1 /* Copyright (C) 2004 Mips Technologies, Inc */
3 #include <linux/kernel.h>
4 #include <linux/sched.h>
5 #include <linux/cpumask.h>
6 #include <linux/interrupt.h>
7 #include <linux/kernel_stat.h>
8 #include <linux/module.h>
11 #include <asm/processor.h>
12 #include <asm/atomic.h>
13 #include <asm/system.h>
14 #include <asm/hardirq.h>
15 #include <asm/hazards.h>
17 #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 * This file should be built into the kernel only if CONFIG_MIPS_MT_SMTC is set.
31 #define MIPS_CPU_IPI_IRQ 1
33 #define LOCK_MT_PRA() \
34 local_irq_save(flags); \
37 #define UNLOCK_MT_PRA() \
39 local_irq_restore(flags)
41 #define LOCK_CORE_PRA() \
42 local_irq_save(flags); \
45 #define UNLOCK_CORE_PRA() \
47 local_irq_restore(flags)
50 * Data structures purely associated with SMTC parallelism
55 * Table for tracking ASIDs whose lifetime is prolonged.
58 asiduse smtc_live_asid[MAX_SMTC_TLBS][MAX_SMTC_ASIDS];
61 * Clock interrupt "latch" buffers, per "CPU"
64 unsigned int ipi_timer_latch[NR_CPUS];
67 * Number of InterProcessor Interupt (IPI) message buffers to allocate
70 #define IPIBUF_PER_CPU 4
72 static struct smtc_ipi_q IPIQ[NR_CPUS];
73 static struct smtc_ipi_q freeIPIq;
76 /* Forward declarations */
78 void ipi_decode(struct smtc_ipi *);
79 static void post_direct_ipi(int cpu, struct smtc_ipi *pipi);
80 static void setup_cross_vpe_interrupts(unsigned int nvpe);
81 void init_smtc_stats(void);
83 /* Global SMTC Status */
85 unsigned int smtc_status = 0;
87 /* Boot command line configuration overrides */
89 static int vpelimit = 0;
90 static int tclimit = 0;
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 maxvpes(char *str)
98 get_option(&str, &vpelimit);
102 static int __init maxtcs(char *str)
104 get_option(&str, &tclimit);
108 static int __init ipibufs(char *str)
110 get_option(&str, &ipibuffers);
114 static int __init stlb_disable(char *s)
120 static int __init asidmask_set(char *str)
122 get_option(&str, &asidmask);
132 smtc_asid_mask = (unsigned long)asidmask;
135 printk("ILLEGAL ASID mask 0x%x from command line\n", asidmask);
140 __setup("maxvpes=", maxvpes);
141 __setup("maxtcs=", maxtcs);
142 __setup("ipibufs=", ipibufs);
143 __setup("nostlb", stlb_disable);
144 __setup("asidmask=", asidmask_set);
146 #ifdef CONFIG_SMTC_IDLE_HOOK_DEBUG
148 static int hang_trig = 0;
150 static int __init hangtrig_enable(char *s)
157 __setup("hangtrig", hangtrig_enable);
159 #define DEFAULT_BLOCKED_IPI_LIMIT 32
161 static int timerq_limit = DEFAULT_BLOCKED_IPI_LIMIT;
163 static int __init tintq(char *str)
165 get_option(&str, &timerq_limit);
169 __setup("tintq=", tintq);
171 int imstuckcount[2][8];
172 /* vpemask represents IM/IE bits of per-VPE Status registers, low-to-high */
173 int vpemask[2][8] = {
174 {0, 0, 1, 0, 0, 0, 0, 1},
175 {0, 0, 0, 0, 0, 0, 0, 1}
177 int tcnoprog[NR_CPUS];
178 static atomic_t idle_hook_initialized = {0};
179 static int clock_hang_reported[NR_CPUS];
181 #endif /* CONFIG_SMTC_IDLE_HOOK_DEBUG */
183 /* Initialize shared TLB - the should probably migrate to smtc_setup_cpus() */
185 void __init sanitize_tlb_entries(void)
187 printk("Deprecated sanitize_tlb_entries() invoked\n");
192 * Configure shared TLB - VPC configuration bit must be set by caller
195 static void smtc_configure_tlb(void)
198 unsigned long mvpconf0;
199 unsigned long config1val;
201 /* Set up ASID preservation table */
202 for (vpes=0; vpes<MAX_SMTC_TLBS; vpes++) {
203 for(i = 0; i < MAX_SMTC_ASIDS; i++) {
204 smtc_live_asid[vpes][i] = 0;
207 mvpconf0 = read_c0_mvpconf0();
209 if ((vpes = ((mvpconf0 & MVPCONF0_PVPE)
210 >> MVPCONF0_PVPE_SHIFT) + 1) > 1) {
211 /* If we have multiple VPEs, try to share the TLB */
212 if ((mvpconf0 & MVPCONF0_TLBS) && !nostlb) {
214 * If TLB sizing is programmable, shared TLB
215 * size is the total available complement.
