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..
10 #include <asm/asm-offsets.h>
11 #include <asm/unistd.h>
12 #include <asm/errno.h>
14 #include <asm/thread_info.h>
16 #include <asm/assembly.h>
17 #include <asm/processor.h>
19 /* We fill the empty parts of the gateway page with
20 * something that will kill the kernel or a
21 * userspace application.
23 #define KILL_INSN break 0,0
25 #include <linux/config.h> /* for CONFIG_SMP */
34 .macro fixup_branch,lbl
38 .macro fixup_branch,lbl
47 .import syscall_exit,code
48 .import syscall_exit_rfi,code
49 .export linux_gateway_page
51 /* Linux gateway page is aliased to virtual page 0 in the kernel
52 * address space. Since it is a gateway page it cannot be
53 * dereferenced, so null pointers will still fault. We start
54 * the actual entry point at 0x100. We put break instructions
55 * at the beginning of the page to trap null indirect function
62 /* ADDRESS 0x00 to 0xb0 = 176 bytes / 4 bytes per insn = 44 insns */
67 /* ADDRESS 0xb0 to 0xb4, lws uses 1 insns for entry */
68 /* Light-weight-syscall entry must always be located at 0xb0 */
69 /* WARNING: Keep this number updated with table size changes */
70 #define __NR_lws_entries (2)
73 /* Unconditional branch to lws_start, located on the
77 /* Fill from 0xb4 to 0xe0 */
82 /* This function MUST be located at 0xe0 for glibc's threading
83 mechanism to work. DO NOT MOVE THIS CODE EVER! */
85 gate .+8, %r0 /* increase privilege */
86 depi 3, 31, 2, %r31 /* Ensure we return into user mode. */
87 be 0(%sr7,%r31) /* return to user space */
88 mtctl %r26, %cr27 /* move arg0 to the control register */
90 /* Increase the chance of trapping if random jumps occur to this
91 address, fill from 0xf0 to 0x100 */
96 /* This address must remain fixed at 0x100 for glibc's syscalls to work */
99 gate .+8, %r0 /* become privileged */
100 mtsp %r0,%sr4 /* get kernel space into sr4 */
101 mtsp %r0,%sr5 /* get kernel space into sr5 */
102 mtsp %r0,%sr6 /* get kernel space into sr6 */
103 mfsp %sr7,%r1 /* save user sr7 */
104 mtsp %r1,%sr3 /* and store it in sr3 */
107 /* for now we can *always* set the W bit on entry to the syscall
108 * since we don't support wide userland processes. We could
109 * also save the current SM other than in r0 and restore it on
110 * exit from the syscall, and also use that value to know
111 * whether to do narrow or wide syscalls. -PB
114 extrd,u %r1,PSW_W_BIT,1,%r1
115 /* sp must be aligned on 4, so deposit the W bit setting into
116 * the bottom of sp temporarily */
119 /* The top halves of argument registers must be cleared on syscall
120 * entry from narrow executable.
122 depdi 0, 31, 32, %r26
123 depdi 0, 31, 32, %r25
124 depdi 0, 31, 32, %r24
125 depdi 0, 31, 32, %r23
126 depdi 0, 31, 32, %r22
127 depdi 0, 31, 32, %r21
131 xor %r1,%r30,%r30 /* ye olde xor trick */
135 ldo THREAD_SZ_ALGN+FRAME_SIZE(%r30),%r30 /* set up kernel stack */
137 /* N.B.: It is critical that we don't set sr7 to 0 until r30
138 * contains a valid kernel stack pointer. It is also
139 * critical that we don't start using the kernel stack
140 * until after sr7 has been set to 0.
