xref: /linux/arch/mips/kernel/vpe.c (revision 4b660dbd9ee2059850fd30e0df420ca7a38a1856)
1 /*
2  * This file is subject to the terms and conditions of the GNU General Public
3  * License.  See the file "COPYING" in the main directory of this archive
4  * for more details.
5  *
6  * Copyright (C) 2004, 2005 MIPS Technologies, Inc.  All rights reserved.
7  * Copyright (C) 2013 Imagination Technologies Ltd.
8  *
9  * VPE support module for loading a MIPS SP program into VPE1. The SP
10  * environment is rather simple since there are no TLBs. It needs
11  * to be relocatable (or partially linked). Initialize your stack in
12  * the startup-code. The loader looks for the symbol __start and sets
13  * up the execution to resume from there. To load and run, simply do
14  * a cat SP 'binary' to the /dev/vpe1 device.
15  */
16 #include <linux/kernel.h>
17 #include <linux/device.h>
18 #include <linux/fs.h>
19 #include <linux/init.h>
20 #include <linux/slab.h>
21 #include <linux/list.h>
22 #include <linux/vmalloc.h>
23 #include <linux/elf.h>
24 #include <linux/seq_file.h>
25 #include <linux/syscalls.h>
26 #include <linux/moduleloader.h>
27 #include <linux/interrupt.h>
28 #include <linux/poll.h>
29 #include <linux/memblock.h>
30 #include <asm/mipsregs.h>
31 #include <asm/mipsmtregs.h>
32 #include <asm/cacheflush.h>
33 #include <linux/atomic.h>
34 #include <asm/mips_mt.h>
35 #include <asm/processor.h>
36 #include <asm/vpe.h>
37 
38 #ifndef ARCH_SHF_SMALL
39 #define ARCH_SHF_SMALL 0
40 #endif
41 
42 /* If this is set, the section belongs in the init part of the module */
43 #define INIT_OFFSET_MASK (1UL << (BITS_PER_LONG-1))
44 
45 struct vpe_control vpecontrol = {
46 	.vpe_list_lock	= __SPIN_LOCK_UNLOCKED(vpe_list_lock),
47 	.vpe_list	= LIST_HEAD_INIT(vpecontrol.vpe_list),
48 	.tc_list_lock	= __SPIN_LOCK_UNLOCKED(tc_list_lock),
49 	.tc_list	= LIST_HEAD_INIT(vpecontrol.tc_list)
50 };
51 
52 /* get the vpe associated with this minor */
53 struct vpe *get_vpe(int minor)
54 {
55 	struct vpe *res, *v;
56 
57 	if (!cpu_has_mipsmt)
58 		return NULL;
59 
60 	res = NULL;
61 	spin_lock(&vpecontrol.vpe_list_lock);
62 	list_for_each_entry(v, &vpecontrol.vpe_list, list) {
63 		if (v->minor == VPE_MODULE_MINOR) {
64 			res = v;
65 			break;
66 		}
67 	}
68 	spin_unlock(&vpecontrol.vpe_list_lock);
69 
70 	return res;
71 }
72 
73 /* get the vpe associated with this minor */
74 struct tc *get_tc(int index)
75 {
76 	struct tc *res, *t;
77 
78 	res = NULL;
79 	spin_lock(&vpecontrol.tc_list_lock);
80 	list_for_each_entry(t, &vpecontrol.tc_list, list) {
81 		if (t->index == index) {
82 			res = t;
83 			break;
84 		}
85 	}
86 	spin_unlock(&vpecontrol.tc_list_lock);
87 
88 	return res;
89 }
90 
91 /* allocate a vpe and associate it with this minor (or index) */
92 struct vpe *alloc_vpe(int minor)
93 {
94 	struct vpe *v;
95 
96 	v = kzalloc(sizeof(struct vpe), GFP_KERNEL);
97 	if (v == NULL)
98 		goto out;
99 
100 	INIT_LIST_HEAD(&v->tc);
101 	spin_lock(&vpecontrol.vpe_list_lock);
102 	list_add_tail(&v->list, &vpecontrol.vpe_list);
103 	spin_unlock(&vpecontrol.vpe_list_lock);
104 
105 	INIT_LIST_HEAD(&v->notify);
106 	v->minor = VPE_MODULE_MINOR;
107 
108 out:
109 	return v;
110 }
111 
