1/* 2 * This file contains the code that gets mapped at the upper end of each task's text 3 * region. For now, it contains the signal trampoline code only. 4 * 5 * Copyright (C) 1999-2003 Hewlett-Packard Co 6 * David Mosberger-Tang <davidm@hpl.hp.com> 7 */ 8 9 10#include <asm/asmmacro.h> 11#include <asm/errno.h> 12#include <asm/asm-offsets.h> 13#include <asm/sigcontext.h> 14#include <asm/unistd.h> 15#include <asm/kregs.h> 16#include <asm/page.h> 17#include "paravirt_inst.h" 18 19/* 20 * We can't easily refer to symbols inside the kernel. To avoid full runtime relocation, 21 * complications with the linker (which likes to create PLT stubs for branches 22 * to targets outside the shared object) and to avoid multi-phase kernel builds, we 23 * simply create minimalistic "patch lists" in special ELF sections. 24 */ 25 .section ".data..patch.fsyscall_table", "a" 26 .previous 27#define LOAD_FSYSCALL_TABLE(reg) \ 28[1:] movl reg=0; \ 29 .xdata4 ".data..patch.fsyscall_table", 1b-. 30 31 .section ".data..patch.brl_fsys_bubble_down", "a" 32 .previous 33#define BRL_COND_FSYS_BUBBLE_DOWN(pr) \ 34[1:](pr)brl.cond.sptk 0; \ 35 ;; \ 36 .xdata4 ".data..patch.brl_fsys_bubble_down", 1b-. 37 38GLOBAL_ENTRY(__kernel_syscall_via_break) 39 .prologue 40 .altrp b6 41 .body 42 /* 43 * Note: for (fast) syscall restart to work, the break instruction must be 44 * the first one in the bundle addressed by syscall_via_break. 45 */ 46{ .mib 47 break 0x100000 48 nop.i 0 49 br.ret.sptk.many b6 50} 51END(__kernel_syscall_via_break) 52 53# define ARG0_OFF (16 + IA64_SIGFRAME_ARG0_OFFSET) 54# define ARG1_OFF (16 + IA64_SIGFRAME_ARG1_OFFSET) 55# define ARG2_OFF (16 + IA64_SIGFRAME_ARG2_OFFSET) 56# define SIGHANDLER_OFF (16 + IA64_SIGFRAME_HANDLER_OFFSET) 57# define SIGCONTEXT_OFF (16 + IA64_SIGFRAME_SIGCONTEXT_OFFSET) 58 59# define FLAGS_OFF IA64_SIGCONTEXT_FLAGS_OFFSET 60# define CFM_OFF IA64_SIGCONTEXT_CFM_OFFSET 61# define FR6_OFF IA64_SIGCONTEXT_FR6_OFFSET 62# define BSP_OFF IA64_SIGCONTEXT_AR_BSP_OFFSET 63# define RNAT_OFF IA64_SIGCONTEXT_AR_RNAT_OFFSET 64# define UNAT_OFF IA64_SIGCONTEXT_AR_UNAT_OFFSET 65# define FPSR_OFF IA64_SIGCONTEXT_AR_FPSR_OFFSET 66# define PR_OFF IA64_SIGCONTEXT_PR_OFFSET 67# define RP_OFF IA64_SIGCONTEXT_IP_OFFSET 68# define SP_OFF IA64_SIGCONTEXT_R12_OFFSET 69# define RBS_BASE_OFF IA64_SIGCONTEXT_RBS_BASE_OFFSET 70# define LOADRS_OFF IA64_SIGCONTEXT_LOADRS_OFFSET 71# define base0 r2 72# define base1 r3 73 /* 74 * When we get here, the memory stack looks like this: 75 * 76 * +===============================+ 77 * | | 78 * // struct sigframe // 79 * | | 80 * +-------------------------------+ <-- sp+16 81 * | 16 byte of scratch | 82 * | space | 83 * +-------------------------------+ <-- sp 84 * 85 * The register stack looks _exactly_ the way it looked at the time the signal 86 * occurred. In other words, we're treading on a potential mine-field: each 87 * incoming general register may be a NaT value (including sp, in which case the 88 * process ends up dying with a SIGSEGV). 