Linux Audio

Check our new training course

Loading...
v3.1
   1/*
   2 * Copyright 2010 Tilera Corporation. All Rights Reserved.
   3 *
   4 *   This program is free software; you can redistribute it and/or
   5 *   modify it under the terms of the GNU General Public License
   6 *   as published by the Free Software Foundation, version 2.
   7 *
   8 *   This program is distributed in the hope that it will be useful, but
   9 *   WITHOUT ANY WARRANTY; without even the implied warranty of
  10 *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
  11 *   NON INFRINGEMENT.  See the GNU General Public License for
  12 *   more details.
  13 *
  14 * Linux interrupt vectors.
  15 */
  16
  17#include <linux/linkage.h>
  18#include <linux/errno.h>
  19#include <linux/init.h>
  20#include <linux/unistd.h>
  21#include <asm/ptrace.h>
  22#include <asm/thread_info.h>
  23#include <asm/irqflags.h>
  24#include <asm/atomic_32.h>
  25#include <asm/asm-offsets.h>
  26#include <hv/hypervisor.h>
  27#include <arch/abi.h>
  28#include <arch/interrupts.h>
  29#include <arch/spr_def.h>
  30
  31#ifdef CONFIG_PREEMPT
  32# error "No support for kernel preemption currently"
  33#endif
  34
  35#define PTREGS_PTR(reg, ptreg) addli reg, sp, C_ABI_SAVE_AREA_SIZE + (ptreg)
  36
  37#define PTREGS_OFFSET_SYSCALL PTREGS_OFFSET_REG(TREG_SYSCALL_NR)
  38
  39#if !CHIP_HAS_WH64()
  40	/* By making this an empty macro, we can use wh64 in the code. */
  41	.macro  wh64 reg
  42	.endm
  43#endif
  44
  45	.macro  push_reg reg, ptr=sp, delta=-4
  46	{
  47	 sw     \ptr, \reg
  48	 addli  \ptr, \ptr, \delta
  49	}
  50	.endm
  51
  52	.macro  pop_reg reg, ptr=sp, delta=4
  53	{
  54	 lw     \reg, \ptr
  55	 addli  \ptr, \ptr, \delta
  56	}
  57	.endm
  58
  59	.macro  pop_reg_zero reg, zreg, ptr=sp, delta=4
  60	{
  61	 move   \zreg, zero
  62	 lw     \reg, \ptr
  63	 addi   \ptr, \ptr, \delta
  64	}
  65	.endm
  66
  67	.macro  push_extra_callee_saves reg
  68	PTREGS_PTR(\reg, PTREGS_OFFSET_REG(51))
  69	push_reg r51, \reg
  70	push_reg r50, \reg
  71	push_reg r49, \reg
  72	push_reg r48, \reg
  73	push_reg r47, \reg
  74	push_reg r46, \reg
  75	push_reg r45, \reg
  76	push_reg r44, \reg
  77	push_reg r43, \reg
  78	push_reg r42, \reg
  79	push_reg r41, \reg
  80	push_reg r40, \reg
  81	push_reg r39, \reg
  82	push_reg r38, \reg
  83	push_reg r37, \reg
  84	push_reg r36, \reg
  85	push_reg r35, \reg
  86	push_reg r34, \reg, PTREGS_OFFSET_BASE - PTREGS_OFFSET_REG(34)
  87	.endm
  88
  89	.macro  panic str
  90	.pushsection .rodata, "a"
  911:
  92	.asciz  "\str"
  93	.popsection
  94	{
  95	 moveli r0, lo16(1b)
  96	}
  97	{
  98	 auli   r0, r0, ha16(1b)
  99	 jal    panic
 100	}
 101	.endm
 102
 103#ifdef __COLLECT_LINKER_FEEDBACK__
 104	.pushsection .text.intvec_feedback,"ax"
 105intvec_feedback:
 106	.popsection
 107#endif
 108
 109	/*
 110	 * Default interrupt handler.
 111	 *
 112	 * vecnum is where we'll put this code.
 113	 * c_routine is the C routine we'll call.
 114	 *
 115	 * The C routine is passed two arguments:
 116	 * - A pointer to the pt_regs state.
 117	 * - The interrupt vector number.
 118	 *
 119	 * The "processing" argument specifies the code for processing
 120	 * the interrupt. Defaults to "handle_interrupt".
 121	 */
 122	.macro  int_hand vecnum, vecname, c_routine, processing=handle_interrupt
 123	.org    (\vecnum << 8)
 124intvec_\vecname:
 125	.ifc    \vecnum, INT_SWINT_1
 126	blz     TREG_SYSCALL_NR_NAME, sys_cmpxchg
 127	.endif
 128
 129	/* Temporarily save a register so we have somewhere to work. */
 130
 131	mtspr   SPR_SYSTEM_SAVE_K_1, r0
 132	mfspr   r0, SPR_EX_CONTEXT_K_1
 133
 134	/* The cmpxchg code clears sp to force us to reset it here on fault. */
 135	{
 136	 bz     sp, 2f
 137	 andi   r0, r0, SPR_EX_CONTEXT_1_1__PL_MASK  /* mask off ICS */
 138	}
 139
 140	.ifc    \vecnum, INT_DOUBLE_FAULT
 141	/*
 142	 * For double-faults from user-space, fall through to the normal
 143	 * register save and stack setup path.  Otherwise, it's the
 144	 * hypervisor giving us one last chance to dump diagnostics, and we
 145	 * branch to the kernel_double_fault routine to do so.
 146	 */
 147	bz      r0, 1f
 148	j       _kernel_double_fault
 1491:
 150	.else
 151	/*
 152	 * If we're coming from user-space, then set sp to the top of
 153	 * the kernel stack.  Otherwise, assume sp is already valid.
 154	 */
 155	{
 156	 bnz    r0, 0f
 157	 move   r0, sp
 158	}
 159	.endif
 160
 161	.ifc    \c_routine, do_page_fault
 162	/*
 163	 * The page_fault handler may be downcalled directly by the
 164	 * hypervisor even when Linux is running and has ICS set.
 165	 *
 166	 * In this case the contents of EX_CONTEXT_K_1 reflect the
 167	 * previous fault and can't be relied on to choose whether or
 168	 * not to reinitialize the stack pointer.  So we add a test
 169	 * to see whether SYSTEM_SAVE_K_2 has the high bit set,
 170	 * and if so we don't reinitialize sp, since we must be coming
 171	 * from Linux.  (In fact the precise case is !(val & ~1),
 172	 * but any Linux PC has to have the high bit set.)
 173	 *
 174	 * Note that the hypervisor *always* sets SYSTEM_SAVE_K_2 for
 175	 * any path that turns into a downcall to one of our TLB handlers.
 176	 */
 177	mfspr   r0, SPR_SYSTEM_SAVE_K_2
 178	{
 179	 blz    r0, 0f    /* high bit in S_S_1_2 is for a PC to use */
 180	 move   r0, sp
 181	}
 182	.endif
 183
 1842:
 185	/*
 186	 * SYSTEM_SAVE_K_0 holds the cpu number in the low bits, and
 187	 * the current stack top in the higher bits.  So we recover
 188	 * our stack top by just masking off the low bits, then
 189	 * point sp at the top aligned address on the actual stack page.
 190	 */
 191	mfspr   r0, SPR_SYSTEM_SAVE_K_0
 192	mm      r0, r0, zero, LOG2_THREAD_SIZE, 31
 193
 1940:
 195	/*
 196	 * Align the stack mod 64 so we can properly predict what
 197	 * cache lines we need to write-hint to reduce memory fetch
 198	 * latency as we enter the kernel.  The layout of memory is
 199	 * as follows, with cache line 0 at the lowest VA, and cache
 200	 * line 4 just below the r0 value this "andi" computes.
 201	 * Note that we never write to cache line 4, and we skip
 202	 * cache line 1 for syscalls.
 203	 *
 204	 *    cache line 4: ptregs padding (two words)
 205	 *    cache line 3: r46...lr, pc, ex1, faultnum, orig_r0, flags, pad
 206	 *    cache line 2: r30...r45
 207	 *    cache line 1: r14...r29
 208	 *    cache line 0: 2 x frame, r0..r13
 209	 */
 
 
 
 210	andi    r0, r0, -64
 211
 212	/*
 213	 * Push the first four registers on the stack, so that we can set
 214	 * them to vector-unique values before we jump to the common code.
 215	 *
 216	 * Registers are pushed on the stack as a struct pt_regs,
 217	 * with the sp initially just above the struct, and when we're
 218	 * done, sp points to the base of the struct, minus
 219	 * C_ABI_SAVE_AREA_SIZE, so we can directly jal to C code.
 220	 *
 221	 * This routine saves just the first four registers, plus the
 222	 * stack context so we can do proper backtracing right away,
 223	 * and defers to handle_interrupt to save the rest.
 224	 * The backtracer needs pc, ex1, lr, sp, r52, and faultnum.
 225	 */
 226	addli   r0, r0, PTREGS_OFFSET_LR - (PTREGS_SIZE + KSTK_PTREGS_GAP)
 227	wh64    r0    /* cache line 3 */
 228	{
 229	 sw     r0, lr
 230	 addli  r0, r0, PTREGS_OFFSET_SP - PTREGS_OFFSET_LR
 231	}
 232	{
 233	 sw     r0, sp
 234	 addli  sp, r0, PTREGS_OFFSET_REG(52) - PTREGS_OFFSET_SP
 235	}
 236	{
 237	 sw     sp, r52
 238	 addli  sp, sp, PTREGS_OFFSET_REG(1) - PTREGS_OFFSET_REG(52)
 239	}
 240	wh64    sp    /* cache line 0 */
 241	{
 242	 sw     sp, r1
 243	 addli  sp, sp, PTREGS_OFFSET_REG(2) - PTREGS_OFFSET_REG(1)
 244	}
 245	{
 246	 sw     sp, r2
 247	 addli  sp, sp, PTREGS_OFFSET_REG(3) - PTREGS_OFFSET_REG(2)
 248	}
 249	{
 250	 sw     sp, r3
 251	 addli  sp, sp, PTREGS_OFFSET_PC - PTREGS_OFFSET_REG(3)
 252	}
 253	mfspr   r0, SPR_EX_CONTEXT_K_0
 254	.ifc \processing,handle_syscall
 255	/*
 256	 * Bump the saved PC by one bundle so that when we return, we won't
 257	 * execute the same swint instruction again.  We need to do this while
 258	 * we're in the critical section.
 259	 */
 260	addi    r0, r0, 8
 261	.endif
 262	{
 263	 sw     sp, r0
 264	 addli  sp, sp, PTREGS_OFFSET_EX1 - PTREGS_OFFSET_PC
 265	}
 266	mfspr   r0, SPR_EX_CONTEXT_K_1
 267	{
 268	 sw     sp, r0
 269	 addi   sp, sp, PTREGS_OFFSET_FAULTNUM - PTREGS_OFFSET_EX1
 270	/*
 271	 * Use r0 for syscalls so it's a temporary; use r1 for interrupts
 272	 * so that it gets passed through unchanged to the handler routine.
 273	 * Note that the .if conditional confusingly spans bundles.
 274	 */
 275	 .ifc \processing,handle_syscall
 276	 movei  r0, \vecnum
 277	}
 278	{
 279	 sw     sp, r0
 280	 .else
 281	 movei  r1, \vecnum
 282	}
 283	{
 284	 sw     sp, r1
 285	 .endif
 286	 addli  sp, sp, PTREGS_OFFSET_REG(0) - PTREGS_OFFSET_FAULTNUM
 287	}
 288	mfspr   r0, SPR_SYSTEM_SAVE_K_1    /* Original r0 */
 289	{
 290	 sw     sp, r0
 291	 addi   sp, sp, -PTREGS_OFFSET_REG(0) - 4
 292	}
 293	{
 294	 sw     sp, zero        /* write zero into "Next SP" frame pointer */
 295	 addi   sp, sp, -4      /* leave SP pointing at bottom of frame */
 296	}
 297	.ifc \processing,handle_syscall
 298	j       handle_syscall
 299	.else
 300	/*
 301	 * Capture per-interrupt SPR context to registers.
 302	 * We overload the meaning of r3 on this path such that if its bit 31
 303	 * is set, we have to mask all interrupts including NMIs before
 304	 * clearing the interrupt critical section bit.
 305	 * See discussion below at "finish_interrupt_save".
 306	 */
 307	.ifc \c_routine, do_page_fault
 308	mfspr   r2, SPR_SYSTEM_SAVE_K_3   /* address of page fault */
 309	mfspr   r3, SPR_SYSTEM_SAVE_K_2   /* info about page fault */
 310	.else
 311	.ifc \vecnum, INT_DOUBLE_FAULT
 312	{
 313	 mfspr  r2, SPR_SYSTEM_SAVE_K_2   /* double fault info from HV */
 314	 movei  r3, 0
 315	}
 316	.else
 317	.ifc \c_routine, do_trap
 318	{
 319	 mfspr  r2, GPV_REASON
 320	 movei  r3, 0
 321	}
 322	.else
 323	.ifc \c_routine, op_handle_perf_interrupt
 324	{
 325	 mfspr  r2, PERF_COUNT_STS
 326	 movei  r3, -1   /* not used, but set for consistency */
 327	}
 328	.else
 329#if CHIP_HAS_AUX_PERF_COUNTERS()
 330	.ifc \c_routine, op_handle_aux_perf_interrupt
 331	{
 332	 mfspr  r2, AUX_PERF_COUNT_STS
 333	 movei  r3, -1   /* not used, but set for consistency */
 334	}
 335	.else
 336#endif
 337	movei   r3, 0
 338#if CHIP_HAS_AUX_PERF_COUNTERS()
 339	.endif
 340#endif
 341	.endif
 342	.endif
 343	.endif
 344	.endif
 345	/* Put function pointer in r0 */
 346	moveli  r0, lo16(\c_routine)
 347	{
 348	 auli   r0, r0, ha16(\c_routine)
 349	 j       \processing
 350	}
 351	.endif
 352	ENDPROC(intvec_\vecname)
 353
 354#ifdef __COLLECT_LINKER_FEEDBACK__
 355	.pushsection .text.intvec_feedback,"ax"
 356	.org    (\vecnum << 5)
 357	FEEDBACK_ENTER_EXPLICIT(intvec_\vecname, .intrpt1, 1 << 8)
 358	jrp     lr
 359	.popsection
 360#endif
 361
 362	.endm
 363
 364
 365	/*
 366	 * Save the rest of the registers that we didn't save in the actual
 367	 * vector itself.  We can't use r0-r10 inclusive here.
 368	 */
 369	.macro  finish_interrupt_save, function
 370
 371	/* If it's a syscall, save a proper orig_r0, otherwise just zero. */
 372	PTREGS_PTR(r52, PTREGS_OFFSET_ORIG_R0)
 373	{
 374	 .ifc \function,handle_syscall
 375	 sw     r52, r0
 376	 .else
 377	 sw     r52, zero
 378	 .endif
 379	 PTREGS_PTR(r52, PTREGS_OFFSET_TP)
 380	}
 381
 382	/*
 383	 * For ordinary syscalls, we save neither caller- nor callee-
 384	 * save registers, since the syscall invoker doesn't expect the
 385	 * caller-saves to be saved, and the called kernel functions will
 386	 * take care of saving the callee-saves for us.
 387	 *
 388	 * For interrupts we save just the caller-save registers.  Saving
 389	 * them is required (since the "caller" can't save them).  Again,
 390	 * the called kernel functions will restore the callee-save
 391	 * registers for us appropriately.
 392	 *
 393	 * On return, we normally restore nothing special for syscalls,
 394	 * and just the caller-save registers for interrupts.
 395	 *
 396	 * However, there are some important caveats to all this:
 397	 *
 398	 * - We always save a few callee-save registers to give us
 399	 *   some scratchpad registers to carry across function calls.
 400	 *
 401	 * - fork/vfork/etc require us to save all the callee-save
 402	 *   registers, which we do in PTREGS_SYSCALL_ALL_REGS, below.
 403	 *
 404	 * - We always save r0..r5 and r10 for syscalls, since we need
 405	 *   to reload them a bit later for the actual kernel call, and
 406	 *   since we might need them for -ERESTARTNOINTR, etc.
 407	 *
 408	 * - Before invoking a signal handler, we save the unsaved
 409	 *   callee-save registers so they are visible to the
 410	 *   signal handler or any ptracer.
 411	 *
 412	 * - If the unsaved callee-save registers are modified, we set
 413	 *   a bit in pt_regs so we know to reload them from pt_regs
 414	 *   and not just rely on the kernel function unwinding.
 415	 *   (Done for ptrace register writes and SA_SIGINFO handler.)
 416	 */
 417	{
 418	 sw     r52, tp
 419	 PTREGS_PTR(r52, PTREGS_OFFSET_REG(33))
 420	}
 421	wh64    r52    /* cache line 2 */
 422	push_reg r33, r52
 423	push_reg r32, r52
 424	push_reg r31, r52
 425	.ifc \function,handle_syscall
 426	push_reg r30, r52, PTREGS_OFFSET_SYSCALL - PTREGS_OFFSET_REG(30)
 427	push_reg TREG_SYSCALL_NR_NAME, r52, \
 428	  PTREGS_OFFSET_REG(5) - PTREGS_OFFSET_SYSCALL
 429	.else
 430
 431	push_reg r30, r52, PTREGS_OFFSET_REG(29) - PTREGS_OFFSET_REG(30)
 432	wh64    r52    /* cache line 1 */
 433	push_reg r29, r52
 434	push_reg r28, r52
 435	push_reg r27, r52
 436	push_reg r26, r52
 437	push_reg r25, r52
 438	push_reg r24, r52
 439	push_reg r23, r52
 440	push_reg r22, r52
 441	push_reg r21, r52
 442	push_reg r20, r52
 443	push_reg r19, r52
 444	push_reg r18, r52
 445	push_reg r17, r52
 446	push_reg r16, r52
 447	push_reg r15, r52
 448	push_reg r14, r52
 449	push_reg r13, r52
 450	push_reg r12, r52
 451	push_reg r11, r52
 452	push_reg r10, r52
 453	push_reg r9, r52
 454	push_reg r8, r52
 455	push_reg r7, r52
 456	push_reg r6, r52
 457
 458	.endif
 459
 460	push_reg r5, r52
 461	sw      r52, r4
 462
 463	/* Load tp with our per-cpu offset. */
 464#ifdef CONFIG_SMP
 465	{
 466	 mfspr  r20, SPR_SYSTEM_SAVE_K_0
 467	 moveli r21, lo16(__per_cpu_offset)
 468	}
 469	{
 470	 auli   r21, r21, ha16(__per_cpu_offset)
 471	 mm     r20, r20, zero, 0, LOG2_THREAD_SIZE-1
 472	}
 473	s2a     r20, r20, r21
 474	lw      tp, r20
 475#else
 476	move    tp, zero
 477#endif
 478
 479	/*
 480	 * If we will be returning to the kernel, we will need to
 481	 * reset the interrupt masks to the state they had before.
 482	 * Set DISABLE_IRQ in flags iff we came from PL1 with irqs disabled.
 483	 * We load flags in r32 here so we can jump to .Lrestore_regs
 484	 * directly after do_page_fault_ics() if necessary.
 485	 */
 486	mfspr   r32, SPR_EX_CONTEXT_K_1
 487	{
 488	 andi   r32, r32, SPR_EX_CONTEXT_1_1__PL_MASK  /* mask off ICS */
 489	 PTREGS_PTR(r21, PTREGS_OFFSET_FLAGS)
 490	}
 491	bzt     r32, 1f       /* zero if from user space */
 492	IRQS_DISABLED(r32)    /* zero if irqs enabled */
 493#if PT_FLAGS_DISABLE_IRQ != 1
 494# error Value of IRQS_DISABLED used to set PT_FLAGS_DISABLE_IRQ; fix
 495#endif
 4961:
 497	.ifnc \function,handle_syscall
 498	/* Record the fact that we saved the caller-save registers above. */
 499	ori     r32, r32, PT_FLAGS_CALLER_SAVES
 500	.endif
 501	sw      r21, r32
 502
 503#ifdef __COLLECT_LINKER_FEEDBACK__
 504	/*
 505	 * Notify the feedback routines that we were in the
 506	 * appropriate fixed interrupt vector area.  Note that we
 507	 * still have ICS set at this point, so we can't invoke any
 508	 * atomic operations or we will panic.  The feedback
 509	 * routines internally preserve r0..r10 and r30 up.
 510	 */
 511	.ifnc \function,handle_syscall
 512	shli    r20, r1, 5
 513	.else
 514	moveli  r20, INT_SWINT_1 << 5
 515	.endif
 516	addli   r20, r20, lo16(intvec_feedback)
 517	auli    r20, r20, ha16(intvec_feedback)
 518	jalr    r20
 519
 520	/* And now notify the feedback routines that we are here. */
 521	FEEDBACK_ENTER(\function)
 522#endif
 523
 524	/*
 525	 * we've captured enough state to the stack (including in
 526	 * particular our EX_CONTEXT state) that we can now release
 527	 * the interrupt critical section and replace it with our
 528	 * standard "interrupts disabled" mask value.  This allows
 529	 * synchronous interrupts (and profile interrupts) to punch
 530	 * through from this point onwards.
 531	 *
 532	 * If bit 31 of r3 is set during a non-NMI interrupt, we know we
 533	 * are on the path where the hypervisor has punched through our
 534	 * ICS with a page fault, so we call out to do_page_fault_ics()
 535	 * to figure out what to do with it.  If the fault was in
 536	 * an atomic op, we unlock the atomic lock, adjust the
 537	 * saved register state a little, and return "zero" in r4,
 538	 * falling through into the normal page-fault interrupt code.
 539	 * If the fault was in a kernel-space atomic operation, then
 540	 * do_page_fault_ics() resolves it itself, returns "one" in r4,
 541	 * and as a result goes directly to restoring registers and iret,
 542	 * without trying to adjust the interrupt masks at all.
 543	 * The do_page_fault_ics() API involves passing and returning
 544	 * a five-word struct (in registers) to avoid writing the
 545	 * save and restore code here.
 546	 */
 547	.ifc \function,handle_nmi
 548	IRQ_DISABLE_ALL(r20)
 549	.else
 550	.ifnc \function,handle_syscall
 551	bgezt   r3, 1f
 552	{
 553	 PTREGS_PTR(r0, PTREGS_OFFSET_BASE)
 554	 jal    do_page_fault_ics
 555	}
 556	FEEDBACK_REENTER(\function)
 557	bzt     r4, 1f
 558	j       .Lrestore_regs
 5591:
 560	.endif
 561	IRQ_DISABLE(r20, r21)
 562	.endif
 563	mtspr   INTERRUPT_CRITICAL_SECTION, zero
 564
 565#if CHIP_HAS_WH64()
 566	/*
 567	 * Prepare the first 256 stack bytes to be rapidly accessible
 568	 * without having to fetch the background data.  We don't really
 569	 * know how far to write-hint, but kernel stacks generally
 570	 * aren't that big, and write-hinting here does take some time.
 571	 */
 572	addi    r52, sp, -64
 573	{
 574	 wh64   r52
 575	 addi   r52, r52, -64
 576	}
 577	{
 578	 wh64   r52
 579	 addi   r52, r52, -64
 580	}
 581	{
 582	 wh64   r52
 583	 addi   r52, r52, -64
 584	}
 585	wh64    r52
 586#endif
 587
 588#ifdef CONFIG_TRACE_IRQFLAGS
 589	.ifnc \function,handle_nmi
 590	/*
 591	 * We finally have enough state set up to notify the irq
 592	 * tracing code that irqs were disabled on entry to the handler.
 593	 * The TRACE_IRQS_OFF call clobbers registers r0-r29.
 594	 * For syscalls, we already have the register state saved away
 595	 * on the stack, so we don't bother to do any register saves here,
 596	 * and later we pop the registers back off the kernel stack.
 597	 * For interrupt handlers, save r0-r3 in callee-saved registers.
 598	 */
 599	.ifnc \function,handle_syscall
 600	{ move r30, r0; move r31, r1 }
 601	{ move r32, r2; move r33, r3 }
 602	.endif
 603	TRACE_IRQS_OFF
 604	.ifnc \function,handle_syscall
 605	{ move r0, r30; move r1, r31 }
 606	{ move r2, r32; move r3, r33 }
 607	.endif
 608	.endif
 609#endif
 610
 611	.endm
 612
 613	.macro  check_single_stepping, kind, not_single_stepping
 614	/*
 615	 * Check for single stepping in user-level priv
 616	 *   kind can be "normal", "ill", or "syscall"
 617	 * At end, if fall-thru
 618	 *   r29: thread_info->step_state
 619	 *   r28: &pt_regs->pc
 620	 *   r27: pt_regs->pc
 621	 *   r26: thread_info->step_state->buffer
 622	 */
 623
 624	/* Check for single stepping */
 625	GET_THREAD_INFO(r29)
 626	{
 627	 /* Get pointer to field holding step state */
 628	 addi   r29, r29, THREAD_INFO_STEP_STATE_OFFSET
 629
 630	 /* Get pointer to EX1 in register state */
 631	 PTREGS_PTR(r27, PTREGS_OFFSET_EX1)
 632	}
 633	{
 634	 /* Get pointer to field holding PC */
 635	 PTREGS_PTR(r28, PTREGS_OFFSET_PC)
 636
 637	 /* Load the pointer to the step state */
 638	 lw     r29, r29
 639	}
 640	/* Load EX1 */
 641	lw      r27, r27
 642	{
 643	 /* Points to flags */
 644	 addi   r23, r29, SINGLESTEP_STATE_FLAGS_OFFSET
 645
 646	 /* No single stepping if there is no step state structure */
 647	 bzt    r29, \not_single_stepping
 648	}
 649	{
 650	 /* mask off ICS and any other high bits */
 651	 andi   r27, r27, SPR_EX_CONTEXT_1_1__PL_MASK
 652
 653	 /* Load pointer to single step instruction buffer */
 654	 lw     r26, r29
 655	}
 656	/* Check priv state */
 657	bnz     r27, \not_single_stepping
 658
 659	/* Get flags */
 660	lw      r22, r23
 661	{
 662	 /* Branch if single-step mode not enabled */
 663	 bbnst  r22, \not_single_stepping
 664
 665	 /* Clear enabled flag */
 666	 andi   r22, r22, ~SINGLESTEP_STATE_MASK_IS_ENABLED
 667	}
 668	.ifc \kind,normal
 669	{
 670	 /* Load PC */
 671	 lw     r27, r28
 672
 673	 /* Point to the entry containing the original PC */
 674	 addi   r24, r29, SINGLESTEP_STATE_ORIG_PC_OFFSET
 675	}
 676	{
 677	 /* Disable single stepping flag */
 678	 sw     r23, r22
 679	}
 680	{
 681	 /* Get the original pc */
 682	 lw     r24, r24
 683
 684	 /* See if the PC is at the start of the single step buffer */
 685	 seq    r25, r26, r27
 686	}
 687	/*
 688	 * NOTE: it is really expected that the PC be in the single step buffer
 689	 *       at this point
 690	 */
 691	bzt     r25, \not_single_stepping
 692
 693	/* Restore the original PC */
 694	sw      r28, r24
 695	.else
 696	.ifc \kind,syscall
 697	{
 698	 /* Load PC */
 699	 lw     r27, r28
 700
 701	 /* Point to the entry containing the next PC */
 702	 addi   r24, r29, SINGLESTEP_STATE_NEXT_PC_OFFSET
 703	}
 704	{
 705	 /* Increment the stopped PC by the bundle size */
 706	 addi   r26, r26, 8
 707
 708	 /* Disable single stepping flag */
 709	 sw     r23, r22
 710	}
 711	{
 712	 /* Get the next pc */
 713	 lw     r24, r24
 714
 715	 /*
 716	  * See if the PC is one bundle past the start of the
 717	  * single step buffer
 718	  */
 719	 seq    r25, r26, r27
 720	}
 721	{
 722	 /*
 723	  * NOTE: it is really expected that the PC be in the
 724	  * single step buffer at this point
 725	  */
 726	 bzt    r25, \not_single_stepping
 727	}
 728	/* Set to the next PC */
 729	sw      r28, r24
 730	.else
 731	{
 732	 /* Point to 3rd bundle in buffer */
 733	 addi   r25, r26, 16
 734
 735	 /* Load PC */
 736	 lw      r27, r28
 737	}
 738	{
 739	 /* Disable single stepping flag */
 740	 sw      r23, r22
 741
 742	 /* See if the PC is in the single step buffer */
 743	 slte_u  r24, r26, r27
 744	}
 745	{
 746	 slte_u r25, r27, r25
 747
 748	 /*
 749	  * NOTE: it is really expected that the PC be in the
 750	  * single step buffer at this point
 751	  */
 752	 bzt    r24, \not_single_stepping
 753	}
 754	bzt     r25, \not_single_stepping
 755	.endif
 756	.endif
 757	.endm
 758
 759	/*
 760	 * Redispatch a downcall.
 761	 */
 762	.macro  dc_dispatch vecnum, vecname
 763	.org    (\vecnum << 8)
 764intvec_\vecname:
 765	j       hv_downcall_dispatch
 766	ENDPROC(intvec_\vecname)
 767	.endm
 768
 769	/*
 770	 * Common code for most interrupts.  The C function we're eventually
 771	 * going to is in r0, and the faultnum is in r1; the original
 772	 * values for those registers are on the stack.
 773	 */
 774	.pushsection .text.handle_interrupt,"ax"
 775handle_interrupt:
 776	finish_interrupt_save handle_interrupt
 777
 778	/*
 779	 * Check for if we are single stepping in user level. If so, then
 780	 * we need to restore the PC.
 781	 */
 782
 783	check_single_stepping normal, .Ldispatch_interrupt
 784.Ldispatch_interrupt:
 785
 786	/* Jump to the C routine; it should enable irqs as soon as possible. */
 787	{
 788	 jalr   r0
 789	 PTREGS_PTR(r0, PTREGS_OFFSET_BASE)
 790	}
 791	FEEDBACK_REENTER(handle_interrupt)
 792	{
 793	 movei  r30, 0   /* not an NMI */
 794	 j      interrupt_return
 795	}
 796	STD_ENDPROC(handle_interrupt)
 797
 798/*
 799 * This routine takes a boolean in r30 indicating if this is an NMI.
 800 * If so, we also expect a boolean in r31 indicating whether to
 801 * re-enable the oprofile interrupts.
 
