x86/kprobes: Fix crash when probing CS CALL instructionsgrafted

When using eBPF to probe CS CALL instructions within a function,
a crash can be triggered.

The eBPF tool probes offset 257 of the __hrtimer_run_queues()
function:

<__hrtimer_run_queues+249>:  nopl   0x0(%rax,%rax,1)
<__hrtimer_run_queues+254>:  mov    %r14,%rdi
<__hrtimer_run_queues+257>:  cs call <__x86_indirect_thunk_r12>
<__hrtimer_run_queues+263>:  mov    %eax,%r12d
<__hrtimer_run_queues+266>:  xchg   %ax,%ax
<__hrtimer_run_queues+268>:  mov    %r13,%rdi

Which triggers this crash:

  BUG: unable to handle page fault for address: 00000000000f41c9
  #PF: supervisor write access in kernel mode
  #PF: error_code(0x0002) - not-present page
  PGD 0 P4D 0
  Oops: 0002 [#1] SMP NOPTI
  CPU: 1 PID: 0 Comm: swapper/1 Kdump: loaded Tainted: P
  RIP: 0010:__hrtimer_run_queues+0x106/0x230

Note that __hrtimer_run_queues+0x106 is __hrtimer_run_queues+262, which is
at the 6th byte of the above CS CALL instruction. Since the CS CALL
instruction occupies 6 bytes, the exception occurred in the middle of that
call instruction.

The root cause is that when using eBPF tools to probe in the middle of a
function, a kprobe with INT3 is used as the underlying implementation.

During single-step emulation of the original CALL instruction,
int3_emulate_call() assumes that the probed CALL instruction is 5 bytes
long. However, the actual CS-prefixed CALL instruction occupies 6 bytes,
so it constructs an incorrect exception return address. When the CPU
returns from the kprobe handler, the next instruction to be executed is at
the address of the last byte of that CS CALL instruction. Coincidentally,
starting from that address, the CPU fetches and decodes a completely
different instruction, which ultimately triggers a kernel crash.

Fix the issue by using the actual instruction length obtained from
the instruction decoder when constructing the exception return
address, rather than relying on the hardcoded CALL_INSN_SIZE macro.

[ mingo: Refined the changelog ]

Fixes: 6256e668b7af ("x86/kprobes: Use int3 instead of debug trap for single-step")
Suggested-by: Masami Hiramatsu (Google) <mhiramat@kernel.org>
Signed-off-by: Jinke Han <jinkehan@didiglobal.com>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Reviewed-by: Masami Hiramatsu (Google) <mhiramat@kernel.org>
Acked-by: Yafang Shao <laoar.shao@gmail.com>
Acked-by: Borislav Petkov <bp@alien8.de>
Cc: Peter Zijlstra <peterz@infradead.org>
Link: https://patch.msgid.link/20260908073742.GA10517@didi-ThinkCentre-M920t-N000

1 files changed, 670 insertions, 0 deletions

diff --git a/arch/mips/mm/page.c b/arch/mips/mm/page.c
new file mode 100644
index 000000000..1df237bd4
--- /dev/null
+++ b/arch/mips/mm/page.c
@@ -0,0 +1,670 @@
+/*
+ * This file is subject to the terms and conditions of the GNU General Public
+ * License.  See the file "COPYING" in the main directory of this archive
+ * for more details.
+ *
+ * Copyright (C) 2003, 04, 05 Ralf Baechle (ralf@linux-mips.org)
+ * Copyright (C) 2007  Maciej W. Rozycki
+ * Copyright (C) 2008  Thiemo Seufer
+ * Copyright (C) 2012  MIPS Technologies, Inc.
