#include <linux/device.h>
#include <linux/io.h>
#include <linux/uaccess.h>
+#include <linux/security.h>
#include <asm/pgtable.h>
#include "gru.h"
#include "grutables.h"
* the GRU, atomic operations must be used to clear bits.
*/
static void get_clear_fault_map(struct gru_state *gru,
- struct gru_tlb_fault_map *map)
+ struct gru_tlb_fault_map *imap,
+ struct gru_tlb_fault_map *dmap)
{
unsigned long i, k;
struct gru_tlb_fault_map *tfm;
k = tfm->fault_bits[i];
if (k)
k = xchg(&tfm->fault_bits[i], 0UL);
- map->fault_bits[i] = k;
+ imap->fault_bits[i] = k;
+ k = tfm->done_bits[i];
+ if (k)
+ k = xchg(&tfm->done_bits[i], 0UL);
+ dmap->fault_bits[i] = k;
}
/*
}
/*
- *
* atomic_pte_lookup
*
* Convert a user virtual address to a physical address
* Only supports Intel large pages (2MB only) on x86_64.
* ZZZ - hugepage support is incomplete
+ *
+ * NOTE: mmap_sem is already held on entry to this function. This
+ * guarantees existence of the page tables.
*/
static int atomic_pte_lookup(struct vm_area_struct *vma, unsigned long vaddr,
int write, unsigned long *paddr, int *pageshift)
pud_t *pudp;
pte_t pte;
- WARN_ON(irqs_disabled()); /* ZZZ debug */
-
- local_irq_disable();
pgdp = pgd_offset(vma->vm_mm, vaddr);
if (unlikely(pgd_none(*pgdp)))
goto err;
#endif
pte = *pte_offset_kernel(pmdp, vaddr);
- local_irq_enable();
-
if (unlikely(!pte_present(pte) ||
(write && (!pte_write(pte) || !pte_dirty(pte)))))
return 1;
*paddr = pte_pfn(pte) << PAGE_SHIFT;
+#ifdef CONFIG_HUGETLB_PAGE
*pageshift = is_vm_hugetlb_page(vma) ? HPAGE_SHIFT : PAGE_SHIFT;
+#else
+ *pageshift = PAGE_SHIFT;
+#endif
return 0;
err:
return 1;
}
+static int gru_vtop(struct gru_thread_state *gts, unsigned long vaddr,
+ int write, int atomic, unsigned long *gpa, int *pageshift)
+{
+ struct mm_struct *mm = gts->ts_mm;
+ struct vm_area_struct *vma;
+ unsigned long paddr;
+ int ret, ps;
+
+ vma = find_vma(mm, vaddr);
+ if (!vma)
+ goto inval;
+
+ /*
+ * Atomic lookup is faster & usually works even if called in non-atomic
+ * context.
+ */
+ rmb(); /* Must/check ms_range_active before loading PTEs */
+ ret = atomic_pte_lookup(vma, vaddr, write, &paddr, &ps);
+ if (ret) {
+ if (atomic)
+ goto upm;
+ if (non_atomic_pte_lookup(vma, vaddr, write, &paddr, &ps))
+ goto inval;
+ }
+ if (is_gru_paddr(paddr))
+ goto inval;
+ paddr = paddr & ~((1UL << ps) - 1);
+ *gpa = uv_soc_phys_ram_to_gpa(paddr);
+ *pageshift = ps;
+ return 0;
+
+inval:
+ return -1;
+upm:
+ return -2;
+}
+
+
/*
* Drop a TLB entry into the GRU. The fault is described by info in an TFH.
* Input:
struct gru_tlb_fault_handle *tfh,
unsigned long __user *cb)
{
- struct mm_struct *mm = gts->ts_mm;
- struct vm_area_struct *vma;
- int pageshift, asid, write, ret;
- unsigned long paddr, gpa, vaddr;
+ int pageshift = 0, asid, write, ret, atomic = !cb;
+ unsigned long gpa = 0, vaddr = 0;
/*
* NOTE: The GRU contains magic hardware that eliminates races between
* Might be a hardware race OR a stupid user. Ignore FMM because FMM
* is a transient state.
