Fix bug in `expr_clone`
[platform/upstream/ltrace.git] / proc.c
diff --git a/proc.c b/proc.c
index bfc6e41..bf2a172 100644 (file)
--- a/proc.c
+++ b/proc.c
+/*
+ * This file is part of ltrace.
+ * Copyright (C) 2011,2012,2013,2014 Petr Machata, Red Hat Inc.
+ * Copyright (C) 2010 Joe Damato
+ * Copyright (C) 1998,2009 Juan Cespedes
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU General Public License as
+ * published by the Free Software Foundation; either version 2 of the
+ * License, or (at your option) any later version.
+ *
+ * This program is distributed in the hope that it will be useful, but
+ * WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * General Public License for more details.
+ *
+ * You should have received a copy of the GNU General Public License
+ * along with this program; if not, write to the Free Software
+ * Foundation, Inc., 51 Franklin St, Fifth Floor, Boston, MA
+ * 02110-1301 USA
+ */
+
+#include "config.h"
+
 #include <sys/types.h>
-#include <string.h>
-#include <stdio.h>
+#include <assert.h>
 #include <errno.h>
+#include <stdio.h>
 #include <stdlib.h>
+#include <string.h>
+
+#include "backend.h"
+#include "breakpoint.h"
+#include "debug.h"
+#include "fetch.h"
+#include "options.h"
+#include "proc.h"
+#include "value_dict.h"
+
+#if defined(HAVE_LIBDW)
+# include "dwarf_prototypes.h"
+#endif /* defined(HAVE_LIBDW) */
+
+#ifndef OS_HAVE_PROCESS_DATA
+int
+os_process_init(struct process *proc)
+{
+       return 0;
+}
+
+void
+os_process_destroy(struct process *proc)
+{
+}
+
+int
+os_process_clone(struct process *retp, struct process *proc)
+{
+       return 0;
+}
+
+int
+os_process_exec(struct process *proc)
+{
+       return 0;
+}
+#endif
+
+#ifndef ARCH_HAVE_PROCESS_DATA
+int
+arch_process_init(struct process *proc)
+{
+       return 0;
+}
+
+void
+arch_process_destroy(struct process *proc)
+{
+}
+
+int
+arch_process_clone(struct process *retp, struct process *proc)
+{
+       return 0;
+}
+
+int
+arch_process_exec(struct process *proc)
+{
+       return 0;
+}
+#endif
+
+#ifndef ARCH_HAVE_DYNLINK_DONE
+void
+arch_dynlink_done(struct process *proc)
+{
+}
+#endif
+
+static int add_process(struct process *proc, int was_exec);
+static void unlist_process(struct process *proc);
+
+static void
+destroy_unwind(struct process *proc)
+{
+#if defined(HAVE_LIBUNWIND)
+       if (proc->unwind_priv != NULL)
+               _UPT_destroy(proc->unwind_priv);
+       if (proc->unwind_as != NULL)
+               unw_destroy_addr_space(proc->unwind_as);
+#endif /* defined(HAVE_LIBUNWIND) */
+
+#if defined(HAVE_LIBDW)
+       if (proc->dwfl != NULL)
+               dwfl_end(proc->dwfl);
+#endif /* defined(HAVE_LIBDW) */
+}
+
+static int
+process_bare_init(struct process *proc, const char *filename,
+                 pid_t pid, int was_exec)
+{
+       if (!was_exec) {
+               memset(proc, 0, sizeof(*proc));
+
+               proc->filename = strdup(filename);
+               if (proc->filename == NULL) {
+               fail:
+                       free(proc->filename);
+                       if (proc->breakpoints != NULL) {
+                               dict_destroy(proc->breakpoints,
+                                            NULL, NULL, NULL);
+                               free(proc->breakpoints);
+                               proc->breakpoints = NULL;
+                       }
+                       return -1;
+               }
+       }
+
+       /* Add process so that we know who the leader is.  */
+       proc->pid = pid;
+       if (add_process(proc, was_exec) < 0)
+               goto fail;
+       if (proc->leader == NULL) {
+       unlist_and_fail:
+               if (!was_exec)
+                       unlist_process(proc);
+               goto fail;
+       }
+
+       if (proc->leader == proc) {
+               proc->breakpoints = malloc(sizeof(*proc->breakpoints));
+               if (proc->breakpoints == NULL)
+                       goto unlist_and_fail;
+               DICT_INIT(proc->breakpoints,
+                         arch_addr_t, struct breakpoint *,
+                         arch_addr_hash, arch_addr_eq, NULL);
+       } else {
+               proc->breakpoints = NULL;
+       }
+
+#if defined(HAVE_LIBUNWIND)
+       if (options.bt_depth > 0) {
+               proc->unwind_priv = _UPT_create(pid);
+               proc->unwind_as = unw_create_addr_space(&_UPT_accessors, 0);
+
+               if (proc->unwind_priv == NULL || proc->unwind_as == NULL) {
