arch/x86/xen/multicalls.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Xen hypercall batching.
   4  *
   5  * Xen allows multiple hypercalls to be issued at once, using the
   6  * multicall interface.  This allows the cost of trapping into the
   7  * hypervisor to be amortized over several calls.
   8  *
   9  * This file implements a simple interface for multicalls.  There's a
  10  * per-cpu buffer of outstanding multicalls.  When you want to queue a
  11  * multicall for issuing, you can allocate a multicall slot for the
  12  * call and its arguments, along with storage for space which is
  13  * pointed to by the arguments (for passing pointers to structures,
  14  * etc).  When the multicall is actually issued, all the space for the
  15  * commands and allocated memory is freed for reuse.
  16  *
  17  * Multicalls are flushed whenever any of the buffers get full, or
  18  * when explicitly requested.  There's no way to get per-multicall
  19  * return results back.  It will BUG if any of the multicalls fail.
  20  *
  21  * Jeremy Fitzhardinge <jeremy@xensource.com>, XenSource Inc, 2007
  22  */
  23 #include <linux/percpu.h>
  24 #include <linux/hardirq.h>
  25 #include <linux/debugfs.h>
  26
  27 #include <asm/xen/hypercall.h>
  28
  29 #include "multicalls.h"
  30 #include "debugfs.h"
  31
  32 #define MC_BATCH        32
  33
  34 #define MC_DEBUG        0
  35
  36 #define MC_ARGS         (MC_BATCH * 16)
  37
  38
  39 struct mc_buffer {
  40         unsigned mcidx, argidx, cbidx;
  41         struct multicall_entry entries[MC_BATCH];
  42 #if MC_DEBUG
  43         struct multicall_entry debug[MC_BATCH];
  44         void *caller[MC_BATCH];
  45 #endif
  46         unsigned char args[MC_ARGS];
  47         struct callback {
  48                 void (*fn)(void *);
  49                 void *data;
  50         } callbacks[MC_BATCH];
  51 };
  52
  53 static DEFINE_PER_CPU(struct mc_buffer, mc_buffer);
  54 DEFINE_PER_CPU(unsigned long, xen_mc_irq_flags);
  55
  56 void xen_mc_flush(void)
  57 {
  58         struct mc_buffer *b = this_cpu_ptr(&mc_buffer);
  59         struct multicall_entry *mc;
  60         int ret = 0;
  61         unsigned long flags;
  62         int i;
  63
  64         BUG_ON(preemptible());
  65
  66         /* Disable interrupts in case someone comes in and queues
  67            something in the middle */
  68         local_irq_save(flags);
  69
  70         trace_xen_mc_flush(b->mcidx, b->argidx, b->cbidx);
  71
  72         switch (b->mcidx) {
  73         case 0:
  74                 /* no-op */
  75                 BUG_ON(b->argidx != 0);
  76                 break;
  77
  78         case 1:
  79                 /* Singleton multicall - bypass multicall machinery
  80                    and just do the call directly. */
  81                 mc = &b->entries[0];
  82
  83                 mc->result = privcmd_call(mc->op,
  84                                           mc->args[0], mc->args[1], mc->args[2],
  85                                           mc->args[3], mc->args[4]);
  86                 ret = mc->result < 0;
  87                 break;
  88
  89         default:
  90 #if MC_DEBUG
  91                 memcpy(b->debug, b->entries,
  92                        b->mcidx * sizeof(struct multicall_entry));
  93 #endif
  94
  95                 if (HYPERVISOR_multicall(b->entries, b->mcidx) != 0)
  96                         BUG();
  97                 for (i = 0; i < b->mcidx; i++)
  98                         if (b->entries[i].result < 0)
  99                                 ret++;
 100
 101 #if MC_DEBUG
 102                 if (ret) {