216 * Otherwise, we have to take the sum of all
217 * static VPE TLB entries.
219 if ((tlbsiz = ((mvpconf0 & MVPCONF0_PTLBE)
220 >> MVPCONF0_PTLBE_SHIFT)) == 0) {
222 * If there's more than one VPE, there had better
223 * be more than one TC, because we need one to bind
224 * to each VPE in turn to be able to read
225 * its configuration state!
228 /* Stop the TC from doing anything foolish */
229 write_tc_c0_tchalt(TCHALT_H);
231 /* No need to un-Halt - that happens later anyway */
232 for (i=0; i < vpes; i++) {
233 write_tc_c0_tcbind(i);
235 * To be 100% sure we're really getting the right
236 * information, we exit the configuration state
237 * and do an IHB after each rebinding.
240 read_c0_mvpcontrol() & ~ MVPCONTROL_VPC );
243 * Only count if the MMU Type indicated is TLB
245 if (((read_vpe_c0_config() & MIPS_CONF_MT) >> 7) == 1) {
246 config1val = read_vpe_c0_config1();
247 tlbsiz += ((config1val >> 25) & 0x3f) + 1;
250 /* Put core back in configuration state */
252 read_c0_mvpcontrol() | MVPCONTROL_VPC );
256 write_c0_mvpcontrol(read_c0_mvpcontrol() | MVPCONTROL_STLB);
260 * Setup kernel data structures to use software total,
261 * rather than read the per-VPE Config1 value. The values
262 * for "CPU 0" gets copied to all the other CPUs as part
263 * of their initialization in smtc_cpu_setup().
266 /* MIPS32 limits TLB indices to 64 */
269 cpu_data[0].tlbsize = current_cpu_data.tlbsize = tlbsiz;
270 smtc_status |= SMTC_TLB_SHARED;
271 local_flush_tlb_all();
273 printk("TLB of %d entry pairs shared by %d VPEs\n",
276 printk("WARNING: TLB Not Sharable on SMTC Boot!\n");
283 * Incrementally build the CPU map out of constituent MIPS MT cores,
284 * using the specified available VPEs and TCs. Plaform code needs
285 * to ensure that each MIPS MT core invokes this routine on reset,
288 * This version of the build_cpu_map and prepare_cpus routines assumes
289 * that *all* TCs of a MIPS MT core will be used for Linux, and that
290 * they will be spread across *all* available VPEs (to minimise the
291 * loss of efficiency due to exception service serialization).
292 * An improved version would pick up configuration information and
293 * possibly leave some TCs/VPEs as "slave" processors.
295 * Use c0_MVPConf0 to find out how many TCs are available, setting up
296 * phys_cpu_present_map and the logical/physical mappings.
299 int __init mipsmt_build_cpu_map(int start_cpu_slot)
304 * The CPU map isn't actually used for anything at this point,
305 * so it's not clear what else we should do apart from set
306 * everything up so that "logical" = "physical".
308 ntcs = ((read_c0_mvpconf0() & MVPCONF0_PTC) >> MVPCONF0_PTC_SHIFT) + 1;
309 for (i=start_cpu_slot; i<NR_CPUS && i<ntcs; i++) {
310 cpu_set(i, phys_cpu_present_map);
311 __cpu_number_map[i] = i;
312 __cpu_logical_map[i] = i;
314 /* Initialize map of CPUs with FPUs */
315 cpus_clear(mt_fpu_cpumask);
317 /* One of those TC's is the one booting, and not a secondary... */
318 printk("%i available secondary CPU TC(s)\n", i - 1);
324 * Common setup before any secondaries are started
325 * Make sure all CPU's are in a sensible state before we boot any of the
328 * For MIPS MT "SMTC" operation, we set up all TCs, spread as evenly
329 * as possible across the available VPEs.
332 static void smtc_tc_setup(int vpe, int tc, int cpu)
335 write_tc_c0_tchalt(TCHALT_H);
337 write_tc_c0_tcstatus((read_tc_c0_tcstatus()
338 & ~(TCSTATUS_TKSU | TCSTATUS_DA | TCSTATUS_IXMT))
340 write_tc_c0_tccontext(0);
342 write_tc_c0_tcbind(vpe);
343 /* In general, all TCs should have the same cpu_data indications */
344 memcpy(&cpu_data[cpu], &cpu_data[0], sizeof(struct cpuinfo_mips));
345 /* For 34Kf, start with TC/CPU 0 as sole owner of single FPU context */
346 if (cpu_data[0].cputype == CPU_34K)
347 cpu_data[cpu].options &= ~MIPS_CPU_FPU;
348 cpu_data[cpu].vpe_id = vpe;
349 cpu_data[cpu].tc_id = tc;
353 void mipsmt_prepare_cpus(void)
355 int i, vpe, tc, ntc, nvpe, tcpervpe, slop, cpu;
359 struct smtc_ipi *pipi;
361 /* disable interrupts so we can disable MT */
362 local_irq_save(flags);
363 /* disable MT so we can configure */
367 spin_lock_init(&freeIPIq.lock);
370 * We probably don't have as many VPEs as we do SMP "CPUs",
371 * but it's possible - and in any case we'll never use more!
