File: /usr/src/linux/arch/sparc64/mm/fault.c

1     /* $Id: fault.c,v 1.58 2001/09/01 00:11:16 kanoj Exp $
2      * arch/sparc64/mm/fault.c: Page fault handlers for the 64-bit Sparc.
3      *
4      * Copyright (C) 1996 David S. Miller (davem@caip.rutgers.edu)
5      * Copyright (C) 1997, 1999 Jakub Jelinek (jj@ultra.linux.cz)
6      */
7     
8     #include <asm/head.h>
9     
10     #include <linux/string.h>
11     #include <linux/types.h>
12     #include <linux/ptrace.h>
13     #include <linux/mman.h>
14     #include <linux/signal.h>
15     #include <linux/mm.h>
16     #include <linux/smp_lock.h>
17     #include <linux/init.h>
18     #include <linux/interrupt.h>
19     
20     #include <asm/page.h>
21     #include <asm/pgtable.h>
22     #include <asm/openprom.h>
23     #include <asm/oplib.h>
24     #include <asm/uaccess.h>
25     #include <asm/asi.h>
26     #include <asm/lsu.h>
27     
28     #define ELEMENTS(arr) (sizeof (arr)/sizeof (arr[0]))
29     
30     extern struct sparc_phys_banks sp_banks[SPARC_PHYS_BANKS];
31     
32     /*
33      * To debug kernel during syscall entry.
34      */
35     void syscall_trace_entry(struct pt_regs *regs)
36     {
37     	printk("scall entry: %s[%d]/cpu%d: %d\n", current->comm, current->pid, smp_processor_id(), (int) regs->u_regs[UREG_G1]);
38     }
39     
40     /*
41      * To debug kernel during syscall exit.
42      */
43     void syscall_trace_exit(struct pt_regs *regs)
44     {
45     	printk("scall exit: %s[%d]/cpu%d: %d\n", current->comm, current->pid, smp_processor_id(), (int) regs->u_regs[UREG_G1]);
46     }
47     
48     /*
49      * To debug kernel to catch accesses to certain virtual/physical addresses.
50      * Mode = 0 selects physical watchpoints, mode = 1 selects virtual watchpoints.
51      * flags = VM_READ watches memread accesses, flags = VM_WRITE watches memwrite accesses.
52      * Caller passes in a 64bit aligned addr, with mask set to the bytes that need to be
53      * watched. This is only useful on a single cpu machine for now. After the watchpoint
54      * is detected, the process causing it will be killed, thus preventing an infinite loop.
55      */
56     void set_brkpt(unsigned long addr, unsigned char mask, int flags, int mode)
57     {
58     	unsigned long lsubits = LSU_CONTROL_IC|LSU_CONTROL_DC|LSU_CONTROL_IM|LSU_CONTROL_DM;
59     
60     	__asm__ __volatile__("stxa	%0, [%1] %2\n\t"
61     			     "membar	#Sync"
62     			     : /* no outputs */
63     			     : "r" (addr), "r" (mode ? VIRT_WATCHPOINT : PHYS_WATCHPOINT),
64     			       "i" (ASI_DMMU));
65     	lsubits |= ((unsigned long)mask << (mode ? 25 : 33));
66     	if (flags & VM_READ)
67     		lsubits |= (mode ? LSU_CONTROL_VR : LSU_CONTROL_PR);
68     	if (flags & VM_WRITE)
69     		lsubits |= (mode ? LSU_CONTROL_VW : LSU_CONTROL_PW);
70     	__asm__ __volatile__("stxa %0, [%%g0] %1\n\t"
71     			     "membar #Sync"
72     			     : /* no outputs */
73     			     : "r" (lsubits), "i" (ASI_LSU_CONTROL)
74     			     : "memory");
75     }
76     
77     /* Nice, simple, prom library does all the sweating for us. ;) */
78     unsigned long __init prom_probe_memory (void)
79     {
80     	register struct linux_mlist_p1275 *mlist;
81     	register unsigned long bytes, base_paddr, tally;
82     	register int i;
83     
84     	i = 0;
85     	mlist = *prom_meminfo()->p1275_available;
86     	bytes = tally = mlist->num_bytes;
87     	base_paddr = mlist->start_adr;
88       
89     	sp_banks[0].base_addr = base_paddr;
90     	sp_banks[0].num_bytes = bytes;
91     
92     	while (mlist->theres_more != (void *) 0) {
93     		i++;
94     		mlist = mlist->theres_more;
95     		bytes = mlist->num_bytes;
96     		tally += bytes;
97     		if (i >= SPARC_PHYS_BANKS-1) {
98     			printk ("The machine has more banks than "
99     				"this kernel can support\n"
100     				"Increase the SPARC_PHYS_BANKS "
101     				"setting (currently %d)\n",
102     				SPARC_PHYS_BANKS);
103     			i = SPARC_PHYS_BANKS-1;
104     			break;
105     		}
106         
107     		sp_banks[i].base_addr = mlist->start_adr;
108     		sp_banks[i].num_bytes = mlist->num_bytes;
109     	}
110     
111     	i++;
112     	sp_banks[i].base_addr = 0xdeadbeefbeefdeadUL;
113     	sp_banks[i].num_bytes = 0;
114     
115     	/* Now mask all bank sizes on a page boundary, it is all we can
116     	 * use anyways.
