xref: /openbmc/linux/net/ceph/osdmap.c (revision 0ed7285e)
13d14c5d2SYehuda Sadeh 
23d14c5d2SYehuda Sadeh #include <linux/ceph/ceph_debug.h>
33d14c5d2SYehuda Sadeh 
43d14c5d2SYehuda Sadeh #include <linux/module.h>
53d14c5d2SYehuda Sadeh #include <linux/slab.h>
63d14c5d2SYehuda Sadeh #include <asm/div64.h>
73d14c5d2SYehuda Sadeh 
83d14c5d2SYehuda Sadeh #include <linux/ceph/libceph.h>
93d14c5d2SYehuda Sadeh #include <linux/ceph/osdmap.h>
103d14c5d2SYehuda Sadeh #include <linux/ceph/decode.h>
113d14c5d2SYehuda Sadeh #include <linux/crush/hash.h>
123d14c5d2SYehuda Sadeh #include <linux/crush/mapper.h>
133d14c5d2SYehuda Sadeh 
143d14c5d2SYehuda Sadeh char *ceph_osdmap_state_str(char *str, int len, int state)
153d14c5d2SYehuda Sadeh {
163d14c5d2SYehuda Sadeh 	int flag = 0;
173d14c5d2SYehuda Sadeh 
183d14c5d2SYehuda Sadeh 	if (!len)
193d14c5d2SYehuda Sadeh 		goto done;
203d14c5d2SYehuda Sadeh 
213d14c5d2SYehuda Sadeh 	*str = '\0';
223d14c5d2SYehuda Sadeh 	if (state) {
233d14c5d2SYehuda Sadeh 		if (state & CEPH_OSD_EXISTS) {
243d14c5d2SYehuda Sadeh 			snprintf(str, len, "exists");
253d14c5d2SYehuda Sadeh 			flag = 1;
263d14c5d2SYehuda Sadeh 		}
273d14c5d2SYehuda Sadeh 		if (state & CEPH_OSD_UP) {
283d14c5d2SYehuda Sadeh 			snprintf(str, len, "%s%s%s", str, (flag ? ", " : ""),
293d14c5d2SYehuda Sadeh 				 "up");
303d14c5d2SYehuda Sadeh 			flag = 1;
313d14c5d2SYehuda Sadeh 		}
323d14c5d2SYehuda Sadeh 	} else {
333d14c5d2SYehuda Sadeh 		snprintf(str, len, "doesn't exist");
343d14c5d2SYehuda Sadeh 	}
353d14c5d2SYehuda Sadeh done:
363d14c5d2SYehuda Sadeh 	return str;
373d14c5d2SYehuda Sadeh }
383d14c5d2SYehuda Sadeh 
393d14c5d2SYehuda Sadeh /* maps */
403d14c5d2SYehuda Sadeh 
4195c96174SEric Dumazet static int calc_bits_of(unsigned int t)
423d14c5d2SYehuda Sadeh {
433d14c5d2SYehuda Sadeh 	int b = 0;
443d14c5d2SYehuda Sadeh 	while (t) {
453d14c5d2SYehuda Sadeh 		t = t >> 1;
463d14c5d2SYehuda Sadeh 		b++;
473d14c5d2SYehuda Sadeh 	}
483d14c5d2SYehuda Sadeh 	return b;
493d14c5d2SYehuda Sadeh }
503d14c5d2SYehuda Sadeh 
513d14c5d2SYehuda Sadeh /*
523d14c5d2SYehuda Sadeh  * the foo_mask is the smallest value 2^n-1 that is >= foo.
533d14c5d2SYehuda Sadeh  */
543d14c5d2SYehuda Sadeh static void calc_pg_masks(struct ceph_pg_pool_info *pi)
553d14c5d2SYehuda Sadeh {
563d14c5d2SYehuda Sadeh 	pi->pg_num_mask = (1 << calc_bits_of(le32_to_cpu(pi->v.pg_num)-1)) - 1;
573d14c5d2SYehuda Sadeh 	pi->pgp_num_mask =
583d14c5d2SYehuda Sadeh 		(1 << calc_bits_of(le32_to_cpu(pi->v.pgp_num)-1)) - 1;
593d14c5d2SYehuda Sadeh 	pi->lpg_num_mask =
603d14c5d2SYehuda Sadeh 		(1 << calc_bits_of(le32_to_cpu(pi->v.lpg_num)-1)) - 1;
613d14c5d2SYehuda Sadeh 	pi->lpgp_num_mask =
623d14c5d2SYehuda Sadeh 		(1 << calc_bits_of(le32_to_cpu(pi->v.lpgp_num)-1)) - 1;
633d14c5d2SYehuda Sadeh }
643d14c5d2SYehuda Sadeh 
653d14c5d2SYehuda Sadeh /*
663d14c5d2SYehuda Sadeh  * decode crush map
673d14c5d2SYehuda Sadeh  */
683d14c5d2SYehuda Sadeh static int crush_decode_uniform_bucket(void **p, void *end,
693d14c5d2SYehuda Sadeh 				       struct crush_bucket_uniform *b)
703d14c5d2SYehuda Sadeh {
713d14c5d2SYehuda Sadeh 	dout("crush_decode_uniform_bucket %p to %p\n", *p, end);
723d14c5d2SYehuda Sadeh 	ceph_decode_need(p, end, (1+b->h.size) * sizeof(u32), bad);
733d14c5d2SYehuda Sadeh 	b->item_weight = ceph_decode_32(p);
743d14c5d2SYehuda Sadeh 	return 0;
753d14c5d2SYehuda Sadeh bad:
763d14c5d2SYehuda Sadeh 	return -EINVAL;
773d14c5d2SYehuda Sadeh }
783d14c5d2SYehuda Sadeh 
793d14c5d2SYehuda Sadeh static int crush_decode_list_bucket(void **p, void *end,
803d14c5d2SYehuda Sadeh 				    struct crush_bucket_list *b)
813d14c5d2SYehuda Sadeh {
823d14c5d2SYehuda Sadeh 	int j;
833d14c5d2SYehuda Sadeh 	dout("crush_decode_list_bucket %p to %p\n", *p, end);
843d14c5d2SYehuda Sadeh 	b->item_weights = kcalloc(b->h.size, sizeof(u32), GFP_NOFS);
853d14c5d2SYehuda Sadeh 	if (b->item_weights == NULL)
863d14c5d2SYehuda Sadeh 		return -ENOMEM;
873d14c5d2SYehuda Sadeh 	b->sum_weights = kcalloc(b->h.size, sizeof(u32), GFP_NOFS);
883d14c5d2SYehuda Sadeh 	if (b->sum_weights == NULL)
893d14c5d2SYehuda Sadeh 		return -ENOMEM;
903d14c5d2SYehuda Sadeh 	ceph_decode_need(p, end, 2 * b->h.size * sizeof(u32), bad);
913d14c5d2SYehuda Sadeh 	for (j = 0; j < b->h.size; j++) {
923d14c5d2SYehuda Sadeh 		b->item_weights[j] = ceph_decode_32(p);
933d14c5d2SYehuda Sadeh 		b->sum_weights[j] = ceph_decode_32(p);
943d14c5d2SYehuda Sadeh 	}
953d14c5d2SYehuda Sadeh 	return 0;
963d14c5d2SYehuda Sadeh bad:
973d14c5d2SYehuda Sadeh 	return -EINVAL;
983d14c5d2SYehuda Sadeh }
993d14c5d2SYehuda Sadeh 
1003d14c5d2SYehuda Sadeh static int crush_decode_tree_bucket(void **p, void *end,
1013d14c5d2SYehuda Sadeh 				    struct crush_bucket_tree *b)
1023d14c5d2SYehuda Sadeh {
1033d14c5d2SYehuda Sadeh 	int j;
1043d14c5d2SYehuda Sadeh 	dout("crush_decode_tree_bucket %p to %p\n", *p, end);
1053d14c5d2SYehuda Sadeh 	ceph_decode_32_safe(p, end, b->num_nodes, bad);
1063d14c5d2SYehuda Sadeh 	b->node_weights = kcalloc(b->num_nodes, sizeof(u32), GFP_NOFS);
1073d14c5d2SYehuda Sadeh 	if (b->node_weights == NULL)
1083d14c5d2SYehuda Sadeh 		return -ENOMEM;
1093d14c5d2SYehuda Sadeh 	ceph_decode_need(p, end, b->num_nodes * sizeof(u32), bad);
1103d14c5d2SYehuda Sadeh 	for (j = 0; j < b->num_nodes; j++)
1113d14c5d2SYehuda Sadeh 		b->node_weights[j] = ceph_decode_32(p);
1123d14c5d2SYehuda Sadeh 	return 0;
1133d14c5d2SYehuda Sadeh bad:
1143d14c5d2SYehuda Sadeh 	return -EINVAL;
1153d14c5d2SYehuda Sadeh }
1163d14c5d2SYehuda Sadeh 
1173d14c5d2SYehuda Sadeh static int crush_decode_straw_bucket(void **p, void *end,
1183d14c5d2SYehuda Sadeh 				     struct crush_bucket_straw *b)
1193d14c5d2SYehuda Sadeh {
1203d14c5d2SYehuda Sadeh 	int j;
1213d14c5d2SYehuda Sadeh 	dout("crush_decode_straw_bucket %p to %p\n", *p, end);
1223d14c5d2SYehuda Sadeh 	b->item_weights = kcalloc(b->h.size, sizeof(u32), GFP_NOFS);
1233d14c5d2SYehuda Sadeh 	if (b->item_weights == NULL)
1243d14c5d2SYehuda Sadeh 		return -ENOMEM;
1253d14c5d2SYehuda Sadeh 	b->straws = kcalloc(b->h.size, sizeof(u32), GFP_NOFS);
1263d14c5d2SYehuda Sadeh 	if (b->straws == NULL)
1273d14c5d2SYehuda Sadeh 		return -ENOMEM;
1283d14c5d2SYehuda Sadeh 	ceph_decode_need(p, end, 2 * b->h.size * sizeof(u32), bad);
1293d14c5d2SYehuda Sadeh 	for (j = 0; j < b->h.size; j++) {
1303d14c5d2SYehuda Sadeh 		b->item_weights[j] = ceph_decode_32(p);
1313d14c5d2SYehuda Sadeh 		b->straws[j] = ceph_decode_32(p);
1323d14c5d2SYehuda Sadeh 	}
1333d14c5d2SYehuda Sadeh 	return 0;
1343d14c5d2SYehuda Sadeh bad:
1353d14c5d2SYehuda Sadeh 	return -EINVAL;
1363d14c5d2SYehuda Sadeh }
1373d14c5d2SYehuda Sadeh 
138546f04efSSage Weil static int skip_name_map(void **p, void *end)
139546f04efSSage Weil {
140546f04efSSage Weil         int len;
141546f04efSSage Weil         ceph_decode_32_safe(p, end, len ,bad);
142546f04efSSage Weil         while (len--) {
143546f04efSSage Weil                 int strlen;
144546f04efSSage Weil                 *p += sizeof(u32);
145546f04efSSage Weil                 ceph_decode_32_safe(p, end, strlen, bad);
146546f04efSSage Weil                 *p += strlen;
147546f04efSSage Weil }
148546f04efSSage Weil         return 0;
149546f04efSSage Weil bad:
150546f04efSSage Weil         return -EINVAL;
151546f04efSSage Weil }
152546f04efSSage Weil 
1533d14c5d2SYehuda Sadeh static struct crush_map *crush_decode(void *pbyval, void *end)
1543d14c5d2SYehuda Sadeh {
