2 * Linux/PA-RISC Project (http://www.parisc-linux.org/)
4 * System call entry code Copyright (c) Matthew Wilcox 1999 <willy@bofh.ai>
5 * Licensed under the GNU GPL.
6 * thanks to Philipp Rumpf, Mike Shaver and various others
7 * sorry about the wall, puffin..
9 #include <linux/config.h> /* for CONFIG_SMP */
11 #include <asm/asm-offsets.h>
12 #include <asm/unistd.h>
13 #include <asm/errno.h>
15 #include <asm/thread_info.h>
17 #include <asm/assembly.h>
18 #include <asm/processor.h>
20 /* We fill the empty parts of the gateway page with
21 * something that will kill the kernel or a
22 * userspace application.
24 #define KILL_INSN break 0,0
33 .macro fixup_branch,lbl
37 .macro fixup_branch,lbl
46 .import syscall_exit,code
47 .import syscall_exit_rfi,code
48 .export linux_gateway_page
50 /* Linux gateway page is aliased to virtual page 0 in the kernel
51 * address space. Since it is a gateway page it cannot be
52 * dereferenced, so null pointers will still fault. We start
53 * the actual entry point at 0x100. We put break instructions
54 * at the beginning of the page to trap null indirect function
61 /* ADDRESS 0x00 to 0xb0 = 176 bytes / 4 bytes per insn = 44 insns */
66 /* ADDRESS 0xb0 to 0xb4, lws uses 1 insns for entry */
67 /* Light-weight-syscall entry must always be located at 0xb0 */
68 /* WARNING: Keep this number updated with table size changes */
69 #define __NR_lws_entries (2)
72 /* Unconditional branch to lws_start, located on the
76 /* Fill from 0xb4 to 0xe0 */
81 /* This function MUST be located at 0xe0 for glibc's threading
82 mechanism to work. DO NOT MOVE THIS CODE EVER! */
84 gate .+8, %r0 /* increase privilege */
85 depi 3, 31, 2, %r31 /* Ensure we return into user mode. */
86 be 0(%sr7,%r31) /* return to user space */
87 mtctl %r26, %cr27 /* move arg0 to the control register */
89 /* Increase the chance of trapping if random jumps occur to this
90 address, fill from 0xf0 to 0x100 */
95 /* This address must remain fixed at 0x100 for glibc's syscalls to work */
98 gate .+8, %r0 /* become privileged */
99 mtsp %r0,%sr4 /* get kernel space into sr4 */
100 mtsp %r0,%sr5 /* get kernel space into sr5 */
101 mtsp %r0,%sr6 /* get kernel space into sr6 */
102 mfsp %sr7,%r1 /* save user sr7 */
103 mtsp %r1,%sr3 /* and store it in sr3 */
106 /* for now we can *always* set the W bit on entry to the syscall
107 * since we don't support wide userland processes. We could
108 * also save the current SM other than in r0 and restore it on
109 * exit from the syscall, and also use that value to know
110 * whether to do narrow or wide syscalls. -PB
113 extrd,u %r1,PSW_W_BIT,1,%r1
114 /* sp must be aligned on 4, so deposit the W bit setting into
115 * the bottom of sp temporarily */
118 /* The top halves of argument registers must be cleared on syscall
119 * entry from narrow executable.
121 depdi 0, 31, 32, %r26
122 depdi 0, 31, 32, %r25
123 depdi 0, 31, 32, %r24
124 depdi 0, 31, 32, %r23
125 depdi 0, 31, 32, %r22
126 depdi 0, 31, 32, %r21
130 xor %r1,%r30,%r30 /* ye olde xor trick */
134 ldo THREAD_SZ_ALGN+FRAME_SIZE(%r30),%r30 /* set up kernel stack */
136 /* N.B.: It is critical that we don't set sr7 to 0 until r30
137 * contains a valid kernel stack pointer. It is also
138 * critical that we don't start using the kernel stack
139 * until after sr7 has been set to 0.
