Revert "Disable vmalloc() direct reclaim"

This reverts commit 2092cf6.  The
use of the PF_MEMALLOC flag was always a hack to work around memory
reclaim deadlocks.  Those issues are believed to be resolved so this
workaround can be safely reverted.

Signed-off-by: Brian Behlendorf <behlendorf1@llnl.gov>

module/spl/spl-kmem.c

@@ -840,31 +840,11 @@ kv_alloc(spl_kmem_cache_t *skc, int size, int flags)
 
 	ASSERT(ISP2(size));
 
-	if (skc->skc_flags & KMC_KMEM) {
+	if (skc->skc_flags & KMC_KMEM)
 		ptr = (void *)__get_free_pages(flags, get_order(size));
-	} else {
-		/*
-		 * As part of vmalloc() an __pte_alloc_kernel() allocation
-		 * may occur.  This internal allocation does not honor the
-		 * gfp flags passed to vmalloc().  This means even when
-		 * vmalloc(GFP_NOFS) is called it is possible synchronous
-		 * reclaim will occur.  This reclaim can trigger file IO
-		 * which can result in a deadlock.  This issue can be avoided
-		 * by explicitly setting PF_MEMALLOC on the process to
-		 * subvert synchronous reclaim.  The following bug has
-		 * been filed at kernel.org to track the issue.
-		 *
-		 * https://bugzilla.kernel.org/show_bug.cgi?id=30702
-		 */
-		if (!(flags & __GFP_FS))
-			current->flags |= PF_MEMALLOC;
-
+	else
 		ptr = __vmalloc(size, flags | __GFP_HIGHMEM, PAGE_KERNEL);
 
-		if (!(flags & __GFP_FS))
-			current->flags &= ~PF_MEMALLOC;
-	}
-
 	/* Resulting allocated memory will be page aligned */
 	ASSERT(IS_P2ALIGNED(ptr, PAGE_SIZE));