]> www.pilppa.org Git - linux-2.6-omap-h63xx.git/commitdiff
GRU Driver: resource management
authorJack Steiner <steiner@sgi.com>
Wed, 30 Jul 2008 05:33:58 +0000 (22:33 -0700)
committerLinus Torvalds <torvalds@linux-foundation.org>
Wed, 30 Jul 2008 16:41:48 +0000 (09:41 -0700)
This file contains functions realted to managing GRU resources provided to
the user.  Examples include GRU context assignment, load, unload,
migration, etc..

Signed-off-by: Jack Steiner <steiner@sgi.com>
Cc: Nick Piggin <nickpiggin@yahoo.com.au>
Cc: Hugh Dickins <hugh@veritas.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
drivers/misc/sgi-gru/grumain.c [new file with mode: 0644]

diff --git a/drivers/misc/sgi-gru/grumain.c b/drivers/misc/sgi-gru/grumain.c
new file mode 100644 (file)
index 0000000..aef6822
--- /dev/null
@@ -0,0 +1,798 @@
+/*
+ * SN Platform GRU Driver
+ *
+ *            DRIVER TABLE MANAGER + GRU CONTEXT LOAD/UNLOAD
+ *
+ * This file is subject to the terms and conditions of the GNU General Public
+ * License.  See the file "COPYING" in the main directory of this archive
+ * for more details.
+ *
+ * Copyright (c) 2008 Silicon Graphics, Inc.  All Rights Reserved.
+ */
+
+#include <linux/kernel.h>
+#include <linux/slab.h>
+#include <linux/mm.h>
+#include <linux/spinlock.h>
+#include <linux/sched.h>
+#include <linux/device.h>
+#include <linux/list.h>
+#include <asm/uv/uv_hub.h>
+#include "gru.h"
+#include "grutables.h"
+#include "gruhandles.h"
+
+unsigned long options __read_mostly;
+
+static struct device_driver gru_driver = {
+       .name = "gru"
+};
+
+static struct device gru_device = {
+       .bus_id = {0},
+       .driver = &gru_driver,
+};
+
+struct device *grudev = &gru_device;
+
+/*
+ * Select a gru fault map to be used by the current cpu. Note that
+ * multiple cpus may be using the same map.
+ *     ZZZ should "shift" be used?? Depends on HT cpu numbering
+ *     ZZZ should be inline but did not work on emulator
+ */
+int gru_cpu_fault_map_id(void)
+{
+       return uv_blade_processor_id() % GRU_NUM_TFM;
+}
+
+/*--------- ASID Management -------------------------------------------
+ *
+ *  Initially, assign asids sequentially from MIN_ASID .. MAX_ASID.
+ *  Once MAX is reached, flush the TLB & start over. However,
+ *  some asids may still be in use. There won't be many (percentage wise) still
+ *  in use. Search active contexts & determine the value of the first
+ *  asid in use ("x"s below). Set "limit" to this value.
+ *  This defines a block of assignable asids.
+ *
+ *  When "limit" is reached, search forward from limit+1 and determine the
+ *  next block of assignable asids.
+ *
+ *  Repeat until MAX_ASID is reached, then start over again.
+ *
+ *  Each time MAX_ASID is reached, increment the asid generation. Since
+ *  the search for in-use asids only checks contexts with GRUs currently
+ *  assigned, asids in some contexts will be missed. Prior to loading
+ *  a context, the asid generation of the GTS asid is rechecked. If it
+ *  doesn't match the current generation, a new asid will be assigned.
+ *
+ *     0---------------x------------x---------------------x----|
+ *       ^-next        ^-limit                                 ^-MAX_ASID
+ *
+ * All asid manipulation & context loading/unloading is protected by the
+ * gs_lock.