 
 
 
 802 */
 803STD_ENTRY(interrupt_return)
 804	/* If we're resuming to kernel space, don't check thread flags. */
 805	{
 806	 bnz    r30, .Lrestore_all  /* NMIs don't special-case user-space */
 807	 PTREGS_PTR(r29, PTREGS_OFFSET_EX1)
 808	}
 809	lw      r29, r29
 810	andi    r29, r29, SPR_EX_CONTEXT_1_1__PL_MASK  /* mask off ICS */
 
 
 
 
 
 
 811	{
 812	 bzt    r29, .Lresume_userspace
 813	 PTREGS_PTR(r29, PTREGS_OFFSET_PC)
 814	}
 
 
 
 
 
 
 
 
 
 
 
 
 
 815
 816	/* If we're resuming to _cpu_idle_nap, bump PC forward by 8. */
 817	{
 818	 lw     r28, r29
 819	 moveli r27, lo16(_cpu_idle_nap)
 820	}
 821	{
 
 822	 auli   r27, r27, ha16(_cpu_idle_nap)
 823	}
 824	{
 825	 seq    r27, r27, r28
 826	}
 827	{
 828	 bbns   r27, .Lrestore_all
 829	 addi   r28, r28, 8
 830	}
 831	sw      r29, r28
 832	j       .Lrestore_all
 833
 834.Lresume_userspace:
 835	FEEDBACK_REENTER(interrupt_return)
 836
 837	/*
 
 
 
 
 
 
 
 
 
 
 
 
 838	 * Disable interrupts so as to make sure we don't
 839	 * miss an interrupt that sets any of the thread flags (like
 840	 * need_resched or sigpending) between sampling and the iret.
 841	 * Routines like schedule() or do_signal() may re-enable
 842	 * interrupts before returning.
 843	 */
 844	IRQ_DISABLE(r20, r21)
 845	TRACE_IRQS_OFF  /* Note: clobbers registers r0-r29 */
 846
 847	/* Get base of stack in r32; note r30/31 are used as arguments here. */
 848	GET_THREAD_INFO(r32)
 849
 850
 851	/* Check to see if there is any work to do before returning to user. */
 852	{
 853	 addi   r29, r32, THREAD_INFO_FLAGS_OFFSET
 854	 moveli r1, lo16(_TIF_ALLWORK_MASK)
 855	}
 856	{
 857	 lw     r29, r29
 858	 auli   r1, r1, ha16(_TIF_ALLWORK_MASK)
 859	}
 860	and     r1, r29, r1
 861	bzt     r1, .Lrestore_all
 862
 863	/*
 864	 * Make sure we have all the registers saved for signal
 865	 * handling or single-step.  Call out to C code to figure out
 866	 * exactly what we need to do for each flag bit, then if
 867	 * necessary, reload the flags and recheck.
 868	 */
 869	push_extra_callee_saves r0
 870	{
 871	 PTREGS_PTR(r0, PTREGS_OFFSET_BASE)
 872	 jal    do_work_pending
 873	}
 874	bnz     r0, .Lresume_userspace
 
 
 
 875
 876	/*
 877	 * In the NMI case we
 878	 * omit the call to single_process_check_nohz, which normally checks
 879	 * to see if we should start or stop the scheduler tick, because
 880	 * we can't call arbitrary Linux code from an NMI context.
 881	 * We always call the homecache TLB deferral code to re-trigger
 882	 * the deferral mechanism.
 883	 *
 884	 * The other chunk of responsibility this code has is to reset the
 885	 * interrupt masks appropriately to reset irqs and NMIs.  We have
 886	 * to call TRACE_IRQS_OFF and TRACE_IRQS_ON to support all the
 887	 * lockdep-type stuff, but we can't set ICS until afterwards, since
 888	 * ICS can only be used in very tight chunks of code to avoid
 889	 * tripping over various assertions that it is off.
 890	 *
 891	 * (There is what looks like a window of vulnerability here since
 892	 * we might take a profile interrupt between the two SPR writes
 893	 * that set the mask, but since we write the low SPR word first,
 894	 * and our interrupt entry code checks the low SPR word, any
 895	 * profile interrupt will actually disable interrupts in both SPRs
 896	 * before returning, which is OK.)
 897	 */
 898.Lrestore_all:
 899	PTREGS_PTR(r0, PTREGS_OFFSET_EX1)
 900	{
 901	 lw     r0, r0
 902	 PTREGS_PTR(r32, PTREGS_OFFSET_FLAGS)
 903	}
 904	{
 905	 andi   r0, r0, SPR_EX_CONTEXT_1_1__PL_MASK
 906	 lw     r32, r32
 907	}
 908	bnz    r0, 1f
 909	j       2f
 910#if PT_FLAGS_DISABLE_IRQ != 1
 911# error Assuming PT_FLAGS_DISABLE_IRQ == 1 so we can use bbnst below
 912#endif
 9131:	bbnst   r32, 2f
 914	IRQ_DISABLE(r20,r21)
 915	TRACE_IRQS_OFF
 916	movei   r0, 1
 917	mtspr   INTERRUPT_CRITICAL_SECTION, r0
 918	bzt     r30, .Lrestore_regs
 919	j       3f
 9202:	TRACE_IRQS_ON
 921	movei   r0, 1
 922	mtspr   INTERRUPT_CRITICAL_SECTION, r0
 923	IRQ_ENABLE(r20, r21)
 924	bzt     r30, .Lrestore_regs
 9253:
 926
 
 
 
 
 
 
 