+                       fprintf(stderr,
+                               "Couldn't initialize unwinding "
+                               "for process %d\n", proc->pid);
+                       destroy_unwind(proc);
+                       proc->unwind_priv = NULL;
+                       proc->unwind_as = NULL;
+               }
+       }
+#endif /* defined(HAVE_LIBUNWIND) */
+
+#if defined(HAVE_LIBDW)
+       proc->dwfl = NULL; /* Initialize for leader only on first library.  */
+       proc->should_attach_dwfl = 1; /* should try to attach the DWFL data */
+#endif /* defined(HAVE_LIBDW) */
+
+       return 0;
+}
+
+static void
+process_bare_destroy(struct process *proc, int was_exec)
+{
+       dict_destroy(proc->breakpoints, NULL, NULL, NULL);
+       free(proc->breakpoints);
+       if (!was_exec) {
+               free(proc->filename);
+               unlist_process(proc);
+               destroy_unwind(proc);
+       }
+}
+
+static int
+process_init_main(struct process *proc)
+{
+       if (breakpoints_init(proc) < 0) {
+               fprintf(stderr, "failed to init breakpoints %d\n",
+                       proc->pid);
+               return -1;
+       }
+
+       return 0;
+}
+
+int
+process_init(struct process *proc, const char *filename, pid_t pid)
+{
+       if (process_bare_init(proc, filename, pid, 0) < 0) {
+       fail:
+               fprintf(stderr, "failed to initialize process %d: %s\n",
+                       pid, strerror(errno));
+               return -1;
+       }
+
+       if (os_process_init(proc) < 0) {
+               process_bare_destroy(proc, 0);
+               goto fail;
+       }
+
+       if (arch_process_init(proc) < 0) {
+               os_process_destroy(proc);
+               process_bare_destroy(proc, 0);
+               goto fail;
+       }
+
+       if (proc->leader != proc) {
+               proc->e_machine = proc->leader->e_machine;
+               proc->e_class = proc->leader->e_class;
+               get_arch_dep(proc);
+       } else if (process_init_main(proc) < 0) {
+               process_bare_destroy(proc, 0);
+               goto fail;
+       }
+       return 0;
+}
+
+static enum callback_status
+destroy_breakpoint_cb(struct process *proc, struct breakpoint *bp, void *data)
+{
+       breakpoint_destroy(bp);
+       free(bp);
+       return CBS_CONT;
+}
+
+// XXX see comment in handle_event.c
+void callstack_pop(struct process *proc);
 
-#include "common.h"
+static void
+private_process_destroy(struct process *proc, int was_exec)
+{
+       /* Pop remaining stack elements.  */
+       while (proc->callstack_depth > 0) {
+               /* When this is called just before a process is
+                * destroyed, the breakpoints should either have been
+                * retracted by now, or were killed by exec.  In any
+                * case, it's safe to pretend that there are no
+                * breakpoints associated with the stack elements, so
+                * that stack_pop doesn't attempt to destroy them.  */
+               size_t i = proc->callstack_depth - 1;
+               if (!proc->callstack[i].is_syscall)
+                       proc->callstack[i].return_addr = 0;
 
-Process *
-open_program(char *filename, pid_t pid) {
-       Process *proc;
-       proc = calloc(sizeof(Process), 1);
-       if (!proc) {
-               perror("malloc");
-               exit(1);
+               callstack_pop(proc);
        }
-       proc->filename = strdup(filename);
-       proc->breakpoints_enabled = -1;
-       if (pid) {
-               proc->pid = pid;
+
+       if (!was_exec)
+               free(proc->filename);
+
+       /* Libraries and symbols.  This is only relevant in
+        * leader.  */
+       struct library *lib;
+       for (lib = proc->libraries; lib != NULL; ) {
+               struct library *next = lib->next;
+               library_destroy(lib);
+               free(lib);
+               lib = next;
        }
-       breakpoints_init(proc);
+       proc->libraries = NULL;
+
+       /* Breakpoints.  */
+       if (proc->breakpoints != NULL) {
+               proc_each_breakpoint(proc, NULL, destroy_breakpoint_cb, NULL);
+               dict_destroy(proc->breakpoints, NULL, NULL, NULL);
+               free(proc->breakpoints);
+               proc->breakpoints = NULL;
+       }
+
+       destroy_unwind(proc);
+}
+
+void
+process_destroy(struct process *proc)
+{
+       arch_process_destroy(proc);
+       os_process_destroy(proc);
+       private_process_destroy(proc, 0);