+ */
+
+/* Hit the asid limit. Start over */
+static int gru_wrap_asid(struct gru_state *gru)
+{
+       gru_dbg(grudev, "gru %p\n", gru);
+       STAT(asid_wrap);
+       gru->gs_asid_gen++;
+       gru_flush_all_tlb(gru);
+       return MIN_ASID;
+}
+
+/* Find the next chunk of unused asids */
+static int gru_reset_asid_limit(struct gru_state *gru, int asid)
+{
+       int i, gid, inuse_asid, limit;
+
+       gru_dbg(grudev, "gru %p, asid 0x%x\n", gru, asid);
+       STAT(asid_next);
+       limit = MAX_ASID;
+       if (asid >= limit)
+               asid = gru_wrap_asid(gru);
+       gid = gru->gs_gid;
+again:
+       for (i = 0; i < GRU_NUM_CCH; i++) {
+               if (!gru->gs_gts[i])
+                       continue;
+               inuse_asid = gru->gs_gts[i]->ts_gms->ms_asids[gid].mt_asid;
+               gru_dbg(grudev, "gru %p, inuse_asid 0x%x, cxtnum %d, gts %p\n",
+                       gru, inuse_asid, i, gru->gs_gts[i]);
+               if (inuse_asid == asid) {
+                       asid += ASID_INC;
+                       if (asid >= limit) {
+                               /*
+                                * empty range: reset the range limit and
+                                * start over
+                                */
+                               limit = MAX_ASID;
+                               if (asid >= MAX_ASID)
+                                       asid = gru_wrap_asid(gru);
+                               goto again;
+                       }
+               }
+
+               if ((inuse_asid > asid) && (inuse_asid < limit))
+                       limit = inuse_asid;
+       }
+       gru->gs_asid_limit = limit;
+       gru->gs_asid = asid;
+       gru_dbg(grudev, "gru %p, new asid 0x%x, new_limit 0x%x\n", gru, asid,
+               limit);
+       return asid;
+}
+
+/* Assign a new ASID to a thread context.  */
+static int gru_assign_asid(struct gru_state *gru)
+{
+       int asid;
+
+       spin_lock(&gru->gs_asid_lock);
+       gru->gs_asid += ASID_INC;
+       asid = gru->gs_asid;
+       if (asid >= gru->gs_asid_limit)
+               asid = gru_reset_asid_limit(gru, asid);
+       spin_unlock(&gru->gs_asid_lock);
+
+       gru_dbg(grudev, "gru %p, asid 0x%x\n", gru, asid);
+       return asid;
+}
+
+/*
+ * Clear n bits in a word. Return a word indicating the bits that were cleared.
+ * Optionally, build an array of chars that contain the bit numbers allocated.
+ */
+static unsigned long reserve_resources(unsigned long *p, int n, int mmax,
+                                      char *idx)
+{
+       unsigned long bits = 0;
+       int i;
+
+       do {
+               i = find_first_bit(p, mmax);
+               if (i == mmax)
+                       BUG();
+               __clear_bit(i, p);
+               __set_bit(i, &bits);
+               if (idx)
+                       *idx++ = i;
+       } while (--n);
+       return bits;
+}
+
+unsigned long reserve_gru_cb_resources(struct gru_state *gru, int cbr_au_count,
+                                      char *cbmap)
+{
+       return reserve_resources(&gru->gs_cbr_map, cbr_au_count, GRU_CBR_AU,
+                                cbmap);
+}
+
+unsigned long reserve_gru_ds_resources(struct gru_state *gru, int dsr_au_count,
+                                      char *dsmap)
+{
+       return reserve_resources(&gru->gs_dsr_map, dsr_au_count, GRU_DSR_AU,
+                                dsmap);
+}
+
+static void reserve_gru_resources(struct gru_state *gru,
+                                 struct gru_thread_state *gts)
+{
+       gru->gs_active_contexts++;
+       gts->ts_cbr_map =
+           reserve_gru_cb_resources(gru, gts->ts_cbr_au_count,
+                                    gts->ts_cbr_idx);
+       gts->ts_dsr_map =
+           reserve_gru_ds_resources(gru, gts->ts_dsr_au_count, NULL);
+}
+
+static void free_gru_resources(struct gru_state *gru,
+                              struct gru_thread_state *gts)
+{
+       gru->gs_active_contexts--;
+       gru->gs_cbr_map |= gts->ts_cbr_map;
+       gru->gs_dsr_map |= gts->ts_dsr_map;
+}
+
+/*
+ * Check if a GRU has sufficient free resources to satisfy an allocation
+ * request. Note: GRU locks may or may not be held when this is called. If
+ * not held, recheck after acquiring the appropriate locks.
