xref: /illumos-gate/usr/src/uts/sun4v/io/vio_util.c (revision 046cfe5c)
1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License (the "License").
6  * You may not use this file except in compliance with the License.
7  *
8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9  * or http://www.opensolaris.org/os/licensing.
10  * See the License for the specific language governing permissions
11  * and limitations under the License.
12  *
13  * When distributing Covered Code, include this CDDL HEADER in each
14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15  * If applicable, add the following below this CDDL HEADER, with the
16  * fields enclosed by brackets "[]" replaced with your own identifying
17  * information: Portions Copyright [yyyy] [name of copyright owner]
18  *
19  * CDDL HEADER END
20  */
21 
22 /*
23  * Copyright (c) 2006, 2010, Oracle and/or its affiliates. All rights reserved.
24  */
25 
26 #include <sys/types.h>
27 #include <sys/sysmacros.h>
28 #include <sys/errno.h>
29 #include <sys/kmem.h>
30 #include <sys/ksynch.h>
31 #include <sys/stream.h>
32 #include <sys/ddi.h>
33 #include <sys/sunddi.h>
34 #include <sys/vio_util.h>
35 
36 static int vio_pool_cleanup_retries = 10;	/* Max retries to free pool */
37 static int vio_pool_cleanup_delay = 10000;	/* 10ms */
38 
39 /*
40  * Create a pool of mblks from which future vio_allocb() requests
41  * will be serviced.
42  *
43  * NOTE: num_mblks has to non-zero and a power-of-2
44  *
45  * Returns
46  *	0 on success
47  *	EINVAL if num_mblks is zero or not a power of 2.
48  *	ENOSPC if the pool could not be created due to alloc failures.
49  */
50 int
vio_create_mblks(uint64_t num_mblks,size_t mblk_size,uint8_t * mblk_datap,vio_mblk_pool_t ** poolp)51 vio_create_mblks(uint64_t num_mblks, size_t mblk_size, uint8_t *mblk_datap,
52     vio_mblk_pool_t **poolp)
53 {
54 	vio_mblk_pool_t		*vmplp;
55 	vio_mblk_t		*vmp;
56 	uint8_t			*datap;
57 	int			i;
58 	int			rv;
59 
60 	if (!(num_mblks) || (!ISP2(num_mblks))) {
61 		*poolp = 0;
62 		return (EINVAL);
63 	}
64 
65 	vmplp = kmem_zalloc(sizeof (*vmplp), KM_SLEEP);
66 	vmplp->quelen = num_mblks;
67 	vmplp->quemask = num_mblks - 1; /* expects quelen is power-of-2 */
68 	vmplp->mblk_size = mblk_size;
69 
70 	mutex_init(&vmplp->hlock, NULL, MUTEX_DRIVER,
71 	    DDI_INTR_PRI(DDI_INTR_SOFTPRI_DEFAULT));
72 	mutex_init(&vmplp->tlock, NULL, MUTEX_DRIVER,
73 	    DDI_INTR_PRI(DDI_INTR_SOFTPRI_DEFAULT));
74 
75 	vmplp->basep = kmem_zalloc(num_mblks * sizeof (vio_mblk_t), KM_SLEEP);
76 	if (mblk_datap == NULL) {
77 		vmplp->datap = kmem_zalloc(num_mblks * mblk_size, KM_SLEEP);
78 	} else {
79 		vmplp->datap = mblk_datap;
80 		vmplp->flag |= VMPL_FLAG_CLIENT_DATA;
81 	}
82 	vmplp->nextp = NULL;
83 
84 	/* create a queue of pointers to free vio_mblk_t's */
85 	vmplp->quep = kmem_zalloc(vmplp->quelen *
86 	    sizeof (vio_mblk_t *), KM_SLEEP);
87 	vmplp->head = 0;
88 	vmplp->tail =  0;
89 
90 	for (i = 0, datap = vmplp->datap; i < num_mblks; i++) {
91 
92 		vmp = &(vmplp->basep[i]);
93 		vmp->vmplp = vmplp;
94 		vmp->datap = datap;
95 		vmp->reclaim.free_func = vio_freeb;
96 		vmp->reclaim.free_arg = (caddr_t)vmp;
97 		vmp->mp = desballoc(vmp->datap, mblk_size, BPRI_MED,
98 		    &vmp->reclaim);
99 
100 		if (vmp->mp == NULL) {
101 			/* reset tail */
102 			vmplp->tail = vmplp->head;
103 
104 			/*
105 			 * vio_destroy_mblks() frees mblks that have been
106 			 * allocated so far and then destroys the pool.
