On 2017/11/1 20:26, Robin Murphy wrote:
> On 01/11/17 07:46, Wei Hu (Xavier) wrote:
>>
>> On 2017/10/12 20:59, Robin Murphy wrote:
>>> On 12/10/17 13:31, Wei Hu (Xavier) wrote:
>>>> On 2017/10/1 0:10, Leon Romanovsky wrote:
>>>>> On Sat, Sep 30, 2017 at 05:28:59PM +0800, Wei Hu (Xavier) wrote:
>>>>>> If the IOMMU is enabled, the length of sg obtained from
>>>>>> __iommu_map_sg_attrs is not 4kB. When the IOVA is set with the sg
>>>>>> dma address, the IOVA will not be page continuous. and the VA
>>>>>> returned from dma_alloc_coherent is a vmalloc address. However,
>>>>>> the VA obtained by the page_address is a discontinuous VA. Under
>>>>>> these circumstances, the IOVA should be calculated based on the
>>>>>> sg length, and record the VA returned from dma_alloc_coherent
>>>>>> in the struct of hem.
>>>>>>
>>>>>> Signed-off-by: Wei Hu (Xavier) <xavier.huwei@xxxxxxxxxx>
>>>>>> Signed-off-by: Shaobo Xu <xushaobo2@xxxxxxxxxx>
>>>>>> Signed-off-by: Lijun Ou <oulijun@xxxxxxxxxx>
>>>>>> ---
>>>>> Doug,
>>>>>
>>>>> I didn't invest time in reviewing it, but having "is_vmalloc_addr" in
>>>>> driver code to deal with dma_alloc_coherent is most probably wrong.
>>>>>
>>>>> Thanks
>>>> Hi, Leon & Doug
>>>> We refered the function named __ttm_dma_alloc_page in the kernel
>>>> code as below:
>>>> And there are similar methods in bch_bio_map and mem_to_page
>>>> functions in current 4.14-rcx.
>>>>
>>>> static struct dma_page *__ttm_dma_alloc_page(struct dma_pool *pool)
>>>> {
>>>> struct dma_page *d_page;
>>>>
>>>> d_page = kmalloc(sizeof(struct dma_page), GFP_KERNEL);
>>>> if (!d_page)
>>>> return NULL;
>>>>
>>>> d_page->vaddr = dma_alloc_coherent(pool->dev, pool->size,
>>>> &d_page->dma,
>>>> pool->gfp_flags);
>>>> if (d_page->vaddr) {
>>>> if (is_vmalloc_addr(d_page->vaddr))
>>>> d_page->p = vmalloc_to_page(d_page->vaddr);
>>>> else
>>>> d_page->p = virt_to_page(d_page->vaddr);
>>> There are cases on various architectures where neither of those is
>>> right. Whether those actually intersect with TTM or RDMA use-cases is
>>> another matter, of course.
>>>
>>> What definitely is a problem is if you ever take that page and end up
>>> accessing it through any virtual address other than the one explicitly
>>> returned by dma_alloc_coherent(). That can blow the coherency wide open
>>> and invite data loss, right up to killing the whole system with a
>>> machine check on certain architectures.
>>>
>>> Robin.
>> Hi, Robin
>> Thanks for your comment.
>>
>> We have one problem and the related code as below.
>> 1. call dma_alloc_coherent function serval times to alloc memory.
>> 2. vmap the allocated memory pages.
>> 3. software access memory by using the return virt addr of vmap
>> and hardware using the dma addr of dma_alloc_coherent.
> The simple answer is "don't do that". Seriously. dma_alloc_coherent()
> gives you a CPU virtual address and a DMA address with which to access
> your buffer, and that is the limit of what you may infer about it. You
> have no guarantee that the virtual address is either in the linear map
> or vmalloc, and not some other special place. You have no guarantee that
> the underlying memory even has an associated struct page at all.
>
>> When IOMMU is disabled in ARM64 architecture, we use virt_to_page()
>> before vmap(), it works. And when IOMMU is enabled using
>> virt_to_page() will cause calltrace later, we found the return
>> addr of dma_alloc_coherent is vmalloc addr, so we add the
>> condition judgement statement as below, it works.
