.2.6.1 ! yamt      597: static void
        !           598: cyclic_expand_xcall(cyc_xcallarg_t *arg)
        !           599: {
        !           600:        cyc_cpu_t *cpu = arg->cyx_cpu;
        !           601:        cyc_index_t new_size = arg->cyx_size, size = cpu->cyp_size, i;
        !           602:        cyc_index_t *new_heap = arg->cyx_heap;
        !           603:        cyclic_t *cyclics = cpu->cyp_cyclics, *new_cyclics = arg->cyx_cyclics;
        !           604:
        !           605:        /* Disable preemption and interrupts. */
        !           606:        mtx_lock_spin(&cpu->cyp_mtx);
        !           607:
        !           608:        /*
        !           609:         * Assert that the new size is a power of 2.
        !           610:         */
        !           611:        ASSERT((new_size & (new_size - 1)) == 0);
        !           612:        ASSERT(new_size == (size << 1));
        !           613:        ASSERT(cpu->cyp_heap != NULL && cpu->cyp_cyclics != NULL);
        !           614:
        !           615:        bcopy(cpu->cyp_heap, new_heap, sizeof (cyc_index_t) * size);
        !           616:        bcopy(cyclics, new_cyclics, sizeof (cyclic_t) * size);
        !           617:
        !           618:        /*
        !           619:         * Set up the free list, and set all of the new cyclics to be CYF_FREE.
        !           620:         */
        !           621:        for (i = size; i < new_size; i++) {
        !           622:                new_heap[i] = i;
        !           623:                new_cyclics[i].cy_flags = CYF_FREE;
        !           624:        }
        !           625:
        !           626:        /*
        !           627:         * We can go ahead and plow the value of cyp_heap and cyp_cyclics;
        !           628:         * cyclic_expand() has kept a copy.
        !           629:         */
        !           630:        cpu->cyp_heap = new_heap;
        !           631:        cpu->cyp_cyclics = new_cyclics;
        !           632:        cpu->cyp_size = new_size;
1.1       darran    633:        mtx_unlock_spin(&cpu->cyp_mtx);
                    634: }
                    635:
                    636: /*
                    637:  * cyclic_expand() will cross call onto the CPU to perform the actual
                    638:  * expand operation.
                    639:  */
                    640: static void
                    641: cyclic_expand(cyc_cpu_t *cpu)
                    642: {
1.2.6.1 ! yamt      643:        cyc_index_t new_size, old_size;
1.1       darran    644:        cyc_index_t *new_heap, *old_heap;
                    645:        cyclic_t *new_cyclics, *old_cyclics;
1.2.6.1 ! yamt      646:        cyc_xcallarg_t arg;
        !           647:        cyc_backend_t *be = cpu->cyp_backend;
1.1       darran    648:
                    649:        ASSERT(MUTEX_HELD(&cpu_lock));
                    650:
1.2.6.1 ! yamt      651:        old_heap = cpu->cyp_heap;
        !           652:        old_cyclics = cpu->cyp_cyclics;
        !           653:
        !           654:        if ((new_size = ((old_size = cpu->cyp_size) << 1)) == 0) {
1.1       darran    655:                new_size = CY_DEFAULT_PERCPU;
1.2.6.1 ! yamt      656:                ASSERT(old_heap == NULL && old_cyclics == NULL);
        !           657:        }
1.1       darran    658:
                    659:        /*
                    660:         * Check that the new_size is a power of 2.
                    661:         */
                    662:        ASSERT(((new_size - 1) & new_size) == 0);
                    663:
                    664:        new_heap = malloc(sizeof(cyc_index_t) * new_size, M_CYCLIC, M_WAITOK);
                    665:        new_cyclics = malloc(sizeof(cyclic_t) * new_size, M_CYCLIC, M_ZERO | M_WAITOK);
                    666:
1.2.6.1 ! yamt      667:        arg.cyx_cpu = cpu;
        !           668:        arg.cyx_heap = new_heap;
        !           669:        arg.cyx_cyclics = new_cyclics;
        !           670:        arg.cyx_size = new_size;
1.1       darran    671:
1.2.6.1 ! yamt      672:        be->cyb_xcall(be->cyb_arg, cpu->cyp_cpu,
        !           673:            (cyc_func_t)cyclic_expand_xcall, &arg);
1.1       darran    674:
                    675:        if (old_cyclics != NULL) {
                    676:                ASSERT(old_heap != NULL);
                    677:                ASSERT(old_size != 0);
                    678:                free(old_cyclics, M_CYCLIC);
                    679:                free(old_heap, M_CYCLIC);
                    680:        }
                    681: }
                    682:
1.2.6.1 ! yamt      683: static void
        !           684: cyclic_add_xcall(cyc_xcallarg_t *arg)
1.1       darran    685: {
1.2.6.1 ! yamt      686:        cyc_cpu_t *cpu = arg->cyx_cpu;
        !           687:        cyc_handler_t *hdlr = arg->cyx_hdlr;
        !           688:        cyc_time_t *when = arg->cyx_when;
        !           689:        cyc_backend_t *be = cpu->cyp_backend;
1.1       darran    690:        cyc_index_t ndx, nelems;
1.2.6.1 ! yamt      691:        cyb_arg_t bar = be->cyb_arg;
1.1       darran    692:        cyclic_t *cyclic;
                    693:
                    694:        ASSERT(cpu->cyp_nelems < cpu->cyp_size);
                    695:
1.2.6.1 ! yamt      696:        /* Disable preemption and interrupts. */
        !           697:        mtx_lock_spin(&cpu->cyp_mtx);
1.1       darran    698:        nelems = cpu->cyp_nelems++;
                    699:
1.2.6.1 ! yamt      700:        if (nelems == 0) {
1.1       darran    701:                /*
                    702:                 * If this is the first element, we need to enable the
                    703:                 * backend on this CPU.