143 mtsp %r0,%sr7 /* get kernel space into sr7 */
144 STREGM %r1,FRAME_SIZE(%r30) /* save r1 (usp) here for now */
145 mfctl %cr30,%r1 /* get task ptr in %r1 */
146 LDREG TI_TASK(%r1),%r1
148 /* Save some registers for sigcontext and potential task
149 switch (see entry.S for the details of which ones are
150 saved/restored). TASK_PT_PSW is zeroed so we can see whether
151 a process is on a syscall or not. For an interrupt the real
152 PSW value is stored. This is needed for gdb and sys_ptrace. */
153 STREG %r0, TASK_PT_PSW(%r1)
154 STREG %r2, TASK_PT_GR2(%r1) /* preserve rp */
155 STREG %r19, TASK_PT_GR19(%r1)
157 LDREGM -FRAME_SIZE(%r30), %r2 /* get users sp back */
159 extrd,u %r2,63,1,%r19 /* W hidden in bottom bit */
161 xor %r19,%r2,%r2 /* clear bottom bit */
163 std %r19,TASK_PT_PSW(%r1)
166 STREG %r2, TASK_PT_GR30(%r1) /* ... and save it */
168 STREG %r20, TASK_PT_GR20(%r1)
169 STREG %r21, TASK_PT_GR21(%r1)
170 STREG %r22, TASK_PT_GR22(%r1)
171 STREG %r23, TASK_PT_GR23(%r1) /* 4th argument */
172 STREG %r24, TASK_PT_GR24(%r1) /* 3rd argument */
173 STREG %r25, TASK_PT_GR25(%r1) /* 2nd argument */
174 STREG %r26, TASK_PT_GR26(%r1) /* 1st argument */
175 STREG %r27, TASK_PT_GR27(%r1) /* user dp */
176 STREG %r28, TASK_PT_GR28(%r1) /* return value 0 */
177 STREG %r28, TASK_PT_ORIG_R28(%r1) /* return value 0 (saved for signals) */
178 STREG %r29, TASK_PT_GR29(%r1) /* return value 1 */
179 STREG %r31, TASK_PT_GR31(%r1) /* preserve syscall return ptr */
181 ldo TASK_PT_FR0(%r1), %r27 /* save fpregs from the kernel */
182 save_fp %r27 /* or potential task switch */
184 mfctl %cr11, %r27 /* i.e. SAR */
185 STREG %r27, TASK_PT_SAR(%r1)
190 ldo -16(%r30),%r29 /* Reference param save area */
191 copy %r19,%r2 /* W bit back to r2 */
193 /* no need to save these on stack in wide mode because the first 8
194 * args are passed in registers */
195 stw %r22, -52(%r30) /* 5th argument */
196 stw %r21, -56(%r30) /* 6th argument */
199 /* Are we being ptraced? */
201 LDREG TI_TASK(%r1),%r1
202 LDREG TASK_PTRACE(%r1), %r1
203 bb,<,n %r1,31,.Ltracesys
205 /* Note! We cannot use the syscall table that is mapped
206 nearby since the gateway page is mapped execute-only. */
209 ldil L%sys_call_table, %r1
211 addil L%(sys_call_table64-sys_call_table), %r1
212 ldo R%sys_call_table(%r1), %r19
214 ldo R%sys_call_table64(%r1), %r19
216 ldil L%sys_call_table, %r1
217 ldo R%sys_call_table(%r1), %r19
219 comiclr,>>= __NR_Linux_syscalls, %r20, %r0
222 LDREGX %r20(%r19), %r19
224 /* If this is a sys_rt_sigreturn call, and the signal was received
225 * when not in_syscall, then we want to return via syscall_exit_rfi,
226 * not syscall_exit. Signal no. in r20, in_syscall in r25 (see
227 * trampoline code in signal.c).
