xref: /openbmc/linux/arch/hexagon/kernel/vm_entry.S (revision 82e6fdd6)
1/*
2 * Event entry/exit for Hexagon
3 *
4 * Copyright (c) 2010-2013, The Linux Foundation. All rights reserved.
5 *
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 and
8 * only version 2 as published by the Free Software Foundation.
9 *
10 * This program is distributed in the hope that it will be useful,
11 * but WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
13 * GNU General Public License for more details.
14 *
15 * You should have received a copy of the GNU General Public License
16 * along with this program; if not, write to the Free Software
17 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA
18 * 02110-1301, USA.
19 */
20
21#include <asm/asm-offsets.h>  /*  assembly-safer versions of C defines */
22#include <asm/mem-layout.h>   /*  sigh, except for page_offset  */
23#include <asm/hexagon_vm.h>
24#include <asm/thread_info.h>
25
26/*
27 * Entry into guest-mode Linux under Hexagon Virtual Machine.
28 * Stack pointer points to event record - build pt_regs on top of it,
29 * set up a plausible C stack frame, and dispatch to the C handler.
30 * On return, do vmrte virtual instruction with SP where we started.
31 *
32 * VM Spec 0.5 uses a trap to fetch HVM record now.
33 */
34
35/*
36 * Save full register state, while setting up thread_info struct
37 * pointer derived from kernel stack pointer in THREADINFO_REG
38 * register, putting prior thread_info.regs pointer in a callee-save
39 * register (R24, which had better not ever be assigned to THREADINFO_REG),
40 * and updating thread_info.regs to point to current stack frame,
41 * so as to support nested events in kernel mode.
42 *
43 * As this is common code, we set the pt_regs system call number
44 * to -1 for all events.  It will be replaced with the system call
45 * number in the case where we decode a system call (trap0(#1)).
46 */
47
48#if CONFIG_HEXAGON_ARCH_VERSION < 4
49#define save_pt_regs()\
50 memd(R0 + #_PT_R3130) = R31:30; \
51 { memw(R0 + #_PT_R2928) = R28; \
52   R31 = memw(R0 + #_PT_ER_VMPSP); }\
53 { memw(R0 + #(_PT_R2928 + 4)) = R31; \
54   R31 = ugp; } \
55 { memd(R0 + #_PT_R2726) = R27:26; \
56   R30 = gp ; } \
57 memd(R0 + #_PT_R2524) = R25:24; \
58 memd(R0 + #_PT_R2322) = R23:22; \
59 memd(R0 + #_PT_R2120) = R21:20; \
60 memd(R0 + #_PT_R1918) = R19:18; \
61 memd(R0 + #_PT_R1716) = R17:16; \
62 memd(R0 + #_PT_R1514) = R15:14; \
63 memd(R0 + #_PT_R1312) = R13:12; \
64 { memd(R0 + #_PT_R1110) = R11:10; \
65   R15 = lc0; } \
66 { memd(R0 + #_PT_R0908) = R9:8; \
67   R14 = sa0; } \
68 { memd(R0 + #_PT_R0706) = R7:6; \
69   R13 = lc1; } \
70 { memd(R0 + #_PT_R0504) = R5:4; \
71   R12 = sa1; } \
72 { memd(R0 + #_PT_GPUGP) = R31:30; \
73   R11 = m1; \
74   R2.H = #HI(_THREAD_SIZE); } \
75 { memd(R0 + #_PT_LC0SA0) = R15:14; \
76   R10 = m0; \
77   R2.L = #LO(_THREAD_SIZE); } \
78 { memd(R0 + #_PT_LC1SA1) = R13:12; \
79   R15 = p3:0; \
