bsd/dev/dtrace/dtrace_ptss.c

   1 /*
   2  * Copyright (c) 2000-2006 Apple Computer, Inc. All rights reserved.
   3  *
   4  * @APPLE_OSREFERENCE_LICENSE_HEADER_START@
   5  *
   6  * This file contains Original Code and/or Modifications of Original Code
   7  * as defined in and that are subject to the Apple Public Source License
   8  * Version 2.0 (the 'License'). You may not use this file except in
   9  * compliance with the License. The rights granted to you under the License
  10  * may not be used to create, or enable the creation or redistribution of,
  11  * unlawful or unlicensed copies of an Apple operating system, or to
  12  * circumvent, violate, or enable the circumvention or violation of, any
  13  * terms of an Apple operating system software license agreement.
  14  *
  15  * Please obtain a copy of the License at
  16  * http://www.opensource.apple.com/apsl/ and read it before using this file.
  17  *
  18  * The Original Code and all software distributed under the License are
  19  * distributed on an 'AS IS' basis, WITHOUT WARRANTY OF ANY KIND, EITHER
  20  * EXPRESS OR IMPLIED, AND APPLE HEREBY DISCLAIMS ALL SUCH WARRANTIES,
  21  * INCLUDING WITHOUT LIMITATION, ANY WARRANTIES OF MERCHANTABILITY,
  22  * FITNESS FOR A PARTICULAR PURPOSE, QUIET ENJOYMENT OR NON-INFRINGEMENT.
  23  * Please see the License for the specific language governing rights and
  24  * limitations under the License.
  25  *
  26  * @APPLE_OSREFERENCE_LICENSE_HEADER_END@
  27  */
  28
  29 #include <sys/types.h>
  30 #include <sys/proc.h>
  31 #include <sys/proc_internal.h>
  32 #include <sys/systm.h>
  33 #include <sys/user.h>
  34 #include <sys/dtrace_ptss.h>
  35
  36 #include <mach/vm_map.h>
  37 #include <mach/vm_param.h>
  38 #include <mach/mach_vm.h>
  39
  40 #include <kern/task.h>
  41
  42 #include <vm/vm_map.h>
  43
  44 /*
  45  * This function requires the sprlock to be held
  46  *
  47  * In general, it will not block. If it needs to allocate a new
  48  * page of memory, the underlying kernel _MALLOC may block.
  49  */
  50 struct dtrace_ptss_page_entry*
  51 dtrace_ptss_claim_entry_locked(struct proc* p)
  52 {
  53         LCK_MTX_ASSERT(&p->p_dtrace_sprlock, LCK_MTX_ASSERT_OWNED);
  54
  55         struct dtrace_ptss_page_entry* entry = NULL;
  56
  57         while (TRUE) {
  58                 struct dtrace_ptss_page_entry* temp = p->p_dtrace_ptss_free_list;
  59
  60                 if (temp == NULL) {
  61                         // Nothing on the free list. Allocate a new page, its okay if multiple threads race here.
  62                         struct dtrace_ptss_page* page = dtrace_ptss_allocate_page(p);
  63
  64                         // Make sure we actually got a page
  65                         if (page == NULL) {
  66                                 return NULL;
  67                         }
  68
  69                         // Add the page to the page list
  70                         page->next = p->p_dtrace_ptss_pages;
  71                         p->p_dtrace_ptss_pages = page;
  72
  73                         // CAS the entries onto the free list.
  74                         do {
  75                                 page->entries[DTRACE_PTSS_ENTRIES_PER_PAGE - 1].next = p->p_dtrace_ptss_free_list;
  76                         } while (!OSCompareAndSwapPtr((void *)page->entries[DTRACE_PTSS_ENTRIES_PER_PAGE - 1].next,
  77                             (void *)&page->entries[0],
  78                             (void * volatile *)&p->p_dtrace_ptss_free_list));
  79
  80                         // Now that we've added to the free list, try again.