229 ldi __NR_rt_sigreturn,%r2
230 comb,= %r2,%r20,.Lrt_sigreturn
232 ldil L%syscall_exit,%r2
234 ldo R%syscall_exit(%r2),%r2
236 comib,<> 0,%r25,.Lin_syscall
237 ldil L%syscall_exit_rfi,%r2
239 ldo R%syscall_exit_rfi(%r2),%r2
241 /* Note! Because we are not running where we were linked, any
242 calls to functions external to this file must be indirect. To
243 be safe, we apply the opposite rule to functions within this
244 file, with local labels given to them to ensure correctness. */
248 ldil L%syscall_exit,%r1
249 be R%syscall_exit(%sr7,%r1)
250 ldo -ENOSYS(%r0),%r28 /* set errno */
253 /* Warning! This trace code is a virtual duplicate of the code above so be
254 * sure to maintain both! */
257 /* Need to save more registers so the debugger can see where we
258 * are. This saves only the lower 8 bits of PSW, so that the C
259 * bit is still clear on syscalls, and the D bit is set if this
260 * full register save path has been executed. We check the D
261 * bit on syscall_return_rfi to determine which registers to
262 * restore. An interrupt results in a full PSW saved with the
263 * C bit set, a non-straced syscall entry results in C and D clear
266 ldo -THREAD_SZ_ALGN-FRAME_SIZE(%r30),%r1 /* get task ptr */
267 LDREG TI_TASK(%r1), %r1
269 STREG %r2,TASK_PT_PSW(%r1) /* Lower 8 bits only!! */
271 STREG %r2,TASK_PT_SR0(%r1)
273 STREG %r2,TASK_PT_SR1(%r1)
275 STREG %r2,TASK_PT_SR2(%r1)
277 STREG %r2,TASK_PT_SR3(%r1)
278 STREG %r2,TASK_PT_SR4(%r1)
279 STREG %r2,TASK_PT_SR5(%r1)
280 STREG %r2,TASK_PT_SR6(%r1)
281 STREG %r2,TASK_PT_SR7(%r1)
282 STREG %r2,TASK_PT_IASQ0(%r1)
283 STREG %r2,TASK_PT_IASQ1(%r1)
284 LDREG TASK_PT_GR31(%r1),%r2
285 STREG %r2,TASK_PT_IAOQ0(%r1)
287 STREG %r2,TASK_PT_IAOQ1(%r1)
288 ldo TASK_REGS(%r1),%r2
290 STREG %r3,PT_GR3(%r2)
291 STREG %r4,PT_GR4(%r2)
292 STREG %r5,PT_GR5(%r2)
293 STREG %r6,PT_GR6(%r2)
294 STREG %r7,PT_GR7(%r2)
295 STREG %r8,PT_GR8(%r2)
296 STREG %r9,PT_GR9(%r2)
297 STREG %r10,PT_GR10(%r2)
298 STREG %r11,PT_GR11(%r2)
299 STREG %r12,PT_GR12(%r2)
300 STREG %r13,PT_GR13(%r2)
301 STREG %r14,PT_GR14(%r2)
302 STREG %r15,PT_GR15(%r2)
303 STREG %r16,PT_GR16(%r2)
304 STREG %r17,PT_GR17(%r2)
305 STREG %r18,PT_GR18(%r2)
306 /* Finished saving things for the debugger */
308 ldil L%syscall_trace,%r1
309 ldil L%tracesys_next,%r2
310 be R%syscall_trace(%sr7,%r1)
311 ldo R%tracesys_next(%r2),%r2
314 ldil L%sys_call_table,%r1
315 ldo R%sys_call_table(%r1), %r19
317 ldo -THREAD_SZ_ALGN-FRAME_SIZE(%r30),%r1 /* get task ptr */
318 LDREG TI_TASK(%r1), %r1
319 LDREG TASK_PT_GR20(%r1), %r20
320 LDREG TASK_PT_GR26(%r1), %r26 /* Restore the users args */
321 LDREG TASK_PT_GR25(%r1), %r25
322 LDREG TASK_PT_GR24(%r1), %r24
323 LDREG TASK_PT_GR23(%r1), %r23
325 LDREG TASK_PT_GR22(%r1), %r22
326 LDREG TASK_PT_GR21(%r1), %r21
327 ldo -16(%r30),%r29 /* Reference param save area */
330 comiclr,>>= __NR_Linux_syscalls, %r20, %r0
333 LDREGX %r20(%r19), %r19
335 /* If this is a sys_rt_sigreturn call, and the signal was received
336 * when not in_syscall, then we want to return via syscall_exit_rfi,
337 * not syscall_exit. Signal no. in r20, in_syscall in r25 (see
338 * trampoline code in signal.c).
340 ldi __NR_rt_sigreturn,%r2
341 comb,= %r2,%r20,.Ltrace_rt_sigreturn
343 ldil L%tracesys_exit,%r2
345 ldo R%tracesys_exit(%r2),%r2
347 /* Do *not* call this function on the gateway page, because it
348 makes a direct call to syscall_trace. */
351 ldo -THREAD_SZ_ALGN-FRAME_SIZE(%r30),%r1 /* get task ptr */
352 LDREG TI_TASK(%r1), %r1
354 ldo -16(%r30),%r29 /* Reference param save area */
356 bl syscall_trace, %r2
357 STREG %r28,TASK_PT_GR28(%r1) /* save return value now */
358 ldo -THREAD_SZ_ALGN-FRAME_SIZE(%r30),%r1 /* get task ptr */
359 LDREG TI_TASK(%r1), %r1
360 LDREG TASK_PT_GR28(%r1), %r28 /* Restore return val. */
362 ldil L%syscall_exit,%r1
363 be,n R%syscall_exit(%sr7,%r1)
365 .Ltrace_rt_sigreturn:
366 comib,<> 0,%r25,.Ltrace_in_syscall
367 ldil L%tracesys_sigexit,%r2
369 ldo R%tracesys_sigexit(%r2),%r2
372 ldo -THREAD_SZ_ALGN-FRAME_SIZE(%r30),%r1 /* get task ptr */
375 ldo -16(%r30),%r29 /* Reference param save area */
377 bl syscall_trace, %r2
380 ldil L%syscall_exit_rfi,%r1
381 be,n R%syscall_exit_rfi(%sr7,%r1)
384 /*********************************************************
385 Light-weight-syscall code
388 r26,r25,r24,r23,r22 - Input registers
389 r28 - Function return register
392 Scracth: Any of the above that aren't being
393 currently used, including r1.