+ */
+#include <linux/kernel.h>
+#include <linux/sched.h>
+#include <linux/smp.h>
+#include <linux/mm.h>
+#include <linux/proc_fs.h>
+
+#include <asm/bugs.h>
+#include <asm/cacheops.h>
+#include <asm/cpu-type.h>
+#include <asm/inst.h>
+#include <asm/io.h>
+#include <asm/page.h>
+#include <asm/prefetch.h>
+#include <asm/bootinfo.h>
+#include <asm/mipsregs.h>
+#include <asm/mmu_context.h>
+#include <asm/regdef.h>
+#include <asm/cpu.h>
+
+#ifdef CONFIG_SIBYTE_DMA_PAGEOPS
+#include <asm/sibyte/sb1250.h>
+#include <asm/sibyte/sb1250_regs.h>
+#include <asm/sibyte/sb1250_dma.h>
+#endif
+
+#include <asm/uasm.h>
+
+/* Handle labels (which must be positive integers). */
+enum label_id {
+	label_clear_nopref = 1,
+	label_clear_pref,
+	label_copy_nopref,
+	label_copy_pref_both,
+	label_copy_pref_store,
+};
+
+UASM_L_LA(_clear_nopref)
+UASM_L_LA(_clear_pref)
+UASM_L_LA(_copy_nopref)
+UASM_L_LA(_copy_pref_both)
+UASM_L_LA(_copy_pref_store)
+
+/* We need one branch and therefore one relocation per target label. */
+static struct uasm_label labels[5];
+static struct uasm_reloc relocs[5];
+
+#define cpu_is_r4600_v1_x()	((read_c0_prid() & 0xfffffff0) == 0x00002010)
+#define cpu_is_r4600_v2_x()	((read_c0_prid() & 0xfffffff0) == 0x00002020)
+
+/*
+ * R6 has a limited offset of the pref instruction.
+ * Skip it if the offset is more than 9 bits.
+ */
+#define _uasm_i_pref(a, b, c, d)		\
+do {						\
+	if (cpu_has_mips_r6) {			\
+		if (c <= 0xff && c >= -0x100)	\
+			uasm_i_pref(a, b, c, d);\
+	} else {				\
+		uasm_i_pref(a, b, c, d);	\
+	}					\
+} while(0)
+
+static int pref_bias_clear_store;
+static int pref_bias_copy_load;
+static int pref_bias_copy_store;
+
+static u32 pref_src_mode;
+static u32 pref_dst_mode;
+
+static int clear_word_size;
+static int copy_word_size;
+
+static int half_clear_loop_size;
+static int half_copy_loop_size;
+
+static int cache_line_size;
+#define cache_line_mask() (cache_line_size - 1)
+
+static inline void
+pg_addiu(u32 **buf, unsigned int reg1, unsigned int reg2, unsigned int off)
+{
+	if (cpu_has_64bit_gp_regs &&
+	    IS_ENABLED(CONFIG_CPU_DADDI_WORKAROUNDS) &&
+	    r4k_daddiu_bug()) {
+		if (off > 0x7fff) {
+			uasm_i_lui(buf, GPR_T9, uasm_rel_hi(off));
+			uasm_i_addiu(buf, GPR_T9, GPR_T9, uasm_rel_lo(off));
+		} else
+			uasm_i_addiu(buf, GPR_T9, GPR_ZERO, off);
+		uasm_i_daddu(buf, reg1, reg2, GPR_T9);
+	} else {
+		if (off > 0x7fff) {
+			uasm_i_lui(buf, GPR_T9, uasm_rel_hi(off));
+			uasm_i_addiu(buf, GPR_T9, GPR_T9, uasm_rel_lo(off));
+			UASM_i_ADDU(buf, reg1, reg2, GPR_T9);
+		} else
+			UASM_i_ADDIU(buf, reg1, reg2, off);
+	}
+}
+
+static void set_prefetch_parameters(void)
+{
+	if (cpu_has_64bit_gp_regs || cpu_has_64bit_zero_reg)
+		clear_word_size = 8;
+	else
+		clear_word_size = 4;
+
+	if (cpu_has_64bit_gp_regs)
+		copy_word_size = 8;
+	else
+		copy_word_size = 4;
+
+	/*
+	 * The pref's used here are using "streaming" hints, which cause the
+	 * copied data to be kicked out of the cache sooner.  A page copy often
+	 * ends up copying a lot more data than is commonly used, so this seems
+	 * to make sense in terms of reducing cache pollution, but I've no real
+	 * performance data to back this up.
