arch/x86/xen/multicalls.c

   1 /*
   2  * Xen hypercall batching.
   3  *
   4  * Xen allows multiple hypercalls to be issued at once, using the
   5  * multicall interface.  This allows the cost of trapping into the
   6  * hypervisor to be amortized over several calls.
   7  *
   8  * This file implements a simple interface for multicalls.  There's a
   9  * per-cpu buffer of outstanding multicalls.  When you want to queue a
  10  * multicall for issuing, you can allocate a multicall slot for the
  11  * call and its arguments, along with storage for space which is
  12  * pointed to by the arguments (for passing pointers to structures,
  13  * etc).  When the multicall is actually issued, all the space for the
  14  * commands and allocated memory is freed for reuse.
  15  *
  16  * Multicalls are flushed whenever any of the buffers get full, or
  17  * when explicitly requested.  There's no way to get per-multicall
  18  * return results back.  It will BUG if any of the multicalls fail.
  19  *
  20  * Jeremy Fitzhardinge <jeremy@xensource.com>, XenSource Inc, 2007
  21  */
  22 #include <linux/percpu.h>
  23 #include <linux/hardirq.h>
  24 #include <linux/debugfs.h>
  25
  26 #include <asm/xen/hypercall.h>
  27
  28 #include "multicalls.h"
  29 #include "debugfs.h"
  30
  31 #define MC_BATCH        32
  32
  33 #define MC_DEBUG        0
  34
  35 #define MC_ARGS         (MC_BATCH * 16)
  36
  37
  38 struct mc_buffer {
  39         unsigned mcidx, argidx, cbidx;
  40         struct multicall_entry entries[MC_BATCH];
  41 #if MC_DEBUG
  42         struct multicall_entry debug[MC_BATCH];
  43         void *caller[MC_BATCH];
  44 #endif
  45         unsigned char args[MC_ARGS];
  46         struct callback {
  47                 void (*fn)(void *);
  48                 void *data;
  49         } callbacks[MC_BATCH];
  50 };
  51
  52 static DEFINE_PER_CPU(struct mc_buffer, mc_buffer);
  53 DEFINE_PER_CPU(unsigned long, xen_mc_irq_flags);
  54
  55 void xen_mc_flush(void)
  56 {
  57         struct mc_buffer *b = &__get_cpu_var(mc_buffer);
  58         struct multicall_entry *mc;
  59         int ret = 0;
  60         unsigned long flags;
  61         int i;
  62
  63         BUG_ON(preemptible());
  64
  65         /* Disable interrupts in case someone comes in and queues
  66            something in the middle */
  67         local_irq_save(flags);
  68
  69         trace_xen_mc_flush(b->mcidx, b->argidx, b->cbidx);
  70
  71         switch (b->mcidx) {
  72         case 0:
  73                 /* no-op */
  74                 BUG_ON(b->argidx != 0);
  75                 break;
  76
  77         case 1:
  78                 /* Singleton multicall - bypass multicall machinery
  79                    and just do the call directly. */
  80                 mc = &b->entries[0];
  81
  82                 mc->result = privcmd_call(mc->op,
  83                                           mc->args[0], mc->args[1], mc->args[2],
  84                                           mc->args[3], mc->args[4]);
  85                 ret = mc->result < 0;
  86                 break;
  87
  88         default:
  89 #if MC_DEBUG
  90                 memcpy(b->debug, b->entries,
  91                        b->mcidx * sizeof(struct multicall_entry));
  92 #endif
  93
  94                 if (HYPERVISOR_multicall(b->entries, b->mcidx) != 0)
  95                         BUG();
  96                 for (i = 0; i < b->mcidx; i++)
  97                         if (b->entries[i].result < 0)
  98                                 ret++;
  99
 100 #if MC_DEBUG
 101                 if (ret) {