+}
 
-       proc->next = list_of_processes;
-       list_of_processes = proc;
+int
+process_exec(struct process *proc)
+{
+       /* Call exec handlers first, before we destroy the main
+        * state.  */
+       if (arch_process_exec(proc) < 0
+           || os_process_exec(proc) < 0)
+               return -1;
+
+       private_process_destroy(proc, 1);
+
+       if (process_bare_init(proc, NULL, proc->pid, 1) < 0)
+               return -1;
+       if (process_init_main(proc) < 0) {
+               process_bare_destroy(proc, 1);
+               return -1;
+       }
+       return 0;
+}
+
+struct process *
+open_program(const char *filename, pid_t pid)
+{
+       assert(pid != 0);
+       struct process *proc = malloc(sizeof(*proc));
+       if (proc == NULL || process_init(proc, filename, pid) < 0) {
+               free(proc);
+               return NULL;
+       }
        return proc;
 }
 
+struct clone_single_bp_data {
+       struct process *old_proc;
+       struct process *new_proc;
+};
+
+static enum callback_status
+clone_single_bp(arch_addr_t *key, struct breakpoint **bpp, void *u)
+{
+       struct breakpoint *bp = *bpp;
+       struct clone_single_bp_data *data = u;
+
+       struct breakpoint *clone = malloc(sizeof(*clone));
+       if (clone == NULL
+           || breakpoint_clone(clone, data->new_proc, bp) < 0) {
+       fail:
+               free(clone);
+               return CBS_STOP;
+       }
+       if (proc_add_breakpoint(data->new_proc->leader, clone) < 0) {
+               breakpoint_destroy(clone);
+               goto fail;
+       }
+       return CBS_CONT;
+}
+
+int
+process_clone(struct process *retp, struct process *proc, pid_t pid)
+{
+       if (process_bare_init(retp, proc->filename, pid, 0) < 0) {
+       fail1:
+               fprintf(stderr, "Failed to clone process %d to %d: %s\n",
+                       proc->pid, pid, strerror(errno));
+               return -1;
+       }
+
+       retp->tracesysgood = proc->tracesysgood;
+       retp->e_machine = proc->e_machine;
+       retp->e_class = proc->e_class;
+
+       /* For non-leader processes, that's all we need to do.  */
+       if (retp->leader != retp)
+               return 0;
+
+       /* Clone symbols first so that we can clone and relink
+        * breakpoints.  */
+       struct library *lib;
+       struct library **nlibp = &retp->libraries;
+       for (lib = proc->leader->libraries; lib != NULL; lib = lib->next) {
+               *nlibp = malloc(sizeof(**nlibp));
+
+               if (*nlibp == NULL
+                   || library_clone(*nlibp, lib) < 0) {
+                       free(*nlibp);
+                       *nlibp = NULL;
+
+               fail2:
+                       process_bare_destroy(retp, 0);
+
+                       /* Error when cloning.  Unroll what was done.  */
+                       for (lib = retp->libraries; lib != NULL; ) {
+                               struct library *next = lib->next;
+                               library_destroy(lib);
+                               free(lib);
+                               lib = next;
+                       }
+                       goto fail1;
+               }
+
+               nlibp = &(*nlibp)->next;
+       }
+
+       /* Now clone breakpoints.  Symbol relinking is done in
+        * clone_single_bp.  */
+       struct clone_single_bp_data data = {
+               .old_proc = proc,
+               .new_proc = retp,
+       };
+       if (DICT_EACH(proc->leader->breakpoints,
+                     arch_addr_t, struct breakpoint *, NULL,
+                     clone_single_bp, &data) != NULL)
+               goto fail2;
+
+       /* And finally the call stack.  */
+       /* XXX clearly the callstack handling should be moved to a
+        * separate module and this whole business extracted to
+        * callstack_clone, or callstack_element_clone.  */
+       memcpy(retp->callstack, proc->callstack, sizeof(retp->callstack));
+       retp->callstack_depth = proc->callstack_depth;
+
+       size_t i;
+       for (i = 0; i < retp->callstack_depth; ++i) {
+               struct callstack_element *elem = &retp->callstack[i];
+               struct fetch_context *ctx = elem->fetch_context;
+               if (ctx != NULL) {
+                       struct fetch_context *nctx = fetch_arg_clone(retp, ctx);
+                       if (nctx == NULL) {
+                               size_t j;
+                       fail3:
+                               for (j = 0; j < i; ++j) {
+                                       nctx = retp->callstack[j].fetch_context;
+                                       fetch_arg_done(nctx);
+                                       elem->fetch_context = NULL;
+                               }
+                               goto fail2;
+                       }
+                       elem->fetch_context = nctx;
+               }
+
+               if (elem->arguments != NULL) {
+                       struct value_dict *nargs = malloc(sizeof(*nargs));
+                       if (nargs == NULL
+                           || val_dict_clone(nargs, elem->arguments) < 0) {
+                               size_t j;
+                               for (j = 0; j < i; ++j) {
+                                       nargs = retp->callstack[j].arguments;
+                                       val_dict_destroy(nargs);
+                                       free(nargs);
+                                       elem->arguments = NULL;
+                               }
+
+                               /* Pretend that this round went well,
+                                * so that fail3 frees I-th
+                                * fetch_context.  */
+                               ++i;
+                               goto fail3;
+                       }
+                       elem->arguments = nargs;
+               }
+
+               /* If it's not a syscall, we need to find the
+                * corresponding library symbol in the cloned
+                * library.  */
+               if (!elem->is_syscall && elem->c_un.libfunc != NULL) {
+                       struct library_symbol *libfunc = elem->c_un.libfunc;
+                       int rc = proc_find_symbol(retp, libfunc,
+                                                 NULL, &elem->c_un.libfunc);
+                       assert(rc == 0);
+               }
+       }
+
+       /* At this point, retp is fully initialized, except for OS and
+        * arch parts, and we can call private_process_destroy.  */
+       if (os_process_clone(retp, proc) < 0) {
+               private_process_destroy(retp, 0);
+               return -1;
+       }
+       if (arch_process_clone(retp, proc) < 0) {
+               os_process_destroy(retp);
+               private_process_destroy(retp, 0);
+               return -1;
+       }
+
+       return 0;
+}
+
+static int
+open_one_pid(pid_t pid)
+{
+       debug(DEBUG_PROCESS, "open_one_pid(pid=%d)", pid);
+
+       /* Get the filename first.  Should the trace_pid fail, we can
+        * easily free it, untracing is more work.  */
+       char *filename = pid2name(pid);
+       if (filename == NULL || trace_pid(pid) < 0) {
+       fail:
+               free(filename);
+               return -1;
+       }
+
+       struct process *proc = open_program(filename, pid);
+       if (proc == NULL)
+               goto fail;
+       free(filename);
+       trace_set_options(proc);
+
+       return 0;
+}
+
+static enum callback_status
+start_one_pid(struct process *proc, void *data)
+{
+       continue_process(proc->pid);
+       return CBS_CONT;
+}
+
+static enum callback_status
+is_main(struct process *proc, struct library *lib, void *data)
+{
+       return CBS_STOP_IF(lib->type == LT_LIBTYPE_MAIN);
+}
+
+void
+process_hit_start(struct process *proc)
+{
+       struct process *leader = proc->leader;
+       assert(leader != NULL);
+
+       struct library *mainlib
+               = proc_each_library(leader, NULL, is_main, NULL);
+       assert(mainlib != NULL);
+       linkmap_init(leader, mainlib->dyn_addr);
+       arch_dynlink_done(leader);
+}
+
 void
-open_pid(pid_t pid) {
-       Process *proc;
-       char *filename;
+open_pid(pid_t pid)
+{
+       debug(DEBUG_PROCESS, "open_pid(pid=%d)", pid);
+       /* If we are already tracing this guy, we should be seeing all
+        * his children via normal tracing route.  */
+       if (pid2proc(pid) != NULL)
+               return;
+
+       /* First, see if we can attach the requested PID itself.  */
+       if (open_one_pid(pid) < 0) {
+               fprintf(stderr, "Cannot attach to pid %u: %s\n",
+                       pid, strerror(errno));
+               trace_fail_warning(pid);
+               return;
+       }
+
+       /* Now attach to all tasks that belong to that PID.  There's a
+        * race between process_tasks and open_one_pid.  So when we
+        * fail in open_one_pid below, we just do another round.