1553d14c5d2SYehuda Sadeh 	struct crush_map *c;
1563d14c5d2SYehuda Sadeh 	int err = -EINVAL;
1573d14c5d2SYehuda Sadeh 	int i, j;
1583d14c5d2SYehuda Sadeh 	void **p = &pbyval;
1593d14c5d2SYehuda Sadeh 	void *start = pbyval;
1603d14c5d2SYehuda Sadeh 	u32 magic;
161546f04efSSage Weil 	u32 num_name_maps;
1623d14c5d2SYehuda Sadeh 
1633d14c5d2SYehuda Sadeh 	dout("crush_decode %p to %p len %d\n", *p, end, (int)(end - *p));
1643d14c5d2SYehuda Sadeh 
1653d14c5d2SYehuda Sadeh 	c = kzalloc(sizeof(*c), GFP_NOFS);
1663d14c5d2SYehuda Sadeh 	if (c == NULL)
1673d14c5d2SYehuda Sadeh 		return ERR_PTR(-ENOMEM);
1683d14c5d2SYehuda Sadeh 
169546f04efSSage Weil         /* set tunables to default values */
170546f04efSSage Weil         c->choose_local_tries = 2;
171546f04efSSage Weil         c->choose_local_fallback_tries = 5;
172546f04efSSage Weil         c->choose_total_tries = 19;
173546f04efSSage Weil 
1743d14c5d2SYehuda Sadeh 	ceph_decode_need(p, end, 4*sizeof(u32), bad);
1753d14c5d2SYehuda Sadeh 	magic = ceph_decode_32(p);
1763d14c5d2SYehuda Sadeh 	if (magic != CRUSH_MAGIC) {
1773d14c5d2SYehuda Sadeh 		pr_err("crush_decode magic %x != current %x\n",
17895c96174SEric Dumazet 		       (unsigned int)magic, (unsigned int)CRUSH_MAGIC);
1793d14c5d2SYehuda Sadeh 		goto bad;
1803d14c5d2SYehuda Sadeh 	}
1813d14c5d2SYehuda Sadeh 	c->max_buckets = ceph_decode_32(p);
1823d14c5d2SYehuda Sadeh 	c->max_rules = ceph_decode_32(p);
1833d14c5d2SYehuda Sadeh 	c->max_devices = ceph_decode_32(p);
1843d14c5d2SYehuda Sadeh 
1853d14c5d2SYehuda Sadeh 	c->buckets = kcalloc(c->max_buckets, sizeof(*c->buckets), GFP_NOFS);
1863d14c5d2SYehuda Sadeh 	if (c->buckets == NULL)
1873d14c5d2SYehuda Sadeh 		goto badmem;
1883d14c5d2SYehuda Sadeh 	c->rules = kcalloc(c->max_rules, sizeof(*c->rules), GFP_NOFS);
1893d14c5d2SYehuda Sadeh 	if (c->rules == NULL)
1903d14c5d2SYehuda Sadeh 		goto badmem;
1913d14c5d2SYehuda Sadeh 
1923d14c5d2SYehuda Sadeh 	/* buckets */
1933d14c5d2SYehuda Sadeh 	for (i = 0; i < c->max_buckets; i++) {
1943d14c5d2SYehuda Sadeh 		int size = 0;
1953d14c5d2SYehuda Sadeh 		u32 alg;
1963d14c5d2SYehuda Sadeh 		struct crush_bucket *b;
1973d14c5d2SYehuda Sadeh 
1983d14c5d2SYehuda Sadeh 		ceph_decode_32_safe(p, end, alg, bad);
1993d14c5d2SYehuda Sadeh 		if (alg == 0) {
2003d14c5d2SYehuda Sadeh 			c->buckets[i] = NULL;
2013d14c5d2SYehuda Sadeh 			continue;
2023d14c5d2SYehuda Sadeh 		}
2033d14c5d2SYehuda Sadeh 		dout("crush_decode bucket %d off %x %p to %p\n",
2043d14c5d2SYehuda Sadeh 		     i, (int)(*p-start), *p, end);
2053d14c5d2SYehuda Sadeh 
2063d14c5d2SYehuda Sadeh 		switch (alg) {
2073d14c5d2SYehuda Sadeh 		case CRUSH_BUCKET_UNIFORM:
2083d14c5d2SYehuda Sadeh 			size = sizeof(struct crush_bucket_uniform);
2093d14c5d2SYehuda Sadeh 			break;
2103d14c5d2SYehuda Sadeh 		case CRUSH_BUCKET_LIST:
2113d14c5d2SYehuda Sadeh 			size = sizeof(struct crush_bucket_list);
2123d14c5d2SYehuda Sadeh 			break;
2133d14c5d2SYehuda Sadeh 		case CRUSH_BUCKET_TREE:
2143d14c5d2SYehuda Sadeh 			size = sizeof(struct crush_bucket_tree);
2153d14c5d2SYehuda Sadeh 			break;
2163d14c5d2SYehuda Sadeh 		case CRUSH_BUCKET_STRAW:
2173d14c5d2SYehuda Sadeh 			size = sizeof(struct crush_bucket_straw);
2183d14c5d2SYehuda Sadeh 			break;
2193d14c5d2SYehuda Sadeh 		default:
2203d14c5d2SYehuda Sadeh 			err = -EINVAL;
2213d14c5d2SYehuda Sadeh 			goto bad;
2223d14c5d2SYehuda Sadeh 		}
2233d14c5d2SYehuda Sadeh 		BUG_ON(size == 0);
2243d14c5d2SYehuda Sadeh 		b = c->buckets[i] = kzalloc(size, GFP_NOFS);
2253d14c5d2SYehuda Sadeh 		if (b == NULL)
2263d14c5d2SYehuda Sadeh 			goto badmem;
2273d14c5d2SYehuda Sadeh 
2283d14c5d2SYehuda Sadeh 		ceph_decode_need(p, end, 4*sizeof(u32), bad);
2293d14c5d2SYehuda Sadeh 		b->id = ceph_decode_32(p);
2303d14c5d2SYehuda Sadeh 		b->type = ceph_decode_16(p);
2313d14c5d2SYehuda Sadeh 		b->alg = ceph_decode_8(p);
2323d14c5d2SYehuda Sadeh 		b->hash = ceph_decode_8(p);
2333d14c5d2SYehuda Sadeh 		b->weight = ceph_decode_32(p);
2343d14c5d2SYehuda Sadeh 		b->size = ceph_decode_32(p);
2353d14c5d2SYehuda Sadeh 
2363d14c5d2SYehuda Sadeh 		dout("crush_decode bucket size %d off %x %p to %p\n",
2373d14c5d2SYehuda Sadeh 		     b->size, (int)(*p-start), *p, end);
2383d14c5d2SYehuda Sadeh 
2393d14c5d2SYehuda Sadeh 		b->items = kcalloc(b->size, sizeof(__s32), GFP_NOFS);
2403d14c5d2SYehuda Sadeh 		if (b->items == NULL)
2413d14c5d2SYehuda Sadeh 			goto badmem;
2423d14c5d2SYehuda Sadeh 		b->perm = kcalloc(b->size, sizeof(u32), GFP_NOFS);
2433d14c5d2SYehuda Sadeh 		if (b->perm == NULL)
2443d14c5d2SYehuda Sadeh 			goto badmem;
2453d14c5d2SYehuda Sadeh 		b->perm_n = 0;
2463d14c5d2SYehuda Sadeh 
2473d14c5d2SYehuda Sadeh 		ceph_decode_need(p, end, b->size*sizeof(u32), bad);
2483d14c5d2SYehuda Sadeh 		for (j = 0; j < b->size; j++)
2493d14c5d2SYehuda Sadeh 			b->items[j] = ceph_decode_32(p);
2503d14c5d2SYehuda Sadeh 
2513d14c5d2SYehuda Sadeh 		switch (b->alg) {
2523d14c5d2SYehuda Sadeh 		case CRUSH_BUCKET_UNIFORM:
2533d14c5d2SYehuda Sadeh 			err = crush_decode_uniform_bucket(p, end,
2543d14c5d2SYehuda Sadeh 				  (struct crush_bucket_uniform *)b);
2553d14c5d2SYehuda Sadeh 			if (err < 0)
2563d14c5d2SYehuda Sadeh 				goto bad;
2573d14c5d2SYehuda Sadeh 			break;
2583d14c5d2SYehuda Sadeh 		case CRUSH_BUCKET_LIST:
2593d14c5d2SYehuda Sadeh 			err = crush_decode_list_bucket(p, end,
2603d14c5d2SYehuda Sadeh 			       (struct crush_bucket_list *)b);
2613d14c5d2SYehuda Sadeh 			if (err < 0)
2623d14c5d2SYehuda Sadeh 				goto bad;
2633d14c5d2SYehuda Sadeh 			break;
2643d14c5d2SYehuda Sadeh 		case CRUSH_BUCKET_TREE:
2653d14c5d2SYehuda Sadeh 			err = crush_decode_tree_bucket(p, end,
2663d14c5d2SYehuda Sadeh 				(struct crush_bucket_tree *)b);
2673d14c5d2SYehuda Sadeh 			if (err < 0)
2683d14c5d2SYehuda Sadeh 				goto bad;
2693d14c5d2SYehuda Sadeh 			break;
2703d14c5d2SYehuda Sadeh 		case CRUSH_BUCKET_STRAW:
2713d14c5d2SYehuda Sadeh 			err = crush_decode_straw_bucket(p, end,
2723d14c5d2SYehuda Sadeh 				(struct crush_bucket_straw *)b);
2733d14c5d2SYehuda Sadeh 			if (err < 0)
2743d14c5d2SYehuda Sadeh 				goto bad;
2753d14c5d2SYehuda Sadeh 			break;
2763d14c5d2SYehuda Sadeh 		}
2773d14c5d2SYehuda Sadeh 	}
2783d14c5d2SYehuda Sadeh 
2793d14c5d2SYehuda Sadeh 	/* rules */
2803d14c5d2SYehuda Sadeh 	dout("rule vec is %p\n", c->rules);
2813d14c5d2SYehuda Sadeh 	for (i = 0; i < c->max_rules; i++) {
2823d14c5d2SYehuda Sadeh 		u32 yes;
2833d14c5d2SYehuda Sadeh 		struct crush_rule *r;
2843d14c5d2SYehuda Sadeh 
2853d14c5d2SYehuda Sadeh 		ceph_decode_32_safe(p, end, yes, bad);
2863d14c5d2SYehuda Sadeh 		if (!yes) {
2873d14c5d2SYehuda Sadeh 			dout("crush_decode NO rule %d off %x %p to %p\n",
2883d14c5d2SYehuda Sadeh 			     i, (int)(*p-start), *p, end);
2893d14c5d2SYehuda Sadeh 			c->rules[i] = NULL;
2903d14c5d2SYehuda Sadeh 			continue;
2913d14c5d2SYehuda Sadeh 		}
2923d14c5d2SYehuda Sadeh 
2933d14c5d2SYehuda Sadeh 		dout("crush_decode rule %d off %x %p to %p\n",
2943d14c5d2SYehuda Sadeh 		     i, (int)(*p-start), *p, end);
2953d14c5d2SYehuda Sadeh 
2963d14c5d2SYehuda Sadeh 		/* len */
2973d14c5d2SYehuda Sadeh 		ceph_decode_32_safe(p, end, yes, bad);
2983d14c5d2SYehuda Sadeh #if BITS_PER_LONG == 32
2993d14c5d2SYehuda Sadeh 		err = -EINVAL;
30064486697SXi Wang 		if (yes > (ULONG_MAX - sizeof(*r))
30164486697SXi Wang 			  / sizeof(struct crush_rule_step))
3023d14c5d2SYehuda Sadeh 			goto bad;
3033d14c5d2SYehuda Sadeh #endif
3043d14c5d2SYehuda Sadeh 		r = c->rules[i] = kmalloc(sizeof(*r) +