395 Return pointer: r31 (Not usable)
397 Error codes returned by entry path:
399 ENOSYS - r20 was an invalid LWS number.
401 *********************************************************/
403 /* Gate and ensure we return to userspace */
405 depi 3, 31, 2, %r31 /* Ensure we return to userspace */
408 /* FIXME: If we are a 64-bit kernel just
409 * turn this on unconditionally.
412 extrd,u %r1,PSW_W_BIT,1,%r1
413 /* sp must be aligned on 4, so deposit the W bit setting into
414 * the bottom of sp temporarily */
417 /* Clip LWS number to a 32-bit value always */
418 depdi 0, 31, 32, %r20
421 /* Is the lws entry number valid? */
422 comiclr,>>= __NR_lws_entries, %r20, %r0
425 /* WARNING: Trashing sr2 and sr3 */
426 mfsp %sr7,%r1 /* get userspace into sr3 */
428 mtsp %r0,%sr2 /* get kernel space into sr2 */
430 /* Load table start */
431 ldil L%lws_table, %r1
432 ldo R%lws_table(%r1), %r28 /* Scratch use of r28 */
433 LDREGX %r20(%sr2,r28), %r21 /* Scratch use of r21 */
435 /* Jump to lws, lws table pointers already relocated */
439 ldo -ENOSYS(%r0),%r21 /* set errno */
440 /* Fall through: Return to userspace */
444 /* decide whether to reset the wide mode bit
446 * For a syscall, the W bit is stored in the lowest bit
447 * of sp. Extract it and reset W if it is zero */
448 extrd,u,*<> %r30,63,1,%r1
450 /* now reset the lowest bit of sp if it was set */
457 /***************************************************
458 Implementing CAS as an atomic operation:
460 %r26 - Address to examine
461 %r25 - Old value to check (old)
462 %r24 - New value to set (new)
463 %r28 - Return prev through this register.
464 %r21 - Kernel error code
466 If debugging is DISabled:
468 %r21 has the following meanings:
470 EAGAIN - CAS is busy, ldcw failed, try again.
471 EFAULT - Read or write failed.
473 If debugging is enabled:
475 EDEADLOCK - CAS called recursively.
476 EAGAIN && r28 == 1 - CAS is busy. Lock contended.
477 EAGAIN && r28 == 2 - CAS is busy. ldcw failed.
478 EFAULT - Read or write failed.
480 Scratch: r20, r28, r1
482 ****************************************************/
484 /* Do not enable LWS debugging */
485 #define ENABLE_LWS_DEBUG 0
487 /* ELF64 Process entry path */
488 lws_compare_and_swap64:
490 b,n lws_compare_and_swap
492 /* If we are not a 64-bit kernel, then we don't
493 * implement having 64-bit input registers
498 /* ELF32 Process entry path */
499 lws_compare_and_swap32:
501 /* Clip all the input registers */
502 depdi 0, 31, 32, %r26
503 depdi 0, 31, 32, %r25
504 depdi 0, 31, 32, %r24
507 lws_compare_and_swap:
509 /* Load start of lock table */
510 ldil L%lws_lock_start, %r20
511 ldo R%lws_lock_start(%r20), %r28
513 /* Extract four bits from r26 and hash lock (Bits 4-7) */
514 extru %r26, 27, 4, %r20
516 /* Find lock to use, the hash is either one of 0 to
517 15, multiplied by 16 (keep it 16-byte aligned)
518 and add to the lock table offset. */
522 # ifdef ENABLE_LWS_DEBUG
524 DEBUG, check for deadlock!
525 If the thread register values are the same
526 then we were the one that locked it last and
527 this is a recurisve call that will deadlock.
