1/* 2 * Linux/PA-RISC Project (http://www.parisc-linux.org/) 3 * 4 * System call entry code / Linux gateway page 5 * Copyright (c) Matthew Wilcox 1999 <willy@bofh.ai> 6 * Licensed under the GNU GPL. 7 * thanks to Philipp Rumpf, Mike Shaver and various others 8 * sorry about the wall, puffin.. 9 */ 10 11/* 12How does the Linux gateway page on PA-RISC work? 13------------------------------------------------ 14The Linux gateway page on PA-RISC is "special". 15It actually has PAGE_GATEWAY bits set (this is linux terminology; in parisc 16terminology it's Execute, promote to PL0) in the page map. So anything 17executing on this page executes with kernel level privilege (there's more to it 18than that: to have this happen, you also have to use a branch with a ,gate 19completer to activate the privilege promotion). The upshot is that everything 20that runs on the gateway page runs at kernel privilege but with the current 21user process address space (although you have access to kernel space via %sr2). 22For the 0x100 syscall entry, we redo the space registers to point to the kernel 23address space (preserving the user address space in %sr3), move to wide mode if 24required, save the user registers and branch into the kernel syscall entry 25point. For all the other functions, we execute at kernel privilege but don't 26flip address spaces. The basic upshot of this is that these code snippets are 27executed atomically (because the kernel can't be pre-empted) and they may 28perform architecturally forbidden (to PL3) operations (like setting control 29registers). 30*/ 31 32 33#include <asm/asm-offsets.h> 34#include <asm/unistd.h> 35#include <asm/errno.h> 36#include <asm/page.h> 37#include <asm/psw.h> 38#include <asm/thread_info.h> 39#include <asm/assembly.h> 40#include <asm/processor.h> 41#include <asm/cache.h> 42 43#include <linux/linkage.h> 44 45 /* We fill the empty parts of the gateway page with 46 * something that will kill the kernel or a 47 * userspace application. 48 */ 49#define KILL_INSN break 0,0 50 51 .level LEVEL 52 53 .text 54 55 .import syscall_exit,code 56 .import syscall_exit_rfi,code 57 58 /* Linux gateway page is aliased to virtual page 0 in the kernel 59 * address space. Since it is a gateway page it cannot be 60 * dereferenced, so null pointers will still fault. We start 61 * the actual entry point at 0x100. We put break instructions 62 * at the beginning of the page to trap null indirect function 63 * pointers. 64 */ 65 66 .align PAGE_SIZE 67ENTRY(linux_gateway_page) 68 69 /* ADDRESS 0x00 to 0xb0 = 176 bytes / 4 bytes per insn = 44 insns */ 70 .rept 44 71 KILL_INSN 72 .endr 73 74 /* ADDRESS 0xb0 to 0xb8, lws uses two insns for entry */ 75 /* Light-weight-syscall entry must always be located at 0xb0 */ 76 /* WARNING: Keep this number updated with table size changes */ 77#define __NR_lws_entries (3) 78 79lws_entry: 80 gate lws_start, %r0 /* increase privilege */ 81 depi 3, 31, 2, %r31 /* Ensure we return into user mode. */ 82 83 /* Fill from 0xb8 to 0xe0 */ 84 .rept 10 85 KILL_INSN 86 .endr 87 88 /* This function MUST be located at 0xe0 for glibc's threading 89 mechanism to work. DO NOT MOVE THIS CODE EVER! */ 90set_thread_pointer: 91 gate .+8, %r0 /* increase privilege */ 92 depi 3, 31, 2, %r31 /* Ensure we return into user mode. */ 93 be 0(%sr7,%r31) /* return to user space */ 94 mtctl %r26, %cr27 /* move arg0 to the control register */ 95 96 /* Increase the chance of trapping if random jumps occur to this 97 address, fill from 0xf0 to 0x100 */ 98 .rept 4 99 KILL_INSN 100 .endr 101 102/* This address must remain fixed at 0x100 for glibc's syscalls to work */ 103 .align 256 104linux_gateway_entry: 105 gate .