 927
 928	/*
 929	 * We now commit to returning from this interrupt, since we will be
 930	 * doing things like setting EX_CONTEXT SPRs and unwinding the stack
 931	 * frame.  No calls should be made to any other code after this point.
 932	 * This code should only be entered with ICS set.
 933	 * r32 must still be set to ptregs.flags.
 934	 * We launch loads to each cache line separately first, so we can
 935	 * get some parallelism out of the memory subsystem.
 936	 * We start zeroing caller-saved registers throughout, since
 937	 * that will save some cycles if this turns out to be a syscall.
 938	 */
 939.Lrestore_regs:
 940	FEEDBACK_REENTER(interrupt_return)   /* called from elsewhere */
 941
 942	/*
 943	 * Rotate so we have one high bit and one low bit to test.
 944	 * - low bit says whether to restore all the callee-saved registers,
 945	 *   or just r30-r33, and r52 up.
 946	 * - high bit (i.e. sign bit) says whether to restore all the
 947	 *   caller-saved registers, or just r0.
 948	 */
 949#if PT_FLAGS_CALLER_SAVES != 2 || PT_FLAGS_RESTORE_REGS != 4
 950# error Rotate trick does not work :-)
 951#endif
 952	{
 953	 rli    r20, r32, 30
 954	 PTREGS_PTR(sp, PTREGS_OFFSET_REG(0))
 955	}
 956
 957	/*
 958	 * Load cache lines 0, 2, and 3 in that order, then use
 959	 * the last loaded value, which makes it likely that the other
 960	 * cache lines have also loaded, at which point we should be
 961	 * able to safely read all the remaining words on those cache
 962	 * lines without waiting for the memory subsystem.
 963	 */
 964	pop_reg_zero r0, r28, sp, PTREGS_OFFSET_REG(30) - PTREGS_OFFSET_REG(0)
 965	pop_reg_zero r30, r2, sp, PTREGS_OFFSET_PC - PTREGS_OFFSET_REG(30)
 966	pop_reg_zero r21, r3, sp, PTREGS_OFFSET_EX1 - PTREGS_OFFSET_PC
 967	pop_reg_zero lr, r4, sp, PTREGS_OFFSET_REG(52) - PTREGS_OFFSET_EX1
 968	{
 969	 mtspr  SPR_EX_CONTEXT_K_0, r21
 970	 move   r5, zero
 971	}
 972	{
 973	 mtspr  SPR_EX_CONTEXT_K_1, lr
 974	 andi   lr, lr, SPR_EX_CONTEXT_1_1__PL_MASK  /* mask off ICS */
 975	}
 976
 977	/* Restore callee-saveds that we actually use. */
 978	pop_reg_zero r52, r6, sp, PTREGS_OFFSET_REG(31) - PTREGS_OFFSET_REG(52)
 979	pop_reg_zero r31, r7
 980	pop_reg_zero r32, r8
 981	pop_reg_zero r33, r9, sp, PTREGS_OFFSET_REG(29) - PTREGS_OFFSET_REG(33)
 982
 983	/*
 984	 * If we modified other callee-saveds, restore them now.
 985	 * This is rare, but could be via ptrace or signal handler.
 986	 */
 987	{
 988	 move   r10, zero
 989	 bbs    r20, .Lrestore_callees
 990	}
 991.Lcontinue_restore_regs:
 992
 993	/* Check if we're returning from a syscall. */
 994	{
 995	 move   r11, zero
 996	 blzt   r20, 1f  /* no, so go restore callee-save registers */
 997	}
 998
 999	/*
1000	 * Check if we're returning to userspace.
1001	 * Note that if we're not, we don't worry about zeroing everything.
1002	 */
1003	{
1004	 addli  sp, sp, PTREGS_OFFSET_LR - PTREGS_OFFSET_REG(29)
1005	 bnz    lr, .Lkernel_return
1006	}
1007
1008	/*
1009	 * On return from syscall, we've restored r0 from pt_regs, but we
1010	 * clear the remainder of the caller-saved registers.  We could
1011	 * restore the syscall arguments, but there's not much point,
1012	 * and it ensures user programs aren't trying to use the
1013	 * caller-saves if we clear them, as well as avoiding leaking
1014	 * kernel pointers into userspace.
1015	 */
1016	pop_reg_zero lr, r12, sp, PTREGS_OFFSET_TP - PTREGS_OFFSET_LR
1017	pop_reg_zero tp, r13, sp, PTREGS_OFFSET_SP - PTREGS_OFFSET_TP
1018	{
1019	 lw     sp, sp
1020	 move   r14, zero
1021	 move   r15, zero
1022	}
1023	{ move r16, zero; move r17, zero }
1024	{ move r18, zero; move r19, zero }
1025	{ move r20, zero; move r21, zero }
1026	{ move r22, zero; move r23, zero }
1027	{ move r24, zero; move r25, zero }
1028	{ move r26, zero; move r27, zero }
1029
1030	/* Set r1 to errno if we are returning an error, otherwise zero. */
1031	{
1032	 moveli r29, 4096
1033	 sub    r1, zero, r0
1034	}
1035	slt_u   r29, r1, r29
1036	{
1037	 mnz    r1, r29, r1
1038	 move   r29, zero
1039	}
1040	iret
1041
1042	/*
1043	 * Not a syscall, so restore caller-saved registers.
1044	 * First kick off a load for cache line 1, which we're touching
1045	 * for the first time here.
1046	 */
1047	.align 64
10481:	pop_reg r29, sp, PTREGS_OFFSET_REG(1) - PTREGS_OFFSET_REG(29)
1049	pop_reg r1
1050	pop_reg r2
1051	pop_reg r3
1052	pop_reg r4
1053	pop_reg r5
1054	pop_reg r6
1055	pop_reg r7
1056	pop_reg r8
1057	pop_reg r9
1058	pop_reg r10
1059	pop_reg r11
1060	pop_reg r12
1061	pop_reg r13
1062	pop_reg r14
1063	pop_reg r15
1064	pop_reg r16
1065	pop_reg r17
1066	pop_reg r18
1067	pop_reg r19
1068	pop_reg r20
1069	pop_reg r21
1070	pop_reg r22
1071	pop_reg r23
1072	pop_reg r24
1073	pop_reg r25
1074	pop_reg r26
1075	pop_reg r27
1076	pop_reg r28, sp, PTREGS_OFFSET_LR - PTREGS_OFFSET_REG(28)
1077	/* r29 already restored above */
1078	bnz     lr, .Lkernel_return
1079	pop_reg lr, sp, PTREGS_OFFSET_TP - PTREGS_OFFSET_LR
1080	pop_reg tp, sp, PTREGS_OFFSET_SP - PTREGS_OFFSET_TP
1081	lw      sp, sp
1082	iret
1083
1084	/*
1085	 * We can't restore tp when in kernel mode, since a thread might
1086	 * have migrated from another cpu and brought a stale tp value.
1087	 */
1088.Lkernel_return:
1089	pop_reg lr, sp, PTREGS_OFFSET_SP - PTREGS_OFFSET_LR
1090	lw      sp, sp
1091	iret
1092
1093	/* Restore callee-saved registers from r34 to r51. */
1094.Lrestore_callees:
1095	addli  sp, sp, PTREGS_OFFSET_REG(34) - PTREGS_OFFSET_REG(29)
1096	pop_reg r34
1097	pop_reg r35
1098	pop_reg r36
1099	pop_reg r37
1100	pop_reg r38
1101	pop_reg r39
1102	pop_reg r40
1103	pop_reg r41
1104	pop_reg r42
1105	pop_reg r43
1106	pop_reg r44
1107	pop_reg r45
1108	pop_reg r46
1109	pop_reg r47
1110	pop_reg r48
1111	pop_reg r49
1112	pop_reg r50
1113	pop_reg r51, sp, PTREGS_OFFSET_REG(29) - PTREGS_OFFSET_REG(51)
1114	j .Lcontinue_restore_regs
1115	STD_ENDPROC(interrupt_return)
1116
1117	/*
1118	 * Some interrupts don't check for single stepping
1119	 */
1120	.pushsection .text.handle_interrupt_no_single_step,"ax"
1121handle_interrupt_no_single_step:
1122	finish_interrupt_save handle_interrupt_no_single_step
1123	{
1124	 jalr   r0
1125	 PTREGS_PTR(r0, PTREGS_OFFSET_BASE)
1126	}
1127	FEEDBACK_REENTER(handle_interrupt_no_single_step)
1128	{
1129	 movei  r30, 0   /* not an NMI */
1130	 j      interrupt_return
1131	}
1132	STD_ENDPROC(handle_interrupt_no_single_step)
1133
1134	/*
1135	 * "NMI" interrupts mask ALL interrupts before calling the
1136	 * handler, and don't check thread flags, etc., on the way
1137	 * back out.  In general, the only things we do here for NMIs
1138	 * are the register save/restore, fixing the PC if we were
1139	 * doing single step, and the dataplane kernel-TLB management.
1140	 * We don't (for example) deal with start/stop of the sched tick.
1141	 */
1142	.pushsection .text.handle_nmi,"ax"
1143handle_nmi:
1144	finish_interrupt_save handle_nmi
1145	check_single_stepping normal, .Ldispatch_nmi
1146.Ldispatch_nmi:
1147	{
1148	 jalr   r0
1149	 PTREGS_PTR(r0, PTREGS_OFFSET_BASE)
1150	}
1151	FEEDBACK_REENTER(handle_nmi)
 
 
 
 
1152	j       interrupt_return
1153	STD_ENDPROC(handle_nmi)
1154
1155	/*
1156	 * Parallel code for syscalls to handle_interrupt.
1157	 */
1158	.pushsection .text.handle_syscall,"ax"
1159handle_syscall:
1160	finish_interrupt_save handle_syscall
1161
1162	/*
1163	 * Check for if we are single stepping in user level. If so, then
1164	 * we need to restore the PC.
1165	 */
1166	check_single_stepping syscall, .Ldispatch_syscall
1167.Ldispatch_syscall:
1168
1169	/* Enable irqs. */
1170	TRACE_IRQS_ON
1171	IRQ_ENABLE(r20, r21)
1172
1173	/* Bump the counter for syscalls made on this tile. */
1174	moveli  r20, lo16(irq_stat + IRQ_CPUSTAT_SYSCALL_COUNT_OFFSET)
1175	auli    r20, r20, ha16(irq_stat + IRQ_CPUSTAT_SYSCALL_COUNT_OFFSET)
1176	add     r20, r20, tp
1177	lw      r21, r20
1178	addi    r21, r21, 1
1179	sw      r20, r21
 
 
 
1180
1181	/* Trace syscalls, if requested. */
1182	GET_THREAD_INFO(r31)
1183	addi	r31, r31, THREAD_INFO_FLAGS_OFFSET
1184	lw	r30, r31
1185	andi    r30, r30, _TIF_SYSCALL_TRACE
1186	bzt	r30, .Lrestore_syscall_regs
1187	jal	do_syscall_trace
 
 
 
1188	FEEDBACK_REENTER(handle_syscall)
1189
1190	/*
1191	 * We always reload our registers from the stack at this
1192	 * point.  They might be valid, if we didn't build with
1193	 * TRACE_IRQFLAGS, and this isn't a dataplane tile, and we're not
1194	 * doing syscall tracing, but there are enough cases now that it
1195	 * seems simplest just to do the reload unconditionally.
1196	 */
1197.Lrestore_syscall_regs:
1198	PTREGS_PTR(r11, PTREGS_OFFSET_REG(0))
1199	pop_reg r0, r11
1200	pop_reg r1, r11
1201	pop_reg r2, r11
1202	pop_reg r3, r11
1203	pop_reg r4, r11
1204	pop_reg r5, r11, PTREGS_OFFSET_SYSCALL - PTREGS_OFFSET_REG(5)
1205	pop_reg TREG_SYSCALL_NR_NAME, r11
1206
1207	/* Ensure that the syscall number is within the legal range. */
1208	moveli  r21, __NR_syscalls
1209	{
1210	 slt_u  r21, TREG_SYSCALL_NR_NAME, r21
1211	 moveli r20, lo16(sys_call_table)
1212	}
1213	{
1214	 bbns   r21, .Linvalid_syscall
1215	 auli   r20, r20, ha16(sys_call_table)
1216	}
1217	s2a     r20, TREG_SYSCALL_NR_NAME, r20
1218	lw      r20, r20
1219
1220	/* Jump to syscall handler. */
1221	jalr    r20
1222.Lhandle_syscall_link: /* value of "lr" after "jalr r20" above */
1223
1224	/*
1225	 * Write our r0 onto the stack so it gets restored instead
1226	 * of whatever the user had there before.
1227	 */
1228	PTREGS_PTR(r29, PTREGS_OFFSET_REG(0))
1229	sw      r29, r0
1230
1231.Lsyscall_sigreturn_skip:
1232	FEEDBACK_REENTER(handle_syscall)
1233
1234	/* Do syscall trace again, if requested. */
1235	lw	r30, r31
1236	andi    r30, r30, _TIF_SYSCALL_TRACE
1237	bzt     r30, 1f
1238	jal	do_syscall_trace
 
 
 
1239	FEEDBACK_REENTER(handle_syscall)
12401:	j       .Lresume_userspace   /* jump into middle of interrupt_return */
 
 
 
1241
1242.Linvalid_syscall:
1243	/* Report an invalid syscall back to the user program */
1244	{
1245	 PTREGS_PTR(r29, PTREGS_OFFSET_REG(0))
1246	 movei  r28, -ENOSYS
1247	}
1248	sw      r29, r28
1249	j       .Lresume_userspace   /* jump into middle of interrupt_return */
 
 
 
1250	STD_ENDPROC(handle_syscall)
1251
1252	/* Return the address for oprofile to suppress in backtraces. */
1253STD_ENTRY_SECTION(handle_syscall_link_address, .text.handle_syscall)
1254	lnk     r0
1255	{
1256	 addli  r0, r0, .Lhandle_syscall_link - .
1257	 jrp    lr
1258	}
1259	STD_ENDPROC(handle_syscall_link_address)
1260
1261STD_ENTRY(ret_from_fork)
1262	jal     sim_notify_fork
1263	jal     schedule_tail
1264	FEEDBACK_REENTER(ret_from_fork)
1265	j       .Lresume_userspace   /* jump into middle of interrupt_return */
 
 
 
1266	STD_ENDPROC(ret_from_fork)
1267
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1268	/*
1269	 * Code for ill interrupt.
1270	 */
1271	.pushsection .text.handle_ill,"ax"
1272handle_ill:
1273	finish_interrupt_save handle_ill
1274
1275	/*
1276	 * Check for if we are single stepping in user level. If so, then
1277	 * we need to restore the PC.
1278	 */
1279	check_single_stepping ill, .Ldispatch_normal_ill
1280
1281	{
1282	 /* See if the PC is the 1st bundle in the buffer */
1283	 seq    r25, r27, r26
1284
1285	 /* Point to the 2nd bundle in the buffer */
1286	 addi   r26, r26, 8
1287	}
1288	{
1289	 /* Point to the original pc */
1290	 addi   r24, r29, SINGLESTEP_STATE_ORIG_PC_OFFSET
1291
1292	 /* Branch if the PC is the 1st bundle in the buffer */
1293	 bnz    r25, 3f
1294	}
1295	{
1296	 /* See if the PC is the 2nd bundle of the buffer */
1297	 seq    r25, r27, r26
1298
1299	 /* Set PC to next instruction */
1300	 addi   r24, r29, SINGLESTEP_STATE_NEXT_PC_OFFSET
1301	}
1302	{
1303	 /* Point to flags */
1304	 addi   r25, r29, SINGLESTEP_STATE_FLAGS_OFFSET
1305
1306	 /* Branch if PC is in the second bundle */
1307	 bz     r25, 2f
1308	}
1309	/* Load flags */
1310	lw      r25, r25
1311	{
1312	 /*
1313	  * Get the offset for the register to restore
1314	  * Note: the lower bound is 2, so we have implicit scaling by 4.
1315	  *  No multiplication of the register number by the size of a register
1316	  *  is needed.
1317	  */
1318	 mm     r27, r25, zero, SINGLESTEP_STATE_TARGET_LB, \
1319		SINGLESTEP_STATE_TARGET_UB
1320
1321	 /* Mask Rewrite_LR */
1322	 andi   r25, r25, SINGLESTEP_STATE_MASK_UPDATE
1323	}
1324	{
1325	 addi   r29, r29, SINGLESTEP_STATE_UPDATE_VALUE_OFFSET
1326
1327	 /* Don't rewrite temp register */
1328	 bz     r25, 3f
1329	}
1330	{
1331	 /* Get the temp value */
1332	 lw     r29, r29
1333
1334	 /* Point to where the register is stored */
1335	 add    r27, r27, sp
1336	}
1337
1338	/* Add in the C ABI save area size to the register offset */
1339	addi    r27, r27, C_ABI_SAVE_AREA_SIZE
1340
1341	/* Restore the user's register with the temp value */
1342	sw      r27, r29
1343	j       3f
1344
13452:
1346	/* Must be in the third bundle */
1347	addi    r24, r29, SINGLESTEP_STATE_BRANCH_NEXT_PC_OFFSET
1348
13493:
1350	/* set PC and continue */
1351	lw      r26, r24
1352	sw      r28, r26
 
 
 
1353
1354	/*
1355	 * Clear TIF_SINGLESTEP to prevent recursion if we execute an ill.
1356	 * The normal non-arch flow redundantly clears TIF_SINGLESTEP, but we
1357	 * need to clear it here and can't really impose on all other arches.
1358	 * So what's another write between friends?
1359	 */
1360	GET_THREAD_INFO(r0)
1361
1362	addi    r1, r0, THREAD_INFO_FLAGS_OFFSET
1363	{
1364	 lw     r2, r1
1365	 addi   r0, r0, THREAD_INFO_TASK_OFFSET  /* currently a no-op */
1366	}
1367	andi    r2, r2, ~_TIF_SINGLESTEP
1368	sw      r1, r2
1369
1370	/* Issue a sigtrap */
1371	{
1372	 lw     r0, r0          /* indirect thru thread_info to get task_info*/
1373	 addi   r1, sp, C_ABI_SAVE_AREA_SIZE  /* put ptregs pointer into r1 */
1374	 move   r2, zero        /* load error code into r2 */
1375	}
1376
1377	jal     send_sigtrap    /* issue a SIGTRAP */
1378	FEEDBACK_REENTER(handle_ill)
1379	j       .Lresume_userspace   /* jump into middle of interrupt_return */
 
 
 
