2 * SN Platform GRU Driver
4 * FAULT HANDLER FOR GRU DETECTED TLB MISSES
6 * This file contains code that handles TLB misses within the GRU.
7 * These misses are reported either via interrupts or user polling of
10 * Copyright (c) 2008 Silicon Graphics, Inc. All Rights Reserved.
12 * This program is free software; you can redistribute it and/or modify
13 * it under the terms of the GNU General Public License as published by
14 * the Free Software Foundation; either version 2 of the License, or
15 * (at your option) any later version.
17 * This program is distributed in the hope that it will be useful,
18 * but WITHOUT ANY WARRANTY; without even the implied warranty of
19 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
20 * GNU General Public License for more details.
22 * You should have received a copy of the GNU General Public License
23 * along with this program; if not, write to the Free Software
24 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
27 #include <linux/kernel.h>
28 #include <linux/errno.h>
29 #include <linux/spinlock.h>
31 #include <linux/hugetlb.h>
32 #include <linux/device.h>
34 #include <linux/uaccess.h>
35 #include <linux/security.h>
36 #include <asm/pgtable.h>
38 #include "grutables.h"
40 #include "gru_instructions.h"
41 #include <asm/uv/uv_hub.h>
44 * Test if a physical address is a valid GRU GSEG address
46 static inline int is_gru_paddr(unsigned long paddr)
48 return paddr >= gru_start_paddr && paddr < gru_end_paddr;
52 * Find the vma of a GRU segment. Caller must hold mmap_sem.
54 struct vm_area_struct *gru_find_vma(unsigned long vaddr)
56 struct vm_area_struct *vma;
58 vma = find_vma(current->mm, vaddr);
59 if (vma && vma->vm_start <= vaddr && vma->vm_ops == &gru_vm_ops)
65 * Find and lock the gts that contains the specified user vaddr.
68 * - *gts with the mmap_sem locked for read and the GTS locked.
69 * - NULL if vaddr invalid OR is not a valid GSEG vaddr.
72 static struct gru_thread_state *gru_find_lock_gts(unsigned long vaddr)
74 struct mm_struct *mm = current->mm;
75 struct vm_area_struct *vma;
76 struct gru_thread_state *gts = NULL;
78 down_read(&mm->mmap_sem);
79 vma = gru_find_vma(vaddr);
81 gts = gru_find_thread_state(vma, TSID(vaddr, vma));
83 mutex_lock(>s->ts_ctxlock);
85 up_read(&mm->mmap_sem);
89 static struct gru_thread_state *gru_alloc_locked_gts(unsigned long vaddr)
91 struct mm_struct *mm = current->mm;
92 struct vm_area_struct *vma;
93 struct gru_thread_state *gts = NULL;
95 down_write(&mm->mmap_sem);
96 vma = gru_find_vma(vaddr);
98 gts = gru_alloc_thread_state(vma, TSID(vaddr, vma));
100 mutex_lock(>s->ts_ctxlock);
101 downgrade_write(&mm->mmap_sem);
103 up_write(&mm->mmap_sem);
110 * Unlock a GTS that was previously locked with gru_find_lock_gts().
112 static void gru_unlock_gts(struct gru_thread_state *gts)
114 mutex_unlock(>s->ts_ctxlock);
115 up_read(¤t->mm->mmap_sem);
119 * Set a CB.istatus to active using a user virtual address. This must be done
120 * just prior to a TFH RESTART. The new cb.istatus is an in-cache status ONLY.
121 * If the line is evicted, the status may be lost. The in-cache update
122 * is necessary to prevent the user from seeing a stale cb.istatus that will
123 * change as soon as the TFH restart is complete. Races may cause an
124 * occasional failure to clear the cb.istatus, but that is ok.
126 * If the cb address is not valid (should not happen, but...), nothing
127 * bad will happen.. The get_user()/put_user() will fail but there
128 * are no bad side-effects.
130 static void gru_cb_set_istatus_active(unsigned long __user *cb)
133 struct gru_instruction_bits bits;
139 u.bits.istatus = CBS_ACTIVE;
145 * Convert a interrupt IRQ to a pointer to the GRU GTS that caused the
146 * interrupt. Interrupts are always sent to a cpu on the blade that contains the
147 * GRU (except for headless blades which are not currently supported). A blade
148 * has N grus; a block of N consecutive IRQs is assigned to the GRUs. The IRQ
149 * number uniquely identifies the GRU chiplet on the local blade that caused the
150 * interrupt. Always called in interrupt context.