+        * Chances are that by then that PID will have gone away, and
+        * that's why we have seen the failure.  The processes that we
+        * manage to open_one_pid are stopped, so we should eventually
+        * reach a point where process_tasks doesn't give any new
+        * processes (because there's nobody left to produce
+        * them).  */
+       size_t old_ntasks = 0;
+       int have_all;
+       while (1) {
+               pid_t *tasks;
+               size_t ntasks;
+               size_t i;
+
+               if (process_tasks(pid, &tasks, &ntasks) < 0) {
+                       fprintf(stderr, "Cannot obtain tasks of pid %u: %s\n",
+                               pid, strerror(errno));
+                       break;
+               }
+
+               have_all = 1;
+               for (i = 0; i < ntasks; ++i)
+                       if (pid2proc(tasks[i]) == NULL
+                           && open_one_pid(tasks[i]) < 0)
+                               have_all = 0;
+
+               free(tasks);
+
+               if (have_all && old_ntasks == ntasks)
+                       break;
+               old_ntasks = ntasks;
+       }
+
+       struct process *leader = pid2proc(pid)->leader;
+
+       /* XXX Is there a way to figure out whether _start has
+        * actually already been hit?  */
+       process_hit_start(leader);
+
+       /* Done.  Continue everyone.  */
+       each_task(leader, NULL, start_one_pid, NULL);
+}
 
-       if (trace_pid(pid) < 0) {
-               fprintf(stderr, "Cannot attach to pid %u: %s\n", pid,
-                       strerror(errno));
+static enum callback_status
+find_proc(struct process *proc, void *data)
+{
+       return CBS_STOP_IF(proc->pid == (pid_t)(uintptr_t)data);
+}
+
+struct process *
+pid2proc(pid_t pid)
+{
+       return each_process(NULL, &find_proc, (void *)(uintptr_t)pid);
+}
+
+static struct process *list_of_processes = NULL;
+
+static void
+unlist_process(struct process *proc)
+{
+       if (list_of_processes == proc) {
+               list_of_processes = list_of_processes->next;
                return;
        }
 
-       filename = pid2name(pid);
+       struct process *tmp;
+       for (tmp = list_of_processes; ; tmp = tmp->next) {
+               /* If the following assert fails, the process wasn't
+                * in the list.  */
+               assert(tmp->next != NULL);
+
+               if (tmp->next == proc) {
+                       tmp->next = tmp->next->next;
+                       return;
+               }
+       }
+}
+
+struct process *
+each_process(struct process *start_after,
+            enum callback_status(*cb)(struct process *proc, void *data),
+            void *data)
+{
+       struct process *it = start_after == NULL ? list_of_processes
+               : start_after->next;
 
-       if (!filename) {
-               fprintf(stderr, "Cannot trace pid %u: %s\n", pid,
-                               strerror(errno));
+       while (it != NULL) {
+               /* Callback might call remove_process.  */
+               struct process *next = it->next;
+               switch ((*cb)(it, data)) {
+               case CBS_FAIL:
+                       /* XXX handle me */
+               case CBS_STOP:
+                       return it;
+               case CBS_CONT:
+                       break;
+               }
+               it = next;
+       }
+       return NULL;
+}
+
+struct process *
+each_task(struct process *proc, struct process *start_after,
+         enum callback_status(*cb)(struct process *proc, void *data),
+         void *data)
+{
+       assert(proc != NULL);
+       struct process *it = start_after == NULL ? proc->leader
+               : start_after->next;
+
+       if (it != NULL) {
+               struct process *leader = it->leader;
+               while (it != NULL && it->leader == leader) {
+                       /* Callback might call remove_process.  */
+                       struct process *next = it->next;
+                       switch ((*cb)(it, data)) {
+                       case CBS_FAIL:
+                               /* XXX handle me */
+                       case CBS_STOP:
+                               return it;
+                       case CBS_CONT:
+                               break;
+                       }
+                       it = next;
+               }
+       }
+       return NULL;
+}
+
+static int
+add_process(struct process *proc, int was_exec)
+{
+       struct process **leaderp = &list_of_processes;
+       if (proc->pid) {
+               pid_t tgid = process_leader(proc->pid);
+               if (tgid == 0)
+                       /* Must have been terminated before we managed
+                        * to fully attach.  */
+                       return -1;
+               if (tgid == proc->pid) {
+                       proc->leader = proc;
+               } else {
+                       struct process *leader = pid2proc(tgid);
+                       proc->leader = leader;
+                       if (leader != NULL)
+                               leaderp = &leader->next;
+               }
+       }
+
+       if (!was_exec) {
+               proc->next = *leaderp;
+               *leaderp = proc;
+       }
+       return 0;
+}
+
+void
+change_process_leader(struct process *proc, struct process *leader)
