arch/powerpc/mm/mmu_context_hash64.c

   1 /*
   2  *  MMU context allocation for 64-bit kernels.
   3  *
   4  *  Copyright (C) 2004 Anton Blanchard, IBM Corp. <anton@samba.org>
   5  *
   6  *  This program is free software; you can redistribute it and/or
   7  *  modify it under the terms of the GNU General Public License
   8  *  as published by the Free Software Foundation; either version
   9  *  2 of the License, or (at your option) any later version.
  10  *
  11  */
  12
  13 #include <linux/sched.h>
  14 #include <linux/kernel.h>
  15 #include <linux/errno.h>
  16 #include <linux/string.h>
  17 #include <linux/types.h>
  18 #include <linux/mm.h>
  19 #include <linux/spinlock.h>
  20 #include <linux/idr.h>
  21 #include <linux/export.h>
  22 #include <linux/gfp.h>
  23 #include <linux/slab.h>
  24
  25 #include <asm/mmu_context.h>
  26
  27 #ifdef CONFIG_PPC_ICSWX
  28 /*
  29  * The processor and its L2 cache cause the icswx instruction to
  30  * generate a COP_REQ transaction on PowerBus. The transaction has
  31  * no address, and the processor does not perform an MMU access
  32  * to authenticate the transaction. The command portion of the
  33  * PowerBus COP_REQ transaction includes the LPAR_ID (LPID) and
  34  * the coprocessor Process ID (PID), which the coprocessor compares
  35  * to the authorized LPID and PID held in the coprocessor, to determine
  36  * if the process is authorized to generate the transaction.
  37  * The data of the COP_REQ transaction is 128-byte or less and is
  38  * placed in cacheable memory on a 128-byte cache line boundary.
  39  *
  40  * The task to use a coprocessor should use use_cop() to allocate
  41  * a coprocessor PID before executing icswx instruction. use_cop()
  42  * also enables the coprocessor context switching. Drop_cop() is
  43  * used to free the coprocessor PID.
  44  *
  45  * Example:
  46  * Host Fabric Interface (HFI) is a PowerPC network coprocessor.
  47  * Each HFI have multiple windows. Each HFI window serves as a
  48  * network device sending to and receiving from HFI network.
  49  * HFI immediate send function uses icswx instruction. The immediate
  50  * send function allows small (single cache-line) packets be sent
  51  * without using the regular HFI send FIFO and doorbell, which are
  52  * much slower than immediate send.
  53  *
  54  * For each task intending to use HFI immediate send, the HFI driver
  55  * calls use_cop() to obtain a coprocessor PID for the task.
  56  * The HFI driver then allocate a free HFI window and save the
  57  * coprocessor PID to the HFI window to allow the task to use the
  58  * HFI window.
  59  *
  60  * The HFI driver repeatedly creates immediate send packets and
  61  * issues icswx instruction to send data through the HFI window.
  62  * The HFI compares the coprocessor PID in the CPU PID register
  63  * to the PID held in the HFI window to determine if the transaction
  64  * is allowed.
  65  *
  66  * When the task to release the HFI window, the HFI driver calls
  67  * drop_cop() to release the coprocessor PID.