3053d14c5d2SYehuda Sadeh 					  yes*sizeof(struct crush_rule_step),
3063d14c5d2SYehuda Sadeh 					  GFP_NOFS);
3073d14c5d2SYehuda Sadeh 		if (r == NULL)
3083d14c5d2SYehuda Sadeh 			goto badmem;
3093d14c5d2SYehuda Sadeh 		dout(" rule %d is at %p\n", i, r);
3103d14c5d2SYehuda Sadeh 		r->len = yes;
3113d14c5d2SYehuda Sadeh 		ceph_decode_copy_safe(p, end, &r->mask, 4, bad); /* 4 u8's */
3123d14c5d2SYehuda Sadeh 		ceph_decode_need(p, end, r->len*3*sizeof(u32), bad);
3133d14c5d2SYehuda Sadeh 		for (j = 0; j < r->len; j++) {
3143d14c5d2SYehuda Sadeh 			r->steps[j].op = ceph_decode_32(p);
3153d14c5d2SYehuda Sadeh 			r->steps[j].arg1 = ceph_decode_32(p);
3163d14c5d2SYehuda Sadeh 			r->steps[j].arg2 = ceph_decode_32(p);
3173d14c5d2SYehuda Sadeh 		}
3183d14c5d2SYehuda Sadeh 	}
3193d14c5d2SYehuda Sadeh 
3203d14c5d2SYehuda Sadeh 	/* ignore trailing name maps. */
321546f04efSSage Weil         for (num_name_maps = 0; num_name_maps < 3; num_name_maps++) {
322546f04efSSage Weil                 err = skip_name_map(p, end);
323546f04efSSage Weil                 if (err < 0)
324546f04efSSage Weil                         goto done;
325546f04efSSage Weil         }
3263d14c5d2SYehuda Sadeh 
327546f04efSSage Weil         /* tunables */
328546f04efSSage Weil         ceph_decode_need(p, end, 3*sizeof(u32), done);
329546f04efSSage Weil         c->choose_local_tries = ceph_decode_32(p);
330546f04efSSage Weil         c->choose_local_fallback_tries =  ceph_decode_32(p);
331546f04efSSage Weil         c->choose_total_tries = ceph_decode_32(p);
332546f04efSSage Weil         dout("crush decode tunable choose_local_tries = %d",
333546f04efSSage Weil              c->choose_local_tries);
334546f04efSSage Weil         dout("crush decode tunable choose_local_fallback_tries = %d",
335546f04efSSage Weil              c->choose_local_fallback_tries);
336546f04efSSage Weil         dout("crush decode tunable choose_total_tries = %d",
337546f04efSSage Weil              c->choose_total_tries);
338546f04efSSage Weil 
339546f04efSSage Weil done:
3403d14c5d2SYehuda Sadeh 	dout("crush_decode success\n");
3413d14c5d2SYehuda Sadeh 	return c;
3423d14c5d2SYehuda Sadeh 
3433d14c5d2SYehuda Sadeh badmem:
3443d14c5d2SYehuda Sadeh 	err = -ENOMEM;
3453d14c5d2SYehuda Sadeh bad:
3463d14c5d2SYehuda Sadeh 	dout("crush_decode fail %d\n", err);
3473d14c5d2SYehuda Sadeh 	crush_destroy(c);
3483d14c5d2SYehuda Sadeh 	return ERR_PTR(err);
3493d14c5d2SYehuda Sadeh }
3503d14c5d2SYehuda Sadeh 
3513d14c5d2SYehuda Sadeh /*
3523d14c5d2SYehuda Sadeh  * rbtree of pg_mapping for handling pg_temp (explicit mapping of pgid
3533d14c5d2SYehuda Sadeh  * to a set of osds)
3543d14c5d2SYehuda Sadeh  */
3553d14c5d2SYehuda Sadeh static int pgid_cmp(struct ceph_pg l, struct ceph_pg r)
3563d14c5d2SYehuda Sadeh {
3573d14c5d2SYehuda Sadeh 	u64 a = *(u64 *)&l;
3583d14c5d2SYehuda Sadeh 	u64 b = *(u64 *)&r;
3593d14c5d2SYehuda Sadeh 
3603d14c5d2SYehuda Sadeh 	if (a < b)
3613d14c5d2SYehuda Sadeh 		return -1;
3623d14c5d2SYehuda Sadeh 	if (a > b)
3633d14c5d2SYehuda Sadeh 		return 1;
3643d14c5d2SYehuda Sadeh 	return 0;
3653d14c5d2SYehuda Sadeh }
3663d14c5d2SYehuda Sadeh 
3673d14c5d2SYehuda Sadeh static int __insert_pg_mapping(struct ceph_pg_mapping *new,
3683d14c5d2SYehuda Sadeh 			       struct rb_root *root)
3693d14c5d2SYehuda Sadeh {
3703d14c5d2SYehuda Sadeh 	struct rb_node **p = &root->rb_node;
3713d14c5d2SYehuda Sadeh 	struct rb_node *parent = NULL;
3723d14c5d2SYehuda Sadeh 	struct ceph_pg_mapping *pg = NULL;
3733d14c5d2SYehuda Sadeh 	int c;
3743d14c5d2SYehuda Sadeh 
3758adc8b3dSSage Weil 	dout("__insert_pg_mapping %llx %p\n", *(u64 *)&new->pgid, new);
3763d14c5d2SYehuda Sadeh 	while (*p) {
3773d14c5d2SYehuda Sadeh 		parent = *p;
3783d14c5d2SYehuda Sadeh 		pg = rb_entry(parent, struct ceph_pg_mapping, node);
3793d14c5d2SYehuda Sadeh 		c = pgid_cmp(new->pgid, pg->pgid);
3803d14c5d2SYehuda Sadeh 		if (c < 0)
3813d14c5d2SYehuda Sadeh 			p = &(*p)->rb_left;
3823d14c5d2SYehuda Sadeh 		else if (c > 0)
3833d14c5d2SYehuda Sadeh 			p = &(*p)->rb_right;
3843d14c5d2SYehuda Sadeh 		else
3853d14c5d2SYehuda Sadeh 			return -EEXIST;
3863d14c5d2SYehuda Sadeh 	}
3873d14c5d2SYehuda Sadeh 
3883d14c5d2SYehuda Sadeh 	rb_link_node(&new->node, parent, p);
3893d14c5d2SYehuda Sadeh 	rb_insert_color(&new->node, root);
3903d14c5d2SYehuda Sadeh 	return 0;
3913d14c5d2SYehuda Sadeh }
3923d14c5d2SYehuda Sadeh 
3933d14c5d2SYehuda Sadeh static struct ceph_pg_mapping *__lookup_pg_mapping(struct rb_root *root,
3943d14c5d2SYehuda Sadeh 						   struct ceph_pg pgid)
3953d14c5d2SYehuda Sadeh {
3963d14c5d2SYehuda Sadeh 	struct rb_node *n = root->rb_node;
3973d14c5d2SYehuda Sadeh 	struct ceph_pg_mapping *pg;
3983d14c5d2SYehuda Sadeh 	int c;
3993d14c5d2SYehuda Sadeh 
4003d14c5d2SYehuda Sadeh 	while (n) {
4013d14c5d2SYehuda Sadeh 		pg = rb_entry(n, struct ceph_pg_mapping, node);
4023d14c5d2SYehuda Sadeh 		c = pgid_cmp(pgid, pg->pgid);
4038adc8b3dSSage Weil 		if (c < 0) {
4043d14c5d2SYehuda Sadeh 			n = n->rb_left;
4058adc8b3dSSage Weil 		} else if (c > 0) {
4063d14c5d2SYehuda Sadeh 			n = n->rb_right;
4078adc8b3dSSage Weil 		} else {
4088adc8b3dSSage Weil 			dout("__lookup_pg_mapping %llx got %p\n",
4098adc8b3dSSage Weil 			     *(u64 *)&pgid, pg);
4103d14c5d2SYehuda Sadeh 			return pg;
4113d14c5d2SYehuda Sadeh 		}
4128adc8b3dSSage Weil 	}
4133d14c5d2SYehuda Sadeh 	return NULL;
4143d14c5d2SYehuda Sadeh }
4153d14c5d2SYehuda Sadeh 
4168adc8b3dSSage Weil static int __remove_pg_mapping(struct rb_root *root, struct ceph_pg pgid)
4178adc8b3dSSage Weil {
4188adc8b3dSSage Weil 	struct ceph_pg_mapping *pg = __lookup_pg_mapping(root, pgid);
4198adc8b3dSSage Weil 
4208adc8b3dSSage Weil 	if (pg) {
4218adc8b3dSSage Weil 		dout("__remove_pg_mapping %llx %p\n", *(u64 *)&pgid, pg);
4228adc8b3dSSage Weil 		rb_erase(&pg->node, root);
4238adc8b3dSSage Weil 		kfree(pg);
4248adc8b3dSSage Weil 		return 0;
4258adc8b3dSSage Weil 	}
4268adc8b3dSSage Weil 	dout("__remove_pg_mapping %llx dne\n", *(u64 *)&pgid);
4278adc8b3dSSage Weil 	return -ENOENT;
4288adc8b3dSSage Weil }
4298adc8b3dSSage Weil 
4303d14c5d2SYehuda Sadeh /*
4313d14c5d2SYehuda Sadeh  * rbtree of pg pool info
4323d14c5d2SYehuda Sadeh  */
4333d14c5d2SYehuda Sadeh static int __insert_pg_pool(struct rb_root *root, struct ceph_pg_pool_info *new)
4343d14c5d2SYehuda Sadeh {
4353d14c5d2SYehuda Sadeh 	struct rb_node **p = &root->rb_node;
4363d14c5d2SYehuda Sadeh 	struct rb_node *parent = NULL;
4373d14c5d2SYehuda Sadeh 	struct ceph_pg_pool_info *pi = NULL;
4383d14c5d2SYehuda Sadeh 
4393d14c5d2SYehuda Sadeh 	while (*p) {
4403d14c5d2SYehuda Sadeh 		parent = *p;
4413d14c5d2SYehuda Sadeh 		pi = rb_entry(parent, struct ceph_pg_pool_info, node);
4423d14c5d2SYehuda Sadeh 		if (new->id < pi->id)
4433d14c5d2SYehuda Sadeh 			p = &(*p)->rb_left;
4443d14c5d2SYehuda Sadeh 		else if (new->id > pi->id)
4453d14c5d2SYehuda Sadeh 			p = &(*p)->rb_right;
4463d14c5d2SYehuda Sadeh 		else
4473d14c5d2SYehuda Sadeh 			return -EEXIST;
4483d14c5d2SYehuda Sadeh 	}
4493d14c5d2SYehuda Sadeh 
4503d14c5d2SYehuda Sadeh 	rb_link_node(&new->node, parent, p);
4513d14c5d2SYehuda Sadeh 	rb_insert_color(&new->node, root);
4523d14c5d2SYehuda Sadeh 	return 0;
4533d14c5d2SYehuda Sadeh }
4543d14c5d2SYehuda Sadeh 
4553d14c5d2SYehuda Sadeh static struct ceph_pg_pool_info *__lookup_pg_pool(struct rb_root *root, int id)
4563d14c5d2SYehuda Sadeh {
4573d14c5d2SYehuda Sadeh 	struct ceph_pg_pool_info *pi;