                    704:                 */
1.2.6.1 ! yamt      705:                be->cyb_enable(bar);
        !           706:        }
1.1       darran    707:
                    708:        ndx = cpu->cyp_heap[nelems];
                    709:        cyclic = &cpu->cyp_cyclics[ndx];
                    710:
                    711:        ASSERT(cyclic->cy_flags == CYF_FREE);
                    712:        cyclic->cy_interval = when->cyt_interval;
                    713:
1.2.6.1 ! yamt      714:        if (when->cyt_when == 0) {
        !           715:                /*
        !           716:                 * If a start time hasn't been explicitly specified, we'll
        !           717:                 * start on the next interval boundary.
        !           718:                 */
        !           719:                cyclic->cy_expire = (gethrtime() / cyclic->cy_interval + 1) *
        !           720:                    cyclic->cy_interval;
        !           721:        } else {
1.1       darran    722:                cyclic->cy_expire = when->cyt_when;
1.2.6.1 ! yamt      723:        }
1.1       darran    724:
                    725:        cyclic->cy_handler = hdlr->cyh_func;
                    726:        cyclic->cy_arg = hdlr->cyh_arg;
1.2.6.1 ! yamt      727:        cyclic->cy_flags = arg->cyx_flags;
1.1       darran    728:
                    729:        if (cyclic_upheap(cpu, nelems)) {
                    730:                hrtime_t exp = cyclic->cy_expire;
                    731:
                    732:                /*
                    733:                 * If our upheap propagated to the root, we need to
                    734:                 * reprogram the interrupt source.
                    735:                 */
1.2.6.1 ! yamt      736:                be->cyb_reprogram(bar, exp);
1.1       darran    737:        }
                    738:        mtx_unlock_spin(&cpu->cyp_mtx);
                    739:
1.2.6.1 ! yamt      740:        arg->cyx_ndx = ndx;
1.1       darran    741: }
                    742:
1.2.6.1 ! yamt      743: static cyc_index_t
        !           744: cyclic_add_here(cyc_cpu_t *cpu, cyc_handler_t *hdlr,
        !           745:     cyc_time_t *when, uint16_t flags)
1.1       darran    746: {
1.2.6.1 ! yamt      747:        cyc_backend_t *be = cpu->cyp_backend;
        !           748:        cyb_arg_t bar = be->cyb_arg;
        !           749:        cyc_xcallarg_t arg;
1.1       darran    750:
                    751:        ASSERT(MUTEX_HELD(&cpu_lock));
1.2.6.1 ! yamt      752:        ASSERT(!(cpu->cyp_cpu->cpu_flags & CPU_OFFLINE));
        !           753:        ASSERT(when->cyt_when >= 0 && when->cyt_interval > 0);
1.1       darran    754:
1.2.6.1 ! yamt      755:        if (cpu->cyp_nelems == cpu->cyp_size) {
        !           756:                /*
        !           757:                 * This is expensive; it will cross call onto the other
        !           758:                 * CPU to perform the expansion.
        !           759:                 */
        !           760:                cyclic_expand(cpu);
        !           761:                ASSERT(cpu->cyp_nelems < cpu->cyp_size);
        !           762:        }
1.1       darran    763:
1.2.6.1 ! yamt      764:        /*
        !           765:         * By now, we know that we're going to be able to successfully
        !           766:         * perform the add.  Now cross call over to the CPU of interest to
        !           767:         * actually add our cyclic.
        !           768:         */
        !           769:        arg.cyx_cpu = cpu;
        !           770:        arg.cyx_hdlr = hdlr;
        !           771:        arg.cyx_when = when;
        !           772:        arg.cyx_flags = flags;
1.1       darran    773:
1.2.6.1 ! yamt      774:        be->cyb_xcall(bar, cpu->cyp_cpu, (cyc_func_t)cyclic_add_xcall, &arg);
        !           775:
        !           776:        return (arg.cyx_ndx);
        !           777: }
1.1       darran    778:
1.2.6.1 ! yamt      779: static void
        !           780: cyclic_remove_xcall(cyc_xcallarg_t *arg)
        !           781: {
        !           782:        cyc_cpu_t *cpu = arg->cyx_cpu;
        !           783:        cyc_backend_t *be = cpu->cyp_backend;
        !           784:        cyb_arg_t bar = be->cyb_arg;
        !           785:        cyc_index_t ndx = arg->cyx_ndx, nelems = cpu->cyp_nelems, i;
        !           786:        cyc_index_t *heap = cpu->cyp_heap, last;
        !           787:        cyclic_t *cyclic;
        !           788:
        !           789:        ASSERT(nelems > 0);
        !           790:
        !           791:        /* Disable preemption and interrupts. */
        !           792:        mtx_lock_spin(&cpu->cyp_mtx);
1.1       darran    793:        cyclic = &cpu->cyp_cyclics[ndx];
                    794:
                    795:        /*
                    796:         * Grab the current expiration time.  If this cyclic is being
                    797:         * removed as part of a juggling operation, the expiration time
                    798:         * will be used when the cyclic is added to the new CPU.