142 mtsp %r0,%sr7 /* get kernel space into sr7 */
143 STREGM %r1,FRAME_SIZE(%r30) /* save r1 (usp) here for now */
144 mfctl %cr30,%r1 /* get task ptr in %r1 */
145 LDREG TI_TASK(%r1),%r1
147 /* Save some registers for sigcontext and potential task
148 switch (see entry.S for the details of which ones are
149 saved/restored). TASK_PT_PSW is zeroed so we can see whether
150 a process is on a syscall or not. For an interrupt the real
151 PSW value is stored. This is needed for gdb and sys_ptrace. */
152 STREG %r0, TASK_PT_PSW(%r1)
153 STREG %r2, TASK_PT_GR2(%r1) /* preserve rp */
154 STREG %r19, TASK_PT_GR19(%r1)
156 LDREGM -FRAME_SIZE(%r30), %r2 /* get users sp back */
158 extrd,u %r2,63,1,%r19 /* W hidden in bottom bit */
160 xor %r19,%r2,%r2 /* clear bottom bit */
162 std %r19,TASK_PT_PSW(%r1)
165 STREG %r2, TASK_PT_GR30(%r1) /* ... and save it */
167 STREG %r20, TASK_PT_GR20(%r1)
168 STREG %r21, TASK_PT_GR21(%r1)
169 STREG %r22, TASK_PT_GR22(%r1)
170 STREG %r23, TASK_PT_GR23(%r1) /* 4th argument */
171 STREG %r24, TASK_PT_GR24(%r1) /* 3rd argument */
172 STREG %r25, TASK_PT_GR25(%r1) /* 2nd argument */
173 STREG %r26, TASK_PT_GR26(%r1) /* 1st argument */
174 STREG %r27, TASK_PT_GR27(%r1) /* user dp */
175 STREG %r28, TASK_PT_GR28(%r1) /* return value 0 */
176 STREG %r28, TASK_PT_ORIG_R28(%r1) /* return value 0 (saved for signals) */
177 STREG %r29, TASK_PT_GR29(%r1) /* return value 1 */
178 STREG %r31, TASK_PT_GR31(%r1) /* preserve syscall return ptr */
180 ldo TASK_PT_FR0(%r1), %r27 /* save fpregs from the kernel */
181 save_fp %r27 /* or potential task switch */
183 mfctl %cr11, %r27 /* i.e. SAR */
184 STREG %r27, TASK_PT_SAR(%r1)
189 ldo -16(%r30),%r29 /* Reference param save area */
190 copy %r19,%r2 /* W bit back to r2 */
192 /* no need to save these on stack in wide mode because the first 8
193 * args are passed in registers */
194 stw %r22, -52(%r30) /* 5th argument */
195 stw %r21, -56(%r30) /* 6th argument */
198 /* Are we being ptraced? */
200 LDREG TI_TASK(%r1),%r1
201 LDREG TASK_PTRACE(%r1), %r1
202 bb,<,n %r1,31,.Ltracesys
204 /* Note! We cannot use the syscall table that is mapped
205 nearby since the gateway page is mapped execute-only. */
208 ldil L%sys_call_table, %r1
210 addil L%(sys_call_table64-sys_call_table), %r1
211 ldo R%sys_call_table(%r1), %r19
213 ldo R%sys_call_table64(%r1), %r19
215 ldil L%sys_call_table, %r1
216 ldo R%sys_call_table(%r1), %r19
218 comiclr,>>= __NR_Linux_syscalls, %r20, %r0
221 LDREGX %r20(%r19), %r19
223 /* If this is a sys_rt_sigreturn call, and the signal was received
224 * when not in_syscall, then we want to return via syscall_exit_rfi,
225 * not syscall_exit. Signal no. in r20, in_syscall in r25 (see
226 * trampoline code in signal.c).