112 /* allocate a tc. At startup only tc0 is running, all other can be halted. */
113 struct tc *alloc_tc(int index)
114 {
115 	struct tc *tc;
116 
117 	tc = kzalloc(sizeof(struct tc), GFP_KERNEL);
118 	if (tc == NULL)
119 		goto out;
120 
121 	INIT_LIST_HEAD(&tc->tc);
122 	tc->index = index;
123 
124 	spin_lock(&vpecontrol.tc_list_lock);
125 	list_add_tail(&tc->list, &vpecontrol.tc_list);
126 	spin_unlock(&vpecontrol.tc_list_lock);
127 
128 out:
129 	return tc;
130 }
131 
132 /* clean up and free everything */
133 void release_vpe(struct vpe *v)
134 {
135 	list_del(&v->list);
136 	if (v->load_addr)
137 		release_progmem(v->load_addr);
138 	kfree(v);
139 }
140 
141 /* Find some VPE program space */
142 void *alloc_progmem(unsigned long len)
143 {
144 	void *addr;
145 
146 #ifdef CONFIG_MIPS_VPE_LOADER_TOM
147 	/*
148 	 * This means you must tell Linux to use less memory than you
149 	 * physically have, for example by passing a mem= boot argument.
150 	 */
151 	addr = pfn_to_kaddr(max_low_pfn);
152 	memset(addr, 0, len);
153 #else
154 	/* simple grab some mem for now */
155 	addr = kzalloc(len, GFP_KERNEL);
156 #endif
157 
158 	return addr;
159 }
160 
161 void release_progmem(void *ptr)
162 {
163 #ifndef CONFIG_MIPS_VPE_LOADER_TOM
164 	kfree(ptr);
165 #endif
166 }
167 
168 /* Update size with this section: return offset. */
169 static long get_offset(unsigned long *size, Elf_Shdr *sechdr)
170 {
171 	long ret;
172 
173 	ret = ALIGN(*size, sechdr->sh_addralign ? : 1);
174 	*size = ret + sechdr->sh_size;
175 	return ret;
176 }
177 
178 /* Lay out the SHF_ALLOC sections in a way not dissimilar to how ld
179    might -- code, read-only data, read-write data, small data.	Tally
180    sizes, and place the offsets into sh_entsize fields: high bit means it
181    belongs in init. */
182 static void layout_sections(struct module *mod, const Elf_Ehdr *hdr,
183 			    Elf_Shdr *sechdrs, const char *secstrings)
184 {
185 	static unsigned long const masks[][2] = {
186 		/* NOTE: all executable code must be the first section
187 		 * in this array; otherwise modify the text_size
188 		 * finder in the two loops below */
189 		{SHF_EXECINSTR | SHF_ALLOC, ARCH_SHF_SMALL},
190 		{SHF_ALLOC, SHF_WRITE | ARCH_SHF_SMALL},
191 		{SHF_WRITE | SHF_ALLOC, ARCH_SHF_SMALL},
192 		{ARCH_SHF_SMALL | SHF_ALLOC, 0}
193 	};
194 	unsigned int m, i;
195 
196 	for (i = 0; i < hdr->e_shnum; i++)
197 		sechdrs[i].sh_entsize = ~0UL;
198 
199 	for (m = 0; m < ARRAY_SIZE(masks); ++m) {
200 		for (i = 0; i < hdr->e_shnum; ++i) {
201 			Elf_Shdr *s = &sechdrs[i];
202 			struct module_memory *mod_mem;
203 
204 			mod_mem = &mod->mem[MOD_TEXT];
205 
206 			if ((s->sh_flags & masks[m][0]) != masks[m][0]
207 			    || (s->sh_flags & masks[m][1])
208 			    || s->sh_entsize != ~0UL)
209 				continue;
210 			s->sh_entsize =
211 				get_offset((unsigned long *)&mod_mem->size, s);
212 		}
213 	}
214 }
215 
216 /* from module-elf32.c, but subverted a little */
217 
218 struct mips_hi16 {
219 	struct mips_hi16 *next;
220 	Elf32_Addr *addr;
221 	Elf32_Addr value;
222 };
223 
224 static struct mips_hi16 *mips_hi16_list;
225 static unsigned int gp_offs, gp_addr;
226 
227 static int apply_r_mips_none(struct module *me, uint32_t *location,
228 			     Elf32_Addr v)
229 {