+ *
+ * Returns 1 if sufficient resources, 0 if not
+ */
+static int check_gru_resources(struct gru_state *gru, int cbr_au_count,
+                              int dsr_au_count, int max_active_contexts)
+{
+       return hweight64(gru->gs_cbr_map) >= cbr_au_count
+               && hweight64(gru->gs_dsr_map) >= dsr_au_count
+               && gru->gs_active_contexts < max_active_contexts;
+}
+
+/*
+ * TLB manangment requires tracking all GRU chiplets that have loaded a GSEG
+ * context.
+ */
+static int gru_load_mm_tracker(struct gru_state *gru, struct gru_mm_struct *gms,
+                              int ctxnum)
+{
+       struct gru_mm_tracker *asids = &gms->ms_asids[gru->gs_gid];
+       unsigned short ctxbitmap = (1 << ctxnum);
+       int asid;
+
+       spin_lock(&gms->ms_asid_lock);
+       asid = asids->mt_asid;
+
+       if (asid == 0 || asids->mt_asid_gen != gru->gs_asid_gen) {
+               asid = gru_assign_asid(gru);
+               asids->mt_asid = asid;
+               asids->mt_asid_gen = gru->gs_asid_gen;
+               STAT(asid_new);
+       } else {
+               STAT(asid_reuse);
+       }
+
+       BUG_ON(asids->mt_ctxbitmap & ctxbitmap);
+       asids->mt_ctxbitmap |= ctxbitmap;
+       if (!test_bit(gru->gs_gid, gms->ms_asidmap))
+               __set_bit(gru->gs_gid, gms->ms_asidmap);
+       spin_unlock(&gms->ms_asid_lock);
+
+       gru_dbg(grudev,
+               "gru %x, gms %p, ctxnum 0x%d, asid 0x%x, asidmap 0x%lx\n",
+               gru->gs_gid, gms, ctxnum, asid, gms->ms_asidmap[0]);
+       return asid;
+}
+
+static void gru_unload_mm_tracker(struct gru_state *gru,
+                                 struct gru_mm_struct *gms, int ctxnum)
+{
+       struct gru_mm_tracker *asids;
+       unsigned short ctxbitmap;
+
+       asids = &gms->ms_asids[gru->gs_gid];
+       ctxbitmap = (1 << ctxnum);
+       spin_lock(&gms->ms_asid_lock);
+       BUG_ON((asids->mt_ctxbitmap & ctxbitmap) != ctxbitmap);
+       asids->mt_ctxbitmap ^= ctxbitmap;
+       gru_dbg(grudev, "gru %x, gms %p, ctxnum 0x%d, asidmap 0x%lx\n",
+               gru->gs_gid, gms, ctxnum, gms->ms_asidmap[0]);
+       spin_unlock(&gms->ms_asid_lock);
+}
+
+/*
+ * Decrement the reference count on a GTS structure. Free the structure
+ * if the reference count goes to zero.
+ */
+void gts_drop(struct gru_thread_state *gts)
+{
+       if (gts && atomic_dec_return(&gts->ts_refcnt) == 0) {
+               gru_drop_mmu_notifier(gts->ts_gms);
+               kfree(gts);
+               STAT(gts_free);
+       }
+}
+
+/*
+ * Locate the GTS structure for the current thread.
+ */
+static struct gru_thread_state *gru_find_current_gts_nolock(struct gru_vma_data
+                           *vdata, int tsid)
+{
+       struct gru_thread_state *gts;
+
+       list_for_each_entry(gts, &vdata->vd_head, ts_next)
+           if (gts->ts_tsid == tsid)
+               return gts;
+       return NULL;
+}
+
+/*
+ * Allocate a thread state structure.