  68  */
  69
  70 #define COP_PID_NONE 0
  71 #define COP_PID_MIN (COP_PID_NONE + 1)
  72 #define COP_PID_MAX (0xFFFF)
  73
  74 static DEFINE_SPINLOCK(mmu_context_acop_lock);
  75 static DEFINE_IDA(cop_ida);
  76
  77 void switch_cop(struct mm_struct *next)
  78 {
  79         mtspr(SPRN_PID, next->context.cop_pid);
  80         mtspr(SPRN_ACOP, next->context.acop);
  81 }
  82
  83 static int new_cop_pid(struct ida *ida, int min_id, int max_id,
  84                        spinlock_t *lock)
  85 {
  86         int index;
  87         int err;
  88
  89 again:
  90         if (!ida_pre_get(ida, GFP_KERNEL))
  91                 return -ENOMEM;
  92
  93         spin_lock(lock);
  94         err = ida_get_new_above(ida, min_id, &index);
  95         spin_unlock(lock);
  96
  97         if (err == -EAGAIN)
  98                 goto again;
  99         else if (err)
 100                 return err;
 101
 102         if (index > max_id) {
 103                 spin_lock(lock);
 104                 ida_remove(ida, index);
 105                 spin_unlock(lock);
 106                 return -ENOMEM;
 107         }
 108
 109         return index;
 110 }
 111
 112 static void sync_cop(void *arg)
 113 {
 114         struct mm_struct *mm = arg;
 115
 116         if (mm == current->active_mm)
 117                 switch_cop(current->active_mm);
 118 }
 119
 120 /**
 121  * Start using a coprocessor.
 122  * @acop: mask of coprocessor to be used.
 123  * @mm: The mm the coprocessor to associate with. Most likely current mm.
 124  *
 125  * Return a positive PID if successful. Negative errno otherwise.
 126  * The returned PID will be fed to the coprocessor to determine if an
 127  * icswx transaction is authenticated.
 128  */
 129 int use_cop(unsigned long acop, struct mm_struct *mm)
 130 {
 131         int ret;
 132
 133         if (!cpu_has_feature(CPU_FTR_ICSWX))
 134                 return -ENODEV;
 135
 136         if (!mm || !acop)
 137                 return -EINVAL;
 138
 139         /* The page_table_lock ensures mm_users won't change under us */
 140         spin_lock(&mm->page_table_lock);
 141         spin_lock(mm->context.cop_lockp);
 142
 143         if (mm->context.cop_pid == COP_PID_NONE) {
 144                 ret = new_cop_pid(&cop_ida, COP_PID_MIN, COP_PID_MAX,
 145                                   &mmu_context_acop_lock);
 146                 if (ret < 0)
 147                         goto out;
 148
 149                 mm->context.cop_pid = ret;
 150         }
 151         mm->context.acop |= acop;
 152
 153         sync_cop(mm);
 154
 155         /*
 156          * If this is a threaded process then there might be other threads
 157          * running. We need to send an IPI to force them to pick up any
 158          * change in PID and ACOP.
 159          */
 160         if (atomic_read(&mm->mm_users) > 1)
 161                 smp_call_function(sync_cop, mm, 1);
 162
 163         ret = mm->context.cop_pid;
 164
 165 out:
 166         spin_unlock(mm->context.cop_lockp);
 167         spin_unlock(&mm->page_table_lock);
 168
 169         return ret;
 170 }
 171 EXPORT_SYMBOL_GPL(use_cop);
 172
 173 /**
 174  * Stop using a coprocessor.
 175  * @acop: mask of coprocessor to be stopped.
 176  * @mm: The mm the coprocessor associated with.
 177  */
 178 void drop_cop(unsigned long acop, struct mm_struct *mm)
 179 {
 180         int free_pid = COP_PID_NONE;
 181
 182         if (!cpu_has_feature(CPU_FTR_ICSWX))
 183                 return;
 184
 185         if (WARN_ON_ONCE(!mm))
 186                 return;
 187
 188         /* The page_table_lock ensures mm_users won't change under us */
 189         spin_lock(&mm->page_table_lock);
 190         spin_lock(mm->context.cop_lockp);
 191
 192         mm->context.acop &= ~acop;
 193
 194         if ((!mm->context.acop) && (mm->context.cop_pid != COP_PID_NONE)) {
 195                 free_pid = mm->context.cop_pid;
 196                 mm->context.cop_pid = COP_PID_NONE;
 197         }
 198
 199         sync_cop(mm);
 200
 201         /*
 202          * If this is a threaded process then there might be other threads
 203          * running. We need to send an IPI to force them to pick up any
 204          * change in PID and ACOP.