                    799:         */
1.2.6.1 ! yamt      800:        if (arg->cyx_when != NULL) {
        !           801:                arg->cyx_when->cyt_when = cyclic->cy_expire;
        !           802:                arg->cyx_when->cyt_interval = cyclic->cy_interval;
1.1       darran    803:        }
                    804:
1.2.6.1 ! yamt      805:        /*
        !           806:         * Now set the flags to CYF_FREE.  We don't need a membar_enter()
        !           807:         * between zeroing pend and setting the flags because we're at
        !           808:         * CY_HIGH_LEVEL (that is, the zeroing of pend and the setting
        !           809:         * of cy_flags appear atomic to softints).
        !           810:         */
1.1       darran    811:        cyclic->cy_flags = CYF_FREE;
                    812:
                    813:        for (i = 0; i < nelems; i++) {
                    814:                if (heap[i] == ndx)
                    815:                        break;
                    816:        }
                    817:
                    818:        if (i == nelems)
                    819:                panic("attempt to remove non-existent cyclic");
                    820:
                    821:        cpu->cyp_nelems = --nelems;
                    822:
1.2.6.1 ! yamt      823:        if (nelems == 0) {
1.1       darran    824:                /*
                    825:                 * If we just removed the last element, then we need to
                    826:                 * disable the backend on this CPU.
                    827:                 */
1.2.6.1 ! yamt      828:                be->cyb_disable(bar);
        !           829:        }
1.1       darran    830:
1.2.6.1 ! yamt      831:        if (i == nelems) {
1.1       darran    832:                /*
                    833:                 * If we just removed the last element of the heap, then
                    834:                 * we don't have to downheap.
                    835:                 */
1.2.6.1 ! yamt      836:                goto out;
        !           837:        }
1.1       darran    838:
                    839:        /*
                    840:         * Swap the last element of the heap with the one we want to
                    841:         * remove, and downheap (this has the implicit effect of putting
                    842:         * the newly freed element on the free list).
                    843:         */
                    844:        heap[i] = (last = heap[nelems]);
                    845:        heap[nelems] = ndx;
                    846:
1.2.6.1 ! yamt      847:        if (i == 0) {
1.1       darran    848:                cyclic_downheap(cpu, 0);
1.2.6.1 ! yamt      849:        } else {
1.1       darran    850:                if (cyclic_upheap(cpu, i) == 0) {
                    851:                        /*
                    852:                         * The upheap didn't propagate to the root; if it
                    853:                         * didn't propagate at all, we need to downheap.
                    854:                         */
1.2.6.1 ! yamt      855:                        if (heap[i] == last) {
1.1       darran    856:                                cyclic_downheap(cpu, i);
1.2.6.1 ! yamt      857:                        }
        !           858:                        goto out;
1.1       darran    859:                }
                    860:        }
                    861:
                    862:        /*
                    863:         * We're here because we changed the root; we need to reprogram
                    864:         * the clock source.
                    865:         */
                    866:        cyclic = &cpu->cyp_cyclics[heap[0]];
                    867:
                    868:        ASSERT(nelems != 0);
1.2.6.1 ! yamt      869:        be->cyb_reprogram(bar, cyclic->cy_expire);
        !           870: out:
1.1       darran    871:        mtx_unlock_spin(&cpu->cyp_mtx);
1.2.6.1 ! yamt      872: }
        !           873:
        !           874: static int
        !           875: cyclic_remove_here(cyc_cpu_t *cpu, cyc_index_t ndx, cyc_time_t *when, int wait)
        !           876: {
        !           877:        cyc_backend_t *be = cpu->cyp_backend;
        !           878:        cyc_xcallarg_t arg;
        !           879:
        !           880:        ASSERT(MUTEX_HELD(&cpu_lock));
        !           881:        ASSERT(wait == CY_WAIT || wait == CY_NOWAIT);
        !           882:
        !           883:        arg.cyx_ndx = ndx;
        !           884:        arg.cyx_cpu = cpu;
        !           885:        arg.cyx_when = when;
        !           886:        arg.cyx_wait = wait;
        !           887:
        !           888:        be->cyb_xcall(be->cyb_arg, cpu->cyp_cpu,
        !           889:            (cyc_func_t)cyclic_remove_xcall, &arg);
1.1       darran    890:
                    891:        return (1);
                    892: }
                    893:
                    894: static void
                    895: cyclic_configure(cpu_t *c)
                    896: {
                    897:        cyc_cpu_t *cpu = malloc(sizeof(cyc_cpu_t), M_CYCLIC, M_ZERO | M_WAITOK);
                    898:        cyc_backend_t *nbe = malloc(sizeof(cyc_backend_t), M_CYCLIC, M_ZERO | M_WAITOK);
                    899:
                    900:        ASSERT(MUTEX_HELD(&cpu_lock));
                    901:
                    902:        if (cyclic_id_cache == NULL)
1.2.6.1 ! yamt      903:                cyclic_id_cache = kmem_cache_create(__UNCONST("cyclic_id_cache"),
1.1       darran    904:                    sizeof (cyc_id_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
                    905:
                    906:        cpu->cyp_cpu = c;
                    907:
                    908:        cpu->cyp_size = 1;
                    909:        cpu->cyp_heap = malloc(sizeof(cyc_index_t), M_CYCLIC, M_ZERO | M_WAITOK);
                    910:        cpu->cyp_cyclics = malloc(sizeof(cyclic_t), M_CYCLIC, M_ZERO | M_WAITOK);
                    911:        cpu->cyp_cyclics->cy_flags = CYF_FREE;
                    912:
                    913:        mtx_init(&cpu->cyp_mtx, "cyclic cpu", NULL, MTX_SPIN);
                    914:
                    915:        /*
                    916:         * Setup the backend for this CPU.