+ */
+static struct gru_thread_state *gru_alloc_gts(struct vm_area_struct *vma,
+                                             struct gru_vma_data *vdata,
+                                             int tsid)
+{
+       struct gru_thread_state *gts;
+       int bytes;
+
+       bytes = DSR_BYTES(vdata->vd_dsr_au_count) +
+                               CBR_BYTES(vdata->vd_cbr_au_count);
+       bytes += sizeof(struct gru_thread_state);
+       gts = kzalloc(bytes, GFP_KERNEL);
+       if (!gts)
+               return NULL;
+
+       STAT(gts_alloc);
+       atomic_set(&gts->ts_refcnt, 1);
+       mutex_init(&gts->ts_ctxlock);
+       gts->ts_cbr_au_count = vdata->vd_cbr_au_count;
+       gts->ts_dsr_au_count = vdata->vd_dsr_au_count;
+       gts->ts_user_options = vdata->vd_user_options;
+       gts->ts_tsid = tsid;
+       gts->ts_user_options = vdata->vd_user_options;
+       gts->ts_ctxnum = NULLCTX;
+       gts->ts_mm = current->mm;
+       gts->ts_vma = vma;
+       gts->ts_tlb_int_select = -1;
+       gts->ts_gms = gru_register_mmu_notifier();
+       if (!gts->ts_gms)
+               goto err;
+
+       gru_dbg(grudev, "alloc vdata %p, new gts %p\n", vdata, gts);
+       return gts;
+
+err:
+       gts_drop(gts);
+       return NULL;
+}
+
+/*
+ * Allocate a vma private data structure.
+ */
+struct gru_vma_data *gru_alloc_vma_data(struct vm_area_struct *vma, int tsid)
+{
+       struct gru_vma_data *vdata = NULL;
+
+       vdata = kmalloc(sizeof(*vdata), GFP_KERNEL);
+       if (!vdata)
+               return NULL;
+
+       INIT_LIST_HEAD(&vdata->vd_head);
+       spin_lock_init(&vdata->vd_lock);
+       gru_dbg(grudev, "alloc vdata %p\n", vdata);
+       return vdata;
+}
+
+/*
+ * Find the thread state structure for the current thread.
+ */
+struct gru_thread_state *gru_find_thread_state(struct vm_area_struct *vma,
+                                       int tsid)
+{
+       struct gru_vma_data *vdata = vma->vm_private_data;
+       struct gru_thread_state *gts;
+
+       spin_lock(&vdata->vd_lock);
+       gts = gru_find_current_gts_nolock(vdata, tsid);
+       spin_unlock(&vdata->vd_lock);
+       gru_dbg(grudev, "vma %p, gts %p\n", vma, gts);
+       return gts;
+}
+
+/*
+ * Allocate a new thread state for a GSEG. Note that races may allow
+ * another thread to race to create a gts.
+ */
+struct gru_thread_state *gru_alloc_thread_state(struct vm_area_struct *vma,
+                                       int tsid)
+{
+       struct gru_vma_data *vdata = vma->vm_private_data;
+       struct gru_thread_state *gts, *ngts;
+
+       gts = gru_alloc_gts(vma, vdata, tsid);
+       if (!gts)
+               return NULL;
+
+       spin_lock(&vdata->vd_lock);
+       ngts = gru_find_current_gts_nolock(vdata, tsid);
+       if (ngts) {
+               gts_drop(gts);
+               gts = ngts;
+               STAT(gts_double_allocate);
+       } else {
+               list_add(&gts->ts_next, &vdata->vd_head);
+       }
+       spin_unlock(&vdata->vd_lock);
+       gru_dbg(grudev, "vma %p, gts %p\n", vma, gts);
+       return gts;
+}
+
+/*
+ * Free the GRU context assigned to the thread state.
