Re: IORING_OP_POLL_ADD slower than linux-aio IOCB_CMD_POLL

[Avi Kivity <avi@scylladb.com>]

On 20/04/2022 15.09, Jens Axboe wrote:
> On 4/20/22 5:55 AM, Avi Kivity wrote:
>> On 19/04/2022 22.58, Jens Axboe wrote:
>>>> I'll try it tomorrow (also the other patch).
>>> Thanks!
>>>
>> With the new kernel, I get
>>
>>
>> io_uring_setup(200, {flags=0, sq_thread_cpu=0, sq_thread_idle=0, sq_entries=256, cq_entries=512, features=IORING_FEAT_SINGLE_MMAP|IORING_FEAT_NODROP|IORING_FEAT_SUBMIT_STABLE|IORING_FEAT_RW_CUR_POS|IORING_FEAT_CUR_PERSONALITY|IORING_FEAT_FAST_POLL|IORING_FEAT_POLL_32BITS|IORING_FEAT_SQPOLL_NONFIXED|IORING_FEAT_EXT_ARG|IORING_FEAT_NATIVE_WORKERS|IORING_FEAT_RSRC_TAGS|0x1800, sq_off={head=0, tail=64, ring_mask=256, ring_entries=264, flags=276, dropped=272, array=8512}, cq_off={head=128, tail=192, ring_mask=260, ring_entries=268, overflow=284, cqes=320, flags=280}}) = 7
>> mmap(NULL, 9536, PROT_READ|PROT_WRITE, MAP_SHARED|MAP_POPULATE, 7, 0) = -1 EACCES (Permission denied)
> That looks odd, and not really related at all?
>

It was selinux. So perhaps there is a regression somewhere that breaks 
selinux+io_uring.


The reason I encountered selinux was that I ran it in a virtual machine 
(nvidia + custom kernel = ...). Now the problem is that the results in 
the VM are very inconsistent, perhaps due to incorrect topology exposed 
to the guest. So I can't provide meaningful results from the patches, at 
least until I find a more workable setup.