107 			 */
108 			rv = vio_destroy_mblks(vmplp);
109 			ASSERT(rv == 0);
110 
111 			*poolp = NULL;
112 			return (ENOSPC);
113 		}
114 
115 		vmp->index = i;
116 		vmp->state = VIO_MBLK_FREE;
117 
118 		/* put this vmp on the free stack */
119 		vmplp->quep[vmplp->tail] = vmp;
120 		vmplp->tail = (vmplp->tail + 1) & vmplp->quemask;
121 
122 		datap += mblk_size;
123 	}
124 
125 	*poolp = vmplp;
126 	return (0);
127 }
128 
129 /*
130  * Destroy the pool of mblks. This can only succeed when
131  * all allocated mblks have been returned to the pool.
132  *
133  * It is up to the caller to ensure that no further mblks are
134  * requested from the pool after destroy has been invoked.
135  *
136  * Returns 0 on success, EINVAL if handle is invalid, or
137  * EBUSY if not all mblks reclaimed yet.
138  */
139 int
vio_destroy_mblks(vio_mblk_pool_t * vmplp)140 vio_destroy_mblks(vio_mblk_pool_t *vmplp)
141 {
142 	uint64_t	i;
143 	uint64_t	num_mblks;
144 	vio_mblk_t	*vmp;
145 	int		pool_cleanup_retries = 0;
146 
147 
148 	if (vmplp == NULL)
149 		return (EINVAL);
150 
151 	/*
152 	 * We can only destroy the pool once all the mblks have
153 	 * been reclaimed.
154 	 */
155 	do {
156 		if (vmplp->head == vmplp->tail) {
157 			break;
158 		}
159 
160 		/* some mblks still in use */
161 		drv_usecwait(vio_pool_cleanup_delay);
162 	} while (++pool_cleanup_retries < vio_pool_cleanup_retries);
163 
164 	if (vmplp->head != vmplp->tail) {
165 		return (EBUSY);
166 	}
167 
168 	num_mblks = vmplp->quelen;
169 
170 	/*
171 	 * Set pool flag to tell vio_freeb() which is invoked from freeb(),
172 	 * that it is being called in the context of vio_destroy_mblks().
173 	 * This results in freeing only mblk_t and dblk_t structures for
174 	 * each mp. The associated data buffers are freed below as one big
175 	 * chunk through kmem_free(vmplp->datap).
176 	 */
177 	vmplp->flag |= VMPL_FLAG_DESTROYING;
178 	for (i = 0; i < num_mblks; i++) {
179 		vmp = &(vmplp->basep[i]);
180 		/*
181 		 * It is possible that mblks have been allocated only upto
182 		 * a certain index and the entire quelen has not been
183 		 * initialized. This might happen due to desballoc() failure
184 		 * while creating the pool. The below check handles this
185 		 * condition.
186 		 */
187 		if (vmp->mp != NULL)
188 			freeb(vmp->mp);
189 	}
190 	vmplp->flag &= ~(VMPL_FLAG_DESTROYING);
191 
192 	kmem_free(vmplp->basep, num_mblks * sizeof (vio_mblk_t));
193 	if ((vmplp->flag & VMPL_FLAG_CLIENT_DATA) == 0) {
194 		kmem_free(vmplp->datap, num_mblks * vmplp->mblk_size);
195 	}
196 	kmem_free(vmplp->quep, num_mblks * sizeof (vio_mblk_t *));
197 
198 	mutex_destroy(&vmplp->hlock);
199 	mutex_destroy(&vmplp->tlock);
200 
201 	kmem_free(vmplp, sizeof (*vmplp));
202 
203 	return (0);
204 }
205 
206 /*
207  * Allocate a vio_mblk from the free pool if one is available.
208  * Otherwise returns NULL.
209  */
210 vio_mblk_t *
vio_allocb(vio_mblk_pool_t * vmplp)211 vio_allocb(vio_mblk_pool_t *vmplp)
212 {
213 	vio_mblk_t	*vmp = NULL;
214 	uint32_t	head;
215 
216 	mutex_enter(&vmplp->hlock);
217 	head = (vmplp->head + 1) & vmplp->quemask;
218 	if (head != vmplp->tail) {
219 		/* we have free mblks */
220 		vmp = vmplp->quep[vmplp->head];
221 		vmplp->head = head;
222 		ASSERT(vmp->state == VIO_MBLK_FREE);
223 		vmp->state = VIO_MBLK_BOUND;
224 	}
225 	mutex_exit(&vmplp->hlock);
226 
227 	return (vmp);
228 }
229 
230 /*
231  * Return a mblk to the free pool. Invoked when the upper IP
232  * layers do freemsg() etc on the mblk they were passed.