+ */
+static void gru_free_gru_context(struct gru_thread_state *gts)
+{
+       struct gru_state *gru;
+
+       gru = gts->ts_gru;
+       gru_dbg(grudev, "gts %p, gru %p\n", gts, gru);
+
+       spin_lock(&gru->gs_lock);
+       gru->gs_gts[gts->ts_ctxnum] = NULL;
+       free_gru_resources(gru, gts);
+       BUG_ON(test_bit(gts->ts_ctxnum, &gru->gs_context_map) == 0);
+       __clear_bit(gts->ts_ctxnum, &gru->gs_context_map);
+       gts->ts_ctxnum = NULLCTX;
+       gts->ts_gru = NULL;
+       spin_unlock(&gru->gs_lock);
+
+       gts_drop(gts);
+       STAT(free_context);
+}
+
+/*
+ * Prefetching cachelines help hardware performance.
+ */
+static void prefetch_data(void *p, int num, int stride)
+{
+       while (num-- > 0) {
+               prefetchw(p);
+               p += stride;
+       }
+}
+
+static inline long gru_copy_handle(void *d, void *s)
+{
+       memcpy(d, s, GRU_HANDLE_BYTES);
+       return GRU_HANDLE_BYTES;
+}
+
+/* rewrite in assembly & use lots of prefetch */
+static void gru_load_context_data(void *save, void *grubase, int ctxnum,
+                                 unsigned long cbrmap, unsigned long dsrmap)
+{
+       void *gseg, *cb, *cbe;
+       unsigned long length;
+       int i, scr;
+
+       gseg = grubase + ctxnum * GRU_GSEG_STRIDE;
+       length = hweight64(dsrmap) * GRU_DSR_AU_BYTES;
+       prefetch_data(gseg + GRU_DS_BASE, length / GRU_CACHE_LINE_BYTES,
+                     GRU_CACHE_LINE_BYTES);
+
+       cb = gseg + GRU_CB_BASE;
+       cbe = grubase + GRU_CBE_BASE;
+       for_each_cbr_in_allocation_map(i, &cbrmap, scr) {
+               prefetch_data(cb, 1, GRU_CACHE_LINE_BYTES);
+               prefetch_data(cbe + i * GRU_HANDLE_STRIDE, 1,
+                             GRU_CACHE_LINE_BYTES);
+               cb += GRU_HANDLE_STRIDE;
+       }
+
+       cb = gseg + GRU_CB_BASE;
+       for_each_cbr_in_allocation_map(i, &cbrmap, scr) {
+               save += gru_copy_handle(cb, save);
+               save += gru_copy_handle(cbe + i * GRU_HANDLE_STRIDE, save);
+               cb += GRU_HANDLE_STRIDE;
+       }
+
+       memcpy(gseg + GRU_DS_BASE, save, length);
+}
+
+static void gru_unload_context_data(void *save, void *grubase, int ctxnum,
+                                   unsigned long cbrmap, unsigned long dsrmap)
+{
+       void *gseg, *cb, *cbe;
+       unsigned long length;
+       int i, scr;
+
+       gseg = grubase + ctxnum * GRU_GSEG_STRIDE;
+
+       cb = gseg + GRU_CB_BASE;
+       cbe = grubase + GRU_CBE_BASE;
+       for_each_cbr_in_allocation_map(i, &cbrmap, scr) {
+               save += gru_copy_handle(save, cb);
+               save += gru_copy_handle(save, cbe + i * GRU_HANDLE_STRIDE);
+               cb += GRU_HANDLE_STRIDE;
+       }
+       length = hweight64(dsrmap) * GRU_DSR_AU_BYTES;
+       memcpy(save, gseg + GRU_DS_BASE, length);
+}
+
+void gru_unload_context(struct gru_thread_state *gts, int savestate)
+{
+       struct gru_state *gru = gts->ts_gru;
+       struct gru_context_configuration_handle *cch;
+       int ctxnum = gts->ts_ctxnum;
+
+       zap_vma_ptes(gts->ts_vma, UGRUADDR(gts), GRU_GSEG_PAGESIZE);
+       cch = get_cch(gru->gs_gru_base_vaddr, ctxnum);
+
+       lock_cch_handle(cch);
+       if (cch_interrupt_sync(cch))
+               BUG();
+       gru_dbg(grudev, "gts %p\n", gts);
+
+       gru_unload_mm_tracker(gru, gts->ts_gms, gts->ts_ctxnum);
+       if (savestate)
+               gru_unload_context_data(gts->ts_gdata, gru->gs_gru_base_vaddr,
+                                       ctxnum, gts->ts_cbr_map,
+                                       gts->ts_dsr_map);
+
+       if (cch_deallocate(cch))
+               BUG();
+       gts->ts_force_unload = 0;       /* ts_force_unload locked by CCH lock */
+       unlock_cch_handle(cch);
+
+       gru_free_gru_context(gts);
+       STAT(unload_context);
+}
+
+/*
+ * Load a GRU context by copying it from the thread data structure in memory
+ * to the GRU.