373 for (i=0; i<NR_CPUS; i++) {
374 IPIQ[i].head = IPIQ[i].tail = NULL;
375 spin_lock_init(&IPIQ[i].lock);
377 ipi_timer_latch[i] = 0;
380 /* cpu_data index starts at zero */
382 cpu_data[cpu].vpe_id = 0;
383 cpu_data[cpu].tc_id = 0;
386 /* Report on boot-time options */
387 mips_mt_set_cpuoptions ();
389 printk("Limit of %d VPEs set\n", vpelimit);
391 printk("Limit of %d TCs set\n", tclimit);
393 printk("Shared TLB Use Inhibited - UNSAFE for Multi-VPE Operation\n");
396 printk("ASID mask value override to 0x%x\n", asidmask);
399 #ifdef CONFIG_SMTC_IDLE_HOOK_DEBUG
401 printk("Logic Analyser Trigger on suspected TC hang\n");
402 #endif /* CONFIG_SMTC_IDLE_HOOK_DEBUG */
404 /* Put MVPE's into 'configuration state' */
405 write_c0_mvpcontrol( read_c0_mvpcontrol() | MVPCONTROL_VPC );
407 val = read_c0_mvpconf0();
408 nvpe = ((val & MVPCONF0_PVPE) >> MVPCONF0_PVPE_SHIFT) + 1;
409 if (vpelimit > 0 && nvpe > vpelimit)
411 ntc = ((val & MVPCONF0_PTC) >> MVPCONF0_PTC_SHIFT) + 1;
414 if (tclimit > 0 && ntc > tclimit)
416 tcpervpe = ntc / nvpe;
417 slop = ntc % nvpe; /* Residual TCs, < NVPE */
419 /* Set up shared TLB */
420 smtc_configure_tlb();
422 for (tc = 0, vpe = 0 ; (vpe < nvpe) && (tc < ntc) ; vpe++) {
427 write_vpe_c0_vpeconf0(read_vpe_c0_vpeconf0() | VPECONF0_MVP);
430 printk("VPE %d: TC", vpe);
431 for (i = 0; i < tcpervpe; i++) {
433 * TC 0 is bound to VPE 0 at reset,
434 * and is presumably executing this
435 * code. Leave it alone!
438 smtc_tc_setup(vpe,tc, cpu);
446 smtc_tc_setup(vpe,tc, cpu);
455 * Clear any stale software interrupts from VPE's Cause
457 write_vpe_c0_cause(0);
460 * Clear ERL/EXL of VPEs other than 0
461 * and set restricted interrupt enable/mask.
463 write_vpe_c0_status((read_vpe_c0_status()
464 & ~(ST0_BEV | ST0_ERL | ST0_EXL | ST0_IM))
465 | (STATUSF_IP0 | STATUSF_IP1 | STATUSF_IP7
468 * set config to be the same as vpe0,
469 * particularly kseg0 coherency alg
471 write_vpe_c0_config(read_c0_config());
472 /* Clear any pending timer interrupt */
473 write_vpe_c0_compare(0);
474 /* Propagate Config7 */
475 write_vpe_c0_config7(read_c0_config7());
476 write_vpe_c0_count(read_c0_count());
478 /* enable multi-threading within VPE */
479 write_vpe_c0_vpecontrol(read_vpe_c0_vpecontrol() | VPECONTROL_TE);
481 write_vpe_c0_vpeconf0(read_vpe_c0_vpeconf0() | VPECONF0_VPA);
485 * Pull any physically present but unused TCs out of circulation.
487 while (tc < (((val & MVPCONF0_PTC) >> MVPCONF0_PTC_SHIFT) + 1)) {
488 cpu_clear(tc, phys_cpu_present_map);
489 cpu_clear(tc, cpu_present_map);
493 /* release config state */
494 write_c0_mvpcontrol( read_c0_mvpcontrol() & ~ MVPCONTROL_VPC );
498 /* Set up coprocessor affinity CPU mask(s) */
500 for (tc = 0; tc < ntc; tc++) {
501 if (cpu_data[tc].options & MIPS_CPU_FPU)
502 cpu_set(tc, mt_fpu_cpumask);
505 /* set up ipi interrupts... */
507 /* If we have multiple VPEs running, set up the cross-VPE interrupt */
509 setup_cross_vpe_interrupts(nvpe);
511 /* Set up queue of free IPI "messages". */
512 nipi = NR_CPUS * IPIBUF_PER_CPU;
516 pipi = kmalloc(nipi *sizeof(struct smtc_ipi), GFP_KERNEL);
518 panic("kmalloc of IPI message buffers failed\n");
520 printk("IPI buffer pool of %d buffers\n", nipi);
521 for (i = 0; i < nipi; i++) {
522 smtc_ipi_nq(&freeIPIq, pipi);
526 /* Arm multithreading and enable other VPEs - but all TCs are Halted */
529 local_irq_restore(flags);
530 /* Initialize SMTC /proc statistics/diagnostics */
536 * Setup the PC, SP, and GP of a secondary processor and start it
538 * smp_bootstrap is the place to resume from
539 * __KSTK_TOS(idle) is apparently the stack pointer
540 * (unsigned long)idle->thread_info the gp
543 void smtc_boot_secondary(int cpu, struct task_struct *idle)
545 extern u32 kernelsp[NR_CPUS];
550 if (cpu_data[cpu].vpe_id != cpu_data[smp_processor_id()].vpe_id) {
553 settc(cpu_data[cpu].tc_id);
556 write_tc_c0_tcrestart((unsigned long)&smp_bootstrap);
559 kernelsp[cpu] = __KSTK_TOS(idle);
560 write_tc_gpr_sp(__KSTK_TOS(idle));
563 write_tc_gpr_gp((unsigned long)task_thread_info(idle));
565 smtc_status |= SMTC_MTC_ACTIVE;
566 write_tc_c0_tchalt(0);
567 if (cpu_data[cpu].vpe_id != cpu_data[smp_processor_id()].vpe_id) {
573 void smtc_init_secondary(void)
576 * Start timer on secondary VPEs if necessary.