152 static inline struct gru_state *irq_to_gru(int irq)
154 return &gru_base[uv_numa_blade_id()]->bs_grus[irq - IRQ_GRU];
160 * The GRU has an array of fault maps. A map is private to a cpu
161 * Only one cpu will be accessing a cpu's fault map.
163 * This function scans the cpu-private fault map & clears all bits that
164 * are set. The function returns a bitmap that indicates the bits that
165 * were cleared. Note that sense the maps may be updated asynchronously by
166 * the GRU, atomic operations must be used to clear bits.
168 static void get_clear_fault_map(struct gru_state *gru,
169 struct gru_tlb_fault_map *imap,
170 struct gru_tlb_fault_map *dmap)
173 struct gru_tlb_fault_map *tfm;
175 tfm = get_tfm_for_cpu(gru, gru_cpu_fault_map_id());
176 prefetchw(tfm); /* Helps on hardware, required for emulator */
177 for (i = 0; i < BITS_TO_LONGS(GRU_NUM_CBE); i++) {
178 k = tfm->fault_bits[i];
180 k = xchg(&tfm->fault_bits[i], 0UL);
181 imap->fault_bits[i] = k;
182 k = tfm->done_bits[i];
184 k = xchg(&tfm->done_bits[i], 0UL);
185 dmap->fault_bits[i] = k;
189 * Not functionally required but helps performance. (Required
192 gru_flush_cache(tfm);
196 * Atomic (interrupt context) & non-atomic (user context) functions to
197 * convert a vaddr into a physical address. The size of the page
198 * is returned in pageshift.
202 * 1 - (atomic only) try again in non-atomic context
204 static int non_atomic_pte_lookup(struct vm_area_struct *vma,
205 unsigned long vaddr, int write,
206 unsigned long *paddr, int *pageshift)
210 /* ZZZ Need to handle HUGE pages */
211 if (is_vm_hugetlb_page(vma))
213 *pageshift = PAGE_SHIFT;
215 (current, current->mm, vaddr, 1, write, 0, &page, NULL) <= 0)
217 *paddr = page_to_phys(page);
225 * Convert a user virtual address to a physical address
226 * Only supports Intel large pages (2MB only) on x86_64.
227 * ZZZ - hugepage support is incomplete
229 * NOTE: mmap_sem is already held on entry to this function. This
230 * guarantees existence of the page tables.
232 static int atomic_pte_lookup(struct vm_area_struct *vma, unsigned long vaddr,
233 int write, unsigned long *paddr, int *pageshift)
240 pgdp = pgd_offset(vma->vm_mm, vaddr);
241 if (unlikely(pgd_none(*pgdp)))
244 pudp = pud_offset(pgdp, vaddr);
245 if (unlikely(pud_none(*pudp)))
248 pmdp = pmd_offset(pudp, vaddr);
249 if (unlikely(pmd_none(*pmdp)))
252 if (unlikely(pmd_large(*pmdp)))
253 pte = *(pte_t *) pmdp;
256 pte = *pte_offset_kernel(pmdp, vaddr);
258 if (unlikely(!pte_present(pte) ||
259 (write && (!pte_write(pte) || !pte_dirty(pte)))))
262 *paddr = pte_pfn(pte) << PAGE_SHIFT;
263 #ifdef CONFIG_HUGETLB_PAGE
264 *pageshift = is_vm_hugetlb_page(vma) ? HPAGE_SHIFT : PAGE_SHIFT;
266 *pageshift = PAGE_SHIFT;
275 static int gru_vtop(struct gru_thread_state *gts, unsigned long vaddr,
276 int write, int atomic, unsigned long *gpa, int *pageshift)
278 struct mm_struct *mm = gts->ts_mm;
279 struct vm_area_struct *vma;
283 vma = find_vma(mm, vaddr);
288 * Atomic lookup is faster & usually works even if called in non-atomic
291 rmb(); /* Must/check ms_range_active before loading PTEs */
292 ret = atomic_pte_lookup(vma, vaddr, write, &paddr, &ps);
296 if (non_atomic_pte_lookup(vma, vaddr, write, &paddr, &ps))
299 if (is_gru_paddr(paddr))
301 paddr = paddr & ~((1UL << ps) - 1);
302 *gpa = uv_soc_phys_ram_to_gpa(paddr);
314 * Drop a TLB entry into the GRU. The fault is described by info in an TFH.
316 * cb Address of user CBR. Null if not running in user context
318 * 0 = dropin, exception, or switch to UPM successful
319 * 1 = range invalidate active
323 static int gru_try_dropin(struct gru_thread_state *gts,
324 struct gru_tlb_fault_handle *tfh,
325 unsigned long __user *cb)
327 int pageshift = 0, asid, write, ret, atomic = !cb;
328 unsigned long gpa = 0, vaddr = 0;
331 * NOTE: The GRU contains magic hardware that eliminates races between
332 * TLB invalidates and TLB dropins. If an invalidate occurs
333 * in the window between reading the TFH and the subsequent TLB dropin,
334 * the dropin is ignored. This eliminates the need for additional locks.
