2 * Copyright (C) 2004, 2005 MIPS Technologies, Inc. All rights reserved.
4 * This program is free software; you can distribute it and/or modify it
5 * under the terms of the GNU General Public License (Version 2) as
6 * published by the Free Software Foundation.
8 * This program is distributed in the hope it will be useful, but WITHOUT
9 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
10 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
13 * You should have received a copy of the GNU General Public License along
14 * with this program; if not, write to the Free Software Foundation, Inc.,
15 * 59 Temple Place - Suite 330, Boston MA 02111-1307, USA.
21 * Provides support for loading a MIPS SP program on VPE1.
22 * The SP environment is rather simple, no tlb's. It needs to be relocatable
23 * (or partially linked). You should initialise your stack in the startup
24 * code. This loader looks for the symbol __start and sets up
25 * execution to resume from there. The MIPS SDE kit contains suitable examples.
27 * To load and run, simply cat a SP 'program file' to /dev/vpe1.
28 * i.e cat spapp >/dev/vpe1.
30 #include <linux/kernel.h>
31 #include <linux/device.h>
33 #include <linux/init.h>
34 #include <asm/uaccess.h>
35 #include <linux/slab.h>
36 #include <linux/list.h>
37 #include <linux/vmalloc.h>
38 #include <linux/elf.h>
39 #include <linux/seq_file.h>
40 #include <linux/syscalls.h>
41 #include <linux/moduleloader.h>
42 #include <linux/interrupt.h>
43 #include <linux/poll.h>
44 #include <linux/bootmem.h>
45 #include <asm/mipsregs.h>
46 #include <asm/mipsmtregs.h>
47 #include <asm/cacheflush.h>
48 #include <linux/atomic.h>
50 #include <asm/mips_mt.h>
51 #include <asm/processor.h>
54 typedef void *vpe_handle;
56 #ifndef ARCH_SHF_SMALL
57 #define ARCH_SHF_SMALL 0
60 /* If this is set, the section belongs in the init part of the module */
61 #define INIT_OFFSET_MASK (1UL << (BITS_PER_LONG-1))
64 * The number of TCs and VPEs physically available on the core
66 static int hw_tcs, hw_vpes;
67 static char module_name[] = "vpe";
69 static const int minor = 1; /* fixed for now */
71 /* grab the likely amount of memory we will need. */
72 #ifdef CONFIG_MIPS_VPE_LOADER_TOM
73 #define P_SIZE (2 * 1024 * 1024)
75 /* add an overhead to the max kmalloc size for non-striped symbols/etc */
76 #define P_SIZE (256 * 1024)
79 extern unsigned long physical_memsize;
82 #define VPE_PATH_MAX 256
100 /* (device) minor associated with this vpe */
103 /* elfloader stuff */
108 char cwd[VPE_PATH_MAX];
110 unsigned long __start;
112 /* tc's associated with this vpe */
115 /* The list of vpe's */
116 struct list_head list;
118 /* shared symbol address */
121 /* the list of who wants to know when something major happens */
122 struct list_head notify;
131 struct vpe *pvpe; /* parent VPE */
132 struct list_head tc; /* The list of TC's with this VPE */
133 struct list_head list; /* The global list of tc's */
137 spinlock_t vpe_list_lock;
138 struct list_head vpe_list; /* Virtual processing elements */
139 spinlock_t tc_list_lock;
140 struct list_head tc_list; /* Thread contexts */
142 .vpe_list_lock = __SPIN_LOCK_UNLOCKED(vpe_list_lock),
143 .vpe_list = LIST_HEAD_INIT(vpecontrol.vpe_list),
144 .tc_list_lock = __SPIN_LOCK_UNLOCKED(tc_list_lock),
145 .tc_list = LIST_HEAD_INIT(vpecontrol.tc_list)
148 static void release_progmem(void *ptr);
150 /* get the vpe associated with this minor */
151 static struct vpe *get_vpe(int minor)
159 spin_lock(&vpecontrol.vpe_list_lock);
160 list_for_each_entry(v, &vpecontrol.vpe_list, list) {
161 if (v->minor == minor) {
166 spin_unlock(&vpecontrol.vpe_list_lock);
171 /* get the vpe associated with this minor */
172 static struct tc *get_tc(int index)
177 spin_lock(&vpecontrol.tc_list_lock);
178 list_for_each_entry(t, &vpecontrol.tc_list, list) {
179 if (t->index == index) {
184 spin_unlock(&vpecontrol.tc_list_lock);
189 /* allocate a vpe and associate it with this minor (or index) */
190 static struct vpe *alloc_vpe(int minor)
194 if ((v = kzalloc(sizeof(struct vpe), GFP_KERNEL)) == NULL)
