1/*
  2 * IOMMU mmap management and range allocation functions.
  3 * Based almost entirely upon the powerpc iommu allocator.
  4 */
  5
  6#include <linux/export.h>
  7#include <linux/bitmap.h>
  8#include <linux/bug.h>
  9#include <linux/iommu-helper.h>
 10#include <linux/iommu-common.h>
 11#include <linux/dma-mapping.h>
 12#include <linux/hash.h>
 13
 14static unsigned long iommu_large_alloc = 15;
 15
 16static	DEFINE_PER_CPU(unsigned int, iommu_hash_common);
 17
 18static inline bool need_flush(struct iommu_map_table *iommu)
 19{
 20	return ((iommu->flags & IOMMU_NEED_FLUSH) != 0);
 21}
 22
 23static inline void set_flush(struct iommu_map_table *iommu)
 24{
 25	iommu->flags |= IOMMU_NEED_FLUSH;
 26}
 27
 28static inline void clear_flush(struct iommu_map_table *iommu)
 29{
 30	iommu->flags &= ~IOMMU_NEED_FLUSH;
 31}
 32
 33static void setup_iommu_pool_hash(void)
 34{
 35	unsigned int i;
 36	static bool do_once;
 37
 38	if (do_once)
 39		return;
 40	do_once = true;
 41	for_each_possible_cpu(i)
 42		per_cpu(iommu_hash_common, i) = hash_32(i, IOMMU_POOL_HASHBITS);
 43}
 44
 45/*
 46 * Initialize iommu_pool entries for the iommu_map_table. `num_entries'
 47 * is the number of table entries. If `large_pool' is set to true,
 48 * the top 1/4 of the table will be set aside for pool allocations
 49 * of more than iommu_large_alloc pages.
 50 */
 51void iommu_tbl_pool_init(struct iommu_map_table *iommu,
 52			 unsigned long num_entries,
 53			 u32 table_shift,
 54			 void (*lazy_flush)(struct iommu_map_table *),
 55			 bool large_pool, u32 npools,
 56			 bool skip_span_boundary_check)
 57{
 58	unsigned int start, i;
 59	struct iommu_pool *p = &(iommu->large_pool);
 60
 61	setup_iommu_pool_hash();
 62	if (npools == 0)
 63		iommu->nr_pools = IOMMU_NR_POOLS;
 64	else
 65		iommu->nr_pools = npools;
 66	BUG_ON(npools > IOMMU_NR_POOLS);
 67
 68	iommu->table_shift = table_shift;
 69	iommu->lazy_flush = lazy_flush;
 70	start = 0;
 71	if (skip_span_boundary_check)
 72		iommu->flags |= IOMMU_NO_SPAN_BOUND;
 73	if (large_pool)
 74		iommu->flags |= IOMMU_HAS_LARGE_POOL;
 75
 76	if (!large_pool)
 77		iommu->poolsize = num_entries/iommu->nr_pools;
 78	else
 79		iommu->poolsize = (num_entries * 3 / 4)/iommu->nr_pools;
 80	for (i = 0; i < iommu->nr_pools; i++) {
 81		spin_lock_init(&(iommu->pools[i].lock));
 82		iommu->pools[i].start = start;
 83		iommu->pools[i].hint = start;
 84		start += iommu->poolsize; /* start for next pool */
 85		iommu->pools[i].end = start - 1;
 86	}
 87	if (!large_pool)
 88		return;
 89	/* initialize large_pool */
 90	spin_lock_init(&(p->lock));
 91	p->start = start;
 92	p->hint = p->start;
 93	p->end = num_entries;
 94}
 95EXPORT_SYMBOL(iommu_tbl_pool_init);
 96
 97unsigned long iommu_tbl_range_alloc(struct device *dev,
 98				struct iommu_map_table *iommu,
 99				unsigned long npages,
100				unsigned long *handle,
101				unsigned long mask,
102				unsigned int align_order)
103{
104	unsigned int pool_hash = __this_cpu_read(iommu_hash_common);
105	unsigned long n, end, start, limit, boundary_size;
106	struct iommu_pool *pool;
107	int pass = 0;
108	unsigned int pool_nr;
109	unsigned int npools = iommu->nr_pools;
110	unsigned long flags;
111	bool large_pool = ((iommu->flags & IOMMU_HAS_LARGE_POOL) != 0);
112	bool largealloc = (large_pool && npages > iommu_large_alloc);
113	unsigned long shift;
114	unsigned long align_mask = 0;
115
116	if (align_order > 0)
117		align_mask = ~0ul >> (BITS_PER_LONG - align_order);
118
119	/* Sanity check */
120	if (unlikely(npages == 0)) {
121		WARN_ON_ONCE(1);
122		return IOMMU_ERROR_CODE;
123	}
124
125	if (largealloc) {
126		pool = &(iommu->large_pool);
127		pool_nr = 0; /* to keep compiler happy */
128	} else {
129		/* pick out pool_nr */
130		pool_nr =  pool_hash & (npools - 1);
131		pool = &(iommu->pools[pool_nr]);
132	}
133	spin_lock_irqsave(&pool->lock, flags);
134
135 again:
136	if (pass == 0 && handle && *handle &&
137	    (*handle >= pool->start) && (*handle < pool->end))
138		start = *handle;
139	else
140		start = pool->hint;
141
142	limit = pool->end;
143
144	/* The case below can happen if we have a small segment appended
145	 * to a large, or when the previous alloc was at the very end of
146	 * the available space. If so, go back to the beginning. If a
147	 * flush is needed, it will get done based on the return value
148	 * from iommu_area_alloc() below.