228 ldi __NR_rt_sigreturn,%r2
229 comb,= %r2,%r20,.Lrt_sigreturn
231 ldil L%syscall_exit,%r2
233 ldo R%syscall_exit(%r2),%r2
235 comib,<> 0,%r25,.Lin_syscall
236 ldil L%syscall_exit_rfi,%r2
238 ldo R%syscall_exit_rfi(%r2),%r2
240 /* Note! Because we are not running where we were linked, any
241 calls to functions external to this file must be indirect. To
242 be safe, we apply the opposite rule to functions within this
243 file, with local labels given to them to ensure correctness. */
247 ldil L%syscall_exit,%r1
248 be R%syscall_exit(%sr7,%r1)
249 ldo -ENOSYS(%r0),%r28 /* set errno */
252 /* Warning! This trace code is a virtual duplicate of the code above so be
253 * sure to maintain both! */
256 /* Need to save more registers so the debugger can see where we
257 * are. This saves only the lower 8 bits of PSW, so that the C
258 * bit is still clear on syscalls, and the D bit is set if this
259 * full register save path has been executed. We check the D
260 * bit on syscall_return_rfi to determine which registers to
261 * restore. An interrupt results in a full PSW saved with the
262 * C bit set, a non-straced syscall entry results in C and D clear
265 ldo -THREAD_SZ_ALGN-FRAME_SIZE(%r30),%r1 /* get task ptr */
266 LDREG TI_TASK(%r1), %r1
268 STREG %r2,TASK_PT_PSW(%r1) /* Lower 8 bits only!! */
270 STREG %r2,TASK_PT_SR0(%r1)
272 STREG %r2,TASK_PT_SR1(%r1)
274 STREG %r2,TASK_PT_SR2(%r1)
276 STREG %r2,TASK_PT_SR3(%r1)
277 STREG %r2,TASK_PT_SR4(%r1)
278 STREG %r2,TASK_PT_SR5(%r1)
279 STREG %r2,TASK_PT_SR6(%r1)
280 STREG %r2,TASK_PT_SR7(%r1)
281 STREG %r2,TASK_PT_IASQ0(%r1)
282 STREG %r2,TASK_PT_IASQ1(%r1)
283 LDREG TASK_PT_GR31(%r1),%r2
284 STREG %r2,TASK_PT_IAOQ0(%r1)
286 STREG %r2,TASK_PT_IAOQ1(%r1)
287 ldo TASK_REGS(%r1),%r2
289 STREG %r3,PT_GR3(%r2)
290 STREG %r4,PT_GR4(%r2)
291 STREG %r5,PT_GR5(%r2)
292 STREG %r6,PT_GR6(%r2)
293 STREG %r7,PT_GR7(%r2)
294 STREG %r8,PT_GR8(%r2)
295 STREG %r9,PT_GR9(%r2)
296 STREG %r10,PT_GR10(%r2)
297 STREG %r11,PT_GR11(%r2)
298 STREG %r12,PT_GR12(%r2)
299 STREG %r13,PT_GR13(%r2)
300 STREG %r14,PT_GR14(%r2)
301 STREG %r15,PT_GR15(%r2)
302 STREG %r16,PT_GR16(%r2)
303 STREG %r17,PT_GR17(%r2)
304 STREG %r18,PT_GR18(%r2)
305 /* Finished saving things for the debugger */
307 ldil L%syscall_trace,%r1
308 ldil L%tracesys_next,%r2
309 be R%syscall_trace(%sr7,%r1)
310 ldo R%tracesys_next(%r2),%r2
313 ldil L%sys_call_table,%r1
314 ldo R%sys_call_table(%r1), %r19
316 ldo -THREAD_SZ_ALGN-FRAME_SIZE(%r30),%r1 /* get task ptr */
317 LDREG TI_TASK(%r1), %r1
318 LDREG TASK_PT_GR20(%r1), %r20
319 LDREG TASK_PT_GR26(%r1), %r26 /* Restore the users args */
320 LDREG TASK_PT_GR25(%r1), %r25
321 LDREG TASK_PT_GR24(%r1), %r24
322 LDREG TASK_PT_GR23(%r1), %r23
324 LDREG TASK_PT_GR22(%r1), %r22
325 LDREG TASK_PT_GR21(%r1), %r21
326 ldo -16(%r30),%r29 /* Reference param save area */
329 comiclr,>>= __NR_Linux_syscalls, %r20, %r0
332 LDREGX %r20(%r19), %r19
334 /* If this is a sys_rt_sigreturn call, and the signal was received
335 * when not in_syscall, then we want to return via syscall_exit_rfi,
336 * not syscall_exit. Signal no. in r20, in_syscall in r25 (see
337 * trampoline code in signal.c).