338 * Error if TFH state is IDLE or FMM mode & the user issuing a UPM call.
339 * Might be a hardware race OR a stupid user. Ignore FMM because FMM
340 * is a transient state.
342 if (tfh->status != TFHSTATUS_EXCEPTION) {
343 gru_flush_cache(tfh);
344 if (tfh->status != TFHSTATUS_EXCEPTION)
345 goto failnoexception;
346 STAT(tfh_stale_on_fault);
348 if (tfh->state == TFHSTATE_IDLE)
350 if (tfh->state == TFHSTATE_MISS_FMM && cb)
353 write = (tfh->cause & TFHCAUSE_TLB_MOD) != 0;
354 vaddr = tfh->missvaddr;
355 asid = tfh->missasid;
359 rmb(); /* TFH must be cache resident before reading ms_range_active */
362 * TFH is cache resident - at least briefly. Fail the dropin
363 * if a range invalidate is active.
365 if (atomic_read(>s->ts_gms->ms_range_active))
368 ret = gru_vtop(gts, vaddr, write, atomic, &gpa, &pageshift);
374 if (!(gts->ts_sizeavail & GRU_SIZEAVAIL(pageshift))) {
375 gts->ts_sizeavail |= GRU_SIZEAVAIL(pageshift);
376 if (atomic || !gru_update_cch(gts, 0)) {
377 gts->ts_force_cch_reload = 1;
381 gru_cb_set_istatus_active(cb);
382 tfh_write_restart(tfh, gpa, GAA_RAM, vaddr, asid, write,
383 GRU_PAGESIZE(pageshift));
386 "%s: tfh 0x%p, vaddr 0x%lx, asid 0x%x, ps %d, gpa 0x%lx\n",
387 ret ? "non-atomic" : "atomic", tfh, vaddr, asid,
392 /* No asid (delayed unload). */
393 STAT(tlb_dropin_fail_no_asid);
394 gru_dbg(grudev, "FAILED no_asid tfh: 0x%p, vaddr 0x%lx\n", tfh, vaddr);
396 tfh_user_polling_mode(tfh);
398 gru_flush_cache(tfh);
402 /* Atomic failure switch CBR to UPM */
403 tfh_user_polling_mode(tfh);
404 STAT(tlb_dropin_fail_upm);
405 gru_dbg(grudev, "FAILED upm tfh: 0x%p, vaddr 0x%lx\n", tfh, vaddr);
409 /* FMM state on UPM call */
410 gru_flush_cache(tfh);
411 STAT(tlb_dropin_fail_fmm);
412 gru_dbg(grudev, "FAILED fmm tfh: 0x%p, state %d\n", tfh, tfh->state);
416 /* TFH status did not show exception pending */
417 gru_flush_cache(tfh);
420 STAT(tlb_dropin_fail_no_exception);
421 gru_dbg(grudev, "FAILED non-exception tfh: 0x%p, status %d, state %d\n", tfh, tfh->status, tfh->state);
425 /* TFH state was idle - no miss pending */
426 gru_flush_cache(tfh);
429 STAT(tlb_dropin_fail_idle);
430 gru_dbg(grudev, "FAILED idle tfh: 0x%p, state %d\n", tfh, tfh->state);
434 /* All errors (atomic & non-atomic) switch CBR to EXCEPTION state */
436 STAT(tlb_dropin_fail_invalid);
437 gru_dbg(grudev, "FAILED inval tfh: 0x%p, vaddr 0x%lx\n", tfh, vaddr);
441 /* Range invalidate active. Switch to UPM iff atomic */
443 tfh_user_polling_mode(tfh);
445 gru_flush_cache(tfh);
446 STAT(tlb_dropin_fail_range_active);
447 gru_dbg(grudev, "FAILED range active: tfh 0x%p, vaddr 0x%lx\n",
453 * Process an external interrupt from the GRU. This interrupt is
454 * caused by a TLB miss.
455 * Note that this is the interrupt handler that is registered with linux
456 * interrupt handlers.