233  */
234 void
vio_freeb(void * arg)235 vio_freeb(void *arg)
236 {
237 	vio_mblk_t	*vmp = (vio_mblk_t *)arg;
238 	vio_mblk_pool_t	*vmplp = vmp->vmplp;
239 
240 	if (vmplp->flag & VMPL_FLAG_DESTROYING) {
241 		/*
242 		 * This flag indicates that freeb() is being called from
243 		 * vio_destroy_mblks().
244 		 * We don't need to alloc a new mblk_t/dblk_t pair for
245 		 * this data buffer, return from here and the data buffer
246 		 * itself will be freed in vio_destroy_mblks().
247 		 */
248 		return;
249 	}
250 
251 	vmp->mp = desballoc(vmp->datap, vmplp->mblk_size,
252 	    BPRI_MED, &vmp->reclaim);
253 	vmp->state = VIO_MBLK_FREE;
254 
255 	mutex_enter(&vmplp->tlock);
256 	vmplp->quep[vmplp->tail] = vmp;
257 	vmplp->tail = (vmplp->tail + 1) & vmplp->quemask;
258 	mutex_exit(&vmplp->tlock);
259 }
260 
261 
262 /*
263  * This function searches the given mblk pool for mblks that are in the
264  * BOUND state and moves them to the FREE state. Note that only clients that
265  * are operating in RxDringData mode use this function. This allows such
266  * clients to reclaim buffers that are provided to the peer as shared memory,
267  * before calling vio_destroy_mblks(). We don't need this in other cases
268  * as the buffer is locally managed.
269  */
270 void
vio_clobber_pool(vio_mblk_pool_t * vmplp)271 vio_clobber_pool(vio_mblk_pool_t *vmplp)
272 {
273 	uint64_t	num_mblks = vmplp->quelen;
274 	uint64_t	i;
275 	vio_mblk_t	*vmp;
276 
277 	mutex_enter(&vmplp->hlock);
278 	mutex_enter(&vmplp->tlock);
279 	for (i = 0; i < num_mblks; i++) {
280 		vmp = &(vmplp->basep[i]);
281 		if ((vmp->state & VIO_MBLK_BOUND) != 0) {
282 			/* put this vmp on the free stack */
283 			vmp->state = VIO_MBLK_FREE;
284 			ASSERT(vmplp->tail != vmplp->head);
285 			vmplp->quep[vmplp->tail] = vmp;
286 			vmplp->tail = (vmplp->tail + 1) & vmplp->quemask;
287 		}
288 	}
289 	mutex_exit(&vmplp->tlock);
290 	mutex_exit(&vmplp->hlock);
291 }
292 
293 /*
294  * Create a multiple pools of mblks from which future vio_allocb()
295  * or vio_multipool_allocb() requests will be serviced.
296  *
297  * Arguments:
298  *	vmultip -- A pointer to vio_multi_pool_t structure.
299  *	num_pools -- Number of the pools.
300  *	... -- Variable arguments consisting a list of buffer sizes for
301  *		each pool and list of number of buffers for each pool.
302  *
303  * NOTE: The restrictions of vio_create_mblks() apply to this interface also.
304  *
305  * Returns 0 on success or an error returned by vio_create_mblks().
306  */
307 int
vio_init_multipools(vio_multi_pool_t * vmultip,int num_pools,...)308 vio_init_multipools(vio_multi_pool_t *vmultip, int num_pools, ...)
309 {
310 	int		i;
311 	int		status;
312 	char		*tbuf;
313 	va_list		vap;
314 	vio_mblk_pool_t *fvmp = NULL;
315 
316 	/*
317 	 * Allocate memory for all of the following in one allocation.