197 INIT_LIST_HEAD(&v->tc);
198 spin_lock(&vpecontrol.vpe_list_lock);
199 list_add_tail(&v->list, &vpecontrol.vpe_list);
200 spin_unlock(&vpecontrol.vpe_list_lock);
202 INIT_LIST_HEAD(&v->notify);
208 /* allocate a tc. At startup only tc0 is running, all other can be halted. */
209 static struct tc *alloc_tc(int index)
213 if ((tc = kzalloc(sizeof(struct tc), GFP_KERNEL)) == NULL)
216 INIT_LIST_HEAD(&tc->tc);
219 spin_lock(&vpecontrol.tc_list_lock);
220 list_add_tail(&tc->list, &vpecontrol.tc_list);
221 spin_unlock(&vpecontrol.tc_list_lock);
227 /* clean up and free everything */
228 static void release_vpe(struct vpe *v)
236 static void __maybe_unused dump_mtregs(void)
240 val = read_c0_config3();
241 printk("config3 0x%lx MT %ld\n", val,
242 (val & CONFIG3_MT) >> CONFIG3_MT_SHIFT);
244 val = read_c0_mvpcontrol();
245 printk("MVPControl 0x%lx, STLB %ld VPC %ld EVP %ld\n", val,
246 (val & MVPCONTROL_STLB) >> MVPCONTROL_STLB_SHIFT,
247 (val & MVPCONTROL_VPC) >> MVPCONTROL_VPC_SHIFT,
248 (val & MVPCONTROL_EVP));
250 val = read_c0_mvpconf0();
251 printk("mvpconf0 0x%lx, PVPE %ld PTC %ld M %ld\n", val,
252 (val & MVPCONF0_PVPE) >> MVPCONF0_PVPE_SHIFT,
253 val & MVPCONF0_PTC, (val & MVPCONF0_M) >> MVPCONF0_M_SHIFT);
256 /* Find some VPE program space */
257 static void *alloc_progmem(unsigned long len)
261 #ifdef CONFIG_MIPS_VPE_LOADER_TOM
263 * This means you must tell Linux to use less memory than you
264 * physically have, for example by passing a mem= boot argument.
266 addr = pfn_to_kaddr(max_low_pfn);
267 memset(addr, 0, len);
269 /* simple grab some mem for now */
270 addr = kzalloc(len, GFP_KERNEL);
276 static void release_progmem(void *ptr)
278 #ifndef CONFIG_MIPS_VPE_LOADER_TOM
283 /* Update size with this section: return offset. */
284 static long get_offset(unsigned long *size, Elf_Shdr * sechdr)
288 ret = ALIGN(*size, sechdr->sh_addralign ? : 1);
289 *size = ret + sechdr->sh_size;
293 /* Lay out the SHF_ALLOC sections in a way not dissimilar to how ld
294 might -- code, read-only data, read-write data, small data. Tally
295 sizes, and place the offsets into sh_entsize fields: high bit means it
297 static void layout_sections(struct module *mod, const Elf_Ehdr * hdr,
298 Elf_Shdr * sechdrs, const char *secstrings)
300 static unsigned long const masks[][2] = {
301 /* NOTE: all executable code must be the first section
302 * in this array; otherwise modify the text_size
303 * finder in the two loops below */
304 {SHF_EXECINSTR | SHF_ALLOC, ARCH_SHF_SMALL},
305 {SHF_ALLOC, SHF_WRITE | ARCH_SHF_SMALL},
306 {SHF_WRITE | SHF_ALLOC, ARCH_SHF_SMALL},
307 {ARCH_SHF_SMALL | SHF_ALLOC, 0}
311 for (i = 0; i < hdr->e_shnum; i++)
312 sechdrs[i].sh_entsize = ~0UL;
314 for (m = 0; m < ARRAY_SIZE(masks); ++m) {
315 for (i = 0; i < hdr->e_shnum; ++i) {
316 Elf_Shdr *s = &sechdrs[i];
318 // || strncmp(secstrings + s->sh_name, ".init", 5) == 0)
319 if ((s->sh_flags & masks[m][0]) != masks[m][0]
320 || (s->sh_flags & masks[m][1])
321 || s->sh_entsize != ~0UL)
324 get_offset((unsigned long *)&mod->core_size, s);
328 mod->core_text_size = mod->core_size;
334 /* from module-elf32.c, but subverted a little */
337 struct mips_hi16 *next;
342 static struct mips_hi16 *mips_hi16_list;
343 static unsigned int gp_offs, gp_addr;
345 static int apply_r_mips_none(struct module *me, uint32_t *location,
351 static int apply_r_mips_gprel16(struct module *me, uint32_t *location,
356 if( !(*location & 0xffff) ) {
357 rel = (int)v - gp_addr;
360 /* .sbss + gp(relative) + offset */
362 rel = (int)(short)((int)v + gp_offs +
363 (int)(short)(*location & 0xffff) - gp_addr);
366 if( (rel > 32768) || (rel < -32768) ) {
367 printk(KERN_DEBUG "VPE loader: apply_r_mips_gprel16: "
368 "relative address 0x%x out of range of gp register\n",
373 *location = (*location & 0xffff0000) | (rel & 0xffff);
378 static int apply_r_mips_pc16(struct module *me, uint32_t *location,
382 rel = (((unsigned int)v - (unsigned int)location));
383 rel >>= 2; // because the offset is in _instructions_ not bytes.