528 We *must* giveup this call and fail.
530 ldw 4(%sr2,%r20), %r28 /* Load thread register */
531 mfctl %cr27, %r21 /* Get current thread register */
532 cmpb,<>,n %r21, %r28, cas_lock /* Called recursive? */
533 b lws_exit /* Return error! */
534 ldo -EDEADLOCK(%r0), %r21
536 cmpb,=,n %r0, %r28, cas_nocontend /* Is nobody using it? */
537 ldo 1(%r0), %r28 /* 1st case */
538 b lws_exit /* Contended... */
539 ldo -EAGAIN(%r0), %r21 /* Spin in userspace */
542 /* ENABLE_LWS_DEBUG */
544 ldcw 0(%sr2,%r20), %r28 /* Try to acquire the lock */
545 cmpb,<>,n %r0, %r28, cas_action /* Did we get it? */
547 ldo 2(%r0), %r28 /* 2nd case */
548 b lws_exit /* Contended... */
549 ldo -EAGAIN(%r0), %r21 /* Spin in userspace */
561 This all works becuse intr_do_signal
562 and schedule both check the return iasq
563 and see that we are on the kernel page
564 so this process is never scheduled off
565 or is ever sent any signal of any sort,
566 thus it is wholly atomic from usrspaces
570 #if defined CONFIG_SMP && defined ENABLE_LWS_DEBUG
573 stw %r1, 4(%sr2,%r20)
575 /* The load and store could fail */
576 1: ldw 0(%sr3,%r26), %r28
577 sub,<> %r28, %r25, %r0
578 2: stw %r24, 0(%sr3,%r26)
581 stw %r20, 0(%sr2,%r20)
582 # ifdef ENABLE_LWS_DEBUG
583 /* Clear thread register indicator */
584 stw %r0, 4(%sr2,%r20)
587 /* Return to userspace, set no error */
592 /* Error occured on load or store */
595 stw %r20, 0(%sr2,%r20)
596 # ifdef ENABLE_LWS_DEBUG
597 stw %r0, 4(%sr2,%r20)
601 ldo -EFAULT(%r0),%r21 /* set errno */
607 /* Two exception table entries, one for the load,
608 the other for the store. Either return -EFAULT.
609 Each of the entries must be relocated. */
610 .section __ex_table,"aw"
612 /* Pad the address calculation */
613 .word 0,(2b - linux_gateway_page)
614 .word 0,(3b - linux_gateway_page)
616 .word (2b - linux_gateway_page)
617 .word (3b - linux_gateway_page)
621 .section __ex_table,"aw"
623 /* Pad the address calculation */
624 .word 0,(1b - linux_gateway_page)
625 .word 0,(3b - linux_gateway_page)
627 .word (1b - linux_gateway_page)
628 .word (3b - linux_gateway_page)
632 end_compare_and_swap:
634 /* Make sure nothing else is placed on this page */
636 .export end_linux_gateway_page
637 end_linux_gateway_page:
639 /* Relocate symbols assuming linux_gateway_page is mapped
640 to virtual address 0x0 */
642 /* FIXME: The code will always be on the gateay page
643 and thus it will be on the first 4k, the
644 assembler seems to think that the final
645 subtraction result is only a word in
646 length, so we pad the value.
648 #define LWS_ENTRY(_name_) .word 0,(lws_##_name_ - linux_gateway_page)
650 #define LWS_ENTRY(_name_) .word (lws_##_name_ - linux_gateway_page)
654 /* Light-weight-syscall table */
655 /* Start of lws table. */
659 LWS_ENTRY(compare_and_swap32) /* 0 - ELF32 Atomic compare and swap */
660 LWS_ENTRY(compare_and_swap64) /* 1 - ELF64 Atomic compare and swap */
661 /* End of lws table */
664 .export sys_call_table
667 #include "syscall_table.S"
671 .export sys_call_table64
674 #define SYSCALL_TABLE_64BIT
675 #include "syscall_table.S"
680 All light-weight-syscall atomic operations
681 will use this set of locks
685 .export lws_lock_start
691 /* Keep locks aligned at 16-bytes */
699 /* CONFIG_SMP for lws_lock_start */