318 	 * 	bufsz_tbl -- sizeof (uint32_t) * num_pools
319 	 * 	nbuf_tbl  -- sizeof (uint32_t) * num_pools
320 	 *	vmpp	  -- sizeof (vio_mblk_pool_t *) * numpools
321 	 */
322 	vmultip->tbsz = (sizeof (uint32_t) * num_pools) +
323 	    (sizeof (uint32_t) * num_pools) +
324 	    (sizeof (vio_mblk_pool_t *) * num_pools);
325 	tbuf = kmem_zalloc(vmultip->tbsz, KM_SLEEP);
326 	vmultip->bufsz_tbl = (uint32_t *)tbuf;
327 	vmultip->nbuf_tbl = (uint32_t *)(tbuf +
328 	    (sizeof (uint32_t) * num_pools));
329 	vmultip->vmpp = (vio_mblk_pool_t **)(tbuf +
330 	    (sizeof (uint32_t) * num_pools * 2));
331 	vmultip->num_pools = num_pools;
332 
333 	/* initialize the array first */
334 	va_start(vap, num_pools);
335 	for (i = 0; i < num_pools; i++) {
336 		vmultip->bufsz_tbl[i] = va_arg(vap, uint32_t);
337 	}
338 	for (i = 0; i < num_pools; i++) {
339 		vmultip->nbuf_tbl[i] = va_arg(vap, uint32_t);
340 	}
341 	va_end(vap);
342 
343 	for (i = 0; i < vmultip->num_pools; i++) {
344 		status = vio_create_mblks(vmultip->nbuf_tbl[i],
345 		    vmultip->bufsz_tbl[i], NULL, &vmultip->vmpp[i]);
346 		if (status != 0) {
347 			vio_destroy_multipools(vmultip, &fvmp);
348 			/* We expect to free the pools without failure here */
349 			ASSERT(fvmp == NULL);
350 			return (status);
351 		}
352 	}
353 	return (0);
354 }
355 
356 /*
357  * Destroy the multiple pools of mblks. This can only succeed when
358  * all allocated mblks have been returned to the pool.
359  *
360  * If a pool of mblks couldn't be destroyed, then the failed vio_mblk_pool_t
361  * pointers are returned via th fvmp list. Its the caller's
362  * responsibility to check this list and free them later at an appropriate
363  * time with vio_destroy_mblks().
364  *
365  * Arguments:
366  *	vmultip -- A pointer to vio_multi_pool_t structure.
367  *	fvmp -- A list in which the pools that couldn't be destroyed are
368  *		returned.
369  */
370 void
vio_destroy_multipools(vio_multi_pool_t * vmultip,vio_mblk_pool_t ** fvmp)371 vio_destroy_multipools(vio_multi_pool_t *vmultip, vio_mblk_pool_t **fvmp)
372 {
373 	int i;
374 	vio_mblk_pool_t *vmp;
375 
376 	for (i = 0; i < vmultip->num_pools; i++) {
377 		if ((vmp = vmultip->vmpp[i]) != NULL) {
378 			if (vio_destroy_mblks(vmp)) {
379 				/*
380 				 * if we cannot reclaim all mblks, then
381 				 * return the pool in the failed vmp
382 				 * list(fvmp).
383 				 */
384 				vmp->nextp =  *fvmp;
385 				*fvmp = vmp;
386 			}
387 		}
388 	}
389 	if (vmultip->tbsz != 0)
390 		kmem_free(vmultip->bufsz_tbl, vmultip->tbsz);
391 	vmultip->bufsz_tbl = NULL;
392 	vmultip->nbuf_tbl = NULL;
393 	vmultip->vmpp = NULL;
394 	vmultip->num_pools = 0;
395 	vmultip->tbsz = 0;
396 }
397 
398 
399 /*
400  * Allocate an vio_mblk from one of the free pools, but tries the pool that
401  * best fits size requested first.
402  */
403 vio_mblk_t *
vio_multipool_allocb(vio_multi_pool_t * vmultip,size_t size)404 vio_multipool_allocb(vio_multi_pool_t *vmultip, size_t size)
405 {
406 	int i;
407 	vio_mblk_t *vmp = NULL;
408 
409 	/* Try allocating any size that fits */
410 	for (i = 0; i < vmultip->num_pools; i++) {
411 		if (size > vmultip->bufsz_tbl[i]) {
412 			continue;
413 		}
414 		vmp = vio_allocb(vmultip->vmpp[i]);
415 		if (vmp != NULL) {
416 			break;
417 		}
418 	}
419 	return (vmp);
420 }
421 
422 /*
423  * -----------------------------------------------------------------------------
424  * LDoms versioning functions
425  *
426  * Future work: the version negotiating code in the various VIO drivers
427  * could be made common and placed here.
428  */
429 
430 /*
431  * Description:
432  *	This function checks to see if the supplied version tuple (major,minor)
433  *	is supported by the version 'ver', negotiated during the handshake
434  *	between the client and the server (ver).
435  *
436  * Assumption:
437  *	This function assumes that backward compatability is not broken in
438  *	newer minor versions of the protocol (e.g. v1.5 & v1.1 support v1.0)
439  *
440  * Return Value:
441  *	B_TRUE		- The (major,minor) version is supported
442  *	B_FALSE		- not supported
443  */
444 boolean_t
vio_ver_is_supported(vio_ver_t ver,uint16_t major,uint16_t minor)445 vio_ver_is_supported(vio_ver_t ver, uint16_t major, uint16_t minor)
446 {
447 	if ((ver.major == major) && (ver.minor >= minor))
448 		return (B_TRUE);
449 
450 	return (B_FALSE);
451 }
452