230 	return 0;
231 }
232 
233 static int apply_r_mips_gprel16(struct module *me, uint32_t *location,
234 				Elf32_Addr v)
235 {
236 	int rel;
237 
238 	if (!(*location & 0xffff)) {
239 		rel = (int)v - gp_addr;
240 	} else {
241 		/* .sbss + gp(relative) + offset */
242 		/* kludge! */
243 		rel =  (int)(short)((int)v + gp_offs +
244 				    (int)(short)(*location & 0xffff) - gp_addr);
245 	}
246 
247 	if ((rel > 32768) || (rel < -32768)) {
248 		pr_debug("VPE loader: apply_r_mips_gprel16: relative address 0x%x out of range of gp register\n",
249 			 rel);
250 		return -ENOEXEC;
251 	}
252 
253 	*location = (*location & 0xffff0000) | (rel & 0xffff);
254 
255 	return 0;
256 }
257 
258 static int apply_r_mips_pc16(struct module *me, uint32_t *location,
259 			     Elf32_Addr v)
260 {
261 	int rel;
262 	rel = (((unsigned int)v - (unsigned int)location));
263 	rel >>= 2; /* because the offset is in _instructions_ not bytes. */
264 	rel -= 1;  /* and one instruction less due to the branch delay slot. */
265 
266 	if ((rel > 32768) || (rel < -32768)) {
267 		pr_debug("VPE loader: apply_r_mips_pc16: relative address out of range 0x%x\n",
268 			 rel);
269 		return -ENOEXEC;
270 	}
271 
272 	*location = (*location & 0xffff0000) | (rel & 0xffff);
273 
274 	return 0;
275 }
276 
277 static int apply_r_mips_32(struct module *me, uint32_t *location,
278 			   Elf32_Addr v)
279 {
280 	*location += v;
281 
282 	return 0;
283 }
284 
285 static int apply_r_mips_26(struct module *me, uint32_t *location,
286 			   Elf32_Addr v)
287 {
288 	if (v % 4) {
289 		pr_debug("VPE loader: apply_r_mips_26: unaligned relocation\n");
290 		return -ENOEXEC;
291 	}
292 
293 /*
294  * Not desperately convinced this is a good check of an overflow condition
295  * anyway. But it gets in the way of handling undefined weak symbols which
296  * we want to set to zero.
297  * if ((v & 0xf0000000) != (((unsigned long)location + 4) & 0xf0000000)) {
298  * printk(KERN_ERR
299  * "module %s: relocation overflow\n",
300  * me->name);
301  * return -ENOEXEC;
302  * }
303  */
304 
305 	*location = (*location & ~0x03ffffff) |
306 		((*location + (v >> 2)) & 0x03ffffff);
307 	return 0;
308 }
309 
310 static int apply_r_mips_hi16(struct module *me, uint32_t *location,
311 			     Elf32_Addr v)
312 {
313 	struct mips_hi16 *n;
314 
315 	/*
316 	 * We cannot relocate this one now because we don't know the value of
317 	 * the carry we need to add.  Save the information, and let LO16 do the
318 	 * actual relocation.
319 	 */
320 	n = kmalloc(sizeof(*n), GFP_KERNEL);
321 	if (!n)
322 		return -ENOMEM;
323 
324 	n->addr = location;
325 	n->value = v;
326 	n->next = mips_hi16_list;
327 	mips_hi16_list = n;
328 
329 	return 0;
330 }
331 
332 static int apply_r_mips_lo16(struct module *me, uint32_t *location,
333 			     Elf32_Addr v)
334 {
335 	unsigned long insnlo = *location;
336 	Elf32_Addr val, vallo;
337 	struct mips_hi16 *l, *next;
338 
339 	/* Sign extend the addend we extract from the lo insn.	*/
340 	vallo = ((insnlo & 0xffff) ^ 0x8000) - 0x8000;
341 
342 	if (mips_hi16_list != NULL) {
343 
344 		l = mips_hi16_list;
345 		while (l != NULL) {
346 			unsigned long insn;
347 
348 			/*
349 			 * The value for the HI16 had best be the same.