89 * 90 * The first thing need to do is a cover to get the registers onto the backing 91 * store. Once that is done, we invoke the signal handler which may modify some 92 * of the machine state. After returning from the signal handler, we return 93 * control to the previous context by executing a sigreturn system call. A signal 94 * handler may call the rt_sigreturn() function to directly return to a given 95 * sigcontext. However, the user-level sigreturn() needs to do much more than 96 * calling the rt_sigreturn() system call as it needs to unwind the stack to 97 * restore preserved registers that may have been saved on the signal handler's 98 * call stack. 99 */ 100 101#define SIGTRAMP_SAVES \ 102 .unwabi 3, 's'; /* mark this as a sigtramp handler (saves scratch regs) */ \ 103 .unwabi @svr4, 's'; /* backwards compatibility with old unwinders (remove in v2.7) */ \ 104 .savesp ar.unat, UNAT_OFF+SIGCONTEXT_OFF; \ 105 .savesp ar.fpsr, FPSR_OFF+SIGCONTEXT_OFF; \ 106 .savesp pr, PR_OFF+SIGCONTEXT_OFF; \ 107 .savesp rp, RP_OFF+SIGCONTEXT_OFF; \ 108 .savesp ar.pfs, CFM_OFF+SIGCONTEXT_OFF; \ 109 .vframesp SP_OFF+SIGCONTEXT_OFF 110 111GLOBAL_ENTRY(__kernel_sigtramp) 112 // describe the state that is active when we get here: 113 .prologue 114 SIGTRAMP_SAVES 115 .body 116 117 .label_state 1 118 119 adds base0=SIGHANDLER_OFF,sp 120 adds base1=RBS_BASE_OFF+SIGCONTEXT_OFF,sp 121 br.call.sptk.many rp=1f 1221: 123 ld8 r17=[base0],(ARG0_OFF-SIGHANDLER_OFF) // get pointer to signal handler's plabel 124 ld8 r15=[base1] // get address of new RBS base (or NULL) 125 cover // push args in interrupted frame onto backing store 126 ;; 127 cmp.ne p1,p0=r15,r0 // do we need to switch rbs? (note: pr is saved by kernel) 128 mov.m r9=ar.bsp // fetch ar.bsp 129 .spillsp.p p1, ar.rnat, RNAT_OFF+SIGCONTEXT_OFF 130(p1) br.cond.spnt setup_rbs // yup -> (clobbers p8, r14-r16, and r18-r20) 131back_from_setup_rbs: 132 alloc r8=ar.pfs,0,0,3,0 133 ld8 out0=[base0],16 // load arg0 (signum) 134 adds base1=(ARG1_OFF-(RBS_BASE_OFF+SIGCONTEXT_OFF)),base1 135 ;; 136 ld8 out1=[base1] // load arg1 (siginfop) 137 ld8 r10=[r17],8 // get signal handler entry point 138 ;; 139 ld8 out2=[base0] // load arg2 (sigcontextp) 140 ld8 gp=[r17] // get signal handler's global pointer 141 adds base0=(BSP_OFF+SIGCONTEXT_OFF),sp 142 ;; 143 .spillsp ar.bsp, BSP_OFF+SIGCONTEXT_OFF 144 st8 [base0]=r9 // save sc_ar_bsp 145 adds base0=(FR6_OFF+SIGCONTEXT_OFF),sp 146 adds base1=(FR6_OFF+16+SIGCONTEXT_OFF),sp 147 ;; 148 stf.spill [base0]=f6,32 149 stf.spill [base1]=f7,32 150 ;; 151 stf.spill [base0]=f8,32 152 stf.spill [base1]=f9,32 153 mov b6=r10 154 ;; 155 stf.spill [base0]=f10,32 156 stf.spill [base1]=f11,32 157 ;; 158 stf.spill [base0]=f12,32 159 stf.spill [base1]=f13,32 160 ;; 161 stf.spill [base0]=f14,32 162 stf.spill [base1]=f15,32 163 br.call.sptk.many rp=b6 // call the signal handler 164.ret0: adds base0=(BSP_OFF+SIGCONTEXT_OFF),sp 165 ;; 166 ld8 r15=[base0] // fetch sc_ar_bsp 167 mov r14=ar.bsp 168 ;; 169 cmp.ne p1,p0=r14,r15 // do we need to restore the rbs? 170(p1) br.cond.spnt restore_rbs // yup -> (clobbers r14-r18, f6 & f7) 171 ;; 172back_from_restore_rbs: 173 adds