4583d14c5d2SYehuda Sadeh 	struct rb_node *n = root->rb_node;
4593d14c5d2SYehuda Sadeh 
4603d14c5d2SYehuda Sadeh 	while (n) {
4613d14c5d2SYehuda Sadeh 		pi = rb_entry(n, struct ceph_pg_pool_info, node);
4623d14c5d2SYehuda Sadeh 		if (id < pi->id)
4633d14c5d2SYehuda Sadeh 			n = n->rb_left;
4643d14c5d2SYehuda Sadeh 		else if (id > pi->id)
4653d14c5d2SYehuda Sadeh 			n = n->rb_right;
4663d14c5d2SYehuda Sadeh 		else
4673d14c5d2SYehuda Sadeh 			return pi;
4683d14c5d2SYehuda Sadeh 	}
4693d14c5d2SYehuda Sadeh 	return NULL;
4703d14c5d2SYehuda Sadeh }
4713d14c5d2SYehuda Sadeh 
4723d14c5d2SYehuda Sadeh int ceph_pg_poolid_by_name(struct ceph_osdmap *map, const char *name)
4733d14c5d2SYehuda Sadeh {
4743d14c5d2SYehuda Sadeh 	struct rb_node *rbp;
4753d14c5d2SYehuda Sadeh 
4763d14c5d2SYehuda Sadeh 	for (rbp = rb_first(&map->pg_pools); rbp; rbp = rb_next(rbp)) {
4773d14c5d2SYehuda Sadeh 		struct ceph_pg_pool_info *pi =
4783d14c5d2SYehuda Sadeh 			rb_entry(rbp, struct ceph_pg_pool_info, node);
4793d14c5d2SYehuda Sadeh 		if (pi->name && strcmp(pi->name, name) == 0)
4803d14c5d2SYehuda Sadeh 			return pi->id;
4813d14c5d2SYehuda Sadeh 	}
4823d14c5d2SYehuda Sadeh 	return -ENOENT;
4833d14c5d2SYehuda Sadeh }
4843d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_pg_poolid_by_name);
4853d14c5d2SYehuda Sadeh 
4863d14c5d2SYehuda Sadeh static void __remove_pg_pool(struct rb_root *root, struct ceph_pg_pool_info *pi)
4873d14c5d2SYehuda Sadeh {
4883d14c5d2SYehuda Sadeh 	rb_erase(&pi->node, root);
4893d14c5d2SYehuda Sadeh 	kfree(pi->name);
4903d14c5d2SYehuda Sadeh 	kfree(pi);
4913d14c5d2SYehuda Sadeh }
4923d14c5d2SYehuda Sadeh 
4933d14c5d2SYehuda Sadeh static int __decode_pool(void **p, void *end, struct ceph_pg_pool_info *pi)
4943d14c5d2SYehuda Sadeh {
49595c96174SEric Dumazet 	unsigned int n, m;
4963d14c5d2SYehuda Sadeh 
4973d14c5d2SYehuda Sadeh 	ceph_decode_copy(p, &pi->v, sizeof(pi->v));
4983d14c5d2SYehuda Sadeh 	calc_pg_masks(pi);
4993d14c5d2SYehuda Sadeh 
5003d14c5d2SYehuda Sadeh 	/* num_snaps * snap_info_t */
5013d14c5d2SYehuda Sadeh 	n = le32_to_cpu(pi->v.num_snaps);
5023d14c5d2SYehuda Sadeh 	while (n--) {
5033d14c5d2SYehuda Sadeh 		ceph_decode_need(p, end, sizeof(u64) + 1 + sizeof(u64) +
5043d14c5d2SYehuda Sadeh 				 sizeof(struct ceph_timespec), bad);
5053d14c5d2SYehuda Sadeh 		*p += sizeof(u64) +       /* key */
5063d14c5d2SYehuda Sadeh 			1 + sizeof(u64) + /* u8, snapid */
5073d14c5d2SYehuda Sadeh 			sizeof(struct ceph_timespec);
5083d14c5d2SYehuda Sadeh 		m = ceph_decode_32(p);    /* snap name */
5093d14c5d2SYehuda Sadeh 		*p += m;
5103d14c5d2SYehuda Sadeh 	}
5113d14c5d2SYehuda Sadeh 
5123d14c5d2SYehuda Sadeh 	*p += le32_to_cpu(pi->v.num_removed_snap_intervals) * sizeof(u64) * 2;
5133d14c5d2SYehuda Sadeh 	return 0;
5143d14c5d2SYehuda Sadeh 
5153d14c5d2SYehuda Sadeh bad:
5163d14c5d2SYehuda Sadeh 	return -EINVAL;
5173d14c5d2SYehuda Sadeh }
5183d14c5d2SYehuda Sadeh 
5193d14c5d2SYehuda Sadeh static int __decode_pool_names(void **p, void *end, struct ceph_osdmap *map)
5203d14c5d2SYehuda Sadeh {
5213d14c5d2SYehuda Sadeh 	struct ceph_pg_pool_info *pi;
5223d14c5d2SYehuda Sadeh 	u32 num, len, pool;
5233d14c5d2SYehuda Sadeh 
5243d14c5d2SYehuda Sadeh 	ceph_decode_32_safe(p, end, num, bad);
5253d14c5d2SYehuda Sadeh 	dout(" %d pool names\n", num);
5263d14c5d2SYehuda Sadeh 	while (num--) {
5273d14c5d2SYehuda Sadeh 		ceph_decode_32_safe(p, end, pool, bad);
5283d14c5d2SYehuda Sadeh 		ceph_decode_32_safe(p, end, len, bad);
5293d14c5d2SYehuda Sadeh 		dout("  pool %d len %d\n", pool, len);
530ad3b904cSXi Wang 		ceph_decode_need(p, end, len, bad);
5313d14c5d2SYehuda Sadeh 		pi = __lookup_pg_pool(&map->pg_pools, pool);
5323d14c5d2SYehuda Sadeh 		if (pi) {
533ad3b904cSXi Wang 			char *name = kstrndup(*p, len, GFP_NOFS);
534ad3b904cSXi Wang 
535ad3b904cSXi Wang 			if (!name)
536ad3b904cSXi Wang 				return -ENOMEM;
5373d14c5d2SYehuda Sadeh 			kfree(pi->name);
538ad3b904cSXi Wang 			pi->name = name;
5393d14c5d2SYehuda Sadeh 			dout("  name is %s\n", pi->name);
5403d14c5d2SYehuda Sadeh 		}
5413d14c5d2SYehuda Sadeh 		*p += len;
5423d14c5d2SYehuda Sadeh 	}
5433d14c5d2SYehuda Sadeh 	return 0;
5443d14c5d2SYehuda Sadeh 
5453d14c5d2SYehuda Sadeh bad:
5463d14c5d2SYehuda Sadeh 	return -EINVAL;
5473d14c5d2SYehuda Sadeh }
5483d14c5d2SYehuda Sadeh 
5493d14c5d2SYehuda Sadeh /*
5503d14c5d2SYehuda Sadeh  * osd map
5513d14c5d2SYehuda Sadeh  */
5523d14c5d2SYehuda Sadeh void ceph_osdmap_destroy(struct ceph_osdmap *map)
5533d14c5d2SYehuda Sadeh {
5543d14c5d2SYehuda Sadeh 	dout("osdmap_destroy %p\n", map);
5553d14c5d2SYehuda Sadeh 	if (map->crush)
5563d14c5d2SYehuda Sadeh 		crush_destroy(map->crush);
5573d14c5d2SYehuda Sadeh 	while (!RB_EMPTY_ROOT(&map->pg_temp)) {
5583d14c5d2SYehuda Sadeh 		struct ceph_pg_mapping *pg =
5593d14c5d2SYehuda Sadeh 			rb_entry(rb_first(&map->pg_temp),
5603d14c5d2SYehuda Sadeh 				 struct ceph_pg_mapping, node);
5613d14c5d2SYehuda Sadeh 		rb_erase(&pg->node, &map->pg_temp);
5623d14c5d2SYehuda Sadeh 		kfree(pg);
5633d14c5d2SYehuda Sadeh 	}
5643d14c5d2SYehuda Sadeh 	while (!RB_EMPTY_ROOT(&map->pg_pools)) {
5653d14c5d2SYehuda Sadeh 		struct ceph_pg_pool_info *pi =
5663d14c5d2SYehuda Sadeh 			rb_entry(rb_first(&map->pg_pools),
5673d14c5d2SYehuda Sadeh 				 struct ceph_pg_pool_info, node);
5683d14c5d2SYehuda Sadeh 		__remove_pg_pool(&map->pg_pools, pi);
5693d14c5d2SYehuda Sadeh 	}
5703d14c5d2SYehuda Sadeh 	kfree(map->osd_state);
5713d14c5d2SYehuda Sadeh 	kfree(map->osd_weight);
5723d14c5d2SYehuda Sadeh 	kfree(map->osd_addr);
5733d14c5d2SYehuda Sadeh 	kfree(map);
5743d14c5d2SYehuda Sadeh }
5753d14c5d2SYehuda Sadeh 
5763d14c5d2SYehuda Sadeh /*
5773d14c5d2SYehuda Sadeh  * adjust max osd value.  reallocate arrays.
5783d14c5d2SYehuda Sadeh  */
5793d14c5d2SYehuda Sadeh static int osdmap_set_max_osd(struct ceph_osdmap *map, int max)
5803d14c5d2SYehuda Sadeh {
5813d14c5d2SYehuda Sadeh 	u8 *state;
5823d14c5d2SYehuda Sadeh 	struct ceph_entity_addr *addr;
5833d14c5d2SYehuda Sadeh 	u32 *weight;
5843d14c5d2SYehuda Sadeh 
5853d14c5d2SYehuda Sadeh 	state = kcalloc(max, sizeof(*state), GFP_NOFS);
5863d14c5d2SYehuda Sadeh 	addr = kcalloc(max, sizeof(*addr), GFP_NOFS);
5873d14c5d2SYehuda Sadeh 	weight = kcalloc(max, sizeof(*weight), GFP_NOFS);
5883d14c5d2SYehuda Sadeh 	if (state == NULL || addr == NULL || weight == NULL) {
5893d14c5d2SYehuda Sadeh 		kfree(state);
5903d14c5d2SYehuda Sadeh 		kfree(addr);
5913d14c5d2SYehuda Sadeh 		kfree(weight);
5923d14c5d2SYehuda Sadeh 		return -ENOMEM;
5933d14c5d2SYehuda Sadeh 	}
5943d14c5d2SYehuda Sadeh 
5953d14c5d2SYehuda Sadeh 	/* copy old? */
5963d14c5d2SYehuda Sadeh 	if (map->osd_state) {
5973d14c5d2SYehuda Sadeh 		memcpy(state, map->osd_state, map->max_osd*sizeof(*state));
5983d14c5d2SYehuda Sadeh 		memcpy(addr, map->osd_addr, map->max_osd*sizeof(*addr));
5993d14c5d2SYehuda Sadeh 		memcpy(weight, map->osd_weight, map->max_osd*sizeof(*weight));
6003d14c5d2SYehuda Sadeh 		kfree(map->osd_state);
6013d14c5d2SYehuda Sadeh 		kfree(map->osd_addr);
6023d14c5d2SYehuda Sadeh 		kfree(map->osd_weight);
6033d14c5d2SYehuda Sadeh 	}
6043d14c5d2SYehuda Sadeh 
6053d14c5d2SYehuda Sadeh 	map->osd_state = state;
6063d14c5d2SYehuda Sadeh 	map->osd_weight = weight;
6073d14c5d2SYehuda Sadeh 	map->osd_addr = addr;
6083d14c5d2SYehuda Sadeh 	map->max_osd = max;
6093d14c5d2SYehuda Sadeh 	return 0;
6103d14c5d2SYehuda Sadeh }
6113d14c5d2SYehuda Sadeh 
6123d14c5d2SYehuda Sadeh /*
6133d14c5d2SYehuda Sadeh  * decode a full map.