  81                         continue;
  82                 }
  83
  84                 // Claim temp
  85                 if (!OSCompareAndSwapPtr((void *)temp, (void *)temp->next, (void * volatile *)&p->p_dtrace_ptss_free_list)) {
  86                         continue;
  87                 }
  88
  89                 // At this point, we own temp.
  90                 entry = temp;
  91
  92                 break;
  93         }
  94
  95         return entry;
  96 }
  97
  98 /*
  99  * This function does not require any locks to be held on entry.
 100  */
 101 struct dtrace_ptss_page_entry*
 102 dtrace_ptss_claim_entry(struct proc* p)
 103 {
 104         // Verify no locks held on entry
 105         LCK_MTX_ASSERT(&p->p_dtrace_sprlock, LCK_MTX_ASSERT_NOTOWNED);
 106         LCK_MTX_ASSERT(&p->p_mlock, LCK_MTX_ASSERT_NOTOWNED);
 107
 108         struct dtrace_ptss_page_entry* entry = NULL;
 109
 110         while (TRUE) {
 111                 struct dtrace_ptss_page_entry* temp = p->p_dtrace_ptss_free_list;
 112
 113                 if (temp == NULL) {
 114                         lck_mtx_lock(&p->p_dtrace_sprlock);
 115                         temp = dtrace_ptss_claim_entry_locked(p);
 116                         lck_mtx_unlock(&p->p_dtrace_sprlock);
 117                         return temp;
 118                 }
 119
 120                 // Claim temp
 121                 if (!OSCompareAndSwapPtr((void *)temp, (void *)temp->next, (void * volatile *)&p->p_dtrace_ptss_free_list)) {
 122                         continue;
 123                 }
 124
 125                 // At this point, we own temp.
 126                 entry = temp;
 127
 128                 break;
 129         }
 130
 131         return entry;
 132 }
 133
 134 /*
 135  * This function does not require any locks to be held on entry.
 136  *
 137  * (PR-11138709) A NULL p->p_dtrace_ptss_pages means the entry can
 138  * no longer be referenced safely. When found in this state, the chore
 139  * of releasing an entry to the free list is ignored.
 140  */
 141 void
 142 dtrace_ptss_release_entry(struct proc* p, struct dtrace_ptss_page_entry* e)
 143 {
 144         if (p && p->p_dtrace_ptss_pages && e) {
 145                 do {
 146                         e->next = p->p_dtrace_ptss_free_list;
 147                 } while (!OSCompareAndSwapPtr((void *)e->next, (void *)e, (void * volatile *)&p->p_dtrace_ptss_free_list));
 148         }
 149 }
 150
 151 /*
 152  * This function allocates a new page in the target process's address space.
 153  *
 154  * It returns a dtrace_ptss_page that has its entries chained, with the last
 155  * entries next field set to NULL. It does not add the page or the entries to
 156  * the process's page/entry lists.
 157  *
 158  * This function does not require that any locks be held when it is invoked.
 159  */
 160 struct dtrace_ptss_page*
 161 dtrace_ptss_allocate_page(struct proc* p)
 162 {
 163         // Allocate the kernel side data
 164         struct dtrace_ptss_page* ptss_page = _MALLOC(sizeof(struct dtrace_ptss_page), M_TEMP, M_ZERO | M_WAITOK);
 165         if (ptss_page == NULL) {
 166                 return NULL;
 167         }
 168
 169         // Now allocate a page in user space and set its protections to allow execute.
 170         task_t task = p->task;
 171         vm_map_t map = get_task_map_reference(task);
 172         if (map == NULL) {
 173                 goto err;
 174         }
 175
 176         mach_vm_size_t size = PAGE_MAX_SIZE;
 177         mach_vm_offset_t addr = 0;
 178         mach_vm_offset_t write_addr = 0;
 179         /*
 180          * The embedded OS has extra permissions for writable and executable pages.
