1 /* Copyright (C) 2004 Mips Technologies, Inc */
3 #include <linux/clockchips.h>
4 #include <linux/kernel.h>
5 #include <linux/sched.h>
6 #include <linux/cpumask.h>
7 #include <linux/interrupt.h>
8 #include <linux/kernel_stat.h>
9 #include <linux/module.h>
12 #include <asm/processor.h>
13 #include <asm/atomic.h>
14 #include <asm/system.h>
15 #include <asm/hardirq.h>
16 #include <asm/hazards.h>
18 #include <asm/mmu_context.h>
19 #include <asm/mipsregs.h>
20 #include <asm/cacheflush.h>
22 #include <asm/addrspace.h>
24 #include <asm/smtc_ipi.h>
25 #include <asm/smtc_proc.h>
28 * SMTC Kernel needs to manipulate low-level CPU interrupt mask
29 * in do_IRQ. These are passed in setup_irq_smtc() and stored
32 unsigned long irq_hwmask[NR_IRQS];
34 #define LOCK_MT_PRA() \
35 local_irq_save(flags); \
38 #define UNLOCK_MT_PRA() \
40 local_irq_restore(flags)
42 #define LOCK_CORE_PRA() \
43 local_irq_save(flags); \
46 #define UNLOCK_CORE_PRA() \
48 local_irq_restore(flags)
51 * Data structures purely associated with SMTC parallelism
56 * Table for tracking ASIDs whose lifetime is prolonged.
59 asiduse smtc_live_asid[MAX_SMTC_TLBS][MAX_SMTC_ASIDS];
62 * Clock interrupt "latch" buffers, per "CPU"
65 static atomic_t ipi_timer_latch[NR_CPUS];
68 * Number of InterProcessor Interrupt (IPI) message buffers to allocate
71 #define IPIBUF_PER_CPU 4
73 static struct smtc_ipi_q IPIQ[NR_CPUS];
74 static struct smtc_ipi_q freeIPIq;
77 /* Forward declarations */
79 void ipi_decode(struct smtc_ipi *);
80 static void post_direct_ipi(int cpu, struct smtc_ipi *pipi);
81 static void setup_cross_vpe_interrupts(unsigned int nvpe);
82 void init_smtc_stats(void);
84 /* Global SMTC Status */
86 unsigned int smtc_status = 0;
88 /* Boot command line configuration overrides */
91 static int ipibuffers = 0;
92 static int nostlb = 0;
93 static int asidmask = 0;
94 unsigned long smtc_asid_mask = 0xff;
96 static int __init vpe0tcs(char *str)
98 get_option(&str, &vpe0limit);
103 static int __init ipibufs(char *str)
105 get_option(&str, &ipibuffers);
109 static int __init stlb_disable(char *s)
115 static int __init asidmask_set(char *str)
117 get_option(&str, &asidmask);
127 smtc_asid_mask = (unsigned long)asidmask;
130 printk("ILLEGAL ASID mask 0x%x from command line\n", asidmask);
135 __setup("vpe0tcs=", vpe0tcs);
136 __setup("ipibufs=", ipibufs);
137 __setup("nostlb", stlb_disable);
138 __setup("asidmask=", asidmask_set);
140 #ifdef CONFIG_SMTC_IDLE_HOOK_DEBUG
142 static int hang_trig = 0;
144 static int __init hangtrig_enable(char *s)
151 __setup("hangtrig", hangtrig_enable);
153 #define DEFAULT_BLOCKED_IPI_LIMIT 32
155 static int timerq_limit = DEFAULT_BLOCKED_IPI_LIMIT;
157 static int __init tintq(char *str)
159 get_option(&str, &timerq_limit);
163 __setup("tintq=", tintq);
165 static int imstuckcount[2][8];
166 /* vpemask represents IM/IE bits of per-VPE Status registers, low-to-high */
167 static int vpemask[2][8] = {
168 {0, 0, 1, 0, 0, 0, 0, 1},
169 {0, 0, 0, 0, 0, 0, 0, 1}
171 int tcnoprog[NR_CPUS];
172 static atomic_t idle_hook_initialized = {0};
173 static int clock_hang_reported[NR_CPUS];
175 #endif /* CONFIG_SMTC_IDLE_HOOK_DEBUG */
178 * Configure shared TLB - VPC configuration bit must be set by caller
181 static void smtc_configure_tlb(void)
184 unsigned long mvpconf0;
185 unsigned long config1val;
187 /* Set up ASID preservation table */
188 for (vpes=0; vpes<MAX_SMTC_TLBS; vpes++) {
189 for(i = 0; i < MAX_SMTC_ASIDS; i++) {
190 smtc_live_asid[vpes][i] = 0;
193 mvpconf0 = read_c0_mvpconf0();
195 if ((vpes = ((mvpconf0 & MVPCONF0_PVPE)
196 >> MVPCONF0_PVPE_SHIFT) + 1) > 1) {
197 /* If we have multiple VPEs, try to share the TLB */
198 if ((mvpconf0 & MVPCONF0_TLBS) && !nostlb) {
200 * If TLB sizing is programmable, shared TLB
201 * size is the total available complement.
202 * Otherwise, we have to take the sum of all
203 * static VPE TLB entries.
205 if ((tlbsiz = ((mvpconf0 & MVPCONF0_PTLBE)
206 >> MVPCONF0_PTLBE_SHIFT)) == 0) {
208 * If there's more than one VPE, there had better
209 * be more than one TC, because we need one to bind
210 * to each VPE in turn to be able to read
211 * its configuration state!