+	 */
+	if (cpu_has_prefetch) {
+		/*
+		 * XXX: Most prefetch bias values in here are based on
+		 * guesswork.
+		 */
+		cache_line_size = cpu_dcache_line_size();
+		switch (current_cpu_type()) {
+		case CPU_R5500:
+		case CPU_TX49XX:
+			/* These processors only support the Pref_Load. */
+			pref_bias_copy_load = 256;
+			break;
+
+		case CPU_R10000:
+		case CPU_R12000:
+		case CPU_R14000:
+		case CPU_R16000:
+			/*
+			 * Those values have been experimentally tuned for an
+			 * Origin 200.
+			 */
+			pref_bias_clear_store = 512;
+			pref_bias_copy_load = 256;
+			pref_bias_copy_store = 256;
+			pref_src_mode = Pref_LoadStreamed;
+			pref_dst_mode = Pref_StoreStreamed;
+			break;
+
+		case CPU_SB1:
+		case CPU_SB1A:
+			pref_bias_clear_store = 128;
+			pref_bias_copy_load = 128;
+			pref_bias_copy_store = 128;
+			/*
+			 * SB1 pass1 Pref_LoadStreamed/Pref_StoreStreamed
+			 * hints are broken.
+			 */
+			if (current_cpu_type() == CPU_SB1 &&
+			    (current_cpu_data.processor_id & 0xff) < 0x02) {
+				pref_src_mode = Pref_Load;
+				pref_dst_mode = Pref_Store;
+			} else {
+				pref_src_mode = Pref_LoadStreamed;
+				pref_dst_mode = Pref_StoreStreamed;
+			}
+			break;
+
+		case CPU_LOONGSON64:
+			/* Loongson-3 only support the Pref_Load/Pref_Store. */
+			pref_bias_clear_store = 128;
+			pref_bias_copy_load = 128;
+			pref_bias_copy_store = 128;
+			pref_src_mode = Pref_Load;
+			pref_dst_mode = Pref_Store;
+			break;
+
+		default:
+			pref_bias_clear_store = 128;
+			pref_bias_copy_load = 256;
+			pref_bias_copy_store = 128;
+			pref_src_mode = Pref_LoadStreamed;
+			if (cpu_has_mips_r6)
+				/*
+				 * Bit 30 (Pref_PrepareForStore) has been
+				 * removed from MIPS R6. Use bit 5
+				 * (Pref_StoreStreamed).
+				 */
+				pref_dst_mode = Pref_StoreStreamed;
+			else
+				pref_dst_mode = Pref_PrepareForStore;
+			break;
+		}
+	} else {
+		if (cpu_has_cache_cdex_s)
+			cache_line_size = cpu_scache_line_size();
+		else if (cpu_has_cache_cdex_p)
+			cache_line_size = cpu_dcache_line_size();
+	}
+	/*
+	 * Too much unrolling will overflow the available space in
+	 * clear_space_array / copy_page_array.