+8, %r0 /* become privileged */ 106 mtsp %r0,%sr4 /* get kernel space into sr4 */ 107 mtsp %r0,%sr5 /* get kernel space into sr5 */ 108 mtsp %r0,%sr6 /* get kernel space into sr6 */ 109 mfsp %sr7,%r1 /* save user sr7 */ 110 mtsp %r1,%sr3 /* and store it in sr3 */ 111 112#ifdef CONFIG_64BIT 113 /* for now we can *always* set the W bit on entry to the syscall 114 * since we don't support wide userland processes. We could 115 * also save the current SM other than in r0 and restore it on 116 * exit from the syscall, and also use that value to know 117 * whether to do narrow or wide syscalls. -PB 118 */ 119 ssm PSW_SM_W, %r1 120 extrd,u %r1,PSW_W_BIT,1,%r1 121 /* sp must be aligned on 4, so deposit the W bit setting into 122 * the bottom of sp temporarily */ 123 or,ev %r1,%r30,%r30 124 b,n 1f 125 /* The top halves of argument registers must be cleared on syscall 126 * entry from narrow executable. 127 */ 128 depdi 0, 31, 32, %r26 129 depdi 0, 31, 32, %r25 130 depdi 0, 31, 32, %r24 131 depdi 0, 31, 32, %r23 132 depdi 0, 31, 32, %r22 133 depdi 0, 31, 32, %r21 1341: 135#endif 136 mfctl %cr30,%r1 137 xor %r1,%r30,%r30 /* ye olde xor trick */ 138 xor %r1,%r30,%r1 139 xor %r1,%r30,%r30 140 141 ldo THREAD_SZ_ALGN+FRAME_SIZE(%r30),%r30 /* set up kernel stack */ 142 143 /* N.B.: It is critical that we don't set sr7 to 0 until r30 144 * contains a valid kernel stack pointer. It is also 145 * critical that we don't start using the kernel stack 146 * until after sr7 has been set to 0. 147 */ 148 149 mtsp %r0,%sr7 /* get kernel space into sr7 */ 150 STREGM %r1,FRAME_SIZE(%r30) /* save r1 (usp) here for now */ 151 mfctl %cr30,%r1 /* get task ptr in %r1 */ 152 LDREG TI_TASK(%r1),%r1 153 154 /* Save some registers for sigcontext and potential task 155 switch (see entry.S for the details of which ones are 156 saved/restored). TASK_PT_PSW is zeroed so we can see whether 157 a process is on a syscall or not. For an interrupt the real 158 PSW value is stored. This is needed for gdb and sys_ptrace. */ 159 STREG %r0, TASK_PT_PSW(%r1) 160 STREG %r2, TASK_PT_GR2(%r1) /* preserve rp */ 161 STREG %r19, TASK_PT_GR19(%r1) 162 163 LDREGM -FRAME_SIZE(%r30), %r2 /* get users sp back */ 164#ifdef CONFIG_64BIT 165 extrd,u %r2,63,1,%r19 /* W hidden in bottom bit */ 166#if 0 167 xor %r19,%r2,%r2 /* clear bottom bit */ 168 depd,z %r19,1,1,%r19 169 std %r19,TASK_PT_PSW(%r1) 170#endif 171#endif 172 STREG %r2, TASK_PT_GR30(%r1) /* ... and save it */ 173 174 STREG %r20, TASK_PT_GR20(%r1) /* Syscall number */ 175 STREG %r21, TASK_PT_GR21(%r1) 176 STREG %r22, TASK_PT_GR22(%r1) 177 STREG %r23, TASK_PT_GR23(%r1) /* 4th argument */ 178 STREG %r24, TASK_PT_GR24(%r1) /* 3rd argument */ 179 STREG %r25, TASK_PT_GR25(%r1) /* 2nd argument */ 180 STREG %r26, TASK_PT_GR26(%r1) /* 1st argument */ 181 STREG %r27, TASK_PT_GR27(%r1) /* user dp */ 182 STREG %r28, TASK_PT_GR28(%r1) /* return value 0 */ 183 STREG %r0, TASK_PT_ORIG_R28(%r1) /* don't prohibit restarts */ 184 STREG %r29, TASK_PT_GR29(%r1) /* return value 1 */ 185 STREG %r31, TASK_PT_GR31(%r1) /* preserve syscall return ptr */ 186 187 ldo TASK_PT_FR0(%r1), %r27 /* save fpregs from the kernel */ 188 save_fp %r27 /* or potential task switch */ 189 190 mfctl %cr11, %r27 /* i.e. SAR */ 191 STREG %r27, TASK_PT_SAR(%r1) 192 193 loadgp 194 195#ifdef CONFIG_64BIT 196 ldo -16(%r30),%r29 /* Reference param save area */ 197 copy %r19,%r2 /* W bit back to r2 */ 198#else 199 /* no need to save these on stack in wide