384 rel -= 1; // and one instruction less due to the branch delay slot.
386 if( (rel > 32768) || (rel < -32768) ) {
387 printk(KERN_DEBUG "VPE loader: "
388 "apply_r_mips_pc16: relative address out of range 0x%x\n", rel);
392 *location = (*location & 0xffff0000) | (rel & 0xffff);
397 static int apply_r_mips_32(struct module *me, uint32_t *location,
405 static int apply_r_mips_26(struct module *me, uint32_t *location,
409 printk(KERN_DEBUG "VPE loader: apply_r_mips_26 "
410 " unaligned relocation\n");
415 * Not desperately convinced this is a good check of an overflow condition
416 * anyway. But it gets in the way of handling undefined weak symbols which
417 * we want to set to zero.
418 * if ((v & 0xf0000000) != (((unsigned long)location + 4) & 0xf0000000)) {
420 * "module %s: relocation overflow\n",
426 *location = (*location & ~0x03ffffff) |
427 ((*location + (v >> 2)) & 0x03ffffff);
431 static int apply_r_mips_hi16(struct module *me, uint32_t *location,
437 * We cannot relocate this one now because we don't know the value of
438 * the carry we need to add. Save the information, and let LO16 do the
441 n = kmalloc(sizeof *n, GFP_KERNEL);
447 n->next = mips_hi16_list;
453 static int apply_r_mips_lo16(struct module *me, uint32_t *location,
456 unsigned long insnlo = *location;
457 Elf32_Addr val, vallo;
458 struct mips_hi16 *l, *next;
460 /* Sign extend the addend we extract from the lo insn. */
461 vallo = ((insnlo & 0xffff) ^ 0x8000) - 0x8000;
463 if (mips_hi16_list != NULL) {
470 * The value for the HI16 had best be the same.
473 printk(KERN_DEBUG "VPE loader: "
474 "apply_r_mips_lo16/hi16: \t"
475 "inconsistent value information\n");
480 * Do the HI16 relocation. Note that we actually don't
481 * need to know anything about the LO16 itself, except
482 * where to find the low 16 bits of the addend needed
486 val = ((insn & 0xffff) << 16) + vallo;
490 * Account for the sign extension that will happen in
493 val = ((val >> 16) + ((val & 0x8000) != 0)) & 0xffff;
495 insn = (insn & ~0xffff) | val;
503 mips_hi16_list = NULL;
507 * Ok, we're done with the HI16 relocs. Now deal with the LO16.
510 insnlo = (insnlo & ~0xffff) | (val & 0xffff);
521 mips_hi16_list = NULL;
526 static int (*reloc_handlers[]) (struct module *me, uint32_t *location,
528 [R_MIPS_NONE] = apply_r_mips_none,
529 [R_MIPS_32] = apply_r_mips_32,
530 [R_MIPS_26] = apply_r_mips_26,
531 [R_MIPS_HI16] = apply_r_mips_hi16,
532 [R_MIPS_LO16] = apply_r_mips_lo16,
533 [R_MIPS_GPREL16] = apply_r_mips_gprel16,
534 [R_MIPS_PC16] = apply_r_mips_pc16
537 static char *rstrs[] = {
538 [R_MIPS_NONE] = "MIPS_NONE",
539 [R_MIPS_32] = "MIPS_32",
540 [R_MIPS_26] = "MIPS_26",
541 [R_MIPS_HI16] = "MIPS_HI16",
542 [R_MIPS_LO16] = "MIPS_LO16",
543 [R_MIPS_GPREL16] = "MIPS_GPREL16",
544 [R_MIPS_PC16] = "MIPS_PC16"
547 static int apply_relocations(Elf32_Shdr *sechdrs,
549 unsigned int symindex,
553 Elf32_Rel *rel = (void *) sechdrs[relsec].sh_addr;
560 for (i = 0; i < sechdrs[relsec].sh_size / sizeof(*rel); i++) {
561 Elf32_Word r_info = rel[i].r_info;
563 /* This is where to make the change */
564 location = (void *)sechdrs[sechdrs[relsec].sh_info].sh_addr
566 /* This is the symbol it is referring to */
567 sym = (Elf32_Sym *)sechdrs[symindex].sh_addr
568 + ELF32_R_SYM(r_info);
570 if (!sym->st_value) {
571 printk(KERN_DEBUG "%s: undefined weak symbol %s\n",
572 me->name, strtab + sym->st_name);
573 /* just print the warning, dont barf */
578 res = reloc_handlers[ELF32_R_TYPE(r_info)](me, location, v);
580 char *r = rstrs[ELF32_R_TYPE(r_info)];
581 printk(KERN_WARNING "VPE loader: .text+0x%x "
582 "relocation type %s for symbol \"%s\" failed\n",
583 rel[i].r_offset, r ? r : "UNKNOWN",
584 strtab + sym->st_name);
592 static inline void save_gp_address(unsigned int secbase, unsigned int rel)
594 gp_addr = secbase + rel;
595 gp_offs = gp_addr - (secbase & 0xffff0000);
597 /* end module-elf32.c */
601 /* Change all symbols so that sh_value encodes the pointer directly. */
602 static void simplify_symbols(Elf_Shdr * sechdrs,
603 unsigned int symindex,
605 const char *secstrings,
606 unsigned int nsecs, struct module *mod)
608 Elf_Sym *sym = (void *)sechdrs[symindex].sh_addr;
609 unsigned long secbase, bssbase = 0;
610 unsigned int i, n = sechdrs[symindex].sh_size / sizeof(Elf_Sym);
613 /* find the .bss section for COMMON symbols */
614 for (i = 0; i < nsecs; i++) {
615 if (strncmp(secstrings + sechdrs[i].sh_name, ".bss", 4) == 0) {
616 bssbase = sechdrs[i].sh_addr;
621 for (i = 1; i < n; i++) {
622 switch (sym[i].st_shndx) {
624 /* Allocate space for the symbol in the .bss section.
