Re: [dpdk-dev] [EXT] Re: [PATCH 2/2] vfio: fix partial DMA unmapping for VFIO type1

["Burakov, Anatoly" <anatoly.burakov@intel.com>]

On 22-Oct-20 1:13 PM, Nithin Dabilpuram wrote:
> Ping.
> 
> On Mon, Oct 19, 2020 at 03:13:15PM +0530, Nithin Dabilpuram wrote:
>> On Sat, Oct 17, 2020 at 05:14:55PM +0100, Burakov, Anatoly wrote:
>>> On 16-Oct-20 8:10 AM, Nithin Dabilpuram wrote:
>>>> On Thu, Oct 15, 2020 at 04:10:31PM +0100, Burakov, Anatoly wrote:
>>>>> On 15-Oct-20 12:57 PM, Nithin Dabilpuram wrote:
>>>>>> On Thu, Oct 15, 2020 at 3:31 PM Burakov, Anatoly
>>>>>> <anatoly.burakov@intel.com> wrote:
>>>>>>>
>>>>>>> On 15-Oct-20 7:09 AM, Nithin Dabilpuram wrote:
>>>>>>>> On Wed, Oct 14, 2020 at 04:07:10PM +0100, Burakov, Anatoly wrote:
>>>>>>>>> External Email
>>>>>>>>>
>>>>>>>>> ----------------------------------------------------------------------
>>>>>>>>> On 12-Oct-20 9:11 AM, Nithin Dabilpuram wrote:
>>>>>>>>>> Partial unmapping is not supported for VFIO IOMMU type1
>>>>>>>>>> by kernel. Though kernel gives return as zero, the unmapped size
>>>>>>>>>> returned will not be same as expected. So check for
>>>>>>>>>> returned unmap size and return error.
>>>>>>>>>>
>>>>>>>>>> For case of DMA map/unmap triggered by heap allocations,
>>>>>>>>>> maintain granularity of memseg page size so that heap
>>>>>>>>>> expansion and contraction does not have this issue.
>>>>>>>>>
>>>>>>>>> This is quite unfortunate, because there was a different bug that had to do
>>>>>>>>> with kernel having a very limited number of mappings available [1], as a
>>>>>>>>> result of which the page concatenation code was added.
>>>>>>>>>
>>>>>>>>> It should therefore be documented that the dma_entry_limit parameter should
>>>>>>>>> be adjusted should the user run out of the DMA entries.
>>>>>>>>>
>>>>>>>>> [1] https://urldefense.proofpoint.com/v2/url?u=https-3A__lore.kernel.org_lkml_155414977872.12780.13728555131525362206.stgit-40gimli.home_T_&d=DwICaQ&c=nKjWec2b6R0mOyPaz7xtfQ&r=FZ_tPCbgFOh18zwRPO9H0yDx8VW38vuapifdDfc8SFQ&m=3GMg-634_cdUCY4WpQPwjzZ_S4ckuMHOnt2FxyyjXMk&s=TJLzppkaDS95VGyRHX2hzflQfb9XLK0OiOszSXoeXKk&e=
>>>>>>>
>>>>>>> <snip>
>>>>>>>
>>>>>>>>>>                        RTE_LOG(ERR, EAL, "  cannot clear DMA remapping, error %i (%s)\n",
>>>>>>>>>>                                        errno, strerror(errno));
>>>>>>>>>>                        return -1;
>>>>>>>>>> +           } else if (dma_unmap.size != len) {
>>>>>>>>>> +                   RTE_LOG(ERR, EAL, "  unexpected size %"PRIu64" of DMA "
>>>>>>>>>> +                           "remapping cleared instead of %"PRIu64"\n",
>>>>>>>>>> +                           (uint64_t)dma_unmap.size, len);
>>>>>>>>>> +                   rte_errno = EIO;
>>>>>>>>>> +                   return -1;
>>>>>>>>>>                }
>>>>>>>>>>        }
>>>>>>>>>> @@ -1853,6 +1869,12 @@ container_dma_unmap(struct vfio_config *vfio_cfg, uint64_t vaddr, uint64_t iova,
>>>>>>>>>>                /* we're partially unmapping a previously mapped region, so we
>>>>>>>>>>                 * need to split entry into two.
>>>>>>>>>>                 */
>>>>>>>>>> +           if (!vfio_cfg->vfio_iommu_type->partial_unmap) {
>>>>>>>>>> +                   RTE_LOG(DEBUG, EAL, "DMA partial unmap unsupported\n");
>>>>>>>>>> +                   rte_errno = ENOTSUP;
>>>>>>>>>> +                   ret = -1;
>>>>>>>>>> +                   goto out;
>>>>>>>>>> +           }
>>>>>>>>>
>>>>>>>>> How would we ever arrive here if we never do more than 1 page worth of
>>>>>>>>> memory anyway? I don't think this is needed.
>>>>>>>>