6143d14c5d2SYehuda Sadeh  */
6153d14c5d2SYehuda Sadeh struct ceph_osdmap *osdmap_decode(void **p, void *end)
6163d14c5d2SYehuda Sadeh {
6173d14c5d2SYehuda Sadeh 	struct ceph_osdmap *map;
6183d14c5d2SYehuda Sadeh 	u16 version;
6193d14c5d2SYehuda Sadeh 	u32 len, max, i;
6203d14c5d2SYehuda Sadeh 	u8 ev;
6213d14c5d2SYehuda Sadeh 	int err = -EINVAL;
6223d14c5d2SYehuda Sadeh 	void *start = *p;
6233d14c5d2SYehuda Sadeh 	struct ceph_pg_pool_info *pi;
6243d14c5d2SYehuda Sadeh 
6253d14c5d2SYehuda Sadeh 	dout("osdmap_decode %p to %p len %d\n", *p, end, (int)(end - *p));
6263d14c5d2SYehuda Sadeh 
6273d14c5d2SYehuda Sadeh 	map = kzalloc(sizeof(*map), GFP_NOFS);
6283d14c5d2SYehuda Sadeh 	if (map == NULL)
6293d14c5d2SYehuda Sadeh 		return ERR_PTR(-ENOMEM);
6303d14c5d2SYehuda Sadeh 	map->pg_temp = RB_ROOT;
6313d14c5d2SYehuda Sadeh 
6323d14c5d2SYehuda Sadeh 	ceph_decode_16_safe(p, end, version, bad);
6333d14c5d2SYehuda Sadeh 	if (version > CEPH_OSDMAP_VERSION) {
6343d14c5d2SYehuda Sadeh 		pr_warning("got unknown v %d > %d of osdmap\n", version,
6353d14c5d2SYehuda Sadeh 			   CEPH_OSDMAP_VERSION);
6363d14c5d2SYehuda Sadeh 		goto bad;
6373d14c5d2SYehuda Sadeh 	}
6383d14c5d2SYehuda Sadeh 
6393d14c5d2SYehuda Sadeh 	ceph_decode_need(p, end, 2*sizeof(u64)+6*sizeof(u32), bad);
6403d14c5d2SYehuda Sadeh 	ceph_decode_copy(p, &map->fsid, sizeof(map->fsid));
6413d14c5d2SYehuda Sadeh 	map->epoch = ceph_decode_32(p);
6423d14c5d2SYehuda Sadeh 	ceph_decode_copy(p, &map->created, sizeof(map->created));
6433d14c5d2SYehuda Sadeh 	ceph_decode_copy(p, &map->modified, sizeof(map->modified));
6443d14c5d2SYehuda Sadeh 
6453d14c5d2SYehuda Sadeh 	ceph_decode_32_safe(p, end, max, bad);
6463d14c5d2SYehuda Sadeh 	while (max--) {
6473d14c5d2SYehuda Sadeh 		ceph_decode_need(p, end, 4 + 1 + sizeof(pi->v), bad);
6480ed7285eSSage Weil 		err = -ENOMEM;
6493d14c5d2SYehuda Sadeh 		pi = kzalloc(sizeof(*pi), GFP_NOFS);
6503d14c5d2SYehuda Sadeh 		if (!pi)
6513d14c5d2SYehuda Sadeh 			goto bad;
6523d14c5d2SYehuda Sadeh 		pi->id = ceph_decode_32(p);
6530ed7285eSSage Weil 		err = -EINVAL;
6543d14c5d2SYehuda Sadeh 		ev = ceph_decode_8(p); /* encoding version */
6553d14c5d2SYehuda Sadeh 		if (ev > CEPH_PG_POOL_VERSION) {
6563d14c5d2SYehuda Sadeh 			pr_warning("got unknown v %d > %d of ceph_pg_pool\n",
6573d14c5d2SYehuda Sadeh 				   ev, CEPH_PG_POOL_VERSION);
6583d14c5d2SYehuda Sadeh 			kfree(pi);
6593d14c5d2SYehuda Sadeh 			goto bad;
6603d14c5d2SYehuda Sadeh 		}
6613d14c5d2SYehuda Sadeh 		err = __decode_pool(p, end, pi);
662b0aee351SJesper Juhl 		if (err < 0) {
663b0aee351SJesper Juhl 			kfree(pi);
6643d14c5d2SYehuda Sadeh 			goto bad;
665b0aee351SJesper Juhl 		}
6663d14c5d2SYehuda Sadeh 		__insert_pg_pool(&map->pg_pools, pi);
6673d14c5d2SYehuda Sadeh 	}
6683d14c5d2SYehuda Sadeh 
6690ed7285eSSage Weil 	if (version >= 5) {
6700ed7285eSSage Weil 		err = __decode_pool_names(p, end, map);
6710ed7285eSSage Weil 		if (err < 0) {
6720ed7285eSSage Weil 			dout("fail to decode pool names");
6733d14c5d2SYehuda Sadeh 			goto bad;
6740ed7285eSSage Weil 		}
6750ed7285eSSage Weil 	}
6763d14c5d2SYehuda Sadeh 
6773d14c5d2SYehuda Sadeh 	ceph_decode_32_safe(p, end, map->pool_max, bad);
6783d14c5d2SYehuda Sadeh 
6793d14c5d2SYehuda Sadeh 	ceph_decode_32_safe(p, end, map->flags, bad);
6803d14c5d2SYehuda Sadeh 
6813d14c5d2SYehuda Sadeh 	max = ceph_decode_32(p);
6823d14c5d2SYehuda Sadeh 
6833d14c5d2SYehuda Sadeh 	/* (re)alloc osd arrays */
6843d14c5d2SYehuda Sadeh 	err = osdmap_set_max_osd(map, max);
6853d14c5d2SYehuda Sadeh 	if (err < 0)
6863d14c5d2SYehuda Sadeh 		goto bad;
6873d14c5d2SYehuda Sadeh 	dout("osdmap_decode max_osd = %d\n", map->max_osd);
6883d14c5d2SYehuda Sadeh 
6893d14c5d2SYehuda Sadeh 	/* osds */
6903d14c5d2SYehuda Sadeh 	err = -EINVAL;
6913d14c5d2SYehuda Sadeh 	ceph_decode_need(p, end, 3*sizeof(u32) +
6923d14c5d2SYehuda Sadeh 			 map->max_osd*(1 + sizeof(*map->osd_weight) +
6933d14c5d2SYehuda Sadeh 				       sizeof(*map->osd_addr)), bad);
6943d14c5d2SYehuda Sadeh 	*p += 4; /* skip length field (should match max) */
6953d14c5d2SYehuda Sadeh 	ceph_decode_copy(p, map->osd_state, map->max_osd);
6963d14c5d2SYehuda Sadeh 
6973d14c5d2SYehuda Sadeh 	*p += 4; /* skip length field (should match max) */
6983d14c5d2SYehuda Sadeh 	for (i = 0; i < map->max_osd; i++)
6993d14c5d2SYehuda Sadeh 		map->osd_weight[i] = ceph_decode_32(p);
7003d14c5d2SYehuda Sadeh 
7013d14c5d2SYehuda Sadeh 	*p += 4; /* skip length field (should match max) */
7023d14c5d2SYehuda Sadeh 	ceph_decode_copy(p, map->osd_addr, map->max_osd*sizeof(*map->osd_addr));
7033d14c5d2SYehuda Sadeh 	for (i = 0; i < map->max_osd; i++)
7043d14c5d2SYehuda Sadeh 		ceph_decode_addr(&map->osd_addr[i]);
7053d14c5d2SYehuda Sadeh 
7063d14c5d2SYehuda Sadeh 	/* pg_temp */
7073d14c5d2SYehuda Sadeh 	ceph_decode_32_safe(p, end, len, bad);
7083d14c5d2SYehuda Sadeh 	for (i = 0; i < len; i++) {
7093d14c5d2SYehuda Sadeh 		int n, j;
7103d14c5d2SYehuda Sadeh 		struct ceph_pg pgid;
7113d14c5d2SYehuda Sadeh 		struct ceph_pg_mapping *pg;
7123d14c5d2SYehuda Sadeh 
7133d14c5d2SYehuda Sadeh 		ceph_decode_need(p, end, sizeof(u32) + sizeof(u64), bad);
7143d14c5d2SYehuda Sadeh 		ceph_decode_copy(p, &pgid, sizeof(pgid));
7153d14c5d2SYehuda Sadeh 		n = ceph_decode_32(p);
716e91a9b63SXi Wang 		err = -EINVAL;
717e91a9b63SXi Wang 		if (n > (UINT_MAX - sizeof(*pg)) / sizeof(u32))
718e91a9b63SXi Wang 			goto bad;
7193d14c5d2SYehuda Sadeh 		ceph_decode_need(p, end, n * sizeof(u32), bad);
7203d14c5d2SYehuda Sadeh 		err = -ENOMEM;
7213d14c5d2SYehuda Sadeh 		pg = kmalloc(sizeof(*pg) + n*sizeof(u32), GFP_NOFS);
7223d14c5d2SYehuda Sadeh 		if (!pg)
7233d14c5d2SYehuda Sadeh 			goto bad;
7243d14c5d2SYehuda Sadeh 		pg->pgid = pgid;
7253d14c5d2SYehuda Sadeh 		pg->len = n;
7263d14c5d2SYehuda Sadeh 		for (j = 0; j < n; j++)
7273d14c5d2SYehuda Sadeh 			pg->osds[j] = ceph_decode_32(p);
7283d14c5d2SYehuda Sadeh 
7293d14c5d2SYehuda Sadeh 		err = __insert_pg_mapping(pg, &map->pg_temp);
7303d14c5d2SYehuda Sadeh 		if (err)
7313d14c5d2SYehuda Sadeh 			goto bad;
7323d14c5d2SYehuda Sadeh 		dout(" added pg_temp %llx len %d\n", *(u64 *)&pgid, len);
7333d14c5d2SYehuda Sadeh 	}
7343d14c5d2SYehuda Sadeh 
7353d14c5d2SYehuda Sadeh 	/* crush */
7363d14c5d2SYehuda Sadeh 	ceph_decode_32_safe(p, end, len, bad);
7373d14c5d2SYehuda Sadeh 	dout("osdmap_decode crush len %d from off 0x%x\n", len,
7383d14c5d2SYehuda Sadeh 	     (int)(*p - start));
7393d14c5d2SYehuda Sadeh 	ceph_decode_need(p, end, len, bad);
7403d14c5d2SYehuda Sadeh 	map->crush = crush_decode(*p, end);
7413d14c5d2SYehuda Sadeh 	*p += len;
7423d14c5d2SYehuda Sadeh 	if (IS_ERR(map->crush)) {
7433d14c5d2SYehuda Sadeh 		err = PTR_ERR(map->crush);
7443d14c5d2SYehuda Sadeh 		map->crush = NULL;
7453d14c5d2SYehuda Sadeh 		goto bad;
7463d14c5d2SYehuda Sadeh 	}
7473d14c5d2SYehuda Sadeh 
7483d14c5d2SYehuda Sadeh 	/* ignore the rest of the map */
7493d14c5d2SYehuda Sadeh 	*p = end;
7503d14c5d2SYehuda Sadeh 
7513d14c5d2SYehuda Sadeh 	dout("osdmap_decode done %p %p\n", *p, end);
7523d14c5d2SYehuda Sadeh 	return map;
7533d14c5d2SYehuda Sadeh 
7543d14c5d2SYehuda Sadeh bad:
7550ed7285eSSage Weil 	dout("osdmap_decode fail err %d\n", err);
7563d14c5d2SYehuda Sadeh 	ceph_osdmap_destroy(map);
7573d14c5d2SYehuda Sadeh 	return ERR_PTR(err);
7583d14c5d2SYehuda Sadeh }
7593d14c5d2SYehuda Sadeh 
7603d14c5d2SYehuda Sadeh /*
7613d14c5d2SYehuda Sadeh  * decode and apply an incremental map update.