80   R2 = neg(R2); } \
81 { memd(R0 + #_PT_M1M0) = R11:10; \
82   R14  = usr; \
83   R2 = and(R0,R2); } \
84 { memd(R0 + #_PT_PREDSUSR) =  R15:14; \
85   THREADINFO_REG = R2; } \
86 { r24 = memw(THREADINFO_REG + #_THREAD_INFO_PT_REGS); \
87   memw(THREADINFO_REG + #_THREAD_INFO_PT_REGS) = R0; \
88   R2 = #-1; } \
89 { memw(R0 + #_PT_SYSCALL_NR) = R2; \
90   R30 = #0; }
91#else
92/* V4+ */
93/* the # ## # syntax inserts a literal ## */
94#define save_pt_regs()\
95	{ memd(R0 + #_PT_R3130) = R31:30; \
96		R30 = memw(R0 + #_PT_ER_VMPSP); }\
97	{ memw(R0 + #_PT_R2928) = R28; \
98		memw(R0 + #(_PT_R2928 + 4)) = R30; }\
99	{ R31:30 = C11:10; \
100		memd(R0 + #_PT_R2726) = R27:26; \
101		memd(R0 + #_PT_R2524) = R25:24; }\
102	{ memd(R0 + #_PT_R2322) = R23:22; \
103		memd(R0 + #_PT_R2120) = R21:20; }\
104	{ memd(R0 + #_PT_R1918) = R19:18; \
105		memd(R0 + #_PT_R1716) = R17:16; }\
106	{ memd(R0 + #_PT_R1514) = R15:14; \
107		memd(R0 + #_PT_R1312) = R13:12; \
108		R17:16 = C13:12; }\
109	{ memd(R0 + #_PT_R1110) = R11:10; \
110		memd(R0 + #_PT_R0908) = R9:8; \
111	  R15:14 = C1:0; } \
112	{ memd(R0 + #_PT_R0706) = R7:6; \
113		memd(R0 + #_PT_R0504) = R5:4; \
114    R13:12 = C3:2; } \
115	{ memd(R0 + #_PT_GPUGP) = R31:30; \
116		memd(R0 + #_PT_LC0SA0) = R15:14; \
117	  R11:10 = C7:6; }\
118	{	THREADINFO_REG = and(R0, # ## #-_THREAD_SIZE); \
119		memd(R0 + #_PT_LC1SA1) = R13:12; \
120	  R15 = p3:0; }\
121	{ memd(R0 + #_PT_M1M0) = R11:10; \
122		memw(R0 + #_PT_PREDSUSR + 4) =  R15; }\
123	{ r24 = memw(THREADINFO_REG + #_THREAD_INFO_PT_REGS); \
124	  memw(THREADINFO_REG + #_THREAD_INFO_PT_REGS) = R0; \
125	  R2 = #-1; } \
126	{ memw(R0 + #_PT_SYSCALL_NR) = R2; \
127		memd(R0 + #_PT_CS1CS0) = R17:16; \
128	  R30 = #0; }
129#endif
130
131/*
132 * Restore registers and thread_info.regs state. THREADINFO_REG
133 * is assumed to still be sane, and R24 to have been correctly
134 * preserved. Don't restore R29 (SP) until later.
135 */
136
137#if CONFIG_HEXAGON_ARCH_VERSION < 4
138#define restore_pt_regs() \
139	{ memw(THREADINFO_REG + #_THREAD_INFO_PT_REGS) = R24; \
140	  R15:14 = memd(R0 + #_PT_PREDSUSR); } \
141	{ R11:10 = memd(R0 + #_PT_M1M0); \
142	  p3:0 = R15; } \
143	{ R13:12 = memd(R0 + #_PT_LC1SA1); \
144	  usr = R14; } \
145	{ R15:14 = memd(R0 + #_PT_LC0SA0); \
146	  m1 = R11; } \
147	{ R3:2 = memd(R0 + #_PT_R0302); \
148	  m0 = R10; } \
149	{ R5:4 = memd(R0 + #_PT_R0504); \
150	  lc1 = R13; } \
151	{ R7:6 = memd(R0 + #_PT_R0706); \
152	  sa1 = R12; } \
153	{ R9:8 = memd(R0 + #_PT_R0908); \
154	  lc0 = R15; } \
155	{ R11:10 = memd(R0 + #_PT_R1110); \
156	  sa0 = R14; } \
157	{ R13:12 = memd(R0 + #_PT_R1312); \
158	  R15:14 = memd(R0 + #_PT_R1514); } \
159	{ R17:16 = memd(R0 + #_PT_R1716); \
160	  R19:18 = memd(R0 + #_PT_R1918); } \
161	{ R21:20 = memd(R0 + #_PT_R2120); \
162	  R23:22 = memd(R0 + #_PT_R2322); } \
163	{ R25:24 = memd(R0 + #_PT_R2524); \
164	  R27:26 = memd(R0 + #_PT_R2726); } \
