Re: [PATCH v3 RESEND] mm: shmem: implement POSIX_FADV_[WILL|DONT]NEED for shmem

[Charan Teja Kalla]

Thanks Mark for the review!!

On 1/7/2022 5:40 PM, Mark Hemment wrote:
> On Thu, 6 Jan 2022 at 17:06, Charan Teja Reddy
> <quic_charante@xxxxxxxxxxx> wrote:
>>
>> From: Charan Teja Reddy <charante@xxxxxxxxxxxxxx>
>>
>> Currently fadvise(2) is supported only for the files that doesn't
>> associated with noop_backing_dev_info thus for the files, like shmem,
>> fadvise results into NOP. But then there is file_operations->fadvise()
>> that lets the file systems to implement their own fadvise
>> implementation. Use this support to implement some of the POSIX_FADV_XXX
>> functionality for shmem files.
>>
>> [snip]
> 
>> +static int shmem_fadvise_willneed(struct address_space *mapping,
>> +                                pgoff_t start, pgoff_t long end)
>> +{
>> +       XA_STATE(xas, &mapping->i_pages, start);
>> +       struct page *page;
>> +
>> +       rcu_read_lock();
>> +       xas_for_each(&xas, page, end) {
>> +               if (!xa_is_value(page))
>> +                       continue;
>> +               xas_pause(&xas);
>> +               rcu_read_unlock();
>> +
>> +               page = shmem_read_mapping_page(mapping, xas.xa_index);
>> +               if (!IS_ERR(page))
>> +                       put_page(page);
>> +
>> +               rcu_read_lock();
>> +               if (need_resched()) {
>> +                       xas_pause(&xas);
>> +                       cond_resched_rcu();
>> +               }
>> +       }
>> +       rcu_read_unlock();
>> +
>> +       return 0;
> 
> I have a doubt on referencing xa_index after calling xas_pause().
> xas_pause() walks xa_index forward, so will not be the value expected
> for the current page.

Agree here. I should have the better test case to verify my changes.

> Also, not necessary to re-call xas_pause() before cond_resched (it is
> a no-op).

In the event when CONFIG_DEBUG_ATOMIC_SLEEP is enabled users may still
need to call the xas_pause(), as we are dropping the rcu lock. NO?

static inline void cond_resched_rcu(void)
{
#if defined(CONFIG_DEBUG_ATOMIC_SLEEP) || !defined(CONFIG_PREEMPT_RCU)
        rcu_read_unlock();
        cond_resched();
        rcu_read_lock();
#endif
}

> Would be better to check need_resched() before
> rcu_read_lock().

Okay, I can directly use cond_resched() if used before rcu_read_lock().

> 
> As this loop may call xas_pause() for most iterations, should consider
> using xa_for_each() instead (I *think* - still getting up to speed
> with XArray).

Even the xarray documentation says that: If most entries found during a
walk require you to call xas_pause(), the xa_for_each() iterator may be
more appropriate.

Since every value entry found in the xarray requires me to do the
xas_pause(), I do agree that xa_for_each() is the appropriate call here.
Will switch to this in the next spin. Waiting for further review
comments on this patch.

> 
> Mark
>