mode because the first 8 200 * args are passed in registers */ 201 stw %r22, -52(%r30) /* 5th argument */ 202 stw %r21, -56(%r30) /* 6th argument */ 203#endif 204 205 /* Are we being ptraced? */ 206 mfctl %cr30, %r1 207 LDREG TI_FLAGS(%r1),%r1 208 ldi _TIF_SYSCALL_TRACE_MASK, %r19 209 and,COND(=) %r1, %r19, %r0 210 b,n .Ltracesys 211 212 /* Note! We cannot use the syscall table that is mapped 213 nearby since the gateway page is mapped execute-only. */ 214 215#ifdef CONFIG_64BIT 216 ldil L%sys_call_table, %r1 217 or,= %r2,%r2,%r2 218 addil L%(sys_call_table64-sys_call_table), %r1 219 ldo R%sys_call_table(%r1), %r19 220 or,= %r2,%r2,%r2 221 ldo R%sys_call_table64(%r1), %r19 222#else 223 ldil L%sys_call_table, %r1 224 ldo R%sys_call_table(%r1), %r19 225#endif 226 comiclr,>> __NR_Linux_syscalls, %r20, %r0 227 b,n .Lsyscall_nosys 228 229 LDREGX %r20(%r19), %r19 230 231 /* If this is a sys_rt_sigreturn call, and the signal was received 232 * when not in_syscall, then we want to return via syscall_exit_rfi, 233 * not syscall_exit. Signal no. in r20, in_syscall in r25 (see 234 * trampoline code in signal.c). 235 */ 236 ldi __NR_rt_sigreturn,%r2 237 comb,= %r2,%r20,.Lrt_sigreturn 238.Lin_syscall: 239 ldil L%syscall_exit,%r2 240 be 0(%sr7,%r19) 241 ldo R%syscall_exit(%r2),%r2 242.Lrt_sigreturn: 243 comib,<> 0,%r25,.Lin_syscall 244 ldil L%syscall_exit_rfi,%r2 245 be 0(%sr7,%r19) 246 ldo R%syscall_exit_rfi(%r2),%r2 247 248 /* Note! Because we are not running where we were linked, any 249 calls to functions external to this file must be indirect. To 250 be safe, we apply the opposite rule to functions within this 251 file, with local labels given to them to ensure correctness. */ 252 253.Lsyscall_nosys: 254syscall_nosys: 255 ldil L%syscall_exit,%r1 256 be R%syscall_exit(%sr7,%r1) 257 ldo -ENOSYS(%r0),%r28 /* set errno */ 258 259 260/* Warning! This trace code is a virtual duplicate of the code above so be 261 * sure to maintain both! */ 262.Ltracesys: 263tracesys: 264 /* Need to save more registers so the debugger can see where we 265 * are. This saves only the lower 8 bits of PSW, so that the C 266 * bit is still clear on syscalls, and the D bit is set if this 267 * full register save path has been executed. We check the D 268 * bit on syscall_return_rfi to determine which registers to 269 * restore. An interrupt results in a full PSW saved with the 270 * C bit set, a non-straced syscall entry results in C and D clear 271 * in the saved PSW. 272 */ 273 ldo -THREAD_SZ_ALGN-FRAME_SIZE(%r30),%r1 /* get task ptr */ 274 LDREG TI_TASK(%r1), %r1 275 ssm 0,%r2 276 STREG %r2,TASK_PT_PSW(%r1) /* Lower 8 bits only!! */ 277 mfsp %sr0,%r2 278 STREG %r2,TASK_PT_SR0(%r1) 279 mfsp %sr1,%r2 280 STREG %r2,TASK_PT_SR1(%r1) 281 mfsp %sr2,%r2 282 STREG %r2,TASK_PT_SR2(%r1) 283 mfsp %sr3,%r2 284 STREG %r2,TASK_PT_SR3(%r1) 285 STREG %r2,TASK_PT_SR4(%r1) 286 STREG %r2,TASK_PT_SR5(%r1) 287 STREG %r2,TASK_PT_SR6(%r1) 288 STREG %r2,TASK_PT_SR7(%r1) 289 STREG %r2,TASK_PT_IASQ0(%r1) 290 STREG %r2,TASK_PT_IASQ1(%r1) 291 LDREG TASK_PT_GR31(%r1),%r2 292 STREG %r2,TASK_PT_IAOQ0(%r1) 293 ldo 4(%r2),%r2 294 STREG %r2,TASK_PT_IAOQ1(%r1) 295 ldo TASK_REGS(%r1),%r2 296 /* reg_save %r2 */ 297 STREG %r3,PT_GR3(%r2) 298 STREG %r4,PT_GR4(%r2) 299 STREG %r5,PT_GR5(%r2) 300 STREG %r6,PT_GR6(%r2) 301 STREG %r7,PT_GR7(%r2) 302 STREG %r8,PT_GR8(%r2) 303 STREG %r9,PT_GR9(%r2) 304 STREG %r10,PT_GR10(%r2) 305 STREG %r11,PT_GR11(%r2) 306 STREG %r12,PT_GR12(%r2) 307 STREG %r13,PT_GR13(%r2) 308 STREG %r14,PT_GR14(%r2) 309 STREG %r15,PT_GR15(%r2) 310 STREG %r16,PT_GR16(%r2) 311 STREG %r17,PT_GR17(%r2) 312 STREG %r18,PT_GR18(%r2) 313 /* Finished saving things for the debugger */ 314 315 copy %r2,%r26 316 ldil L%do_syscall_trace_enter,%r1 317 ldil L%tracesys_next,%r2 318 be R%do_syscall_trace_enter(%sr7,%r1) 319 ldo R%tracesys_next(%r2),%r2 320 321tracesys_next: 322 /* do_syscall_trace_enter either returned the syscallno, or -1L, 323 * so we skip restoring the PT_GR20 below, since we pulled it from 324 * task->thread.regs.gr[20] above. 