1380
1381.Ldispatch_normal_ill:
1382	{
1383	 jalr   r0
1384	 PTREGS_PTR(r0, PTREGS_OFFSET_BASE)
1385	}
1386	FEEDBACK_REENTER(handle_ill)
1387	{
1388	 movei  r30, 0   /* not an NMI */
1389	 j      interrupt_return
1390	}
1391	STD_ENDPROC(handle_ill)
1392
1393/* Various stub interrupt handlers and syscall handlers */
1394
1395STD_ENTRY_LOCAL(_kernel_double_fault)
1396	mfspr   r1, SPR_EX_CONTEXT_K_0
1397	move    r2, lr
1398	move    r3, sp
1399	move    r4, r52
1400	addi    sp, sp, -C_ABI_SAVE_AREA_SIZE
1401	j       kernel_double_fault
1402	STD_ENDPROC(_kernel_double_fault)
1403
1404STD_ENTRY_LOCAL(bad_intr)
1405	mfspr   r2, SPR_EX_CONTEXT_K_0
1406	panic   "Unhandled interrupt %#x: PC %#lx"
1407	STD_ENDPROC(bad_intr)
1408
1409/* Put address of pt_regs in reg and jump. */
1410#define PTREGS_SYSCALL(x, reg)                          \
1411	STD_ENTRY(_##x);                                \
1412	{                                               \
1413	 PTREGS_PTR(reg, PTREGS_OFFSET_BASE);           \
1414	 j      x                                       \
1415	};                                              \
1416	STD_ENDPROC(_##x)
1417
1418/*
1419 * Special-case sigreturn to not write r0 to the stack on return.
1420 * This is technically more efficient, but it also avoids difficulties
1421 * in the 64-bit OS when handling 32-bit compat code, since we must not
1422 * sign-extend r0 for the sigreturn return-value case.
1423 */
1424#define PTREGS_SYSCALL_SIGRETURN(x, reg)                \
1425	STD_ENTRY(_##x);                                \
1426	addli   lr, lr, .Lsyscall_sigreturn_skip - .Lhandle_syscall_link; \
1427	{                                               \
1428	 PTREGS_PTR(reg, PTREGS_OFFSET_BASE);           \
1429	 j      x                                       \
1430	};                                              \
1431	STD_ENDPROC(_##x)
1432
1433PTREGS_SYSCALL(sys_execve, r3)
1434PTREGS_SYSCALL(sys_sigaltstack, r2)
1435PTREGS_SYSCALL_SIGRETURN(sys_rt_sigreturn, r0)
1436PTREGS_SYSCALL(sys_cmpxchg_badaddr, r1)
1437
1438/* Save additional callee-saves to pt_regs, put address in r4 and jump. */
1439STD_ENTRY(_sys_clone)
1440	push_extra_callee_saves r4
1441	j       sys_clone
1442	STD_ENDPROC(_sys_clone)
1443
1444/*
1445 * This entrypoint is taken for the cmpxchg and atomic_update fast
1446 * swints.  We may wish to generalize it to other fast swints at some
1447 * point, but for now there are just two very similar ones, which
1448 * makes it faster.
1449 *
1450 * The fast swint code is designed to have a small footprint.  It does
1451 * not save or restore any GPRs, counting on the caller-save registers
1452 * to be available to it on entry.  It does not modify any callee-save
1453 * registers (including "lr").  It does not check what PL it is being
1454 * called at, so you'd better not call it other than at PL0.
1455 * The <atomic.h> wrapper assumes it only clobbers r20-r29, so if
1456 * it ever is necessary to use more registers, be aware.
1457 *
1458 * It does not use the stack, but since it might be re-interrupted by
1459 * a page fault which would assume the stack was valid, it does
1460 * save/restore the stack pointer and zero it out to make sure it gets reset.
1461 * Since we always keep interrupts disabled, the hypervisor won't
1462 * clobber our EX_CONTEXT_K_x registers, so we don't save/restore them
1463 * (other than to advance the PC on return).
1464 *
1465 * We have to manually validate the user vs kernel address range
1466 * (since at PL1 we can read/write both), and for performance reasons
1467 * we don't allow cmpxchg on the fc000000 memory region, since we only
1468 * validate that the user address is below PAGE_OFFSET.
1469 *
1470 * We place it in the __HEAD section to ensure it is relatively
1471 * near to the intvec_SWINT_1 code (reachable by a conditional branch).
1472 *
1473 * Our use of ATOMIC_LOCK_REG here must match do_page_fault_ics().
1474 *
1475 * As we do in lib/atomic_asm_32.S, we bypass a store if the value we
1476 * would store is the same as the value we just loaded.
1477 */
1478	__HEAD
1479	.align 64
1480	/* Align much later jump on the start of a cache line. */
1481#if !ATOMIC_LOCKS_FOUND_VIA_TABLE()
1482	nop
1483#if PAGE_SIZE >= 0x10000
1484	nop
1485#endif
1486#endif
1487ENTRY(sys_cmpxchg)
1488
1489	/*
1490	 * Save "sp" and set it zero for any possible page fault.
1491	 *
1492	 * HACK: We want to both zero sp and check r0's alignment,
1493	 * so we do both at once. If "sp" becomes nonzero we
1494	 * know r0 is unaligned and branch to the error handler that
1495	 * restores sp, so this is OK.
1496	 *
1497	 * ICS is disabled right now so having a garbage but nonzero
1498	 * sp is OK, since we won't execute any faulting instructions
1499	 * when it is nonzero.
1500	 */
1501	{
1502	 move   r27, sp
1503	 andi	sp, r0, 3
1504	}
1505
1506	/*
1507	 * Get the lock address in ATOMIC_LOCK_REG, and also validate that the
1508	 * address is less than PAGE_OFFSET, since that won't trap at PL1.
1509	 * We only use bits less than PAGE_SHIFT to avoid having to worry
1510	 * about aliasing among multiple mappings of the same physical page,
1511	 * and we ignore the low 3 bits so we have one lock that covers
1512	 * both a cmpxchg64() and a cmpxchg() on either its low or high word.
1513	 * NOTE: this must match __atomic_hashed_lock() in lib/atomic_32.c.
1514	 */
1515
1516#if (PAGE_OFFSET & 0xffff) != 0
1517# error Code here assumes PAGE_OFFSET can be loaded with just hi16()
1518#endif
1519
1520#if ATOMIC_LOCKS_FOUND_VIA_TABLE()
1521	{
1522	 /* Check for unaligned input. */
1523	 bnz    sp, .Lcmpxchg_badaddr
1524	 mm     r25, r0, zero, 3, PAGE_SHIFT-1
1525	}
1526	{
1527	 crc32_32 r25, zero, r25
1528	 moveli r21, lo16(atomic_lock_ptr)
1529	}
1530	{
1531	 auli   r21, r21, ha16(atomic_lock_ptr)
1532	 auli   r23, zero, hi16(PAGE_OFFSET)  /* hugepage-aligned */
1533	}
1534	{
1535	 shri	r20, r25, 32 - ATOMIC_HASH_L1_SHIFT
1536	 slt_u  r23, r0, r23
1537	 lw	r26, r0  /* see comment in the "#else" for the "lw r26". */
1538	}
1539	{
1540	 s2a    r21, r20, r21
1541	 bbns   r23, .Lcmpxchg_badaddr
1542	}
1543	{
1544	 lw     r21, r21
1545	 seqi	r23, TREG_SYSCALL_NR_NAME, __NR_FAST_cmpxchg64
1546	 andi	r25, r25, ATOMIC_HASH_L2_SIZE - 1
1547	}
1548	{
1549	 /* Branch away at this point if we're doing a 64-bit cmpxchg. */
1550	 bbs    r23, .Lcmpxchg64
1551	 andi   r23, r0, 7       /* Precompute alignment for cmpxchg64. */
1552	}
1553	{
1554	 s2a	ATOMIC_LOCK_REG_NAME, r25, r21
1555	 j      .Lcmpxchg32_tns   /* see comment in the #else for the jump. */
1556	}
1557
1558#else /* ATOMIC_LOCKS_FOUND_VIA_TABLE() */
1559	{
1560	 /* Check for unaligned input. */
1561	 bnz    sp, .Lcmpxchg_badaddr
1562	 auli   r23, zero, hi16(PAGE_OFFSET)  /* hugepage-aligned */
1563	}
1564	{
1565	 /*
1566	  * Slide bits into position for 'mm'. We want to ignore
1567	  * the low 3 bits of r0, and consider only the next
1568	  * ATOMIC_HASH_SHIFT bits.
1569	  * Because of C pointer arithmetic, we want to compute this:
1570	  *
1571	  * ((char*)atomic_locks +
1572	  *  (((r0 >> 3) & (1 << (ATOMIC_HASH_SIZE - 1))) << 2))
1573	  *
1574	  * Instead of two shifts we just ">> 1", and use 'mm'
1575	  * to ignore the low and high bits we don't want.
1576	  */
1577	 shri	r25, r0, 1
1578
1579	 slt_u  r23, r0, r23
1580
1581	 /*
1582	  * Ensure that the TLB is loaded before we take out the lock.
1583	  * On tilepro, this will start fetching the value all the way
1584	  * into our L1 as well (and if it gets modified before we
1585	  * grab the lock, it will be invalidated from our cache
1586	  * before we reload it).  On tile64, we'll start fetching it
1587	  * into our L1 if we're the home, and if we're not, we'll
1588	  * still at least start fetching it into the home's L2.
1589	  */
1590	 lw	r26, r0
1591	}
1592	{
1593	 auli	r21, zero, ha16(atomic_locks)
1594
1595	 bbns   r23, .Lcmpxchg_badaddr
1596	}
1597#if PAGE_SIZE < 0x10000
1598	/* atomic_locks is page-aligned so for big pages we don't need this. */
1599	addli   r21, r21, lo16(atomic_locks)
1600#endif
1601	{
1602	 /*
1603	  * Insert the hash bits into the page-aligned pointer.
1604	  * ATOMIC_HASH_SHIFT is so big that we don't actually hash
1605	  * the unmasked address bits, as that may cause unnecessary
1606	  * collisions.
1607	  */
1608	 mm	ATOMIC_LOCK_REG_NAME, r25, r21, 2, (ATOMIC_HASH_SHIFT + 2) - 1
1609
1610	 seqi	r23, TREG_SYSCALL_NR_NAME, __NR_FAST_cmpxchg64
1611	}
1612	{
1613	 /* Branch away at this point if we're doing a 64-bit cmpxchg. */
1614	 bbs    r23, .Lcmpxchg64
1615	 andi   r23, r0, 7       /* Precompute alignment for cmpxchg64. */
1616	}
1617	{
1618	 /*
1619	  * We very carefully align the code that actually runs with
1620	  * the lock held (twelve bundles) so that we know it is all in
1621	  * the icache when we start.  This instruction (the jump) is
1622	  * at the start of the first cache line, address zero mod 64;
1623	  * we jump to the very end of the second cache line to get that
1624	  * line loaded in the icache, then fall through to issue the tns
1625	  * in the third cache line, at which point it's all cached.
1626	  * Note that is for performance, not correctness.
1627	  */
1628	 j      .Lcmpxchg32_tns
1629	}
1630
1631#endif /* ATOMIC_LOCKS_FOUND_VIA_TABLE() */
1632
1633/* Symbol for do_page_fault_ics() to use to compare against the PC. */
1634.global __sys_cmpxchg_grab_lock
1635__sys_cmpxchg_grab_lock:
1636
1637	/*
1638	 * Perform the actual cmpxchg or atomic_update.
1639	 */
1640.Ldo_cmpxchg32:
1641	{
1642	 lw     r21, r0
1643	 seqi	r23, TREG_SYSCALL_NR_NAME, __NR_FAST_atomic_update
1644	 move	r24, r2
1645	}
1646	{
1647	 seq    r22, r21, r1     /* See if cmpxchg matches. */
1648	 and	r25, r21, r1     /* If atomic_update, compute (*mem & mask) */
1649	}
1650	{
1651	 or	r22, r22, r23    /* Skip compare branch for atomic_update. */
1652	 add	r25, r25, r2     /* Compute (*mem & mask) + addend. */
1653	}
1654	{
1655	 mvnz	r24, r23, r25    /* Use atomic_update value if appropriate. */
1656	 bbns   r22, .Lcmpxchg32_nostore
1657	}
1658	seq     r22, r24, r21    /* Are we storing the value we loaded? */
1659	bbs     r22, .Lcmpxchg32_nostore
1660	sw      r0, r24
1661
1662	/* The following instruction is the start of the second cache line. */
1663	/* Do slow mtspr here so the following "mf" waits less. */
1664	{
1665	 move   sp, r27
1666	 mtspr  SPR_EX_CONTEXT_K_0, r28
1667	}
1668	mf
1669
1670	{
1671	 move   r0, r21
1672	 sw     ATOMIC_LOCK_REG_NAME, zero
1673	}
1674	iret
1675
1676	/* Duplicated code here in the case where we don't overlap "mf" */
1677.Lcmpxchg32_nostore:
1678	{
1679	 move   r0, r21
1680	 sw     ATOMIC_LOCK_REG_NAME, zero
1681	}
1682	{
1683	 move   sp, r27
1684	 mtspr  SPR_EX_CONTEXT_K_0, r28
1685	}
1686	iret
1687
1688	/*
1689	 * The locking code is the same for 32-bit cmpxchg/atomic_update,
1690	 * and for 64-bit cmpxchg.  We provide it as a macro and put
1691	 * it into both versions.  We can't share the code literally
1692	 * since it depends on having the right branch-back address.
1693	 */
1694	.macro  cmpxchg_lock, bitwidth
1695
1696	/* Lock; if we succeed, jump back up to the read-modify-write. */
1697#ifdef CONFIG_SMP
1698	tns     r21, ATOMIC_LOCK_REG_NAME
1699#else
1700	/*
1701	 * Non-SMP preserves all the lock infrastructure, to keep the
1702	 * code simpler for the interesting (SMP) case.  However, we do
1703	 * one small optimization here and in atomic_asm.S, which is
1704	 * to fake out acquiring the actual lock in the atomic_lock table.
1705	 */
1706	movei	r21, 0
1707#endif
1708
1709	/* Issue the slow SPR here while the tns result is in flight. */
1710	mfspr   r28, SPR_EX_CONTEXT_K_0
1711
1712	{
1713	 addi   r28, r28, 8    /* return to the instruction after the swint1 */
1714	 bzt    r21, .Ldo_cmpxchg\bitwidth
1715	}
1716	/*
1717	 * The preceding instruction is the last thing that must be
1718	 * hot in the icache before we do the "tns" above.
1719	 */
1720
1721#ifdef CONFIG_SMP
1722	/*
1723	 * We failed to acquire the tns lock on our first try.  Now use
1724	 * bounded exponential backoff to retry, like __atomic_spinlock().
1725	 */
1726	{
1727	 moveli r23, 2048       /* maximum backoff time in cycles */
1728	 moveli r25, 32         /* starting backoff time in cycles */
1729	}
17301:	mfspr   r26, CYCLE_LOW  /* get start point for this backoff */
17312:	mfspr   r22, CYCLE_LOW  /* test to see if we've backed off enough */
1732	sub     r22, r22, r26
1733	slt     r22, r22, r25
1734	bbst    r22, 2b
1735	{
1736	 shli   r25, r25, 1     /* double the backoff; retry the tns */
1737	 tns    r21, ATOMIC_LOCK_REG_NAME
1738	}
1739	slt     r26, r23, r25   /* is the proposed backoff too big? */
1740	{
1741	 mvnz   r25, r26, r23
1742	 bzt    r21, .Ldo_cmpxchg\bitwidth
1743	}
1744	j       1b
1745#endif /* CONFIG_SMP */
1746	.endm
1747
1748.Lcmpxchg32_tns:
1749	/*
1750	 * This is the last instruction on the second cache line.
1751	 * The nop here loads the second line, then we fall through
1752	 * to the tns to load the third line before we take the lock.
1753	 */
1754	nop
1755	cmpxchg_lock 32
1756
1757	/*
1758	 * This code is invoked from sys_cmpxchg after most of the
1759	 * preconditions have been checked.  We still need to check
1760	 * that r0 is 8-byte aligned, since if it's not we won't
1761	 * actually be atomic.  However, ATOMIC_LOCK_REG has the atomic
1762	 * lock pointer and r27/r28 have the saved SP/PC.
1763	 * r23 is holding "r0 & 7" so we can test for alignment.
1764	 * The compare value is in r2/r3; the new value is in r4/r5.
1765	 * On return, we must put the old value in r0/r1.
1766	 */
1767	.align 64
1768.Lcmpxchg64:
1769	{
1770#if ATOMIC_LOCKS_FOUND_VIA_TABLE()
1771	 s2a	ATOMIC_LOCK_REG_NAME, r25, r21
1772#endif
1773	 bzt     r23, .Lcmpxchg64_tns
1774	}
1775	j       .Lcmpxchg_badaddr
1776
1777.Ldo_cmpxchg64:
1778	{
1779	 lw     r21, r0
1780	 addi   r25, r0, 4
1781	}
1782	{
1783	 lw     r1, r25
1784	}
1785	seq     r26, r21, r2
1786	{
1787	 bz     r26, .Lcmpxchg64_mismatch
1788	 seq    r26, r1, r3
1789	}
1790	{
1791	 bz     r26, .Lcmpxchg64_mismatch
1792	}
1793	sw      r0, r4
1794	sw      r25, r5
1795
1796	/*
1797	 * The 32-bit path provides optimized "match" and "mismatch"
1798	 * iret paths, but we don't have enough bundles in this cache line
1799	 * to do that, so we just make even the "mismatch" path do an "mf".
1800	 */
1801.Lcmpxchg64_mismatch:
1802	{
1803	 move   sp, r27
1804	 mtspr  SPR_EX_CONTEXT_K_0, r28
1805	}
1806	mf
1807	{
1808	 move   r0, r21
1809	 sw     ATOMIC_LOCK_REG_NAME, zero
1810	}
1811	iret
1812
1813.Lcmpxchg64_tns:
1814	cmpxchg_lock 64
1815
1816
1817	/*
1818	 * Reset sp and revector to sys_cmpxchg_badaddr(), which will
1819	 * just raise the appropriate signal and exit.  Doing it this
1820	 * way means we don't have to duplicate the code in intvec.S's
1821	 * int_hand macro that locates the top of the stack.
1822	 */
1823.Lcmpxchg_badaddr:
1824	{
1825	 moveli TREG_SYSCALL_NR_NAME, __NR_cmpxchg_badaddr
1826	 move   sp, r27
1827	}
1828	j       intvec_SWINT_1
1829	ENDPROC(sys_cmpxchg)
1830	ENTRY(__sys_cmpxchg_end)
1831
1832
1833/* The single-step support may need to read all the registers. */
1834int_unalign:
1835	push_extra_callee_saves r0
1836	j       do_trap
1837
1838/* Include .intrpt1 array of interrupt vectors */
1839	.section ".intrpt1", "ax"
1840
1841#define op_handle_perf_interrupt bad_intr
1842#define op_handle_aux_perf_interrupt bad_intr
 
1843
1844#ifndef CONFIG_HARDWALL
1845#define do_hardwall_trap bad_intr
1846#endif
1847
1848	int_hand     INT_ITLB_MISS, ITLB_MISS, \
1849		     do_page_fault, handle_interrupt_no_single_step
1850	int_hand     INT_MEM_ERROR, MEM_ERROR, bad_intr
1851	int_hand     INT_ILL, ILL, do_trap, handle_ill
1852	int_hand     INT_GPV, GPV, do_trap
1853	int_hand     INT_SN_ACCESS, SN_ACCESS, do_trap
1854	int_hand     INT_IDN_ACCESS, IDN_ACCESS, do_trap
1855	int_hand     INT_UDN_ACCESS, UDN_ACCESS, do_trap
1856	int_hand     INT_IDN_REFILL, IDN_REFILL, bad_intr
1857	int_hand     INT_UDN_REFILL, UDN_REFILL, bad_intr
1858	int_hand     INT_IDN_COMPLETE, IDN_COMPLETE, bad_intr
1859	int_hand     INT_UDN_COMPLETE, UDN_COMPLETE, bad_intr
1860	int_hand     INT_SWINT_3, SWINT_3, do_trap
1861	int_hand     INT_SWINT_2, SWINT_2, do_trap
1862	int_hand     INT_SWINT_1, SWINT_1, SYSCALL, handle_syscall
1863	int_hand     INT_SWINT_0, SWINT_0, do_trap
1864	int_hand     INT_UNALIGN_DATA, UNALIGN_DATA, int_unalign
1865	int_hand     INT_DTLB_MISS, DTLB_MISS, do_page_fault
1866	int_hand     INT_DTLB_ACCESS, DTLB_ACCESS, do_page_fault
1867	int_hand     INT_DMATLB_MISS, DMATLB_MISS, do_page_fault
1868	int_hand     INT_DMATLB_ACCESS, DMATLB_ACCESS, do_page_fault
1869	int_hand     INT_SNITLB_MISS, SNITLB_MISS, do_page_fault
1870	int_hand     INT_SN_NOTIFY, SN_NOTIFY, bad_intr
1871	int_hand     INT_SN_FIREWALL, SN_FIREWALL, do_hardwall_trap
1872	int_hand     INT_IDN_FIREWALL, IDN_FIREWALL, bad_intr
1873	int_hand     INT_UDN_FIREWALL, UDN_FIREWALL, do_hardwall_trap
1874	int_hand     INT_TILE_TIMER, TILE_TIMER, do_timer_interrupt
1875	int_hand     INT_IDN_TIMER, IDN_TIMER, bad_intr
1876	int_hand     INT_UDN_TIMER, UDN_TIMER, bad_intr
1877	int_hand     INT_DMA_NOTIFY, DMA_NOTIFY, bad_intr
1878	int_hand     INT_IDN_CA, IDN_CA, bad_intr
1879	int_hand     INT_UDN_CA, UDN_CA, bad_intr
1880	int_hand     INT_IDN_AVAIL, IDN_AVAIL, bad_intr
1881	int_hand     INT_UDN_AVAIL, UDN_AVAIL, bad_intr
1882	int_hand     INT_PERF_COUNT, PERF_COUNT, \
1883		     op_handle_perf_interrupt, handle_nmi
1884	int_hand     INT_INTCTRL_3, INTCTRL_3, bad_intr
1885#if CONFIG_KERNEL_PL == 2
1886	dc_dispatch  INT_INTCTRL_2, INTCTRL_2
1887	int_hand     INT_INTCTRL_1, INTCTRL_1, bad_intr
1888#else
1889	int_hand     INT_INTCTRL_2, INTCTRL_2, bad_intr
1890	dc_dispatch  INT_INTCTRL_1, INTCTRL_1
1891#endif
1892	int_hand     INT_INTCTRL_0, INTCTRL_0, bad_intr
1893	int_hand     INT_MESSAGE_RCV_DWNCL, MESSAGE_RCV_DWNCL, \
1894		     hv_message_intr
1895	int_hand     INT_DEV_INTR_DWNCL, DEV_INTR_DWNCL, \
1896		     tile_dev_intr
1897	int_hand     INT_I_ASID, I_ASID, bad_intr
1898	int_hand     INT_D_ASID, D_ASID, bad_intr
1899	int_hand     INT_DMATLB_MISS_DWNCL, DMATLB_MISS_DWNCL, \
1900		     do_page_fault
1901	int_hand     INT_SNITLB_MISS_DWNCL, SNITLB_MISS_DWNCL, \
1902		     do_page_fault
1903	int_hand     INT_DMATLB_ACCESS_DWNCL, DMATLB_ACCESS_DWNCL, \
1904		     do_page_fault
1905	int_hand     INT_SN_CPL, SN_CPL, bad_intr
1906	int_hand     INT_DOUBLE_FAULT, DOUBLE_FAULT, do_trap
1907#if CHIP_HAS_AUX_PERF_COUNTERS()
1908	int_hand     INT_AUX_PERF_COUNT, AUX_PERF_COUNT, \
1909		     op_handle_aux_perf_interrupt, handle_nmi
1910#endif
1911
1912	/* Synthetic interrupt delivered only by the simulator */
1913	int_hand     INT_BREAKPOINT, BREAKPOINT, do_breakpoint
v3.15
   1/*
   2 * Copyright 2010 Tilera Corporation. All Rights Reserved.
   3 *
   4 *   This program is free software; you can redistribute it and/or
   5 *   modify it under the terms of the GNU General Public License
   6 *   as published by the Free Software Foundation, version 2.
   7 *
   8 *   This program is distributed in the hope that it will be useful, but
   9 *   WITHOUT ANY WARRANTY; without even the implied warranty of
  10 *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
  11 *   NON INFRINGEMENT.  See the GNU General Public License for
  12 *   more details.
  13 *
  14 * Linux interrupt vectors.
  15 */
  16
  17#include <linux/linkage.h>
  18#include <linux/errno.h>
  19#include <linux/init.h>
  20#include <linux/unistd.h>
  21#include <asm/ptrace.h>
  22#include <asm/thread_info.h>
  23#include <asm/irqflags.h>
  24#include <asm/atomic_32.h>
  25#include <asm/asm-offsets.h>
  26#include <hv/hypervisor.h>
  27#include <arch/abi.h>
  28#include <arch/interrupts.h>
  29#include <arch/spr_def.h>
  30
 