>>>>>>>> container_dma_unmap() is called by user via rte_vfio_container_dma_unmap()
>>>>>>>> and when he maps we don't split it as we don't about his memory.
>>>>>>>> So if he maps multiple pages and tries to unmap partially, then we should fail.
>>>>>>>
>>>>>>> Should we map it in page granularity then, instead of adding this
>>>>>>> discrepancy between EAL and user mapping? I.e. instead of adding a
>>>>>>> workaround, how about we just do the same thing for user mem mappings?
>>>>>>>
>>>>>> In heap mapping's we map and unmap it at huge page granularity as we will always
>>>>>> maintain that.
>>>>>>
>>>>>> But here I think we don't know if user's allocation is huge page or
>>>>>> collection of system
>>>>>> pages. Only thing we can do here is map it at system page granularity which
>>>>>> could waste entries if he say really is working with hugepages. Isn't ?
>>>>>>
>>>>>
>>>>> Yeah we do. The API mandates the pages granularity, and it will check
>>>>> against page size and number of IOVA entries, so yes, we do enforce the fact
>>>>> that the IOVA addresses supplied by the user have to be page addresses.
>>>>
>>>> If I see rte_vfio_container_dma_map(), there is no mention of Huge page size
>>>> user is providing or we computing. He can call rte_vfio_container_dma_map()
>>>> with 1GB huge page or 4K system page.
>>>>
>>>> Am I missing something ?
>>>
>>> Are you suggesting that a DMA mapping for hugepage-backed memory will be
>>> made at system page size granularity? E.g. will a 1GB page-backed segment be
>>> mapped for DMA as a contiguous 4K-based block?
>>
>> I'm not suggesting anything. My only thought is how to solve below problem.
>> Say application does the following.
>>
>> #1 Allocate 1GB memory from huge page or some external mem.
>> #2 Do rte_vfio_container_dma_map(RTE_VFIO_DEFAULT_CONTAINER_FD, mem, mem, 1GB)
>>     In linux/eal_vfio.c, we map it is as single VFIO DMA entry of 1 GB as we
>>     don't know where this memory is coming from or backed by what.
>> #3 After a while call rte_vfio_container_dma_unmap(RTE_VFIO_DEFAULT_CONTAINER_FD, mem+4KB, mem+4KB, 4KB)
>>   
>> Though rte_vfio_container_dma_unmap() supports #3 by splitting entry as shown below,
>> In VFIO type1 iommu, #3 cannot be supported by current kernel interface. So how
>> can we allow #3 ?
>>
>>
>> static int
>> container_dma_unmap(struct vfio_config *vfio_cfg, uint64_t vaddr, uint64_t iova,
>>                  uint64_t len)
>> {
>>          struct user_mem_map *map, *new_map = NULL;
>>          struct user_mem_maps *user_mem_maps;
>>          int ret = 0;
>>
>>          user_mem_maps = &vfio_cfg->mem_maps;
>>          rte_spinlock_recursive_lock(&user_mem_maps->lock);
>>
>>          /* find our mapping */
>>          map = find_user_mem_map(user_mem_maps, vaddr, iova, len);
>>          if (!map) {
>>                  RTE_LOG(ERR, EAL, "Couldn't find previously mapped region\n");
>>                  rte_errno = EINVAL;
>>                  ret = -1;
>>                  goto out;
>>          }
>>          if (map->addr != vaddr || map->iova != iova || map->len != len) {
>>                  /* we're partially unmapping a previously mapped region, so we
>>                   * need to split entry into two.
>>                   */

Hi,

Apologies, i was on vacation.

Yes, I can see the problem now. Does VFIO even support non-system page 
sizes? Like, if i allocated a 1GB page, would i be able to map *this 
page* for DMA, as opposed to first 4K of this page? I suspect that the 
mapping doesn't support page sizes other than the system page size.

-- 
Thanks,
Anatoly