325 */ 326 copy %ret0,%r20 327 ldil L%sys_call_table,%r1 328 ldo R%sys_call_table(%r1), %r19 329 330 ldo -THREAD_SZ_ALGN-FRAME_SIZE(%r30),%r1 /* get task ptr */ 331 LDREG TI_TASK(%r1), %r1 332 LDREG TASK_PT_GR26(%r1), %r26 /* Restore the users args */ 333 LDREG TASK_PT_GR25(%r1), %r25 334 LDREG TASK_PT_GR24(%r1), %r24 335 LDREG TASK_PT_GR23(%r1), %r23 336 LDREG TASK_PT_GR22(%r1), %r22 337 LDREG TASK_PT_GR21(%r1), %r21 338#ifdef CONFIG_64BIT 339 ldo -16(%r30),%r29 /* Reference param save area */ 340#else 341 stw %r22, -52(%r30) /* 5th argument */ 342 stw %r21, -56(%r30) /* 6th argument */ 343#endif 344 345 comiclr,>>= __NR_Linux_syscalls, %r20, %r0 346 b,n .Ltracesys_nosys 347 348 LDREGX %r20(%r19), %r19 349 350 /* If this is a sys_rt_sigreturn call, and the signal was received 351 * when not in_syscall, then we want to return via syscall_exit_rfi, 352 * not syscall_exit. Signal no. in r20, in_syscall in r25 (see 353 * trampoline code in signal.c). 354 */ 355 ldi __NR_rt_sigreturn,%r2 356 comb,= %r2,%r20,.Ltrace_rt_sigreturn 357.Ltrace_in_syscall: 358 ldil L%tracesys_exit,%r2 359 be 0(%sr7,%r19) 360 ldo R%tracesys_exit(%r2),%r2 361 362.Ltracesys_nosys: 363 ldo -ENOSYS(%r0),%r28 /* set errno */ 364 365 /* Do *not* call this function on the gateway page, because it 366 makes a direct call to syscall_trace. */ 367 368tracesys_exit: 369 ldo -THREAD_SZ_ALGN-FRAME_SIZE(%r30),%r1 /* get task ptr */ 370 LDREG TI_TASK(%r1), %r1 371#ifdef CONFIG_64BIT 372 ldo -16(%r30),%r29 /* Reference param save area */ 373#endif 374 ldo TASK_REGS(%r1),%r26 375 BL do_syscall_trace_exit,%r2 376 STREG %r28,TASK_PT_GR28(%r1) /* save return value now */ 377 ldo -THREAD_SZ_ALGN-FRAME_SIZE(%r30),%r1 /* get task ptr */ 378 LDREG TI_TASK(%r1), %r1 379 LDREG TASK_PT_GR28(%r1), %r28 /* Restore return val. */ 380 381 ldil L%syscall_exit,%r1 382 be,n R%syscall_exit(%sr7,%r1) 383 384.Ltrace_rt_sigreturn: 385 comib,<> 0,%r25,.Ltrace_in_syscall 386 ldil L%tracesys_sigexit,%r2 387 be 0(%sr7,%r19) 388 ldo R%tracesys_sigexit(%r2),%r2 389 390tracesys_sigexit: 391 ldo -THREAD_SZ_ALGN-FRAME_SIZE(%r30),%r1 /* get task ptr */ 392 LDREG TI_TASK(%r1), %r1 393#ifdef CONFIG_64BIT 394 ldo -16(%r30),%r29 /* Reference param save area */ 395#endif 396 BL do_syscall_trace_exit,%r2 397 ldo TASK_REGS(%r1),%r26 398 399 ldil L%syscall_exit_rfi,%r1 400 be,n R%syscall_exit_rfi(%sr7,%r1) 401 402 403 /********************************************************* 404 32/64-bit Light-Weight-Syscall ABI 405 406 * - Indicates a hint for userspace inline asm 407 implementations. 408 409 Syscall number (caller-saves) 410 - %r20 411 * In asm clobber. 412 413 Argument registers (caller-saves) 414 - %r26, %r25, %r24, %r23, %r22 415 * In asm input. 416 417 Return registers (caller-saves) 418 - %r28 (return), %r21 (errno) 419 * In asm output. 420 421 Caller-saves registers 422 - %r1, %r27, %r29 423 - %r2 (return pointer) 424 - %r31 (ble link register) 425 * In asm clobber. 426 427 Callee-saves registers 428 - %r3-%r18 429 - %r30 (stack pointer) 430 * Not in asm clobber. 