214 /* Stop the TC from doing anything foolish */
215 write_tc_c0_tchalt(TCHALT_H);
217 /* No need to un-Halt - that happens later anyway */
218 for (i=0; i < vpes; i++) {
219 write_tc_c0_tcbind(i);
221 * To be 100% sure we're really getting the right
222 * information, we exit the configuration state
223 * and do an IHB after each rebinding.
226 read_c0_mvpcontrol() & ~ MVPCONTROL_VPC );
229 * Only count if the MMU Type indicated is TLB
231 if (((read_vpe_c0_config() & MIPS_CONF_MT) >> 7) == 1) {
232 config1val = read_vpe_c0_config1();
233 tlbsiz += ((config1val >> 25) & 0x3f) + 1;
236 /* Put core back in configuration state */
238 read_c0_mvpcontrol() | MVPCONTROL_VPC );
242 write_c0_mvpcontrol(read_c0_mvpcontrol() | MVPCONTROL_STLB);
246 * Setup kernel data structures to use software total,
247 * rather than read the per-VPE Config1 value. The values
248 * for "CPU 0" gets copied to all the other CPUs as part
249 * of their initialization in smtc_cpu_setup().
252 /* MIPS32 limits TLB indices to 64 */
255 cpu_data[0].tlbsize = current_cpu_data.tlbsize = tlbsiz;
256 smtc_status |= SMTC_TLB_SHARED;
257 local_flush_tlb_all();
259 printk("TLB of %d entry pairs shared by %d VPEs\n",
262 printk("WARNING: TLB Not Sharable on SMTC Boot!\n");
269 * Incrementally build the CPU map out of constituent MIPS MT cores,
270 * using the specified available VPEs and TCs. Plaform code needs
271 * to ensure that each MIPS MT core invokes this routine on reset,
274 * This version of the build_cpu_map and prepare_cpus routines assumes
275 * that *all* TCs of a MIPS MT core will be used for Linux, and that
276 * they will be spread across *all* available VPEs (to minimise the
277 * loss of efficiency due to exception service serialization).
278 * An improved version would pick up configuration information and
279 * possibly leave some TCs/VPEs as "slave" processors.
281 * Use c0_MVPConf0 to find out how many TCs are available, setting up
282 * phys_cpu_present_map and the logical/physical mappings.
285 int __init mipsmt_build_cpu_map(int start_cpu_slot)
290 * The CPU map isn't actually used for anything at this point,
291 * so it's not clear what else we should do apart from set
292 * everything up so that "logical" = "physical".
294 ntcs = ((read_c0_mvpconf0() & MVPCONF0_PTC) >> MVPCONF0_PTC_SHIFT) + 1;
295 for (i=start_cpu_slot; i<NR_CPUS && i<ntcs; i++) {
296 cpu_set(i, phys_cpu_present_map);
297 __cpu_number_map[i] = i;
298 __cpu_logical_map[i] = i;
300 #ifdef CONFIG_MIPS_MT_FPAFF
301 /* Initialize map of CPUs with FPUs */
302 cpus_clear(mt_fpu_cpumask);
305 /* One of those TC's is the one booting, and not a secondary... */
306 printk("%i available secondary CPU TC(s)\n", i - 1);
312 * Common setup before any secondaries are started
313 * Make sure all CPU's are in a sensible state before we boot any of the
316 * For MIPS MT "SMTC" operation, we set up all TCs, spread as evenly
317 * as possible across the available VPEs.
320 static void smtc_tc_setup(int vpe, int tc, int cpu)
323 write_tc_c0_tchalt(TCHALT_H);
325 write_tc_c0_tcstatus((read_tc_c0_tcstatus()
326 & ~(TCSTATUS_TKSU | TCSTATUS_DA | TCSTATUS_IXMT))
328 write_tc_c0_tccontext(0);
330 write_tc_c0_tcbind(vpe);
331 /* In general, all TCs should have the same cpu_data indications */
332 memcpy(&cpu_data[cpu], &cpu_data[0], sizeof(struct cpuinfo_mips));
333 /* For 34Kf, start with TC/CPU 0 as sole owner of single FPU context */
334 if (cpu_data[0].cputype == CPU_34K ||
335 cpu_data[0].cputype == CPU_1004K)
336 cpu_data[cpu].options &= ~MIPS_CPU_FPU;
337 cpu_data[cpu].vpe_id = vpe;
338 cpu_data[cpu].tc_id = tc;
342 void mipsmt_prepare_cpus(void)
344 int i, vpe, tc, ntc, nvpe, tcpervpe[NR_CPUS], slop, cpu;
348 struct smtc_ipi *pipi;
350 /* disable interrupts so we can disable MT */
351 local_irq_save(flags);
352 /* disable MT so we can configure */
356 spin_lock_init(&freeIPIq.lock);
359 * We probably don't have as many VPEs as we do SMP "CPUs",
360 * but it's possible - and in any case we'll never use more!