 205          */
 206         if (atomic_read(&mm->mm_users) > 1)
 207                 smp_call_function(sync_cop, mm, 1);
 208
 209         if (free_pid != COP_PID_NONE) {
 210                 spin_lock(&mmu_context_acop_lock);
 211                 ida_remove(&cop_ida, free_pid);
 212                 spin_unlock(&mmu_context_acop_lock);
 213         }
 214
 215         spin_unlock(mm->context.cop_lockp);
 216         spin_unlock(&mm->page_table_lock);
 217 }
 218 EXPORT_SYMBOL_GPL(drop_cop);
 219
 220 #endif /* CONFIG_PPC_ICSWX */
 221
 222 static DEFINE_SPINLOCK(mmu_context_lock);
 223 static DEFINE_IDA(mmu_context_ida);
 224
 225 /*
 226  * The proto-VSID space has 2^35 - 1 segments available for user mappings.
 227  * Each segment contains 2^28 bytes.  Each context maps 2^44 bytes,
 228  * so we can support 2^19-1 contexts (19 == 35 + 28 - 44).
 229  */
 230 #define MAX_CONTEXT     ((1UL << 19) - 1)
 231
 232 int __init_new_context(void)
 233 {
 234         int index;
 235         int err;
 236
 237 again:
 238         if (!ida_pre_get(&mmu_context_ida, GFP_KERNEL))
 239                 return -ENOMEM;
 240
 241         spin_lock(&mmu_context_lock);
 242         err = ida_get_new_above(&mmu_context_ida, 1, &index);
 243         spin_unlock(&mmu_context_lock);
 244
 245         if (err == -EAGAIN)
 246                 goto again;
 247         else if (err)
 248                 return err;
 249
 250         if (index > MAX_CONTEXT) {
 251                 spin_lock(&mmu_context_lock);
 252                 ida_remove(&mmu_context_ida, index);
 253                 spin_unlock(&mmu_context_lock);
 254                 return -ENOMEM;
 255         }
 256
 257         return index;
 258 }
 259 EXPORT_SYMBOL_GPL(__init_new_context);
 260
 261 int init_new_context(struct task_struct *tsk, struct mm_struct *mm)
 262 {
 263         int index;
 264
 265         index = __init_new_context();
 266         if (index < 0)
 267                 return index;
 268
 269         /* The old code would re-promote on fork, we don't do that
 270          * when using slices as it could cause problem promoting slices
 271          * that have been forced down to 4K
 272          */
 273         if (slice_mm_new_context(mm))
 274                 slice_set_user_psize(mm, mmu_virtual_psize);
 275         subpage_prot_init_new_context(mm);
 276         mm->context.id = index;
 277 #ifdef CONFIG_PPC_ICSWX
 278         mm->context.cop_lockp = kmalloc(sizeof(spinlock_t), GFP_KERNEL);
 279         if (!mm->context.cop_lockp) {
 280                 __destroy_context(index);
 281                 subpage_prot_free(mm);
 282                 mm->context.id = MMU_NO_CONTEXT;
 283                 return -ENOMEM;
 284         }
 285         spin_lock_init(mm->context.cop_lockp);
 286 #endif /* CONFIG_PPC_ICSWX */
 287
 288         return 0;
 289 }
 290
 291 void __destroy_context(int context_id)
 292 {
 293         spin_lock(&mmu_context_lock);
 294         ida_remove(&mmu_context_ida, context_id);
 295         spin_unlock(&mmu_context_lock);
 296 }
 297 EXPORT_SYMBOL_GPL(__destroy_context);
 298
 299 void destroy_context(struct mm_struct *mm)
 300 {
 301 #ifdef CONFIG_PPC_ICSWX
 302         drop_cop(mm->context.acop, mm);
 303         kfree(mm->context.cop_lockp);
 304         mm->context.cop_lockp = NULL;
 305 #endif /* CONFIG_PPC_ICSWX */
 306         __destroy_context(mm->context.id);
 307         subpage_prot_free(mm);
 308         mm->context.id = MMU_NO_CONTEXT;
 309 }