117     	 */
118     	for (i = 0; sp_banks[i].num_bytes != 0; i++)
119     		sp_banks[i].num_bytes &= PAGE_MASK;
120     
121     	return tally;
122     }
123     
124     void unhandled_fault(unsigned long address, struct task_struct *tsk,
125                          struct pt_regs *regs)
126     {
127     	if ((unsigned long) address < PAGE_SIZE) {
128     		printk(KERN_ALERT "Unable to handle kernel NULL "
129     		       "pointer dereference\n");
130     	} else {
131     		printk(KERN_ALERT "Unable to handle kernel paging request "
132     		       "at virtual address %016lx\n", (unsigned long)address);
133     	}
134     	printk(KERN_ALERT "tsk->{mm,active_mm}->context = %016lx\n",
135     	       (tsk->mm ? tsk->mm->context : tsk->active_mm->context));
136     	printk(KERN_ALERT "tsk->{mm,active_mm}->pgd = %016lx\n",
137     	       (tsk->mm ? (unsigned long) tsk->mm->pgd :
138     		          (unsigned long) tsk->active_mm->pgd));
139     	die_if_kernel("Oops", regs);
140     }
141     
142     /*
143      * We now make sure that mmap_sem is held in all paths that call 
144      * this. Additionally, to prevent kswapd from ripping ptes from
145      * under us, raise interrupts around the time that we look at the
146      * pte, kswapd will have to wait to get his smp ipi response from
147      * us. This saves us having to get page_table_lock.
148      */
149     static unsigned int get_user_insn(unsigned long tpc)
150     {
151     	pgd_t *pgdp = pgd_offset(current->mm, tpc);
152     	pmd_t *pmdp;
153     	pte_t *ptep, pte;
154     	unsigned long pa;
155     	u32 insn = 0;
156     	unsigned long pstate;
157     
158     	if (pgd_none(*pgdp))
159     		goto outret;
160     	pmdp = pmd_offset(pgdp, tpc);
161     	if (pmd_none(*pmdp))
162     		goto outret;
163     	ptep = pte_offset(pmdp, tpc);
164     	__asm__ __volatile__("rdpr %%pstate, %0" : "=r" (pstate));
165     	__asm__ __volatile__("wrpr %0, %1, %%pstate"
166     				: : "r" (pstate), "i" (PSTATE_IE));
167     	pte = *ptep;
168     	if (!pte_present(pte))
169     		goto out;
170     
171     	pa  = (pte_val(pte) & _PAGE_PADDR);
172     	pa += (tpc & ~PAGE_MASK);
173     
174     	/* Use phys bypass so we don't pollute dtlb/dcache. */
175     	__asm__ __volatile__("lduwa [%1] %2, %0"
176     			     : "=r" (insn)
177     			     : "r" (pa), "i" (ASI_PHYS_USE_EC));
178     
179     out:
180     	__asm__ __volatile__("wrpr %0, 0x0, %%pstate" : : "r" (pstate));
181     outret:
182     	return insn;
183     }
184     
185     static void do_fault_siginfo(int code, int sig, unsigned long address)
186     {
187     	siginfo_t info;
188     
189     	info.si_code = code;
190     	info.si_signo = sig;
191     	info.si_errno = 0;
192     	info.si_addr = (void *) address;
193     	info.si_trapno = 0;
194     	force_sig_info(sig, &info, current);
195     }
196     
197     extern int handle_ldf_stq(u32, struct pt_regs *);
198     extern int handle_ld_nf(u32, struct pt_regs *);
199     
200     static inline unsigned int get_fault_insn(struct pt_regs *regs, unsigned int insn)
201     {
202     	if (!insn) {
203     		if (!regs->tpc || (regs->tpc & 0x3))
204     			return 0;
205     		if (regs->tstate & TSTATE_PRIV) {
206     			insn = *(unsigned int *)regs->tpc;
207     		} else {
208     			insn = get_user_insn(regs->tpc);
209     		}
210     	}
211     	return insn;
212     }
213     
214     static void do_kernel_fault(struct pt_regs *regs, int si_code, int fault_code,
215     			    unsigned int insn, unsigned long address)
216     {
217     	unsigned long g2;
218     	unsigned char asi = ASI_P;
219      
220     	if ((!insn) && (regs->tstate & TSTATE_PRIV))
221     		goto cannot_handle;
222     
223     	/* If user insn could be read (thus insn is zero), that
224     	 * is fine.  We will just gun down the process with a signal
225     	 * in that case.