625 st_value is currently size.
626 We want it to have the address of the symbol. */
628 size = sym[i].st_value;
629 sym[i].st_value = bssbase;
635 /* Don't need to do anything */
642 case SHN_MIPS_SCOMMON:
643 printk(KERN_DEBUG "simplify_symbols: ignoring SHN_MIPS_SCOMMON "
644 "symbol <%s> st_shndx %d\n", strtab + sym[i].st_name,
650 secbase = sechdrs[sym[i].st_shndx].sh_addr;
652 if (strncmp(strtab + sym[i].st_name, "_gp", 3) == 0) {
653 save_gp_address(secbase, sym[i].st_value);
656 sym[i].st_value += secbase;
662 #ifdef DEBUG_ELFLOADER
663 static void dump_elfsymbols(Elf_Shdr * sechdrs, unsigned int symindex,
664 const char *strtab, struct module *mod)
666 Elf_Sym *sym = (void *)sechdrs[symindex].sh_addr;
667 unsigned int i, n = sechdrs[symindex].sh_size / sizeof(Elf_Sym);
669 printk(KERN_DEBUG "dump_elfsymbols: n %d\n", n);
670 for (i = 1; i < n; i++) {
671 printk(KERN_DEBUG " i %d name <%s> 0x%x\n", i,
672 strtab + sym[i].st_name, sym[i].st_value);
677 /* We are prepared so configure and start the VPE... */
678 static int vpe_run(struct vpe * v)
680 unsigned long flags, val, dmt_flag;
681 struct vpe_notifications *n;
682 unsigned int vpeflags;
685 /* check we are the Master VPE */
686 local_irq_save(flags);
687 val = read_c0_vpeconf0();
688 if (!(val & VPECONF0_MVP)) {
690 "VPE loader: only Master VPE's are allowed to configure MT\n");
691 local_irq_restore(flags);
699 if (list_empty(&v->tc)) {
702 local_irq_restore(flags);
705 "VPE loader: No TC's associated with VPE %d\n",
711 t = list_first_entry(&v->tc, struct tc, tc);
713 /* Put MVPE's into 'configuration state' */
714 set_c0_mvpcontrol(MVPCONTROL_VPC);
718 /* should check it is halted, and not activated */
719 if ((read_tc_c0_tcstatus() & TCSTATUS_A) || !(read_tc_c0_tchalt() & TCHALT_H)) {
722 local_irq_restore(flags);
724 printk(KERN_WARNING "VPE loader: TC %d is already active!\n",
730 /* Write the address we want it to start running from in the TCPC register. */
731 write_tc_c0_tcrestart((unsigned long)v->__start);
732 write_tc_c0_tccontext((unsigned long)0);
735 * Mark the TC as activated, not interrupt exempt and not dynamically
738 val = read_tc_c0_tcstatus();
739 val = (val & ~(TCSTATUS_DA | TCSTATUS_IXMT)) | TCSTATUS_A;
740 write_tc_c0_tcstatus(val);
742 write_tc_c0_tchalt(read_tc_c0_tchalt() & ~TCHALT_H);
745 * The sde-kit passes 'memsize' to __start in $a3, so set something
746 * here... Or set $a3 to zero and define DFLT_STACK_SIZE and
747 * DFLT_HEAP_SIZE when you compile your program
750 mttgpr(7, physical_memsize);
754 * bind the TC to VPE 1 as late as possible so we only have the final
755 * VPE registers to set up, and so an EJTAG probe can trigger on it
757 write_tc_c0_tcbind((read_tc_c0_tcbind() & ~TCBIND_CURVPE) | 1);
759 write_vpe_c0_vpeconf0(read_vpe_c0_vpeconf0() & ~(VPECONF0_VPA));
761 back_to_back_c0_hazard();
763 /* Set up the XTC bit in vpeconf0 to point at our tc */
764 write_vpe_c0_vpeconf0( (read_vpe_c0_vpeconf0() & ~(VPECONF0_XTC))
765 | (t->index << VPECONF0_XTC_SHIFT));
767 back_to_back_c0_hazard();
769 /* enable this VPE */
770 write_vpe_c0_vpeconf0(read_vpe_c0_vpeconf0() | VPECONF0_VPA);
772 /* clear out any left overs from a previous program */
773 write_vpe_c0_status(0);
774 write_vpe_c0_cause(0);
776 /* take system out of configuration state */
777 clear_c0_mvpcontrol(MVPCONTROL_VPC);
780 * SMTC/SMVP kernels manage VPE enable independently,
781 * but uniprocessor kernels need to turn it on, even
782 * if that wasn't the pre-dvpe() state.