165	R31:30 = memd(R0 + #_PT_GPUGP); \
166	{ R28 = memw(R0 + #_PT_R2928); \
167	  ugp = R31; } \
168	{ R31:30 = memd(R0 + #_PT_R3130); \
169	  gp = R30; }
170#else
171/* V4+ */
172#define restore_pt_regs() \
173	{ memw(THREADINFO_REG + #_THREAD_INFO_PT_REGS) = R24; \
174	  R15:14 = memd(R0 + #_PT_PREDSUSR); } \
175	{ R11:10 = memd(R0 + #_PT_M1M0); \
176		R13:12 = memd(R0 + #_PT_LC1SA1); \
177		p3:0 = R15; } \
178	{ R15:14 = memd(R0 + #_PT_LC0SA0); \
179		R3:2 = memd(R0 + #_PT_R0302); \
180		usr = R14; } \
181	{ R5:4 = memd(R0 + #_PT_R0504); \
182		R7:6 = memd(R0 + #_PT_R0706); \
183		C7:6 = R11:10; }\
184	{ R9:8 = memd(R0 + #_PT_R0908); \
185		R11:10 = memd(R0 + #_PT_R1110); \
186    C3:2 = R13:12; }\
187	{ R13:12 = memd(R0 + #_PT_R1312); \
188	  R15:14 = memd(R0 + #_PT_R1514); \
189		C1:0 = R15:14; }\
190	{ R17:16 = memd(R0 + #_PT_R1716); \
191	  R19:18 = memd(R0 + #_PT_R1918); } \
192	{ R21:20 = memd(R0 + #_PT_R2120); \
193	  R23:22 = memd(R0 + #_PT_R2322); } \
194	{ R25:24 = memd(R0 + #_PT_R2524); \
195	  R27:26 = memd(R0 + #_PT_R2726); } \
196	R31:30 = memd(R0 + #_PT_CS1CS0); \
197	{ C13:12 = R31:30; \
198		R31:30 = memd(R0 + #_PT_GPUGP) ; \
199		R28 = memw(R0 + #_PT_R2928); }\
200	{ C11:10 = R31:30; \
201		R31:30 = memd(R0 + #_PT_R3130); }
202#endif
203
204	/*
205	 * Clears off enough space for the rest of pt_regs; evrec is a part
206	 * of pt_regs in HVM mode.  Save R0/R1, set handler's address in R1.
207	 * R0 is the address of pt_regs and is the parameter to save_pt_regs.
208	 */
209
210/*
211 * Since the HVM isn't automagically pushing the EVREC onto the stack anymore,
212 * we'll subract the entire size out and then fill it in ourselves.
213 * Need to save off R0, R1, R2, R3 immediately.
214 */
215
216#if CONFIG_HEXAGON_ARCH_VERSION < 4
217#define	vm_event_entry(CHandler) \
218	{ \
219		R29 = add(R29, #-(_PT_REGS_SIZE)); \
220		memd(R29 + #(_PT_R0100 + -_PT_REGS_SIZE)) = R1:0; \
221	} \
222	{ \
223		memd(R29 +#_PT_R0302) = R3:2; \
224	} \
225	trap1(#HVM_TRAP1_VMGETREGS); \
226	{ \
227		memd(R29 + #_PT_ER_VMEL) = R1:0; \
228		R0 = R29; \
229		R1.L = #LO(CHandler); \
230	} \
231	{ \
232		memd(R29 + #_PT_ER_VMPSP) = R3:2; \
233		R1.H = #HI(CHandler); \
234		jump event_dispatch; \
235	}
236#else
237/* V4+ */
238/* turn on I$ prefetch early */
239/* the # ## # syntax inserts a literal ## */
240#define	vm_event_entry(CHandler) \
241	{ \
242		R29 = add(R29, #-(_PT_REGS_SIZE)); \
243		memd(R29 + #(_PT_R0100 + -_PT_REGS_SIZE)) = R1:0; \
244		memd(R29 + #(_PT_R0302 + -_PT_REGS_SIZE)) = R3:2; \
245		R0 = usr; \
246	} \
247	{ \
248		memw(R29 + #_PT_PREDSUSR) = R0; \
249		R0 = setbit(R0, #16); \
250	} \
251	usr = R0; \
252	R1:0 = G1:0; \
253	{ \
254		memd(R29 + #_PT_ER_VMEL) = R1:0; \
255		R1 = # ## #(CHandler); \
256		R3:2 = G3:2; \
257	} \
258	{ \
259		R0 = R29; \
260		memd(R29 + #_PT_ER_VMPSP) = R3:2; \
261		jump event_dispatch; \
262	}
263#endif
264
265.text
266	/*
267	 * Do bulk save/restore in one place.
