drivers/parisc/iommu-helpers.h

0001 /* SPDX-License-Identifier: GPL-2.0 */
0002 #include <linux/prefetch.h>
0003
0004 /**
0005  * iommu_fill_pdir - Insert coalesced scatter/gather chunks into the I/O Pdir.
0006  * @ioc: The I/O Controller.
0007  * @startsg: The scatter/gather list of coalesced chunks.
0008  * @nents: The number of entries in the scatter/gather list.
0009  * @hint: The DMA Hint.
0010  *
0011  * This function inserts the coalesced scatter/gather list chunks into the
0012  * I/O Controller's I/O Pdir.
0013  */
0014 static inline unsigned int
0015 iommu_fill_pdir(struct ioc *ioc, struct scatterlist *startsg, int nents,
0016         unsigned long hint,
0017         void (*iommu_io_pdir_entry)(u64 *, space_t, unsigned long,
0018                         unsigned long))
0019 {
0020     struct scatterlist *dma_sg = startsg;   /* pointer to current DMA */
0021     unsigned int n_mappings = 0;
0022     unsigned long dma_offset = 0, dma_len = 0;
0023     u64 *pdirp = NULL;
0024
0025     /* Horrible hack.  For efficiency's sake, dma_sg starts one
0026      * entry below the true start (it is immediately incremented
0027      * in the loop) */
0028      dma_sg--;
0029
0030     while (nents-- > 0) {
0031         unsigned long vaddr;
0032         long size;
0033
0034         DBG_RUN_SG(" %d : %08lx/%05x %p/%05x\n", nents,
0035                (unsigned long)sg_dma_address(startsg), cnt,
0036                sg_virt(startsg), startsg->length
0037         );
0038
0039
0040         /*
0041         ** Look for the start of a new DMA stream
0042         */
0043
0044         if (sg_dma_address(startsg) & PIDE_FLAG) {
0045             u32 pide = sg_dma_address(startsg) & ~PIDE_FLAG;
0046
0047             BUG_ON(pdirp && (dma_len != sg_dma_len(dma_sg)));
0048
0049             dma_sg++;
0050
0051             dma_len = sg_dma_len(startsg);
0052             sg_dma_len(startsg) = 0;
0053             dma_offset = (unsigned long) pide & ~IOVP_MASK;
0054             n_mappings++;
0055 #if defined(ZX1_SUPPORT)
0056             /* Pluto IOMMU IO Virt Address is not zero based */
0057             sg_dma_address(dma_sg) = pide | ioc->ibase;
0058 #else
0059             /* SBA, ccio, and dino are zero based.
0060              * Trying to save a few CPU cycles for most users.
0061              */
0062             sg_dma_address(dma_sg) = pide;
0063 #endif
0064             pdirp = &(ioc->pdir_base[pide >> IOVP_SHIFT]);
0065             prefetchw(pdirp);
0066         }
0067
0068         BUG_ON(pdirp == NULL);
0069
0070         vaddr = (unsigned long)sg_virt(startsg);
0071         sg_dma_len(dma_sg) += startsg->length;
0072         size = startsg->length + dma_offset;
0073         dma_offset = 0;
0074 #ifdef IOMMU_MAP_STATS
0075         ioc->msg_pages += startsg->length >> IOVP_SHIFT;
0076 #endif
0077         do {
0078             iommu_io_pdir_entry(pdirp, KERNEL_SPACE,
0079                         vaddr, hint);
0080             vaddr += IOVP_SIZE;
0081             size -= IOVP_SIZE;
0082             pdirp++;
0083         } while(unlikely(size > 0));
0084         startsg++;
0085     }
0086     return(n_mappings);
0087 }
0088
0089
0090 /*
0091 ** First pass is to walk the SG list and determine where the breaks are
0092 ** in the DMA stream. Allocates PDIR entries but does not fill them.
0093 ** Returns the number of DMA chunks.
0094 **
0095 ** Doing the fill separate from the coalescing/allocation keeps the
0096 ** code simpler. Future enhancement could make one pass through
0097 ** the sglist do both.
0098 */
0099
0100 static inline unsigned int
0101 iommu_coalesce_chunks(struct ioc *ioc, struct device *dev,
0102         struct scatterlist *startsg, int nents,
0103         int (*iommu_alloc_range)(struct ioc *, struct device *, size_t))
0104 {
0105     struct scatterlist *contig_sg;     /* contig chunk head */
0106     unsigned long dma_offset, dma_len; /* start/len of DMA stream */
0107     unsigned int n_mappings = 0;
0108     unsigned int max_seg_size = min(dma_get_max_seg_size(dev),
0109                     (unsigned)DMA_CHUNK_SIZE);
0110     unsigned int max_seg_boundary = dma_get_seg_boundary(dev) + 1;
0111     if (max_seg_boundary)   /* check if the addition above didn't overflow */
0112         max_seg_size = min(max_seg_size, max_seg_boundary);
0113
0114     while (nents > 0) {
0115
0116         /*
0117         ** Prepare for first/next DMA stream
0118         */
0119         contig_sg = startsg;
0120         dma_len = startsg->length;
0121         dma_offset = startsg->offset;
0122
0123         /* PARANOID: clear entries */
0124         sg_dma_address(startsg) = 0;
0125         sg_dma_len(startsg) = 0;
0126
0127         /*
0128         ** This loop terminates one iteration "early" since
0129         ** it's always looking one "ahead".
0130         */
0131         while(--nents > 0) {
0132             unsigned long prev_end, sg_start;
0133
0134             prev_end = (unsigned long)sg_virt(startsg) +
0135                             startsg->length;
0136
0137             startsg++;
0138             sg_start = (unsigned long)sg_virt(startsg);
0139
0140             /* PARANOID: clear entries */
0141             sg_dma_address(startsg) = 0;
0142             sg_dma_len(startsg) = 0;
0143
0144             /*
0145             ** First make sure current dma stream won't
0146             ** exceed max_seg_size if we coalesce the
0147             ** next entry.
0148             */
0149             if (unlikely(ALIGN(dma_len + dma_offset + startsg->length, IOVP_SIZE) >
0150                      max_seg_size))
0151                 break;
0152
0153             /*
0154             * Next see if we can append the next chunk (i.e.
0155             * it must end on one page and begin on another, or
0156             * it must start on the same address as the previous
0157             * entry ended.
0158             */
0159             if (unlikely((prev_end != sg_start) ||
0160                 ((prev_end | sg_start) & ~PAGE_MASK)))
0161                 break;
0162
0163             dma_len += startsg->length;
0164         }
0165
0166         /*
0167         ** End of DMA Stream
0168         ** Terminate last VCONTIG block.
0169         ** Allocate space for DMA stream.
0170         */
0171         sg_dma_len(contig_sg) = dma_len;
0172         dma_len = ALIGN(dma_len + dma_offset, IOVP_SIZE);
0173         sg_dma_address(contig_sg) =
0174             PIDE_FLAG
0175             | (iommu_alloc_range(ioc, dev, dma_len) << IOVP_SHIFT)
0176             | dma_offset;
0177         n_mappings++;
0178     }
0179
0180     return n_mappings;
0181 }
0182