790 local_irq_restore(flags);
792 list_for_each_entry(n, &v->notify, list)
798 static int find_vpe_symbols(struct vpe * v, Elf_Shdr * sechdrs,
799 unsigned int symindex, const char *strtab,
802 Elf_Sym *sym = (void *)sechdrs[symindex].sh_addr;
803 unsigned int i, n = sechdrs[symindex].sh_size / sizeof(Elf_Sym);
805 for (i = 1; i < n; i++) {
806 if (strcmp(strtab + sym[i].st_name, "__start") == 0) {
807 v->__start = sym[i].st_value;
810 if (strcmp(strtab + sym[i].st_name, "vpe_shared") == 0) {
811 v->shared_ptr = (void *)sym[i].st_value;
815 if ( (v->__start == 0) || (v->shared_ptr == NULL))
822 * Allocates a VPE with some program code space(the load address), copies the
823 * contents of the program (p)buffer performing relocatations/etc, free's it
826 static int vpe_elfload(struct vpe * v)
831 char *secstrings, *strtab = NULL;
832 unsigned int len, i, symindex = 0, strindex = 0, relocate = 0;
833 struct module mod; // so we can re-use the relocations code
835 memset(&mod, 0, sizeof(struct module));
836 strcpy(mod.name, "VPE loader");
838 hdr = (Elf_Ehdr *) v->pbuffer;
841 /* Sanity checks against insmoding binaries or wrong arch,
843 if (memcmp(hdr->e_ident, ELFMAG, SELFMAG) != 0
844 || (hdr->e_type != ET_REL && hdr->e_type != ET_EXEC)
845 || !elf_check_arch(hdr)
846 || hdr->e_shentsize != sizeof(*sechdrs)) {
848 "VPE loader: program wrong arch or weird elf version\n");
853 if (hdr->e_type == ET_REL)
856 if (len < hdr->e_shoff + hdr->e_shnum * sizeof(Elf_Shdr)) {
857 printk(KERN_ERR "VPE loader: program length %u truncated\n",
863 /* Convenience variables */
864 sechdrs = (void *)hdr + hdr->e_shoff;
865 secstrings = (void *)hdr + sechdrs[hdr->e_shstrndx].sh_offset;
866 sechdrs[0].sh_addr = 0;
868 /* And these should exist, but gcc whinges if we don't init them */
869 symindex = strindex = 0;
872 for (i = 1; i < hdr->e_shnum; i++) {
873 if (sechdrs[i].sh_type != SHT_NOBITS
874 && len < sechdrs[i].sh_offset + sechdrs[i].sh_size) {
875 printk(KERN_ERR "VPE program length %u truncated\n",
880 /* Mark all sections sh_addr with their address in the
882 sechdrs[i].sh_addr = (size_t) hdr + sechdrs[i].sh_offset;
884 /* Internal symbols and strings. */
885 if (sechdrs[i].sh_type == SHT_SYMTAB) {
887 strindex = sechdrs[i].sh_link;
888 strtab = (char *)hdr + sechdrs[strindex].sh_offset;
891 layout_sections(&mod, hdr, sechdrs, secstrings);
894 v->load_addr = alloc_progmem(mod.core_size);
898 pr_info("VPE loader: loading to %p\n", v->load_addr);
901 for (i = 0; i < hdr->e_shnum; i++) {
904 if (!(sechdrs[i].sh_flags & SHF_ALLOC))
907 dest = v->load_addr + sechdrs[i].sh_entsize;
909 if (sechdrs[i].sh_type != SHT_NOBITS)
910 memcpy(dest, (void *)sechdrs[i].sh_addr,
912 /* Update sh_addr to point to copy in image. */
913 sechdrs[i].sh_addr = (unsigned long)dest;
915 printk(KERN_DEBUG " section sh_name %s sh_addr 0x%x\n",
916 secstrings + sechdrs[i].sh_name, sechdrs[i].sh_addr);
919 /* Fix up syms, so that st_value is a pointer to location. */
920 simplify_symbols(sechdrs, symindex, strtab, secstrings,
923 /* Now do relocations. */
924 for (i = 1; i < hdr->e_shnum; i++) {
925 const char *strtab = (char *)sechdrs[strindex].sh_addr;
926 unsigned int info = sechdrs[i].sh_info;
928 /* Not a valid relocation section? */
929 if (info >= hdr->e_shnum)
932 /* Don't bother with non-allocated sections */
933 if (!(sechdrs[info].sh_flags & SHF_ALLOC))
936 if (sechdrs[i].sh_type == SHT_REL)
937 err = apply_relocations(sechdrs, strtab, symindex, i,
939 else if (sechdrs[i].sh_type == SHT_RELA)
940 err = apply_relocate_add(sechdrs, strtab, symindex, i,