                    917:         */
                    918:        bcopy(&cyclic_backend, nbe, sizeof (cyc_backend_t));
                    919:        if (nbe->cyb_configure != NULL)
                    920:                nbe->cyb_arg = nbe->cyb_configure(c);
                    921:        cpu->cyp_backend = nbe;
                    922:
                    923:        /*
                    924:         * On platforms where stray interrupts may be taken during startup,
                    925:         * the CPU's cpu_cyclic pointer serves as an indicator that the
                    926:         * cyclic subsystem for this CPU is prepared to field interrupts.
                    927:         */
                    928:        membar_producer();
                    929:
                    930:        c->cpu_cyclic = cpu;
                    931: }
                    932:
                    933: static void
                    934: cyclic_unconfigure(cpu_t *c)
                    935: {
                    936:        cyc_cpu_t *cpu = c->cpu_cyclic;
                    937:        cyc_backend_t *be = cpu->cyp_backend;
                    938:        cyb_arg_t bar = be->cyb_arg;
                    939:
                    940:        ASSERT(MUTEX_HELD(&cpu_lock));
                    941:
                    942:        c->cpu_cyclic = NULL;
                    943:
                    944:        /*
                    945:         * Let the backend know that the CPU is being yanked, and free up
                    946:         * the backend structure.
                    947:         */
                    948:        if (be->cyb_unconfigure != NULL)
                    949:                be->cyb_unconfigure(bar);
                    950:        free(be, M_CYCLIC);
                    951:        cpu->cyp_backend = NULL;
                    952:
                    953:        mtx_destroy(&cpu->cyp_mtx);
                    954:
                    955:        /* Finally, clean up our remaining dynamic structures. */
                    956:        free(cpu->cyp_cyclics, M_CYCLIC);
                    957:        free(cpu->cyp_heap, M_CYCLIC);
                    958:        free(cpu, M_CYCLIC);
                    959: }
                    960:
                    961: static void
                    962: cyclic_omni_start(cyc_id_t *idp, cyc_cpu_t *cpu)
                    963: {
                    964:        cyc_omni_handler_t *omni = &idp->cyi_omni_hdlr;
                    965:        cyc_omni_cpu_t *ocpu = malloc(sizeof(cyc_omni_cpu_t), M_CYCLIC , M_WAITOK);
                    966:        cyc_handler_t hdlr;
                    967:        cyc_time_t when;
                    968:
                    969:        ASSERT(MUTEX_HELD(&cpu_lock));
                    970:        ASSERT(idp->cyi_cpu == NULL);
                    971:
                    972:        hdlr.cyh_func = NULL;
                    973:        hdlr.cyh_arg = NULL;
                    974:
                    975:        when.cyt_when = 0;
                    976:        when.cyt_interval = 0;
                    977:
                    978:        omni->cyo_online(omni->cyo_arg, cpu->cyp_cpu, &hdlr, &when);
                    979:
                    980:        ASSERT(hdlr.cyh_func != NULL);
                    981:        ASSERT(when.cyt_when >= 0 && when.cyt_interval > 0);
                    982:
                    983:        ocpu->cyo_cpu = cpu;
                    984:        ocpu->cyo_arg = hdlr.cyh_arg;
                    985:        ocpu->cyo_ndx = cyclic_add_here(cpu, &hdlr, &when, 0);
                    986:        ocpu->cyo_next = idp->cyi_omni_list;
                    987:        idp->cyi_omni_list = ocpu;
                    988: }
                    989:
                    990: static void
                    991: cyclic_omni_stop(cyc_id_t *idp, cyc_cpu_t *cpu)
                    992: {
                    993:        cyc_omni_handler_t *omni = &idp->cyi_omni_hdlr;
                    994:        cyc_omni_cpu_t *ocpu = idp->cyi_omni_list, *prev = NULL;
                    995:
                    996:        ASSERT(MUTEX_HELD(&cpu_lock));
                    997:        ASSERT(idp->cyi_cpu == NULL);
                    998:        ASSERT(ocpu != NULL);
                    999:
                   1000:        while (ocpu != NULL && ocpu->cyo_cpu != cpu) {
                   1001:                prev = ocpu;
                   1002:                ocpu = ocpu->cyo_next;
                   1003:        }
                   1004:
                   1005:        /*
                   1006:         * We _must_ have found an cyc_omni_cpu which corresponds to this
                   1007:         * CPU -- the definition of an omnipresent cyclic is that it runs
                   1008:         * on all online CPUs.
                   1009:         */
                   1010:        ASSERT(ocpu != NULL);
                   1011:
                   1012:        if (prev == NULL) {
                   1013:                idp->cyi_omni_list = ocpu->cyo_next;
                   1014:        } else {
                   1015:                prev->cyo_next = ocpu->cyo_next;
                   1016:        }
                   1017:
                   1018:        (void) cyclic_remove_here(ocpu->cyo_cpu, ocpu->cyo_ndx, NULL, CY_WAIT);
                   1019:
                   1020:        /*
                   1021:         * The cyclic has been removed from this CPU; time to call the
                   1022:         * omnipresent offline handler.