+{
+       struct process **leaderp = &list_of_processes;
+       if (proc->leader == leader)
                return;
+
+       assert(leader != NULL);
+       unlist_process(proc);
+       if (proc != leader)
+               leaderp = &leader->next;
+
+       proc->leader = leader;
+       proc->next = *leaderp;
+       *leaderp = proc;
+}
+
+static enum callback_status
+clear_leader(struct process *proc, void *data)
+{
+       debug(DEBUG_FUNCTION, "detach_task %d from leader %d",
+             proc->pid, proc->leader->pid);
+       proc->leader = NULL;
+       return CBS_CONT;
+}
+
+void
+remove_process(struct process *proc)
+{
+       debug(DEBUG_FUNCTION, "remove_proc(pid=%d)", proc->pid);
+
+       if (proc->leader == proc)
+               each_task(proc, NULL, &clear_leader, NULL);
+
+       unlist_process(proc);
+       process_removed(proc);
+       process_destroy(proc);
+       free(proc);
+}
+
+void
+install_event_handler(struct process *proc, struct event_handler *handler)
+{
+       debug(DEBUG_FUNCTION, "install_event_handler(pid=%d, %p)", proc->pid, handler);
+       assert(proc->event_handler == NULL);
+       proc->event_handler = handler;
+}
+
+void
+destroy_event_handler(struct process *proc)
+{
+       struct event_handler *handler = proc->event_handler;
+       debug(DEBUG_FUNCTION, "destroy_event_handler(pid=%d, %p)", proc->pid, handler);
+       assert(handler != NULL);
+       if (handler->destroy != NULL)
+               handler->destroy(handler);
+       free(handler);
+       proc->event_handler = NULL;
+}
+
+static int
+breakpoint_for_symbol(struct library_symbol *libsym, struct process *proc)
+{
+       arch_addr_t bp_addr;
+       assert(proc->leader == proc);
+
+       /* Don't enable latent or delayed symbols.  */
+       if (libsym->latent || libsym->delayed) {
+               debug(DEBUG_FUNCTION,
+                     "delayed and/or latent breakpoint pid=%d, %s@%p",
+                     proc->pid, libsym->name, libsym->enter_addr);
+               return 0;
+       }
+
+       bp_addr = sym2addr(proc, libsym);
+
+       /* If there is an artificial breakpoint on the same address,
+        * its libsym will be NULL, and we can smuggle our libsym
+        * there.  That artificial breakpoint is there presumably for
+        * the callbacks, which we don't touch.  If there is a real
+        * breakpoint, then this is a bug.  ltrace-elf.c should filter
+        * symbols and ignore extra symbol aliases.
+        *
+        * The other direction is more complicated and currently not
+        * supported.  If a breakpoint has custom callbacks, it might
+        * be also custom-allocated, and we would really need to swap
+        * the two: delete the one now in the dictionary, swap values
+        * around, and put the new breakpoint back in.  */
+       struct breakpoint *bp;
+       if (DICT_FIND_VAL(proc->breakpoints, &bp_addr, &bp) == 0) {
+               /* MIPS backend makes duplicate requests.  This is
+                * likely a bug in the backend.  Currently there's no
+                * point assigning more than one symbol to a
+                * breakpoint, because when it hits, we won't know
+                * what to print out.  But it's easier to fix it here
+                * before someone who understands MIPS has the time to
+                * look into it.  So turn the sanity check off on
+                * MIPS.  References:
+                *
+                *   http://lists.alioth.debian.org/pipermail/ltrace-devel/2012-November/000764.html
+                *   http://lists.alioth.debian.org/pipermail/ltrace-devel/2012-November/000770.html
+                */
+#ifndef __mips__
+               assert(bp->libsym == NULL);
+               bp->libsym = libsym;
+#endif
+               return 0;
+       }
+
+       bp = malloc(sizeof(*bp));
+       if (bp == NULL
+           || breakpoint_init(bp, proc, bp_addr, libsym) < 0) {
+       fail:
+               free(bp);
+               return -1;
+       }
+       if (proc_add_breakpoint(proc, bp) < 0) {
+               breakpoint_destroy(bp);
+               goto fail;
+       }
+
+       if (breakpoint_turn_on(bp, proc) < 0) {
+               proc_remove_breakpoint(proc, bp);
+               breakpoint_destroy(bp);
+               goto fail;
+       }
+
+       return 0;
+}
+
+static enum callback_status
+cb_breakpoint_for_symbol(struct library_symbol *libsym, void *data)
+{
+       return CBS_STOP_IF(breakpoint_for_symbol(libsym, data) < 0);
+}
+
+static int
+proc_activate_latent_symbol(struct process *proc,
+                           struct library_symbol *libsym)
+{
+       assert(libsym->latent);
+       libsym->latent = 0;
+       debug(DEBUG_FUNCTION, "activated latent symbol");
+       return breakpoint_for_symbol(libsym, proc);
+}
+
+int
+proc_activate_delayed_symbol(struct process *proc,
+                            struct library_symbol *libsym)
+{
+       assert(libsym->delayed);
+       libsym->delayed = 0;
+       debug(DEBUG_FUNCTION, "activated delayed symbol");
+       return breakpoint_for_symbol(libsym, proc);
+}
+
+
+struct activate_latent_in_context
+{
+       struct process *proc;
+       struct library_exported_names *exported_names;
+};
+static enum callback_status
+activate_latent_in_cb(struct library_symbol *libsym, void *data)