339 ldi __NR_rt_sigreturn,%r2
340 comb,= %r2,%r20,.Ltrace_rt_sigreturn
342 ldil L%tracesys_exit,%r2
344 ldo R%tracesys_exit(%r2),%r2
346 /* Do *not* call this function on the gateway page, because it
347 makes a direct call to syscall_trace. */
350 ldo -THREAD_SZ_ALGN-FRAME_SIZE(%r30),%r1 /* get task ptr */
351 LDREG TI_TASK(%r1), %r1
353 ldo -16(%r30),%r29 /* Reference param save area */
355 bl syscall_trace, %r2
356 STREG %r28,TASK_PT_GR28(%r1) /* save return value now */
357 ldo -THREAD_SZ_ALGN-FRAME_SIZE(%r30),%r1 /* get task ptr */
358 LDREG TI_TASK(%r1), %r1
359 LDREG TASK_PT_GR28(%r1), %r28 /* Restore return val. */
361 ldil L%syscall_exit,%r1
362 be,n R%syscall_exit(%sr7,%r1)
364 .Ltrace_rt_sigreturn:
365 comib,<> 0,%r25,.Ltrace_in_syscall
366 ldil L%tracesys_sigexit,%r2
368 ldo R%tracesys_sigexit(%r2),%r2
371 ldo -THREAD_SZ_ALGN-FRAME_SIZE(%r30),%r1 /* get task ptr */
374 ldo -16(%r30),%r29 /* Reference param save area */
376 bl syscall_trace, %r2
379 ldil L%syscall_exit_rfi,%r1
380 be,n R%syscall_exit_rfi(%sr7,%r1)
383 /*********************************************************
384 Light-weight-syscall code
387 r26,r25,r24,r23,r22 - Input registers
388 r28 - Function return register
391 Scracth: Any of the above that aren't being
392 currently used, including r1.
394 Return pointer: r31 (Not usable)
396 Error codes returned by entry path:
398 ENOSYS - r20 was an invalid LWS number.
400 *********************************************************/
402 /* Gate and ensure we return to userspace */
404 depi 3, 31, 2, %r31 /* Ensure we return to userspace */
407 /* FIXME: If we are a 64-bit kernel just
408 * turn this on unconditionally.
411 extrd,u %r1,PSW_W_BIT,1,%r1
412 /* sp must be aligned on 4, so deposit the W bit setting into
413 * the bottom of sp temporarily */
416 /* Clip LWS number to a 32-bit value always */
417 depdi 0, 31, 32, %r20
420 /* Is the lws entry number valid? */
421 comiclr,>>= __NR_lws_entries, %r20, %r0
424 /* WARNING: Trashing sr2 and sr3 */
425 mfsp %sr7,%r1 /* get userspace into sr3 */
427 mtsp %r0,%sr2 /* get kernel space into sr2 */
429 /* Load table start */
430 ldil L%lws_table, %r1
431 ldo R%lws_table(%r1), %r28 /* Scratch use of r28 */
432 LDREGX %r20(%sr2,r28), %r21 /* Scratch use of r21 */
434 /* Jump to lws, lws table pointers already relocated */
438 ldo -ENOSYS(%r0),%r21 /* set errno */
439 /* Fall through: Return to userspace */
443 /* decide whether to reset the wide mode bit
445 * For a syscall, the W bit is stored in the lowest bit
446 * of sp. Extract it and reset W if it is zero */
447 extrd,u,*<> %r30,63,1,%r1
449 /* now reset the lowest bit of sp if it was set */
456 /***************************************************
457 Implementing CAS as an atomic operation:
459 %r26 - Address to examine
460 %r25 - Old value to check (old)
461 %r24 - New value to set (new)
462 %r28 - Return prev through this register.
463 %r21 - Kernel error code
465 If debugging is DISabled:
467 %r21 has the following meanings:
469 EAGAIN - CAS is busy, ldcw failed, try again.
470 EFAULT - Read or write failed.
472 If debugging is enabled:
474 EDEADLOCK - CAS called recursively.
475 EAGAIN && r28 == 1 - CAS is busy. Lock contended.
476 EAGAIN && r28 == 2 - CAS is busy. ldcw failed.
477 EFAULT - Read or write failed.
479 Scratch: r20, r28, r1
481 ****************************************************/
483 /* Do not enable LWS debugging */
484 #define ENABLE_LWS_DEBUG 0
486 /* ELF64 Process entry path */
487 lws_compare_and_swap64:
489 b,n lws_compare_and_swap
491 /* If we are not a 64-bit kernel, then we don't
492 * implement having 64-bit input registers
497 /* ELF32 Process entry path */
498 lws_compare_and_swap32:
500 /* Clip all the input registers */
501 depdi 0, 31, 32, %r26
502 depdi 0, 31, 32, %r25
503 depdi 0, 31, 32, %r24
506 lws_compare_and_swap:
508 /* Load start of lock table */
509 ldil L%lws_lock_start, %r20
510 ldo R%lws_lock_start(%r20), %r28
512 /* Extract four bits from r26 and hash lock (Bits 4-7) */
513 extru %r26, 27, 4, %r20
515 /* Find lock to use, the hash is either one of 0 to
516 15, multiplied by 16 (keep it 16-byte aligned)
517 and add to the lock table offset. */
521 # ifdef ENABLE_LWS_DEBUG
523 DEBUG, check for deadlock!
524 If the thread register values are the same
525 then we were the one that locked it last and
526 this is a recurisve call that will deadlock.
