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 .Lsyscall_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 /* Do *not* call this function on the gateway page, because it 363 makes a direct call to syscall_trace. */ 364 365tracesys_exit: 366 ldo -THREAD_SZ_ALGN-FRAME_SIZE(%r30),%r1 /* get task ptr */ 367 LDREG TI_TASK(%r1), %r1 368#ifdef CONFIG_64BIT 369 ldo -16(%r30),%r29 /* Reference param save area */ 370#endif 371 ldo TASK_REGS(%r1),%r26 372 BL do_syscall_trace_exit,%r2 373 STREG %r28,TASK_PT_GR28(%r1) /* save return value now */ 374 ldo -THREAD_SZ_ALGN-FRAME_SIZE(%r30),%r1 /* get task ptr */ 375 LDREG TI_TASK(%r1), %r1 376 LDREG TASK_PT_GR28(%r1), %r28 /* Restore return val. */ 377 378 ldil L%syscall_exit,%r1 379 be,n R%syscall_exit(%sr7,%r1) 380 381.Ltrace_rt_sigreturn: 382 comib,<> 0,%r25,.Ltrace_in_syscall 383 ldil L%tracesys_sigexit,%r2 384 be 0(%sr7,%r19) 385 ldo R%tracesys_sigexit(%r2),%r2 386 387tracesys_sigexit: 388 ldo -THREAD_SZ_ALGN-FRAME_SIZE(%r30),%r1 /* get task ptr */ 389 LDREG TI_TASK(%r1), %r1 390#ifdef CONFIG_64BIT 391 ldo -16(%r30),%r29 /* Reference param save area */ 392#endif 393 BL do_syscall_trace_exit,%r2 394 ldo TASK_REGS(%r1),%r26 395 396 ldil L%syscall_exit_rfi,%r1 397 be,n R%syscall_exit_rfi(%sr7,%r1) 398 399 400 /********************************************************* 401 32/64-bit Light-Weight-Syscall ABI 402 403 * - Indicates a hint for userspace inline asm 404 implementations. 405 406 Syscall number (caller-saves) 407 - %r20 408 * In asm clobber. 409 410 Argument registers (caller-saves) 411 - %r26, %r25, %r24, %r23, %r22 412 * In asm input. 413 414 Return registers (caller-saves) 415 - %r28 (return), %r21 (errno) 416 * In asm output. 417 418 Caller-saves registers 419 - %r1, %r27, %r29 420 - %r2 (return pointer) 421 - %r31 (ble link register) 422 * In asm clobber. 423 424 Callee-saves registers 425 - %r3-%r18 426 - %r30 (stack pointer) 427 * Not in asm clobber. 428 429 If userspace is 32-bit: 430 Callee-saves registers 431 - %r19 (32-bit PIC register) 432 433 Differences from 32-bit calling convention: 434 - Syscall number in %r20 435 - Additional argument register %r22 (arg4) 436 - Callee-saves %r19. 437 438 If userspace is 64-bit: 439 Callee-saves registers 440 - %r27 (64-bit PIC register) 441 442 Differences from 64-bit calling convention: 443 - Syscall number in %r20 444 - Additional argument register %r22 (arg4) 445 - Callee-saves %r27. 446 447 Error codes returned by entry path: 448 449 ENOSYS - r20 was an invalid LWS number. 450 451 *********************************************************/ 452lws_start: 453 454#ifdef CONFIG_64BIT 455 /* FIXME: If we are a 64-bit kernel just 456 * turn this on unconditionally. 457 */ 458 ssm PSW_SM_W, %r1 459 extrd,u %r1,PSW_W_BIT,1,%r1 460 /* sp must be aligned on 4, so deposit the W bit setting into 461 * the bottom of sp temporarily */ 462 or,ev %r1,%r30,%r30 463 464 /* Clip LWS number to a 32-bit value always */ 465 depdi 0, 31, 32, %r20 466#endif 467 468 /* Is the lws entry number valid? */ 469 comiclr,>> __NR_lws_entries, %r20, %r0 470 b,n lws_exit_nosys 471 472 /* WARNING: Trashing sr2 and sr3 */ 473 mfsp %sr7,%r1 /* get userspace into sr3 */ 474 mtsp %r1,%sr3 475 mtsp %r0,%sr2 /* get kernel space into