7623d14c5d2SYehuda Sadeh  */
7633d14c5d2SYehuda Sadeh struct ceph_osdmap *osdmap_apply_incremental(void **p, void *end,
7643d14c5d2SYehuda Sadeh 					     struct ceph_osdmap *map,
7653d14c5d2SYehuda Sadeh 					     struct ceph_messenger *msgr)
7663d14c5d2SYehuda Sadeh {
7673d14c5d2SYehuda Sadeh 	struct crush_map *newcrush = NULL;
7683d14c5d2SYehuda Sadeh 	struct ceph_fsid fsid;
7693d14c5d2SYehuda Sadeh 	u32 epoch = 0;
7703d14c5d2SYehuda Sadeh 	struct ceph_timespec modified;
7713d14c5d2SYehuda Sadeh 	u32 len, pool;
7723d14c5d2SYehuda Sadeh 	__s32 new_pool_max, new_flags, max;
7733d14c5d2SYehuda Sadeh 	void *start = *p;
7743d14c5d2SYehuda Sadeh 	int err = -EINVAL;
7753d14c5d2SYehuda Sadeh 	u16 version;
7763d14c5d2SYehuda Sadeh 
7773d14c5d2SYehuda Sadeh 	ceph_decode_16_safe(p, end, version, bad);
7783d14c5d2SYehuda Sadeh 	if (version > CEPH_OSDMAP_INC_VERSION) {
7793d14c5d2SYehuda Sadeh 		pr_warning("got unknown v %d > %d of inc osdmap\n", version,
7803d14c5d2SYehuda Sadeh 			   CEPH_OSDMAP_INC_VERSION);
7813d14c5d2SYehuda Sadeh 		goto bad;
7823d14c5d2SYehuda Sadeh 	}
7833d14c5d2SYehuda Sadeh 
7843d14c5d2SYehuda Sadeh 	ceph_decode_need(p, end, sizeof(fsid)+sizeof(modified)+2*sizeof(u32),
7853d14c5d2SYehuda Sadeh 			 bad);
7863d14c5d2SYehuda Sadeh 	ceph_decode_copy(p, &fsid, sizeof(fsid));
7873d14c5d2SYehuda Sadeh 	epoch = ceph_decode_32(p);
7883d14c5d2SYehuda Sadeh 	BUG_ON(epoch != map->epoch+1);
7893d14c5d2SYehuda Sadeh 	ceph_decode_copy(p, &modified, sizeof(modified));
7903d14c5d2SYehuda Sadeh 	new_pool_max = ceph_decode_32(p);
7913d14c5d2SYehuda Sadeh 	new_flags = ceph_decode_32(p);
7923d14c5d2SYehuda Sadeh 
7933d14c5d2SYehuda Sadeh 	/* full map? */
7943d14c5d2SYehuda Sadeh 	ceph_decode_32_safe(p, end, len, bad);
7953d14c5d2SYehuda Sadeh 	if (len > 0) {
7963d14c5d2SYehuda Sadeh 		dout("apply_incremental full map len %d, %p to %p\n",
7973d14c5d2SYehuda Sadeh 		     len, *p, end);
7983d14c5d2SYehuda Sadeh 		return osdmap_decode(p, min(*p+len, end));
7993d14c5d2SYehuda Sadeh 	}
8003d14c5d2SYehuda Sadeh 
8013d14c5d2SYehuda Sadeh 	/* new crush? */
8023d14c5d2SYehuda Sadeh 	ceph_decode_32_safe(p, end, len, bad);
8033d14c5d2SYehuda Sadeh 	if (len > 0) {
8043d14c5d2SYehuda Sadeh 		dout("apply_incremental new crush map len %d, %p to %p\n",
8053d14c5d2SYehuda Sadeh 		     len, *p, end);
8063d14c5d2SYehuda Sadeh 		newcrush = crush_decode(*p, min(*p+len, end));
8073d14c5d2SYehuda Sadeh 		if (IS_ERR(newcrush))
8083d14c5d2SYehuda Sadeh 			return ERR_CAST(newcrush);
8093d14c5d2SYehuda Sadeh 		*p += len;
8103d14c5d2SYehuda Sadeh 	}
8113d14c5d2SYehuda Sadeh 
8123d14c5d2SYehuda Sadeh 	/* new flags? */
8133d14c5d2SYehuda Sadeh 	if (new_flags >= 0)
8143d14c5d2SYehuda Sadeh 		map->flags = new_flags;
8153d14c5d2SYehuda Sadeh 	if (new_pool_max >= 0)
8163d14c5d2SYehuda Sadeh 		map->pool_max = new_pool_max;
8173d14c5d2SYehuda Sadeh 
8183d14c5d2SYehuda Sadeh 	ceph_decode_need(p, end, 5*sizeof(u32), bad);
8193d14c5d2SYehuda Sadeh 
8203d14c5d2SYehuda Sadeh 	/* new max? */
8213d14c5d2SYehuda Sadeh 	max = ceph_decode_32(p);
8223d14c5d2SYehuda Sadeh 	if (max >= 0) {
8233d14c5d2SYehuda Sadeh 		err = osdmap_set_max_osd(map, max);
8243d14c5d2SYehuda Sadeh 		if (err < 0)
8253d14c5d2SYehuda Sadeh 			goto bad;
8263d14c5d2SYehuda Sadeh 	}
8273d14c5d2SYehuda Sadeh 
8283d14c5d2SYehuda Sadeh 	map->epoch++;
82931456665SSage Weil 	map->modified = modified;
8303d14c5d2SYehuda Sadeh 	if (newcrush) {
8313d14c5d2SYehuda Sadeh 		if (map->crush)
8323d14c5d2SYehuda Sadeh 			crush_destroy(map->crush);
8333d14c5d2SYehuda Sadeh 		map->crush = newcrush;
8343d14c5d2SYehuda Sadeh 		newcrush = NULL;
8353d14c5d2SYehuda Sadeh 	}
8363d14c5d2SYehuda Sadeh 
8373d14c5d2SYehuda Sadeh 	/* new_pool */
8383d14c5d2SYehuda Sadeh 	ceph_decode_32_safe(p, end, len, bad);
8393d14c5d2SYehuda Sadeh 	while (len--) {
8403d14c5d2SYehuda Sadeh 		__u8 ev;
8413d14c5d2SYehuda Sadeh 		struct ceph_pg_pool_info *pi;
8423d14c5d2SYehuda Sadeh 
8433d14c5d2SYehuda Sadeh 		ceph_decode_32_safe(p, end, pool, bad);
8443d14c5d2SYehuda Sadeh 		ceph_decode_need(p, end, 1 + sizeof(pi->v), bad);
8453d14c5d2SYehuda Sadeh 		ev = ceph_decode_8(p);  /* encoding version */
8463d14c5d2SYehuda Sadeh 		if (ev > CEPH_PG_POOL_VERSION) {
8473d14c5d2SYehuda Sadeh 			pr_warning("got unknown v %d > %d of ceph_pg_pool\n",
8483d14c5d2SYehuda Sadeh 				   ev, CEPH_PG_POOL_VERSION);
8490ed7285eSSage Weil 			err = -EINVAL;
8503d14c5d2SYehuda Sadeh 			goto bad;
8513d14c5d2SYehuda Sadeh 		}
8523d14c5d2SYehuda Sadeh 		pi = __lookup_pg_pool(&map->pg_pools, pool);
8533d14c5d2SYehuda Sadeh 		if (!pi) {
8543d14c5d2SYehuda Sadeh 			pi = kzalloc(sizeof(*pi), GFP_NOFS);
8553d14c5d2SYehuda Sadeh 			if (!pi) {
8563d14c5d2SYehuda Sadeh 				err = -ENOMEM;
8573d14c5d2SYehuda Sadeh 				goto bad;
8583d14c5d2SYehuda Sadeh 			}
8593d14c5d2SYehuda Sadeh 			pi->id = pool;
8603d14c5d2SYehuda Sadeh 			__insert_pg_pool(&map->pg_pools, pi);
8613d14c5d2SYehuda Sadeh 		}
8623d14c5d2SYehuda Sadeh 		err = __decode_pool(p, end, pi);
8633d14c5d2SYehuda Sadeh 		if (err < 0)
8643d14c5d2SYehuda Sadeh 			goto bad;
8653d14c5d2SYehuda Sadeh 	}
8660ed7285eSSage Weil 	if (version >= 5) {
8670ed7285eSSage Weil 		err = __decode_pool_names(p, end, map);
8680ed7285eSSage Weil 		if (err < 0)
8693d14c5d2SYehuda Sadeh 			goto bad;
8700ed7285eSSage Weil 	}
8713d14c5d2SYehuda Sadeh 
8723d14c5d2SYehuda Sadeh 	/* old_pool */
8733d14c5d2SYehuda Sadeh 	ceph_decode_32_safe(p, end, len, bad);
8743d14c5d2SYehuda Sadeh 	while (len--) {
8753d14c5d2SYehuda Sadeh 		struct ceph_pg_pool_info *pi;
8763d14c5d2SYehuda Sadeh 
8773d14c5d2SYehuda Sadeh 		ceph_decode_32_safe(p, end, pool, bad);
8783d14c5d2SYehuda Sadeh 		pi = __lookup_pg_pool(&map->pg_pools, pool);
8793d14c5d2SYehuda Sadeh 		if (pi)
8803d14c5d2SYehuda Sadeh 			__remove_pg_pool(&map->pg_pools, pi);
8813d14c5d2SYehuda Sadeh 	}
8823d14c5d2SYehuda Sadeh 
8833d14c5d2SYehuda Sadeh 	/* new_up */
8843d14c5d2SYehuda Sadeh 	err = -EINVAL;
8853d14c5d2SYehuda Sadeh 	ceph_decode_32_safe(p, end, len, bad);
8863d14c5d2SYehuda Sadeh 	while (len--) {
8873d14c5d2SYehuda Sadeh 		u32 osd;
8883d14c5d2SYehuda Sadeh 		struct ceph_entity_addr addr;
8893d14c5d2SYehuda Sadeh 		ceph_decode_32_safe(p, end, osd, bad);
8903d14c5d2SYehuda Sadeh 		ceph_decode_copy_safe(p, end, &addr, sizeof(addr), bad);
8913d14c5d2SYehuda Sadeh 		ceph_decode_addr(&addr);
8923d14c5d2SYehuda Sadeh 		pr_info("osd%d up\n", osd);
8933d14c5d2SYehuda Sadeh 		BUG_ON(osd >= map->max_osd);
8943d14c5d2SYehuda Sadeh 		map->osd_state[osd] |= CEPH_OSD_UP;
8953d14c5d2SYehuda Sadeh 		map->osd_addr[osd] = addr;
8963d14c5d2SYehuda Sadeh 	}
8973d14c5d2SYehuda Sadeh 
8987662d8ffSSage Weil 	/* new_state */
8993d14c5d2SYehuda Sadeh 	ceph_decode_32_safe(p, end, len, bad);
9003d14c5d2SYehuda Sadeh 	while (len--) {
9013d14c5d2SYehuda Sadeh 		u32 osd;
9027662d8ffSSage Weil 		u8 xorstate;
9033d14c5d2SYehuda Sadeh 		ceph_decode_32_safe(p, end, osd, bad);
9047662d8ffSSage Weil 		xorstate = **(u8 **)p;