+	 */
+	half_clear_loop_size = min(16 * clear_word_size,
+				   max(cache_line_size >> 1,
+				       4 * clear_word_size));
+	half_copy_loop_size = min(16 * copy_word_size,
+				  max(cache_line_size >> 1,
+				      4 * copy_word_size));
+}
+
+static void build_clear_store(u32 **buf, int off)
+{
+	if (cpu_has_64bit_gp_regs || cpu_has_64bit_zero_reg) {
+		uasm_i_sd(buf, GPR_ZERO, off, GPR_A0);
+	} else {
+		uasm_i_sw(buf, GPR_ZERO, off, GPR_A0);
+	}
+}
+
+static inline void build_clear_pref(u32 **buf, int off)
+{
+	if (off & cache_line_mask())
+		return;
+
+	if (pref_bias_clear_store) {
+		_uasm_i_pref(buf, pref_dst_mode, pref_bias_clear_store + off,
+			    GPR_A0);
+	} else if (cache_line_size == (half_clear_loop_size << 1)) {
+		if (cpu_has_cache_cdex_s) {
+			uasm_i_cache(buf, Create_Dirty_Excl_SD, off, GPR_A0);
+		} else if (cpu_has_cache_cdex_p) {
+			if (IS_ENABLED(CONFIG_WAR_R4600_V1_HIT_CACHEOP) &&
+			    cpu_is_r4600_v1_x()) {
+				uasm_i_nop(buf);
+				uasm_i_nop(buf);
+				uasm_i_nop(buf);
+				uasm_i_nop(buf);
+			}
+
+			if (IS_ENABLED(CONFIG_WAR_R4600_V2_HIT_CACHEOP) &&
+			    cpu_is_r4600_v2_x())
+				uasm_i_lw(buf, GPR_ZERO, GPR_ZERO, GPR_AT);
+
+			uasm_i_cache(buf, Create_Dirty_Excl_D, off, GPR_A0);
+		}
+	}
+}
+
+extern u32 __clear_page_start;
+extern u32 __clear_page_end;
+extern u32 __copy_page_start;
+extern u32 __copy_page_end;
+
+void build_clear_page(void)
+{
+	int off;
+	u32 *buf = &__clear_page_start;
+	struct uasm_label *l = labels;
+	struct uasm_reloc *r = relocs;
+	int i;
+	static atomic_t run_once = ATOMIC_INIT(0);
+
+	if (atomic_xchg(&run_once, 1)) {
+		return;
+	}
+
+	memset(labels, 0, sizeof(labels));
+	memset(relocs, 0, sizeof(relocs));
+
+	set_prefetch_parameters();
+
+	/*
+	 * This algorithm makes the following assumptions:
+	 *   - The prefetch bias is a multiple of 2 words.
+	 *   - The prefetch bias is less than one page.
+	 */
+	BUG_ON(pref_bias_clear_store % (2 * clear_word_size));
+	BUG_ON(PAGE_SIZE < pref_bias_clear_store);
+
+	off = PAGE_SIZE - pref_bias_clear_store;
+	if (off > 0xffff || !pref_bias_clear_store)
+		pg_addiu(&buf, GPR_A2, GPR_A0, off);
+	else
+		uasm_i_ori(&buf, GPR_A2, GPR_A0, off);
+
+	if (IS_ENABLED(CONFIG_WAR_R4600_V2_HIT_CACHEOP) && cpu_is_r4600_v2_x())
+		uasm_i_lui(&buf, GPR_AT, uasm_rel_hi(0xa0000000));
+
+	off = cache_line_size ? min(8, pref_bias_clear_store / cache_line_size)
+				* cache_line_size : 0;
+	while (off) {
+		build_clear_pref(&buf, -off);
+		off -= cache_line_size;
+	}
+	uasm_l_clear_pref(&l, buf);
+	do {
+		build_clear_pref(&buf, off);
+		build_clear_store(&buf, off);
+		off += clear_word_size;
+	} while (off < half_clear_loop_size);
+	pg_addiu(&buf, GPR_A0, GPR_A0, 2 * off);
+	off = -off;
+	do {
+		build_clear_pref(&buf, off);
+		if (off == -clear_word_size)
+			uasm_il_bne(&buf, &r, GPR_A0, GPR_A2, label_clear_pref);
+		build_clear_store(&buf, off);
+		off += clear_word_size;
+	} while (off < 0);
