asynchronous readahead prefetcher operation

asynchronous readahead prefetcher operation

[Vasilis Dimitsas]

[-- Attachment #1: Type: text/plain, Size: 684 bytes --]

Good evening,

I am currently working on a project which is related to the operation of
the linux readahead prefetcher. As a result, I am trying to understand its
operation. Having read thoroughly the relevant part in the kernel code, I
realize, from the comments, that part of the prefetching occurs
asynchronously. The problem is that I can not verify this from the code.

Even if you call page_cache_sync_readahead() or
page_cache_async_readahead(), then both will end up in ra_submit(), in
which, the operation is common for both cases.

So, please could you tell me at which point does the operation of
prefetching occurs asynchronously?

Thank you in advance,

Vasilis Dimitsas

[-- Attachment #2: Type: text/html, Size: 829 bytes --]

Re: asynchronous readahead prefetcher operation

[Matthew Wilcox]

On Wed, Jul 12, 2017 at 11:31:21PM +0300, Vasilis Dimitsas wrote:
> I am currently working on a project which is related to the operation of
> the linux readahead prefetcher. As a result, I am trying to understand its
> operation. Having read thoroughly the relevant part in the kernel code, I
> realize, from the comments, that part of the prefetching occurs
> asynchronously. The problem is that I can not verify this from the code.
> 
> Even if you call page_cache_sync_readahead() or
> page_cache_async_readahead(), then both will end up in ra_submit(), in
> which, the operation is common for both cases.
> 
> So, please could you tell me at which point does the operation of
> prefetching occurs asynchronously?

The prefetching operation always occurs asynchronously; the
I/O is submitted and then both page_cache_sync_readahead() and
page_cache_async_readahead() return to the caller.  They use slightly
different algorithms, which is why they're different functions, but the
I/O is not waited for.  It's up to the caller to do that.

I imagine you're looking at filemap_fault(), and it happens like this:

        page = find_get_page(mapping, offset);
(returns NULL because there's no page in the cache)
                do_sync_mmap_readahead(vmf->vma, ra, file, offset);
(will create pages and put them in the page cache, taking PageLock on each page)
                page = find_get_page(mapping, offset);
(finds the page that was just created)
        if (!lock_page_or_retry(page, vmf->vma->vm_mm, vmf->flags)) {
(will attempt to lock the page ... if it's locked and the fault lets us retry,
fails so we can handle retries at the higher level.  If it's locked and the
fault says we can't retry, then sleeps until unlocked.  If/once it's unlocked,
will return success)

When the I/O completes, the page will be unlocked, usually by calling
page_endio().

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>

Re: asynchronous readahead prefetcher operation

[Vasilis Dimitsas]

[-- Attachment #1: Type: text/plain, Size: 2559 bytes --]

Hello Matthew,

Thank you for your response. Since at user level I am using the pread()
function, in kernel level, unless I am making a mistake, the
do_generic_file_read() is being called. Inside this, the find_get_page() is
called and if the page is not in the page cache then
page_cache_sync_readahead() is called or page_cache_async_readahead() if
the page is marked with the PG_readahead flag. So, I would like to find in
which exact part of the code can someone understand that the I/O is not
waited for.

Thank you again,

Vasilis

On Thu, Jul 13, 2017 at 7:34 PM, Matthew Wilcox <willy@infradead.org> wrote:

> On Wed, Jul 12, 2017 at 11:31:21PM +0300, Vasilis Dimitsas wrote:
> > I am currently working on a project which is related to the operation of
> > the linux readahead prefetcher. As a result, I am trying to understand
> its
> > operation. Having read thoroughly the relevant part in the kernel code, I
> > realize, from the comments, that part of the prefetching occurs
> > asynchronously. The problem is that I can not verify this from the code.
> >
> > Even if you call page_cache_sync_readahead() or
> > page_cache_async_readahead(), then both will end up in ra_submit(), in
> > which, the operation is common for both cases.
> >
> > So, please could you tell me at which point does the operation of
> > prefetching occurs asynchronously?
>
> The prefetching operation always occurs asynchronously; the
> I/O is submitted and then both page_cache_sync_readahead() and
> page_cache_async_readahead() return to the caller.  They use slightly
> different algorithms, which is why they're different functions, but the
> I/O is not waited for.  It's up to the caller to do that.
>
> I imagine you're looking at filemap_fault(), and it happens like this:
>
>         page = find_get_page(mapping, offset);
> (returns NULL because there's no page in the cache)
>                 do_sync_mmap_readahead(vmf->vma, ra, file, offset);
> (will create pages and put them in the page cache, taking PageLock on each
> page)
>                 page = find_get_page(mapping, offset);
> (finds the page that was just created)
>         if (!lock_page_or_retry(page, vmf->vma->vm_mm, vmf->flags)) {
> (will attempt to lock the page ... if it's locked and the fault lets us
> retry,
> fails so we can handle retries at the higher level.  If it's locked and the
> fault says we can't retry, then sleeps until unlocked.  If/once it's
> unlocked,
> will return success)
>
> When the I/O completes, the page will be unlocked, usually by calling
> page_endio().
>

[-- Attachment #2: Type: text/html, Size: 3162 bytes --]

Re: asynchronous readahead prefetcher operation

[Matthew Wilcox]

On Thu, Jul 13, 2017 at 11:36:22PM +0300, Vasilis Dimitsas wrote:
> Hello Matthew,
> 
> Thank you for your response. Since at user level I am using the pread()
> function, in kernel level, unless I am making a mistake, the
> do_generic_file_read() is being called. Inside this, the find_get_page() is
> called and if the page is not in the page cache then
> page_cache_sync_readahead() is called or page_cache_async_readahead() if
> the page is marked with the PG_readahead flag. So, I would like to find in
> which exact part of the code can someone understand that the I/O is not
> waited for.

As I said, the I/O is never waited for by either of these functions.
The _sync_readahead vs _async_readahead calls just use a slightly
different algorithm for choosing which pages to bring in.  In the case
of do_generic_file_read(), if the I/O has not finished by the time the
call returns, then the page will not be marked Uptodate, so we follow
this path:

                if (!PageUptodate(page)) {
                        error = wait_on_page_locked_killable(page);

and that is where we wait for the I/O to complete, no matter whether the
I/O was triggered by the sync or async calls.

> Thank you again,
> 
> Vasilis
> 
> On Thu, Jul 13, 2017 at 7:34 PM, Matthew Wilcox <willy@infradead.org> wrote:
> 
> > On Wed, Jul 12, 2017 at 11:31:21PM +0300, Vasilis Dimitsas wrote:
> > > I am currently working on a project which is related to the operation of
> > > the linux readahead prefetcher. As a result, I am trying to understand
> > its
> > > operation. Having read thoroughly the relevant part in the kernel code, I
> > > realize, from the comments, that part of the prefetching occurs
> > > asynchronously. The problem is that I can not verify this from the code.
> > >
> > > Even if you call page_cache_sync_readahead() or
> > > page_cache_async_readahead(), then both will end up in ra_submit(), in
> > > which, the operation is common for both cases.
> > >
> > > So, please could you tell me at which point does the operation of
> > > prefetching occurs asynchronously?
> >
> > The prefetching operation always occurs asynchronously; the
> > I/O is submitted and then both page_cache_sync_readahead() and
> > page_cache_async_readahead() return to the caller.  They use slightly
> > different algorithms, which is why they're different functions, but the
> > I/O is not waited for.  It's up to the caller to do that.
> >
> > I imagine you're looking at filemap_fault(), and it happens like this:
> >
> >         page = find_get_page(mapping, offset);
> > (returns NULL because there's no page in the cache)
> >                 do_sync_mmap_readahead(vmf->vma, ra, file, offset);
> > (will create pages and put them in the page cache, taking PageLock on each
> > page)
> >                 page = find_get_page(mapping, offset);
> > (finds the page that was just created)
> >         if (!lock_page_or_retry(page, vmf->vma->vm_mm, vmf->flags)) {
> > (will attempt to lock the page ... if it's locked and the fault lets us
> > retry,
> > fails so we can handle retries at the higher level.  If it's locked and the
> > fault says we can't retry, then sleeps until unlocked.  If/once it's
> > unlocked,
> > will return success)
> >
> > When the I/O completes, the page will be unlocked, usually by calling
> > page_endio().
> >

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>