base0=(FR6_OFF+SIGCONTEXT_OFF),sp 174 adds base1=(FR6_OFF+16+SIGCONTEXT_OFF),sp 175 ;; 176 ldf.fill f6=[base0],32 177 ldf.fill f7=[base1],32 178 ;; 179 ldf.fill f8=[base0],32 180 ldf.fill f9=[base1],32 181 ;; 182 ldf.fill f10=[base0],32 183 ldf.fill f11=[base1],32 184 ;; 185 ldf.fill f12=[base0],32 186 ldf.fill f13=[base1],32 187 ;; 188 ldf.fill f14=[base0],32 189 ldf.fill f15=[base1],32 190 mov r15=__NR_rt_sigreturn 191 .restore sp // pop .prologue 192 break __BREAK_SYSCALL 193 194 .prologue 195 SIGTRAMP_SAVES 196setup_rbs: 197 mov ar.rsc=0 // put RSE into enforced lazy mode 198 ;; 199 .save ar.rnat, r19 200 mov r19=ar.rnat // save RNaT before switching backing store area 201 adds r14=(RNAT_OFF+SIGCONTEXT_OFF),sp 202 203 mov r18=ar.bspstore 204 mov ar.bspstore=r15 // switch over to new register backing store area 205 ;; 206 207 .spillsp ar.rnat, RNAT_OFF+SIGCONTEXT_OFF 208 st8 [r14]=r19 // save sc_ar_rnat 209 .body 210 mov.m r16=ar.bsp // sc_loadrs <- (new bsp - new bspstore) << 16 211 adds r14=(LOADRS_OFF+SIGCONTEXT_OFF),sp 212 ;; 213 invala 214 sub r15=r16,r15 215 extr.u r20=r18,3,6 216 ;; 217 mov ar.rsc=0xf // set RSE into eager mode, pl 3 218 cmp.eq p8,p0=63,r20 219 shl r15=r15,16 220 ;; 221 st8 [r14]=r15 // save sc_loadrs 222(p8) st8 [r18]=r19 // if bspstore points at RNaT slot, store RNaT there now 223 .restore sp // pop .prologue 224 br.cond.sptk back_from_setup_rbs 225 226 .prologue 227 SIGTRAMP_SAVES 228 .spillsp ar.rnat, RNAT_OFF+SIGCONTEXT_OFF 229 .body 230restore_rbs: 231 // On input: 232 // r14 = bsp1 (bsp at the time of return from signal handler) 233 // r15 = bsp0 (bsp at the time the signal occurred) 234 // 235 // Here, we need to calculate bspstore0, the value that ar.bspstore needs 236 // to be set to, based on bsp0 and the size of the dirty partition on 237 // the alternate stack (sc_loadrs >> 16). This can be done with the 238 // following algorithm: 239 // 240 // bspstore0 = rse_skip_regs(bsp0, -rse_num_regs(bsp1 - (loadrs >> 19), bsp1)); 241 // 242 // This is what the code below does. 243 // 244 alloc r2=ar.pfs,0,0,0,0 // alloc null frame 245 adds r16=(LOADRS_OFF+SIGCONTEXT_OFF),sp 246 adds r18=(RNAT_OFF+SIGCONTEXT_OFF),sp 247 ;; 248 ld8 r17=[r16] 249 ld8 r16=[r18] // get new rnat 250 extr.u r18=r15,3,6 // r18 <- rse_slot_num(bsp0) 251 ;; 252 mov ar.rsc=r17 // put RSE into enforced lazy mode 253 shr.u r17=r17,16 254 ;; 255 sub r14=r14,r17 // r14 (bspstore1) <- bsp1 - (sc_loadrs >> 16) 256 shr.u r17=r17,3 // r17 <- (sc_loadrs >> 19) 257 ;; 258 loadrs // restore dirty partition 259 extr.u r14=r14,3,6 // r14 <- rse_slot_num(bspstore1) 260 ;; 261 add r14=r14,r17 // r14 <- rse_slot_num(bspstore1) + (sc_loadrs >> 19) 262 ;; 263 shr.u r14=r14,6 // r14 <- (rse_slot_num(bspstore1) + (sc_loadrs >> 19))/0x40 264 ;; 265 sub r14=r14,r17 // r14 <- -rse_num_regs(bspstore1, bsp1) 266 movl r17=0x8208208208208209 267 ;; 268 add r18=r18,r14 // r18 (delta) <- rse_slot_num(bsp0) - rse_num_regs(bspstore1,bsp1) 269 setf.sig f7=r17 270 cmp.lt p7,p0=r14,r0 // p7 <- (r14 < 0)? 