350 			 */
351 			if (v != l->value) {
352 				pr_debug("VPE loader: apply_r_mips_lo16/hi16: inconsistent value information\n");
353 				goto out_free;
354 			}
355 
356 			/*
357 			 * Do the HI16 relocation.  Note that we actually don't
358 			 * need to know anything about the LO16 itself, except
359 			 * where to find the low 16 bits of the addend needed
360 			 * by the LO16.
361 			 */
362 			insn = *l->addr;
363 			val = ((insn & 0xffff) << 16) + vallo;
364 			val += v;
365 
366 			/*
367 			 * Account for the sign extension that will happen in
368 			 * the low bits.
369 			 */
370 			val = ((val >> 16) + ((val & 0x8000) != 0)) & 0xffff;
371 
372 			insn = (insn & ~0xffff) | val;
373 			*l->addr = insn;
374 
375 			next = l->next;
376 			kfree(l);
377 			l = next;
378 		}
379 
380 		mips_hi16_list = NULL;
381 	}
382 
383 	/*
384 	 * Ok, we're done with the HI16 relocs.	 Now deal with the LO16.
385 	 */
386 	val = v + vallo;
387 	insnlo = (insnlo & ~0xffff) | (val & 0xffff);
388 	*location = insnlo;
389 
390 	return 0;
391 
392 out_free:
393 	while (l != NULL) {
394 		next = l->next;
395 		kfree(l);
396 		l = next;
397 	}
398 	mips_hi16_list = NULL;
399 
400 	return -ENOEXEC;
401 }
402 
403 static int (*reloc_handlers[]) (struct module *me, uint32_t *location,
404 				Elf32_Addr v) = {
405 	[R_MIPS_NONE]	= apply_r_mips_none,
406 	[R_MIPS_32]	= apply_r_mips_32,
407 	[R_MIPS_26]	= apply_r_mips_26,
408 	[R_MIPS_HI16]	= apply_r_mips_hi16,
409 	[R_MIPS_LO16]	= apply_r_mips_lo16,
410 	[R_MIPS_GPREL16] = apply_r_mips_gprel16,
411 	[R_MIPS_PC16] = apply_r_mips_pc16
412 };
413 
414 static char *rstrs[] = {
415 	[R_MIPS_NONE]	= "MIPS_NONE",
416 	[R_MIPS_32]	= "MIPS_32",
417 	[R_MIPS_26]	= "MIPS_26",
418 	[R_MIPS_HI16]	= "MIPS_HI16",
419 	[R_MIPS_LO16]	= "MIPS_LO16",
420 	[R_MIPS_GPREL16] = "MIPS_GPREL16",
421 	[R_MIPS_PC16] = "MIPS_PC16"
422 };
423 
424 static int apply_relocations(Elf32_Shdr *sechdrs,
425 		      const char *strtab,
426 		      unsigned int symindex,
427 		      unsigned int relsec,
428 		      struct module *me)
429 {
430 	Elf32_Rel *rel = (void *) sechdrs[relsec].sh_addr;
431 	Elf32_Sym *sym;
432 	uint32_t *location;
433 	unsigned int i;
434 	Elf32_Addr v;
435 	int res;
436 
437 	for (i = 0; i < sechdrs[relsec].sh_size / sizeof(*rel); i++) {
438 		Elf32_Word r_info = rel[i].r_info;
439 
440 		/* This is where to make the change */
441 		location = (void *)sechdrs[sechdrs[relsec].sh_info].sh_addr
442 			+ rel[i].r_offset;
443 		/* This is the symbol it is referring to */
444 		sym = (Elf32_Sym *)sechdrs[symindex].sh_addr
445 			+ ELF32_R_SYM(r_info);
446 
447 		if (!sym->st_value) {
448 			pr_debug("%s: undefined weak symbol %s\n",
449 				 me->name, strtab + sym->st_name);
450 			/* just print the warning, dont barf */
451 		}
452 
453 		v = sym->st_value;
454 
455 		res = reloc_handlers[ELF32_R_TYPE(r_info)](me, location, v);
456 		if (res) {
457 			char *r = rstrs[ELF32_R_TYPE(r_info)];
458 			pr_warn("VPE loader: .text+0x%x relocation type %s for symbol \"%s\" failed\n",
459 				rel[i].r_offset, r ? r : "UNKNOWN",
460 				strtab + sym->st_name);
461 			return res;
462 		}
463 	}
464 
465 	return 0;
466 }
467 