+ */
+static void gru_load_context(struct gru_thread_state *gts)
+{
+       struct gru_state *gru = gts->ts_gru;
+       struct gru_context_configuration_handle *cch;
+       int err, asid, ctxnum = gts->ts_ctxnum;
+
+       gru_dbg(grudev, "gts %p\n", gts);
+       cch = get_cch(gru->gs_gru_base_vaddr, ctxnum);
+
+       lock_cch_handle(cch);
+       asid = gru_load_mm_tracker(gru, gts->ts_gms, gts->ts_ctxnum);
+       cch->tfm_fault_bit_enable =
+           (gts->ts_user_options == GRU_OPT_MISS_FMM_POLL
+            || gts->ts_user_options == GRU_OPT_MISS_FMM_INTR);
+       cch->tlb_int_enable = (gts->ts_user_options == GRU_OPT_MISS_FMM_INTR);
+       if (cch->tlb_int_enable) {
+               gts->ts_tlb_int_select = gru_cpu_fault_map_id();
+               cch->tlb_int_select = gts->ts_tlb_int_select;
+       }
+       cch->tfm_done_bit_enable = 0;
+       err = cch_allocate(cch, asid, gts->ts_cbr_map, gts->ts_dsr_map);
+       if (err) {
+               gru_dbg(grudev,
+                       "err %d: cch %p, gts %p, cbr 0x%lx, dsr 0x%lx\n",
+                       err, cch, gts, gts->ts_cbr_map, gts->ts_dsr_map);
+               BUG();
+       }
+
+       gru_load_context_data(gts->ts_gdata, gru->gs_gru_base_vaddr, ctxnum,
+                             gts->ts_cbr_map, gts->ts_dsr_map);
+
+       if (cch_start(cch))
+               BUG();
+       unlock_cch_handle(cch);
+
+       STAT(load_context);
+}
+
+/*
+ * Update fields in an active CCH:
+ *     - retarget interrupts on local blade
+ *     - force a delayed context unload by clearing the CCH asids. This
+ *       forces TLB misses for new GRU instructions. The context is unloaded
+ *       when the next TLB miss occurs.
+ */
+static int gru_update_cch(struct gru_thread_state *gts, int int_select)
+{
+       struct gru_context_configuration_handle *cch;
+       struct gru_state *gru = gts->ts_gru;
+       int i, ctxnum = gts->ts_ctxnum, ret = 0;
+
+       cch = get_cch(gru->gs_gru_base_vaddr, ctxnum);
+
+       lock_cch_handle(cch);
+       if (cch->state == CCHSTATE_ACTIVE) {
+               if (gru->gs_gts[gts->ts_ctxnum] != gts)
+                       goto exit;
+               if (cch_interrupt(cch))
+                       BUG();
+               if (int_select >= 0) {
+                       gts->ts_tlb_int_select = int_select;
+                       cch->tlb_int_select = int_select;
+               } else {
+                       for (i = 0; i < 8; i++)
+                               cch->asid[i] = 0;
+                       cch->tfm_fault_bit_enable = 0;
+                       cch->tlb_int_enable = 0;
+                       gts->ts_force_unload = 1;
+               }
+               if (cch_start(cch))
+                       BUG();
+               ret = 1;
+       }
+exit:
+       unlock_cch_handle(cch);
+       return ret;
+}
+
+/*
+ * Update CCH tlb interrupt select. Required when all the following is true:
+ *     - task's GRU context is loaded into a GRU
+ *     - task is using interrupt notification for TLB faults
+ *     - task has migrated to a different cpu on the same blade where
+ *       it was previously running.