 102                         printk(KERN_ERR "%d multicall(s) failed: cpu %d\n",
 103                                ret, smp_processor_id());
 104                         dump_stack();
 105                         for (i = 0; i < b->mcidx; i++) {
 106                                 printk(KERN_DEBUG "  call %2d/%d: op=%lu arg=[%lx] result=%ld\t%pF\n",
 107                                        i+1, b->mcidx,
 108                                        b->debug[i].op,
 109                                        b->debug[i].args[0],
 110                                        b->entries[i].result,
 111                                        b->caller[i]);
 112                         }
 113                 }
 114 #endif
 115         }
 116
 117         b->mcidx = 0;
 118         b->argidx = 0;
 119
 120         for (i = 0; i < b->cbidx; i++) {
 121                 struct callback *cb = &b->callbacks[i];
 122
 123                 (*cb->fn)(cb->data);
 124         }
 125         b->cbidx = 0;
 126
 127         local_irq_restore(flags);
 128
 129         WARN_ON(ret);
 130 }
 131
 132 struct multicall_space __xen_mc_entry(size_t args)
 133 {
 134         struct mc_buffer *b = &__get_cpu_var(mc_buffer);
 135         struct multicall_space ret;
 136         unsigned argidx = roundup(b->argidx, sizeof(u64));
 137
 138         trace_xen_mc_entry_alloc(args);
 139
 140         BUG_ON(preemptible());
 141         BUG_ON(b->argidx >= MC_ARGS);
 142
 143         if (unlikely(b->mcidx == MC_BATCH ||
 144                      (argidx + args) >= MC_ARGS)) {
 145                 trace_xen_mc_flush_reason((b->mcidx == MC_BATCH) ?
 146                                           XEN_MC_FL_BATCH : XEN_MC_FL_ARGS);
 147                 xen_mc_flush();
 148                 argidx = roundup(b->argidx, sizeof(u64));
 149         }
 150
 151         ret.mc = &b->entries[b->mcidx];
 152 #if MC_DEBUG
 153         b->caller[b->mcidx] = __builtin_return_address(0);
 154 #endif
 155         b->mcidx++;
 156         ret.args = &b->args[argidx];
 157         b->argidx = argidx + args;
 158
 159         BUG_ON(b->argidx >= MC_ARGS);
 160         return ret;
 161 }
 162
 163 struct multicall_space xen_mc_extend_args(unsigned long op, size_t size)
 164 {
 165         struct mc_buffer *b = &__get_cpu_var(mc_buffer);
 166         struct multicall_space ret = { NULL, NULL };
 167
 168         BUG_ON(preemptible());
 169         BUG_ON(b->argidx >= MC_ARGS);
 170
 171         if (unlikely(b->mcidx == 0 ||
 172                      b->entries[b->mcidx - 1].op != op)) {
 173                 trace_xen_mc_extend_args(op, size, XEN_MC_XE_BAD_OP);
 174                 goto out;
 175         }
 176
 177         if (unlikely((b->argidx + size) >= MC_ARGS)) {
 178                 trace_xen_mc_extend_args(op, size, XEN_MC_XE_NO_SPACE);
 179                 goto out;
 180         }
 181
 182         ret.mc = &b->entries[b->mcidx - 1];
 183         ret.args = &b->args[b->argidx];
 184         b->argidx += size;
 185
 186         BUG_ON(b->argidx >= MC_ARGS);
 187
 188         trace_xen_mc_extend_args(op, size, XEN_MC_XE_OK);
 189 out:
 190         return ret;
 191 }
 192
 193 void xen_mc_callback(void (*fn)(void *), void *data)
 194 {
 195         struct mc_buffer *b = &__get_cpu_var(mc_buffer);
 196         struct callback *cb;
 197
 198         if (b->cbidx == MC_BATCH) {
 199                 trace_xen_mc_flush_reason(XEN_MC_FL_CALLBACK);
 200                 xen_mc_flush();
 201         }
 202
 203         trace_xen_mc_callback(fn, data);
 204
 205         cb = &b->callbacks[b->cbidx++];
 206         cb->fn = fn;
 207         cb->data = data;
 208 }