431 432 If userspace is 32-bit: 433 Callee-saves registers 434 - %r19 (32-bit PIC register) 435 436 Differences from 32-bit calling convention: 437 - Syscall number in %r20 438 - Additional argument register %r22 (arg4) 439 - Callee-saves %r19. 440 441 If userspace is 64-bit: 442 Callee-saves registers 443 - %r27 (64-bit PIC register) 444 445 Differences from 64-bit calling convention: 446 - Syscall number in %r20 447 - Additional argument register %r22 (arg4) 448 - Callee-saves %r27. 449 450 Error codes returned by entry path: 451 452 ENOSYS - r20 was an invalid LWS number. 453 454 *********************************************************/ 455lws_start: 456 457#ifdef CONFIG_64BIT 458 /* FIXME: If we are a 64-bit kernel just 459 * turn this on unconditionally. 460 */ 461 ssm PSW_SM_W, %r1 462 extrd,u %r1,PSW_W_BIT,1,%r1 463 /* sp must be aligned on 4, so deposit the W bit setting into 464 * the bottom of sp temporarily */ 465 or,ev %r1,%r30,%r30 466 467 /* Clip LWS number to a 32-bit value always */ 468 depdi 0, 31, 32, %r20 469#endif 470 471 /* Is the lws entry number valid? */ 472 comiclr,>> __NR_lws_entries, %r20, %r0 473 b,n lws_exit_nosys 474 475 /* WARNING: Trashing sr2 and sr3 */ 476 mfsp %sr7,%r1 /* get userspace into sr3 */ 477 mtsp %r1,%sr3 478 mtsp %r0,%sr2 /* get kernel space into sr2 */ 479 480 /* Load table start */ 481 ldil L%lws_table, %r1 482 ldo R%lws_table(%r1), %r28 /* Scratch use of r28 */ 483 LDREGX %r20(%sr2,r28), %r21 /* Scratch use of r21 */ 484 485 /* Jump to lws, lws table pointers already relocated */ 486 be,n 0(%sr2,%r21) 487 488lws_exit_nosys: 489 ldo -ENOSYS(%r0),%r21 /* set errno */ 490 /* Fall through: Return to userspace */ 491 492lws_exit: 493#ifdef CONFIG_64BIT 494 /* decide whether to reset the wide mode bit 495 * 496 * For a syscall, the W bit is stored in the lowest bit 497 * of sp. Extract it and reset W if it is zero */ 498 extrd,u,*<> %r30,63,1,%r1 499 rsm PSW_SM_W, %r0 500 /* now reset the lowest bit of sp if it was set */ 501 xor %r30,%r1,%r30 502#endif 503 be,n 0(%sr7, %r31) 504 505 506 507 /*************************************************** 508 Implementing 32bit CAS as an atomic operation: 509 510 %r26 - Address to examine 511 %r25 - Old value to check (old) 512 %r24 - New value to set (new) 513 %r28 - Return prev through this register. 514 %r21 - Kernel error code 515 516 If debugging is DISabled: 517 518 %r21 has the following meanings: 519 520 EAGAIN - CAS is busy, ldcw failed, try again. 521 EFAULT - Read or write failed. 522 523 If debugging is enabled: 524 525 EDEADLOCK - CAS called recursively. 526 EAGAIN && r28 == 1 - CAS is busy. Lock contended. 527 EAGAIN && r28 == 2 - CAS is busy. ldcw failed. 528 EFAULT - Read or write failed. 529 530 Scratch: r20, r28, r1 531 532 ****************************************************/ 533 534 /* Do not enable LWS debugging */ 535#define ENABLE_LWS_DEBUG 0 536 537 /* ELF64 Process entry path */ 538lws_compare_and_swap64: 539#ifdef CONFIG_64BIT 540 b,n lws_compare_and_swap 541#else 542 /* If we are not a 64-bit kernel, then we don't 543 * have 64-bit input registers, and calling 544 * the 64-bit LWS CAS returns ENOSYS. 545 */ 546 b,n lws_exit_nosys 547#endif 548 549 /* ELF32 Process entry path */ 550lws_compare_and_swap32: 551#ifdef CONFIG_64BIT 552 /* Clip all the input registers */ 553 depdi 0, 31, 32, %r26 554 depdi 0, 31, 32, %r25 555 depdi 0, 31, 32, %r24 556#endif 557 558lws_compare_and_swap: 559 /* Load start of lock table */ 560 ldil L%lws_lock_start, %r20 561 ldo R%lws_lock_start(%r20), %r28 562 563 /* Extract four bits from r26 and hash lock (Bits 4-7) */ 564 extru %r26, 27, 4, %r20 565 566 /* Find lock to use, the hash is either one of 0 to 567 15, multiplied by 16 (keep it 16-byte aligned) 568 and add to the lock table offset. */ 569 shlw %r20, 4, %r20 570 add %r20, %r28, %r20 571 572# if ENABLE_LWS_DEBUG 573 /* 574 DEBUG, check for deadlock! 