>> for (i = 0; i < buf->nbufs; ++i)
>> pages[i] =
>> is_vmalloc_addr(buf->page_list[i].buf) ?
>> vmalloc_to_page(buf->page_list[i].buf) :
>> virt_to_page(buf->page_list[i].buf);
>> Can you give us suggestion? better method?
> Oh my goodness, having now taken a closer look at this driver, I'm lost
> for words in disbelief. To pick just one example:
>
> u32 bits_per_long = BITS_PER_LONG;
> ...
> if (bits_per_long == 64) {
> /* memory mapping nonsense */
> }
>
> WTF does the size of a long have to do with DMA buffer management!?
>
> Of course I can guess that it might be trying to make some tortuous
> inference about vmalloc space being constrained on 32-bit platforms, but
> still...
>
>> The related code as below:
>> buf->page_list = kcalloc(buf->nbufs, sizeof(*buf->page_list),
>> GFP_KERNEL);
>> if (!buf->page_list)
>> return -ENOMEM;
>>
>> for (i = 0; i < buf->nbufs; ++i) {
>> buf->page_list[i].buf = dma_alloc_coherent(dev,
>> page_size, &t,
>> GFP_KERNEL);
>> if (!buf->page_list[i].buf)
>> goto err_free;
>>
>> buf->page_list[i].map = t;
>> memset(buf->page_list[i].buf, 0, page_size);
>> }
>>
>> pages = kmalloc_array(buf->nbufs, sizeof(*pages),
>> GFP_KERNEL);
>> if (!pages)
>> goto err_free;
>>
>> for (i = 0; i < buf->nbufs; ++i)
>> pages[i] =
>> is_vmalloc_addr(buf->page_list[i].buf) ?
>> vmalloc_to_page(buf->page_list[i].buf) :
>> virt_to_page(buf->page_list[i].buf);
>>
>> buf->direct.buf = vmap(pages, buf->nbufs, VM_MAP,
>> PAGE_KERNEL);
>> kfree(pages);
>> if (!buf->direct.buf)
>> goto err_free;
> OK, this is complete crap. As above, you cannot assume that a struct
> page even exists; even if it does you cannot assume that using a
> PAGE_KERNEL mapping will not result in mismatched attributes,
> unpredictable behaviour and data loss. Trying to remap coherent DMA
> allocations like this is just egregiously wrong.
>
> What I do like is that you can seemingly fix all this by simply deleting
> hns_roce_buf::direct and all the garbage code related to it, and using
> the page_list entries consistently because the alternate paths involving
> those appear to do the right thing already.
>
> That is, of course, assuming that the buffers involved can be so large
> that it's not practical to just always make a single allocation and
> fragment it into multiple descriptors if the hardware does have some
> maximum length constraint - frankly I'm a little puzzled by the
> PAGE_SIZE * 2 threshold, given that that's not a fixed size.
>
> Robin.
Hi，Robin
We reconstruct the code as below:
It replaces dma_alloc_coherent with __get_free_pages and
dma_map_single
functions. So, we can vmap serveral ptrs returned by
__get_free_pages, right?
buf->page_list = kcalloc(buf->nbufs, sizeof(*buf->page_list),
GFP_KERNEL);
if (!buf->page_list)
return -ENOMEM;
for (i = 0; i < buf->nbufs; ++i) {
ptr = (void *)__get_free_pages(GFP_KERNEL | __GFP_ZERO,
get_order(page_size));
if (!ptr) {
dev_err(dev, "Alloc pages error.\n");
goto err_free;
}
t = dma_map_single(dev, ptr, page_size,
DMA_BIDIRECTIONAL);
if (dma_mapping_error(dev, t)) {
dev_err(dev, "DMA mapping error.\n");
free_pages((unsigned long)ptr,
get_order(page_size));
goto err_free;
}
buf->page_list[i].buf = ptr;
buf->page_list[i].map = t;
}
pages = kmalloc_array(buf->nbufs, sizeof(*pages),
GFP_KERNEL);
if (!pages)
goto err_free;
for (i = 0; i < buf->nbufs; ++i)
pages[i] = virt_to_page(buf->page_list[i].buf);
buf->direct.buf = vmap(pages, buf->nbufs, VM_MAP,
PAGE_KERNEL);
kfree(pages);
if (!buf->direct.buf)
goto err_free;
Regards
Wei Hu
>> Regards
>> Wei Hu
>>>> } else {
>>>> kfree(d_page);
>>>> d_page = NULL;
>>>> }
>>>> return d_page;
>>>> }
>>>>
>>>> Regards
>>>> Wei Hu
>>>>>> drivers/infiniband/hw/hns/hns_roce_alloc.c | 5 ++++-
>>>>>> drivers/infiniband/hw/hns/hns_roce_hem.c | 30
>>>>>> +++++++++++++++++++++++++++---
>>>>>> drivers/infiniband/hw/hns/hns_roce_hem.h | 6 ++++++
>>>>>> drivers/infiniband/hw/hns/hns_roce_hw_v2.c | 22 +++++++++++++++-------
>>>>>> 4 files changed, 52 insertions(+), 11 deletions(-)
>>>>>>
>>>>>> diff --git a/drivers/infiniband/hw/hns/hns_roce_alloc.c
>>>>>> b/drivers/infiniband/hw/hns/hns_roce_alloc.c
>>>>>> index 3e4c525..a69cd4b 100644
>>>>>> --- a/drivers/infiniband/hw/hns/hns_roce_alloc.c
>>>>>> +++ b/drivers/infiniband/hw/hns/hns_roce_alloc.c
>>>>>> @@ -243,7 +243,10 @@ int hns_roce_buf_alloc(struct hns_roce_dev
>>>>>> *hr_dev, u32 size, u32 max_direct,
>>>>>> goto err_free;
>>>>>>
>>>>>> for (i = 0; i < buf->nbufs; ++i)
>>>>>> - pages[i] = virt_to_page(buf->page_list[i].buf);
>>>>>> + pages[i] =
>>>>>> + is_vmalloc_addr(buf->page_list[i].buf) ?
>>>>>> + vmalloc_to_page(buf->page_list[i].buf) :
>>>>>> + virt_to_page(buf->page_list[i].buf);
>>>>>>
>>>>>> buf->direct.buf = vmap(pages, buf->nbufs, VM_MAP,
>>>>>> PAGE_KERNEL);
>>>>>> diff --git a/drivers/infiniband/hw/hns/hns_roce_hem.c
>>>>>> b/drivers/infiniband/hw/hns/hns_roce_hem.c
>>>>>> index 8388ae2..4a3d1d4 100644
>>>>>> --- a/drivers/infiniband/hw/hns/hns_roce_hem.c
>>>>>> +++ b/drivers/infiniband/hw/hns/hns_roce_hem.c