577 * plat_timer_setup has already have been invoked by init/main
578 * on "boot" TC. Like per_cpu_trap_init() hack, this assumes that
579 * SMTC init code assigns TCs consdecutively and in ascending order
580 * to across available VPEs.
582 if (((read_c0_tcbind() & TCBIND_CURTC) != 0) &&
583 ((read_c0_tcbind() & TCBIND_CURVPE)
584 != cpu_data[smp_processor_id() - 1].vpe_id)){
585 write_c0_compare (read_c0_count() + mips_hpt_frequency/HZ);
591 void smtc_smp_finish(void)
593 printk("TC %d going on-line as CPU %d\n",
594 cpu_data[smp_processor_id()].tc_id, smp_processor_id());
597 void smtc_cpus_done(void)
602 * Support for SMTC-optimized driver IRQ registration
606 * SMTC Kernel needs to manipulate low-level CPU interrupt mask
607 * in do_IRQ. These are passed in setup_irq_smtc() and stored
611 int setup_irq_smtc(unsigned int irq, struct irqaction * new,
612 unsigned long hwmask)
614 #ifdef CONFIG_SMTC_IDLE_HOOK_DEBUG
615 unsigned int vpe = current_cpu_data.vpe_id;
617 vpemask[vpe][irq - MIPS_CPU_IRQ_BASE] = 1;
619 irq_hwmask[irq] = hwmask;
621 return setup_irq(irq, new);
625 * IPI model for SMTC is tricky, because interrupts aren't TC-specific.
626 * Within a VPE one TC can interrupt another by different approaches.
627 * The easiest to get right would probably be to make all TCs except
628 * the target IXMT and set a software interrupt, but an IXMT-based
629 * scheme requires that a handler must run before a new IPI could
630 * be sent, which would break the "broadcast" loops in MIPS MT.
631 * A more gonzo approach within a VPE is to halt the TC, extract
632 * its Restart, Status, and a couple of GPRs, and program the Restart
633 * address to emulate an interrupt.
635 * Within a VPE, one can be confident that the target TC isn't in
636 * a critical EXL state when halted, since the write to the Halt
637 * register could not have issued on the writing thread if the
638 * halting thread had EXL set. So k0 and k1 of the target TC
639 * can be used by the injection code. Across VPEs, one can't
640 * be certain that the target TC isn't in a critical exception
641 * state. So we try a two-step process of sending a software
642 * interrupt to the target VPE, which either handles the event
643 * itself (if it was the target) or injects the event within
647 static void smtc_ipi_qdump(void)
651 for (i = 0; i < NR_CPUS ;i++) {
652 printk("IPIQ[%d]: head = 0x%x, tail = 0x%x, depth = %d\n",
653 i, (unsigned)IPIQ[i].head, (unsigned)IPIQ[i].tail,
659 * The standard atomic.h primitives don't quite do what we want
660 * here: We need an atomic add-and-return-previous-value (which
661 * could be done with atomic_add_return and a decrement) and an
662 * atomic set/zero-and-return-previous-value (which can't really
663 * be done with the atomic.h primitives). And since this is
664 * MIPS MT, we can assume that we have LL/SC.