149	 */
150	if (start >= limit)
151		start = pool->start;
152	shift = iommu->table_map_base >> iommu->table_shift;
153	if (limit + shift > mask) {
154		limit = mask - shift + 1;
155		/* If we're constrained on address range, first try
156		 * at the masked hint to avoid O(n) search complexity,
157		 * but on second pass, start at 0 in pool 0.
158		 */
159		if ((start & mask) >= limit || pass > 0) {
160			spin_unlock(&(pool->lock));
161			pool = &(iommu->pools[0]);
162			spin_lock(&(pool->lock));
163			start = pool->start;
164		} else {
165			start &= mask;
166		}
167	}
168
169	if (dev)
170		boundary_size = ALIGN(dma_get_seg_boundary(dev) + 1,
171				      1 << iommu->table_shift);
172	else
173		boundary_size = ALIGN(1ULL << 32, 1 << iommu->table_shift);
174
175	boundary_size = boundary_size >> iommu->table_shift;
176	/*
177	 * if the skip_span_boundary_check had been set during init, we set
178	 * things up so that iommu_is_span_boundary() merely checks if the
179	 * (index + npages) < num_tsb_entries
180	 */
181	if ((iommu->flags & IOMMU_NO_SPAN_BOUND) != 0) {
182		shift = 0;
183		boundary_size = iommu->poolsize * iommu->nr_pools;
184	}
185	n = iommu_area_alloc(iommu->map, limit, start, npages, shift,
186			     boundary_size, align_mask);
187	if (n == -1) {
188		if (likely(pass == 0)) {
189			/* First failure, rescan from the beginning.  */
190			pool->hint = pool->start;
191			set_flush(iommu);
192			pass++;
193			goto again;
194		} else if (!largealloc && pass <= iommu->nr_pools) {
195			spin_unlock(&(pool->lock));
196			pool_nr = (pool_nr + 1) & (iommu->nr_pools - 1);
197			pool = &(iommu->pools[pool_nr]);
198			spin_lock(&(pool->lock));
199			pool->hint = pool->start;
200			set_flush(iommu);
201			pass++;
202			goto again;
203		} else {
204			/* give up */
205			n = IOMMU_ERROR_CODE;
206			goto bail;
207		}
208	}
209	if (iommu->lazy_flush &&
210	    (n < pool->hint || need_flush(iommu))) {
211		clear_flush(iommu);
212		iommu->lazy_flush(iommu);
213	}
214
215	end = n + npages;
216	pool->hint = end;
217
218	/* Update handle for SG allocations */
219	if (handle)
220		*handle = end;
221bail:
222	spin_unlock_irqrestore(&(pool->lock), flags);
223
224	return n;
225}
226EXPORT_SYMBOL(iommu_tbl_range_alloc);
227
228static struct iommu_pool *get_pool(struct iommu_map_table *tbl,
229				   unsigned long entry)
230{
231	struct iommu_pool *p;
232	unsigned long largepool_start = tbl->large_pool.start;
233	bool large_pool = ((tbl->flags & IOMMU_HAS_LARGE_POOL) != 0);
234
235	/* The large pool is the last pool at the top of the table */
236	if (large_pool && entry >= largepool_start) {
237		p = &tbl->large_pool;
238	} else {
239		unsigned int pool_nr = entry / tbl->poolsize;
240
241		BUG_ON(pool_nr >= tbl->nr_pools);
242		p = &tbl->pools[pool_nr];
243	}
244	return p;
245}
246
247/* Caller supplies the index of the entry into the iommu map table
248 * itself when the mapping from dma_addr to the entry is not the
249 * default addr->entry mapping below.
250 */
251void iommu_tbl_range_free(struct iommu_map_table *iommu, u64 dma_addr,
252			  unsigned long npages, unsigned long entry)
253{
254	struct iommu_pool *pool;
255	unsigned long flags;
256	unsigned long shift = iommu->table_shift;
257
258	if (entry == IOMMU_ERROR_CODE) /* use default addr->entry mapping */
259		entry = (dma_addr - iommu->table_map_base) >> shift;
260	pool = get_pool(iommu, entry);
261
262	spin_lock_irqsave(&(pool->lock), flags);
263	bitmap_clear(iommu->map, entry, npages);
264	spin_unlock_irqrestore(&(pool->lock), flags);
265}
266EXPORT_SYMBOL(iommu_tbl_range_free);