947 struct elf_phdr *phdr = (struct elf_phdr *) ((char *)hdr + hdr->e_phoff);
949 for (i = 0; i < hdr->e_phnum; i++) {
950 if (phdr->p_type == PT_LOAD) {
951 memcpy((void *)phdr->p_paddr,
952 (char *)hdr + phdr->p_offset,
954 memset((void *)phdr->p_paddr + phdr->p_filesz,
955 0, phdr->p_memsz - phdr->p_filesz);
960 for (i = 0; i < hdr->e_shnum; i++) {
961 /* Internal symbols and strings. */
962 if (sechdrs[i].sh_type == SHT_SYMTAB) {
964 strindex = sechdrs[i].sh_link;
965 strtab = (char *)hdr + sechdrs[strindex].sh_offset;
967 /* mark the symtab's address for when we try to find the
969 sechdrs[i].sh_addr = (size_t) hdr + sechdrs[i].sh_offset;
974 /* make sure it's physically written out */
975 flush_icache_range((unsigned long)v->load_addr,
976 (unsigned long)v->load_addr + v->len);
978 if ((find_vpe_symbols(v, sechdrs, symindex, strtab, &mod)) < 0) {
979 if (v->__start == 0) {
980 printk(KERN_WARNING "VPE loader: program does not contain "
981 "a __start symbol\n");
985 if (v->shared_ptr == NULL)
986 printk(KERN_WARNING "VPE loader: "
987 "program does not contain vpe_shared symbol.\n"
988 " Unable to use AMVP (AP/SP) facilities.\n");
991 printk(" elf loaded\n");
995 static void cleanup_tc(struct tc *tc)
998 unsigned int mtflags, vpflags;
1001 local_irq_save(flags);
1004 /* Put MVPE's into 'configuration state' */
1005 set_c0_mvpcontrol(MVPCONTROL_VPC);
1008 tmp = read_tc_c0_tcstatus();
1010 /* mark not allocated and not dynamically allocatable */
1011 tmp &= ~(TCSTATUS_A | TCSTATUS_DA);
1012 tmp |= TCSTATUS_IXMT; /* interrupt exempt */
1013 write_tc_c0_tcstatus(tmp);
1015 write_tc_c0_tchalt(TCHALT_H);
1018 /* bind it to anything other than VPE1 */
1019 // write_tc_c0_tcbind(read_tc_c0_tcbind() & ~TCBIND_CURVPE); // | TCBIND_CURVPE
1021 clear_c0_mvpcontrol(MVPCONTROL_VPC);
1024 local_irq_restore(flags);
1027 static int getcwd(char *buff, int size)
1029 mm_segment_t old_fs;
1035 ret = sys_getcwd(buff, size);
1042 /* checks VPE is unused and gets ready to load program */
1043 static int vpe_open(struct inode *inode, struct file *filp)
1045 enum vpe_state state;
1046 struct vpe_notifications *not;
1050 if (minor != iminor(inode)) {
1051 /* assume only 1 device at the moment. */
1052 pr_warning("VPE loader: only vpe1 is supported\n");
1057 if ((v = get_vpe(tclimit)) == NULL) {
1058 pr_warning("VPE loader: unable to get vpe\n");
1063 state = xchg(&v->state, VPE_STATE_INUSE);
1064 if (state != VPE_STATE_UNUSED) {
1065 printk(KERN_DEBUG "VPE loader: tc in use dumping regs\n");
1067 list_for_each_entry(not, &v->notify, list) {
1071 release_progmem(v->load_addr);
1072 cleanup_tc(get_tc(tclimit));
1075 /* this of-course trashes what was there before... */
1076 v->pbuffer = vmalloc(P_SIZE);
1078 pr_warning("VPE loader: unable to allocate memory\n");
1082 v->load_addr = NULL;
1086 ret = getcwd(v->cwd, VPE_PATH_MAX);
1088 printk(KERN_WARNING "VPE loader: open, getcwd returned %d\n", ret);
1090 v->shared_ptr = NULL;
1096 static int vpe_release(struct inode *inode, struct file *filp)
1102 v = get_vpe(tclimit);
1106 hdr = (Elf_Ehdr *) v->pbuffer;
1107 if (memcmp(hdr->e_ident, ELFMAG, SELFMAG) == 0) {
1108 if (vpe_elfload(v) >= 0) {
1111 printk(KERN_WARNING "VPE loader: ELF load failed.\n");
1115 printk(KERN_WARNING "VPE loader: only elf files are supported\n");
1119 /* It's good to be able to run the SP and if it chokes have a look at
1120 the /dev/rt?. But if we reset the pointer to the shared struct we
1121 lose what has happened. So perhaps if garbage is sent to the vpe
1122 device, use it as a trigger for the reset. Hopefully a nice
1123 executable will be along shortly. */