9053d14c5d2SYehuda Sadeh 		(*p)++;  /* clean flag */
9067662d8ffSSage Weil 		if (xorstate == 0)
9077662d8ffSSage Weil 			xorstate = CEPH_OSD_UP;
9087662d8ffSSage Weil 		if (xorstate & CEPH_OSD_UP)
9093d14c5d2SYehuda Sadeh 			pr_info("osd%d down\n", osd);
9103d14c5d2SYehuda Sadeh 		if (osd < map->max_osd)
9117662d8ffSSage Weil 			map->osd_state[osd] ^= xorstate;
9123d14c5d2SYehuda Sadeh 	}
9133d14c5d2SYehuda Sadeh 
9143d14c5d2SYehuda Sadeh 	/* new_weight */
9153d14c5d2SYehuda Sadeh 	ceph_decode_32_safe(p, end, len, bad);
9163d14c5d2SYehuda Sadeh 	while (len--) {
9173d14c5d2SYehuda Sadeh 		u32 osd, off;
9183d14c5d2SYehuda Sadeh 		ceph_decode_need(p, end, sizeof(u32)*2, bad);
9193d14c5d2SYehuda Sadeh 		osd = ceph_decode_32(p);
9203d14c5d2SYehuda Sadeh 		off = ceph_decode_32(p);
9213d14c5d2SYehuda Sadeh 		pr_info("osd%d weight 0x%x %s\n", osd, off,
9223d14c5d2SYehuda Sadeh 		     off == CEPH_OSD_IN ? "(in)" :
9233d14c5d2SYehuda Sadeh 		     (off == CEPH_OSD_OUT ? "(out)" : ""));
9243d14c5d2SYehuda Sadeh 		if (osd < map->max_osd)
9253d14c5d2SYehuda Sadeh 			map->osd_weight[osd] = off;
9263d14c5d2SYehuda Sadeh 	}
9273d14c5d2SYehuda Sadeh 
9283d14c5d2SYehuda Sadeh 	/* new_pg_temp */
9293d14c5d2SYehuda Sadeh 	ceph_decode_32_safe(p, end, len, bad);
9303d14c5d2SYehuda Sadeh 	while (len--) {
9313d14c5d2SYehuda Sadeh 		struct ceph_pg_mapping *pg;
9323d14c5d2SYehuda Sadeh 		int j;
9333d14c5d2SYehuda Sadeh 		struct ceph_pg pgid;
9343d14c5d2SYehuda Sadeh 		u32 pglen;
9353d14c5d2SYehuda Sadeh 		ceph_decode_need(p, end, sizeof(u64) + sizeof(u32), bad);
9363d14c5d2SYehuda Sadeh 		ceph_decode_copy(p, &pgid, sizeof(pgid));
9373d14c5d2SYehuda Sadeh 		pglen = ceph_decode_32(p);
9383d14c5d2SYehuda Sadeh 
9393d14c5d2SYehuda Sadeh 		if (pglen) {
9403d14c5d2SYehuda Sadeh 			ceph_decode_need(p, end, pglen*sizeof(u32), bad);
9416bd9adbdSSage Weil 
9426bd9adbdSSage Weil 			/* removing existing (if any) */
9436bd9adbdSSage Weil 			(void) __remove_pg_mapping(&map->pg_temp, pgid);
9446bd9adbdSSage Weil 
9456bd9adbdSSage Weil 			/* insert */
946a5506049SXi Wang 			err = -EINVAL;
9470ed7285eSSage Weil 			if (pglen > (UINT_MAX - sizeof(*pg)) / sizeof(u32))
948a5506049SXi Wang 				goto bad;
9493d14c5d2SYehuda Sadeh 			err = -ENOMEM;
9500ed7285eSSage Weil 			pg = kmalloc(sizeof(*pg) + sizeof(u32)*pglen, GFP_NOFS);
9510ed7285eSSage Weil 			if (!pg)
9523d14c5d2SYehuda Sadeh 				goto bad;
9533d14c5d2SYehuda Sadeh 			pg->pgid = pgid;
9543d14c5d2SYehuda Sadeh 			pg->len = pglen;
9553d14c5d2SYehuda Sadeh 			for (j = 0; j < pglen; j++)
9563d14c5d2SYehuda Sadeh 				pg->osds[j] = ceph_decode_32(p);
9573d14c5d2SYehuda Sadeh 			err = __insert_pg_mapping(pg, &map->pg_temp);
9583d14c5d2SYehuda Sadeh 			if (err) {
9593d14c5d2SYehuda Sadeh 				kfree(pg);
9603d14c5d2SYehuda Sadeh 				goto bad;
9613d14c5d2SYehuda Sadeh 			}
9623d14c5d2SYehuda Sadeh 			dout(" added pg_temp %llx len %d\n", *(u64 *)&pgid,
9633d14c5d2SYehuda Sadeh 			     pglen);
9648adc8b3dSSage Weil 		} else {
9658adc8b3dSSage Weil 			/* remove */
9668adc8b3dSSage Weil 			__remove_pg_mapping(&map->pg_temp, pgid);
9673d14c5d2SYehuda Sadeh 		}
9683d14c5d2SYehuda Sadeh 	}
9693d14c5d2SYehuda Sadeh 
9703d14c5d2SYehuda Sadeh 	/* ignore the rest */
9713d14c5d2SYehuda Sadeh 	*p = end;
9723d14c5d2SYehuda Sadeh 	return map;
9733d14c5d2SYehuda Sadeh 
9743d14c5d2SYehuda Sadeh bad:
9753d14c5d2SYehuda Sadeh 	pr_err("corrupt inc osdmap epoch %d off %d (%p of %p-%p)\n",
9763d14c5d2SYehuda Sadeh 	       epoch, (int)(*p - start), *p, start, end);
9773d14c5d2SYehuda Sadeh 	print_hex_dump(KERN_DEBUG, "osdmap: ",
9783d14c5d2SYehuda Sadeh 		       DUMP_PREFIX_OFFSET, 16, 1,
9793d14c5d2SYehuda Sadeh 		       start, end - start, true);
9803d14c5d2SYehuda Sadeh 	if (newcrush)
9813d14c5d2SYehuda Sadeh 		crush_destroy(newcrush);
9823d14c5d2SYehuda Sadeh 	return ERR_PTR(err);
9833d14c5d2SYehuda Sadeh }
9843d14c5d2SYehuda Sadeh 
9853d14c5d2SYehuda Sadeh 
9863d14c5d2SYehuda Sadeh 
9873d14c5d2SYehuda Sadeh 
9883d14c5d2SYehuda Sadeh /*
9893d14c5d2SYehuda Sadeh  * calculate file layout from given offset, length.
9903d14c5d2SYehuda Sadeh  * fill in correct oid, logical length, and object extent
9913d14c5d2SYehuda Sadeh  * offset, length.
9923d14c5d2SYehuda Sadeh  *
9933d14c5d2SYehuda Sadeh  * for now, we write only a single su, until we can
9943d14c5d2SYehuda Sadeh  * pass a stride back to the caller.
9953d14c5d2SYehuda Sadeh  */
996d63b77f4SSage Weil int ceph_calc_file_object_mapping(struct ceph_file_layout *layout,
9973d14c5d2SYehuda Sadeh 				   u64 off, u64 *plen,
9983d14c5d2SYehuda Sadeh 				   u64 *ono,
9993d14c5d2SYehuda Sadeh 				   u64 *oxoff, u64 *oxlen)
10003d14c5d2SYehuda Sadeh {
10013d14c5d2SYehuda Sadeh 	u32 osize = le32_to_cpu(layout->fl_object_size);
10023d14c5d2SYehuda Sadeh 	u32 su = le32_to_cpu(layout->fl_stripe_unit);
10033d14c5d2SYehuda Sadeh 	u32 sc = le32_to_cpu(layout->fl_stripe_count);
10043d14c5d2SYehuda Sadeh 	u32 bl, stripeno, stripepos, objsetno;
10053d14c5d2SYehuda Sadeh 	u32 su_per_object;
10063d14c5d2SYehuda Sadeh 	u64 t, su_offset;
10073d14c5d2SYehuda Sadeh 
10083d14c5d2SYehuda Sadeh 	dout("mapping %llu~%llu  osize %u fl_su %u\n", off, *plen,
10093d14c5d2SYehuda Sadeh 	     osize, su);
1010d63b77f4SSage Weil 	if (su == 0 || sc == 0)
1011d63b77f4SSage Weil 		goto invalid;
10123d14c5d2SYehuda Sadeh 	su_per_object = osize / su;
1013d63b77f4SSage Weil 	if (su_per_object == 0)
1014d63b77f4SSage Weil 		goto invalid;
10153d14c5d2SYehuda Sadeh 	dout("osize %u / su %u = su_per_object %u\n", osize, su,
10163d14c5d2SYehuda Sadeh 	     su_per_object);
10173d14c5d2SYehuda Sadeh 
1018d63b77f4SSage Weil 	if ((su & ~PAGE_MASK) != 0)
1019d63b77f4SSage Weil 		goto invalid;
1020d63b77f4SSage Weil 
10213d14c5d2SYehuda Sadeh 	/* bl = *off / su; */
10223d14c5d2SYehuda Sadeh 	t = off;
10233d14c5d2SYehuda Sadeh 	do_div(t, su);
10243d14c5d2SYehuda Sadeh 	bl = t;
10253d14c5d2SYehuda Sadeh 	dout("off %llu / su %u = bl %u\n", off, su, bl);
10263d14c5d2SYehuda Sadeh 
10273d14c5d2SYehuda Sadeh 	stripeno = bl / sc;
10283d14c5d2SYehuda Sadeh 	stripepos = bl % sc;
10293d14c5d2SYehuda Sadeh 	objsetno = stripeno / su_per_object;
10303d14c5d2SYehuda Sadeh 
10313d14c5d2SYehuda Sadeh 	*ono = objsetno * sc + stripepos;
103295c96174SEric Dumazet 	dout("objset %u * sc %u = ono %u\n", objsetno, sc, (unsigned int)*ono);
10333d14c5d2SYehuda Sadeh 
10343d14c5d2SYehuda Sadeh 	/* *oxoff = *off % layout->fl_stripe_unit;  # offset in su */
10353d14c5d2SYehuda Sadeh 	t = off;
10363d14c5d2SYehuda Sadeh 	su_offset = do_div(t, su);
10373d14c5d2SYehuda Sadeh 	*oxoff = su_offset + (stripeno % su_per_object) * su;
10383d14c5d2SYehuda Sadeh 
10393d14c5d2SYehuda Sadeh 	/*
10403d14c5d2SYehuda Sadeh 	 * Calculate the length of the extent being written to the selected
10413d14c5d2SYehuda Sadeh 	 * object. This is the minimum of the full length requested (plen) or
10423d14c5d2SYehuda Sadeh 	 * the remainder of the current stripe being written to.