666 static __inline__ int atomic_postincrement(unsigned int *pv)
668 unsigned long result;
672 __asm__ __volatile__(
678 : "=&r" (result), "=&r" (temp), "=m" (*pv)
685 void smtc_send_ipi(int cpu, int type, unsigned int action)
688 struct smtc_ipi *pipi;
692 if (cpu == smp_processor_id()) {
693 printk("Cannot Send IPI to self!\n");
696 /* Set up a descriptor, to be delivered either promptly or queued */
697 pipi = smtc_ipi_dq(&freeIPIq);
700 mips_mt_regdump(dvpe());
701 panic("IPI Msg. Buffers Depleted\n");
704 pipi->arg = (void *)action;
706 if (cpu_data[cpu].vpe_id != cpu_data[smp_processor_id()].vpe_id) {
707 /* If not on same VPE, enqueue and send cross-VPE interupt */
708 smtc_ipi_nq(&IPIQ[cpu], pipi);
710 settc(cpu_data[cpu].tc_id);
711 write_vpe_c0_cause(read_vpe_c0_cause() | C_SW1);
715 * Not sufficient to do a LOCK_MT_PRA (dmt) here,
716 * since ASID shootdown on the other VPE may
717 * collide with this operation.
720 settc(cpu_data[cpu].tc_id);
721 /* Halt the targeted TC */
722 write_tc_c0_tchalt(TCHALT_H);
726 * Inspect TCStatus - if IXMT is set, we have to queue
727 * a message. Otherwise, we set up the "interrupt"
730 tcstatus = read_tc_c0_tcstatus();
732 if ((tcstatus & TCSTATUS_IXMT) != 0) {
734 * Spin-waiting here can deadlock,
735 * so we queue the message for the target TC.
737 write_tc_c0_tchalt(0);
739 /* Try to reduce redundant timer interrupt messages */
740 if (type == SMTC_CLOCK_TICK) {
741 if (atomic_postincrement(&ipi_timer_latch[cpu])!=0){
742 smtc_ipi_nq(&freeIPIq, pipi);
746 smtc_ipi_nq(&IPIQ[cpu], pipi);
748 post_direct_ipi(cpu, pipi);
749 write_tc_c0_tchalt(0);
756 * Send IPI message to Halted TC, TargTC/TargVPE already having been set
758 static void post_direct_ipi(int cpu, struct smtc_ipi *pipi)
760 struct pt_regs *kstack;
761 unsigned long tcstatus;
762 unsigned long tcrestart;
763 extern u32 kernelsp[NR_CPUS];
764 extern void __smtc_ipi_vector(void);
766 /* Extract Status, EPC from halted TC */
767 tcstatus = read_tc_c0_tcstatus();
768 tcrestart = read_tc_c0_tcrestart();
769 /* If TCRestart indicates a WAIT instruction, advance the PC */
770 if ((tcrestart & 0x80000000)
771 && ((*(unsigned int *)tcrestart & 0xfe00003f) == 0x42000020)) {
775 * Save on TC's future kernel stack
777 * CU bit of Status is indicator that TC was
778 * already running on a kernel stack...
780 if (tcstatus & ST0_CU0) {
781 /* Note that this "- 1" is pointer arithmetic */
782 kstack = ((struct pt_regs *)read_tc_gpr_sp()) - 1;
784 kstack = ((struct pt_regs *)kernelsp[cpu]) - 1;
787 kstack->cp0_epc = (long)tcrestart;
789 kstack->cp0_tcstatus = tcstatus;
790 /* Pass token of operation to be performed kernel stack pad area */
791 kstack->pad0[4] = (unsigned long)pipi;
792 /* Pass address of function to be called likewise */
793 kstack->pad0[5] = (unsigned long)&ipi_decode;
794 /* Set interrupt exempt and kernel mode */
795 tcstatus |= TCSTATUS_IXMT;
796 tcstatus &= ~TCSTATUS_TKSU;
797 write_tc_c0_tcstatus(tcstatus);
799 /* Set TC Restart address to be SMTC IPI vector */
800 write_tc_c0_tcrestart(__smtc_ipi_vector);
803 static void ipi_resched_interrupt(void)
805 /* Return from interrupt should be enough to cause scheduler check */
809 static void ipi_call_interrupt(void)
811 /* Invoke generic function invocation code in smp.c */
812 smp_call_function_interrupt();
815 void ipi_decode(struct smtc_ipi *pipi)
817 void *arg_copy = pipi->arg;
818 int type_copy = pipi->type;
819 int dest_copy = pipi->dest;
821 smtc_ipi_nq(&freeIPIq, pipi);
823 case SMTC_CLOCK_TICK:
825 kstat_this_cpu.irqs[MIPS_CPU_IRQ_BASE + cp0_compare_irq]++;
826 /* Invoke Clock "Interrupt" */
827 ipi_timer_latch[dest_copy] = 0;
828 #ifdef CONFIG_SMTC_IDLE_HOOK_DEBUG
829 clock_hang_reported[dest_copy] = 0;
830 #endif /* CONFIG_SMTC_IDLE_HOOK_DEBUG */
831 local_timer_interrupt(0, NULL);
835 switch ((int)arg_copy) {
836 case SMP_RESCHEDULE_YOURSELF:
837 ipi_resched_interrupt();
839 case SMP_CALL_FUNCTION:
840 ipi_call_interrupt();
843 printk("Impossible SMTC IPI Argument 0x%x\n",
849 printk("Impossible SMTC IPI Type 0x%x\n", type_copy);
854 void deferred_smtc_ipi(void)
856 struct smtc_ipi *pipi;
859 int q = smp_processor_id();
862 * Test is not atomic, but much faster than a dequeue,
863 * and the vast majority of invocations will have a null queue.