 181          * To ensure correct permissions, we must set the page protections separately.
 182          */
 183         vm_prot_t cur_protection = VM_PROT_READ | VM_PROT_EXECUTE;
 184         vm_prot_t max_protection = VM_PROT_READ | VM_PROT_EXECUTE | VM_PROT_WRITE;
 185
 186         kern_return_t kr = mach_vm_map_kernel(map, &addr, size, 0, VM_FLAGS_ANYWHERE, VM_MAP_KERNEL_FLAGS_NONE, VM_KERN_MEMORY_NONE, IPC_PORT_NULL, 0, FALSE, cur_protection, max_protection, VM_INHERIT_DEFAULT);
 187         if (kr != KERN_SUCCESS) {
 188                 goto err;
 189         }
 190         /*
 191          * If on embedded, remap the scratch space as writable at another
 192          * virtual address
 193          */
 194         kr = mach_vm_remap_kernel(map, &write_addr, size, 0, VM_FLAGS_ANYWHERE, VM_KERN_MEMORY_NONE, map, addr, FALSE, &cur_protection, &max_protection, VM_INHERIT_DEFAULT);
 195         if (kr != KERN_SUCCESS || !(max_protection & VM_PROT_WRITE)) {
 196                 goto err;
 197         }
 198
 199         kr = mach_vm_protect(map, (mach_vm_offset_t)write_addr, (mach_vm_size_t)size, 0, VM_PROT_READ | VM_PROT_WRITE);
 200         if (kr != KERN_SUCCESS) {
 201                 goto err;
 202         }
 203
 204         // Chain the page entries.
 205         int i;
 206         for (i = 0; i < DTRACE_PTSS_ENTRIES_PER_PAGE; i++) {
 207                 ptss_page->entries[i].addr = addr + (i * DTRACE_PTSS_SCRATCH_SPACE_PER_THREAD);
 208                 ptss_page->entries[i].write_addr = write_addr + (i * DTRACE_PTSS_SCRATCH_SPACE_PER_THREAD);
 209                 ptss_page->entries[i].next = &ptss_page->entries[i + 1];
 210         }
 211
 212         // The last entry should point to NULL
 213         ptss_page->entries[DTRACE_PTSS_ENTRIES_PER_PAGE - 1].next = NULL;
 214
 215         vm_map_deallocate(map);
 216
 217         return ptss_page;
 218
 219 err:
 220         _FREE(ptss_page, M_TEMP);
 221
 222         if (map) {
 223                 vm_map_deallocate(map);
 224         }
 225
 226         return NULL;
 227 }
 228
 229 /*
 230  * This function frees an existing page in the target process's address space.
 231  *
 232  * It does not alter any of the process's page/entry lists.
 233  *
 234  * TODO: Inline in dtrace_ptrace_exec_exit?
 235  */
 236 void
 237 dtrace_ptss_free_page(struct proc* p, struct dtrace_ptss_page* ptss_page)
 238 {
 239         // Grab the task and get a reference to its vm_map
 240         task_t task = p->task;
 241         vm_map_t map = get_task_map_reference(task);
 242
 243         mach_vm_address_t addr = ptss_page->entries[0].addr;
 244         mach_vm_size_t size = PAGE_SIZE; // We need some way to assert that this matches vm_map_round_page() !!!
 245
 246         // Silent failures, no point in checking return code.
 247         mach_vm_deallocate(map, addr, size);
 248
 249         mach_vm_address_t write_addr = ptss_page->entries[0].write_addr;
 250         mach_vm_deallocate(map, write_addr, size);
 251
 252         vm_map_deallocate(map);
 253 }
 254
 255 /*
 256  * This function assumes that the target process has been
 257  * suspended, and the proc_lock & sprlock is held
 258  */
 259 void
 260 dtrace_ptss_enable(struct proc* p)
 261 {
 262         LCK_MTX_ASSERT(&p->p_dtrace_sprlock, LCK_MTX_ASSERT_OWNED);
 263         LCK_MTX_ASSERT(&p->p_mlock, LCK_MTX_ASSERT_OWNED);
 264
 265         struct uthread* uth;
 266         /*
 267          * XXX There has been a concern raised about holding the proc_lock
 268          * while calling dtrace_ptss_claim_entry(), due to the fact
 269          * that dtrace_ptss_claim_entry() can potentially malloc.