1125 v->shared_ptr = NULL;
1133 static ssize_t vpe_write(struct file *file, const char __user * buffer,
1134 size_t count, loff_t * ppos)
1139 if (iminor(file_inode(file)) != minor)
1142 v = get_vpe(tclimit);
1146 if ((count + v->len) > v->plen) {
1148 "VPE loader: elf size too big. Perhaps strip uneeded symbols\n");
1152 count -= copy_from_user(v->pbuffer + v->len, buffer, count);
1160 static const struct file_operations vpe_fops = {
1161 .owner = THIS_MODULE,
1163 .release = vpe_release,
1165 .llseek = noop_llseek,
1168 /* module wrapper entry points */
1170 vpe_handle vpe_alloc(void)
1176 for (i = 1; i < MAX_VPES; i++) {
1177 if ((v = get_vpe(i)) != NULL) {
1178 v->state = VPE_STATE_INUSE;
1185 EXPORT_SYMBOL(vpe_alloc);
1187 /* start running from here */
1188 int vpe_start(vpe_handle vpe, unsigned long start)
1190 struct vpe *v = vpe;
1196 EXPORT_SYMBOL(vpe_start);
1198 /* halt it for now */
1199 int vpe_stop(vpe_handle vpe)
1201 struct vpe *v = vpe;
1203 unsigned int evpe_flags;
1205 evpe_flags = dvpe();
1207 if ((t = list_entry(v->tc.next, struct tc, tc)) != NULL) {
1210 write_vpe_c0_vpeconf0(read_vpe_c0_vpeconf0() & ~VPECONF0_VPA);
1218 EXPORT_SYMBOL(vpe_stop);
1220 /* I've done with it thank you */
1221 int vpe_free(vpe_handle vpe)
1223 struct vpe *v = vpe;
1225 unsigned int evpe_flags;
1227 if ((t = list_entry(v->tc.next, struct tc, tc)) == NULL) {
1231 evpe_flags = dvpe();
1233 /* Put MVPE's into 'configuration state' */
1234 set_c0_mvpcontrol(MVPCONTROL_VPC);
1237 write_vpe_c0_vpeconf0(read_vpe_c0_vpeconf0() & ~VPECONF0_VPA);
1240 write_tc_c0_tchalt(TCHALT_H);
1243 /* mark the TC unallocated */
1244 write_tc_c0_tcstatus(read_tc_c0_tcstatus() & ~TCSTATUS_A);
1246 v->state = VPE_STATE_UNUSED;
1248 clear_c0_mvpcontrol(MVPCONTROL_VPC);
1254 EXPORT_SYMBOL(vpe_free);
1256 void *vpe_get_shared(int index)
1260 if ((v = get_vpe(index)) == NULL)
1263 return v->shared_ptr;
1266 EXPORT_SYMBOL(vpe_get_shared);
1268 int vpe_notify(int index, struct vpe_notifications *notify)
1272 if ((v = get_vpe(index)) == NULL)
1275 list_add(¬ify->list, &v->notify);
1279 EXPORT_SYMBOL(vpe_notify);
1281 char *vpe_getcwd(int index)
1285 if ((v = get_vpe(index)) == NULL)
1291 EXPORT_SYMBOL(vpe_getcwd);
1293 static ssize_t store_kill(struct device *dev, struct device_attribute *attr,
1294 const char *buf, size_t len)
1296 struct vpe *vpe = get_vpe(tclimit);
1297 struct vpe_notifications *not;
1299 list_for_each_entry(not, &vpe->notify, list) {
1303 release_progmem(vpe->load_addr);
1304 cleanup_tc(get_tc(tclimit));
1310 static DEVICE_ATTR(kill, S_IWUSR, NULL, store_kill);
1312 static ssize_t ntcs_show(struct device *cd, struct device_attribute *attr,
1315 struct vpe *vpe = get_vpe(tclimit);
1317 return sprintf(buf, "%d\n", vpe->ntcs);
1320 static ssize_t ntcs_store(struct device *dev, struct device_attribute *attr,
1321 const char *buf, size_t len)
1323 struct vpe *vpe = get_vpe(tclimit);
1327 new = simple_strtoul(buf, &endp, 0);
1331 if (new == 0 || new > (hw_tcs - tclimit))
1341 static DEVICE_ATTR_RW(ntcs);
1343 static struct attribute *vpe_attrs[] = {
1344 &dev_attr_kill.attr,
1345 &dev_attr_ntcs.attr,
1348 ATTRIBUTE_GROUPS(vpe);
1350 static void vpe_device_release(struct device *cd)
1355 struct class vpe_class = {
1357 .owner = THIS_MODULE,
1358 .dev_release = vpe_device_release,
1359 .dev_groups = vpe_groups,
1362 struct device vpe_device;
1364 static int __init vpe_module_init(void)
1366 unsigned int mtflags, vpflags;
1367 unsigned long flags, val;
1368 struct vpe *v = NULL;
1372 if (!cpu_has_mipsmt) {
1373 printk("VPE loader: not a MIPS MT capable processor\n");