*/
+ if (tfh->status != TFHSTATUS_EXCEPTION)
+ goto failnoexception;
if (tfh->state == TFHSTATE_IDLE)
goto failidle;
if (tfh->state == TFHSTATE_MISS_FMM && cb)
if (atomic_read(>s->ts_gms->ms_range_active))
goto failactive;
- vma = find_vma(mm, vaddr);
- if (!vma)
+ ret = gru_vtop(gts, vaddr, write, atomic, &gpa, &pageshift);
+ if (ret == -1)
goto failinval;
+ if (ret == -2)
+ goto failupm;
- /*
- * Atomic lookup is faster & usually works even if called in non-atomic
- * context.
- */
- ret = atomic_pte_lookup(vma, vaddr, write, &paddr, &pageshift);
- if (ret) {
- if (!cb)
+ if (!(gts->ts_sizeavail & GRU_SIZEAVAIL(pageshift))) {
+ gts->ts_sizeavail |= GRU_SIZEAVAIL(pageshift);
+ if (atomic || !gru_update_cch(gts, 0)) {
+ gts->ts_force_cch_reload = 1;
goto failupm;
- if (non_atomic_pte_lookup(vma, vaddr, write, &paddr,
- &pageshift))
- goto failinval;
+ }
}
- if (is_gru_paddr(paddr))
- goto failinval;
-
- paddr = paddr & ~((1UL << pageshift) - 1);
- gpa = uv_soc_phys_ram_to_gpa(paddr);
gru_cb_set_istatus_active(cb);
tfh_write_restart(tfh, gpa, GAA_RAM, vaddr, asid, write,
GRU_PAGESIZE(pageshift));
failfmm:
/* FMM state on UPM call */
+ gru_flush_cache(tfh);
STAT(tlb_dropin_fail_fmm);
gru_dbg(grudev, "FAILED fmm tfh: 0x%p, state %d\n", tfh, tfh->state);
return 0;
+failnoexception:
+ /* TFH status did not show exception pending */
+ gru_flush_cache(tfh);
+ if (cb)
+ gru_flush_cache(cb);
+ STAT(tlb_dropin_fail_no_exception);
+ gru_dbg(grudev, "FAILED non-exception tfh: 0x%p, status %d, state %d\n", tfh, tfh->status, tfh->state);
+ return 0;
+
failidle:
- /* TFH was idle - no miss pending */
+ /* TFH state was idle - no miss pending */
gru_flush_cache(tfh);
if (cb)
gru_flush_cache(cb);
irqreturn_t gru_intr(int irq, void *dev_id)
{
struct gru_state *gru;
- struct gru_tlb_fault_map map;
+ struct gru_tlb_fault_map imap, dmap;
struct gru_thread_state *gts;
struct gru_tlb_fault_handle *tfh = NULL;
int cbrnum, ctxnum;
raw_smp_processor_id(), irq);
return IRQ_NONE;
}
- get_clear_fault_map(gru, &map);
- gru_dbg(grudev, "irq %d, gru %x, map 0x%lx\n", irq, gru->gs_gid,
- map.fault_bits[0]);
+ get_clear_fault_map(gru, &imap, &dmap);
+ gru_dbg(grudev,
+ "irq %d, gid %d, imap %016lx %016lx, dmap %016lx %016lx\n",
+ irq, gru->gs_gid, dmap.fault_bits[0], dmap.fault_bits[1],
+ dmap.fault_bits[0], dmap.fault_bits[1]);
+
+ for_each_cbr_in_tfm(cbrnum, dmap.fault_bits) {
+ complete(gru->gs_blade->bs_async_wq);
+ gru_dbg(grudev, "gid %d, cbr_done %d, done %d\n",
+ gru->gs_gid, cbrnum, gru->gs_blade->bs_async_wq->done);
+ }
- for_each_cbr_in_tfm(cbrnum, map.fault_bits) {
+ for_each_cbr_in_tfm(cbrnum, imap.fault_bits) {
tfh = get_tfh_by_index(gru, cbrnum);
prefetchw(tfh); /* Helps on hdw, required for emulator */
* This is running in interrupt context. Trylock the mmap_sem.
* If it fails, retry the fault in user context.
*/
- if (down_read_trylock(>s->ts_mm->mmap_sem)) {
+ if (!gts->ts_force_cch_reload &&
+ down_read_trylock(>s->ts_mm->mmap_sem)) {
gru_try_dropin(gts, tfh, NULL);
up_read(>s->ts_mm->mmap_sem);
} else {
tfh_user_polling_mode(tfh);
+ STAT(intr_mm_lock_failed);
}
}
return IRQ_HANDLED;
if (!gts)
return -EINVAL;
- if (ucbnum >= gts->ts_cbr_au_count * GRU_CBR_AU_SIZE) {
- ret = -EINVAL;
+ if (ucbnum >= gts->ts_cbr_au_count * GRU_CBR_AU_SIZE)
goto exit;
- }
/*
* If force_unload is set, the UPM TLB fault is phony. The task
* unload the context. The task will page fault and assign a new
* context.