sr2 */ 476 477 /* Load table start */ 478 ldil L%lws_table, %r1 479 ldo R%lws_table(%r1), %r28 /* Scratch use of r28 */ 480 LDREGX %r20(%sr2,r28), %r21 /* Scratch use of r21 */ 481 482 /* Jump to lws, lws table pointers already relocated */ 483 be,n 0(%sr2,%r21) 484 485lws_exit_nosys: 486 ldo -ENOSYS(%r0),%r21 /* set errno */ 487 /* Fall through: Return to userspace */ 488 489lws_exit: 490#ifdef CONFIG_64BIT 491 /* decide whether to reset the wide mode bit 492 * 493 * For a syscall, the W bit is stored in the lowest bit 494 * of sp. Extract it and reset W if it is zero */ 495 extrd,u,*<> %r30,63,1,%r1 496 rsm PSW_SM_W, %r0 497 /* now reset the lowest bit of sp if it was set */ 498 xor %r30,%r1,%r30 499#endif 500 be,n 0(%sr7, %r31) 501 502 503 504 /*************************************************** 505 Implementing 32bit CAS as an atomic operation: 506 507 %r26 - Address to examine 508 %r25 - Old value to check (old) 509 %r24 - New value to set (new) 510 %r28 - Return prev through this register. 511 %r21 - Kernel error code 512 513 If debugging is DISabled: 514 515 %r21 has the following meanings: 516 517 EAGAIN - CAS is busy, ldcw failed, try again. 518 EFAULT - Read or write failed. 519 520 If debugging is enabled: 521 522 EDEADLOCK - CAS called recursively. 523 EAGAIN && r28 == 1 - CAS is busy. Lock contended. 524 EAGAIN && r28 == 2 - CAS is busy. ldcw failed. 525 EFAULT - Read or write failed. 526 527 Scratch: r20, r28, r1 528 529 ****************************************************/ 530 531 /* Do not enable LWS debugging */ 532#define ENABLE_LWS_DEBUG 0 533 534 /* ELF64 Process entry path */ 535lws_compare_and_swap64: 536#ifdef CONFIG_64BIT 537 b,n lws_compare_and_swap 538#else 539 /* If we are not a 64-bit kernel, then we don't 540 * have 64-bit input registers, and calling 541 * the 64-bit LWS CAS returns ENOSYS. 542 */ 543 b,n lws_exit_nosys 544#endif 545 546 /* ELF32 Process entry path */ 547lws_compare_and_swap32: 548#ifdef CONFIG_64BIT 549 /* Clip all the input registers */ 550 depdi 0, 31, 32, %r26 551 depdi 0, 31, 32, %r25 552 depdi 0, 31, 32, %r24 553#endif 554 555lws_compare_and_swap: 556 /* Load start of lock table */ 557 ldil L%lws_lock_start, %r20 558 ldo R%lws_lock_start(%r20), %r28 559 560 /* Extract four bits from r26 and hash lock (Bits 4-7) */ 561 extru %r26, 27, 4, %r20 562 563 /* Find lock to use, the hash is either one of 0 to 564 15, multiplied by 16 (keep it 16-byte aligned) 565 and add to the lock table offset. */ 566 shlw %r20, 4, %r20 567 add %r20, %r28, %r20 568 569# if ENABLE_LWS_DEBUG 570 /* 571 DEBUG, check for deadlock! 572 If the thread register values are the same 573 then we were the one that locked it last and 574 this is a recurisve call that will deadlock. 575 We *must* giveup this call and fail. 576 */ 577 ldw 4(%sr2,%r20), %r28 /* Load thread register */ 578 /* WARNING: If cr27 cycles to the same value we have problems */ 579 mfctl %cr27, %r21 /* Get current thread register */ 580 cmpb,<>,n %r21, %r28, cas_lock /* Called recursive? */ 581 b lws_exit /* Return error! */ 582 ldo -EDEADLOCK(%r0), %r21 583cas_lock: 584 cmpb,=,n %r0, %r28, cas_nocontend /* Is nobody using it? */ 585 ldo 1(%r0), %r28 /* 1st case */ 586 b lws_exit /* Contended... */ 587 ldo -EAGAIN(%r0), %r21 /* Spin in userspace */ 588cas_nocontend: 589# endif 590/* ENABLE_LWS_DEBUG */ 591 592 rsm PSW_SM_I, %r0 /* Disable interrupts */ 593 /* COW breaks can cause contention on UP systems */ 594 LDCW 0(%sr2,%r20), %r28 /* Try to acquire the lock */ 595 cmpb,<>,n %r0, %r28, cas_action /* Did we get it? */ 596cas_wouldblock: 597 ldo 2(%r0), %r28 /* 2nd case */ 598 ssm PSW_SM_I, %r0 599 b lws_exit /* Contended... */ 600 ldo -EAGAIN(%r0), %r21 /* Spin in userspace */ 601 602 /* 603 prev = *addr; 604 if ( prev == old ) 605 *addr = new; 606 return prev; 607 */ 608 609 /* NOTES: 610 This all works becuse intr_do_signal 611 and schedule both check the return iasq 612 and see that we are on the kernel page 613 so this process is never scheduled off 614 or is ever sent any signal of any sort, 615 thus it is wholly atomic from usrspaces 616 perspective 617 */ 618cas_action: 619#if defined CONFIG_SMP && ENABLE_LWS_DEBUG 620 /* DEBUG */ 621 mfctl %cr27, %r1 622 stw %r1, 4(%sr2,%r20) 623#endif 624 /* The load and store could fail */ 6251: ldw,ma 0(%sr3,%r26), %r28 626 sub,<> %r28, %r25, %r0 6272: stw,ma %r24, 0(%sr3,%r26) 628 /* Free lock */ 629 stw,ma %r20, 0(%sr2,%r20) 630#if ENABLE_LWS_DEBUG 631 /* Clear thread register indicator */ 632 stw %r0, 4(%sr2,%r20) 633#endif 634 /* Enable interrupts */ 635 ssm PSW_SM_I, %r0 636 /* Return to userspace, set no error */ 637 b lws_exit 638 copy %r0, %r21 639 6403: 641 /* Error occurred on load or store */ 642 /* Free lock */ 643 stw %r20, 0(%sr2,%r20) 644#if ENABLE_LWS_DEBUG 645 stw %r0, 4(%sr2,%r20) 646#endif 647 ssm PSW_SM_I, %r0 648 b lws_exit 649 ldo -EFAULT(%r0),%r21 /* set errno */ 650 nop 651 nop 652 nop 653 nop 654 655 /* Two exception table entries, one for the load, 656 the other for the store. Either return -EFAULT. 657 Each of the entries must be relocated. */ 658 ASM_EXCEPTIONTABLE_ENTRY(1b-linux_gateway_page, 3b-linux_gateway_page) 659 ASM_EXCEPTIONTABLE_ENTRY(2b-linux_gateway_page, 3b-linux_gateway_page) 660 661 662 /*************************************************** 663 New CAS implementation which uses pointers and variable size 664 information. The value pointed by old and new MUST NOT change 665 while performing CAS. The lock only protect the value at %r26. 666 667 %r26 - Address to examine 668 %r25 - Pointer to the value to check (old) 669 %r24 - Pointer to the value to set (new) 670 %r23 - Size of the variable (0/1/2/3 for 8/16/32/64 bit) 671 %r28 - Return non-zero on failure 672 %r21 - Kernel error code 673 674 %r21 has the following meanings: 675 676 EAGAIN - CAS is busy, ldcw failed, try again. 677 EFAULT - Read or write failed. 678 679 Scratch: r20, r22, r28, r29, r1, fr4 (32bit for 64bit CAS only) 680 681 ****************************************************/ 682 683 /* ELF32 Process entry path */ 684lws_compare_and_swap_2: 685#ifdef CONFIG_64BIT 686 /* Clip the input registers */ 687 depdi 0, 31, 32, %r26 688 depdi 0, 31, 32, %r25 689 depdi 0, 31, 32, %r24 690 depdi 0, 31, 32, %r23 691#endif 692 693 /* Check the validity of the size pointer */ 694 subi,>>= 4, %r23, %r0 695 b,n lws_exit_nosys 696 697 /* Jump to the functions which will load the old and new values into 698 registers depending on the their size */ 699 shlw %r23, 2, %r29 700 