362 for (i=0; i<NR_CPUS; i++) {
363 IPIQ[i].head = IPIQ[i].tail = NULL;
364 spin_lock_init(&IPIQ[i].lock);
366 atomic_set(&ipi_timer_latch[i], 0);
369 /* cpu_data index starts at zero */
371 cpu_data[cpu].vpe_id = 0;
372 cpu_data[cpu].tc_id = 0;
375 /* Report on boot-time options */
376 mips_mt_set_cpuoptions();
378 printk("Limit of %d VPEs set\n", vpelimit);
380 printk("Limit of %d TCs set\n", tclimit);
382 printk("Shared TLB Use Inhibited - UNSAFE for Multi-VPE Operation\n");
385 printk("ASID mask value override to 0x%x\n", asidmask);
388 #ifdef CONFIG_SMTC_IDLE_HOOK_DEBUG
390 printk("Logic Analyser Trigger on suspected TC hang\n");
391 #endif /* CONFIG_SMTC_IDLE_HOOK_DEBUG */
393 /* Put MVPE's into 'configuration state' */
394 write_c0_mvpcontrol( read_c0_mvpcontrol() | MVPCONTROL_VPC );
396 val = read_c0_mvpconf0();
397 nvpe = ((val & MVPCONF0_PVPE) >> MVPCONF0_PVPE_SHIFT) + 1;
398 if (vpelimit > 0 && nvpe > vpelimit)
400 ntc = ((val & MVPCONF0_PTC) >> MVPCONF0_PTC_SHIFT) + 1;
403 if (tclimit > 0 && ntc > tclimit)
406 for (i = 0; i < nvpe; i++) {
407 tcpervpe[i] = ntc / nvpe;
409 if((slop - i) > 0) tcpervpe[i]++;
412 /* Handle command line override for VPE0 */
413 if (vpe0limit > ntc) vpe0limit = ntc;
416 if (vpe0limit < tcpervpe[0]) {
417 /* Reducing TC count - distribute to others */
418 slop = tcpervpe[0] - vpe0limit;
419 slopslop = slop % (nvpe - 1);
420 tcpervpe[0] = vpe0limit;
421 for (i = 1; i < nvpe; i++) {
422 tcpervpe[i] += slop / (nvpe - 1);
423 if(slopslop && ((slopslop - (i - 1) > 0)))
426 } else if (vpe0limit > tcpervpe[0]) {
427 /* Increasing TC count - steal from others */
428 slop = vpe0limit - tcpervpe[0];
429 slopslop = slop % (nvpe - 1);
430 tcpervpe[0] = vpe0limit;
431 for (i = 1; i < nvpe; i++) {
432 tcpervpe[i] -= slop / (nvpe - 1);
433 if(slopslop && ((slopslop - (i - 1) > 0)))
439 /* Set up shared TLB */
440 smtc_configure_tlb();
442 for (tc = 0, vpe = 0 ; (vpe < nvpe) && (tc < ntc) ; vpe++) {
447 write_vpe_c0_vpeconf0(read_vpe_c0_vpeconf0() | VPECONF0_MVP);
450 printk("VPE %d: TC", vpe);
451 for (i = 0; i < tcpervpe[vpe]; i++) {
453 * TC 0 is bound to VPE 0 at reset,
454 * and is presumably executing this
455 * code. Leave it alone!
458 smtc_tc_setup(vpe, tc, cpu);
466 * Clear any stale software interrupts from VPE's Cause
468 write_vpe_c0_cause(0);
471 * Clear ERL/EXL of VPEs other than 0
472 * and set restricted interrupt enable/mask.
474 write_vpe_c0_status((read_vpe_c0_status()
475 & ~(ST0_BEV | ST0_ERL | ST0_EXL | ST0_IM))
476 | (STATUSF_IP0 | STATUSF_IP1 | STATUSF_IP7
479 * set config to be the same as vpe0,
480 * particularly kseg0 coherency alg
482 write_vpe_c0_config(read_c0_config());
483 /* Clear any pending timer interrupt */
484 write_vpe_c0_compare(0);
485 /* Propagate Config7 */
486 write_vpe_c0_config7(read_c0_config7());
487 write_vpe_c0_count(read_c0_count());
489 /* enable multi-threading within VPE */
490 write_vpe_c0_vpecontrol(read_vpe_c0_vpecontrol() | VPECONTROL_TE);
492 write_vpe_c0_vpeconf0(read_vpe_c0_vpeconf0() | VPECONF0_VPA);
496 * Pull any physically present but unused TCs out of circulation.