10433d14c5d2SYehuda Sadeh 	 */
10443d14c5d2SYehuda Sadeh 	*oxlen = min_t(u64, *plen, su - su_offset);
10453d14c5d2SYehuda Sadeh 	*plen = *oxlen;
10463d14c5d2SYehuda Sadeh 
10473d14c5d2SYehuda Sadeh 	dout(" obj extent %llu~%llu\n", *oxoff, *oxlen);
1048d63b77f4SSage Weil 	return 0;
1049d63b77f4SSage Weil 
1050d63b77f4SSage Weil invalid:
1051d63b77f4SSage Weil 	dout(" invalid layout\n");
1052d63b77f4SSage Weil 	*ono = 0;
1053d63b77f4SSage Weil 	*oxoff = 0;
1054d63b77f4SSage Weil 	*oxlen = 0;
1055d63b77f4SSage Weil 	return -EINVAL;
10563d14c5d2SYehuda Sadeh }
10573d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_calc_file_object_mapping);
10583d14c5d2SYehuda Sadeh 
10593d14c5d2SYehuda Sadeh /*
10603d14c5d2SYehuda Sadeh  * calculate an object layout (i.e. pgid) from an oid,
10613d14c5d2SYehuda Sadeh  * file_layout, and osdmap
10623d14c5d2SYehuda Sadeh  */
10633d14c5d2SYehuda Sadeh int ceph_calc_object_layout(struct ceph_object_layout *ol,
10643d14c5d2SYehuda Sadeh 			    const char *oid,
10653d14c5d2SYehuda Sadeh 			    struct ceph_file_layout *fl,
10663d14c5d2SYehuda Sadeh 			    struct ceph_osdmap *osdmap)
10673d14c5d2SYehuda Sadeh {
106895c96174SEric Dumazet 	unsigned int num, num_mask;
10693d14c5d2SYehuda Sadeh 	struct ceph_pg pgid;
10703d14c5d2SYehuda Sadeh 	int poolid = le32_to_cpu(fl->fl_pg_pool);
10713d14c5d2SYehuda Sadeh 	struct ceph_pg_pool_info *pool;
107295c96174SEric Dumazet 	unsigned int ps;
10733d14c5d2SYehuda Sadeh 
10743d14c5d2SYehuda Sadeh 	BUG_ON(!osdmap);
10753d14c5d2SYehuda Sadeh 
10763d14c5d2SYehuda Sadeh 	pool = __lookup_pg_pool(&osdmap->pg_pools, poolid);
10773d14c5d2SYehuda Sadeh 	if (!pool)
10783d14c5d2SYehuda Sadeh 		return -EIO;
10793d14c5d2SYehuda Sadeh 	ps = ceph_str_hash(pool->v.object_hash, oid, strlen(oid));
10803d14c5d2SYehuda Sadeh 	num = le32_to_cpu(pool->v.pg_num);
10813d14c5d2SYehuda Sadeh 	num_mask = pool->pg_num_mask;
10823d14c5d2SYehuda Sadeh 
10833d14c5d2SYehuda Sadeh 	pgid.ps = cpu_to_le16(ps);
10843469ac1aSSage Weil 	pgid.preferred = cpu_to_le16(-1);
10853d14c5d2SYehuda Sadeh 	pgid.pool = fl->fl_pg_pool;
10863d14c5d2SYehuda Sadeh 	dout("calc_object_layout '%s' pgid %d.%x\n", oid, poolid, ps);
10873d14c5d2SYehuda Sadeh 
10883d14c5d2SYehuda Sadeh 	ol->ol_pgid = pgid;
10893d14c5d2SYehuda Sadeh 	ol->ol_stripe_unit = fl->fl_object_stripe_unit;
10903d14c5d2SYehuda Sadeh 	return 0;
10913d14c5d2SYehuda Sadeh }
10923d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_calc_object_layout);
10933d14c5d2SYehuda Sadeh 
10943d14c5d2SYehuda Sadeh /*
10953d14c5d2SYehuda Sadeh  * Calculate raw osd vector for the given pgid.  Return pointer to osd
10963d14c5d2SYehuda Sadeh  * array, or NULL on failure.
10973d14c5d2SYehuda Sadeh  */
10983d14c5d2SYehuda Sadeh static int *calc_pg_raw(struct ceph_osdmap *osdmap, struct ceph_pg pgid,
10993d14c5d2SYehuda Sadeh 			int *osds, int *num)
11003d14c5d2SYehuda Sadeh {
11013d14c5d2SYehuda Sadeh 	struct ceph_pg_mapping *pg;
11023d14c5d2SYehuda Sadeh 	struct ceph_pg_pool_info *pool;
11033d14c5d2SYehuda Sadeh 	int ruleno;
1104af56e0aaSLinus Torvalds 	unsigned int poolid, ps, pps, t, r;
11053d14c5d2SYehuda Sadeh 
1106782e182eSSage Weil 	poolid = le32_to_cpu(pgid.pool);
1107782e182eSSage Weil 	ps = le16_to_cpu(pgid.ps);
1108782e182eSSage Weil 
1109782e182eSSage Weil 	pool = __lookup_pg_pool(&osdmap->pg_pools, poolid);
1110782e182eSSage Weil 	if (!pool)
1111782e182eSSage Weil 		return NULL;
1112782e182eSSage Weil 
11133d14c5d2SYehuda Sadeh 	/* pg_temp? */
1114782e182eSSage Weil 	t = ceph_stable_mod(ps, le32_to_cpu(pool->v.pg_num),
1115782e182eSSage Weil 			    pool->pgp_num_mask);
1116782e182eSSage Weil 	pgid.ps = cpu_to_le16(t);
11173d14c5d2SYehuda Sadeh 	pg = __lookup_pg_mapping(&osdmap->pg_temp, pgid);
11183d14c5d2SYehuda Sadeh 	if (pg) {
11193d14c5d2SYehuda Sadeh 		*num = pg->len;
11203d14c5d2SYehuda Sadeh 		return pg->osds;
11213d14c5d2SYehuda Sadeh 	}
11223d14c5d2SYehuda Sadeh 
11233d14c5d2SYehuda Sadeh 	/* crush */
11243d14c5d2SYehuda Sadeh 	ruleno = crush_find_rule(osdmap->crush, pool->v.crush_ruleset,
11253d14c5d2SYehuda Sadeh 				 pool->v.type, pool->v.size);
11263d14c5d2SYehuda Sadeh 	if (ruleno < 0) {
11273d14c5d2SYehuda Sadeh 		pr_err("no crush rule pool %d ruleset %d type %d size %d\n",
11283d14c5d2SYehuda Sadeh 		       poolid, pool->v.crush_ruleset, pool->v.type,
11293d14c5d2SYehuda Sadeh 		       pool->v.size);
11303d14c5d2SYehuda Sadeh 		return NULL;
11313d14c5d2SYehuda Sadeh 	}
11323d14c5d2SYehuda Sadeh 
11333d14c5d2SYehuda Sadeh 	pps = ceph_stable_mod(ps,
11343d14c5d2SYehuda Sadeh 			      le32_to_cpu(pool->v.pgp_num),
11353d14c5d2SYehuda Sadeh 			      pool->pgp_num_mask);
11363d14c5d2SYehuda Sadeh 	pps += poolid;
11378b393269SSage Weil 	r = crush_do_rule(osdmap->crush, ruleno, pps, osds,
11383d14c5d2SYehuda Sadeh 			  min_t(int, pool->v.size, *num),
113941ebcc09SSage Weil 			  osdmap->osd_weight);
11408b393269SSage Weil 	if (r < 0) {
11418b393269SSage Weil 		pr_err("error %d from crush rule: pool %d ruleset %d type %d"
11428b393269SSage Weil 		       " size %d\n", r, poolid, pool->v.crush_ruleset,
11438b393269SSage Weil 		       pool->v.type, pool->v.size);
11448b393269SSage Weil 		return NULL;
11458b393269SSage Weil 	}
11468b393269SSage Weil 	*num = r;
11473d14c5d2SYehuda Sadeh 	return osds;
11483d14c5d2SYehuda Sadeh }
11493d14c5d2SYehuda Sadeh 
11503d14c5d2SYehuda Sadeh /*
11513d14c5d2SYehuda Sadeh  * Return acting set for given pgid.
11523d14c5d2SYehuda Sadeh  */
11533d14c5d2SYehuda Sadeh int ceph_calc_pg_acting(struct ceph_osdmap *osdmap, struct ceph_pg pgid,
11543d14c5d2SYehuda Sadeh 			int *acting)
11553d14c5d2SYehuda Sadeh {
11563d14c5d2SYehuda Sadeh 	int rawosds[CEPH_PG_MAX_SIZE], *osds;
11573d14c5d2SYehuda Sadeh 	int i, o, num = CEPH_PG_MAX_SIZE;
11583d14c5d2SYehuda Sadeh 
11593d14c5d2SYehuda Sadeh 	osds = calc_pg_raw(osdmap, pgid, rawosds, &num);
11603d14c5d2SYehuda Sadeh 	if (!osds)
11613d14c5d2SYehuda Sadeh 		return -1;
11623d14c5d2SYehuda Sadeh 
11633d14c5d2SYehuda Sadeh 	/* primary is first up osd */
11643d14c5d2SYehuda Sadeh 	o = 0;
11653d14c5d2SYehuda Sadeh 	for (i = 0; i < num; i++)
11663d14c5d2SYehuda Sadeh 		if (ceph_osd_is_up(osdmap, osds[i]))
11673d14c5d2SYehuda Sadeh 			acting[o++] = osds[i];
11683d14c5d2SYehuda Sadeh 	return o;
11693d14c5d2SYehuda Sadeh }
11703d14c5d2SYehuda Sadeh 
11713d14c5d2SYehuda Sadeh /*
11723d14c5d2SYehuda Sadeh  * Return primary osd for given pgid, or -1 if none.
11733d14c5d2SYehuda Sadeh  */
11743d14c5d2SYehuda Sadeh int ceph_calc_pg_primary(struct ceph_osdmap *osdmap, struct ceph_pg pgid)
11753d14c5d2SYehuda Sadeh {
11763d14c5d2SYehuda Sadeh 	int rawosds[CEPH_PG_MAX_SIZE], *osds;
11773d14c5d2SYehuda Sadeh 	int i, num = CEPH_PG_MAX_SIZE;
11783d14c5d2SYehuda Sadeh 
11793d14c5d2SYehuda Sadeh 	osds = calc_pg_raw(osdmap, pgid, rawosds, &num);
11803d14c5d2SYehuda Sadeh 	if (!osds)
11813d14c5d2SYehuda Sadeh 		return -1;
11823d14c5d2SYehuda Sadeh 
11833d14c5d2SYehuda Sadeh 	/* primary is first up osd */
11843d14c5d2SYehuda Sadeh 	for (i = 0; i < num; i++)
11853d14c5d2SYehuda Sadeh 		if (ceph_osd_is_up(osdmap, osds[i]))
11863d14c5d2SYehuda Sadeh 			return osds[i];
11873d14c5d2SYehuda Sadeh 	return -1;
11883d14c5d2SYehuda Sadeh }
11893d14c5d2SYehuda Sadeh EXPORT_SYMBOL(ceph_calc_pg_primary);
1190