468 static inline void save_gp_address(unsigned int secbase, unsigned int rel)
469 {
470 	gp_addr = secbase + rel;
471 	gp_offs = gp_addr - (secbase & 0xffff0000);
472 }
473 /* end module-elf32.c */
474 
475 /* Change all symbols so that sh_value encodes the pointer directly. */
476 static void simplify_symbols(Elf_Shdr *sechdrs,
477 			    unsigned int symindex,
478 			    const char *strtab,
479 			    const char *secstrings,
480 			    unsigned int nsecs, struct module *mod)
481 {
482 	Elf_Sym *sym = (void *)sechdrs[symindex].sh_addr;
483 	unsigned long secbase, bssbase = 0;
484 	unsigned int i, n = sechdrs[symindex].sh_size / sizeof(Elf_Sym);
485 	int size;
486 
487 	/* find the .bss section for COMMON symbols */
488 	for (i = 0; i < nsecs; i++) {
489 		if (strncmp(secstrings + sechdrs[i].sh_name, ".bss", 4) == 0) {
490 			bssbase = sechdrs[i].sh_addr;
491 			break;
492 		}
493 	}
494 
495 	for (i = 1; i < n; i++) {
496 		switch (sym[i].st_shndx) {
497 		case SHN_COMMON:
498 			/* Allocate space for the symbol in the .bss section.
499 			   st_value is currently size.
500 			   We want it to have the address of the symbol. */
501 
502 			size = sym[i].st_value;
503 			sym[i].st_value = bssbase;
504 
505 			bssbase += size;
506 			break;
507 
508 		case SHN_ABS:
509 			/* Don't need to do anything */
510 			break;
511 
512 		case SHN_UNDEF:
513 			/* ret = -ENOENT; */
514 			break;
515 
516 		case SHN_MIPS_SCOMMON:
517 			pr_debug("simplify_symbols: ignoring SHN_MIPS_SCOMMON symbol <%s> st_shndx %d\n",
518 				 strtab + sym[i].st_name, sym[i].st_shndx);
519 			/* .sbss section */
520 			break;
521 
522 		default:
523 			secbase = sechdrs[sym[i].st_shndx].sh_addr;
524 
525 			if (strncmp(strtab + sym[i].st_name, "_gp", 3) == 0)
526 				save_gp_address(secbase, sym[i].st_value);
527 
528 			sym[i].st_value += secbase;
529 			break;
530 		}
531 	}
532 }
533 
534 #ifdef DEBUG_ELFLOADER
535 static void dump_elfsymbols(Elf_Shdr *sechdrs, unsigned int symindex,
536 			    const char *strtab, struct module *mod)
537 {
538 	Elf_Sym *sym = (void *)sechdrs[symindex].sh_addr;
539 	unsigned int i, n = sechdrs[symindex].sh_size / sizeof(Elf_Sym);
540 
541 	pr_debug("dump_elfsymbols: n %d\n", n);
542 	for (i = 1; i < n; i++) {
543 		pr_debug(" i %d name <%s> 0x%x\n", i, strtab + sym[i].st_name,
544 			 sym[i].st_value);
545 	}
546 }
547 #endif
548 
549 static int find_vpe_symbols(struct vpe *v, Elf_Shdr *sechdrs,
550 				      unsigned int symindex, const char *strtab,
551 				      struct module *mod)
552 {
553 	Elf_Sym *sym = (void *)sechdrs[symindex].sh_addr;
554 	unsigned int i, n = sechdrs[symindex].sh_size / sizeof(Elf_Sym);
555 
556 	for (i = 1; i < n; i++) {
557 		if (strcmp(strtab + sym[i].st_name, "__start") == 0)
558 			v->__start = sym[i].st_value;
559 
560 		if (strcmp(strtab + sym[i].st_name, "vpe_shared") == 0)
561 			v->shared_ptr = (void *)sym[i].st_value;
562 	}
563 
564 	if ((v->__start == 0) || (v->shared_ptr == NULL))
565 		return -1;
566 
567 	return 0;
568 }
569 
570 /*
571  * Allocates a VPE with some program code space(the load address), copies the
572  * contents of the program (p)buffer performing relocatations/etc, free's it
573  * when finished.