+{
+       struct activate_latent_in_context *ctx =
+               (struct activate_latent_in_context*)data;
+
+       if (libsym->latent &&
+           library_exported_names_contains(ctx->exported_names,
+                                           libsym->name) != 0)
+               proc_activate_latent_symbol(ctx->proc, libsym);
+
+       return CBS_CONT;
+}
+
+static enum callback_status
+activate_latent_in(struct process *proc, struct library *lib, void *data)
+{
+       struct library_symbol *libsym = NULL;
+
+       struct library_exported_names *exported_names =
+               (struct library_exported_names*)data;
+
+       struct activate_latent_in_context ctx =
+               {.proc = proc,
+                .exported_names = exported_names};
+
+       if (library_each_symbol(lib, libsym,
+                               activate_latent_in_cb,
+                               &ctx) != NULL)
+               return CBS_FAIL;
+
+       return CBS_CONT;
+}
+
+void
+proc_add_library(struct process *proc, struct library *lib)
+{
+       assert(lib->next == NULL);
+       lib->next = proc->libraries;
+       proc->libraries = lib;
+       debug(DEBUG_PROCESS, "added library %s@%p (%s) to %d",
+             lib->soname, lib->base, lib->pathname, proc->pid);
+
+#if defined(HAVE_LIBDW)
+       Dwfl *dwfl = NULL;
+       Dwfl_Module *dwfl_module = NULL;
+
+       /* Setup module tracking for libdwfl unwinding.  */
+       struct process *leader = proc->leader;
+       dwfl = leader->dwfl;
+       if (dwfl == NULL) {
+               static const Dwfl_Callbacks proc_callbacks = {
+                       .find_elf = dwfl_linux_proc_find_elf,
+                       .find_debuginfo = dwfl_standard_find_debuginfo
+               };
+               dwfl = dwfl_begin(&proc_callbacks);
+               if (dwfl == NULL)
+                       fprintf(stderr,
+                               "Couldn't initialize libdwfl unwinding "
+                               "for process %d: %s\n", leader->pid,
+                               dwfl_errmsg (-1));
        }
 
-       proc = open_program(filename, pid);
-       continue_process(pid);
-       proc->breakpoints_enabled = 1;
+       if (dwfl != NULL) {
+               dwfl_report_begin_add(dwfl);
+               dwfl_module =
+                       dwfl_report_elf(dwfl, lib->soname,
+                                       lib->pathname, -1,
+                                       (GElf_Addr) lib->base,
+                                       false);
+               if (dwfl_module == NULL)
+                       fprintf(stderr,
+                               "dwfl_report_elf %s@%p (%s) %d: %s\n",
+                               lib->soname, lib->base, lib->pathname,
+                               proc->pid, dwfl_errmsg (-1));
+
+               dwfl_report_end(dwfl, NULL, NULL);
+
+               if (options.bt_depth > 0) {
+                       if (proc->should_attach_dwfl) {
+                               int r = dwfl_linux_proc_attach(dwfl,
+                                                              leader->pid,
+                                                              true);
+                               proc->should_attach_dwfl = 0;
+                               if (r != 0) {
+                                       const char *msg;
+                                       dwfl_end(dwfl);
+                                       if (r < 0)
+                                               msg = dwfl_errmsg(-1);
+                                       else
+                                               msg = strerror(r);
+                                       fprintf(stderr, "Couldn't initialize "
+                                               "libdwfl (unwinding, prototype "
+                                               "import) for process %d: %s\n",
+                                               leader->pid, msg);
+                               }
+                       }
+               }
+       }
+
+       lib->dwfl_module = dwfl_module;
+       leader->dwfl = dwfl;
+
+#endif /* defined(HAVE_LIBDW) */
+
+       /* Insert breakpoints for all active (non-latent) symbols.  */
+       struct library_symbol *libsym = NULL;
+       while ((libsym = library_each_symbol(lib, libsym,
+                                            cb_breakpoint_for_symbol,
+                                            proc)) != NULL)
+               fprintf(stderr,
+                       "Couldn't insert breakpoint for %s to %d: %s.\n",
+                       libsym->name, proc->pid, strerror(errno));
+
+       if (lib->should_activate_latent != 0) {
+               /* Look through export list of the new library and compare it
+                * with latent symbols of all libraries (including this
+                * library itself).  */
+               struct library *lib2 = NULL;
+
+               while ((lib2 = proc_each_library(proc, lib2, activate_latent_in,
+                                                &lib->exported_names)) != NULL)
+                       fprintf(stderr,
+                               "Couldn't activate latent symbols "
+                               "for %s in %d: %s.\n",
+                               lib2->soname, proc->pid, strerror(errno));
+       }
 }
 
-Process *
-pid2proc(pid_t pid) {