 
 
 
  31#define PTREGS_PTR(reg, ptreg) addli reg, sp, C_ABI_SAVE_AREA_SIZE + (ptreg)
  32
  33#define PTREGS_OFFSET_SYSCALL PTREGS_OFFSET_REG(TREG_SYSCALL_NR)
  34
 
 
 
 
 
 
  35	.macro  push_reg reg, ptr=sp, delta=-4
  36	{
  37	 sw     \ptr, \reg
  38	 addli  \ptr, \ptr, \delta
  39	}
  40	.endm
  41
  42	.macro  pop_reg reg, ptr=sp, delta=4
  43	{
  44	 lw     \reg, \ptr
  45	 addli  \ptr, \ptr, \delta
  46	}
  47	.endm
  48
  49	.macro  pop_reg_zero reg, zreg, ptr=sp, delta=4
  50	{
  51	 move   \zreg, zero
  52	 lw     \reg, \ptr
  53	 addi   \ptr, \ptr, \delta
  54	}
  55	.endm
  56
  57	.macro  push_extra_callee_saves reg
  58	PTREGS_PTR(\reg, PTREGS_OFFSET_REG(51))
  59	push_reg r51, \reg
  60	push_reg r50, \reg
  61	push_reg r49, \reg
  62	push_reg r48, \reg
  63	push_reg r47, \reg
  64	push_reg r46, \reg
  65	push_reg r45, \reg
  66	push_reg r44, \reg
  67	push_reg r43, \reg
  68	push_reg r42, \reg
  69	push_reg r41, \reg
  70	push_reg r40, \reg
  71	push_reg r39, \reg
  72	push_reg r38, \reg
  73	push_reg r37, \reg
  74	push_reg r36, \reg
  75	push_reg r35, \reg
  76	push_reg r34, \reg, PTREGS_OFFSET_BASE - PTREGS_OFFSET_REG(34)
  77	.endm
  78
  79	.macro  panic str
  80	.pushsection .rodata, "a"
  811:
  82	.asciz  "\str"
  83	.popsection
  84	{
  85	 moveli r0, lo16(1b)
  86	}
  87	{
  88	 auli   r0, r0, ha16(1b)
  89	 jal    panic
  90	}
  91	.endm
  92
  93#ifdef __COLLECT_LINKER_FEEDBACK__
  94	.pushsection .text.intvec_feedback,"ax"
  95intvec_feedback:
  96	.popsection
  97#endif
  98
  99	/*
 100	 * Default interrupt handler.
 101	 *
 102	 * vecnum is where we'll put this code.
 103	 * c_routine is the C routine we'll call.
 104	 *
 105	 * The C routine is passed two arguments:
 106	 * - A pointer to the pt_regs state.
 107	 * - The interrupt vector number.
 108	 *
 109	 * The "processing" argument specifies the code for processing
 110	 * the interrupt. Defaults to "handle_interrupt".
 111	 */
 112	.macro  int_hand vecnum, vecname, c_routine, processing=handle_interrupt
 113	.org    (\vecnum << 8)
 114intvec_\vecname:
 115	.ifc    \vecnum, INT_SWINT_1
 116	blz     TREG_SYSCALL_NR_NAME, sys_cmpxchg
 117	.endif
 118
 119	/* Temporarily save a register so we have somewhere to work. */
 120
 121	mtspr   SPR_SYSTEM_SAVE_K_1, r0
 122	mfspr   r0, SPR_EX_CONTEXT_K_1
 123
 124	/* The cmpxchg code clears sp to force us to reset it here on fault. */
 125	{
 126	 bz     sp, 2f
 127	 andi   r0, r0, SPR_EX_CONTEXT_1_1__PL_MASK  /* mask off ICS */
 128	}
 129
 130	.ifc    \vecnum, INT_DOUBLE_FAULT
 131	/*
 132	 * For double-faults from user-space, fall through to the normal
 133	 * register save and stack setup path.  Otherwise, it's the
 134	 * hypervisor giving us one last chance to dump diagnostics, and we
 135	 * branch to the kernel_double_fault routine to do so.
 136	 */
 137	bz      r0, 1f
 138	j       _kernel_double_fault
 1391:
 140	.else
 141	/*
 142	 * If we're coming from user-space, then set sp to the top of
 143	 * the kernel stack.  Otherwise, assume sp is already valid.
 144	 */
 145	{
 146	 bnz    r0, 0f
 147	 move   r0, sp
 148	}
 149	.endif
 150
 151	.ifc    \c_routine, do_page_fault
 152	/*
 153	 * The page_fault handler may be downcalled directly by the
 154	 * hypervisor even when Linux is running and has ICS set.
 155	 *
 156	 * In this case the contents of EX_CONTEXT_K_1 reflect the
 157	 * previous fault and can't be relied on to choose whether or
 158	 * not to reinitialize the stack pointer.  So we add a test
 159	 * to see whether SYSTEM_SAVE_K_2 has the high bit set,
 160	 * and if so we don't reinitialize sp, since we must be coming
 161	 * from Linux.  (In fact the precise case is !(val & ~1),
 162	 * but any Linux PC has to have the high bit set.)
 163	 *
 164	 * Note that the hypervisor *always* sets SYSTEM_SAVE_K_2 for
 165	 * any path that turns into a downcall to one of our TLB handlers.
 166	 */
 167	mfspr   r0, SPR_SYSTEM_SAVE_K_2
 168	{
 169	 blz    r0, 0f    /* high bit in S_S_1_2 is for a PC to use */
 170	 move   r0, sp
 171	}
 172	.endif
 173
 1742:
 175	/*
 176	 * SYSTEM_SAVE_K_0 holds the cpu number in the low bits, and
 177	 * the current stack top in the higher bits.  So we recover
 178	 * our stack top by just masking off the low bits, then
 179	 * point sp at the top aligned address on the actual stack page.
 180	 */
 181	mfspr   r0, SPR_SYSTEM_SAVE_K_0
 182	mm      r0, r0, zero, LOG2_NR_CPU_IDS, 31
 183
 1840:
 185	/*
 186	 * Align the stack mod 64 so we can properly predict what
 187	 * cache lines we need to write-hint to reduce memory fetch
 188	 * latency as we enter the kernel.  The layout of memory is
 189	 * as follows, with cache line 0 at the lowest VA, and cache
 190	 * line 4 just below the r0 value this "andi" computes.
 191	 * Note that we never write to cache line 4, and we skip
 192	 * cache line 1 for syscalls.
 193	 *
 194	 *    cache line 4: ptregs padding (two words)
 195	 *    cache line 3: r46...lr, pc, ex1, faultnum, orig_r0, flags, pad
 196	 *    cache line 2: r30...r45
 197	 *    cache line 1: r14...r29
 198	 *    cache line 0: 2 x frame, r0..r13
 199	 */
 200#if STACK_TOP_DELTA != 64
 201#error STACK_TOP_DELTA must be 64 for assumptions here and in task_pt_regs()
 202#endif
 203	andi    r0, r0, -64
 204
 205	/*
 206	 * Push the first four registers on the stack, so that we can set
 207	 * them to vector-unique values before we jump to the common code.
 208	 *
 209	 * Registers are pushed on the stack as a struct pt_regs,
 210	 * with the sp initially just above the struct, and when we're
 211	 * done, sp points to the base of the struct, minus
 212	 * C_ABI_SAVE_AREA_SIZE, so we can directly jal to C code.
 213	 *
 214	 * This routine saves just the first four registers, plus the
 215	 * stack context so we can do proper backtracing right away,
 216	 * and defers to handle_interrupt to save the rest.
 217	 * The backtracer needs pc, ex1, lr, sp, r52, and faultnum.
 218	 */
 219	addli   r0, r0, PTREGS_OFFSET_LR - (PTREGS_SIZE + KSTK_PTREGS_GAP)
 220	wh64    r0    /* cache line 3 */
 221	{
 222	 sw     r0, lr
 223	 addli  r0, r0, PTREGS_OFFSET_SP - PTREGS_OFFSET_LR
 224	}
 225	{
 226	 sw     r0, sp
 227	 addli  sp, r0, PTREGS_OFFSET_REG(52) - PTREGS_OFFSET_SP
 228	}
 229	{
 230	 sw     sp, r52
 231	 addli  sp, sp, PTREGS_OFFSET_REG(1) - PTREGS_OFFSET_REG(52)
 232	}
 233	wh64    sp    /* cache line 0 */
 234	{
 235	 sw     sp, r1
 236	 addli  sp, sp, PTREGS_OFFSET_REG(2) - PTREGS_OFFSET_REG(1)
 237	}
 238	{
 239	 sw     sp, r2
 240	 addli  sp, sp, PTREGS_OFFSET_REG(3) - PTREGS_OFFSET_REG(2)
 241	}
 242	{
 243	 sw     sp, r3
 244	 addli  sp, sp, PTREGS_OFFSET_PC - PTREGS_OFFSET_REG(3)
 245	}
 246	mfspr   r0, SPR_EX_CONTEXT_K_0
 247	.ifc \processing,handle_syscall
 248	/*
 249	 * Bump the saved PC by one bundle so that when we return, we won't
 250	 * execute the same swint instruction again.  We need to do this while
 251	 * we're in the critical section.
 252	 */
 253	addi    r0, r0, 8
 254	.endif
 255	{
 256	 sw     sp, r0
 257	 addli  sp, sp, PTREGS_OFFSET_EX1 - PTREGS_OFFSET_PC
 258	}
 259	mfspr   r0, SPR_EX_CONTEXT_K_1
 260	{
 261	 sw     sp, r0
 262	 addi   sp, sp, PTREGS_OFFSET_FAULTNUM - PTREGS_OFFSET_EX1
 263	/*
 264	 * Use r0 for syscalls so it's a temporary; use r1 for interrupts
 265	 * so that it gets passed through unchanged to the handler routine.
 266	 * Note that the .if conditional confusingly spans bundles.
 267	 */
 268	 .ifc \processing,handle_syscall
 269	 movei  r0, \vecnum
 270	}
 271	{
 272	 sw     sp, r0
 273	 .else
 274	 movei  r1, \vecnum
 275	}
 276	{
 277	 sw     sp, r1
 278	 .endif
 279	 addli  sp, sp, PTREGS_OFFSET_REG(0) - PTREGS_OFFSET_FAULTNUM
 280	}
 281	mfspr   r0, SPR_SYSTEM_SAVE_K_1    /* Original r0 */
 282	{
 283	 sw     sp, r0
 284	 addi   sp, sp, -PTREGS_OFFSET_REG(0) - 4
 285	}
 286	{
 287	 sw     sp, zero        /* write zero into "Next SP" frame pointer */
 288	 addi   sp, sp, -4      /* leave SP pointing at bottom of frame */
 289	}
 290	.ifc \processing,handle_syscall
 291	j       handle_syscall
 292	.else
 293	/*
 294	 * Capture per-interrupt SPR context to registers.
 295	 * We overload the meaning of r3 on this path such that if its bit 31
 296	 * is set, we have to mask all interrupts including NMIs before
 297	 * clearing the interrupt critical section bit.
 298	 * See discussion below at "finish_interrupt_save".
 299	 */
 300	.ifc \c_routine, do_page_fault
 301	mfspr   r2, SPR_SYSTEM_SAVE_K_3   /* address of page fault */
 302	mfspr   r3, SPR_SYSTEM_SAVE_K_2   /* info about page fault */
 303	.else
 304	.ifc \vecnum, INT_DOUBLE_FAULT
 305	{
 306	 mfspr  r2, SPR_SYSTEM_SAVE_K_2   /* double fault info from HV */
 307	 movei  r3, 0
 308	}
 309	.else
 310	.ifc \c_routine, do_trap
 311	{
 312	 mfspr  r2, GPV_REASON
 313	 movei  r3, 0
 314	}
 315	.else
 316	.ifc \c_routine, handle_perf_interrupt
 317	{
 318	 mfspr  r2, PERF_COUNT_STS
 319	 movei  r3, -1   /* not used, but set for consistency */
 320	}
 321	.else
 322	.ifc \c_routine, handle_perf_interrupt
 
 323	{
 324	 mfspr  r2, AUX_PERF_COUNT_STS
 325	 movei  r3, -1   /* not used, but set for consistency */
 326	}
 327	.else
 
 328	movei   r3, 0
 
 329	.endif
 
 330	.endif
 331	.endif
 332	.endif
 333	.endif
 334	/* Put function pointer in r0 */
 335	moveli  r0, lo16(\c_routine)
 336	{
 337	 auli   r0, r0, ha16(\c_routine)
 338	 j       \processing
 339	}
 340	.endif
 341	ENDPROC(intvec_\vecname)
 342
 343#ifdef __COLLECT_LINKER_FEEDBACK__
 344	.pushsection .text.intvec_feedback,"ax"
 345	.org    (\vecnum << 5)
 346	FEEDBACK_ENTER_EXPLICIT(intvec_\vecname, .intrpt, 1 << 8)
 347	jrp     lr
 348	.popsection
 349#endif
 350
 351	.endm
 352
 353
 354	/*
 355	 * Save the rest of the registers that we didn't save in the actual
 356	 * vector itself.  We can't use r0-r10 inclusive here.
 357	 */
 358	.macro  finish_interrupt_save, function
 359
 360	/* If it's a syscall, save a proper orig_r0, otherwise just zero. */
 361	PTREGS_PTR(r52, PTREGS_OFFSET_ORIG_R0)
 362	{
 363	 .ifc \function,handle_syscall
 364	 sw     r52, r0
 365	 .else
 366	 sw     r52, zero
 367	 .endif
 368	 PTREGS_PTR(r52, PTREGS_OFFSET_TP)
 369	}
 370
 371	/*
 372	 * For ordinary syscalls, we save neither caller- nor callee-
 373	 * save registers, since the syscall invoker doesn't expect the
 374	 * caller-saves to be saved, and the called kernel functions will
 375	 * take care of saving the callee-saves for us.
 376	 *
 377	 * For interrupts we save just the caller-save registers.  Saving
 378	 * them is required (since the "caller" can't save them).  Again,
 379	 * the called kernel functions will restore the callee-save
 380	 * registers for us appropriately.
 381	 *
 382	 * On return, we normally restore nothing special for syscalls,
 383	 * and just the caller-save registers for interrupts.
 384	 *
 385	 * However, there are some important caveats to all this:
 386	 *
 387	 * - We always save a few callee-save registers to give us
 388	 *   some scratchpad registers to carry across function calls.
 389	 *
 390	 * - fork/vfork/etc require us to save all the callee-save
 391	 *   registers, which we do in PTREGS_SYSCALL_ALL_REGS, below.
 392	 *
 393	 * - We always save r0..r5 and r10 for syscalls, since we need
 394	 *   to reload them a bit later for the actual kernel call, and
 395	 *   since we might need them for -ERESTARTNOINTR, etc.
 396	 *
 397	 * - Before invoking a signal handler, we save the unsaved
 398	 *   callee-save registers so they are visible to the
 399	 *   signal handler or any ptracer.
 400	 *
 401	 * - If the unsaved callee-save registers are modified, we set
 402	 *   a bit in pt_regs so we know to reload them from pt_regs
 403	 *   and not just rely on the kernel function unwinding.
 404	 *   (Done for ptrace register writes and SA_SIGINFO handler.)
 405	 */
 406	{
 407	 sw     r52, tp
 408	 PTREGS_PTR(r52, PTREGS_OFFSET_REG(33))
 409	}
 410	wh64    r52    /* cache line 2 */
 411	push_reg r33, r52
 412	push_reg r32, r52
 413	push_reg r31, r52
 414	.ifc \function,handle_syscall
 415	push_reg r30, r52, PTREGS_OFFSET_SYSCALL - PTREGS_OFFSET_REG(30)
 416	push_reg TREG_SYSCALL_NR_NAME, r52, \
 417	  PTREGS_OFFSET_REG(5) - PTREGS_OFFSET_SYSCALL
 418	.else
 419
 420	push_reg r30, r52, PTREGS_OFFSET_REG(29) - PTREGS_OFFSET_REG(30)
 421	wh64    r52    /* cache line 1 */
 422	push_reg r29, r52
 423	push_reg r28, r52
 424	push_reg r27, r52
 425	push_reg r26, r52
 426	push_reg r25, r52
 427	push_reg r24, r52
 428	push_reg r23, r52
 429	push_reg r22, r52
 430	push_reg r21, r52
 431	push_reg r20, r52
 432	push_reg r19, r52
 433	push_reg r18, r52
 434	push_reg r17, r52
 435	push_reg r16, r52
 436	push_reg r15, r52
 437	push_reg r14, r52
 438	push_reg r13, r52
 439	push_reg r12, r52
 440	push_reg r11, r52
 441	push_reg r10, r52
 442	push_reg r9, r52
 443	push_reg r8, r52
 444	push_reg r7, r52
 445	push_reg r6, r52
 446
 447	.endif
 448
 449	push_reg r5, r52
 450	sw      r52, r4
 451
 452	/* Load tp with our per-cpu offset. */
 453#ifdef CONFIG_SMP
 454	{
 455	 mfspr  r20, SPR_SYSTEM_SAVE_K_0
 456	 moveli r21, lo16(__per_cpu_offset)
 457	}
 458	{
 459	 auli   r21, r21, ha16(__per_cpu_offset)
 460	 mm     r20, r20, zero, 0, LOG2_NR_CPU_IDS-1
 461	}
 462	s2a     r20, r20, r21
 463	lw      tp, r20
 464#else
 465	move    tp, zero
 466#endif
 467
 468	/*
 469	 * If we will be returning to the kernel, we will need to
 470	 * reset the interrupt masks to the state they had before.
 471	 * Set DISABLE_IRQ in flags iff we came from PL1 with irqs disabled.
 472	 * We load flags in r32 here so we can jump to .Lrestore_regs
 473	 * directly after do_page_fault_ics() if necessary.
 474	 */
 475	mfspr   r32, SPR_EX_CONTEXT_K_1
 476	{
 477	 andi   r32, r32, SPR_EX_CONTEXT_1_1__PL_MASK  /* mask off ICS */
 478	 PTREGS_PTR(r21, PTREGS_OFFSET_FLAGS)
 479	}
 480	bzt     r32, 1f       /* zero if from user space */
 481	IRQS_DISABLED(r32)    /* zero if irqs enabled */
 482#if PT_FLAGS_DISABLE_IRQ != 1
 483# error Value of IRQS_DISABLED used to set PT_FLAGS_DISABLE_IRQ; fix
 484#endif
 4851:
 486	.ifnc \function,handle_syscall
 487	/* Record the fact that we saved the caller-save registers above. */
 488	ori     r32, r32, PT_FLAGS_CALLER_SAVES
 489	.endif
 490	sw      r21, r32
 491
 492#ifdef __COLLECT_LINKER_FEEDBACK__
 493	/*
 494	 * Notify the feedback routines that we were in the
 495	 * appropriate fixed interrupt vector area.  Note that we
 496	 * still have ICS set at this point, so we can't invoke any
 497	 * atomic operations or we will panic.  The feedback
 498	 * routines internally preserve r0..r10 and r30 up.
 499	 */
 500	.ifnc \function,handle_syscall
 501	shli    r20, r1, 5
 502	.else
 503	moveli  r20, INT_SWINT_1 << 5
 504	.endif
 505	addli   r20, r20, lo16(intvec_feedback)
 506	auli    r20, r20, ha16(intvec_feedback)
 507	jalr    r20
 508
 509	/* And now notify the feedback routines that we are here. */
 510	FEEDBACK_ENTER(\function)
 511#endif
 512
 513	/*
 514	 * we've captured enough state to the stack (including in
 515	 * particular our EX_CONTEXT state) that we can now release
 516	 * the interrupt critical section and replace it with our
 517	 * standard "interrupts disabled" mask value.  This allows
 518	 * synchronous interrupts (and profile interrupts) to punch
 519	 * through from this point onwards.
 520	 *
 521	 * If bit 31 of r3 is set during a non-NMI interrupt, we know we
 522	 * are on the path where the hypervisor has punched through our
 523	 * ICS with a page fault, so we call out to do_page_fault_ics()
 524	 * to figure out what to do with it.  If the fault was in
 525	 * an atomic op, we unlock the atomic lock, adjust the
 526	 * saved register state a little, and return "zero" in r4,
 527	 * falling through into the normal page-fault interrupt code.
 528	 * If the fault was in a kernel-space atomic operation, then
 529	 * do_page_fault_ics() resolves it itself, returns "one" in r4,
 530	 * and as a result goes directly to restoring registers and iret,
 531	 * without trying to adjust the interrupt masks at all.
 532	 * The do_page_fault_ics() API involves passing and returning
 533	 * a five-word struct (in registers) to avoid writing the
 534	 * save and restore code here.
 535	 */
 536	.ifc \function,handle_nmi
 537	IRQ_DISABLE_ALL(r20)
 538	.else
 539	.ifnc \function,handle_syscall
 540	bgezt   r3, 1f
 541	{
 542	 PTREGS_PTR(r0, PTREGS_OFFSET_BASE)
 543	 jal    do_page_fault_ics
 544	}
 545	FEEDBACK_REENTER(\function)
 546	bzt     r4, 1f
 547	j       .Lrestore_regs
 5481:
 549	.endif
 550	IRQ_DISABLE(r20, r21)
 551	.endif
 552	mtspr   INTERRUPT_CRITICAL_SECTION, zero
 553
 