574  */
575 static int vpe_elfload(struct vpe *v)
576 {
577 	Elf_Ehdr *hdr;
578 	Elf_Shdr *sechdrs;
579 	long err = 0;
580 	char *secstrings, *strtab = NULL;
581 	unsigned int len, i, symindex = 0, strindex = 0, relocate = 0;
582 	struct module mod; /* so we can re-use the relocations code */
583 
584 	memset(&mod, 0, sizeof(struct module));
585 	strcpy(mod.name, "VPE loader");
586 
587 	hdr = (Elf_Ehdr *) v->pbuffer;
588 	len = v->plen;
589 
590 	/* Sanity checks against insmoding binaries or wrong arch,
591 	   weird elf version */
592 	if (memcmp(hdr->e_ident, ELFMAG, SELFMAG) != 0
593 	    || (hdr->e_type != ET_REL && hdr->e_type != ET_EXEC)
594 	    || !elf_check_arch(hdr)
595 	    || hdr->e_shentsize != sizeof(*sechdrs)) {
596 		pr_warn("VPE loader: program wrong arch or weird elf version\n");
597 
598 		return -ENOEXEC;
599 	}
600 
601 	if (hdr->e_type == ET_REL)
602 		relocate = 1;
603 
604 	if (len < hdr->e_shoff + hdr->e_shnum * sizeof(Elf_Shdr)) {
605 		pr_err("VPE loader: program length %u truncated\n", len);
606 
607 		return -ENOEXEC;
608 	}
609 
610 	/* Convenience variables */
611 	sechdrs = (void *)hdr + hdr->e_shoff;
612 	secstrings = (void *)hdr + sechdrs[hdr->e_shstrndx].sh_offset;
613 	sechdrs[0].sh_addr = 0;
614 
615 	/* And these should exist, but gcc whinges if we don't init them */
616 	symindex = strindex = 0;
617 
618 	if (relocate) {
619 		for (i = 1; i < hdr->e_shnum; i++) {
620 			if ((sechdrs[i].sh_type != SHT_NOBITS) &&
621 			    (len < sechdrs[i].sh_offset + sechdrs[i].sh_size)) {
622 				pr_err("VPE program length %u truncated\n",
623 				       len);
624 				return -ENOEXEC;
625 			}
626 
627 			/* Mark all sections sh_addr with their address in the
628 			   temporary image. */
629 			sechdrs[i].sh_addr = (size_t) hdr +
630 				sechdrs[i].sh_offset;
631 
632 			/* Internal symbols and strings. */
633 			if (sechdrs[i].sh_type == SHT_SYMTAB) {
634 				symindex = i;
635 				strindex = sechdrs[i].sh_link;
636 				strtab = (char *)hdr +
637 					sechdrs[strindex].sh_offset;
638 			}
639 		}
640 		layout_sections(&mod, hdr, sechdrs, secstrings);
641 	}
642 
643 	v->load_addr = alloc_progmem(mod.mem[MOD_TEXT].size);
644 	if (!v->load_addr)
645 		return -ENOMEM;
646 
647 	pr_info("VPE loader: loading to %p\n", v->load_addr);
648 
649 	if (relocate) {
650 		for (i = 0; i < hdr->e_shnum; i++) {
651 			void *dest;
652 
653 			if (!(sechdrs[i].sh_flags & SHF_ALLOC))
654 				continue;
655 
656 			dest = v->load_addr + sechdrs[i].sh_entsize;
657 
658 			if (sechdrs[i].sh_type != SHT_NOBITS)
659 				memcpy(dest, (void *)sechdrs[i].sh_addr,
660 				       sechdrs[i].sh_size);
661 			/* Update sh_addr to point to copy in image. */
662 			sechdrs[i].sh_addr = (unsigned long)dest;
663 
664 			pr_debug(" section sh_name %s sh_addr 0x%x\n",
665 				 secstrings + sechdrs[i].sh_name,
666 				 sechdrs[i].sh_addr);
667 		}
668 
669 		/* Fix up syms, so that st_value is a pointer to location. */
670 		simplify_symbols(sechdrs, symindex, strtab, secstrings,
671 				 hdr->e_shnum, &mod);
672 