                   1023:         */
                   1024:        if (omni->cyo_offline != NULL)
                   1025:                omni->cyo_offline(omni->cyo_arg, cpu->cyp_cpu, ocpu->cyo_arg);
                   1026:
                   1027:        free(ocpu, M_CYCLIC);
                   1028: }
                   1029:
                   1030: static cyc_id_t *
                   1031: cyclic_new_id(void)
                   1032: {
                   1033:        cyc_id_t *idp;
                   1034:
                   1035:        ASSERT(MUTEX_HELD(&cpu_lock));
                   1036:
                   1037:        idp = kmem_cache_alloc(cyclic_id_cache, KM_SLEEP);
                   1038:
                   1039:        /*
                   1040:         * The cyi_cpu field of the cyc_id_t structure tracks the CPU
                   1041:         * associated with the cyclic.  If and only if this field is NULL, the
                   1042:         * cyc_id_t is an omnipresent cyclic.  Note that cyi_omni_list may be
                   1043:         * NULL for an omnipresent cyclic while the cyclic is being created
                   1044:         * or destroyed.
                   1045:         */
                   1046:        idp->cyi_cpu = NULL;
                   1047:        idp->cyi_ndx = 0;
                   1048:
                   1049:        idp->cyi_next = cyclic_id_head;
                   1050:        idp->cyi_prev = NULL;
                   1051:        idp->cyi_omni_list = NULL;
                   1052:
                   1053:        if (cyclic_id_head != NULL) {
                   1054:                ASSERT(cyclic_id_head->cyi_prev == NULL);
                   1055:                cyclic_id_head->cyi_prev = idp;
                   1056:        }
                   1057:
                   1058:        cyclic_id_head = idp;
                   1059:
                   1060:        return (idp);
                   1061: }
                   1062:
                   1063: /*
                   1064:  *  cyclic_id_t cyclic_add(cyc_handler_t *, cyc_time_t *)
                   1065:  *
                   1066:  *  Overview
                   1067:  *
                   1068:  *    cyclic_add() will create an unbound cyclic with the specified handler and
                   1069:  *    interval.  The cyclic will run on a CPU which both has interrupts enabled
                   1070:  *    and is in the system CPU partition.
                   1071:  *
                   1072:  *  Arguments and notes
                   1073:  *
                   1074:  *    As its first argument, cyclic_add() takes a cyc_handler, which has the
                   1075:  *    following members:
                   1076:  *
                   1077:  *      cyc_func_t cyh_func    <-- Cyclic handler
                   1078:  *      void *cyh_arg          <-- Argument to cyclic handler
                   1079:  *
                   1080:  *    In addition to a cyc_handler, cyclic_add() takes a cyc_time, which
                   1081:  *    has the following members:
                   1082:  *
                   1083:  *       hrtime_t cyt_when     <-- Absolute time, in nanoseconds since boot, at
                   1084:  *                                 which to start firing
                   1085:  *       hrtime_t cyt_interval <-- Length of interval, in nanoseconds
                   1086:  *
                   1087:  *    gethrtime() is the time source for nanoseconds since boot.  If cyt_when
                   1088:  *    is set to 0, the cyclic will start to fire when cyt_interval next
                   1089:  *    divides the number of nanoseconds since boot.
                   1090:  *
                   1091:  *    The cyt_interval field _must_ be filled in by the caller; one-shots are
                   1092:  *    _not_ explicitly supported by the cyclic subsystem (cyclic_add() will
                   1093:  *    assert that cyt_interval is non-zero).  The maximum value for either
                   1094:  *    field is INT64_MAX; the caller is responsible for assuring that
                   1095:  *    cyt_when + cyt_interval <= INT64_MAX.  Neither field may be negative.
                   1096:  *
                   1097:  *    For an arbitrary time t in the future, the cyclic handler is guaranteed
                   1098:  *    to have been called (t - cyt_when) / cyt_interval times.  This will
                   1099:  *    be true even if interrupts have been disabled for periods greater than
                   1100:  *    cyt_interval nanoseconds.  In order to compensate for such periods,
                   1101:  *    the cyclic handler may be called a finite number of times with an
                   1102:  *    arbitrarily small interval.
                   1103:  *
                   1104:  *    The cyclic subsystem will not enforce any lower bound on the interval;
                   1105:  *    if the interval is less than the time required to process an interrupt,
                   1106:  *    the CPU will wedge.  It's the responsibility of the caller to assure that
                   1107:  *    either the value of the interval is sane, or that its caller has
                   1108:  *    sufficient privilege to deny service (i.e. its caller is root).
                   1109:  *
                   1110:  *  Return value
                   1111:  *
                   1112:  *    cyclic_add() returns a cyclic_id_t, which is guaranteed to be a value
                   1113:  *    other than CYCLIC_NONE.  cyclic_add() cannot fail.
