Linux kernel mirror (for testing)
git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git
kernel
os
linux
Merge tag 'dma-mapping-6.6-2023-09-30' of git://git.infradead.org/users/hch/dma-mapping
Pull dma-mapping fixes from Christoph Hellwig:
- fix the narea calculation in swiotlb initialization (Ross Lagerwall)
- fix the check whether a device has used swiotlb (Petr Tesarik)
* tag 'dma-mapping-6.6-2023-09-30' of git://git.infradead.org/users/hch/dma-mapping:
swiotlb: fix the check whether a device has used software IO TLB
swiotlb: use the calculated number of areas
+38
-16
+16
-7
include/linux/swiotlb.h
+16
-7
include/linux/swiotlb.h
···
172
172
if (!mem)
173
173
return false;
174
174
175
-
if (IS_ENABLED(CONFIG_SWIOTLB_DYNAMIC)) {
176
-
/* Pairs with smp_wmb() in swiotlb_find_slots() and
177
-
* swiotlb_dyn_alloc(), which modify the RCU lists.
178
-
*/
179
-
smp_rmb();
180
-
return swiotlb_find_pool(dev, paddr);
181
-
}
175
+
#ifdef CONFIG_SWIOTLB_DYNAMIC
176
+
/*
177
+
* All SWIOTLB buffer addresses must have been returned by
178
+
* swiotlb_tbl_map_single() and passed to a device driver.
179
+
* If a SWIOTLB address is checked on another CPU, then it was
180
+
* presumably loaded by the device driver from an unspecified private
181
+
* data structure. Make sure that this load is ordered before reading
182
+
* dev->dma_uses_io_tlb here and mem->pools in swiotlb_find_pool().
183
+
*
184
+
* This barrier pairs with smp_mb() in swiotlb_find_slots().
185
+
*/
186
+
smp_rmb();
187
+
return READ_ONCE(dev->dma_uses_io_tlb) &&
188
+
swiotlb_find_pool(dev, paddr);
189
+
#else
182
190
return paddr >= mem->defpool.start && paddr < mem->defpool.end;
191
+
#endif
183
192
}
184
193
185
194
static inline bool is_swiotlb_force_bounce(struct device *dev)
+22
-9
kernel/dma/swiotlb.c
+22
-9
kernel/dma/swiotlb.c
···
399
399
}
400
400
401
401
mem->areas = memblock_alloc(array_size(sizeof(struct io_tlb_area),
402
-
default_nareas), SMP_CACHE_BYTES);
402
+
nareas), SMP_CACHE_BYTES);
403
403
if (!mem->areas) {
404
404
pr_warn("%s: Failed to allocate mem->areas.\n", __func__);
405
405
return;
406
406
}
407
407
408
-
swiotlb_init_io_tlb_pool(mem, __pa(tlb), nslabs, false,
409
-
default_nareas);
408
+
swiotlb_init_io_tlb_pool(mem, __pa(tlb), nslabs, false, nareas);
410
409
add_mem_pool(&io_tlb_default_mem, mem);
411
410
412
411
if (flags & SWIOTLB_VERBOSE)
···
728
729
}
729
730
730
731
add_mem_pool(mem, pool);
731
-
732
-
/* Pairs with smp_rmb() in is_swiotlb_buffer(). */
733
-
smp_wmb();
734
732
}
735
733
736
734
/**
···
1148
1152
spin_unlock_irqrestore(&dev->dma_io_tlb_lock, flags);
1149
1153
1150
1154
found:
1151
-
dev->dma_uses_io_tlb = true;
1152
-
/* Pairs with smp_rmb() in is_swiotlb_buffer() */
1153
-
smp_wmb();
1155
+
WRITE_ONCE(dev->dma_uses_io_tlb, true);
1156
+
1157
+
/*
1158
+
* The general barrier orders reads and writes against a presumed store
1159
+
* of the SWIOTLB buffer address by a device driver (to a driver private
1160
+
* data structure). It serves two purposes.
1161
+
*
1162
+
* First, the store to dev->dma_uses_io_tlb must be ordered before the
1163
+
* presumed store. This guarantees that the returned buffer address
1164
+
* cannot be passed to another CPU before updating dev->dma_uses_io_tlb.
1165
+
*
1166
+
* Second, the load from mem->pools must be ordered before the same
1167
+
* presumed store. This guarantees that the returned buffer address
1168
+
* cannot be observed by another CPU before an update of the RCU list
1169
+
* that was made by swiotlb_dyn_alloc() on a third CPU (cf. multicopy
1170
+
* atomicity).
1171
+
*
1172
+
* See also the comment in is_swiotlb_buffer().
1173
+
*/
1174
+
smp_mb();
1154
1175
1155
1176
*retpool = pool;
1156
1177
return index;