x86, olpc: Speed up device tree creation during boot

Calling alloc_bootmem() for tiny chunks of memory over and over is really
slow; on an XO-1, it caused the time between when the kernel started
booting and when the display came alive (post-lxfb probe) to increase
to 44s.  This patch optimizes the prom_early_alloc function by
calling alloc_bootmem for 4k-sized blocks of memory, and handing out
chunks of that to callers.  With this patch, the time between kernel load
and display initialization decreased to 23s.  If there's a better way to
do this early in the boot process, please let me know.

(Note: increasing the chunk size to 16k didn't noticably affect boot time,
and wasted 9k.)

v4: clarify comment, requested by hpa
v3: fix wasted memory buglet found by Milton Miller, and style fix.
v2: reorder prom_early_alloc as suggested by Grant.

Signed-off-by: Andres Salomon <dilinger@queued.net>
LKML-Reference: <20101129153951.74202a84@queued.net>
Signed-off-by: H. Peter Anvin <hpa@linux.intel.com>

arch/x86/platform/olpc/olpc_dt.c

@@ -126,14 +126,32 @@ static unsigned int prom_early_allocated __initdata;
 
 void * __init prom_early_alloc(unsigned long size)
 {
+	static u8 *mem;
+	static size_t free_mem;
 	void *res;
 
-	res = alloc_bootmem(size);
-	if (res)
-		memset(res, 0, size);
-
-	prom_early_allocated += size;
+	if (free_mem < size) {
+		const size_t chunk_size = max(PAGE_SIZE, size);
+
+		/*
+		 * To mimimize the number of allocations, grab at least
+		 * PAGE_SIZE of memory (that's an arbitrary choice that's
+		 * fast enough on the platforms we care about while minimizing
+		 * wasted bootmem) and hand off chunks of it to callers.
+		 */
+		res = alloc_bootmem(chunk_size);
+		if (!res)
+			return NULL;
+		prom_early_allocated += chunk_size;
+		memset(res, 0, chunk_size);
+		free_mem = chunk_size;
+		mem = res;
+	}
 
+	/* allocate from the local cache */
+	free_mem -= size;
+	res = mem;
+	mem += size;
 	return res;
 }