 554	/*
 555	 * Prepare the first 256 stack bytes to be rapidly accessible
 556	 * without having to fetch the background data.  We don't really
 557	 * know how far to write-hint, but kernel stacks generally
 558	 * aren't that big, and write-hinting here does take some time.
 559	 */
 560	addi    r52, sp, -64
 561	{
 562	 wh64   r52
 563	 addi   r52, r52, -64
 564	}
 565	{
 566	 wh64   r52
 567	 addi   r52, r52, -64
 568	}
 569	{
 570	 wh64   r52
 571	 addi   r52, r52, -64
 572	}
 573	wh64    r52
 
 574
 575#ifdef CONFIG_TRACE_IRQFLAGS
 576	.ifnc \function,handle_nmi
 577	/*
 578	 * We finally have enough state set up to notify the irq
 579	 * tracing code that irqs were disabled on entry to the handler.
 580	 * The TRACE_IRQS_OFF call clobbers registers r0-r29.
 581	 * For syscalls, we already have the register state saved away
 582	 * on the stack, so we don't bother to do any register saves here,
 583	 * and later we pop the registers back off the kernel stack.
 584	 * For interrupt handlers, save r0-r3 in callee-saved registers.
 585	 */
 586	.ifnc \function,handle_syscall
 587	{ move r30, r0; move r31, r1 }
 588	{ move r32, r2; move r33, r3 }
 589	.endif
 590	TRACE_IRQS_OFF
 591	.ifnc \function,handle_syscall
 592	{ move r0, r30; move r1, r31 }
 593	{ move r2, r32; move r3, r33 }
 594	.endif
 595	.endif
 596#endif
 597
 598	.endm
 599
 600	.macro  check_single_stepping, kind, not_single_stepping
 601	/*
 602	 * Check for single stepping in user-level priv
 603	 *   kind can be "normal", "ill", or "syscall"
 604	 * At end, if fall-thru
 605	 *   r29: thread_info->step_state
 606	 *   r28: &pt_regs->pc
 607	 *   r27: pt_regs->pc
 608	 *   r26: thread_info->step_state->buffer
 609	 */
 610
 611	/* Check for single stepping */
 612	GET_THREAD_INFO(r29)
 613	{
 614	 /* Get pointer to field holding step state */
 615	 addi   r29, r29, THREAD_INFO_STEP_STATE_OFFSET
 616
 617	 /* Get pointer to EX1 in register state */
 618	 PTREGS_PTR(r27, PTREGS_OFFSET_EX1)
 619	}
 620	{
 621	 /* Get pointer to field holding PC */
 622	 PTREGS_PTR(r28, PTREGS_OFFSET_PC)
 623
 624	 /* Load the pointer to the step state */
 625	 lw     r29, r29
 626	}
 627	/* Load EX1 */
 628	lw      r27, r27
 629	{
 630	 /* Points to flags */
 631	 addi   r23, r29, SINGLESTEP_STATE_FLAGS_OFFSET
 632
 633	 /* No single stepping if there is no step state structure */
 634	 bzt    r29, \not_single_stepping
 635	}
 636	{
 637	 /* mask off ICS and any other high bits */
 638	 andi   r27, r27, SPR_EX_CONTEXT_1_1__PL_MASK
 639
 640	 /* Load pointer to single step instruction buffer */
 641	 lw     r26, r29
 642	}
 643	/* Check priv state */
 644	bnz     r27, \not_single_stepping
 645
 646	/* Get flags */
 647	lw      r22, r23
 648	{
 649	 /* Branch if single-step mode not enabled */
 650	 bbnst  r22, \not_single_stepping
 651
 652	 /* Clear enabled flag */
 653	 andi   r22, r22, ~SINGLESTEP_STATE_MASK_IS_ENABLED
 654	}
 655	.ifc \kind,normal
 656	{
 657	 /* Load PC */
 658	 lw     r27, r28
 659
 660	 /* Point to the entry containing the original PC */
 661	 addi   r24, r29, SINGLESTEP_STATE_ORIG_PC_OFFSET
 662	}
 663	{
 664	 /* Disable single stepping flag */
 665	 sw     r23, r22
 666	}
 667	{
 668	 /* Get the original pc */
 669	 lw     r24, r24
 670
 671	 /* See if the PC is at the start of the single step buffer */
 672	 seq    r25, r26, r27
 673	}
 674	/*
 675	 * NOTE: it is really expected that the PC be in the single step buffer
 676	 *       at this point
 677	 */
 678	bzt     r25, \not_single_stepping
 679
 680	/* Restore the original PC */
 681	sw      r28, r24
 682	.else
 683	.ifc \kind,syscall
 684	{
 685	 /* Load PC */
 686	 lw     r27, r28
 687
 688	 /* Point to the entry containing the next PC */
 689	 addi   r24, r29, SINGLESTEP_STATE_NEXT_PC_OFFSET
 690	}
 691	{
 692	 /* Increment the stopped PC by the bundle size */
 693	 addi   r26, r26, 8
 694
 695	 /* Disable single stepping flag */
 696	 sw     r23, r22
 697	}
 698	{
 699	 /* Get the next pc */
 700	 lw     r24, r24
 701
 702	 /*
 703	  * See if the PC is one bundle past the start of the
 704	  * single step buffer
 705	  */
 706	 seq    r25, r26, r27
 707	}
 708	{
 709	 /*
 710	  * NOTE: it is really expected that the PC be in the
 711	  * single step buffer at this point
 712	  */
 713	 bzt    r25, \not_single_stepping
 714	}
 715	/* Set to the next PC */
 716	sw      r28, r24
 717	.else
 718	{
 719	 /* Point to 3rd bundle in buffer */
 720	 addi   r25, r26, 16
 721
 722	 /* Load PC */
 723	 lw      r27, r28
 724	}
 725	{
 726	 /* Disable single stepping flag */
 727	 sw      r23, r22
 728
 729	 /* See if the PC is in the single step buffer */
 730	 slte_u  r24, r26, r27
 731	}
 732	{
 733	 slte_u r25, r27, r25
 734
 735	 /*
 736	  * NOTE: it is really expected that the PC be in the
 737	  * single step buffer at this point
 738	  */
 739	 bzt    r24, \not_single_stepping
 740	}
 741	bzt     r25, \not_single_stepping
 742	.endif
 743	.endif
 744	.endm
 745
 746	/*
 747	 * Redispatch a downcall.
 748	 */
 749	.macro  dc_dispatch vecnum, vecname
 750	.org    (\vecnum << 8)
 751intvec_\vecname:
 752	j       _hv_downcall_dispatch
 753	ENDPROC(intvec_\vecname)
 754	.endm
 755
 756	/*
 757	 * Common code for most interrupts.  The C function we're eventually
 758	 * going to is in r0, and the faultnum is in r1; the original
 759	 * values for those registers are on the stack.
 760	 */
 761	.pushsection .text.handle_interrupt,"ax"
 762handle_interrupt:
 763	finish_interrupt_save handle_interrupt
 764
 765	/*
 766	 * Check for if we are single stepping in user level. If so, then
 767	 * we need to restore the PC.
 768	 */
 769
 770	check_single_stepping normal, .Ldispatch_interrupt
 771.Ldispatch_interrupt:
 772
 773	/* Jump to the C routine; it should enable irqs as soon as possible. */
 774	{
 775	 jalr   r0
 776	 PTREGS_PTR(r0, PTREGS_OFFSET_BASE)
 777	}
 778	FEEDBACK_REENTER(handle_interrupt)
 779	{
 780	 movei  r30, 0   /* not an NMI */
 781	 j      interrupt_return
 782	}
 783	STD_ENDPROC(handle_interrupt)
 784
 785/*
 786 * This routine takes a boolean in r30 indicating if this is an NMI.
 787 * If so, we also expect a boolean in r31 indicating whether to
 788 * re-enable the oprofile interrupts.
 789 *
 790 * Note that .Lresume_userspace is jumped to directly in several
 791 * places, and we need to make sure r30 is set correctly in those
 792 * callers as well.
 793 */
 794STD_ENTRY(interrupt_return)
 795	/* If we're resuming to kernel space, don't check thread flags. */
 796	{
 797	 bnz    r30, .Lrestore_all  /* NMIs don't special-case user-space */
 798	 PTREGS_PTR(r29, PTREGS_OFFSET_EX1)
 799	}
 800	lw      r29, r29
 801	andi    r29, r29, SPR_EX_CONTEXT_1_1__PL_MASK  /* mask off ICS */
 802	bzt     r29, .Lresume_userspace
 803
 804#ifdef CONFIG_PREEMPT
 805	/* Returning to kernel space. Check if we need preemption. */
 806	GET_THREAD_INFO(r29)
 807	addli   r28, r29, THREAD_INFO_FLAGS_OFFSET
 808	{
 809	 lw     r28, r28
 810	 addli  r29, r29, THREAD_INFO_PREEMPT_COUNT_OFFSET
 811	}
 812	{
 813	 andi   r28, r28, _TIF_NEED_RESCHED
 814	 lw     r29, r29
 815	}
 816	bzt     r28, 1f
 817	bnz     r29, 1f
 818	/* Disable interrupts explicitly for preemption. */
 819	IRQ_DISABLE(r20,r21)
 820	TRACE_IRQS_OFF
 821	jal     preempt_schedule_irq
 822	FEEDBACK_REENTER(interrupt_return)
 8231:
 824#endif
 825
 826	/* If we're resuming to _cpu_idle_nap, bump PC forward by 8. */
 827	{
 828	 PTREGS_PTR(r29, PTREGS_OFFSET_PC)
 829	 moveli r27, lo16(_cpu_idle_nap)
 830	}
 831	{
 832	 lw     r28, r29
 833	 auli   r27, r27, ha16(_cpu_idle_nap)
 834	}
 835	{
 836	 seq    r27, r27, r28
 837	}
 838	{
 839	 bbns   r27, .Lrestore_all
 840	 addi   r28, r28, 8
 841	}
 842	sw      r29, r28
 843	j       .Lrestore_all
 844
 845.Lresume_userspace:
 846	FEEDBACK_REENTER(interrupt_return)
 847
 848	/*
 849	 * Use r33 to hold whether we have already loaded the callee-saves
 850	 * into ptregs.  We don't want to do it twice in this loop, since
 851	 * then we'd clobber whatever changes are made by ptrace, etc.
 852	 * Get base of stack in r32.
 853	 */
 854	{
 855	 GET_THREAD_INFO(r32)
 856	 movei  r33, 0
 857	}
 858
 859.Lretry_work_pending:
 860	/*
 861	 * Disable interrupts so as to make sure we don't
 862	 * miss an interrupt that sets any of the thread flags (like
 863	 * need_resched or sigpending) between sampling and the iret.
 864	 * Routines like schedule() or do_signal() may re-enable
 865	 * interrupts before returning.
 866	 */
 867	IRQ_DISABLE(r20, r21)
 868	TRACE_IRQS_OFF  /* Note: clobbers registers r0-r29 */
 869
 
 
 
 870
 871	/* Check to see if there is any work to do before returning to user. */
 872	{
 873	 addi   r29, r32, THREAD_INFO_FLAGS_OFFSET
 874	 moveli r1, lo16(_TIF_ALLWORK_MASK)
 875	}
 876	{
 877	 lw     r29, r29
 878	 auli   r1, r1, ha16(_TIF_ALLWORK_MASK)
 879	}
 880	and     r1, r29, r1
 881	bzt     r1, .Lrestore_all
 882
 883	/*
 884	 * Make sure we have all the registers saved for signal
 885	 * handling, notify-resume, or single-step.  Call out to C
 886	 * code to figure out exactly what we need to do for each flag bit,
 887	 * then if necessary, reload the flags and recheck.
 888	 */
 