*/
+ if (gts->ts_tgid_owner == current->tgid && gts->ts_blade >= 0 &&
+ gts->ts_blade != uv_numa_blade_id()) {
+ STAT(call_os_offnode_reference);
+ gts->ts_force_unload = 1;
+ }
+
+ /*
+ * CCH may contain stale data if ts_force_cch_reload is set.
+ */
+ if (gts->ts_gru && gts->ts_force_cch_reload) {
+ gts->ts_force_cch_reload = 0;
+ gru_update_cch(gts, 0);
+ }
+
ret = -EAGAIN;
cbrnum = thread_cbr_number(gts, ucbnum);
if (gts->ts_force_unload) {
if (!gts)
return -EINVAL;
- if (gts->ts_gru) {
- ucbnum = get_cb_number((void *)excdet.cb);
+ ucbnum = get_cb_number((void *)excdet.cb);
+ if (ucbnum >= gts->ts_cbr_au_count * GRU_CBR_AU_SIZE) {
+ ret = -EINVAL;
+ } else if (gts->ts_gru) {
cbrnum = thread_cbr_number(gts, ucbnum);
cbe = get_cbe_by_index(gts->ts_gru, cbrnum);
+ prefetchw(cbe);/* Harmless on hardware, required for emulator */
excdet.opc = cbe->opccpy;
excdet.exopc = cbe->exopccpy;
excdet.ecause = cbe->ecause;
excdet.exceptdet0 = cbe->idef1upd;
excdet.exceptdet1 = cbe->idef3upd;
+ excdet.cbrstate = cbe->cbrstate;
+ excdet.cbrexecstatus = cbe->cbrexecstatus;
ret = 0;
} else {
ret = -EAGAIN;
}
gru_unlock_gts(gts);
- gru_dbg(grudev, "address 0x%lx, ecause 0x%x\n", excdet.cb,
- excdet.ecause);
+ gru_dbg(grudev,
+ "cb 0x%lx, op %d, exopc %d, cbrstate %d, cbrexecstatus 0x%x, ecause 0x%x, "
+ "exdet0 0x%lx, exdet1 0x%x\n",
+ excdet.cb, excdet.opc, excdet.exopc, excdet.cbrstate, excdet.cbrexecstatus,
+ excdet.ecause, excdet.exceptdet0, excdet.exceptdet1);
if (!ret && copy_to_user((void __user *)arg, &excdet, sizeof(excdet)))
ret = -EFAULT;
return ret;
/*
* User request to unload a context. Content is saved for possible reload.
*/
+static int gru_unload_all_contexts(void)
+{
+ struct gru_thread_state *gts;
+ struct gru_state *gru;
+ int gid, ctxnum;
+
+ if (!capable(CAP_SYS_ADMIN))
+ return -EPERM;
+ foreach_gid(gid) {
+ gru = GID_TO_GRU(gid);
+ spin_lock(&gru->gs_lock);
+ for (ctxnum = 0; ctxnum < GRU_NUM_CCH; ctxnum++) {
+ gts = gru->gs_gts[ctxnum];
+ if (gts && mutex_trylock(>s->ts_ctxlock)) {
+ spin_unlock(&gru->gs_lock);
+ gru_unload_context(gts, 1);
+ mutex_unlock(>s->ts_ctxlock);
+ spin_lock(&gru->gs_lock);
+ }
+ }
+ spin_unlock(&gru->gs_lock);
+ }
+ return 0;
+}
+
int gru_user_unload_context(unsigned long arg)
{
struct gru_thread_state *gts;
gru_dbg(grudev, "gseg 0x%lx\n", req.gseg);
+ if (!req.gseg)
+ return gru_unload_all_contexts();
+
gts = gru_find_lock_gts(req.gseg);
if (!gts)
return -EINVAL;
if (!gts)
return -EINVAL;
- gru_flush_tlb_range(gts->ts_gms, req.vaddr, req.vaddr + req.len);
+ gru_flush_tlb_range(gts->ts_gms, req.vaddr, req.len);
gru_unlock_gts(gts);
return 0;