+ */
+static int gru_retarget_intr(struct gru_thread_state *gts)
+{
+       if (gts->ts_tlb_int_select < 0
+           || gts->ts_tlb_int_select == gru_cpu_fault_map_id())
+               return 0;
+
+       gru_dbg(grudev, "retarget from %d to %d\n", gts->ts_tlb_int_select,
+               gru_cpu_fault_map_id());
+       return gru_update_cch(gts, gru_cpu_fault_map_id());
+}
+
+
+/*
+ * Insufficient GRU resources available on the local blade. Steal a context from
+ * a process. This is a hack until a _real_ resource scheduler is written....
+ */
+#define next_ctxnum(n) ((n) <  GRU_NUM_CCH - 2 ? (n) + 1 : 0)
+#define next_gru(b, g) (((g) < &(b)->bs_grus[GRU_CHIPLETS_PER_BLADE - 1]) ?  \
+                                ((g)+1) : &(b)->bs_grus[0])
+
+static void gru_steal_context(struct gru_thread_state *gts)
+{
+       struct gru_blade_state *blade;
+       struct gru_state *gru, *gru0;
+       struct gru_thread_state *ngts = NULL;
+       int ctxnum, ctxnum0, flag = 0, cbr, dsr;
+
+       cbr = gts->ts_cbr_au_count;
+       dsr = gts->ts_dsr_au_count;
+
+       preempt_disable();
+       blade = gru_base[uv_numa_blade_id()];
+       spin_lock(&blade->bs_lock);
+
+       ctxnum = next_ctxnum(blade->bs_lru_ctxnum);
+       gru = blade->bs_lru_gru;
+       if (ctxnum == 0)
+               gru = next_gru(blade, gru);
+       ctxnum0 = ctxnum;
+       gru0 = gru;
+       while (1) {
+               if (check_gru_resources(gru, cbr, dsr, GRU_NUM_CCH))
+                       break;
+               spin_lock(&gru->gs_lock);
+               for (; ctxnum < GRU_NUM_CCH; ctxnum++) {
+                       if (flag && gru == gru0 && ctxnum == ctxnum0)
+                               break;
+                       ngts = gru->gs_gts[ctxnum];
+                       /*
+                        * We are grabbing locks out of order, so trylock is
+                        * needed. GTSs are usually not locked, so the odds of
+                        * success are high. If trylock fails, try to steal a
+                        * different GSEG.
+                        */
+                       if (ngts && mutex_trylock(&ngts->ts_ctxlock))
+                               break;
+                       ngts = NULL;
+                       flag = 1;
+               }
+               spin_unlock(&gru->gs_lock);
+               if (ngts || (flag && gru == gru0 && ctxnum == ctxnum0))
+                       break;
+               ctxnum = 0;
+               gru = next_gru(blade, gru);
+       }
+       blade->bs_lru_gru = gru;
+       blade->bs_lru_ctxnum = ctxnum;
+       spin_unlock(&blade->bs_lock);
+       preempt_enable();
+
+       if (ngts) {
+               STAT(steal_context);
+               ngts->ts_steal_jiffies = jiffies;
+               gru_unload_context(ngts, 1);
+               mutex_unlock(&ngts->ts_ctxlock);
+       } else {
+               STAT(steal_context_failed);
+       }
+       gru_dbg(grudev,
+               "stole gru %x, ctxnum %d from gts %p. Need cb %d, ds %d;"
+               " avail cb %ld, ds %ld\n",
+               gru->gs_gid, ctxnum, ngts, cbr, dsr, hweight64(gru->gs_cbr_map),
+               hweight64(gru->gs_dsr_map));
+}
+
+/*
+ * Scan the GRUs on the local blade & assign a GRU context.