865 if (IPIQ[q].head != NULL) {
866 while((pipi = smtc_ipi_dq(&IPIQ[q])) != NULL) {
867 /* ipi_decode() should be called with interrupts off */
868 local_irq_save(flags);
870 local_irq_restore(flags);
876 * Send clock tick to all TCs except the one executing the funtion
879 void smtc_timer_broadcast(int vpe)
882 int myTC = cpu_data[smp_processor_id()].tc_id;
883 int myVPE = cpu_data[smp_processor_id()].vpe_id;
885 smtc_cpu_stats[smp_processor_id()].timerints++;
887 for_each_online_cpu(cpu) {
888 if (cpu_data[cpu].vpe_id == myVPE &&
889 cpu_data[cpu].tc_id != myTC)
890 smtc_send_ipi(cpu, SMTC_CLOCK_TICK, 0);
895 * Cross-VPE interrupts in the SMTC prototype use "software interrupts"
896 * set via cross-VPE MTTR manipulation of the Cause register. It would be
897 * in some regards preferable to have external logic for "doorbell" hardware
901 static int cpu_ipi_irq = MIPS_CPU_IRQ_BASE + MIPS_CPU_IPI_IRQ;
903 static irqreturn_t ipi_interrupt(int irq, void *dev_idm)
905 int my_vpe = cpu_data[smp_processor_id()].vpe_id;
906 int my_tc = cpu_data[smp_processor_id()].tc_id;
908 struct smtc_ipi *pipi;
909 unsigned long tcstatus;
912 unsigned int mtflags;
913 unsigned int vpflags;
916 * So long as cross-VPE interrupts are done via
917 * MFTR/MTTR read-modify-writes of Cause, we need
918 * to stop other VPEs whenever the local VPE does
921 local_irq_save(flags);
923 clear_c0_cause(0x100 << MIPS_CPU_IPI_IRQ);
924 set_c0_status(0x100 << MIPS_CPU_IPI_IRQ);
927 local_irq_restore(flags);
930 * Cross-VPE Interrupt handler: Try to directly deliver IPIs
931 * queued for TCs on this VPE other than the current one.
932 * Return-from-interrupt should cause us to drain the queue
933 * for the current TC, so we ought not to have to do it explicitly here.
936 for_each_online_cpu(cpu) {
937 if (cpu_data[cpu].vpe_id != my_vpe)
940 pipi = smtc_ipi_dq(&IPIQ[cpu]);
942 if (cpu_data[cpu].tc_id != my_tc) {
945 settc(cpu_data[cpu].tc_id);
946 write_tc_c0_tchalt(TCHALT_H);
948 tcstatus = read_tc_c0_tcstatus();
949 if ((tcstatus & TCSTATUS_IXMT) == 0) {
950 post_direct_ipi(cpu, pipi);
953 write_tc_c0_tchalt(0);
956 smtc_ipi_req(&IPIQ[cpu], pipi);
960 * ipi_decode() should be called
961 * with interrupts off
963 local_irq_save(flags);
965 local_irq_restore(flags);
973 static void ipi_irq_dispatch(void)
978 static struct irqaction irq_ipi;
980 static void setup_cross_vpe_interrupts(unsigned int nvpe)
986 panic("SMTC Kernel requires Vectored Interupt support");
988 set_vi_handler(MIPS_CPU_IPI_IRQ, ipi_irq_dispatch);
990 irq_ipi.handler = ipi_interrupt;
991 irq_ipi.flags = IRQF_DISABLED;
992 irq_ipi.name = "SMTC_IPI";
994 setup_irq_smtc(cpu_ipi_irq, &irq_ipi, (0x100 << MIPS_CPU_IPI_IRQ));
996 irq_desc[cpu_ipi_irq].status |= IRQ_PER_CPU;
997 set_irq_handler(cpu_ipi_irq, handle_percpu_irq);
1001 * SMTC-specific hacks invoked from elsewhere in the kernel.
1003 * smtc_ipi_replay is called from raw_local_irq_restore which is only ever
1004 * called with interrupts disabled. We do rely on interrupts being disabled
1005 * here because using spin_lock_irqsave()/spin_unlock_irqrestore() would
1006 * result in a recursive call to raw_local_irq_restore().
1009 static void __smtc_ipi_replay(void)
1011 unsigned int cpu = smp_processor_id();
1014 * To the extent that we've ever turned interrupts off,
1015 * we may have accumulated deferred IPIs. This is subtle.
1016 * If we use the smtc_ipi_qdepth() macro, we'll get an
1017 * exact number - but we'll also disable interrupts
1018 * and create a window of failure where a new IPI gets
1019 * queued after we test the depth but before we re-enable
1020 * interrupts. So long as IXMT never gets set, however,
1021 * we should be OK: If we pick up something and dispatch
1022 * it here, that's great. If we see nothing, but concurrent
1023 * with this operation, another TC sends us an IPI, IXMT
1024 * is clear, and we'll handle it as a real pseudo-interrupt
1025 * and not a pseudo-pseudo interrupt.