673 		/* Now do relocations. */
674 		for (i = 1; i < hdr->e_shnum; i++) {
675 			const char *strtab = (char *)sechdrs[strindex].sh_addr;
676 			unsigned int info = sechdrs[i].sh_info;
677 
678 			/* Not a valid relocation section? */
679 			if (info >= hdr->e_shnum)
680 				continue;
681 
682 			/* Don't bother with non-allocated sections */
683 			if (!(sechdrs[info].sh_flags & SHF_ALLOC))
684 				continue;
685 
686 			if (sechdrs[i].sh_type == SHT_REL)
687 				err = apply_relocations(sechdrs, strtab,
688 							symindex, i, &mod);
689 			else if (sechdrs[i].sh_type == SHT_RELA)
690 				err = apply_relocate_add(sechdrs, strtab,
691 							 symindex, i, &mod);
692 			if (err < 0)
693 				return err;
694 
695 		}
696 	} else {
697 		struct elf_phdr *phdr = (struct elf_phdr *)
698 						((char *)hdr + hdr->e_phoff);
699 
700 		for (i = 0; i < hdr->e_phnum; i++) {
701 			if (phdr->p_type == PT_LOAD) {
702 				memcpy((void *)phdr->p_paddr,
703 				       (char *)hdr + phdr->p_offset,
704 				       phdr->p_filesz);
705 				memset((void *)phdr->p_paddr + phdr->p_filesz,
706 				       0, phdr->p_memsz - phdr->p_filesz);
707 		    }
708 		    phdr++;
709 		}
710 
711 		for (i = 0; i < hdr->e_shnum; i++) {
712 			/* Internal symbols and strings. */
713 			if (sechdrs[i].sh_type == SHT_SYMTAB) {
714 				symindex = i;
715 				strindex = sechdrs[i].sh_link;
716 				strtab = (char *)hdr +
717 					sechdrs[strindex].sh_offset;
718 
719 				/*
720 				 * mark symtab's address for when we try
721 				 * to find the magic symbols
722 				 */
723 				sechdrs[i].sh_addr = (size_t) hdr +
724 					sechdrs[i].sh_offset;
725 			}
726 		}
727 	}
728 
729 	/* make sure it's physically written out */
730 	flush_icache_range((unsigned long)v->load_addr,
731 			   (unsigned long)v->load_addr + v->len);
732 
733 	if ((find_vpe_symbols(v, sechdrs, symindex, strtab, &mod)) < 0) {
734 		if (v->__start == 0) {
735 			pr_warn("VPE loader: program does not contain a __start symbol\n");
736 			return -ENOEXEC;
737 		}
738 
739 		if (v->shared_ptr == NULL)
740 			pr_warn("VPE loader: program does not contain vpe_shared symbol.\n"
741 				" Unable to use AMVP (AP/SP) facilities.\n");
742 	}
743 
744 	pr_info(" elf loaded\n");
745 	return 0;
746 }
747 
748 /* checks VPE is unused and gets ready to load program	*/
749 static int vpe_open(struct inode *inode, struct file *filp)
750 {
751 	enum vpe_state state;
752 	struct vpe_notifications *notifier;
753 	struct vpe *v;
754 
755 	if (VPE_MODULE_MINOR != iminor(inode)) {
756 		/* assume only 1 device at the moment. */
757 		pr_warn("VPE loader: only vpe1 is supported\n");
758 
759 		return -ENODEV;
760 	}
761 
762 	v = get_vpe(aprp_cpu_index());
763 	if (v == NULL) {
764 		pr_warn("VPE loader: unable to get vpe\n");
765 
766 		return -ENODEV;
767 	}
768 
769 	state = xchg(&v->state, VPE_STATE_INUSE);
770 	if (state != VPE_STATE_UNUSED) {
771 		pr_debug("VPE loader: tc in use dumping regs\n");
772 
773 		list_for_each_entry(notifier, &v->notify, list)
774 			notifier->stop(aprp_cpu_index());
775 
776 		release_progmem(v->load_addr);
