Re: [BUG] commit 3c517a61, slab: better fallback allocation behavior

[Christoph Lameter <clameter@sgi.com>]

On Sun, 10 Dec 2006, Andrew Morton wrote:

> > I fixed the cpuset_zone_allowed() call from fallback_alloc() to avoid
> > sleeping.  Notice the __GFP_HARDWALL added in Linus's version, or the
> > new function cpuset_zone_allowed_hardwall() in Andrew's version, all
> > done in the last week.
> > 
> > But apparently kmem_getpages() can also sleep, as it calls __alloc_pages().

Yes fallback_alloc calls kmem_getpages() and if one does not set the gfp 
flags properly __alloc_pages will sleep. fallback_alloc() uses the passed 
flags so if kmem_cache_alloc is called with the proper flags set then 
__alloc_pages() will not sleep.

In this case kmem_cache_alloc() is called from 
journal_alloc_journal_head() with GFP_NOFS set which is 

#define GFP_NOFS        (__GFP_WAIT | __GFP_IO)

__GFP_WAIT means

#define __GFP_WAIT      ((__force gfp_t)0x10u)  /* Can wait and reschedule? */

Thus the caller allows __alloc_pages to sleep.


/*
 * journal_head splicing and dicing
 */
static struct journal_head *journal_alloc_journal_head(void)
{
        struct journal_head *ret;
        static unsigned long last_warning;

#ifdef CONFIG_JBD_DEBUG
        atomic_inc(&nr_journal_heads);
#endif
        ret = kmem_cache_alloc(journal_head_cache, GFP_NOFS);
        if (ret == 0) {
                jbd_debug(1, "out of memory for journal_head\n");
                if (time_after(jiffies, last_warning + 5*HZ)) {
                        printk(KERN_NOTICE "ENOMEM in %s, retrying.\n",
                               __FUNCTION__);
                        last_warning = jiffies;
                }
                while (ret == 0) {
                        yield();
                        ret = kmem_cache_alloc(journal_head_cache, 
GFP_NOFS);
                }
        }
        return ret;
}