575 If the thread register values are the same 576 then we were the one that locked it last and 577 this is a recurisve call that will deadlock. 578 We *must* giveup this call and fail. 579 */ 580 ldw 4(%sr2,%r20), %r28 /* Load thread register */ 581 /* WARNING: If cr27 cycles to the same value we have problems */ 582 mfctl %cr27, %r21 /* Get current thread register */ 583 cmpb,<>,n %r21, %r28, cas_lock /* Called recursive? */ 584 b lws_exit /* Return error! */ 585 ldo -EDEADLOCK(%r0), %r21 586cas_lock: 587 cmpb,=,n %r0, %r28, cas_nocontend /* Is nobody using it? */ 588 ldo 1(%r0), %r28 /* 1st case */ 589 b lws_exit /* Contended... */ 590 ldo -EAGAIN(%r0), %r21 /* Spin in userspace */ 591cas_nocontend: 592# endif 593/* ENABLE_LWS_DEBUG */ 594 595 rsm PSW_SM_I, %r0 /* Disable interrupts */ 596 /* COW breaks can cause contention on UP systems */ 597 LDCW 0(%sr2,%r20), %r28 /* Try to acquire the lock */ 598 cmpb,<>,n %r0, %r28, cas_action /* Did we get it? */ 599cas_wouldblock: 600 ldo 2(%r0), %r28 /* 2nd case */ 601 ssm PSW_SM_I, %r0 602 b lws_exit /* Contended... */ 603 ldo -EAGAIN(%r0), %r21 /* Spin in userspace */ 604 605 /* 606 prev = *addr; 607 if ( prev == old ) 608 *addr = new; 609 return prev; 610 */ 611 612 /* NOTES: 613 This all works becuse intr_do_signal 614 and schedule both check the return iasq 615 and see that we are on the kernel page 616 so this process is never scheduled off 617 or is ever sent any signal of any sort, 618 thus it is wholly atomic from usrspaces 619 perspective 620 */ 621cas_action: 622#if defined CONFIG_SMP && ENABLE_LWS_DEBUG 623 /* DEBUG */ 624 mfctl %cr27, %r1 625 stw %r1, 4(%sr2,%r20) 626#endif 627 /* The load and store could fail */ 6281: ldw,ma 0(%sr3,%r26), %r28 629 sub,<> %r28, %r25, %r0 6302: stw,ma %r24, 0(%sr3,%r26) 631 /* Free lock */ 632 stw,ma %r20, 0(%sr2,%r20) 633#if ENABLE_LWS_DEBUG 634 /* Clear thread register indicator */ 635 stw %r0, 4(%sr2,%r20) 636#endif 637 /* Enable interrupts */ 638 ssm PSW_SM_I, %r0 639 /* Return to userspace, set no error */ 640 b lws_exit 641 copy %r0, %r21 642 6433: 644 /* Error occurred on load or store */ 645 /* Free lock */ 646 stw %r20, 0(%sr2,%r20) 647#if ENABLE_LWS_DEBUG 648 stw %r0, 4(%sr2,%r20) 649#endif 650 ssm PSW_SM_I, %r0 651 b lws_exit 652 ldo -EFAULT(%r0),%r21 /* set errno */ 653 nop 654 nop 655 nop 656 nop 657 658 /* Two exception table entries, one for the load, 659 the other for the store. Either return -EFAULT. 660 Each of the entries must be relocated. */ 661 ASM_EXCEPTIONTABLE_ENTRY(1b-linux_gateway_page, 3b-linux_gateway_page) 662 ASM_EXCEPTIONTABLE_ENTRY(2b-linux_gateway_page, 3b-linux_gateway_page) 663 664 665 /*************************************************** 666 New CAS implementation which uses pointers and variable size 667 information. The value pointed by old and new MUST NOT change 668 while performing CAS. The lock only protect the value at %r26. 669 670 %r26 - Address to examine 671 %r25 - Pointer to the value to check (old) 672 %r24 - Pointer to the value to set (new) 673 %r23 - Size of the variable (0/1/2/3 for 8/16/32/64 bit) 674 %r28 - Return non-zero on failure 675 %r21 - Kernel error code 676 677 %r21 has the following meanings: 678 679 EAGAIN - CAS is busy, ldcw failed, try again. 680 EFAULT - Read or write failed. 