                   1114:  *
                   1115:  *  Caller's context
                   1116:  *
                   1117:  *    cpu_lock must be held by the caller, and the caller must not be in
                   1118:  *    interrupt context.  cyclic_add() will perform a KM_SLEEP kernel
                   1119:  *    memory allocation, so the usual rules (e.g. p_lock cannot be held)
                   1120:  *    apply.  A cyclic may be added even in the presence of CPUs that have
                   1121:  *    not been configured with respect to the cyclic subsystem, but only
                   1122:  *    configured CPUs will be eligible to run the new cyclic.
                   1123:  *
                   1124:  *  Cyclic handler's context
                   1125:  *
                   1126:  *    Cyclic handlers will be executed in the interrupt context corresponding
                   1127:  *    to the specified level (i.e. either high, lock or low level).  The
                   1128:  *    usual context rules apply.
                   1129:  *
                   1130:  *    A cyclic handler may not grab ANY locks held by the caller of any of
                   1131:  *    cyclic_add() or cyclic_remove(); the implementation of these functions
                   1132:  *    may require blocking on cyclic handler completion.
                   1133:  *    Moreover, cyclic handlers may not make any call back into the cyclic
                   1134:  *    subsystem.
                   1135:  */
                   1136: cyclic_id_t
                   1137: cyclic_add(cyc_handler_t *hdlr, cyc_time_t *when)
                   1138: {
                   1139:        cyc_id_t *idp = cyclic_new_id();
1.2.6.1 ! yamt     1140:        solaris_cpu_t *c = &solaris_cpu[cpu_number()];
1.1       darran   1141:
                   1142:        ASSERT(MUTEX_HELD(&cpu_lock));
                   1143:        ASSERT(when->cyt_when >= 0 && when->cyt_interval > 0);
                   1144:
                   1145:        idp->cyi_cpu = c->cpu_cyclic;
                   1146:        idp->cyi_ndx = cyclic_add_here(idp->cyi_cpu, hdlr, when, 0);
                   1147:
                   1148:        return ((uintptr_t)idp);
                   1149: }
                   1150:
                   1151: /*
                   1152:  *  cyclic_id_t cyclic_add_omni(cyc_omni_handler_t *)
                   1153:  *
                   1154:  *  Overview
                   1155:  *
                   1156:  *    cyclic_add_omni() will create an omnipresent cyclic with the specified
                   1157:  *    online and offline handlers.  Omnipresent cyclics run on all online
                   1158:  *    CPUs, including CPUs which have unbound interrupts disabled.
                   1159:  *
                   1160:  *  Arguments
                   1161:  *
                   1162:  *    As its only argument, cyclic_add_omni() takes a cyc_omni_handler, which
                   1163:  *    has the following members:
                   1164:  *
                   1165:  *      void (*cyo_online)()   <-- Online handler
                   1166:  *      void (*cyo_offline)()  <-- Offline handler
                   1167:  *      void *cyo_arg          <-- Argument to be passed to on/offline handlers
                   1168:  *
                   1169:  *  Online handler
                   1170:  *
                   1171:  *    The cyo_online member is a pointer to a function which has the following
                   1172:  *    four arguments:
                   1173:  *
                   1174:  *      void *                 <-- Argument (cyo_arg)
                   1175:  *      cpu_t *                <-- Pointer to CPU about to be onlined
                   1176:  *      cyc_handler_t *        <-- Pointer to cyc_handler_t; must be filled in
                   1177:  *                                 by omni online handler
                   1178:  *      cyc_time_t *           <-- Pointer to cyc_time_t; must be filled in by
                   1179:  *                                 omni online handler
                   1180:  *
                   1181:  *    The omni cyclic online handler is always called _before_ the omni
                   1182:  *    cyclic begins to fire on the specified CPU.  As the above argument
                   1183:  *    description implies, the online handler must fill in the two structures
                   1184:  *    passed to it:  the cyc_handler_t and the cyc_time_t.  These are the
                   1185:  *    same two structures passed to cyclic_add(), outlined above.  This
                   1186:  *    allows the omni cyclic to have maximum flexibility; different CPUs may
                   1187:  *    optionally
                   1188:  *
                   1189:  *      (a)  have different intervals
                   1190:  *      (b)  be explicitly in or out of phase with one another
                   1191:  *      (c)  have different handlers
                   1192:  *      (d)  have different handler arguments
                   1193:  *      (e)  fire at different levels
                   1194:  *
                   1195:  *    Of these, (e) seems somewhat dubious, but is nonetheless allowed.
                   1196:  *
                   1197:  *    The omni online handler is called in the same context as cyclic_add(),
                   1198:  *    and has the same liberties:  omni online handlers may perform KM_SLEEP
                   1199:  *    kernel memory allocations, and may grab locks which are also acquired
                   1200:  *    by cyclic handlers.  However, omni cyclic online handlers may _not_
                   1201:  *    call back into the cyclic subsystem, and should be generally careful
                   1202:  *    about calling into arbitrary kernel subsystems.
                   1203:  *
                   1204:  *  Offline handler
                   1205:  *
                   1206:  *    The cyo_offline member is a pointer to a function which has the following
                   1207:  *    three arguments:
                   1208:  *
                   1209:  *      void *                 <-- Argument (cyo_arg)
                   1210:  *      cpu_t *                <-- Pointer to CPU about to be offlined
                   1211:  *      void *                 <-- CPU's cyclic argument (that is, value
                   1212:  *                                 to which cyh_arg member of the cyc_handler_t
                   1213:  *                                 was set in the omni online handler)
                   1214:  *
                   1215:  *    The omni cyclic offline handler is always called _after_ the omni
                   1216:  *    cyclic has ceased firing on the specified CPU.  Its purpose is to
                   1217:  *    allow cleanup of any resources dynamically allocated in the omni cyclic
                   1218:  *    online handler.  The context of the offline handler is identical to
                   1219:  *    that of the online handler; the same constraints and liberties apply.