271 ;; 272(p7) adds r18=-62,r18 // delta -= 62 273 ;; 274 setf.sig f6=r18 275 ;; 276 xmpy.h f6=f6,f7 277 ;; 278 getf.sig r17=f6 279 ;; 280 add r17=r17,r18 281 shr r18=r18,63 282 ;; 283 shr r17=r17,5 284 ;; 285 sub r17=r17,r18 // r17 = delta/63 286 ;; 287 add r17=r14,r17 // r17 <- delta/63 - rse_num_regs(bspstore1, bsp1) 288 ;; 289 shladd r15=r17,3,r15 // r15 <- bsp0 + 8*(delta/63 - rse_num_regs(bspstore1, bsp1)) 290 ;; 291 mov ar.bspstore=r15 // switch back to old register backing store area 292 ;; 293 mov ar.rnat=r16 // restore RNaT 294 mov ar.rsc=0xf // (will be restored later on from sc_ar_rsc) 295 // invala not necessary as that will happen when returning to user-mode 296 br.cond.sptk back_from_restore_rbs 297END(__kernel_sigtramp) 298 299/* 300 * On entry: 301 * r11 = saved ar.pfs 302 * r15 = system call # 303 * b0 = saved return address 304 * b6 = return address 305 * On exit: 306 * r11 = saved ar.pfs 307 * r15 = system call # 308 * b0 = saved return address 309 * all other "scratch" registers: undefined 310 * all "preserved" registers: same as on entry 311 */ 312 313GLOBAL_ENTRY(__kernel_syscall_via_epc) 314 .prologue 315 .altrp b6 316 .body 317{ 318 /* 319 * Note: the kernel cannot assume that the first two instructions in this 320 * bundle get executed. The remaining code must be safe even if 321 * they do not get executed. 322 */ 323 adds r17=-1024,r15 // A 324 mov r10=0 // A default to successful syscall execution 325 epc // B causes split-issue 326} 327 ;; 328 RSM_PSR_BE_I(r20, r22) // M2 (5 cyc to srlz.d) 329 LOAD_FSYSCALL_TABLE(r14) // X 330 ;; 331 mov r16=IA64_KR(CURRENT) // M2 (12 cyc) 332 shladd r18=r17,3,r14 // A 333 mov r19=NR_syscalls-1 // A 334 ;; 335 lfetch [r18] // M0|1 336 MOV_FROM_PSR(p0, r29, r8) // M2 (12 cyc) 337 // If r17 is a NaT, p6 will be zero 338 cmp.geu p6,p7=r19,r17 // A (sysnr > 0 && sysnr < 1024+NR_syscalls)? 339 ;; 340 mov r21=ar.fpsr // M2 (12 cyc) 341 tnat.nz p10,p9=r15 // I0 342 mov.i r26=ar.pfs // I0 (would stall anyhow due to srlz.d...) 343 ;; 344 srlz.d // M0 (forces split-issue) ensure PSR.BE==0 345(p6) ld8 r18=[r18] // M0|1 346 nop.i 0 347 ;; 348 nop.m 0 349(p6) tbit.z.unc p8,p0=r18,0 // I0 (dual-issues with "mov b7=r18"!) 350 nop.i 0 351 ;; 352 SSM_PSR_I(p8, p14, r25) 353(p6) mov b7=r18 // I0 354(p8) br.dptk.many b7 // B 355 356 mov r27=ar.rsc // M2 (12 cyc) 357/* 358 * brl.cond doesn't work as intended because the linker would convert this branch 359 * into a branch to a PLT. Perhaps there will be a way to avoid this with some 360 * future version of the linker. In the meantime, we just use an indirect branch 361 * instead. 362 */ 363#ifdef CONFIG_ITANIUM 364(p6) add r14=-8,r14 // r14 <- addr of fsys_bubble_down entry 365 ;; 366(p6) ld8 r14=[r14] // r14 <- fsys_bubble_down 367 ;; 368(p6) mov b7=r14 369(p6) br.sptk.many b7 370#else 371 BRL_COND_FSYS_BUBBLE_DOWN(p6) 372#endif 373 SSM_PSR_I(p0, p14, r10) 374 mov r10=-1 375(p10) mov r8=EINVAL 376(p9) mov r8=ENOSYS 377 FSYS_RETURN 378 379#ifdef CONFIG_PARAVIRT 380 /* 381 * padd to make the size of this symbol constant 382 * independent of paravirtualization. 383 */ 384 .align PAGE_SIZE / 8 385#endif 386END(__kernel_syscall_via_epc) 387