+
+	if (pref_bias_clear_store) {
+		pg_addiu(&buf, GPR_A2, GPR_A0, pref_bias_clear_store);
+		uasm_l_clear_nopref(&l, buf);
+		off = 0;
+		do {
+			build_clear_store(&buf, off);
+			off += clear_word_size;
+		} while (off < half_clear_loop_size);
+		pg_addiu(&buf, GPR_A0, GPR_A0, 2 * off);
+		off = -off;
+		do {
+			if (off == -clear_word_size)
+				uasm_il_bne(&buf, &r, GPR_A0, GPR_A2,
+					    label_clear_nopref);
+			build_clear_store(&buf, off);
+			off += clear_word_size;
+		} while (off < 0);
+	}
+
+	uasm_i_jr(&buf, GPR_RA);
+	uasm_i_nop(&buf);
+
+	BUG_ON(buf > &__clear_page_end);
+
+	uasm_resolve_relocs(relocs, labels);
+
+	pr_debug("Synthesized clear page handler (%u instructions).\n",
+		 (u32)(buf - &__clear_page_start));
+
+	pr_debug("\t.set push\n");
+	pr_debug("\t.set noreorder\n");
+	for (i = 0; i < (buf - &__clear_page_start); i++)
+		pr_debug("\t.word 0x%08x\n", (&__clear_page_start)[i]);
+	pr_debug("\t.set pop\n");
+}
+
+static void build_copy_load(u32 **buf, int reg, int off)
+{
+	if (cpu_has_64bit_gp_regs) {
+		uasm_i_ld(buf, reg, off, GPR_A1);
+	} else {
+		uasm_i_lw(buf, reg, off, GPR_A1);
+	}
+}
+
+static void build_copy_store(u32 **buf, int reg, int off)
+{
+	if (cpu_has_64bit_gp_regs) {
+		uasm_i_sd(buf, reg, off, GPR_A0);
+	} else {
+		uasm_i_sw(buf, reg, off, GPR_A0);
+	}
+}
+
+static inline void build_copy_load_pref(u32 **buf, int off)
+{
+	if (off & cache_line_mask())
+		return;
+
+	if (pref_bias_copy_load)
+		_uasm_i_pref(buf, pref_src_mode, pref_bias_copy_load + off, GPR_A1);
+}
+
+static inline void build_copy_store_pref(u32 **buf, int off)
+{
+	if (off & cache_line_mask())
+		return;
+
+	if (pref_bias_copy_store) {
+		_uasm_i_pref(buf, pref_dst_mode, pref_bias_copy_store + off,
+			    GPR_A0);
+	} else if (cache_line_size == (half_copy_loop_size << 1)) {
+		if (cpu_has_cache_cdex_s) {
+			uasm_i_cache(buf, Create_Dirty_Excl_SD, off, GPR_A0);
+		} else if (cpu_has_cache_cdex_p) {
+			if (IS_ENABLED(CONFIG_WAR_R4600_V1_HIT_CACHEOP) &&
+			    cpu_is_r4600_v1_x()) {
+				uasm_i_nop(buf);
+				uasm_i_nop(buf);
+				uasm_i_nop(buf);
+				uasm_i_nop(buf);
+			}
+
+			if (IS_ENABLED(CONFIG_WAR_R4600_V2_HIT_CACHEOP) &&
+			    cpu_is_r4600_v2_x())
+				uasm_i_lw(buf, GPR_ZERO, GPR_ZERO, GPR_AT);
+
+			uasm_i_cache(buf, Create_Dirty_Excl_D, off, GPR_A0);
+		}
+	}
+}
+
+void build_copy_page(void)
+{
+	int off;
+	u32 *buf = &__copy_page_start;
+	struct uasm_label *l = labels;
+	struct uasm_reloc *r = relocs;
+	int i;
+	static atomic_t run_once = ATOMIC_INIT(0);
+
+	if (atomic_xchg(&run_once, 1)) {
+		return;
+	}
+
+	memset(labels, 0, sizeof(labels));
+	memset(relocs, 0, sizeof(relocs));
+
+	set_prefetch_parameters();
+
+	/*
+	 * This algorithm makes the following assumptions:
+	 *   - All prefetch biases are multiples of 8 words.
+	 *   - The prefetch biases are less than one page.
+	 *   - The store prefetch bias isn't greater than the load
+	 *     prefetch bias.