                   1220:  *
                   1221:  *    The offline handler is optional; it may be NULL.
                   1222:  *
                   1223:  *  Return value
                   1224:  *
                   1225:  *    cyclic_add_omni() returns a cyclic_id_t, which is guaranteed to be a
                   1226:  *    value other than CYCLIC_NONE.  cyclic_add_omni() cannot fail.
                   1227:  *
                   1228:  *  Caller's context
                   1229:  *
                   1230:  *    The caller's context is identical to that of cyclic_add(), specified
                   1231:  *    above.
                   1232:  */
                   1233: cyclic_id_t
                   1234: cyclic_add_omni(cyc_omni_handler_t *omni)
                   1235: {
                   1236:        cyc_id_t *idp = cyclic_new_id();
                   1237:        cyc_cpu_t *cpu;
                   1238:        cpu_t *c;
                   1239:        int i;
                   1240:
                   1241:        ASSERT(MUTEX_HELD(&cpu_lock));
                   1242:        ASSERT(omni != NULL && omni->cyo_online != NULL);
                   1243:
                   1244:        idp->cyi_omni_hdlr = *omni;
                   1245:
1.2.6.1 ! yamt     1246:        CPU_FOREACH(i) {
        !          1247:                i = cpu_index(ci);
1.1       darran   1248:                c = &solaris_cpu[i];
                   1249:                if ((cpu = c->cpu_cyclic) == NULL)
                   1250:                        continue;
                   1251:                cyclic_omni_start(idp, cpu);
                   1252:        }
                   1253:
                   1254:        /*
                   1255:         * We must have found at least one online CPU on which to run
                   1256:         * this cyclic.
                   1257:         */
                   1258:        ASSERT(idp->cyi_omni_list != NULL);
                   1259:        ASSERT(idp->cyi_cpu == NULL);
                   1260:
                   1261:        return ((uintptr_t)idp);
                   1262: }
                   1263:
                   1264: /*
                   1265:  *  void cyclic_remove(cyclic_id_t)
                   1266:  *
                   1267:  *  Overview
                   1268:  *
                   1269:  *    cyclic_remove() will remove the specified cyclic from the system.
                   1270:  *
                   1271:  *  Arguments and notes
                   1272:  *
                   1273:  *    The only argument is a cyclic_id returned from either cyclic_add() or
                   1274:  *    cyclic_add_omni().
                   1275:  *
                   1276:  *    By the time cyclic_remove() returns, the caller is guaranteed that the
                   1277:  *    removed cyclic handler has completed execution (this is the same
                   1278:  *    semantic that untimeout() provides).  As a result, cyclic_remove() may
                   1279:  *    need to block, waiting for the removed cyclic to complete execution.
                   1280:  *    This leads to an important constraint on the caller:  no lock may be
                   1281:  *    held across cyclic_remove() that also may be acquired by a cyclic
                   1282:  *    handler.
                   1283:  *
                   1284:  *  Return value
                   1285:  *
                   1286:  *    None; cyclic_remove() always succeeds.
                   1287:  *
                   1288:  *  Caller's context
                   1289:  *
                   1290:  *    cpu_lock must be held by the caller, and the caller must not be in
                   1291:  *    interrupt context.  The caller may not hold any locks which are also
                   1292:  *    grabbed by any cyclic handler.  See "Arguments and notes", above.
                   1293:  */
                   1294: void
                   1295: cyclic_remove(cyclic_id_t id)
                   1296: {
                   1297:        cyc_id_t *idp = (cyc_id_t *)id;
                   1298:        cyc_id_t *prev = idp->cyi_prev, *next = idp->cyi_next;
                   1299:        cyc_cpu_t *cpu = idp->cyi_cpu;
                   1300:
                   1301:        ASSERT(MUTEX_HELD(&cpu_lock));
                   1302:
                   1303:        if (cpu != NULL) {
                   1304:                (void) cyclic_remove_here(cpu, idp->cyi_ndx, NULL, CY_WAIT);
                   1305:        } else {
                   1306:                ASSERT(idp->cyi_omni_list != NULL);
                   1307:                while (idp->cyi_omni_list != NULL)
                   1308:                        cyclic_omni_stop(idp, idp->cyi_omni_list->cyo_cpu);
                   1309:        }
                   1310:
                   1311:        if (prev != NULL) {
                   1312:                ASSERT(cyclic_id_head != idp);
                   1313:                prev->cyi_next = next;
                   1314:        } else {
                   1315:                ASSERT(cyclic_id_head == idp);
                   1316:                cyclic_id_head = next;
                   1317:        }
                   1318:
                   1319:        if (next != NULL)
                   1320:                next->cyi_prev = prev;
                   1321:
                   1322:        kmem_cache_free(cyclic_id_cache, idp);
                   1323: }
                   1324:
                   1325: static void
                   1326: cyclic_init(cyc_backend_t *be)
                   1327: {
                   1328:        ASSERT(MUTEX_HELD(&cpu_lock));
                   1329:
                   1330:        /*
                   1331:         * Copy the passed cyc_backend into the backend template.  This must
                   1332:         * be done before the CPU can be configured.