blr %r29, %r0 701 nop 702 703 /* 8bit load */ 7044: ldb 0(%sr3,%r25), %r25 705 b cas2_lock_start 7065: ldb 0(%sr3,%r24), %r24 707 nop 708 nop 709 nop 710 nop 711 nop 712 713 /* 16bit load */ 7146: ldh 0(%sr3,%r25), %r25 715 b cas2_lock_start 7167: ldh 0(%sr3,%r24), %r24 717 nop 718 nop 719 nop 720 nop 721 nop 722 723 /* 32bit load */ 7248: ldw 0(%sr3,%r25), %r25 725 b cas2_lock_start 7269: ldw 0(%sr3,%r24), %r24 727 nop 728 nop 729 nop 730 nop 731 nop 732 733 /* 64bit load */ 734#ifdef CONFIG_64BIT 73510: ldd 0(%sr3,%r25), %r25 73611: ldd 0(%sr3,%r24), %r24 737#else 738 /* Load new value into r22/r23 - high/low */ 73910: ldw 0(%sr3,%r25), %r22 74011: ldw 4(%sr3,%r25), %r23 741 /* Load new value into fr4 for atomic store later */ 74212: flddx 0(%sr3,%r24), %fr4 743#endif 744 745cas2_lock_start: 746 /* Load start of lock table */ 747 ldil L%lws_lock_start, %r20 748 ldo R%lws_lock_start(%r20), %r28 749 750 /* Extract four bits from r26 and hash lock (Bits 4-7) */ 751 extru %r26, 27, 4, %r20 752 753 /* Find lock to use, the hash is either one of 0 to 754 15, multiplied by 16 (keep it 16-byte aligned) 755 and add to the lock table offset. */ 756 shlw %r20, 4, %r20 757 add %r20, %r28, %r20 758 759 rsm PSW_SM_I, %r0 /* Disable interrupts */ 760 /* COW breaks can cause contention on UP systems */ 761 LDCW 0(%sr2,%r20), %r28 /* Try to acquire the lock */ 762 cmpb,<>,n %r0, %r28, cas2_action /* Did we get it? */ 763cas2_wouldblock: 764 ldo 2(%r0), %r28 /* 2nd case */ 765 ssm PSW_SM_I, %r0 766 b lws_exit /* Contended... */ 767 ldo -EAGAIN(%r0), %r21 /* Spin in userspace */ 768 769 /* 770 prev = *addr; 771 if ( prev == old ) 772 *addr = new; 773 return prev; 774 */ 775 776 /* NOTES: 777 This all works becuse intr_do_signal 778 and schedule both check the return iasq 779 and see that we are on the kernel page 780 so this process is never scheduled off 781 or is ever sent any signal of any sort, 782 thus it is wholly atomic from usrspaces 783 perspective 784 */ 785cas2_action: 786 /* Jump to the correct function */ 787 blr %r29, %r0 788 /* Set %r28 as non-zero for now */ 789 ldo 1(%r0),%r28 790 791 /* 8bit CAS */ 79213: ldb,ma 0(%sr3,%r26), %r29 793 sub,= %r29, %r25, %r0 794 b,n cas2_end 79514: stb,ma %r24, 0(%sr3,%r26) 796 b cas2_end 797 copy %r0, %r28 798 nop 799 nop 800 801 /* 16bit CAS */ 80215: ldh,ma 0(%sr3,%r26), %r29 803 sub,= %r29, %r25, %r0 804 b,n cas2_end 80516: sth,ma %r24, 0(%sr3,%r26) 806 b cas2_end 807 copy %r0, %r28 808 nop 809 nop 810 811 /* 32bit CAS */ 81217: ldw,ma 0(%sr3,%r26), %r29 813 sub,= %r29, %r25, %r0 814 b,n cas2_end 81518: stw,ma %r24, 0(%sr3,%r26) 816 b cas2_end 817 copy %r0, %r28 818 nop 819 nop 820 821 /* 64bit CAS */ 822#ifdef CONFIG_64BIT 82319: ldd,ma 0(%sr3,%r26), %r29 824 sub,*= %r29, %r25, %r0 825 b,n cas2_end 82620: std,ma %r24, 0(%sr3,%r26) 827 copy %r0, %r28 828#else 829 /* Compare first word */ 83019: ldw,ma 0(%sr3,%r26), %r29 831 sub,= %r29, %r22, %r0 832 b,n cas2_end 833 /* Compare second word */ 83420: ldw,ma 4(%sr3,%r26), %r29 835 sub,= %r29, %r23, %r0 836 b,n cas2_end 837 /* Perform the store */ 83821: fstdx %fr4, 0(%sr3,%r26) 839 copy %r0, %r28 840#endif 841 842cas2_end: 843 /* Free lock */ 844 stw,ma %r20, 0(%sr2,%r20) 845 /* Enable interrupts */ 846 ssm PSW_SM_I, %r0 847 /* Return to userspace, set no error */ 848 b lws_exit 849 copy %r0, %r21 850 85122: 852 /* Error occurred on load or store */ 853 /* Free lock */ 854 stw %r20, 0(%sr2,%r20) 855 ssm PSW_SM_I, %r0 856 ldo 1(%r0),%r28 857 b lws_exit 858 ldo -EFAULT(%r0),%r21 /* set errno */ 859 nop 860 nop 861 nop 862 863 /* Exception table entries, for the load and store, return EFAULT. 