+	 */
+	BUG_ON(pref_bias_copy_load % (8 * copy_word_size));
+	BUG_ON(pref_bias_copy_store % (8 * copy_word_size));
+	BUG_ON(PAGE_SIZE < pref_bias_copy_load);
+	BUG_ON(pref_bias_copy_store > pref_bias_copy_load);
+
+	off = PAGE_SIZE - pref_bias_copy_load;
+	if (off > 0xffff || !pref_bias_copy_load)
+		pg_addiu(&buf, GPR_A2, GPR_A0, off);
+	else
+		uasm_i_ori(&buf, GPR_A2, GPR_A0, off);
+
+	if (IS_ENABLED(CONFIG_WAR_R4600_V2_HIT_CACHEOP) && cpu_is_r4600_v2_x())
+		uasm_i_lui(&buf, GPR_AT, uasm_rel_hi(0xa0000000));
+
+	off = cache_line_size ? min(8, pref_bias_copy_load / cache_line_size) *
+				cache_line_size : 0;
+	while (off) {
+		build_copy_load_pref(&buf, -off);
+		off -= cache_line_size;
+	}
+	off = cache_line_size ? min(8, pref_bias_copy_store / cache_line_size) *
+				cache_line_size : 0;
+	while (off) {
+		build_copy_store_pref(&buf, -off);
+		off -= cache_line_size;
+	}
+	uasm_l_copy_pref_both(&l, buf);
+	do {
+		build_copy_load_pref(&buf, off);
+		build_copy_load(&buf, GPR_T0, off);
+		build_copy_load_pref(&buf, off + copy_word_size);
+		build_copy_load(&buf, GPR_T1, off + copy_word_size);
+		build_copy_load_pref(&buf, off + 2 * copy_word_size);
+		build_copy_load(&buf, GPR_T2, off + 2 * copy_word_size);
+		build_copy_load_pref(&buf, off + 3 * copy_word_size);
+		build_copy_load(&buf, GPR_T3, off + 3 * copy_word_size);
+		build_copy_store_pref(&buf, off);
+		build_copy_store(&buf, GPR_T0, off);
+		build_copy_store_pref(&buf, off + copy_word_size);
+		build_copy_store(&buf, GPR_T1, off + copy_word_size);
+		build_copy_store_pref(&buf, off + 2 * copy_word_size);
+		build_copy_store(&buf, GPR_T2, off + 2 * copy_word_size);
+		build_copy_store_pref(&buf, off + 3 * copy_word_size);
+		build_copy_store(&buf, GPR_T3, off + 3 * copy_word_size);
+		off += 4 * copy_word_size;
+	} while (off < half_copy_loop_size);
+	pg_addiu(&buf, GPR_A1, GPR_A1, 2 * off);
+	pg_addiu(&buf, GPR_A0, GPR_A0, 2 * off);
+	off = -off;
+	do {
+		build_copy_load_pref(&buf, off);
+		build_copy_load(&buf, GPR_T0, off);
+		build_copy_load_pref(&buf, off + copy_word_size);
+		build_copy_load(&buf, GPR_T1, off + copy_word_size);
+		build_copy_load_pref(&buf, off + 2 * copy_word_size);
+		build_copy_load(&buf, GPR_T2, off + 2 * copy_word_size);
+		build_copy_load_pref(&buf, off + 3 * copy_word_size);
+		build_copy_load(&buf, GPR_T3, off + 3 * copy_word_size);
+		build_copy_store_pref(&buf, off);
+		build_copy_store(&buf, GPR_T0, off);
+		build_copy_store_pref(&buf, off + copy_word_size);
+		build_copy_store(&buf, GPR_T1, off + copy_word_size);
+		build_copy_store_pref(&buf, off + 2 * copy_word_size);
+		build_copy_store(&buf, GPR_T2, off + 2 * copy_word_size);
+		build_copy_store_pref(&buf, off + 3 * copy_word_size);
+		if (off == -(4 * copy_word_size))
+			uasm_il_bne(&buf, &r, GPR_A2, GPR_A0, label_copy_pref_both);
+		build_copy_store(&buf, GPR_T3, off + 3 * copy_word_size);
+		off += 4 * copy_word_size;
+	} while (off < 0);
+
+	if (pref_bias_copy_load - pref_bias_copy_store) {
+		pg_addiu(&buf, GPR_A2, GPR_A0,
+			 pref_bias_copy_load - pref_bias_copy_store);
+		uasm_l_copy_pref_store(&l, buf);
+		off = 0;
+		do {
+			build_copy_load(&buf, GPR_T0, off);
+			build_copy_load(&buf, GPR_T1, off + copy_word_size);
+			build_copy_load(&buf, GPR_T2, off + 2 * copy_word_size);
+			build_copy_load(&buf, GPR_T3, off + 3 * copy_word_size);
+			build_copy_store_pref(&buf, off);