226     	 */
227     
228     	if (!(fault_code & FAULT_CODE_WRITE) &&
229     	    (insn & 0xc0800000) == 0xc0800000) {
230     		if (insn & 0x2000)
231     			asi = (regs->tstate >> 24);
232     		else
233     			asi = (insn >> 5);
234     		if ((asi & 0xf2) == 0x82) {
235     			if (insn & 0x1000000) {
236     				handle_ldf_stq(insn, regs);
237     			} else {
238     				/* This was a non-faulting load. Just clear the
239     				 * destination register(s) and continue with the next
240     				 * instruction. -jj
241     				 */
242     				handle_ld_nf(insn, regs);
243     			}
244     			return;
245     		}
246     	}
247     		
248     	g2 = regs->u_regs[UREG_G2];
249     
250     	/* Is this in ex_table? */
251     	if (regs->tstate & TSTATE_PRIV) {
252     		unsigned long fixup;
253     
254     		if (asi == ASI_P && (insn & 0xc0800000) == 0xc0800000) {
255     			if (insn & 0x2000)
256     				asi = (regs->tstate >> 24);
257     			else
258     				asi = (insn >> 5);
259     		}
260     	
261     		/* Look in asi.h: All _S asis have LS bit set */
262     		if ((asi & 0x1) &&
263     		    (fixup = search_exception_table (regs->tpc, &g2))) {
264     			regs->tpc = fixup;
265     			regs->tnpc = regs->tpc + 4;
266     			regs->u_regs[UREG_G2] = g2;
267     			return;
268     		}
269     	} else {
270     		/* The si_code was set to make clear whether
271     		 * this was a SEGV_MAPERR or SEGV_ACCERR fault.
272     		 */
273     		do_fault_siginfo(si_code, SIGSEGV, address);
274     		return;
275     	}
276     
277     cannot_handle:
278     	unhandled_fault (address, current, regs);
279     }
280     
281     asmlinkage void do_sparc64_fault(struct pt_regs *regs)
282     {
283     	struct mm_struct *mm = current->mm;
284     	struct vm_area_struct *vma;
285     	unsigned int insn = 0;
286     	int si_code, fault_code;
287     	unsigned long address;
288     
289     	si_code = SEGV_MAPERR;
290     	fault_code = current->thread.fault_code;
291     	address = current->thread.fault_address;
292     
293     	if ((fault_code & FAULT_CODE_ITLB) &&
294     	    (fault_code & FAULT_CODE_DTLB))
295     		BUG();
296     
297     	/*
298     	 * If we're in an interrupt or have no user
299     	 * context, we must not take the fault..
300     	 */
301     	if (in_interrupt() || !mm)
302     		goto intr_or_no_mm;
303     
304     	if ((current->thread.flags & SPARC_FLAG_32BIT) != 0) {
305     		regs->tpc &= 0xffffffff;
306     		address &= 0xffffffff;
307     	}
308     
309     	down_read(&mm->mmap_sem);
310     	vma = find_vma(mm, address);
311     	if (!vma)
312     		goto bad_area;
313     
314     	/* Pure DTLB misses do not tell us whether the fault causing
315     	 * load/store/atomic was a write or not, it only says that there
316     	 * was no match.  So in such a case we (carefully) read the
317     	 * instruction to try and figure this out.  It's an optimization
318     	 * so it's ok if we can't do this.