498 while (tc < (((val & MVPCONF0_PTC) >> MVPCONF0_PTC_SHIFT) + 1)) {
499 cpu_clear(tc, phys_cpu_present_map);
500 cpu_clear(tc, cpu_present_map);
504 /* release config state */
505 write_c0_mvpcontrol( read_c0_mvpcontrol() & ~ MVPCONTROL_VPC );
509 /* Set up coprocessor affinity CPU mask(s) */
511 #ifdef CONFIG_MIPS_MT_FPAFF
512 for (tc = 0; tc < ntc; tc++) {
513 if (cpu_data[tc].options & MIPS_CPU_FPU)
514 cpu_set(tc, mt_fpu_cpumask);
518 /* set up ipi interrupts... */
520 /* If we have multiple VPEs running, set up the cross-VPE interrupt */
522 setup_cross_vpe_interrupts(nvpe);
524 /* Set up queue of free IPI "messages". */
525 nipi = NR_CPUS * IPIBUF_PER_CPU;
529 pipi = kmalloc(nipi *sizeof(struct smtc_ipi), GFP_KERNEL);
531 panic("kmalloc of IPI message buffers failed\n");
533 printk("IPI buffer pool of %d buffers\n", nipi);
534 for (i = 0; i < nipi; i++) {
535 smtc_ipi_nq(&freeIPIq, pipi);
539 /* Arm multithreading and enable other VPEs - but all TCs are Halted */
542 local_irq_restore(flags);
543 /* Initialize SMTC /proc statistics/diagnostics */
549 * Setup the PC, SP, and GP of a secondary processor and start it
551 * smp_bootstrap is the place to resume from
552 * __KSTK_TOS(idle) is apparently the stack pointer
553 * (unsigned long)idle->thread_info the gp
556 void __cpuinit smtc_boot_secondary(int cpu, struct task_struct *idle)
558 extern u32 kernelsp[NR_CPUS];
563 if (cpu_data[cpu].vpe_id != cpu_data[smp_processor_id()].vpe_id) {
566 settc(cpu_data[cpu].tc_id);
569 write_tc_c0_tcrestart((unsigned long)&smp_bootstrap);
572 kernelsp[cpu] = __KSTK_TOS(idle);
573 write_tc_gpr_sp(__KSTK_TOS(idle));
576 write_tc_gpr_gp((unsigned long)task_thread_info(idle));
578 smtc_status |= SMTC_MTC_ACTIVE;
579 write_tc_c0_tchalt(0);
580 if (cpu_data[cpu].vpe_id != cpu_data[smp_processor_id()].vpe_id) {
586 void smtc_init_secondary(void)
589 * Start timer on secondary VPEs if necessary.
590 * plat_timer_setup has already have been invoked by init/main
591 * on "boot" TC. Like per_cpu_trap_init() hack, this assumes that
592 * SMTC init code assigns TCs consdecutively and in ascending order
593 * to across available VPEs.
595 if (((read_c0_tcbind() & TCBIND_CURTC) != 0) &&
596 ((read_c0_tcbind() & TCBIND_CURVPE)
597 != cpu_data[smp_processor_id() - 1].vpe_id)){
598 write_c0_compare(read_c0_count() + mips_hpt_frequency/HZ);
604 void smtc_smp_finish(void)
606 printk("TC %d going on-line as CPU %d\n",
607 cpu_data[smp_processor_id()].tc_id, smp_processor_id());
610 void smtc_cpus_done(void)
615 * Support for SMTC-optimized driver IRQ registration
619 * SMTC Kernel needs to manipulate low-level CPU interrupt mask
620 * in do_IRQ. These are passed in setup_irq_smtc() and stored
624 int setup_irq_smtc(unsigned int irq, struct irqaction * new,
625 unsigned long hwmask)
627 #ifdef CONFIG_SMTC_IDLE_HOOK_DEBUG
628 unsigned int vpe = current_cpu_data.vpe_id;
630 vpemask[vpe][irq - MIPS_CPU_IRQ_BASE] = 1;
632 irq_hwmask[irq] = hwmask;
634 return setup_irq(irq, new);
637 #ifdef CONFIG_MIPS_MT_SMTC_IRQAFF
639 * Support for IRQ affinity to TCs
642 void smtc_set_irq_affinity(unsigned int irq, cpumask_t affinity)
645 * If a "fast path" cache of quickly decodable affinity state
646 * is maintained, this is where it gets done, on a call up
647 * from the platform affinity code.
651 void smtc_forward_irq(unsigned int irq)
656 * OK wise guy, now figure out how to get the IRQ
657 * to be serviced on an authorized "CPU".
659 * Ideally, to handle the situation where an IRQ has multiple
660 * eligible CPUS, we would maintain state per IRQ that would
661 * allow a fair distribution of service requests. Since the
662 * expected use model is any-or-only-one, for simplicity
663 * and efficiency, we just pick the easiest one to find.
666 target = first_cpu(irq_desc[irq].affinity);
669 * We depend on the platform code to have correctly processed
670 * IRQ affinity change requests to ensure that the IRQ affinity
671 * mask has been purged of bits corresponding to nonexistent and
672 * offline "CPUs", and to TCs bound to VPEs other than the VPE
673 * connected to the physical interrupt input for the interrupt
674 * in question. Otherwise we have a nasty problem with interrupt
675 * mask management. This is best handled in non-performance-critical
676 * platform IRQ affinity setting code, to minimize interrupt-time
680 /* If no one is eligible, service locally */
681 if (target >= NR_CPUS) {
682 do_IRQ_no_affinity(irq);
686 smtc_send_ipi(target, IRQ_AFFINITY_IPI, irq);
689 #endif /* CONFIG_MIPS_MT_SMTC_IRQAFF */
692 * IPI model for SMTC is tricky, because interrupts aren't TC-specific.
693 * Within a VPE one TC can interrupt another by different approaches.
694 * The easiest to get right would probably be to make all TCs except
695 * the target IXMT and set a software interrupt, but an IXMT-based
696 * scheme requires that a handler must run before a new IPI could
697 * be sent, which would break the "broadcast" loops in MIPS MT.
698 * A more gonzo approach within a VPE is to halt the TC, extract
699 * its Restart, Status, and a couple of GPRs, and program the Restart
700 * address to emulate an interrupt.