777 		cleanup_tc(get_tc(aprp_cpu_index()));
778 	}
779 
780 	/* this of-course trashes what was there before... */
781 	v->pbuffer = vmalloc(P_SIZE);
782 	if (!v->pbuffer) {
783 		pr_warn("VPE loader: unable to allocate memory\n");
784 		return -ENOMEM;
785 	}
786 	v->plen = P_SIZE;
787 	v->load_addr = NULL;
788 	v->len = 0;
789 	v->shared_ptr = NULL;
790 	v->__start = 0;
791 
792 	return 0;
793 }
794 
795 static int vpe_release(struct inode *inode, struct file *filp)
796 {
797 #ifdef CONFIG_MIPS_VPE_LOADER_MT
798 	struct vpe *v;
799 	Elf_Ehdr *hdr;
800 	int ret = 0;
801 
802 	v = get_vpe(aprp_cpu_index());
803 	if (v == NULL)
804 		return -ENODEV;
805 
806 	hdr = (Elf_Ehdr *) v->pbuffer;
807 	if (memcmp(hdr->e_ident, ELFMAG, SELFMAG) == 0) {
808 		if (vpe_elfload(v) >= 0) {
809 			vpe_run(v);
810 		} else {
811 			pr_warn("VPE loader: ELF load failed.\n");
812 			ret = -ENOEXEC;
813 		}
814 	} else {
815 		pr_warn("VPE loader: only elf files are supported\n");
816 		ret = -ENOEXEC;
817 	}
818 
819 	/* It's good to be able to run the SP and if it chokes have a look at
820 	   the /dev/rt?. But if we reset the pointer to the shared struct we
821 	   lose what has happened. So perhaps if garbage is sent to the vpe
822 	   device, use it as a trigger for the reset. Hopefully a nice
823 	   executable will be along shortly. */
824 	if (ret < 0)
825 		v->shared_ptr = NULL;
826 
827 	vfree(v->pbuffer);
828 	v->plen = 0;
829 
830 	return ret;
831 #else
832 	pr_warn("VPE loader: ELF load failed.\n");
833 	return -ENOEXEC;
834 #endif
835 }
836 
837 static ssize_t vpe_write(struct file *file, const char __user *buffer,
838 			 size_t count, loff_t *ppos)
839 {
840 	size_t ret = count;
841 	struct vpe *v;
842 
843 	if (iminor(file_inode(file)) != VPE_MODULE_MINOR)
844 		return -ENODEV;
845 
846 	v = get_vpe(aprp_cpu_index());
847 
848 	if (v == NULL)
849 		return -ENODEV;
850 
851 	if ((count + v->len) > v->plen) {
852 		pr_warn("VPE loader: elf size too big. Perhaps strip unneeded symbols\n");
853 		return -ENOMEM;
854 	}
855 
856 	count -= copy_from_user(v->pbuffer + v->len, buffer, count);
857 	if (!count)
858 		return -EFAULT;
859 
860 	v->len += count;
861 	return ret;
862 }
863 
864 const struct file_operations vpe_fops = {
865 	.owner = THIS_MODULE,
866 	.open = vpe_open,
867 	.release = vpe_release,
868 	.write = vpe_write,
869 	.llseek = noop_llseek,
870 };
871 
872 void *vpe_get_shared(int index)
873 {
874 	struct vpe *v = get_vpe(index);
875 
876 	if (v == NULL)
877 		return NULL;
878 
879 	return v->shared_ptr;
880 }
881 EXPORT_SYMBOL(vpe_get_shared);
882 
883 int vpe_notify(int index, struct vpe_notifications *notify)
884 {
885 	struct vpe *v = get_vpe(index);
886 
887 	if (v == NULL)
888 		return -1;
889 
890 	list_add(&notify->list, &v->notify);
891 	return 0;
892 }
893 EXPORT_SYMBOL(vpe_notify);
894 
895 module_init(vpe_module_init);
896 module_exit(vpe_module_exit);
897 MODULE_DESCRIPTION("MIPS VPE Loader");
898 MODULE_AUTHOR("Elizabeth Oldham, MIPS Technologies, Inc.");
899 MODULE_LICENSE("GPL");
900