1377 if (vpelimit == 0) {
1378 printk(KERN_WARNING "No VPEs reserved for AP/SP, not "
1379 "initializing VPE loader.\nPass maxvpes=<n> argument as "
1380 "kernel argument\n");
1386 printk(KERN_WARNING "No TCs reserved for AP/SP, not "
1387 "initializing VPE loader.\nPass maxtcs=<n> argument as "
1388 "kernel argument\n");
1393 major = register_chrdev(0, module_name, &vpe_fops);
1395 printk("VPE loader: unable to register character device\n");
1399 err = class_register(&vpe_class);
1401 printk(KERN_ERR "vpe_class registration failed\n");
1405 device_initialize(&vpe_device);
1406 vpe_device.class = &vpe_class,
1407 vpe_device.parent = NULL,
1408 dev_set_name(&vpe_device, "vpe1");
1409 vpe_device.devt = MKDEV(major, minor);
1410 err = device_add(&vpe_device);
1412 printk(KERN_ERR "Adding vpe_device failed\n");
1416 local_irq_save(flags);
1420 /* Put MVPE's into 'configuration state' */
1421 set_c0_mvpcontrol(MVPCONTROL_VPC);
1423 /* dump_mtregs(); */
1425 val = read_c0_mvpconf0();
1426 hw_tcs = (val & MVPCONF0_PTC) + 1;
1427 hw_vpes = ((val & MVPCONF0_PVPE) >> MVPCONF0_PVPE_SHIFT) + 1;
1429 for (tc = tclimit; tc < hw_tcs; tc++) {
1431 * Must re-enable multithreading temporarily or in case we
1432 * reschedule send IPIs or similar we might hang.
1434 clear_c0_mvpcontrol(MVPCONTROL_VPC);
1437 local_irq_restore(flags);
1444 local_irq_save(flags);
1447 set_c0_mvpcontrol(MVPCONTROL_VPC);
1453 if ((v = alloc_vpe(tc)) == NULL) {
1454 printk(KERN_WARNING "VPE: unable to allocate VPE\n");
1459 v->ntcs = hw_tcs - tclimit;
1461 /* add the tc to the list of this vpe's tc's. */
1462 list_add(&t->tc, &v->tc);
1464 /* deactivate all but vpe0 */
1465 if (tc >= tclimit) {
1466 unsigned long tmp = read_vpe_c0_vpeconf0();
1468 tmp &= ~VPECONF0_VPA;
1471 tmp |= VPECONF0_MVP;
1472 write_vpe_c0_vpeconf0(tmp);
1475 /* disable multi-threading with TC's */
1476 write_vpe_c0_vpecontrol(read_vpe_c0_vpecontrol() & ~VPECONTROL_TE);
1478 if (tc >= vpelimit) {
1480 * Set config to be the same as vpe0,
1481 * particularly kseg0 coherency alg
1483 write_vpe_c0_config(read_c0_config());
1488 t->pvpe = v; /* set the parent vpe */
1490 if (tc >= tclimit) {
1495 /* Any TC that is bound to VPE0 gets left as is - in case
1496 we are running SMTC on VPE0. A TC that is bound to any
1497 other VPE gets bound to VPE0, ideally I'd like to make
1498 it homeless but it doesn't appear to let me bind a TC
1499 to a non-existent VPE. Which is perfectly reasonable.
1501 The (un)bound state is visible to an EJTAG probe so may
1505 if (((tmp = read_tc_c0_tcbind()) & TCBIND_CURVPE)) {
1506 /* tc is bound >vpe0 */
1507 write_tc_c0_tcbind(tmp & ~TCBIND_CURVPE);
1509 t->pvpe = get_vpe(0); /* set the parent vpe */
1513 write_tc_c0_tchalt(TCHALT_H);
1516 tmp = read_tc_c0_tcstatus();
1518 /* mark not activated and not dynamically allocatable */
1519 tmp &= ~(TCSTATUS_A | TCSTATUS_DA);
1520 tmp |= TCSTATUS_IXMT; /* interrupt exempt */
1521 write_tc_c0_tcstatus(tmp);
1526 /* release config state */
1527 clear_c0_mvpcontrol(MVPCONTROL_VPC);
1531 local_irq_restore(flags);
1536 class_unregister(&vpe_class);
1538 unregister_chrdev(major, module_name);
1544 static void __exit vpe_module_exit(void)
1548 device_del(&vpe_device);
1549 unregister_chrdev(major, module_name);
1551 /* No locking needed here */
1552 list_for_each_entry_safe(v, n, &vpecontrol.vpe_list, list) {
1553 if (v->state != VPE_STATE_UNUSED)
1558 module_init(vpe_module_init);
1559 module_exit(vpe_module_exit);
1560 MODULE_DESCRIPTION("MIPS VPE Loader");
1561 MODULE_AUTHOR("Elizabeth Oldham, MIPS Technologies, Inc.");
1562 MODULE_LICENSE("GPL");