1027 if (IPIQ[cpu].depth > 0) {
1029 struct smtc_ipi_q *q = &IPIQ[cpu];
1030 struct smtc_ipi *pipi;
1031 extern void self_ipi(struct smtc_ipi *);
1033 spin_lock(&q->lock);
1034 pipi = __smtc_ipi_dq(q);
1035 spin_unlock(&q->lock);
1040 smtc_cpu_stats[cpu].selfipis++;
1045 void smtc_ipi_replay(void)
1047 raw_local_irq_disable();
1048 __smtc_ipi_replay();
1051 EXPORT_SYMBOL(smtc_ipi_replay);
1053 void smtc_idle_loop_hook(void)
1055 #ifdef CONFIG_SMTC_IDLE_HOOK_DEBUG
1064 * printk within DMT-protected regions can deadlock,
1065 * so buffer diagnostic messages for later output.
1068 char id_ho_db_msg[768]; /* worst-case use should be less than 700 */
1070 if (atomic_read(&idle_hook_initialized) == 0) { /* fast test */
1071 if (atomic_add_return(1, &idle_hook_initialized) == 1) {
1073 /* Tedious stuff to just do once */
1074 mvpconf0 = read_c0_mvpconf0();
1075 hook_ntcs = ((mvpconf0 & MVPCONF0_PTC) >> MVPCONF0_PTC_SHIFT) + 1;
1076 if (hook_ntcs > NR_CPUS)
1077 hook_ntcs = NR_CPUS;
1078 for (tc = 0; tc < hook_ntcs; tc++) {
1080 clock_hang_reported[tc] = 0;
1082 for (vpe = 0; vpe < 2; vpe++)
1083 for (im = 0; im < 8; im++)
1084 imstuckcount[vpe][im] = 0;
1085 printk("Idle loop test hook initialized for %d TCs\n", hook_ntcs);
1086 atomic_set(&idle_hook_initialized, 1000);
1088 /* Someone else is initializing in parallel - let 'em finish */
1089 while (atomic_read(&idle_hook_initialized) < 1000)
1094 /* Have we stupidly left IXMT set somewhere? */
1095 if (read_c0_tcstatus() & 0x400) {
1096 write_c0_tcstatus(read_c0_tcstatus() & ~0x400);
1098 printk("Dangling IXMT in cpu_idle()\n");
1101 /* Have we stupidly left an IM bit turned off? */
1102 #define IM_LIMIT 2000
1103 local_irq_save(flags);
1105 pdb_msg = &id_ho_db_msg[0];
1106 im = read_c0_status();
1107 vpe = current_cpu_data.vpe_id;
1108 for (bit = 0; bit < 8; bit++) {
1110 * In current prototype, I/O interrupts
1111 * are masked for VPE > 0
1113 if (vpemask[vpe][bit]) {
1114 if (!(im & (0x100 << bit)))
1115 imstuckcount[vpe][bit]++;
1117 imstuckcount[vpe][bit] = 0;
1118 if (imstuckcount[vpe][bit] > IM_LIMIT) {
1119 set_c0_status(0x100 << bit);
1121 imstuckcount[vpe][bit] = 0;
1122 pdb_msg += sprintf(pdb_msg,
1123 "Dangling IM %d fixed for VPE %d\n", bit,
1130 * Now that we limit outstanding timer IPIs, check for hung TC
1132 for (tc = 0; tc < NR_CPUS; tc++) {
1133 /* Don't check ourself - we'll dequeue IPIs just below */
1134 if ((tc != smp_processor_id()) &&
1135 ipi_timer_latch[tc] > timerq_limit) {
1136 if (clock_hang_reported[tc] == 0) {
1137 pdb_msg += sprintf(pdb_msg,
1138 "TC %d looks hung with timer latch at %d\n",
1139 tc, ipi_timer_latch[tc]);
1140 clock_hang_reported[tc]++;
1145 local_irq_restore(flags);
1146 if (pdb_msg != &id_ho_db_msg[0])
1147 printk("CPU%d: %s", smp_processor_id(), id_ho_db_msg);
1148 #endif /* CONFIG_SMTC_IDLE_HOOK_DEBUG */
1151 * Replay any accumulated deferred IPIs. If "Instant Replay"
1152 * is in use, there should never be any.