+			build_copy_store(&buf, GPR_T0, off);
+			build_copy_store_pref(&buf, off + copy_word_size);
+			build_copy_store(&buf, GPR_T1, off + copy_word_size);
+			build_copy_store_pref(&buf, off + 2 * copy_word_size);
+			build_copy_store(&buf, GPR_T2, off + 2 * copy_word_size);
+			build_copy_store_pref(&buf, off + 3 * copy_word_size);
+			build_copy_store(&buf, GPR_T3, off + 3 * copy_word_size);
+			off += 4 * copy_word_size;
+		} while (off < half_copy_loop_size);
+		pg_addiu(&buf, GPR_A1, GPR_A1, 2 * off);
+		pg_addiu(&buf, GPR_A0, GPR_A0, 2 * off);
+		off = -off;
+		do {
+			build_copy_load(&buf, GPR_T0, off);
+			build_copy_load(&buf, GPR_T1, off + copy_word_size);
+			build_copy_load(&buf, GPR_T2, off + 2 * copy_word_size);
+			build_copy_load(&buf, GPR_T3, off + 3 * copy_word_size);
+			build_copy_store_pref(&buf, off);
+			build_copy_store(&buf, GPR_T0, off);
+			build_copy_store_pref(&buf, off + copy_word_size);
+			build_copy_store(&buf, GPR_T1, off + copy_word_size);
+			build_copy_store_pref(&buf, off + 2 * copy_word_size);
+			build_copy_store(&buf, GPR_T2, off + 2 * copy_word_size);
+			build_copy_store_pref(&buf, off + 3 * copy_word_size);
+			if (off == -(4 * copy_word_size))
+				uasm_il_bne(&buf, &r, GPR_A2, GPR_A0,
+					    label_copy_pref_store);
+			build_copy_store(&buf, GPR_T3, off + 3 * copy_word_size);
+			off += 4 * copy_word_size;
+		} while (off < 0);
+	}
+
+	if (pref_bias_copy_store) {
+		pg_addiu(&buf, GPR_A2, GPR_A0, pref_bias_copy_store);
+		uasm_l_copy_nopref(&l, buf);
+		off = 0;
+		do {
+			build_copy_load(&buf, GPR_T0, off);
+			build_copy_load(&buf, GPR_T1, off + copy_word_size);
+			build_copy_load(&buf, GPR_T2, off + 2 * copy_word_size);
+			build_copy_load(&buf, GPR_T3, off + 3 * copy_word_size);
+			build_copy_store(&buf, GPR_T0, off);
+			build_copy_store(&buf, GPR_T1, off + copy_word_size);
+			build_copy_store(&buf, GPR_T2, off + 2 * copy_word_size);
+			build_copy_store(&buf, GPR_T3, off + 3 * copy_word_size);
+			off += 4 * copy_word_size;
+		} while (off < half_copy_loop_size);
+		pg_addiu(&buf, GPR_A1, GPR_A1, 2 * off);
+		pg_addiu(&buf, GPR_A0, GPR_A0, 2 * off);
+		off = -off;
+		do {
+			build_copy_load(&buf, GPR_T0, off);
+			build_copy_load(&buf, GPR_T1, off + copy_word_size);
+			build_copy_load(&buf, GPR_T2, off + 2 * copy_word_size);
+			build_copy_load(&buf, GPR_T3, off + 3 * copy_word_size);
+			build_copy_store(&buf, GPR_T0, off);
+			build_copy_store(&buf, GPR_T1, off + copy_word_size);
+			build_copy_store(&buf, GPR_T2, off + 2 * copy_word_size);
+			if (off == -(4 * copy_word_size))
+				uasm_il_bne(&buf, &r, GPR_A2, GPR_A0,
+					    label_copy_nopref);
+			build_copy_store(&buf, GPR_T3, off + 3 * copy_word_size);
+			off += 4 * copy_word_size;
+		} while (off < 0);
+	}
+
+	uasm_i_jr(&buf, GPR_RA);
+	uasm_i_nop(&buf);
+
+	BUG_ON(buf > &__copy_page_end);
+
+	uasm_resolve_relocs(relocs, labels);
+
+	pr_debug("Synthesized copy page handler (%u instructions).\n",
+		 (u32)(buf - &__copy_page_start));
+
+	pr_debug("\t.set push\n");
+	pr_debug("\t.set noreorder\n");
+	for (i = 0; i < (buf - &__copy_page_start); i++)
+		pr_debug("\t.word 0x%08x\n", (&__copy_page_start)[i]);
+	pr_debug("\t.set pop\n");
+}
+
+#ifdef CONFIG_SIBYTE_DMA_PAGEOPS
+extern void clear_page_cpu(void *page);
+extern void copy_page_cpu(void *to, void *from);
+
+/*
+ * Pad descriptors to cacheline, since each is exclusively owned by a
+ * particular CPU.
