x86: Implement arch_phys_memset so that it can wipe memory above 4GB

Implement arch_phys_memset so that it can set memory at physical addresses
above 4GB using PAE paging. Because there are only 5 page tables in PAE mode,
1 PDPT and 4 PDTs, those tables are statically allocated in the BSS. The
tables must be 4K page aligned and are declared that way, and because U-Boot
starts as 4K aligned and the relocation code relocates it to a 4K aligned
address, the tables work as intended.

While paging is turned on, all 4GB are identity mapped except for one 2MB
page which is used as the window into high memory. This way, U-Boot will
continue to work as expected when running code that expects to access memory
freely, but the code can still get at high memory through its window.

The window is put at 2MB so that it's 2MB page aligned, low in memory to be
out of the way of things U-Boot is likely to care about, and above the lowest
1MB where lots of random things live.

Signed-off-by: Gabe Black <gabeblack@chromium.org>

Signed-off-by: Simon Glass <sjg@chromium.org>

2 files changed, 229 insertions, 0 deletions

diff --git a/arch/x86/lib/Makefile b/arch/x86/lib/Makefile
index 4325b2502e..2a3e8f00cd 100644
--- a/arch/x86/lib/Makefile
+++ b/arch/x86/lib/Makefile
@@ -47,6 +47,7 @@ COBJS-$(CONFIG_SYS_GENERIC_TIMER) += pcat_timer.o
 COBJS-$(CONFIG_PCI) += pci.o
 COBJS-$(CONFIG_PCI) += pci_type1.o
 COBJS-y	+= relocate.o
+COBJS-y += physmem.o
 COBJS-y	+= string.o
 COBJS-$(CONFIG_SYS_X86_ISR_TIMER)	+= timer.o
 COBJS-$(CONFIG_VIDEO)	+= video.o
diff --git a/arch/x86/lib/physmem.c b/arch/x86/lib/physmem.c
new file mode 100644
index 0000000000..18f0e62acd
--- /dev/null
+++ b/arch/x86/lib/physmem.c
@@ -0,0 +1,228 @@
+/*
+ * Copyright (c) 2012 The Chromium OS Authors. All rights reserved.
+ * Use of this source code is governed by a BSD-style license that can be
+ * found in the LICENSE file.
+ *
+ * Alternatively, this software may be distributed under the terms of the
+ * GNU General Public License ("GPL") version 2 as published by the Free
+ * Software Foundation.
+ */
+
+#include <common.h>
+#include <physmem.h>
+#include <linux/compiler.h>
+
+/* Large pages are 2MB. */
+#define LARGE_PAGE_SIZE ((1 << 20) * 2)
+
+/*
+ * Paging data structures.
+ */
+
+struct pdpe {
+	uint64_t p:1;
+	uint64_t mbz_0:2;
+	uint64_t pwt:1;
+	uint64_t pcd:1;
+	uint64_t mbz_1:4;
+	uint64_t avl:3;
+	uint64_t base:40;
+	uint64_t mbz_2:12;
+};
+
+typedef struct pdpe pdpt_t[512];
+
+struct pde {
+	uint64_t p:1;      /* present */
+	uint64_t rw:1;     /* read/write */
+	uint64_t us:1;     /* user/supervisor */
+	uint64_t pwt:1;    /* page-level writethrough */
+	uint64_t pcd:1;    /* page-level cache disable */
+	uint64_t a:1;      /* accessed */
+	uint64_t d:1;      /* dirty */
+	uint64_t ps:1;     /* page size */
+	uint64_t g:1;      /* global page */
+	uint64_t avl:3;    /* available to software */
+	uint64_t pat:1;    /* page-attribute table */
+	uint64_t mbz_0:8;  /* must be zero */
+	uint64_t base:31;  /* base address */
+};
+
+typedef struct pde pdt_t[512];
+
+static pdpt_t pdpt __aligned(4096);
+static pdt_t pdts[4] __aligned(4096);
+
+/*
+ * Map a virtual address to a physical address and optionally invalidate any
+ * old mapping.
+ *
+ * @param virt		The virtual address to use.
+ * @param phys		The physical address to use.
+ * @param invlpg	Whether to use invlpg to clear any old mappings.
+ */
+static void x86_phys_map_page(uintptr_t virt, phys_addr_t phys, int invlpg)
+{
+	/* Extract the two bit PDPT index and the 9 bit PDT index. */
+	uintptr_t pdpt_idx = (virt >> 30) & 0x3;
+	uintptr_t pdt_idx = (virt >> 21) & 0x1ff;
+
+	/* Set up a handy pointer to the appropriate PDE. */
+	struct pde *pde = &(pdts[pdpt_idx][pdt_idx]);
+
+	memset(pde, 0, sizeof(struct pde));
+	pde->p = 1;
+	pde->rw = 1;
+	pde->us = 1;
+	pde->ps = 1;
+	pde->base = phys >> 21;
+
+	if (invlpg) {