268	 * Adds a jump to dispatch latency, but
269	 * saves hundreds of bytes.
270	 */
271
272event_dispatch:
273	save_pt_regs()
274	callr	r1
275
276	/*
277	 * Coming back from the C-world, our thread info pointer
278	 * should be in the designated register (usually R19)
279	 *
280	 * If we were in kernel mode, we don't need to check scheduler
281	 * or signals if CONFIG_PREEMPT is not set.  If set, then it has
282	 * to jump to a need_resched kind of block.
283	 * BTW, CONFIG_PREEMPT is not supported yet.
284	 */
285
286#ifdef CONFIG_PREEMPT
287	R0 = #VM_INT_DISABLE
288	trap1(#HVM_TRAP1_VMSETIE)
289#endif
290
291	/*  "Nested control path" -- if the previous mode was kernel  */
292	{
293		R0 = memw(R29 + #_PT_ER_VMEST);
294		R26.L = #LO(do_work_pending);
295	}
296	{
297		P0 = tstbit(R0, #HVM_VMEST_UM_SFT);
298		if (!P0.new) jump:nt restore_all;
299		R26.H = #HI(do_work_pending);
300		R0 = #VM_INT_DISABLE;
301	}
302
303	/*
304	 * Check also the return from fork/system call, normally coming back from
305	 * user mode
306	 *
307	 * R26 needs to have do_work_pending, and R0 should have VM_INT_DISABLE
308	 */
309
310check_work_pending:
311	/*  Disable interrupts while checking TIF  */
312	trap1(#HVM_TRAP1_VMSETIE)
313	{
314		R0 = R29;  /*  regs should still be at top of stack  */
315		R1 = memw(THREADINFO_REG + #_THREAD_INFO_FLAGS);
316		callr R26;
317	}
318
319	{
320		P0 = cmp.eq(R0, #0); if (!P0.new) jump:nt check_work_pending;
321		R0 = #VM_INT_DISABLE;
322	}
323
324restore_all:
325	/*
326	 * Disable interrupts, if they weren't already, before reg restore.
327	 * R0 gets preloaded with #VM_INT_DISABLE before we get here.
328	 */
329	trap1(#HVM_TRAP1_VMSETIE)
330
331	/*  do the setregs here for VM 0.5  */
332	/*  R29 here should already be pointing at pt_regs  */
333	{
334		R1:0 = memd(R29 + #_PT_ER_VMEL);
335		R3:2 = memd(R29 + #_PT_ER_VMPSP);
336	}
337#if CONFIG_HEXAGON_ARCH_VERSION < 4
338	trap1(#HVM_TRAP1_VMSETREGS);
339#else
340	G1:0 = R1:0;
341	G3:2 = R3:2;
342#endif
343
344	R0 = R29
345	restore_pt_regs()
346	{
347		R1:0 = memd(R29 + #_PT_R0100);
348		R29 = add(R29, #_PT_REGS_SIZE);
349	}
350	trap1(#HVM_TRAP1_VMRTE)
351	/* Notreached */
352
353
354	.globl _K_enter_genex
355_K_enter_genex:
356	vm_event_entry(do_genex)
357
358	.globl _K_enter_interrupt
359_K_enter_interrupt:
360	vm_event_entry(arch_do_IRQ)
361
362	.globl _K_enter_trap0
363_K_enter_trap0:
364	vm_event_entry(do_trap0)
365
366	.globl _K_enter_machcheck
367_K_enter_machcheck:
368	vm_event_entry(do_machcheck)
369
370	.globl _K_enter_debug
371_K_enter_debug:
372	vm_event_entry(do_debug_exception)
373
374	.globl ret_from_fork
375ret_from_fork:
376	{
377		call schedule_tail
378		R26.H = #HI(do_work_pending);
379	}
380	{
381		P0 = cmp.eq(R24, #0);
382		R26.L = #LO(do_work_pending);
383		R0 = #VM_INT_DISABLE;
384	}
385	if P0 jump check_work_pending
386	{
387		R0 = R25;
388		callr R24
389	}
390	{
391		jump check_work_pending
392		R0 = #VM_INT_DISABLE;
393	}
394