 270          */
 271         TAILQ_FOREACH(uth, &p->p_uthlist, uu_list) {
 272                 uth->t_dtrace_scratch = dtrace_ptss_claim_entry_locked(p);
 273         }
 274 }
 275
 276 /*
 277  * This function is not thread safe.
 278  *
 279  * It assumes the sprlock is held, and the proc_lock is not.
 280  */
 281 void
 282 dtrace_ptss_exec_exit(struct proc* p)
 283 {
 284         /*
 285          * Should hold sprlock to touch the pages list. Must not
 286          * hold the proc lock to avoid deadlock.
 287          */
 288         LCK_MTX_ASSERT(&p->p_dtrace_sprlock, LCK_MTX_ASSERT_OWNED);
 289         LCK_MTX_ASSERT(&p->p_mlock, LCK_MTX_ASSERT_NOTOWNED);
 290
 291         p->p_dtrace_ptss_free_list = NULL;
 292
 293         struct dtrace_ptss_page* temp = p->p_dtrace_ptss_pages;
 294         p->p_dtrace_ptss_pages = NULL;
 295
 296         while (temp != NULL) {
 297                 struct dtrace_ptss_page* next = temp->next;
 298
 299                 // Do we need to specifically mach_vm_deallocate the user pages?
 300                 // This can be called when the process is exiting, I believe the proc's
 301                 // vm_map_t may already be toast.
 302
 303                 // Must be certain to free the kernel memory!
 304                 _FREE(temp, M_TEMP);
 305                 temp = next;
 306         }
 307 }
 308
 309 /*
 310  * This function is not thread safe. It is not used for vfork.
 311  *
 312  * The child proc ptss fields are initialized to NULL at fork time.
 313  * Pages allocated in the parent are copied as part of the vm_map copy, though.
 314  * We need to deallocate those pages.
 315  *
 316  * Parent and child sprlock should be held, and proc_lock must NOT be held.
 317  */
 318 void
 319 dtrace_ptss_fork(struct proc* parent, struct proc* child)
 320 {
 321         // The child should not have any pages/entries allocated at this point.
 322         // ASSERT(child->p_dtrace_ptss_pages == NULL);
 323         // ASSERT(child->p_dtrace_ptss_free_list == NULL);
 324
 325         /*
 326          * The parent's sprlock should be held, to protect its pages list
 327          * from changing while the child references it. The child's sprlock
 328          * must also be held, because we are modifying its pages list.
 329          * Finally, to prevent a deadlock with the fasttrap cleanup code,
 330          * neither the parent or child proc_lock should be held.
 331          */
 332         LCK_MTX_ASSERT(&parent->p_dtrace_sprlock, LCK_MTX_ASSERT_OWNED);
 333         LCK_MTX_ASSERT(&parent->p_mlock, LCK_MTX_ASSERT_NOTOWNED);
 334         LCK_MTX_ASSERT(&child->p_dtrace_sprlock, LCK_MTX_ASSERT_OWNED);
 335         LCK_MTX_ASSERT(&child->p_mlock, LCK_MTX_ASSERT_NOTOWNED);
 336
 337         // Get page list from *PARENT*
 338         struct dtrace_ptss_page* temp = parent->p_dtrace_ptss_pages;
 339
 340         while (temp != NULL) {
 341                 // Freeing the page in the *CHILD*
 342                 dtrace_ptss_free_page(child, temp);
 343
 344                 // Do not free the kernel memory, it belong to the parent.
 345                 temp = temp->next;
 346         }
 347 }