在 2019/5/27 20:07, Jason Gunthorpe 写道:
> On Sat, May 25, 2019 at 05:50:07PM +0800, Lijun Ou wrote:
>> Currently, the code for getting umem and kmem buffers exist many files,
>> this patch adds a group interfaces to simplify the buffers getting flow.
>>
>> Signed-off-by: Xi Wang <wangxi11@xxxxxxxxxx>
>> Signed-off-by: Lijun Ou <oulijun@xxxxxxxxxx>
>> drivers/infiniband/hw/hns/hns_roce_alloc.c | 131 ++++++++++++++++++++++++++++
>> drivers/infiniband/hw/hns/hns_roce_device.h | 12 +++
>> 2 files changed, 143 insertions(+)
>>
>> diff --git a/drivers/infiniband/hw/hns/hns_roce_alloc.c b/drivers/infiniband/hw/hns/hns_roce_alloc.c
>> index dac058d..7a08064 100644
>> +++ b/drivers/infiniband/hw/hns/hns_roce_alloc.c
>> @@ -34,6 +34,7 @@
>> #include <linux/platform_device.h>
>> #include <linux/vmalloc.h>
>> #include "hns_roce_device.h"
>> +#include <rdma/ib_umem.h>
>>
>> int hns_roce_bitmap_alloc(struct hns_roce_bitmap *bitmap, unsigned long *obj)
>> {
>> @@ -238,6 +239,136 @@ int hns_roce_buf_alloc(struct hns_roce_dev *hr_dev, u32 size, u32 max_direct,
>> return -ENOMEM;
>> }
>>
>> +int hns_roce_get_kmem_bufs(struct hns_roce_dev *hr_dev, dma_addr_t *bufs,
>> + int buf_cnt, int start, struct hns_roce_buf *buf)
>> +{
>> + int i, end;
>> + int total;
>> +
>> + end = start + buf_cnt;
>> + if (end > buf->npages) {
>> + dev_err(hr_dev->dev,
>> + "invalid kmem region,offset %d,buf_cnt %d,total %d!\n",
>> + start, buf_cnt, buf->npages);
>> + return -EINVAL;
>> + }
>> +
>> + total = 0;
>> + for (i = start; i < end; i++)
>> + if (buf->nbufs == 1)
>> + bufs[total++] = buf->direct.map +
>> + (i << buf->page_shift);
>> + else
>> + bufs[total++] = buf->page_list[i].map;
>> +
>> + return total;
>> +}
>> +
>> +int hns_roce_get_umem_bufs(struct hns_roce_dev *hr_dev, dma_addr_t *bufs,
>> + int buf_cnt, int start, struct ib_umem *umem,
>> + int page_shift)
>> +{
>> + struct scatterlist *sg;
>> + int npage_per_buf;
>> + int npage_per_sg;
>> + dma_addr_t addr;
>> + int n, entry;
>> + int idx, end;
>> + int npage;
>> + int total;
>> +
>> + if (page_shift < PAGE_SHIFT || page_shift > umem->page_shift) {
>> + dev_err(hr_dev->dev, "invalid page shift %d, umem shift %d!\n",
>> + page_shift, umem->page_shift);
>> + return -EINVAL;
>> + }
>> +
>> + /* convert system page cnt to hw page cnt */
>> + npage_per_buf = (1 << (page_shift - PAGE_SHIFT));
>> + total = DIV_ROUND_UP(ib_umem_page_count(umem), npage_per_buf);
>> + end = start + buf_cnt;
>> + if (end > total) {
>> + dev_err(hr_dev->dev,
>> + "invalid umem region,offset %d,buf_cnt %d,total %d!\n",
>> + start, buf_cnt, total);
>> + return -EINVAL;
>> + }
>> +
>> + idx = 0;
>> + npage = 0;
>> + total = 0;
>> + for_each_sg(umem->sg_head.sgl, sg, umem->nmap, entry) {
> You should convert this driver to use the new APIs, see something like
>
> commit eb52c0333f06b88bca5bac0dc0aeca729de6eb11
> Author: Shiraz Saleem <shiraz.saleem@xxxxxxxxx>
> Date: Mon May 6 08:53:34 2019 -0500
>
> RDMA/i40iw: Use core helpers to get aligned DMA address within a supported page size
>
> Call the core helpers to retrieve the HW aligned address to use for the
> MR, within a supported i40iw page size.
>
> Remove code in i40iw to determine when MR is backed by 2M huge pages which
> involves checking the umem->hugetlb flag and VMA inspection. The new DMA
> iterator will return the 2M aligned address if the MR is backed by 2M
> pages.
>
> Fixes: f26c7c83395b ("i40iw: Add 2MB page support")
> Reviewed-by: Michael J. Ruhl <michael.j.ruhl@xxxxxxxxx>
> Signed-off-by: Shiraz Saleem <shiraz.saleem@xxxxxxxxx>
> Signed-off-by: Jason Gunthorpe <jgg@xxxxxxxxxxxx>
>
>
Thanks. We will fix it and send the V2.
>> + npage_per_sg = sg_dma_len(sg) >> PAGE_SHIFT;
>> + for (n = 0; n < npage_per_sg; n++) {
>> + if (!(npage % npage_per_buf)) {
>> + addr = sg_dma_address(sg) +
>> + (n << umem->page_shift);
> No new references to umem->page_shift, this won't apply.
>
> Jason
>
> .
>
