Re: [PATCH] mm: don't cap request size based on read-ahead setting

From: Jens Axboe
Date: Wed Nov 16 2016 - 13:39:10 EST

On 11/16/2016 08:12 AM, Jens Axboe wrote:
On 11/16/2016 12:17 AM, Hillf Danton wrote:
On Wednesday, November 16, 2016 12:31 PM Jens Axboe wrote:
@@ -369,10 +369,25 @@ ondemand_readahead(struct address_space *mapping,
bool hit_readahead_marker, pgoff_t offset,
unsigned long req_size)
- unsigned long max = ra->ra_pages;
+ unsigned long io_pages, max_pages;
pgoff_t prev_offset;

+ * If bdi->io_pages is set, that indicates the (soft) max IO size
+ * per command for that device. If we have that available, use
+ * that as the max suitable read-ahead size for this IO. Instead of
+ * capping read-ahead at ra_pages if req_size is larger, we can go
+ * up to io_pages. If io_pages isn't set, fall back to using
+ * ra_pages as a safe max.
+ */
+ io_pages = inode_to_bdi(mapping->host)->io_pages;
+ if (io_pages) {
+ max_pages = max_t(unsigned long, ra->ra_pages, req_size);
+ io_pages = min(io_pages, max_pages);

Doubt if you mean
max_pages = min(io_pages, max_pages);

No, that is what I mean. We want the maximum of the RA setting and the
user IO size, but the minimum of that and the device max command size.

Johannes pointed out that I'm an idiot - a last minute edit introduced
this typo, and I was too blind to spot it when you sent that email this
morning. So yes, it should of course be:

max_pages = min(io_pages, max_pages);

like the first version I posted. I'll post a v3...

Jens Axboe