>>>>>> @@ -200,6 +200,7 @@ static struct hns_roce_hem
>>>>>> *hns_roce_alloc_hem(struct hns_roce_dev *hr_dev,
>>>>>> gfp_t gfp_mask)
>>>>>> {
>>>>>> struct hns_roce_hem_chunk *chunk = NULL;
>>>>>> + struct hns_roce_vmalloc *vmalloc;
>>>>>> struct hns_roce_hem *hem;
>>>>>> struct scatterlist *mem;
>>>>>> int order;
>>>>>> @@ -227,6 +228,7 @@ static struct hns_roce_hem
>>>>>> *hns_roce_alloc_hem(struct hns_roce_dev *hr_dev,
>>>>>> sg_init_table(chunk->mem, HNS_ROCE_HEM_CHUNK_LEN);
>>>>>> chunk->npages = 0;
>>>>>> chunk->nsg = 0;
>>>>>> + memset(chunk->vmalloc, 0, sizeof(chunk->vmalloc));
>>>>>> list_add_tail(&chunk->list, &hem->chunk_list);
>>>>>> }
>>>>>>
>>>>>> @@ -243,7 +245,15 @@ static struct hns_roce_hem
>>>>>> *hns_roce_alloc_hem(struct hns_roce_dev *hr_dev,
>>>>>> if (!buf)
>>>>>> goto fail;
>>>>>>
>>>>>> - sg_set_buf(mem, buf, PAGE_SIZE << order);
>>>>>> + if (is_vmalloc_addr(buf)) {
>>>>>> + vmalloc = &chunk->vmalloc[chunk->npages];
>>>>>> + vmalloc->is_vmalloc_addr = true;
>>>>>> + vmalloc->vmalloc_addr = buf;
>>>>>> + sg_set_page(mem, vmalloc_to_page(buf),
>>>>>> + PAGE_SIZE << order, offset_in_page(buf));
>>>>>> + } else {
>>>>>> + sg_set_buf(mem, buf, PAGE_SIZE << order);
>>>>>> + }
>>>>>> WARN_ON(mem->offset);
>>>>>> sg_dma_len(mem) = PAGE_SIZE << order;
>>>>>>
>>>>>> @@ -262,17 +272,25 @@ static struct hns_roce_hem
>>>>>> *hns_roce_alloc_hem(struct hns_roce_dev *hr_dev,
>>>>>> void hns_roce_free_hem(struct hns_roce_dev *hr_dev, struct
>>>>>> hns_roce_hem *hem)
>>>>>> {
>>>>>> struct hns_roce_hem_chunk *chunk, *tmp;
>>>>>> + void *cpu_addr;
>>>>>> int i;
>>>>>>
>>>>>> if (!hem)
>>>>>> return;
>>>>>>
>>>>>> list_for_each_entry_safe(chunk, tmp, &hem->chunk_list, list) {
>>>>>> - for (i = 0; i < chunk->npages; ++i)
>>>>>> + for (i = 0; i < chunk->npages; ++i) {
>>>>>> + if (chunk->vmalloc[i].is_vmalloc_addr)
>>>>>> + cpu_addr = chunk->vmalloc[i].vmalloc_addr;
>>>>>> + else
>>>>>> + cpu_addr =
>>>>>> + lowmem_page_address(sg_page(&chunk->mem[i]));
>>>>>> +
>>>>>> dma_free_coherent(hr_dev->dev,
>>>>>> chunk->mem[i].length,
>>>>>> - lowmem_page_address(sg_page(&chunk->mem[i])),
>>>>>> + cpu_addr,
>>>>>> sg_dma_address(&chunk->mem[i]));
>>>>>> + }
>>>>>> kfree(chunk);
>>>>>> }
>>>>>>
>>>>>> @@ -774,6 +792,12 @@ void *hns_roce_table_find(struct hns_roce_dev
>>>>>> *hr_dev,
>>>>>>
>>>>>> if (chunk->mem[i].length > (u32)offset) {
>>>>>> page = sg_page(&chunk->mem[i]);
>>>>>> + if (chunk->vmalloc[i].is_vmalloc_addr) {
>>>>>> + mutex_unlock(&table->mutex);
>>>>>> + return page ?