458 irqreturn_t gru_intr(int irq, void *dev_id)
460 struct gru_state *gru;
461 struct gru_tlb_fault_map imap, dmap;
462 struct gru_thread_state *gts;
463 struct gru_tlb_fault_handle *tfh = NULL;
468 gru = irq_to_gru(irq);
470 dev_err(grudev, "GRU: invalid interrupt: cpu %d, irq %d\n",
471 raw_smp_processor_id(), irq);
474 get_clear_fault_map(gru, &imap, &dmap);
476 for_each_cbr_in_tfm(cbrnum, dmap.fault_bits) {
477 complete(gru->gs_blade->bs_async_wq);
478 gru_dbg(grudev, "gid %d, cbr_done %d, done %d\n",
479 gru->gs_gid, cbrnum, gru->gs_blade->bs_async_wq->done);
482 for_each_cbr_in_tfm(cbrnum, imap.fault_bits) {
483 tfh = get_tfh_by_index(gru, cbrnum);
484 prefetchw(tfh); /* Helps on hdw, required for emulator */
487 * When hardware sets a bit in the faultmap, it implicitly
488 * locks the GRU context so that it cannot be unloaded.
489 * The gts cannot change until a TFH start/writestart command
492 ctxnum = tfh->ctxnum;
493 gts = gru->gs_gts[ctxnum];
496 * This is running in interrupt context. Trylock the mmap_sem.
497 * If it fails, retry the fault in user context.
499 if (!gts->ts_force_cch_reload &&
500 down_read_trylock(>s->ts_mm->mmap_sem)) {
501 gts->ustats.fmm_tlbdropin++;
502 gru_try_dropin(gts, tfh, NULL);
503 up_read(>s->ts_mm->mmap_sem);
505 tfh_user_polling_mode(tfh);
506 STAT(intr_mm_lock_failed);
513 static int gru_user_dropin(struct gru_thread_state *gts,
514 struct gru_tlb_fault_handle *tfh,
515 unsigned long __user *cb)
517 struct gru_mm_struct *gms = gts->ts_gms;
520 gts->ustats.upm_tlbdropin++;
522 wait_event(gms->ms_wait_queue,
523 atomic_read(&gms->ms_range_active) == 0);
524 prefetchw(tfh); /* Helps on hdw, required for emulator */
525 ret = gru_try_dropin(gts, tfh, cb);
528 STAT(call_os_wait_queue);
533 * This interface is called as a result of a user detecting a "call OS" bit
534 * in a user CB. Normally means that a TLB fault has occurred.
535 * cb - user virtual address of the CB
537 int gru_handle_user_call_os(unsigned long cb)
539 struct gru_tlb_fault_handle *tfh;
540 struct gru_thread_state *gts;
541 unsigned long __user *cbp;
542 int ucbnum, cbrnum, ret = -EINVAL;
545 gru_dbg(grudev, "address 0x%lx\n", cb);
547 /* sanity check the cb pointer */
548 ucbnum = get_cb_number((void *)cb);
549 if ((cb & (GRU_HANDLE_STRIDE - 1)) || ucbnum >= GRU_NUM_CB)
551 cbp = (unsigned long *)cb;
553 gts = gru_find_lock_gts(cb);
557 if (ucbnum >= gts->ts_cbr_au_count * GRU_CBR_AU_SIZE)
561 * If force_unload is set, the UPM TLB fault is phony. The task
562 * has migrated to another node and the GSEG must be moved. Just
563 * unload the context. The task will page fault and assign a new
566 if (gts->ts_tgid_owner == current->tgid && gts->ts_blade >= 0 &&
567 gts->ts_blade != uv_numa_blade_id()) {
568 STAT(call_os_offnode_reference);
569 gts->ts_force_unload = 1;
573 * CCH may contain stale data if ts_force_cch_reload is set.