+ */
+static struct gru_state *gru_assign_gru_context(struct gru_thread_state *gts)
+{
+       struct gru_state *gru, *grux;
+       int i, max_active_contexts;
+
+       preempt_disable();
+
+again:
+       gru = NULL;
+       max_active_contexts = GRU_NUM_CCH;
+       for_each_gru_on_blade(grux, uv_numa_blade_id(), i) {
+               if (check_gru_resources(grux, gts->ts_cbr_au_count,
+                                       gts->ts_dsr_au_count,
+                                       max_active_contexts)) {
+                       gru = grux;
+                       max_active_contexts = grux->gs_active_contexts;
+                       if (max_active_contexts == 0)
+                               break;
+               }
+       }
+
+       if (gru) {
+               spin_lock(&gru->gs_lock);
+               if (!check_gru_resources(gru, gts->ts_cbr_au_count,
+                                        gts->ts_dsr_au_count, GRU_NUM_CCH)) {
+                       spin_unlock(&gru->gs_lock);
+                       goto again;
+               }
+               reserve_gru_resources(gru, gts);
+               gts->ts_gru = gru;
+               gts->ts_ctxnum =
+                   find_first_zero_bit(&gru->gs_context_map, GRU_NUM_CCH);
+               BUG_ON(gts->ts_ctxnum == GRU_NUM_CCH);
+               atomic_inc(&gts->ts_refcnt);
+               gru->gs_gts[gts->ts_ctxnum] = gts;
+               __set_bit(gts->ts_ctxnum, &gru->gs_context_map);
+               spin_unlock(&gru->gs_lock);
+
+               STAT(assign_context);
+               gru_dbg(grudev,
+                       "gseg %p, gts %p, gru %x, ctx %d, cbr %d, dsr %d\n",
+                       gseg_virtual_address(gts->ts_gru, gts->ts_ctxnum), gts,
+                       gts->ts_gru->gs_gid, gts->ts_ctxnum,
+                       gts->ts_cbr_au_count, gts->ts_dsr_au_count);
+       } else {
+               gru_dbg(grudev, "failed to allocate a GTS %s\n", "");
+               STAT(assign_context_failed);
+       }
+
+       preempt_enable();
+       return gru;
+}
+
+/*
+ * gru_nopage
+ *
+ * Map the user's GRU segment
+ */
+int gru_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
+{
+       struct gru_thread_state *gts;
+       unsigned long paddr, vaddr;
+
+       vaddr = (unsigned long)vmf->virtual_address;
+       gru_dbg(grudev, "vma %p, vaddr 0x%lx (0x%lx)\n",
+               vma, vaddr, GSEG_BASE(vaddr));
+       STAT(nopfn);
+
+       gts = gru_find_thread_state(vma, TSID(vaddr, vma));
+       if (!gts)
+               return VM_FAULT_SIGBUS;
+
+again:
+       preempt_disable();
+       mutex_lock(&gts->ts_ctxlock);
+       if (gts->ts_gru) {
+               if (gts->ts_gru->gs_blade_id != uv_numa_blade_id()) {
+                       STAT(migrated_nopfn_unload);
+                       gru_unload_context(gts, 1);
+               } else {
+                       if (gru_retarget_intr(gts))
+                               STAT(migrated_nopfn_retarget);
+               }
+       }
+
+       if (!gts->ts_gru) {
+               while (!gru_assign_gru_context(gts)) {
+                       mutex_unlock(&gts->ts_ctxlock);
+                       preempt_enable();
+                       schedule_timeout(GRU_ASSIGN_DELAY);  /* true hack ZZZ */
+                       if (gts->ts_steal_jiffies + GRU_STEAL_DELAY < jiffies)
+                               gru_steal_context(gts);
+                       goto again;
+               }
+               gru_load_context(gts);
+               paddr = gseg_physical_address(gts->ts_gru, gts->ts_ctxnum);
+               remap_pfn_range(vma, vaddr & ~(GRU_GSEG_PAGESIZE - 1),
+                               paddr >> PAGE_SHIFT, GRU_GSEG_PAGESIZE,
+                               vma->vm_page_prot);
+       }
+
+       mutex_unlock(&gts->ts_ctxlock);
+       preempt_enable();
+
+       return VM_FAULT_NOPAGE;
+}
+