527 We *must* giveup this call and fail.
529 ldw 4(%sr2,%r20), %r28 /* Load thread register */
530 mfctl %cr27, %r21 /* Get current thread register */
531 cmpb,<>,n %r21, %r28, cas_lock /* Called recursive? */
532 b lws_exit /* Return error! */
533 ldo -EDEADLOCK(%r0), %r21
535 cmpb,=,n %r0, %r28, cas_nocontend /* Is nobody using it? */
536 ldo 1(%r0), %r28 /* 1st case */
537 b lws_exit /* Contended... */
538 ldo -EAGAIN(%r0), %r21 /* Spin in userspace */
541 /* ENABLE_LWS_DEBUG */
543 ldcw 0(%sr2,%r20), %r28 /* Try to acquire the lock */
544 cmpb,<>,n %r0, %r28, cas_action /* Did we get it? */
546 ldo 2(%r0), %r28 /* 2nd case */
547 b lws_exit /* Contended... */
548 ldo -EAGAIN(%r0), %r21 /* Spin in userspace */
560 This all works becuse intr_do_signal
561 and schedule both check the return iasq
562 and see that we are on the kernel page
563 so this process is never scheduled off
564 or is ever sent any signal of any sort,
565 thus it is wholly atomic from usrspaces
569 #if defined CONFIG_SMP && defined ENABLE_LWS_DEBUG
572 stw %r1, 4(%sr2,%r20)
574 /* The load and store could fail */
575 1: ldw 0(%sr3,%r26), %r28
576 sub,<> %r28, %r25, %r0
577 2: stw %r24, 0(%sr3,%r26)
580 stw %r20, 0(%sr2,%r20)
581 # ifdef ENABLE_LWS_DEBUG
582 /* Clear thread register indicator */
583 stw %r0, 4(%sr2,%r20)
586 /* Return to userspace, set no error */
591 /* Error occured on load or store */
594 stw %r20, 0(%sr2,%r20)
595 # ifdef ENABLE_LWS_DEBUG
596 stw %r0, 4(%sr2,%r20)
600 ldo -EFAULT(%r0),%r21 /* set errno */
606 /* Two exception table entries, one for the load,
607 the other for the store. Either return -EFAULT.
608 Each of the entries must be relocated. */
609 .section __ex_table,"aw"
611 /* Pad the address calculation */
612 .word 0,(2b - linux_gateway_page)
613 .word 0,(3b - linux_gateway_page)
615 .word (2b - linux_gateway_page)
616 .word (3b - linux_gateway_page)
620 .section __ex_table,"aw"
622 /* Pad the address calculation */
623 .word 0,(1b - linux_gateway_page)
624 .word 0,(3b - linux_gateway_page)
626 .word (1b - linux_gateway_page)
627 .word (3b - linux_gateway_page)
631 end_compare_and_swap:
633 /* Make sure nothing else is placed on this page */
635 .export end_linux_gateway_page
636 end_linux_gateway_page:
638 /* Relocate symbols assuming linux_gateway_page is mapped
639 to virtual address 0x0 */
641 /* FIXME: The code will always be on the gateay page
642 and thus it will be on the first 4k, the
643 assembler seems to think that the final
644 subtraction result is only a word in
645 length, so we pad the value.
647 #define LWS_ENTRY(_name_) .word 0,(lws_##_name_ - linux_gateway_page)
649 #define LWS_ENTRY(_name_) .word (lws_##_name_ - linux_gateway_page)
653 /* Light-weight-syscall table */
654 /* Start of lws table. */
658 LWS_ENTRY(compare_and_swap32) /* 0 - ELF32 Atomic compare and swap */
659 LWS_ENTRY(compare_and_swap64) /* 1 - ELF64 Atomic compare and swap */
660 /* End of lws table */
663 .export sys_call_table
666 #include "syscall_table.S"
670 .export sys_call_table64
673 #define SYSCALL_TABLE_64BIT
674 #include "syscall_table.S"
679 All light-weight-syscall atomic operations
680 will use this set of locks
684 .export lws_lock_start
690 /* Keep locks aligned at 16-bytes */
698 /* CONFIG_SMP for lws_lock_start */