702 * Within a VPE, one can be confident that the target TC isn't in
703 * a critical EXL state when halted, since the write to the Halt
704 * register could not have issued on the writing thread if the
705 * halting thread had EXL set. So k0 and k1 of the target TC
706 * can be used by the injection code. Across VPEs, one can't
707 * be certain that the target TC isn't in a critical exception
708 * state. So we try a two-step process of sending a software
709 * interrupt to the target VPE, which either handles the event
710 * itself (if it was the target) or injects the event within
714 static void smtc_ipi_qdump(void)
718 for (i = 0; i < NR_CPUS ;i++) {
719 printk("IPIQ[%d]: head = 0x%x, tail = 0x%x, depth = %d\n",
720 i, (unsigned)IPIQ[i].head, (unsigned)IPIQ[i].tail,
726 * The standard atomic.h primitives don't quite do what we want
727 * here: We need an atomic add-and-return-previous-value (which
728 * could be done with atomic_add_return and a decrement) and an
729 * atomic set/zero-and-return-previous-value (which can't really
730 * be done with the atomic.h primitives). And since this is
731 * MIPS MT, we can assume that we have LL/SC.
733 static inline int atomic_postincrement(atomic_t *v)
735 unsigned long result;
739 __asm__ __volatile__(
745 : "=&r" (result), "=&r" (temp), "=m" (v->counter)
752 void smtc_send_ipi(int cpu, int type, unsigned int action)
755 struct smtc_ipi *pipi;
759 if (cpu == smp_processor_id()) {
760 printk("Cannot Send IPI to self!\n");
763 /* Set up a descriptor, to be delivered either promptly or queued */
764 pipi = smtc_ipi_dq(&freeIPIq);
767 mips_mt_regdump(dvpe());
768 panic("IPI Msg. Buffers Depleted\n");
771 pipi->arg = (void *)action;
773 if (cpu_data[cpu].vpe_id != cpu_data[smp_processor_id()].vpe_id) {
774 if (type == SMTC_CLOCK_TICK)
775 atomic_inc(&ipi_timer_latch[cpu]);
776 /* If not on same VPE, enqueue and send cross-VPE interrupt */
777 smtc_ipi_nq(&IPIQ[cpu], pipi);
779 settc(cpu_data[cpu].tc_id);
780 write_vpe_c0_cause(read_vpe_c0_cause() | C_SW1);
784 * Not sufficient to do a LOCK_MT_PRA (dmt) here,
785 * since ASID shootdown on the other VPE may
786 * collide with this operation.
789 settc(cpu_data[cpu].tc_id);
790 /* Halt the targeted TC */
791 write_tc_c0_tchalt(TCHALT_H);
795 * Inspect TCStatus - if IXMT is set, we have to queue
796 * a message. Otherwise, we set up the "interrupt"
799 tcstatus = read_tc_c0_tcstatus();
801 if ((tcstatus & TCSTATUS_IXMT) != 0) {
803 * Spin-waiting here can deadlock,
804 * so we queue the message for the target TC.
806 write_tc_c0_tchalt(0);
808 /* Try to reduce redundant timer interrupt messages */
809 if (type == SMTC_CLOCK_TICK) {
810 if (atomic_postincrement(&ipi_timer_latch[cpu])!=0){
811 smtc_ipi_nq(&freeIPIq, pipi);
815 smtc_ipi_nq(&IPIQ[cpu], pipi);
817 if (type == SMTC_CLOCK_TICK)
818 atomic_inc(&ipi_timer_latch[cpu]);
819 post_direct_ipi(cpu, pipi);
820 write_tc_c0_tchalt(0);
827 * Send IPI message to Halted TC, TargTC/TargVPE already having been set
829 static void post_direct_ipi(int cpu, struct smtc_ipi *pipi)
831 struct pt_regs *kstack;
832 unsigned long tcstatus;
833 unsigned long tcrestart;
834 extern u32 kernelsp[NR_CPUS];
835 extern void __smtc_ipi_vector(void);
836 //printk("%s: on %d for %d\n", __func__, smp_processor_id(), cpu);
838 /* Extract Status, EPC from halted TC */
839 tcstatus = read_tc_c0_tcstatus();
840 tcrestart = read_tc_c0_tcrestart();
841 /* If TCRestart indicates a WAIT instruction, advance the PC */
842 if ((tcrestart & 0x80000000)
843 && ((*(unsigned int *)tcrestart & 0xfe00003f) == 0x42000020)) {
847 * Save on TC's future kernel stack
849 * CU bit of Status is indicator that TC was
850 * already running on a kernel stack...