575 if (gts->ts_gru && gts->ts_force_cch_reload) {
576 gts->ts_force_cch_reload = 0;
577 gru_update_cch(gts, 0);
581 cbrnum = thread_cbr_number(gts, ucbnum);
582 if (gts->ts_force_unload) {
583 gru_unload_context(gts, 1);
584 } else if (gts->ts_gru) {
585 tfh = get_tfh_by_index(gts->ts_gru, cbrnum);
586 ret = gru_user_dropin(gts, tfh, cbp);
594 * Fetch the exception detail information for a CB that terminated with
597 int gru_get_exception_detail(unsigned long arg)
599 struct control_block_extended_exc_detail excdet;
600 struct gru_control_block_extended *cbe;
601 struct gru_thread_state *gts;
602 int ucbnum, cbrnum, ret;
604 STAT(user_exception);
605 if (copy_from_user(&excdet, (void __user *)arg, sizeof(excdet)))
608 gru_dbg(grudev, "address 0x%lx\n", excdet.cb);
609 gts = gru_find_lock_gts(excdet.cb);
613 ucbnum = get_cb_number((void *)excdet.cb);
614 if (ucbnum >= gts->ts_cbr_au_count * GRU_CBR_AU_SIZE) {
616 } else if (gts->ts_gru) {
617 cbrnum = thread_cbr_number(gts, ucbnum);
618 cbe = get_cbe_by_index(gts->ts_gru, cbrnum);
619 gru_flush_cache(cbe); /* CBE not coherent */
620 excdet.opc = cbe->opccpy;
621 excdet.exopc = cbe->exopccpy;
622 excdet.ecause = cbe->ecause;
623 excdet.exceptdet0 = cbe->idef1upd;
624 excdet.exceptdet1 = cbe->idef3upd;
625 excdet.cbrstate = cbe->cbrstate;
626 excdet.cbrexecstatus = cbe->cbrexecstatus;
627 gru_flush_cache(cbe);
635 "cb 0x%lx, op %d, exopc %d, cbrstate %d, cbrexecstatus 0x%x, ecause 0x%x, "
636 "exdet0 0x%lx, exdet1 0x%x\n",
637 excdet.cb, excdet.opc, excdet.exopc, excdet.cbrstate, excdet.cbrexecstatus,
638 excdet.ecause, excdet.exceptdet0, excdet.exceptdet1);
639 if (!ret && copy_to_user((void __user *)arg, &excdet, sizeof(excdet)))
645 * User request to unload a context. Content is saved for possible reload.
647 static int gru_unload_all_contexts(void)
649 struct gru_thread_state *gts;
650 struct gru_state *gru;
653 if (!capable(CAP_SYS_ADMIN))
656 gru = GID_TO_GRU(gid);
657 spin_lock(&gru->gs_lock);
658 for (ctxnum = 0; ctxnum < GRU_NUM_CCH; ctxnum++) {
659 gts = gru->gs_gts[ctxnum];
660 if (gts && mutex_trylock(>s->ts_ctxlock)) {
661 spin_unlock(&gru->gs_lock);
662 gru_unload_context(gts, 1);
663 mutex_unlock(>s->ts_ctxlock);
664 spin_lock(&gru->gs_lock);
667 spin_unlock(&gru->gs_lock);
672 int gru_user_unload_context(unsigned long arg)
674 struct gru_thread_state *gts;
675 struct gru_unload_context_req req;
677 STAT(user_unload_context);
678 if (copy_from_user(&req, (void __user *)arg, sizeof(req)))
681 gru_dbg(grudev, "gseg 0x%lx\n", req.gseg);
684 return gru_unload_all_contexts();
686 gts = gru_find_lock_gts(req.gseg);
691 gru_unload_context(gts, 1);
698 * User request to flush a range of virtual addresses from the GRU TLB
699 * (Mainly for testing).
701 int gru_user_flush_tlb(unsigned long arg)
703 struct gru_thread_state *gts;
704 struct gru_flush_tlb_req req;
705 struct gru_mm_struct *gms;
707 STAT(user_flush_tlb);
708 if (copy_from_user(&req, (void __user *)arg, sizeof(req)))
711 gru_dbg(grudev, "gseg 0x%lx, vaddr 0x%lx, len 0x%lx\n", req.gseg,
714 gts = gru_find_lock_gts(req.gseg);
720 gru_flush_tlb_range(gms, req.vaddr, req.len);
726 * Fetch GSEG statisticss
728 long gru_get_gseg_statistics(unsigned long arg)
730 struct gru_thread_state *gts;
731 struct gru_get_gseg_statistics_req req;
733 if (copy_from_user(&req, (void __user *)arg, sizeof(req)))
736 gts = gru_find_lock_gts(req.gseg);
738 memcpy(&req.stats, >s->ustats, sizeof(gts->ustats));
741 memset(&req.stats, 0, sizeof(gts->ustats));
744 if (copy_to_user((void __user *)arg, &req, sizeof(req)))
751 * Register the current task as the user of the GSEG slice.
752 * Needed for TLB fault interrupt targeting.
754 int gru_set_context_option(unsigned long arg)
756 struct gru_thread_state *gts;
757 struct gru_set_context_option_req req;
760 STAT(set_context_option);
761 if (copy_from_user(&req, (void __user *)arg, sizeof(req)))
763 gru_dbg(grudev, "op %d, gseg 0x%lx, value1 0x%lx\n", req.op, req.gseg, req.val1);
765 gts = gru_alloc_locked_gts(req.gseg);
771 /* Register the current task as the GSEG owner */
772 gts->ts_tgid_owner = current->tgid;
774 case sco_cch_req_slice:
775 /* Set the CCH slice option */
776 gts->ts_cch_req_slice = req.val1 & 3;