319     	 *
320     	 * Special hack, window spill/fill knows the exact fault type.
321     	 */
322     	if (((fault_code &
323     	      (FAULT_CODE_DTLB | FAULT_CODE_WRITE | FAULT_CODE_WINFIXUP)) == FAULT_CODE_DTLB) &&
324     	    (vma->vm_flags & VM_WRITE) != 0) {
325     		insn = get_fault_insn(regs, 0);
326     		if (!insn)
327     			goto continue_fault;
328     		if ((insn & 0xc0200000) == 0xc0200000 &&
329     		    (insn & 0x1780000) != 0x1680000) {
330     			/* Don't bother updating thread struct value,
331     			 * because update_mmu_cache only cares which tlb
332     			 * the access came from.
333     			 */
334     			fault_code |= FAULT_CODE_WRITE;
335     		}
336     	}
337     continue_fault:
338     
339     	if (vma->vm_start <= address)
340     		goto good_area;
341     	if (!(vma->vm_flags & VM_GROWSDOWN))
342     		goto bad_area;
343     	if (expand_stack(vma, address))
344     		goto bad_area;
345     	/*
346     	 * Ok, we have a good vm_area for this memory access, so
347     	 * we can handle it..
348     	 */
349     good_area:
350     	si_code = SEGV_ACCERR;
351     	if (fault_code & FAULT_CODE_WRITE) {
352     		if (!(vma->vm_flags & VM_WRITE))
353     			goto bad_area;
354     
355     		/* Spitfire has an icache which does not snoop
356     		 * processor stores.  Later processors do...
357     		 */
358     		if (tlb_type == spitfire &&
359     		    (vma->vm_flags & VM_EXEC) != 0 &&
360     		    vma->vm_file != NULL)
361     			current->thread.use_blkcommit = 1;
362     	} else {
363     		/* Allow reads even for write-only mappings */
364     		if (!(vma->vm_flags & (VM_READ | VM_EXEC)))
365     			goto bad_area;
366     	}
367     
368     	switch (handle_mm_fault(mm, vma, address, (fault_code & FAULT_CODE_WRITE))) {
369     	case 1:
370     		current->min_flt++;
371     		break;
372     	case 2:
373     		current->maj_flt++;
374     		break;
375     	case 0:
376     		goto do_sigbus;
377     	default:
378     		goto out_of_memory;
379     	}
380     
381     	up_read(&mm->mmap_sem);
382     	goto fault_done;
383     
384     	/*
385     	 * Something tried to access memory that isn't in our memory map..
386     	 * Fix it, but check if it's kernel or user first..
387     	 */
388     bad_area:
389     	insn = get_fault_insn(regs, insn);
390     	up_read(&mm->mmap_sem);
391     
392     handle_kernel_fault:
393     	do_kernel_fault(regs, si_code, fault_code, insn, address);
394     
395     	goto fault_done;
396     
397     /*
398      * We ran out of memory, or some other thing happened to us that made
399      * us unable to handle the page fault gracefully.
400      */
401     out_of_memory:
402     	insn = get_fault_insn(regs, insn);
403     	up_read(&mm->mmap_sem);
404     	printk("VM: killing process %s\n", current->comm);
405     	if (!(regs->tstate & TSTATE_PRIV))
406     		do_exit(SIGKILL);
407     	goto handle_kernel_fault;
408     
409     intr_or_no_mm:
410     	insn = get_fault_insn(regs, 0);
411     	goto handle_kernel_fault;
412     
413     do_sigbus:
414     	insn = get_fault_insn(regs, insn);
415     	up_read(&mm->mmap_sem);
416     
417     	/*
418     	 * Send a sigbus, regardless of whether we were in kernel
419     	 * or user mode.
420     	 */
421     	do_fault_siginfo(BUS_ADRERR, SIGBUS, address);
422     
423     	/* Kernel mode? Handle exceptions or die */
424     	if (regs->tstate & TSTATE_PRIV)
425     		goto handle_kernel_fault;
426     
427     fault_done:
428     	/* These values are no longer needed, clear them. */
429     	current->thread.fault_code = 0;
430     	current->thread.use_blkcommit = 0;
431     	current->thread.fault_address = 0;
432     }
433