852 if (tcstatus & ST0_CU0) {
853 /* Note that this "- 1" is pointer arithmetic */
854 kstack = ((struct pt_regs *)read_tc_gpr_sp()) - 1;
856 kstack = ((struct pt_regs *)kernelsp[cpu]) - 1;
859 kstack->cp0_epc = (long)tcrestart;
861 kstack->cp0_tcstatus = tcstatus;
862 /* Pass token of operation to be performed kernel stack pad area */
863 kstack->pad0[4] = (unsigned long)pipi;
864 /* Pass address of function to be called likewise */
865 kstack->pad0[5] = (unsigned long)&ipi_decode;
866 /* Set interrupt exempt and kernel mode */
867 tcstatus |= TCSTATUS_IXMT;
868 tcstatus &= ~TCSTATUS_TKSU;
869 write_tc_c0_tcstatus(tcstatus);
871 /* Set TC Restart address to be SMTC IPI vector */
872 write_tc_c0_tcrestart(__smtc_ipi_vector);
875 static void ipi_resched_interrupt(void)
877 /* Return from interrupt should be enough to cause scheduler check */
880 static void ipi_call_interrupt(void)
882 /* Invoke generic function invocation code in smp.c */
883 smp_call_function_interrupt();
886 DECLARE_PER_CPU(struct clock_event_device, smtc_dummy_clockevent_device);
888 void ipi_decode(struct smtc_ipi *pipi)
890 unsigned int cpu = smp_processor_id();
891 struct clock_event_device *cd;
892 void *arg_copy = pipi->arg;
893 int type_copy = pipi->type;
896 smtc_ipi_nq(&freeIPIq, pipi);
898 case SMTC_CLOCK_TICK:
900 kstat_this_cpu.irqs[MIPS_CPU_IRQ_BASE + 1]++;
901 cd = &per_cpu(smtc_dummy_clockevent_device, cpu);
902 ticks = atomic_read(&ipi_timer_latch[cpu]);
903 atomic_sub(ticks, &ipi_timer_latch[cpu]);
905 cd->event_handler(cd);
912 switch ((int)arg_copy) {
913 case SMP_RESCHEDULE_YOURSELF:
914 ipi_resched_interrupt();
916 case SMP_CALL_FUNCTION:
917 ipi_call_interrupt();
920 printk("Impossible SMTC IPI Argument 0x%x\n",
925 #ifdef CONFIG_MIPS_MT_SMTC_IRQAFF
926 case IRQ_AFFINITY_IPI:
928 * Accept a "forwarded" interrupt that was initially
929 * taken by a TC who doesn't have affinity for the IRQ.
931 do_IRQ_no_affinity((int)arg_copy);
933 #endif /* CONFIG_MIPS_MT_SMTC_IRQAFF */
935 printk("Impossible SMTC IPI Type 0x%x\n", type_copy);
940 void deferred_smtc_ipi(void)
942 struct smtc_ipi *pipi;
945 int q = smp_processor_id();
948 * Test is not atomic, but much faster than a dequeue,
949 * and the vast majority of invocations will have a null queue.
951 if (IPIQ[q].head != NULL) {
952 while((pipi = smtc_ipi_dq(&IPIQ[q])) != NULL) {
953 /* ipi_decode() should be called with interrupts off */
954 local_irq_save(flags);
956 local_irq_restore(flags);
962 * Cross-VPE interrupts in the SMTC prototype use "software interrupts"
963 * set via cross-VPE MTTR manipulation of the Cause register. It would be
964 * in some regards preferable to have external logic for "doorbell" hardware
968 static int cpu_ipi_irq = MIPS_CPU_IRQ_BASE + MIPS_CPU_IPI_IRQ;
970 static irqreturn_t ipi_interrupt(int irq, void *dev_idm)
972 int my_vpe = cpu_data[smp_processor_id()].vpe_id;
973 int my_tc = cpu_data[smp_processor_id()].tc_id;
975 struct smtc_ipi *pipi;
976 unsigned long tcstatus;
979 unsigned int mtflags;
980 unsigned int vpflags;
983 * So long as cross-VPE interrupts are done via
984 * MFTR/MTTR read-modify-writes of Cause, we need
985 * to stop other VPEs whenever the local VPE does
988 local_irq_save(flags);
990 clear_c0_cause(0x100 << MIPS_CPU_IPI_IRQ);
991 set_c0_status(0x100 << MIPS_CPU_IPI_IRQ);
994 local_irq_restore(flags);
997 * Cross-VPE Interrupt handler: Try to directly deliver IPIs
998 * queued for TCs on this VPE other than the current one.
999 * Return-from-interrupt should cause us to drain the queue
1000 * for the current TC, so we ought not to have to do it explicitly here.
1003 for_each_online_cpu(cpu) {
1004 if (cpu_data[cpu].vpe_id != my_vpe)
1007 pipi = smtc_ipi_dq(&IPIQ[cpu]);
1009 if (cpu_data[cpu].tc_id != my_tc) {
1012 settc(cpu_data[cpu].tc_id);
1013 write_tc_c0_tchalt(TCHALT_H);
1015 tcstatus = read_tc_c0_tcstatus();
1016 if ((tcstatus & TCSTATUS_IXMT) == 0) {
1017 post_direct_ipi(cpu, pipi);
1020 write_tc_c0_tchalt(0);
1023 smtc_ipi_req(&IPIQ[cpu], pipi);
1027 * ipi_decode() should be called
1028 * with interrupts off
1030 local_irq_save(flags);
1032 local_irq_restore(flags);
1040 static void ipi_irq_dispatch(void)
1042 do_IRQ(cpu_ipi_irq);
1045 static struct irqaction irq_ipi = {
1046 .handler = ipi_interrupt,
1047 .flags = IRQF_DISABLED,
1049 .flags = IRQF_PERCPU
1052 static void setup_cross_vpe_interrupts(unsigned int nvpe)
1058 panic("SMTC Kernel requires Vectored Interrupt support");
1060 set_vi_handler(MIPS_CPU_IPI_IRQ, ipi_irq_dispatch);
1062 setup_irq_smtc(cpu_ipi_irq, &irq_ipi, (0x100 << MIPS_CPU_IPI_IRQ));
1064 set_irq_handler(cpu_ipi_irq, handle_percpu_irq);
1068 * SMTC-specific hacks invoked from elsewhere in the kernel.
1070 * smtc_ipi_replay is called from raw_local_irq_restore which is only ever
1071 * called with interrupts disabled. We do rely on interrupts being disabled
1072 * here because using spin_lock_irqsave()/spin_unlock_irqrestore() would
1073 * result in a recursive call to raw_local_irq_restore().