 889	{
 890	 PTREGS_PTR(r0, PTREGS_OFFSET_BASE)
 891	 bnz    r33, 1f
 892	}
 893	push_extra_callee_saves r0
 894	movei   r33, 1
 8951:	jal     do_work_pending
 896	bnz     r0, .Lretry_work_pending
 897
 898	/*
 899	 * In the NMI case we
 900	 * omit the call to single_process_check_nohz, which normally checks
 901	 * to see if we should start or stop the scheduler tick, because
 902	 * we can't call arbitrary Linux code from an NMI context.
 903	 * We always call the homecache TLB deferral code to re-trigger
 904	 * the deferral mechanism.
 905	 *
 906	 * The other chunk of responsibility this code has is to reset the
 907	 * interrupt masks appropriately to reset irqs and NMIs.  We have
 908	 * to call TRACE_IRQS_OFF and TRACE_IRQS_ON to support all the
 909	 * lockdep-type stuff, but we can't set ICS until afterwards, since
 910	 * ICS can only be used in very tight chunks of code to avoid
 911	 * tripping over various assertions that it is off.
 912	 *
 913	 * (There is what looks like a window of vulnerability here since
 914	 * we might take a profile interrupt between the two SPR writes
 915	 * that set the mask, but since we write the low SPR word first,
 916	 * and our interrupt entry code checks the low SPR word, any
 917	 * profile interrupt will actually disable interrupts in both SPRs
 918	 * before returning, which is OK.)
 919	 */
 920.Lrestore_all:
 921	PTREGS_PTR(r0, PTREGS_OFFSET_EX1)
 922	{
 923	 lw     r0, r0
 924	 PTREGS_PTR(r32, PTREGS_OFFSET_FLAGS)
 925	}
 926	{
 927	 andi   r0, r0, SPR_EX_CONTEXT_1_1__PL_MASK
 928	 lw     r32, r32
 929	}
 930	bnz    r0, 1f
 931	j       2f
 932#if PT_FLAGS_DISABLE_IRQ != 1
 933# error Assuming PT_FLAGS_DISABLE_IRQ == 1 so we can use bbnst below
 934#endif
 9351:	bbnst   r32, 2f
 936	IRQ_DISABLE(r20,r21)
 937	TRACE_IRQS_OFF
 938	movei   r0, 1
 939	mtspr   INTERRUPT_CRITICAL_SECTION, r0
 940	bzt     r30, .Lrestore_regs
 941	j       3f
 9422:	TRACE_IRQS_ON
 943	movei   r0, 1
 944	mtspr   INTERRUPT_CRITICAL_SECTION, r0
 945	IRQ_ENABLE(r20, r21)
 946	bzt     r30, .Lrestore_regs
 9473:
 948
 949	/* We are relying on INT_PERF_COUNT at 33, and AUX_PERF_COUNT at 48 */
 950	{
 951	 moveli r0, lo16(1 << (INT_PERF_COUNT - 32))
 952	 bz     r31, .Lrestore_regs
 953	}
 954	auli    r0, r0, ha16(1 << (INT_AUX_PERF_COUNT - 32))
 955	mtspr   SPR_INTERRUPT_MASK_RESET_K_1, r0
 956
 957	/*
 958	 * We now commit to returning from this interrupt, since we will be
 959	 * doing things like setting EX_CONTEXT SPRs and unwinding the stack
 960	 * frame.  No calls should be made to any other code after this point.
 961	 * This code should only be entered with ICS set.
 962	 * r32 must still be set to ptregs.flags.
 963	 * We launch loads to each cache line separately first, so we can
 964	 * get some parallelism out of the memory subsystem.
 965	 * We start zeroing caller-saved registers throughout, since
 966	 * that will save some cycles if this turns out to be a syscall.
 967	 */
 968.Lrestore_regs:
 969	FEEDBACK_REENTER(interrupt_return)   /* called from elsewhere */
 970
 971	/*
 972	 * Rotate so we have one high bit and one low bit to test.
 973	 * - low bit says whether to restore all the callee-saved registers,
 974	 *   or just r30-r33, and r52 up.
 975	 * - high bit (i.e. sign bit) says whether to restore all the
 976	 *   caller-saved registers, or just r0.
 977	 */
 978#if PT_FLAGS_CALLER_SAVES != 2 || PT_FLAGS_RESTORE_REGS != 4
 979# error Rotate trick does not work :-)
 980#endif
 981	{
 982	 rli    r20, r32, 30
 983	 PTREGS_PTR(sp, PTREGS_OFFSET_REG(0))
 984	}
 985
 986	/*
 987	 * Load cache lines 0, 2, and 3 in that order, then use
 988	 * the last loaded value, which makes it likely that the other
 989	 * cache lines have also loaded, at which point we should be
 990	 * able to safely read all the remaining words on those cache
 991	 * lines without waiting for the memory subsystem.
 992	 */
 993	pop_reg_zero r0, r28, sp, PTREGS_OFFSET_REG(30) - PTREGS_OFFSET_REG(0)
 994	pop_reg_zero r30, r2, sp, PTREGS_OFFSET_PC - PTREGS_OFFSET_REG(30)
 995	pop_reg_zero r21, r3, sp, PTREGS_OFFSET_EX1 - PTREGS_OFFSET_PC
 996	pop_reg_zero lr, r4, sp, PTREGS_OFFSET_REG(52) - PTREGS_OFFSET_EX1
 997	{
 998	 mtspr  SPR_EX_CONTEXT_K_0, r21
 999	 move   r5, zero
1000	}
1001	{
1002	 mtspr  SPR_EX_CONTEXT_K_1, lr
1003	 andi   lr, lr, SPR_EX_CONTEXT_1_1__PL_MASK  /* mask off ICS */
1004	}
1005
1006	/* Restore callee-saveds that we actually use. */
1007	pop_reg_zero r52, r6, sp, PTREGS_OFFSET_REG(31) - PTREGS_OFFSET_REG(52)
1008	pop_reg_zero r31, r7
1009	pop_reg_zero r32, r8
1010	pop_reg_zero r33, r9, sp, PTREGS_OFFSET_REG(29) - PTREGS_OFFSET_REG(33)
1011
1012	/*
1013	 * If we modified other callee-saveds, restore them now.
1014	 * This is rare, but could be via ptrace or signal handler.
1015	 */
1016	{
1017	 move   r10, zero
1018	 bbs    r20, .Lrestore_callees
1019	}
1020.Lcontinue_restore_regs:
1021
1022	/* Check if we're returning from a syscall. */
1023	{
1024	 move   r11, zero
1025	 blzt   r20, 1f  /* no, so go restore callee-save registers */
1026	}
1027
1028	/*
1029	 * Check if we're returning to userspace.
1030	 * Note that if we're not, we don't worry about zeroing everything.
1031	 */
1032	{
1033	 addli  sp, sp, PTREGS_OFFSET_LR - PTREGS_OFFSET_REG(29)
1034	 bnz    lr, .Lkernel_return
1035	}
1036
1037	/*
1038	 * On return from syscall, we've restored r0 from pt_regs, but we
1039	 * clear the remainder of the caller-saved registers.  We could
1040	 * restore the syscall arguments, but there's not much point,
1041	 * and it ensures user programs aren't trying to use the
1042	 * caller-saves if we clear them, as well as avoiding leaking
1043	 * kernel pointers into userspace.
1044	 */
1045	pop_reg_zero lr, r12, sp, PTREGS_OFFSET_TP - PTREGS_OFFSET_LR
1046	pop_reg_zero tp, r13, sp, PTREGS_OFFSET_SP - PTREGS_OFFSET_TP
1047	{
1048	 lw     sp, sp
1049	 move   r14, zero
1050	 move   r15, zero
1051	}
1052	{ move r16, zero; move r17, zero }
1053	{ move r18, zero; move r19, zero }
1054	{ move r20, zero; move r21, zero }
1055	{ move r22, zero; move r23, zero }
1056	{ move r24, zero; move r25, zero }
1057	{ move r26, zero; move r27, zero }
1058
1059	/* Set r1 to errno if we are returning an error, otherwise zero. */
1060	{
1061	 moveli r29, 4096
1062	 sub    r1, zero, r0
1063	}
1064	slt_u   r29, r1, r29
1065	{
1066	 mnz    r1, r29, r1
1067	 move   r29, zero
1068	}
1069	iret
1070
1071	/*
1072	 * Not a syscall, so restore caller-saved registers.
1073	 * First kick off a load for cache line 1, which we're touching
1074	 * for the first time here.
1075	 */
1076	.align 64
10771:	pop_reg r29, sp, PTREGS_OFFSET_REG(1) - PTREGS_OFFSET_REG(29)
1078	pop_reg r1
1079	pop_reg r2
1080	pop_reg r3
1081	pop_reg r4
1082	pop_reg r5
1083	pop_reg r6
1084	pop_reg r7
1085	pop_reg r8
1086	pop_reg r9
1087	pop_reg r10
1088	pop_reg r11
1089	pop_reg r12
1090	pop_reg r13
1091	pop_reg r14
1092	pop_reg r15
1093	pop_reg r16
1094	pop_reg r17
1095	pop_reg r18
1096	pop_reg r19
1097	pop_reg r20
1098	pop_reg r21
1099	pop_reg r22
1100	pop_reg r23
1101	pop_reg r24
1102	pop_reg r25
1103	pop_reg r26
1104	pop_reg r27
1105	pop_reg r28, sp, PTREGS_OFFSET_LR - PTREGS_OFFSET_REG(28)
1106	/* r29 already restored above */
1107	bnz     lr, .Lkernel_return
1108	pop_reg lr, sp, PTREGS_OFFSET_TP - PTREGS_OFFSET_LR
1109	pop_reg tp, sp, PTREGS_OFFSET_SP - PTREGS_OFFSET_TP
1110	lw      sp, sp
1111	iret
1112
1113	/*
1114	 * We can't restore tp when in kernel mode, since a thread might
1115	 * have migrated from another cpu and brought a stale tp value.
1116	 */
1117.Lkernel_return:
1118	pop_reg lr, sp, PTREGS_OFFSET_SP - PTREGS_OFFSET_LR
1119	lw      sp, sp
1120	iret
1121
1122	/* Restore callee-saved registers from r34 to r51. */
1123.Lrestore_callees:
1124	addli  sp, sp, PTREGS_OFFSET_REG(34) - PTREGS_OFFSET_REG(29)
1125	pop_reg r34
1126	pop_reg r35
1127	pop_reg r36
1128	pop_reg r37
1129	pop_reg r38
1130	pop_reg r39
1131	pop_reg r40
1132	pop_reg r41
1133	pop_reg r42
1134	pop_reg r43
1135	pop_reg r44
1136	pop_reg r45
1137	pop_reg r46
1138	pop_reg r47
1139	pop_reg r48
1140	pop_reg r49
1141	pop_reg r50
1142	pop_reg r51, sp, PTREGS_OFFSET_REG(29) - PTREGS_OFFSET_REG(51)
1143	j .Lcontinue_restore_regs
1144	STD_ENDPROC(interrupt_return)
1145
1146	/*
1147	 * Some interrupts don't check for single stepping
1148	 */
1149	.pushsection .text.handle_interrupt_no_single_step,"ax"
1150handle_interrupt_no_single_step:
1151	finish_interrupt_save handle_interrupt_no_single_step
1152	{
1153	 jalr   r0
1154	 PTREGS_PTR(r0, PTREGS_OFFSET_BASE)
1155	}
1156	FEEDBACK_REENTER(handle_interrupt_no_single_step)
1157	{
1158	 movei  r30, 0   /* not an NMI */
1159	 j      interrupt_return
1160	}
1161	STD_ENDPROC(handle_interrupt_no_single_step)
1162
1163	/*
1164	 * "NMI" interrupts mask ALL interrupts before calling the
1165	 * handler, and don't check thread flags, etc., on the way
1166	 * back out.  In general, the only things we do here for NMIs
1167	 * are the register save/restore, fixing the PC if we were
1168	 * doing single step, and the dataplane kernel-TLB management.
1169	 * We don't (for example) deal with start/stop of the sched tick.
1170	 */
1171	.pushsection .text.handle_nmi,"ax"
1172handle_nmi:
1173	finish_interrupt_save handle_nmi
1174	check_single_stepping normal, .Ldispatch_nmi
1175.Ldispatch_nmi:
1176	{
1177	 jalr   r0
1178	 PTREGS_PTR(r0, PTREGS_OFFSET_BASE)
1179	}
1180	FEEDBACK_REENTER(handle_nmi)
1181	{
1182	 movei  r30, 1
1183	 seq    r31, r0, zero
1184	}
1185	j       interrupt_return
1186	STD_ENDPROC(handle_nmi)
1187
1188	/*
1189	 * Parallel code for syscalls to handle_interrupt.
1190	 */
1191	.pushsection .text.handle_syscall,"ax"
1192handle_syscall:
1193	finish_interrupt_save handle_syscall
1194
1195	/*
1196	 * Check for if we are single stepping in user level. If so, then
1197	 * we need to restore the PC.
1198	 */
1199	check_single_stepping syscall, .Ldispatch_syscall
1200.Ldispatch_syscall:
1201
1202	/* Enable irqs. */
1203	TRACE_IRQS_ON
1204	IRQ_ENABLE(r20, r21)
1205
1206	/* Bump the counter for syscalls made on this tile. */
1207	moveli  r20, lo16(irq_stat + IRQ_CPUSTAT_SYSCALL_COUNT_OFFSET)
1208	auli    r20, r20, ha16(irq_stat + IRQ_CPUSTAT_SYSCALL_COUNT_OFFSET)
1209	add     r20, r20, tp
1210	lw      r21, r20
1211	addi    r21, r21, 1
1212	{
1213	 sw     r20, r21
1214	 GET_THREAD_INFO(r31)
1215	}
1216
1217	/* Trace syscalls, if requested. */
 
1218	addi	r31, r31, THREAD_INFO_FLAGS_OFFSET
1219	lw	r30, r31
1220	andi    r30, r30, _TIF_SYSCALL_TRACE
1221	bzt	r30, .Lrestore_syscall_regs
1222	{
1223	 PTREGS_PTR(r0, PTREGS_OFFSET_BASE)
1224	 jal    do_syscall_trace_enter
1225	}
1226	FEEDBACK_REENTER(handle_syscall)
1227
1228	/*
1229	 * We always reload our registers from the stack at this
1230	 * point.  They might be valid, if we didn't build with
1231	 * TRACE_IRQFLAGS, and this isn't a dataplane tile, and we're not
1232	 * doing syscall tracing, but there are enough cases now that it
1233	 * seems simplest just to do the reload unconditionally.
1234	 */
1235.Lrestore_syscall_regs:
1236	PTREGS_PTR(r11, PTREGS_OFFSET_REG(0))
1237	pop_reg r0, r11
1238	pop_reg r1, r11
1239	pop_reg r2, r11
1240	pop_reg r3, r11
1241	pop_reg r4, r11
1242	pop_reg r5, r11, PTREGS_OFFSET_SYSCALL - PTREGS_OFFSET_REG(5)
1243	pop_reg TREG_SYSCALL_NR_NAME, r11
1244
1245	/* Ensure that the syscall number is within the legal range. */
1246	moveli  r21, __NR_syscalls
1247	{
1248	 slt_u  r21, TREG_SYSCALL_NR_NAME, r21
1249	 moveli r20, lo16(sys_call_table)
1250	}
1251	{
1252	 bbns   r21, .Linvalid_syscall
1253	 auli   r20, r20, ha16(sys_call_table)
1254	}
1255	s2a     r20, TREG_SYSCALL_NR_NAME, r20
1256	lw      r20, r20
1257
1258	/* Jump to syscall handler. */
1259	jalr    r20
1260.Lhandle_syscall_link: /* value of "lr" after "jalr r20" above */
1261
1262	/*
1263	 * Write our r0 onto the stack so it gets restored instead
1264	 * of whatever the user had there before.
1265	 */
1266	PTREGS_PTR(r29, PTREGS_OFFSET_REG(0))
1267	sw      r29, r0
1268
1269.Lsyscall_sigreturn_skip:
1270	FEEDBACK_REENTER(handle_syscall)
1271
1272	/* Do syscall trace again, if requested. */
1273	lw	r30, r31
1274	andi    r30, r30, _TIF_SYSCALL_TRACE
1275	bzt     r30, 1f
1276	{
1277	 PTREGS_PTR(r0, PTREGS_OFFSET_BASE)
1278	 jal    do_syscall_trace_exit
1279	}
1280	FEEDBACK_REENTER(handle_syscall)
12811:	{
1282	 movei  r30, 0               /* not an NMI */
1283	 j      .Lresume_userspace   /* jump into middle of interrupt_return */
1284	}
1285
1286.Linvalid_syscall:
1287	/* Report an invalid syscall back to the user program */
1288	{
1289	 PTREGS_PTR(r29, PTREGS_OFFSET_REG(0))
1290	 movei  r28, -ENOSYS
1291	}
1292	sw      r29, r28
1293	{
1294	 movei  r30, 0               /* not an NMI */
1295	 j      .Lresume_userspace   /* jump into middle of interrupt_return */
1296	}
1297	STD_ENDPROC(handle_syscall)
1298
1299	/* Return the address for oprofile to suppress in backtraces. */
1300STD_ENTRY_SECTION(handle_syscall_link_address, .text.handle_syscall)
1301	lnk     r0
1302	{
1303	 addli  r0, r0, .Lhandle_syscall_link - .
1304	 jrp    lr
1305	}
1306	STD_ENDPROC(handle_syscall_link_address)
1307
1308STD_ENTRY(ret_from_fork)
1309	jal     sim_notify_fork
1310	jal     schedule_tail
1311	FEEDBACK_REENTER(ret_from_fork)
1312	{
1313	 movei  r30, 0               /* not an NMI */
1314	 j      .Lresume_userspace   /* jump into middle of interrupt_return */
1315	}
1316	STD_ENDPROC(ret_from_fork)
1317
1318STD_ENTRY(ret_from_kernel_thread)
1319	jal     sim_notify_fork
1320	jal     schedule_tail
1321	FEEDBACK_REENTER(ret_from_fork)
1322	{
1323	 move   r0, r31
1324	 jalr   r30
1325	}
1326	FEEDBACK_REENTER(ret_from_kernel_thread)
1327	{
1328	 movei  r30, 0               /* not an NMI */
1329	 j      .Lresume_userspace   /* jump into middle of interrupt_return */
1330	}
1331	STD_ENDPROC(ret_from_kernel_thread)
1332
1333	/*
1334	 * Code for ill interrupt.
1335	 */
1336	.pushsection .text.handle_ill,"ax"
1337handle_ill:
1338	finish_interrupt_save handle_ill
1339
1340	/*
1341	 * Check for if we are single stepping in user level. If so, then
1342	 * we need to restore the PC.
1343	 */
1344	check_single_stepping ill, .Ldispatch_normal_ill
1345
1346	{
1347	 /* See if the PC is the 1st bundle in the buffer */
1348	 seq    r25, r27, r26
1349
1350	 /* Point to the 2nd bundle in the buffer */
1351	 addi   r26, r26, 8
1352	}
1353	{
1354	 /* Point to the original pc */
1355	 addi   r24, r29, SINGLESTEP_STATE_ORIG_PC_OFFSET
1356
1357	 /* Branch if the PC is the 1st bundle in the buffer */
1358	 bnz    r25, 3f
1359	}
1360	{
1361	 /* See if the PC is the 2nd bundle of the buffer */
1362	 seq    r25, r27, r26
1363
1364	 /* Set PC to next instruction */
1365	 addi   r24, r29, SINGLESTEP_STATE_NEXT_PC_OFFSET
1366	}
1367	{
1368	 /* Point to flags */
1369	 addi   r25, r29, SINGLESTEP_STATE_FLAGS_OFFSET
1370
1371	 /* Branch if PC is in the second bundle */
1372	 bz     r25, 2f
1373	}
1374	/* Load flags */
1375	lw      r25, r25
1376	{
1377	 /*
1378	  * Get the offset for the register to restore
1379	  * Note: the lower bound is 2, so we have implicit scaling by 4.
1380	  *  No multiplication of the register number by the size of a register
1381	  *  is needed.
1382	  */
1383	 mm     r27, r25, zero, SINGLESTEP_STATE_TARGET_LB, \
1384		SINGLESTEP_STATE_TARGET_UB
1385
1386	 /* Mask Rewrite_LR */
1387	 andi   r25, r25, SINGLESTEP_STATE_MASK_UPDATE
1388	}
1389	{
1390	 addi   r29, r29, SINGLESTEP_STATE_UPDATE_VALUE_OFFSET
1391
1392	 /* Don't rewrite temp register */
1393	 bz     r25, 3f
1394	}
1395	{
1396	 /* Get the temp value */
1397	 lw     r29, r29
1398
1399	 /* Point to where the register is stored */
1400	 add    r27, r27, sp
1401	}
1402
1403	/* Add in the C ABI save area size to the register offset */
1404	addi    r27, r27, C_ABI_SAVE_AREA_SIZE
1405
1406	/* Restore the user's register with the temp value */
1407	sw      r27, r29
1408	j       3f
1409
14102:
1411	/* Must be in the third bundle */
1412	addi    r24, r29, SINGLESTEP_STATE_BRANCH_NEXT_PC_OFFSET
1413
14143:
1415	/* set PC and continue */
1416	lw      r26, r24
1417	{
1418	 sw     r28, r26
1419	 GET_THREAD_INFO(r0)
1420	}
1421
1422	/*
1423	 * Clear TIF_SINGLESTEP to prevent recursion if we execute an ill.
1424	 * The normal non-arch flow redundantly clears TIF_SINGLESTEP, but we
1425	 * need to clear it here and can't really impose on all other arches.
1426	 * So what's another write between friends?
1427	 */
 
1428
1429	addi    r1, r0, THREAD_INFO_FLAGS_OFFSET
1430	{
1431	 lw     r2, r1
1432	 addi   r0, r0, THREAD_INFO_TASK_OFFSET  /* currently a no-op */
1433	}
1434	andi    r2, r2, ~_TIF_SINGLESTEP
1435	sw      r1, r2
1436
1437	/* Issue a sigtrap */
1438	{
1439	 lw     r0, r0          /* indirect thru thread_info to get task_info*/
1440	 addi   r1, sp, C_ABI_SAVE_AREA_SIZE  /* put ptregs pointer into r1 */
 
1441	}
1442
1443	jal     send_sigtrap    /* issue a SIGTRAP */
1444	FEEDBACK_REENTER(handle_ill)
1445	{
1446	 movei  r30, 0               /* not an NMI */
1447	 j      .Lresume_userspace   /* jump into middle of interrupt_return */
1448	}
1449
1450.Ldispatch_normal_ill:
1451	{
1452	 jalr   r0
1453	 PTREGS_PTR(r0, PTREGS_OFFSET_BASE)
1454	}
1455	FEEDBACK_REENTER(handle_ill)
1456	{
1457	 movei  r30, 0   /* not an NMI */
1458	 j      interrupt_return
1459	}
1460	STD_ENDPROC(handle_ill)
1461
1462/* Various stub interrupt handlers and syscall handlers */
1463
1464STD_ENTRY_LOCAL(_kernel_double_fault)
1465	mfspr   r1, SPR_EX_CONTEXT_K_0
1466	move    r2, lr
1467	move    r3, sp
1468	move    r4, r52
1469	addi    sp, sp, -C_ABI_SAVE_AREA_SIZE
1470	j       kernel_double_fault
1471	STD_ENDPROC(_kernel_double_fault)
1472
1473STD_ENTRY_LOCAL(bad_intr)
1474	mfspr   r2, SPR_EX_CONTEXT_K_0
1475	panic   "Unhandled interrupt %#x: PC %#lx"
1476	STD_ENDPROC(bad_intr)
1477
 
 
 
 
 
 
 
 
 
1478/*
1479 * Special-case sigreturn to not write r0 to the stack on return.
1480 * This is technically more efficient, but it also avoids difficulties
1481 * in the 64-bit OS when handling 32-bit compat code, since we must not
1482 * sign-extend r0 for the sigreturn return-value case.
1483 */
1484#define PTREGS_SYSCALL_SIGRETURN(x, reg)                \
1485	STD_ENTRY(_##x);                                \
1486	addli   lr, lr, .Lsyscall_sigreturn_skip - .Lhandle_syscall_link; \
1487	{                                               \
1488	 PTREGS_PTR(reg, PTREGS_OFFSET_BASE);           \
1489	 j      x                                       \
1490	};                                              \
1491	STD_ENDPROC(_##x)
1492
 