681 682 Scratch: r20, r22, r28, r29, r1, fr4 (32bit for 64bit CAS only) 683 684 ****************************************************/ 685 686 /* ELF32 Process entry path */ 687lws_compare_and_swap_2: 688#ifdef CONFIG_64BIT 689 /* Clip the input registers */ 690 depdi 0, 31, 32, %r26 691 depdi 0, 31, 32, %r25 692 depdi 0, 31, 32, %r24 693 depdi 0, 31, 32, %r23 694#endif 695 696 /* Check the validity of the size pointer */ 697 subi,>>= 4, %r23, %r0 698 b,n lws_exit_nosys 699 700 /* Jump to the functions which will load the old and new values into 701 registers depending on the their size */ 702 shlw %r23, 2, %r29 703 blr %r29, %r0 704 nop 705 706 /* 8bit load */ 7074: ldb 0(%sr3,%r25), %r25 708 b cas2_lock_start 7095: ldb 0(%sr3,%r24), %r24 710 nop 711 nop 712 nop 713 nop 714 nop 715 716 /* 16bit load */ 7176: ldh 0(%sr3,%r25), %r25 718 b cas2_lock_start 7197: ldh 0(%sr3,%r24), %r24 720 nop 721 nop 722 nop 723 nop 724 nop 725 726 /* 32bit load */ 7278: ldw 0(%sr3,%r25), %r25 728 b cas2_lock_start 7299: ldw 0(%sr3,%r24), %r24 730 nop 731 nop 732 nop 733 nop 734 nop 735 736 /* 64bit load */ 737#ifdef CONFIG_64BIT 73810: ldd 0(%sr3,%r25), %r25 73911: ldd 0(%sr3,%r24), %r24 740#else 741 /* Load new value into r22/r23 - high/low */ 74210: ldw 0(%sr3,%r25), %r22 74311: ldw 4(%sr3,%r25), %r23 744 /* Load new value into fr4 for atomic store later */ 74512: flddx 0(%sr3,%r24), %fr4 746#endif 747 748cas2_lock_start: 749 /* Load start of lock table */ 750 ldil L%lws_lock_start, %r20 751 ldo R%lws_lock_start(%r20), %r28 752 753 /* Extract four bits from r26 and hash lock (Bits 4-7) */ 754 extru %r26, 27, 4, %r20 755 756 /* Find lock to use, the hash is either one of 0 to 757 15, multiplied by 16 (keep it 16-byte aligned) 758 and add to the lock table offset. */ 759 shlw %r20, 4, %r20 760 add %r20, %r28, %r20 761 762 rsm PSW_SM_I, %r0 /* Disable interrupts */ 763 /* COW breaks can cause contention on UP systems */ 764 LDCW 0(%sr2,%r20), %r28 /* Try to acquire the lock */ 765 cmpb,<>,n %r0, %r28, cas2_action /* Did we get it? */ 766cas2_wouldblock: 767 ldo 2(%r0), %r28 /* 2nd case */ 768 ssm PSW_SM_I, %r0 769 b lws_exit /* Contended... */ 770 ldo -EAGAIN(%r0), %r21 /* Spin in userspace */ 771 772 /* 773 prev = *addr; 774 if ( prev == old ) 775 *addr = new; 776 return prev; 777 */ 778 779 /* NOTES: 780 This all works becuse intr_do_signal 781 and schedule both check the return iasq 782 and see that we are on the kernel page 783 so this process is never scheduled off 784 or is ever sent any signal of any sort, 785 thus it is wholly atomic from usrspaces 786 perspective 787 */ 788cas2_action: 789 /* Jump to the correct function */ 790 blr %r29, %r0 791 /* Set %r28 as non-zero for now */ 792 ldo 1(%r0),%r28 793 794 /* 8bit CAS */ 79513: ldb,ma 0(%sr3,%r26), %r29 796 sub,= %r29, %r25, %r0 797 b,n cas2_end 79814: stb,ma %r24, 0(%sr3,%r26) 799 b cas2_end 800 copy %r0, %r28 801 nop 802 nop 803 804 /* 16bit CAS */ 80515: ldh,ma 0(%sr3,%r26), %r29 806 sub,= %r29, %r25, %r0 807 b,n cas2_end 80816: sth,ma %r24, 0(%sr3,%r26) 809 b cas2_end 810 copy %r0, %r28 811 nop 812 nop 813 814 /* 32bit CAS */ 81517: ldw,ma 0(%sr3,%r26), %r29 816 sub,= %r29, %r25, %r0 817 b,n cas2_end 81818: stw,ma %r24, 0(%sr3,%r26) 819 b cas2_end 820 copy %r0, %r28 821 nop 822 nop 823 824 /* 64bit CAS */ 825#ifdef CONFIG_64BIT 82619: ldd,ma 0(%sr3,%r26), %r29 827 sub,*= %r29, %r25, %r0 828 b,n cas2_end 82920: std,ma %r24, 0(%sr3,%r26) 830 copy %r0, %r28 831#else 832 /* Compare first word */ 83319: ldw,ma 0(%sr3,%r26), %r29 834 sub,= %r29, %r22, %r0 835 b,n