1076 static void __smtc_ipi_replay(void)
1078 unsigned int cpu = smp_processor_id();
1081 * To the extent that we've ever turned interrupts off,
1082 * we may have accumulated deferred IPIs. This is subtle.
1083 * If we use the smtc_ipi_qdepth() macro, we'll get an
1084 * exact number - but we'll also disable interrupts
1085 * and create a window of failure where a new IPI gets
1086 * queued after we test the depth but before we re-enable
1087 * interrupts. So long as IXMT never gets set, however,
1088 * we should be OK: If we pick up something and dispatch
1089 * it here, that's great. If we see nothing, but concurrent
1090 * with this operation, another TC sends us an IPI, IXMT
1091 * is clear, and we'll handle it as a real pseudo-interrupt
1092 * and not a pseudo-pseudo interrupt.
1094 if (IPIQ[cpu].depth > 0) {
1096 struct smtc_ipi_q *q = &IPIQ[cpu];
1097 struct smtc_ipi *pipi;
1098 extern void self_ipi(struct smtc_ipi *);
1100 spin_lock(&q->lock);
1101 pipi = __smtc_ipi_dq(q);
1102 spin_unlock(&q->lock);
1107 smtc_cpu_stats[cpu].selfipis++;
1112 void smtc_ipi_replay(void)
1114 raw_local_irq_disable();
1115 __smtc_ipi_replay();
1118 EXPORT_SYMBOL(smtc_ipi_replay);
1120 void smtc_idle_loop_hook(void)
1122 #ifdef CONFIG_SMTC_IDLE_HOOK_DEBUG
1131 * printk within DMT-protected regions can deadlock,
1132 * so buffer diagnostic messages for later output.
1135 char id_ho_db_msg[768]; /* worst-case use should be less than 700 */
1137 if (atomic_read(&idle_hook_initialized) == 0) { /* fast test */
1138 if (atomic_add_return(1, &idle_hook_initialized) == 1) {
1140 /* Tedious stuff to just do once */
1141 mvpconf0 = read_c0_mvpconf0();
1142 hook_ntcs = ((mvpconf0 & MVPCONF0_PTC) >> MVPCONF0_PTC_SHIFT) + 1;
1143 if (hook_ntcs > NR_CPUS)
1144 hook_ntcs = NR_CPUS;
1145 for (tc = 0; tc < hook_ntcs; tc++) {
1147 clock_hang_reported[tc] = 0;
1149 for (vpe = 0; vpe < 2; vpe++)
1150 for (im = 0; im < 8; im++)
1151 imstuckcount[vpe][im] = 0;
1152 printk("Idle loop test hook initialized for %d TCs\n", hook_ntcs);
1153 atomic_set(&idle_hook_initialized, 1000);
1155 /* Someone else is initializing in parallel - let 'em finish */
1156 while (atomic_read(&idle_hook_initialized) < 1000)
1161 /* Have we stupidly left IXMT set somewhere? */
1162 if (read_c0_tcstatus() & 0x400) {
1163 write_c0_tcstatus(read_c0_tcstatus() & ~0x400);
1165 printk("Dangling IXMT in cpu_idle()\n");
1168 /* Have we stupidly left an IM bit turned off? */
1169 #define IM_LIMIT 2000
1170 local_irq_save(flags);
1172 pdb_msg = &id_ho_db_msg[0];
1173 im = read_c0_status();
1174 vpe = current_cpu_data.vpe_id;
1175 for (bit = 0; bit < 8; bit++) {
1177 * In current prototype, I/O interrupts
1178 * are masked for VPE > 0
1180 if (vpemask[vpe][bit]) {
1181 if (!(im & (0x100 << bit)))
1182 imstuckcount[vpe][bit]++;
1184 imstuckcount[vpe][bit] = 0;
1185 if (imstuckcount[vpe][bit] > IM_LIMIT) {
1186 set_c0_status(0x100 << bit);
1188 imstuckcount[vpe][bit] = 0;
1189 pdb_msg += sprintf(pdb_msg,
1190 "Dangling IM %d fixed for VPE %d\n", bit,
1197 * Now that we limit outstanding timer IPIs, check for hung TC
1199 for (tc = 0; tc < NR_CPUS; tc++) {
1200 /* Don't check ourself - we'll dequeue IPIs just below */
1201 if ((tc != smp_processor_id()) &&
1202 atomic_read(&ipi_timer_latch[tc]) > timerq_limit) {
1203 if (clock_hang_reported[tc] == 0) {
1204 pdb_msg += sprintf(pdb_msg,
1205 "TC %d looks hung with timer latch at %d\n",
1206 tc, atomic_read(&ipi_timer_latch[tc]));
1207 clock_hang_reported[tc]++;
1212 local_irq_restore(flags);
1213 if (pdb_msg != &id_ho_db_msg[0])
1214 printk("CPU%d: %s", smp_processor_id(), id_ho_db_msg);
1215 #endif /* CONFIG_SMTC_IDLE_HOOK_DEBUG */
1218 * Replay any accumulated deferred IPIs. If "Instant Replay"
1219 * is in use, there should never be any.