>>>>>> + chunk->vmalloc[i].vmalloc_addr
>>>>>> + + offset : NULL;
>>>>>> + }
>>>>>> goto out;
>>>>>> }
>>>>>> offset -= chunk->mem[i].length;
>>>>>> diff --git a/drivers/infiniband/hw/hns/hns_roce_hem.h
>>>>>> b/drivers/infiniband/hw/hns/hns_roce_hem.h
>>>>>> index af28bbf..62d712a 100644
>>>>>> --- a/drivers/infiniband/hw/hns/hns_roce_hem.h
>>>>>> +++ b/drivers/infiniband/hw/hns/hns_roce_hem.h
>>>>>> @@ -72,11 +72,17 @@ enum {
>>>>>> HNS_ROCE_HEM_PAGE_SIZE = 1 << HNS_ROCE_HEM_PAGE_SHIFT,
>>>>>> };
>>>>>>
>>>>>> +struct hns_roce_vmalloc {
>>>>>> + bool is_vmalloc_addr;
>>>>>> + void *vmalloc_addr;
>>>>>> +};
>>>>>> +
>>>>>> struct hns_roce_hem_chunk {
>>>>>> struct list_head list;
>>>>>> int npages;
>>>>>> int nsg;
>>>>>> struct scatterlist mem[HNS_ROCE_HEM_CHUNK_LEN];
>>>>>> + struct hns_roce_vmalloc vmalloc[HNS_ROCE_HEM_CHUNK_LEN];
>>>>>> };
>>>>>>
>>>>>> struct hns_roce_hem {
>>>>>> diff --git a/drivers/infiniband/hw/hns/hns_roce_hw_v2.c
>>>>>> b/drivers/infiniband/hw/hns/hns_roce_hw_v2.c
>>>>>> index b99d70a..9e19bf1 100644
>>>>>> --- a/drivers/infiniband/hw/hns/hns_roce_hw_v2.c
>>>>>> +++ b/drivers/infiniband/hw/hns/hns_roce_hw_v2.c
>>>>>> @@ -1093,9 +1093,11 @@ static int hns_roce_v2_write_mtpt(void
>>>>>> *mb_buf, struct hns_roce_mr *mr,
>>>>>> {
>>>>>> struct hns_roce_v2_mpt_entry *mpt_entry;
>>>>>> struct scatterlist *sg;
>>>>>> + u64 page_addr = 0;
>>>>>> u64 *pages;
>>>>>> + int i = 0, j = 0;
>>>>>> + int len = 0;
>>>>>> int entry;
>>>>>> - int i;
>>>>>>
>>>>>> mpt_entry = mb_buf;
>>>>>> memset(mpt_entry, 0, sizeof(*mpt_entry));
>>>>>> @@ -1153,14 +1155,20 @@ static int hns_roce_v2_write_mtpt(void
>>>>>> *mb_buf, struct hns_roce_mr *mr,
>>>>>>
>>>>>> i = 0;
>>>>>> for_each_sg(mr->umem->sg_head.sgl, sg, mr->umem->nmap, entry) {
>>>>>> - pages[i] = ((u64)sg_dma_address(sg)) >> 6;
>>>>>> -
>>>>>> - /* Record the first 2 entry directly to MTPT table */
>>>>>> - if (i >= HNS_ROCE_V2_MAX_INNER_MTPT_NUM - 1)
>>>>>> - break;
>>>>>> - i++;
>>>>>> + len = sg_dma_len(sg) >> PAGE_SHIFT;
>>>>>> + for (j = 0; j < len; ++j) {
>>>>>> + page_addr = sg_dma_address(sg) +
>>>>>> + (j << mr->umem->page_shift);
>>>>>> + pages[i] = page_addr >> 6;
>>>>>> +
>>>>>> + /* Record the first 2 entry directly to MTPT table */
>>>>>> + if (i >= HNS_ROCE_V2_MAX_INNER_MTPT_NUM - 1)
>>>>>> + goto found;
>>>>>> + i++;
>>>>>> + }
>>>>>> }
>>>>>>
>>>>>> +found:
>>>>>> mpt_entry->pa0_l = cpu_to_le32(lower_32_bits(pages[0]));
>>>>>> roce_set_field(mpt_entry->byte_56_pa0_h, V2_MPT_BYTE_56_PA0_H_M,
>>>>>> V2_MPT_BYTE_56_PA0_H_S,
>>>>>> --
>>>>>> 1.9.1
>>>>>>
>>>> _______________________________________________
>>>> iommu mailing list
>>>> iommu@xxxxxxxxxxxxxxxxxxxxxxxxxx
>>>> https://lists.linuxfoundation.org/mailman/listinfo/iommu
>>> .
>>>
>>
> --
> To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
> the body of a message to majordomo@xxxxxxxxxxxxxxx
> More majordomo info at http://vger.kernel.org/majordomo-info.html
>
> .
>
--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