cas2_end 836 /* Compare second word */ 83720: ldw,ma 4(%sr3,%r26), %r29 838 sub,= %r29, %r23, %r0 839 b,n cas2_end 840 /* Perform the store */ 84121: fstdx %fr4, 0(%sr3,%r26) 842 copy %r0, %r28 843#endif 844 845cas2_end: 846 /* Free lock */ 847 stw,ma %r20, 0(%sr2,%r20) 848 /* Enable interrupts */ 849 ssm PSW_SM_I, %r0 850 /* Return to userspace, set no error */ 851 b lws_exit 852 copy %r0, %r21 853 85422: 855 /* Error occurred on load or store */ 856 /* Free lock */ 857 stw %r20, 0(%sr2,%r20) 858 ssm PSW_SM_I, %r0 859 ldo 1(%r0),%r28 860 b lws_exit 861 ldo -EFAULT(%r0),%r21 /* set errno */ 862 nop 863 nop 864 nop 865 866 /* Exception table entries, for the load and store, return EFAULT. 867 Each of the entries must be relocated. */ 868 ASM_EXCEPTIONTABLE_ENTRY(4b-linux_gateway_page, 22b-linux_gateway_page) 869 ASM_EXCEPTIONTABLE_ENTRY(5b-linux_gateway_page, 22b-linux_gateway_page) 870 ASM_EXCEPTIONTABLE_ENTRY(6b-linux_gateway_page, 22b-linux_gateway_page) 871 ASM_EXCEPTIONTABLE_ENTRY(7b-linux_gateway_page, 22b-linux_gateway_page) 872 ASM_EXCEPTIONTABLE_ENTRY(8b-linux_gateway_page, 22b-linux_gateway_page) 873 ASM_EXCEPTIONTABLE_ENTRY(9b-linux_gateway_page, 22b-linux_gateway_page) 874 ASM_EXCEPTIONTABLE_ENTRY(10b-linux_gateway_page, 22b-linux_gateway_page) 875 ASM_EXCEPTIONTABLE_ENTRY(11b-linux_gateway_page, 22b-linux_gateway_page) 876 ASM_EXCEPTIONTABLE_ENTRY(13b-linux_gateway_page, 22b-linux_gateway_page) 877 ASM_EXCEPTIONTABLE_ENTRY(14b-linux_gateway_page, 22b-linux_gateway_page) 878 ASM_EXCEPTIONTABLE_ENTRY(15b-linux_gateway_page, 22b-linux_gateway_page) 879 ASM_EXCEPTIONTABLE_ENTRY(16b-linux_gateway_page, 22b-linux_gateway_page) 880 ASM_EXCEPTIONTABLE_ENTRY(17b-linux_gateway_page, 22b-linux_gateway_page) 881 ASM_EXCEPTIONTABLE_ENTRY(18b-linux_gateway_page, 22b-linux_gateway_page) 882 ASM_EXCEPTIONTABLE_ENTRY(19b-linux_gateway_page, 22b-linux_gateway_page) 883 ASM_EXCEPTIONTABLE_ENTRY(20b-linux_gateway_page, 22b-linux_gateway_page) 884#ifndef CONFIG_64BIT 885 ASM_EXCEPTIONTABLE_ENTRY(12b-linux_gateway_page, 22b-linux_gateway_page) 886 ASM_EXCEPTIONTABLE_ENTRY(21b-linux_gateway_page, 22b-linux_gateway_page) 887#endif 888 889 /* Make sure nothing else is placed on this page */ 890 .align PAGE_SIZE 891END(linux_gateway_page) 892ENTRY(end_linux_gateway_page) 893 894 /* Relocate symbols assuming linux_gateway_page is mapped 895 to virtual address 0x0 */ 896 897#define LWS_ENTRY(_name_) ASM_ULONG_INSN (lws_##_name_ - linux_gateway_page) 898 899 .section .rodata,"a" 900 901 .align 8 902 /* Light-weight-syscall table */ 903 /* Start of lws table. */ 904ENTRY(lws_table) 905 LWS_ENTRY(compare_and_swap32) /* 0 - ELF32 Atomic 32bit CAS */ 906 LWS_ENTRY(compare_and_swap64) /* 1 - ELF64 Atomic 32bit CAS */ 907 LWS_ENTRY(compare_and_swap_2) /* 2 - ELF32 Atomic 64bit CAS */ 908END(lws_table) 909 /* End of lws table */ 910 911 .align 8 912ENTRY(sys_call_table) 913#include "syscall_table.S" 914END(sys_call_table) 915 916#ifdef CONFIG_64BIT 917 .align 8 918ENTRY(sys_call_table64) 919#define SYSCALL_TABLE_64BIT 920#include "syscall_table.S" 921END(sys_call_table64) 922#endif 923 924 /* 925 All light-weight-syscall atomic operations 926 will use this set of locks 927 928 NOTE: The lws_lock_start symbol must be 929 at least 16-byte aligned for safe use 930 with ldcw. 931 */ 932 .section .data 933 .align L1_CACHE_BYTES 934ENTRY(lws_lock_start) 935 /* lws locks */ 936 .rept 16 937 /* Keep locks aligned at 16-bytes */ 938 .word 1 939 .word 0 940 .word 0 941 .word 0 942 .endr 943END(lws_lock_start) 944 .previous 945 946.end 947 948 949