1221 #ifndef CONFIG_MIPS_MT_SMTC_INSTANT_REPLAY
1223 unsigned long flags;
1225 local_irq_save(flags);
1226 __smtc_ipi_replay();
1227 local_irq_restore(flags);
1229 #endif /* CONFIG_MIPS_MT_SMTC_INSTANT_REPLAY */
1232 void smtc_soft_dump(void)
1236 printk("Counter Interrupts taken per CPU (TC)\n");
1237 for (i=0; i < NR_CPUS; i++) {
1238 printk("%d: %ld\n", i, smtc_cpu_stats[i].timerints);
1240 printk("Self-IPI invocations:\n");
1241 for (i=0; i < NR_CPUS; i++) {
1242 printk("%d: %ld\n", i, smtc_cpu_stats[i].selfipis);
1245 printk("Timer IPI Backlogs:\n");
1246 for (i=0; i < NR_CPUS; i++) {
1247 printk("%d: %d\n", i, atomic_read(&ipi_timer_latch[i]));
1249 printk("%d Recoveries of \"stolen\" FPU\n",
1250 atomic_read(&smtc_fpu_recoveries));
1255 * TLB management routines special to SMTC
1258 void smtc_get_new_mmu_context(struct mm_struct *mm, unsigned long cpu)
1260 unsigned long flags, mtflags, tcstat, prevhalt, asid;
1264 * It would be nice to be able to use a spinlock here,
1265 * but this is invoked from within TLB flush routines
1266 * that protect themselves with DVPE, so if a lock is
1267 * held by another TC, it'll never be freed.
1269 * DVPE/DMT must not be done with interrupts enabled,
1270 * so even so most callers will already have disabled
1271 * them, let's be really careful...
1274 local_irq_save(flags);
1275 if (smtc_status & SMTC_TLB_SHARED) {
1280 tlb = cpu_data[cpu].vpe_id;
1282 asid = asid_cache(cpu);
1285 if (!((asid += ASID_INC) & ASID_MASK) ) {
1286 if (cpu_has_vtag_icache)
1288 /* Traverse all online CPUs (hack requires contigous range) */
1289 for_each_online_cpu(i) {
1291 * We don't need to worry about our own CPU, nor those of
1292 * CPUs who don't share our TLB.
1294 if ((i != smp_processor_id()) &&
1295 ((smtc_status & SMTC_TLB_SHARED) ||
1296 (cpu_data[i].vpe_id == cpu_data[cpu].vpe_id))) {
1297 settc(cpu_data[i].tc_id);
1298 prevhalt = read_tc_c0_tchalt() & TCHALT_H;
1300 write_tc_c0_tchalt(TCHALT_H);
1303 tcstat = read_tc_c0_tcstatus();
1304 smtc_live_asid[tlb][(tcstat & ASID_MASK)] |= (asiduse)(0x1 << i);
1306 write_tc_c0_tchalt(0);
1309 if (!asid) /* fix version if needed */
1310 asid = ASID_FIRST_VERSION;
1311 local_flush_tlb_all(); /* start new asid cycle */
1313 } while (smtc_live_asid[tlb][(asid & ASID_MASK)]);
1316 * SMTC shares the TLB within VPEs and possibly across all VPEs.
1318 for_each_online_cpu(i) {
1319 if ((smtc_status & SMTC_TLB_SHARED) ||
1320 (cpu_data[i].vpe_id == cpu_data[cpu].vpe_id))
1321 cpu_context(i, mm) = asid_cache(i) = asid;
1324 if (smtc_status & SMTC_TLB_SHARED)
1328 local_irq_restore(flags);
1332 * Invoked from macros defined in mmu_context.h
1333 * which must already have disabled interrupts
1334 * and done a DVPE or DMT as appropriate.
1337 void smtc_flush_tlb_asid(unsigned long asid)
1342 entry = read_c0_wired();
1344 /* Traverse all non-wired entries */
1345 while (entry < current_cpu_data.tlbsize) {
1346 write_c0_index(entry);
1350 ehi = read_c0_entryhi();
1351 if ((ehi & ASID_MASK) == asid) {
1353 * Invalidate only entries with specified ASID,
1354 * makiing sure all entries differ.
1356 write_c0_entryhi(CKSEG0 + (entry << (PAGE_SHIFT + 1)));
1357 write_c0_entrylo0(0);
1358 write_c0_entrylo1(0);
1360 tlb_write_indexed();
1364 write_c0_index(PARKED_INDEX);
1369 * Support for single-threading cache flush operations.
1372 static int halt_state_save[NR_CPUS];
1375 * To really, really be sure that nothing is being done
1376 * by other TCs, halt them all. This code assumes that
1377 * a DVPE has already been done, so while their Halted
1378 * state is theoretically architecturally unstable, in
1379 * practice, it's not going to change while we're looking
1383 void smtc_cflush_lockdown(void)
1387 for_each_online_cpu(cpu) {
1388 if (cpu != smp_processor_id()) {
1389 settc(cpu_data[cpu].tc_id);
1390 halt_state_save[cpu] = read_tc_c0_tchalt();
1391 write_tc_c0_tchalt(TCHALT_H);
1397 /* It would be cheating to change the cpu_online states during a flush! */
1399 void smtc_cflush_release(void)
1404 * Start with a hazard barrier to ensure
1405 * that all CACHE ops have played through.
1409 for_each_online_cpu(cpu) {
1410 if (cpu != smp_processor_id()) {
1411 settc(cpu_data[cpu].tc_id);
1412 write_tc_c0_tchalt(halt_state_save[cpu]);