 
1493PTREGS_SYSCALL_SIGRETURN(sys_rt_sigreturn, r0)
 
1494
1495/* Save additional callee-saves to pt_regs and jump to standard function. */
1496STD_ENTRY(_sys_clone)
1497	push_extra_callee_saves r4
1498	j       sys_clone
1499	STD_ENDPROC(_sys_clone)
1500
1501/*
1502 * This entrypoint is taken for the cmpxchg and atomic_update fast
1503 * swints.  We may wish to generalize it to other fast swints at some
1504 * point, but for now there are just two very similar ones, which
1505 * makes it faster.
1506 *
1507 * The fast swint code is designed to have a small footprint.  It does
1508 * not save or restore any GPRs, counting on the caller-save registers
1509 * to be available to it on entry.  It does not modify any callee-save
1510 * registers (including "lr").  It does not check what PL it is being
1511 * called at, so you'd better not call it other than at PL0.
1512 * The <atomic.h> wrapper assumes it only clobbers r20-r29, so if
1513 * it ever is necessary to use more registers, be aware.
1514 *
1515 * It does not use the stack, but since it might be re-interrupted by
1516 * a page fault which would assume the stack was valid, it does
1517 * save/restore the stack pointer and zero it out to make sure it gets reset.
1518 * Since we always keep interrupts disabled, the hypervisor won't
1519 * clobber our EX_CONTEXT_K_x registers, so we don't save/restore them
1520 * (other than to advance the PC on return).
1521 *
1522 * We have to manually validate the user vs kernel address range
1523 * (since at PL1 we can read/write both), and for performance reasons
1524 * we don't allow cmpxchg on the fc000000 memory region, since we only
1525 * validate that the user address is below PAGE_OFFSET.
1526 *
1527 * We place it in the __HEAD section to ensure it is relatively
1528 * near to the intvec_SWINT_1 code (reachable by a conditional branch).
1529 *
1530 * Our use of ATOMIC_LOCK_REG here must match do_page_fault_ics().
1531 *
1532 * As we do in lib/atomic_asm_32.S, we bypass a store if the value we
1533 * would store is the same as the value we just loaded.
1534 */
1535	__HEAD
1536	.align 64
1537	/* Align much later jump on the start of a cache line. */
 
1538	nop
1539#if PAGE_SIZE >= 0x10000
1540	nop
1541#endif
 
1542ENTRY(sys_cmpxchg)
1543
1544	/*
1545	 * Save "sp" and set it zero for any possible page fault.
1546	 *
1547	 * HACK: We want to both zero sp and check r0's alignment,
1548	 * so we do both at once. If "sp" becomes nonzero we
1549	 * know r0 is unaligned and branch to the error handler that
1550	 * restores sp, so this is OK.
1551	 *
1552	 * ICS is disabled right now so having a garbage but nonzero
1553	 * sp is OK, since we won't execute any faulting instructions
1554	 * when it is nonzero.
1555	 */
1556	{
1557	 move   r27, sp
1558	 andi	sp, r0, 3
1559	}
1560
1561	/*
1562	 * Get the lock address in ATOMIC_LOCK_REG, and also validate that the
1563	 * address is less than PAGE_OFFSET, since that won't trap at PL1.
1564	 * We only use bits less than PAGE_SHIFT to avoid having to worry
1565	 * about aliasing among multiple mappings of the same physical page,
1566	 * and we ignore the low 3 bits so we have one lock that covers
1567	 * both a cmpxchg64() and a cmpxchg() on either its low or high word.
1568	 * NOTE: this must match __atomic_hashed_lock() in lib/atomic_32.c.
1569	 */
1570
1571#if (PAGE_OFFSET & 0xffff) != 0
1572# error Code here assumes PAGE_OFFSET can be loaded with just hi16()
1573#endif
1574
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1575	{
1576	 /* Check for unaligned input. */
1577	 bnz    sp, .Lcmpxchg_badaddr
1578	 auli   r23, zero, hi16(PAGE_OFFSET)  /* hugepage-aligned */
1579	}
1580	{
1581	 /*
1582	  * Slide bits into position for 'mm'. We want to ignore
1583	  * the low 3 bits of r0, and consider only the next
1584	  * ATOMIC_HASH_SHIFT bits.
1585	  * Because of C pointer arithmetic, we want to compute this:
1586	  *
1587	  * ((char*)atomic_locks +
1588	  *  (((r0 >> 3) & ((1 << ATOMIC_HASH_SHIFT) - 1)) << 2))
1589	  *
1590	  * Instead of two shifts we just ">> 1", and use 'mm'
1591	  * to ignore the low and high bits we don't want.
1592	  */
1593	 shri	r25, r0, 1
1594
1595	 slt_u  r23, r0, r23
1596
1597	 /*
1598	  * Ensure that the TLB is loaded before we take out the lock.
1599	  * This will start fetching the value all the way into our L1
1600	  * as well (and if it gets modified before we grab the lock,
1601	  * it will be invalidated from our cache before we reload it).
 
 
 
1602	  */
1603	 lw	r26, r0
1604	}
1605	{
1606	 auli	r21, zero, ha16(atomic_locks)
1607
1608	 bbns   r23, .Lcmpxchg_badaddr
1609	}
1610#if PAGE_SIZE < 0x10000
1611	/* atomic_locks is page-aligned so for big pages we don't need this. */
1612	addli   r21, r21, lo16(atomic_locks)
1613#endif
1614	{
1615	 /*
1616	  * Insert the hash bits into the page-aligned pointer.
1617	  * ATOMIC_HASH_SHIFT is so big that we don't actually hash
1618	  * the unmasked address bits, as that may cause unnecessary
1619	  * collisions.
1620	  */
1621	 mm	ATOMIC_LOCK_REG_NAME, r25, r21, 2, (ATOMIC_HASH_SHIFT + 2) - 1
1622
1623	 seqi	r23, TREG_SYSCALL_NR_NAME, __NR_FAST_cmpxchg64
1624	}
1625	{
1626	 /* Branch away at this point if we're doing a 64-bit cmpxchg. */
1627	 bbs    r23, .Lcmpxchg64
1628	 andi   r23, r0, 7       /* Precompute alignment for cmpxchg64. */
1629	}
1630	{
1631	 /*
1632	  * We very carefully align the code that actually runs with
1633	  * the lock held (twelve bundles) so that we know it is all in
1634	  * the icache when we start.  This instruction (the jump) is
1635	  * at the start of the first cache line, address zero mod 64;
1636	  * we jump to the very end of the second cache line to get that
1637	  * line loaded in the icache, then fall through to issue the tns
1638	  * in the third cache line, at which point it's all cached.
1639	  * Note that is for performance, not correctness.
1640	  */
1641	 j      .Lcmpxchg32_tns
1642	}
1643
 
 
1644/* Symbol for do_page_fault_ics() to use to compare against the PC. */
1645.global __sys_cmpxchg_grab_lock
1646__sys_cmpxchg_grab_lock:
1647
1648	/*
1649	 * Perform the actual cmpxchg or atomic_update.
1650	 */
1651.Ldo_cmpxchg32:
1652	{
1653	 lw     r21, r0
1654	 seqi	r23, TREG_SYSCALL_NR_NAME, __NR_FAST_atomic_update
1655	 move	r24, r2
1656	}
1657	{
1658	 seq    r22, r21, r1     /* See if cmpxchg matches. */
1659	 and	r25, r21, r1     /* If atomic_update, compute (*mem & mask) */
1660	}
1661	{
1662	 or	r22, r22, r23    /* Skip compare branch for atomic_update. */
1663	 add	r25, r25, r2     /* Compute (*mem & mask) + addend. */
1664	}
1665	{
1666	 mvnz	r24, r23, r25    /* Use atomic_update value if appropriate. */
1667	 bbns   r22, .Lcmpxchg32_nostore
1668	}
1669	seq     r22, r24, r21    /* Are we storing the value we loaded? */
1670	bbs     r22, .Lcmpxchg32_nostore
1671	sw      r0, r24
1672
1673	/* The following instruction is the start of the second cache line. */
1674	/* Do slow mtspr here so the following "mf" waits less. */
1675	{
1676	 move   sp, r27
1677	 mtspr  SPR_EX_CONTEXT_K_0, r28
1678	}
1679	mf
1680
1681	{
1682	 move   r0, r21
1683	 sw     ATOMIC_LOCK_REG_NAME, zero
1684	}
1685	iret
1686
1687	/* Duplicated code here in the case where we don't overlap "mf" */
1688.Lcmpxchg32_nostore:
1689	{
1690	 move   r0, r21
1691	 sw     ATOMIC_LOCK_REG_NAME, zero
1692	}
1693	{
1694	 move   sp, r27
1695	 mtspr  SPR_EX_CONTEXT_K_0, r28
1696	}
1697	iret
1698
1699	/*
1700	 * The locking code is the same for 32-bit cmpxchg/atomic_update,
1701	 * and for 64-bit cmpxchg.  We provide it as a macro and put
1702	 * it into both versions.  We can't share the code literally
1703	 * since it depends on having the right branch-back address.
1704	 */
1705	.macro  cmpxchg_lock, bitwidth
1706
1707	/* Lock; if we succeed, jump back up to the read-modify-write. */
1708#ifdef CONFIG_SMP
1709	tns     r21, ATOMIC_LOCK_REG_NAME
1710#else
1711	/*
1712	 * Non-SMP preserves all the lock infrastructure, to keep the
1713	 * code simpler for the interesting (SMP) case.  However, we do
1714	 * one small optimization here and in atomic_asm.S, which is
1715	 * to fake out acquiring the actual lock in the atomic_lock table.
1716	 */
1717	movei	r21, 0
1718#endif
1719
1720	/* Issue the slow SPR here while the tns result is in flight. */
1721	mfspr   r28, SPR_EX_CONTEXT_K_0
1722
1723	{
1724	 addi   r28, r28, 8    /* return to the instruction after the swint1 */
1725	 bzt    r21, .Ldo_cmpxchg\bitwidth
1726	}
1727	/*
1728	 * The preceding instruction is the last thing that must be
1729	 * hot in the icache before we do the "tns" above.
1730	 */
1731
1732#ifdef CONFIG_SMP
1733	/*
1734	 * We failed to acquire the tns lock on our first try.  Now use
1735	 * bounded exponential backoff to retry, like __atomic_spinlock().
1736	 */
1737	{
1738	 moveli r23, 2048       /* maximum backoff time in cycles */
1739	 moveli r25, 32         /* starting backoff time in cycles */
1740	}
17411:	mfspr   r26, CYCLE_LOW  /* get start point for this backoff */
17422:	mfspr   r22, CYCLE_LOW  /* test to see if we've backed off enough */
1743	sub     r22, r22, r26
1744	slt     r22, r22, r25
1745	bbst    r22, 2b
1746	{
1747	 shli   r25, r25, 1     /* double the backoff; retry the tns */
1748	 tns    r21, ATOMIC_LOCK_REG_NAME
1749	}
1750	slt     r26, r23, r25   /* is the proposed backoff too big? */
1751	{
1752	 mvnz   r25, r26, r23
1753	 bzt    r21, .Ldo_cmpxchg\bitwidth
1754	}
1755	j       1b
1756#endif /* CONFIG_SMP */
1757	.endm
1758
1759.Lcmpxchg32_tns:
1760	/*
1761	 * This is the last instruction on the second cache line.
1762	 * The nop here loads the second line, then we fall through
1763	 * to the tns to load the third line before we take the lock.
1764	 */
1765	nop
1766	cmpxchg_lock 32
1767
1768	/*
1769	 * This code is invoked from sys_cmpxchg after most of the
1770	 * preconditions have been checked.  We still need to check
1771	 * that r0 is 8-byte aligned, since if it's not we won't
1772	 * actually be atomic.  However, ATOMIC_LOCK_REG has the atomic
1773	 * lock pointer and r27/r28 have the saved SP/PC.
1774	 * r23 is holding "r0 & 7" so we can test for alignment.
1775	 * The compare value is in r2/r3; the new value is in r4/r5.
1776	 * On return, we must put the old value in r0/r1.
1777	 */
1778	.align 64
1779.Lcmpxchg64:
1780	{
 
 
 
1781	 bzt     r23, .Lcmpxchg64_tns
1782	}
1783	j       .Lcmpxchg_badaddr
1784
1785.Ldo_cmpxchg64:
1786	{
1787	 lw     r21, r0
1788	 addi   r25, r0, 4
1789	}
1790	{
1791	 lw     r1, r25
1792	}
1793	seq     r26, r21, r2
1794	{
1795	 bz     r26, .Lcmpxchg64_mismatch
1796	 seq    r26, r1, r3
1797	}
1798	{
1799	 bz     r26, .Lcmpxchg64_mismatch
1800	}
1801	sw      r0, r4
1802	sw      r25, r5
1803
1804	/*
1805	 * The 32-bit path provides optimized "match" and "mismatch"
1806	 * iret paths, but we don't have enough bundles in this cache line
1807	 * to do that, so we just make even the "mismatch" path do an "mf".
1808	 */
1809.Lcmpxchg64_mismatch:
1810	{
1811	 move   sp, r27
1812	 mtspr  SPR_EX_CONTEXT_K_0, r28
1813	}
1814	mf
1815	{
1816	 move   r0, r21
1817	 sw     ATOMIC_LOCK_REG_NAME, zero
1818	}
1819	iret
1820
1821.Lcmpxchg64_tns:
1822	cmpxchg_lock 64
1823
1824
1825	/*
1826	 * Reset sp and revector to sys_cmpxchg_badaddr(), which will
1827	 * just raise the appropriate signal and exit.  Doing it this
1828	 * way means we don't have to duplicate the code in intvec.S's
1829	 * int_hand macro that locates the top of the stack.
1830	 */
1831.Lcmpxchg_badaddr:
1832	{
1833	 moveli TREG_SYSCALL_NR_NAME, __NR_cmpxchg_badaddr
1834	 move   sp, r27
1835	}
1836	j       intvec_SWINT_1
1837	ENDPROC(sys_cmpxchg)
1838	ENTRY(__sys_cmpxchg_end)
1839
1840
1841/* The single-step support may need to read all the registers. */
1842int_unalign:
1843	push_extra_callee_saves r0
1844	j       do_trap
1845
1846/* Include .intrpt array of interrupt vectors */
1847	.section ".intrpt", "ax"
1848
1849#ifndef CONFIG_USE_PMC
1850#define handle_perf_interrupt bad_intr
1851#endif
1852
1853#ifndef CONFIG_HARDWALL
1854#define do_hardwall_trap bad_intr
1855#endif
1856
1857	int_hand     INT_ITLB_MISS, ITLB_MISS, \
1858		     do_page_fault, handle_interrupt_no_single_step
1859	int_hand     INT_MEM_ERROR, MEM_ERROR, bad_intr
1860	int_hand     INT_ILL, ILL, do_trap, handle_ill
1861	int_hand     INT_GPV, GPV, do_trap
1862	int_hand     INT_SN_ACCESS, SN_ACCESS, do_trap
1863	int_hand     INT_IDN_ACCESS, IDN_ACCESS, do_trap
1864	int_hand     INT_UDN_ACCESS, UDN_ACCESS, do_trap
1865	int_hand     INT_IDN_REFILL, IDN_REFILL, bad_intr
1866	int_hand     INT_UDN_REFILL, UDN_REFILL, bad_intr
1867	int_hand     INT_IDN_COMPLETE, IDN_COMPLETE, bad_intr
1868	int_hand     INT_UDN_COMPLETE, UDN_COMPLETE, bad_intr
1869	int_hand     INT_SWINT_3, SWINT_3, do_trap
1870	int_hand     INT_SWINT_2, SWINT_2, do_trap
1871	int_hand     INT_SWINT_1, SWINT_1, SYSCALL, handle_syscall
1872	int_hand     INT_SWINT_0, SWINT_0, do_trap
1873	int_hand     INT_UNALIGN_DATA, UNALIGN_DATA, int_unalign
1874	int_hand     INT_DTLB_MISS, DTLB_MISS, do_page_fault
1875	int_hand     INT_DTLB_ACCESS, DTLB_ACCESS, do_page_fault
1876	int_hand     INT_DMATLB_MISS, DMATLB_MISS, do_page_fault
1877	int_hand     INT_DMATLB_ACCESS, DMATLB_ACCESS, do_page_fault
1878	int_hand     INT_SNITLB_MISS, SNITLB_MISS, do_page_fault
1879	int_hand     INT_SN_NOTIFY, SN_NOTIFY, bad_intr
1880	int_hand     INT_SN_FIREWALL, SN_FIREWALL, do_hardwall_trap
1881	int_hand     INT_IDN_FIREWALL, IDN_FIREWALL, bad_intr
1882	int_hand     INT_UDN_FIREWALL, UDN_FIREWALL, do_hardwall_trap
1883	int_hand     INT_TILE_TIMER, TILE_TIMER, do_timer_interrupt
1884	int_hand     INT_IDN_TIMER, IDN_TIMER, bad_intr
1885	int_hand     INT_UDN_TIMER, UDN_TIMER, bad_intr
1886	int_hand     INT_DMA_NOTIFY, DMA_NOTIFY, bad_intr
1887	int_hand     INT_IDN_CA, IDN_CA, bad_intr
1888	int_hand     INT_UDN_CA, UDN_CA, bad_intr
1889	int_hand     INT_IDN_AVAIL, IDN_AVAIL, bad_intr
1890	int_hand     INT_UDN_AVAIL, UDN_AVAIL, bad_intr
1891	int_hand     INT_PERF_COUNT, PERF_COUNT, \
1892		     handle_perf_interrupt, handle_nmi
1893	int_hand     INT_INTCTRL_3, INTCTRL_3, bad_intr
1894#if CONFIG_KERNEL_PL == 2
1895	dc_dispatch  INT_INTCTRL_2, INTCTRL_2
1896	int_hand     INT_INTCTRL_1, INTCTRL_1, bad_intr
1897#else
1898	int_hand     INT_INTCTRL_2, INTCTRL_2, bad_intr
1899	dc_dispatch  INT_INTCTRL_1, INTCTRL_1
1900#endif
1901	int_hand     INT_INTCTRL_0, INTCTRL_0, bad_intr
1902	int_hand     INT_MESSAGE_RCV_DWNCL, MESSAGE_RCV_DWNCL, \
1903		     hv_message_intr
1904	int_hand     INT_DEV_INTR_DWNCL, DEV_INTR_DWNCL, \
1905		     tile_dev_intr
1906	int_hand     INT_I_ASID, I_ASID, bad_intr
1907	int_hand     INT_D_ASID, D_ASID, bad_intr
1908	int_hand     INT_DMATLB_MISS_DWNCL, DMATLB_MISS_DWNCL, \
1909		     do_page_fault
1910	int_hand     INT_SNITLB_MISS_DWNCL, SNITLB_MISS_DWNCL, \
1911		     do_page_fault
1912	int_hand     INT_DMATLB_ACCESS_DWNCL, DMATLB_ACCESS_DWNCL, \
1913		     do_page_fault
1914	int_hand     INT_SN_CPL, SN_CPL, bad_intr
1915	int_hand     INT_DOUBLE_FAULT, DOUBLE_FAULT, do_trap
 
1916	int_hand     INT_AUX_PERF_COUNT, AUX_PERF_COUNT, \
1917		     handle_perf_interrupt, handle_nmi
 
1918
1919	/* Synthetic interrupt delivered only by the simulator */
1920	int_hand     INT_BREAKPOINT, BREAKPOINT, do_breakpoint