Re: [dpdk-dev] [PATCH] vhost: fix mmap failure as len not aligned with hugepage size

["Xie, Huawei" <huawei.xie@intel.com>]

On 11/12/2015 10:35 AM, Tan, Jianfeng wrote:
>
>> -----Original Message-----
>> From: Xie, Huawei
>> Sent: Wednesday, November 11, 2015 11:57 AM
>> To: Tan, Jianfeng; dev@dpdk.org
>> Subject: Re: [dpdk-dev] [PATCH] vhost: fix mmap failure as len not aligned with
>> hugepage size
>>
>> On 10/30/2015 2:52 PM, Jianfeng Tan wrote:
>>> This patch fixes a bug under lower version linux kernel, mmap() fails
>>> when
>> Since which version Linux hugetlbfs changes the requirement of size alignment?
>>> length is not aligned with hugepage size.
> This link shows this bug was fixed in Linux kernel commit: dab2d3dc45ae7343216635d981d43637e1cb7d45
> After my check, that patch was applied to long term version 3.4.110+
> So distributions using 2.6.32 and 3.2.72 need this patch to make vhost work well.
> https://bugzilla.kernel.org/show_bug.cgi?id=56881
OK, please add this in commit message, remove unnecessary RTE_ALIGN in
free_memory_region, and add comment to the code because our fix is a
workaround to kernel hugetlbfs implementation issue.
>
>>> Signed-off-by: Jianfeng Tan <jianfeng.tan@intel.com>
>>> ---
>>>  lib/librte_vhost/vhost_user/virtio-net-user.c | 12 +++++++++---
>>>  1 file changed, 9 insertions(+), 3 deletions(-)
>>>
>>> diff --git a/lib/librte_vhost/vhost_user/virtio-net-user.c
>>> b/lib/librte_vhost/vhost_user/virtio-net-user.c
>>> index a998ad8..641561c 100644
>>> --- a/lib/librte_vhost/vhost_user/virtio-net-user.c
>>> +++ b/lib/librte_vhost/vhost_user/virtio-net-user.c
>>> @@ -147,6 +147,10 @@ user_set_mem_table(struct vhost_device_ctx ctx,
>> struct VhostUserMsg *pmsg)
>>>  		/* This is ugly */
>>>  		mapped_size = memory.regions[idx].memory_size +
>>>  			memory.regions[idx].mmap_offset;
>>> +
>>> +		alignment = get_blk_size(pmsg->fds[idx]);
>>> +		mapped_size = RTE_ALIGN_CEIL(mapped_size, alignment);
>> Probably we could remove the alignment of mapped size in free_mem_region as
>> well.
> Yes, after aligning mapped_address when mmap(), this address does not need to be aligned again
> when munmap(). But this will effect nothing, or incur any performance issue. I'm prone to take no
> change to it.
>
>>    RTE_ALIGN_CEIL(
>>         region[idx].mapped_size, alignment) If we are not sure, leave it as it is.
>>> +
>>>  		mapped_address = (uint64_t)(uintptr_t)mmap(NULL,
>>>  			mapped_size,
>>>  			PROT_READ | PROT_WRITE, MAP_SHARED, @@ -154,9
>> +158,11 @@
>>> user_set_mem_table(struct vhost_device_ctx ctx, struct VhostUserMsg *pmsg)
>>>  			0);
>>>
>>>  		RTE_LOG(INFO, VHOST_CONFIG,
>>> -			"mapped region %d fd:%d to %p sz:0x%"PRIx64"
>> off:0x%"PRIx64"\n",
>>> +			"mapped region %d fd:%d to:%p sz:0x%"PRIx64" "
>>> +			"off:0x%"PRIx64" align:0x%"PRIx64"\n",
>>>  			idx, pmsg->fds[idx], (void *)(uintptr_t)mapped_address,
>>> -			mapped_size, memory.regions[idx].mmap_offset);
>>> +			mapped_size, memory.regions[idx].mmap_offset,
>>> +			alignment);
>>>
>>>  		if (mapped_address == (uint64_t)(uintptr_t)MAP_FAILED) {
>>>  			RTE_LOG(ERR, VHOST_CONFIG,
>>> @@ -166,7 +172,7 @@ user_set_mem_table(struct vhost_device_ctx ctx,
>>> struct VhostUserMsg *pmsg)
>>>
>>>  		pregion_orig[idx].mapped_address = mapped_address;
>>>  		pregion_orig[idx].mapped_size = mapped_size;
>>> -		pregion_orig[idx].blksz = get_blk_size(pmsg->fds[idx]);
>>> +		pregion_orig[idx].blksz = alignment;
>>>  		pregion_orig[idx].fd = pmsg->fds[idx];
>>>
>>>  		mapped_address +=  memory.regions[idx].mmap_offset;
>