1154 #ifndef CONFIG_MIPS_MT_SMTC_INSTANT_REPLAY
1156 unsigned long flags;
1158 local_irq_save(flags);
1159 __smtc_ipi_replay();
1160 local_irq_restore(flags);
1162 #endif /* CONFIG_MIPS_MT_SMTC_INSTANT_REPLAY */
1165 void smtc_soft_dump(void)
1169 printk("Counter Interrupts taken per CPU (TC)\n");
1170 for (i=0; i < NR_CPUS; i++) {
1171 printk("%d: %ld\n", i, smtc_cpu_stats[i].timerints);
1173 printk("Self-IPI invocations:\n");
1174 for (i=0; i < NR_CPUS; i++) {
1175 printk("%d: %ld\n", i, smtc_cpu_stats[i].selfipis);
1178 printk("Timer IPI Backlogs:\n");
1179 for (i=0; i < NR_CPUS; i++) {
1180 printk("%d: %d\n", i, ipi_timer_latch[i]);
1182 printk("%d Recoveries of \"stolen\" FPU\n",
1183 atomic_read(&smtc_fpu_recoveries));
1188 * TLB management routines special to SMTC
1191 void smtc_get_new_mmu_context(struct mm_struct *mm, unsigned long cpu)
1193 unsigned long flags, mtflags, tcstat, prevhalt, asid;
1197 * It would be nice to be able to use a spinlock here,
1198 * but this is invoked from within TLB flush routines
1199 * that protect themselves with DVPE, so if a lock is
1200 * held by another TC, it'll never be freed.
1202 * DVPE/DMT must not be done with interrupts enabled,
1203 * so even so most callers will already have disabled
1204 * them, let's be really careful...
1207 local_irq_save(flags);
1208 if (smtc_status & SMTC_TLB_SHARED) {
1213 tlb = cpu_data[cpu].vpe_id;
1215 asid = asid_cache(cpu);
1218 if (!((asid += ASID_INC) & ASID_MASK) ) {
1219 if (cpu_has_vtag_icache)
1221 /* Traverse all online CPUs (hack requires contigous range) */
1222 for (i = 0; i < num_online_cpus(); i++) {
1224 * We don't need to worry about our own CPU, nor those of
1225 * CPUs who don't share our TLB.
1227 if ((i != smp_processor_id()) &&
1228 ((smtc_status & SMTC_TLB_SHARED) ||
1229 (cpu_data[i].vpe_id == cpu_data[cpu].vpe_id))) {
1230 settc(cpu_data[i].tc_id);
1231 prevhalt = read_tc_c0_tchalt() & TCHALT_H;
1233 write_tc_c0_tchalt(TCHALT_H);
1236 tcstat = read_tc_c0_tcstatus();
1237 smtc_live_asid[tlb][(tcstat & ASID_MASK)] |= (asiduse)(0x1 << i);
1239 write_tc_c0_tchalt(0);
1242 if (!asid) /* fix version if needed */
1243 asid = ASID_FIRST_VERSION;
1244 local_flush_tlb_all(); /* start new asid cycle */
1246 } while (smtc_live_asid[tlb][(asid & ASID_MASK)]);
1249 * SMTC shares the TLB within VPEs and possibly across all VPEs.
1251 for (i = 0; i < num_online_cpus(); i++) {
1252 if ((smtc_status & SMTC_TLB_SHARED) ||
1253 (cpu_data[i].vpe_id == cpu_data[cpu].vpe_id))
1254 cpu_context(i, mm) = asid_cache(i) = asid;
1257 if (smtc_status & SMTC_TLB_SHARED)
1261 local_irq_restore(flags);
1265 * Invoked from macros defined in mmu_context.h
1266 * which must already have disabled interrupts
1267 * and done a DVPE or DMT as appropriate.
1270 void smtc_flush_tlb_asid(unsigned long asid)
1275 entry = read_c0_wired();
1277 /* Traverse all non-wired entries */
1278 while (entry < current_cpu_data.tlbsize) {
1279 write_c0_index(entry);
1283 ehi = read_c0_entryhi();
1284 if ((ehi & ASID_MASK) == asid) {
1286 * Invalidate only entries with specified ASID,
1287 * makiing sure all entries differ.
1289 write_c0_entryhi(CKSEG0 + (entry << (PAGE_SHIFT + 1)));
1290 write_c0_entrylo0(0);
1291 write_c0_entrylo1(0);
1293 tlb_write_indexed();
1297 write_c0_index(PARKED_INDEX);
1302 * Support for single-threading cache flush operations.
1305 static int halt_state_save[NR_CPUS];
1308 * To really, really be sure that nothing is being done
1309 * by other TCs, halt them all. This code assumes that
1310 * a DVPE has already been done, so while their Halted
1311 * state is theoretically architecturally unstable, in
1312 * practice, it's not going to change while we're looking
1316 void smtc_cflush_lockdown(void)
1320 for_each_online_cpu(cpu) {
1321 if (cpu != smp_processor_id()) {
1322 settc(cpu_data[cpu].tc_id);
1323 halt_state_save[cpu] = read_tc_c0_tchalt();
1324 write_tc_c0_tchalt(TCHALT_H);
1330 /* It would be cheating to change the cpu_online states during a flush! */
1332 void smtc_cflush_release(void)
1337 * Start with a hazard barrier to ensure
1338 * that all CACHE ops have played through.
1342 for_each_online_cpu(cpu) {
1343 if (cpu != smp_processor_id()) {
1344 settc(cpu_data[cpu].tc_id);
1345 write_tc_c0_tchalt(halt_state_save[cpu]);