Re: [GIT PULL] io_uring fixes for 5.10-rc

[Jens Axboe]

On 11/21/20 11:07 AM, Linus Torvalds wrote:
> On Fri, Nov 20, 2020 at 7:00 PM Jens Axboe <axboe@xxxxxxxxx> wrote:
>>
>> Actually, I think we can do even better. How about just having
>> do_filp_open() exit after LOOKUP_RCU fails, if LOOKUP_RCU was already
>> set in the lookup flags? Then we don't need to change much else, and
>> most of it falls out naturally.
> 
> So I was thinking doing the RCU lookup unconditionally, and then doing
> the nn-RCU lookup if that fails afterwards.
> 
> But your patch looks good to me.
> 
> Except for the issue you noticed.

After having taken a closer look, I think the saner approach is
LOOKUP_NONBLOCK instead of using LOOKUP_RCU which is used more as
a state than lookup flag. I'll try and hack something up that looks
passable.

>> Except it seems that should work, except LOOKUP_RCU does not guarantee
>> that we're not going to do IO:
> 
> Well, almost nothing guarantees lack of IO, since allocations etc can
> still block, but..

Sure, and we can't always avoid that - but blatant block on waiting
for IO should be avoided.

>> [   20.463195]  schedule+0x5f/0xd0
>> [   20.463444]  io_schedule+0x45/0x70
>> [   20.463712]  bit_wait_io+0x11/0x50
>> [   20.463981]  __wait_on_bit+0x2c/0x90
>> [   20.464264]  out_of_line_wait_on_bit+0x86/0x90
>> [   20.464611]  ? var_wake_function+0x30/0x30
>> [   20.464932]  __ext4_find_entry+0x2b5/0x410
>> [   20.465254]  ? d_alloc_parallel+0x241/0x4e0
>> [   20.465581]  ext4_lookup+0x51/0x1b0
>> [   20.465855]  ? __d_lookup+0x77/0x120
>> [   20.466136]  path_openat+0x4e8/0xe40
>> [   20.466417]  do_filp_open+0x79/0x100
> 
> Hmm. Is this perhaps an O_CREAT case? I think we only do the dcache
> lookups under RCU, not the final path component creation.

It's just a basic test that opens all files under a directory. So
no O_CREAT, it's all existing files. I think this is just a case of not
aborting early enough for LOOKUP_NONBLOCK, and we've obviously already
dropped LOOKUP_RCU (and done rcu_read_unlock() again) at this point.

> And there are probably lots of other situations where we finish with
> LOOKUP_RCU (with unlazy_walk()), and then continue.> 
> Example: look at "may_lookup()" - if inode_permission() says "I can't
> do this without blocking" the logic actually just tries to validate
> the current state (that "unlazy_walk()" thing), and then continue
> without RCU.
> 
> It obviously hasn't been about lockless semantics, it's been about
> really being lockless. So LOOKUP_RCU has been a "try to do this
> locklessly" rather than "you cannot take any locks".
> 
> I guess we would have to add a LOOKUP_NOBLOCK thing to actually then
> say "if the RCU lookup fails, return -EAGAIN".
> 
> That's probably not a huge undertaking, but yeah, I didn't think of
> it. I think this is a "we need to have Al tell us if it's reasonable".

Definitely. I did have a weak attempt at LOOKUP_NONBLOCK earlier, I'll
try and resurrect it and see what that leads to. Outside of just pure
lookup, the d_revalidate() was a bit interesting as it may block for
certain cases, but those should be (hopefully) detectable upfront.

-- 
Jens Axboe