+ */
+struct dmadscr {
+	u64 dscr_a;
+	u64 dscr_b;
+	u64 pad_a;
+	u64 pad_b;
+} ____cacheline_aligned_in_smp page_descr[DM_NUM_CHANNELS];
+
+void clear_page(void *page)
+{
+	u64 to_phys = CPHYSADDR((unsigned long)page);
+	unsigned int cpu = smp_processor_id();
+
+	/* if the page is not in KSEG0, use old way */
+	if ((long)KSEGX((unsigned long)page) != (long)CKSEG0)
+		return clear_page_cpu(page);
+
+	page_descr[cpu].dscr_a = to_phys | M_DM_DSCRA_ZERO_MEM |
+				 M_DM_DSCRA_L2C_DEST | M_DM_DSCRA_INTERRUPT;
+	page_descr[cpu].dscr_b = V_DM_DSCRB_SRC_LENGTH(PAGE_SIZE);
+	__raw_writeq(1, IOADDR(A_DM_REGISTER(cpu, R_DM_DSCR_COUNT)));
+
+	/*
+	 * Don't really want to do it this way, but there's no
+	 * reliable way to delay completion detection.
+	 */
+	while (!(__raw_readq(IOADDR(A_DM_REGISTER(cpu, R_DM_DSCR_BASE_DEBUG)))
+		 & M_DM_DSCR_BASE_INTERRUPT))
+		;
+	__raw_readq(IOADDR(A_DM_REGISTER(cpu, R_DM_DSCR_BASE)));
+}
+EXPORT_SYMBOL(clear_page);
+
+void copy_page(void *to, void *from)
+{
+	u64 from_phys = CPHYSADDR((unsigned long)from);
+	u64 to_phys = CPHYSADDR((unsigned long)to);
+	unsigned int cpu = smp_processor_id();
+
+	/* if any page is not in KSEG0, use old way */
+	if ((long)KSEGX((unsigned long)to) != (long)CKSEG0
+	    || (long)KSEGX((unsigned long)from) != (long)CKSEG0)
+		return copy_page_cpu(to, from);
+
+	page_descr[cpu].dscr_a = to_phys | M_DM_DSCRA_L2C_DEST |
+				 M_DM_DSCRA_INTERRUPT;
+	page_descr[cpu].dscr_b = from_phys | V_DM_DSCRB_SRC_LENGTH(PAGE_SIZE);
+	__raw_writeq(1, IOADDR(A_DM_REGISTER(cpu, R_DM_DSCR_COUNT)));
+
+	/*
+	 * Don't really want to do it this way, but there's no
+	 * reliable way to delay completion detection.
+	 */
+	while (!(__raw_readq(IOADDR(A_DM_REGISTER(cpu, R_DM_DSCR_BASE_DEBUG)))
+		 & M_DM_DSCR_BASE_INTERRUPT))
+		;
+	__raw_readq(IOADDR(A_DM_REGISTER(cpu, R_DM_DSCR_BASE)));
+}
+EXPORT_SYMBOL(copy_page);
+
+#endif /* CONFIG_SIBYTE_DMA_PAGEOPS */