864 Each of the entries must be relocated. */ 865 ASM_EXCEPTIONTABLE_ENTRY(4b-linux_gateway_page, 22b-linux_gateway_page) 866 ASM_EXCEPTIONTABLE_ENTRY(5b-linux_gateway_page, 22b-linux_gateway_page) 867 ASM_EXCEPTIONTABLE_ENTRY(6b-linux_gateway_page, 22b-linux_gateway_page) 868 ASM_EXCEPTIONTABLE_ENTRY(7b-linux_gateway_page, 22b-linux_gateway_page) 869 ASM_EXCEPTIONTABLE_ENTRY(8b-linux_gateway_page, 22b-linux_gateway_page) 870 ASM_EXCEPTIONTABLE_ENTRY(9b-linux_gateway_page, 22b-linux_gateway_page) 871 ASM_EXCEPTIONTABLE_ENTRY(10b-linux_gateway_page, 22b-linux_gateway_page) 872 ASM_EXCEPTIONTABLE_ENTRY(11b-linux_gateway_page, 22b-linux_gateway_page) 873 ASM_EXCEPTIONTABLE_ENTRY(13b-linux_gateway_page, 22b-linux_gateway_page) 874 ASM_EXCEPTIONTABLE_ENTRY(14b-linux_gateway_page, 22b-linux_gateway_page) 875 ASM_EXCEPTIONTABLE_ENTRY(15b-linux_gateway_page, 22b-linux_gateway_page) 876 ASM_EXCEPTIONTABLE_ENTRY(16b-linux_gateway_page, 22b-linux_gateway_page) 877 ASM_EXCEPTIONTABLE_ENTRY(17b-linux_gateway_page, 22b-linux_gateway_page) 878 ASM_EXCEPTIONTABLE_ENTRY(18b-linux_gateway_page, 22b-linux_gateway_page) 879 ASM_EXCEPTIONTABLE_ENTRY(19b-linux_gateway_page, 22b-linux_gateway_page) 880 ASM_EXCEPTIONTABLE_ENTRY(20b-linux_gateway_page, 22b-linux_gateway_page) 881#ifndef CONFIG_64BIT 882 ASM_EXCEPTIONTABLE_ENTRY(12b-linux_gateway_page, 22b-linux_gateway_page) 883 ASM_EXCEPTIONTABLE_ENTRY(21b-linux_gateway_page, 22b-linux_gateway_page) 884#endif 885 886 /* Make sure nothing else is placed on this page */ 887 .align PAGE_SIZE 888END(linux_gateway_page) 889ENTRY(end_linux_gateway_page) 890 891 /* Relocate symbols assuming linux_gateway_page is mapped 892 to virtual address 0x0 */ 893 894#define LWS_ENTRY(_name_) ASM_ULONG_INSN (lws_##_name_ - linux_gateway_page) 895 896 .section .rodata,"a" 897 898 .align 8 899 /* Light-weight-syscall table */ 900 /* Start of lws table. */ 901ENTRY(lws_table) 902 LWS_ENTRY(compare_and_swap32) /* 0 - ELF32 Atomic 32bit CAS */ 903 LWS_ENTRY(compare_and_swap64) /* 1 - ELF64 Atomic 32bit CAS */ 904 LWS_ENTRY(compare_and_swap_2) /* 2 - ELF32 Atomic 64bit CAS */ 905END(lws_table) 906 /* End of lws table */ 907 908 .align 8 909ENTRY(sys_call_table) 910#include "syscall_table.S" 911END(sys_call_table) 912 913#ifdef CONFIG_64BIT 914 .align 8 915ENTRY(sys_call_table64) 916#define SYSCALL_TABLE_64BIT 917#include "syscall_table.S" 918END(sys_call_table64) 919#endif 920 921 /* 922 All light-weight-syscall atomic operations 923 will use this set of locks 924 925 NOTE: The lws_lock_start symbol must be 926 at least 16-byte aligned for safe use 927 with ldcw. 928 */ 929 .section .data 930 .align L1_CACHE_BYTES 931ENTRY(lws_lock_start) 932 /* lws locks */ 933 .rept 16 934 /* Keep locks aligned at 16-bytes */ 935 .word 1 936 .word 0 937 .word 0 938 .word 0 939 .endr 940END(lws_lock_start) 941 .previous 942 943.end 944 945 946