                   1333:         */
                   1334:        bcopy(be, &cyclic_backend, sizeof (cyc_backend_t));
                   1335:
1.2.6.1 ! yamt     1336:        cyclic_configure(&solaris_cpu[cpu_number()]);
1.1       darran   1337: }
                   1338:
                   1339: /*
                   1340:  * It is assumed that cyclic_mp_init() is called some time after cyclic
                   1341:  * init (and therefore, after cpu0 has been initialized).  We grab cpu_lock,
                   1342:  * find the already initialized CPU, and initialize every other CPU with the
                   1343:  * same backend.
                   1344:  */
                   1345: static void
                   1346: cyclic_mp_init(void)
                   1347: {
                   1348:        cpu_t *c;
                   1349:        int i;
                   1350:
1.2.6.1 ! yamt     1351: #ifndef __NetBSD__
1.1       darran   1352:        mutex_enter(&cpu_lock);
1.2.6.1 ! yamt     1353: #endif
1.1       darran   1354:
1.2.6.1 ! yamt     1355:        CPU_FOREACH(i) {
        !          1356:                i = cpu_index(ci);
1.1       darran   1357:                c = &solaris_cpu[i];
                   1358:                if (c->cpu_cyclic == NULL)
                   1359:                        cyclic_configure(c);
                   1360:        }
                   1361:
1.2.6.1 ! yamt     1362: #ifndef __NetBSD__
1.1       darran   1363:        mutex_exit(&cpu_lock);
1.2.6.1 ! yamt     1364: #endif
1.1       darran   1365: }
                   1366:
                   1367: static void
                   1368: cyclic_uninit(void)
                   1369: {
                   1370:        cpu_t *c;
                   1371:        int id;
                   1372:
1.2.6.1 ! yamt     1373:        CPU_FOREACH(id) {
        !          1374:                id = cpu_index(ci);
1.1       darran   1375:                c = &solaris_cpu[id];
                   1376:                if (c->cpu_cyclic == NULL)
                   1377:                        continue;
                   1378:                cyclic_unconfigure(c);
                   1379:        }
                   1380:
                   1381:        if (cyclic_id_cache != NULL)
                   1382:                kmem_cache_destroy(cyclic_id_cache);
                   1383: }
                   1384:
                   1385: #include "cyclic_machdep.c"
                   1386:
                   1387: /*
                   1388:  *  Cyclic subsystem initialisation.
                   1389:  */
                   1390: static void
                   1391: cyclic_load(void *dummy)
                   1392: {
                   1393:        mutex_enter(&cpu_lock);
                   1394:
                   1395:        /* Initialise the machine-dependent backend. */
                   1396:        cyclic_machdep_init();
                   1397:
                   1398:        mutex_exit(&cpu_lock);
                   1399: }
                   1400:
                   1401: SYSINIT(cyclic_register, SI_SUB_CYCLIC, SI_ORDER_SECOND, cyclic_load, NULL);
                   1402:
                   1403: static void
                   1404: cyclic_unload(void)
                   1405: {
                   1406:        mutex_enter(&cpu_lock);
                   1407:
                   1408:        /* Uninitialise the machine-dependent backend. */
                   1409:        cyclic_machdep_uninit();
                   1410:
                   1411:        mutex_exit(&cpu_lock);
                   1412: }
                   1413:
                   1414: SYSUNINIT(cyclic_unregister, SI_SUB_CYCLIC, SI_ORDER_SECOND, cyclic_unload, NULL);
                   1415:
1.2.6.1 ! yamt     1416: #ifdef __FreeBSD__
1.1       darran   1417: /* ARGSUSED */
                   1418: static int
                   1419: cyclic_modevent(module_t mod __unused, int type, void *data __unused)
                   1420: {
                   1421:        int error = 0;
                   1422:
                   1423:        switch (type) {
                   1424:        case MOD_LOAD:
                   1425:                break;
                   1426:
                   1427:        case MOD_UNLOAD:
                   1428:                break;
                   1429:
                   1430:        case MOD_SHUTDOWN:
                   1431:                break;
                   1432:
                   1433:        default:
                   1434:                error = EOPNOTSUPP;
                   1435:                break;
                   1436:
                   1437:        }
                   1438:        return (error);
                   1439: }
                   1440:
                   1441: DEV_MODULE(cyclic, cyclic_modevent, NULL);
                   1442: MODULE_VERSION(cyclic, 1);
                   1443: MODULE_DEPEND(cyclic, opensolaris, 1, 1, 1);
1.2.6.1 ! yamt     1444: #endif
        !          1445:
        !          1446: #ifdef __NetBSD__
        !          1447: static int
        !          1448: cyclic_modcmd(modcmd_t cmd, void *data)
        !          1449: {
        !          1450:        switch (cmd) {
        !          1451:        case MODULE_CMD_INIT:
        !          1452:                cyclic_load(NULL);
        !          1453:                return 0;
        !          1454:
        !          1455:        case MODULE_CMD_FINI:
        !          1456:                cyclic_unload();
        !          1457:                return 0;
        !          1458:        default:
        !          1459:                return ENOTTY;
        !          1460:        }
        !          1461: }
        !          1462:
        !          1463: MODULE(MODULE_CLASS_MISC, cyclic, "dtrace");
        !          1464: #endif

CVSweb <webmaster@jp.NetBSD.org>