 103                         printk(KERN_ERR "%d multicall(s) failed: cpu %d\n",
 104                                ret, smp_processor_id());
 105                         dump_stack();
 106                         for (i = 0; i < b->mcidx; i++) {
 107                                 printk(KERN_DEBUG "  call %2d/%d: op=%lu arg=[%lx] result=%ld\t%pF\n",
 108                                        i+1, b->mcidx,
 109                                        b->debug[i].op,
 110                                        b->debug[i].args[0],
 111                                        b->entries[i].result,
 112                                        b->caller[i]);
 113                         }
 114                 }
 115 #endif
 116         }
 117
 118         b->mcidx = 0;
 119         b->argidx = 0;
 120
 121         for (i = 0; i < b->cbidx; i++) {
 122                 struct callback *cb = &b->callbacks[i];
 123
 124                 (*cb->fn)(cb->data);
 125         }
 126         b->cbidx = 0;
 127
 128         local_irq_restore(flags);
 129
 130         WARN_ON(ret);
 131 }
 132
 133 struct multicall_space __xen_mc_entry(size_t args)
 134 {
 135         struct mc_buffer *b = this_cpu_ptr(&mc_buffer);
 136         struct multicall_space ret;
 137         unsigned argidx = roundup(b->argidx, sizeof(u64));
 138
 139         trace_xen_mc_entry_alloc(args);
 140
 141         BUG_ON(preemptible());
 142         BUG_ON(b->argidx >= MC_ARGS);
 143
 144         if (unlikely(b->mcidx == MC_BATCH ||
 145                      (argidx + args) >= MC_ARGS)) {
 146                 trace_xen_mc_flush_reason((b->mcidx == MC_BATCH) ?
 147                                           XEN_MC_FL_BATCH : XEN_MC_FL_ARGS);
 148                 xen_mc_flush();
 149                 argidx = roundup(b->argidx, sizeof(u64));
 150         }
 151
 152         ret.mc = &b->entries[b->mcidx];
 153 #if MC_DEBUG
 154         b->caller[b->mcidx] = __builtin_return_address(0);
 155 #endif
 156         b->mcidx++;
 157         ret.args = &b->args[argidx];
 158         b->argidx = argidx + args;
 159
 160         BUG_ON(b->argidx >= MC_ARGS);
 161         return ret;
 162 }
 163
 164 struct multicall_space xen_mc_extend_args(unsigned long op, size_t size)
 165 {
 166         struct mc_buffer *b = this_cpu_ptr(&mc_buffer);
 167         struct multicall_space ret = { NULL, NULL };
 168
 169         BUG_ON(preemptible());
 170         BUG_ON(b->argidx >= MC_ARGS);
 171
 172         if (unlikely(b->mcidx == 0 ||
 173                      b->entries[b->mcidx - 1].op != op)) {
 174                 trace_xen_mc_extend_args(op, size, XEN_MC_XE_BAD_OP);
 175                 goto out;
 176         }
 177
 178         if (unlikely((b->argidx + size) >= MC_ARGS)) {
 179                 trace_xen_mc_extend_args(op, size, XEN_MC_XE_NO_SPACE);
 180                 goto out;
 181         }
 182
 183         ret.mc = &b->entries[b->mcidx - 1];
 184         ret.args = &b->args[b->argidx];
 185         b->argidx += size;
 186
 187         BUG_ON(b->argidx >= MC_ARGS);
 188
 189         trace_xen_mc_extend_args(op, size, XEN_MC_XE_OK);
 190 out:
 191         return ret;
 192 }
 193
 194 void xen_mc_callback(void (*fn)(void *), void *data)
 195 {
 196         struct mc_buffer *b = this_cpu_ptr(&mc_buffer);
 197         struct callback *cb;
 198
 199         if (b->cbidx == MC_BATCH) {
 200                 trace_xen_mc_flush_reason(XEN_MC_FL_CALLBACK);
 201                 xen_mc_flush();
 202         }
 203
 204         trace_xen_mc_callback(fn, data);
 205
 206         cb = &b->callbacks[b->cbidx++];
 207         cb->fn = fn;
 208         cb->data = data;
 209 }