Date:   Thu, 20 Feb 2020 20:19:39 -0800
From:   John Hubbard <jhubbard@...dia.com>
To:     Matthew Wilcox <willy@...radead.org>
CC:     <linux-fsdevel@...r.kernel.org>, <linux-mm@...ck.org>,
        <linux-kernel@...r.kernel.org>, <linux-btrfs@...r.kernel.org>,
        <linux-erofs@...ts.ozlabs.org>, <linux-ext4@...r.kernel.org>,
        <linux-f2fs-devel@...ts.sourceforge.net>,
        <cluster-devel@...hat.com>, <ocfs2-devel@....oracle.com>,
        <linux-xfs@...r.kernel.org>
Subject: Re: [PATCH v7 09/24] mm: Put readahead pages in cache earlier

On 2/20/20 7:43 PM, Matthew Wilcox wrote:
> On Thu, Feb 20, 2020 at 07:19:58PM -0800, John Hubbard wrote:
>>> +static inline struct page *readahead_page(struct readahead_control *rac)
>>> +{
>>> +	struct page *page;
>>> +
>>> +	BUG_ON(rac->_batch_count > rac->_nr_pages);
>>> +	rac->_nr_pages -= rac->_batch_count;
>>> +	rac->_index += rac->_batch_count;
>>> +	rac->_batch_count = 0;
>>
>>
>> Is it intentional, to set rac->_batch_count twice (here, and below)? The
>> only reason I can see is if a caller needs to use ->_batch_count in the
>> "return NULL" case, which doesn't seem to come up...
> 
> Ah, but it does.  Not in this patch, but the next one ...
> 
> +       if (aops->readahead) {
> +               aops->readahead(rac);
> +               /* Clean up the remaining pages */
> +               while ((page = readahead_page(rac))) {
> +                       unlock_page(page);
> +                       put_page(page);
> +               }
> 
> In the normal case, the ->readahead method will consume all the pages,
> and we need readahead_page() to do nothing if it is called again.
> 
>>> +	if (!rac->_nr_pages)
>>> +		return NULL;
> 
> ... admittedly I could do:
> 
> 	if (!rac->_nr_pages) {
> 		rac->_batch_count = 0;
> 		return NULL;
> 	}
> 
> which might be less confusing.


Yes, that would be a nice bit of polish if you end up doing another revision for other
reasons.


> 
>>> @@ -130,23 +129,23 @@ static void read_pages(struct readahead_control *rac, struct list_head *pages,
>>>  				readahead_count(rac));
>>>  		/* Clean up the remaining pages */
>>>  		put_pages_list(pages);
>>> -		goto out;
>>> -	}
>>> -
>>> -	for (page_idx = 0; page_idx < readahead_count(rac); page_idx++) {
>>> -		struct page *page = lru_to_page(pages);
>>> -		list_del(&page->lru);
>>> -		if (!add_to_page_cache_lru(page, rac->mapping, page->index,
>>> -				gfp))
>>> +		rac->_index += rac->_nr_pages;
>>> +		rac->_nr_pages = 0;
>>> +	} else {
>>> +		while ((page = readahead_page(rac))) {
>>>  			aops->readpage(rac->file, page);
>>> -		put_page(page);
>>> +			put_page(page);
>>> +		}
>>>  	}
>>>  
>>> -out:
>>>  	blk_finish_plug(&plug);
>>>  
>>>  	BUG_ON(!list_empty(pages));
>>> -	rac->_nr_pages = 0;
>>> +	BUG_ON(readahead_count(rac));
>>> +
>>> +out:
>>> +	/* If we were called due to a conflicting page, skip over it */
>>
>> Tiny documentation nit: What if we were *not* called due to a conflicting page? 
>> (And what is a "conflicting page", in this context, btw?) The next line unconditionally 
>> moves the index ahead, so the "if" part of the comment really confuses me.
> 
> By the end of the series, read_pages() is called in three places:
> 
> 1.              if (page && !xa_is_value(page)) {
>                         read_pages(&rac, &page_pool);
> 
> 2.              } else if (add_to_page_cache_lru(page, mapping, index + i,
>                                         gfp_mask) < 0) {
>                         put_page(page);
>                         read_pages(&rac, &page_pool);
> 
> 3.      read_pages(&rac, &page_pool);
> 
> In the first two cases, there's an existing page in the page cache
> (which conflicts with this readahead operation), and so we need to
> advance index.  In the third case, we're exiting the function, so it
> does no harm to advance index one further.


OK, I see. As you know, I tend toward maybe over-documenting, but what about
adding just a *few* hints to help new readers, like this approximately (maybe
it should be pared down):


diff --git a/mm/readahead.c b/mm/readahead.c
index 9fb5f77dcf69..0dd5b09c376e 100644
--- a/mm/readahead.c
+++ b/mm/readahead.c
@@ -114,6 +114,10 @@ int read_cache_pages(struct address_space *mapping, struct list_head *pages,
 
 EXPORT_SYMBOL(read_cache_pages);
 
+/*
+ * Read pages into the page cache, OR skip over a page if it is already in the
+ * page cache.
+ */
 static void read_pages(struct readahead_control *rac, struct list_head *pages)
 {
        const struct address_space_operations *aops = rac->mapping->a_ops;
@@ -152,7 +156,11 @@ static void read_pages(struct readahead_control *rac, struct list_head *pages)
        BUG_ON(readahead_count(rac));
 
 out:
-       /* If we were called due to a conflicting page, skip over it */
+       /*
+        * This routine might have been called in order to skip over a page
+        * that is already in the page cache. And for other cases, the index is
+        * ignored by the caller. So just increment unconditionally:
+        */
        rac->_index++;
 }


?

> 
>>> +		} else if (add_to_page_cache_lru(page, mapping, index + i,
>>> +					gfp_mask) < 0) {
>>
>> I still think you'll want to compare against !=0, rather than < 0, here.
> 
> I tend to prefer < 0 when checking for an error value in case the function
> decides to start using positive numbers to mean something.  I don't think
> it's a particularly important preference though (after all, returning 1
> might mean "failed, but for this weird reason rather than an errno").
> 
>>> +			put_page(page);
>>> +			read_pages(&rac, &page_pool);
>>
>> Doing a read_pages() in the error case is because...actually, I'm not sure yet.
>> Why do we do this? Effectively it's a retry?
> 
> Same as the reason we call read_pages() if we found a page in the page
> cache earlier -- we're sending down a set of pages which are consecutive
> in the file's address space, and now we have to skip one.  At least one ;-)
> 

Got it. Finally. :)


thanks,
-- 
John Hubbard
NVIDIA

Powered by blists - more mailing lists