+		/* Flush any stale mapping out of the TLBs. */
+		__asm__ __volatile__(
+			"invlpg %0\n\t"
+			:
+			: "m" (*(uint8_t *)virt)
+		);
+	}
+}
+
+/* Identity map the lower 4GB and turn on paging with PAE. */
+static void x86_phys_enter_paging(void)
+{
+	phys_addr_t page_addr;
+	unsigned i;
+
+	/* Zero out the page tables. */
+	memset(pdpt, 0, sizeof(pdpt));
+	memset(pdts, 0, sizeof(pdts));
+
+	/* Set up the PDPT. */
+	for (i = 0; i < ARRAY_SIZE(pdts); i++) {
+		pdpt[i].p = 1;
+		pdpt[i].base = ((uintptr_t)&pdts[i]) >> 12;
+	}
+
+	/* Identity map everything up to 4GB. */
+	for (page_addr = 0; page_addr < (1ULL << 32);
+			page_addr += LARGE_PAGE_SIZE) {
+		/* There's no reason to invalidate the TLB with paging off. */
+		x86_phys_map_page(page_addr, page_addr, 0);
+	}
+
+	/* Turn on paging */
+	__asm__ __volatile__(
+		/* Load the page table address */
+		"movl	%0, %%cr3\n\t"
+		/* Enable pae */
+		"movl	%%cr4, %%eax\n\t"
+		"orl	$0x00000020, %%eax\n\t"
+		"movl	%%eax, %%cr4\n\t"
+		/* Enable paging */
+		"movl	%%cr0, %%eax\n\t"
+		"orl	$0x80000000, %%eax\n\t"
+		"movl	%%eax, %%cr0\n\t"
+		:
+		: "r" (pdpt)
+		: "eax"
+	);
+}
+
+/* Disable paging and PAE mode. */
+static void x86_phys_exit_paging(void)
+{
+	/* Turn off paging */
+	__asm__ __volatile__ (
+		/* Disable paging */
+		"movl	%%cr0, %%eax\n\t"
+		"andl	$0x7fffffff, %%eax\n\t"
+		"movl	%%eax, %%cr0\n\t"
+		/* Disable pae */
+		"movl	%%cr4, %%eax\n\t"
+		"andl	$0xffffffdf, %%eax\n\t"
+		"movl	%%eax, %%cr4\n\t"
+		:
+		:
+		: "eax"
+	);
+}
+
+/*
+ * Set physical memory to a particular value when the whole region fits on one
+ * page.
+ *
+ * @param map_addr	The address that starts the physical page.
+ * @param offset	How far into that page to start setting a value.
+ * @param c		The value to set memory to.
+ * @param size		The size in bytes of the area to set.
+ */
+static void x86_phys_memset_page(phys_addr_t map_addr, uintptr_t offset, int c,
+				 unsigned size)
+{
+	/*
+	 * U-Boot should be far away from the beginning of memory, so that's a
+	 * good place to map our window on top of.
+	 */
+	const uintptr_t window = LARGE_PAGE_SIZE;
+
+	/* Make sure the window is below U-Boot. */
+	assert(window + LARGE_PAGE_SIZE <
+	       gd->relocaddr - CONFIG_SYS_MALLOC_LEN - CONFIG_SYS_STACK_SIZE);
+	/* Map the page into the window and then memset the appropriate part. */
+	x86_phys_map_page(window, map_addr, 1);
+	memset((void *)(window + offset), c, size);
+}
+
+/*
+ * A physical memory anologue to memset with matching parameters and return
+ * value.
+ */
+phys_addr_t arch_phys_memset(phys_addr_t start, int c, phys_size_t size)
+{
+	const phys_addr_t max_addr = (phys_addr_t)~(uintptr_t)0;
+	const phys_addr_t orig_start = start;
+
+	if (!size)
+		return orig_start;
+
+	/* Handle memory below 4GB. */
+	if (start <= max_addr) {
+		phys_size_t low_size = MIN(max_addr + 1 - start, size);
+		void *start_ptr = (void *)(uintptr_t)start;
+
+		assert(((phys_addr_t)(uintptr_t)start) == start);
+		memset(start_ptr, c, low_size);
+		start += low_size;
+		size -= low_size;
+	}
+
+	/* Use paging and PAE to handle memory above 4GB up to 64GB. */
+	if (size) {
+		phys_addr_t map_addr = start & ~(LARGE_PAGE_SIZE - 1);
+		phys_addr_t offset = start - map_addr;
+
+		x86_phys_enter_paging();
+
+		/* Handle the first partial page. */
+		if (offset) {
+			phys_addr_t end =
+				MIN(map_addr + LARGE_PAGE_SIZE, start + size);
+			phys_size_t cur_size = end - start;
+			x86_phys_memset_page(map_addr, offset, c, cur_size);
+			size -= cur_size;
+			map_addr += LARGE_PAGE_SIZE;
+		}
+		/* Handle the complete pages. */
+		while (size > LARGE_PAGE_SIZE) {
+			x86_phys_memset_page(map_addr, 0, c, LARGE_PAGE_SIZE);
+			size -= LARGE_PAGE_SIZE;
+			map_addr += LARGE_PAGE_SIZE;
+		}
+		/* Handle the last partial page. */
+		if (size)
+			x86_phys_memset_page(map_addr, 0, c, size);
+
+		x86_phys_exit_paging();
+	}
+	return orig_start;
+}