-       Process *tmp;
+int
+proc_remove_library(struct process *proc, struct library *lib)
+{
+       struct library **libp;
+       for (libp = &proc->libraries; *libp != NULL; libp = &(*libp)->next)
+               if (*libp == lib) {
+                       *libp = lib->next;
+                       return 0;
+               }
+       return -1;
+}
+
+struct library *
+proc_each_library(struct process *proc, struct library *it,
+                 enum callback_status (*cb)(struct process *proc,
+                                            struct library *lib, void *data),
+                 void *data)
+{
+       if (it == NULL)
+               it = proc->libraries;
+       else
+               it = it->next;
+
+       while (it != NULL) {
+               struct library *next = it->next;
 
-       tmp = list_of_processes;
-       while (tmp) {
-               if (pid == tmp->pid) {
-                       return tmp;
+               switch (cb(proc, it, data)) {
+               case CBS_FAIL:
+                       /* XXX handle me */
+               case CBS_STOP:
+                       return it;
+               case CBS_CONT:
+                       break;
                }
-               tmp = tmp->next;
+
+               it = next;
        }
+
        return NULL;
 }
+
+static void
+check_leader(struct process *proc)
+{
+       /* Only the group leader should be getting the breakpoints and
+        * thus have ->breakpoint initialized.  */
+       assert(proc->leader != NULL);
+       assert(proc->leader == proc);
+       assert(proc->breakpoints != NULL);
+}
+
+int
+proc_add_breakpoint(struct process *proc, struct breakpoint *bp)
+{
+       debug(DEBUG_FUNCTION, "proc_add_breakpoint(pid=%d, %s@%p)",
+             proc->pid, breakpoint_name(bp), bp->addr);
+       check_leader(proc);
+
+       /* XXX We might merge bp->libsym instead of the following
+        * assert, but that's not necessary right now.  Read the
+        * comment in breakpoint_for_symbol.  */
+       assert(dict_find(proc->breakpoints, &bp->addr) == NULL);
+
+       if (DICT_INSERT(proc->breakpoints, &bp->addr, &bp) < 0) {
+               fprintf(stderr,
+                       "couldn't enter breakpoint %s@%p to dictionary: %s\n",
+                       breakpoint_name(bp), bp->addr, strerror(errno));
+               return -1;
+       }
+
+       return 0;
+}
+
+void
+proc_remove_breakpoint(struct process *proc, struct breakpoint *bp)
+{
+       debug(DEBUG_FUNCTION, "proc_remove_breakpoint(pid=%d, %s@%p)",
+             proc->pid, breakpoint_name(bp), bp->addr);
+       check_leader(proc);
+       int rc = DICT_ERASE(proc->breakpoints, &bp->addr, struct breakpoint *,
+                           NULL, NULL, NULL);
+       assert(rc == 0);
+}
+
+struct each_breakpoint_data
+{
+       struct process *proc;
+       enum callback_status (*cb)(struct process *proc,
+                                  struct breakpoint *bp,
+                                  void *data);
+       void *cb_data;
+};
+
+static enum callback_status
+each_breakpoint_cb(arch_addr_t *key, struct breakpoint **bpp, void *d)
+{
+       struct each_breakpoint_data *data = d;
+       return data->cb(data->proc, *bpp, data->cb_data);
+}
+
+arch_addr_t *
+proc_each_breakpoint(struct process *proc, arch_addr_t *start,
+                    enum callback_status (*cb)(struct process *proc,
+                                               struct breakpoint *bp,
+                                               void *data), void *data)
+{
+       struct each_breakpoint_data dd = {
+               .proc = proc,
+               .cb = cb,
+               .cb_data = data,
+       };
+       return DICT_EACH(proc->breakpoints,
+                        arch_addr_t, struct breakpoint *, start,
+                        &each_breakpoint_cb, &dd);
+}
+
+int
+proc_find_symbol(struct process *proc, struct library_symbol *sym,
+                struct library **retlib, struct library_symbol **retsym)
+{
+       struct library *lib = sym->lib;
+       assert(lib != NULL);
+
+       struct library *flib
+               = proc_each_library(proc, NULL, library_with_key_cb, &lib->key);
+       if (flib == NULL)
+               return -1;
+
+       struct library_symbol *fsym
+               = library_each_symbol(flib, NULL, library_symbol_named_cb,
+                                     (char *)sym->name);
+       if (fsym == NULL)
+               return -1;
+
+       if (retlib != NULL)
+               *retlib = flib;
+       if (retsym != NULL)
+               *retsym = fsym;
+
+       return 0;
+}
+
+struct library_symbol *
+proc_each_symbol(struct process *proc, struct library_symbol *start_after,
+                enum callback_status (*cb)(struct library_symbol *, void *),
+                void *data)
+{
+       struct library *lib;
+       for (lib = start_after != NULL ? start_after->lib : proc->libraries;
+            lib != NULL; lib = lib->next) {
+               start_after = library_each_symbol(lib, start_after, cb, data);
+               if (start_after != NULL)
+                       return start_after;
+       }
+
+       return NULL;
+}
+
+#define DEF_READER(NAME, SIZE)                                         \
+       int                                                             \
+       NAME(struct process *proc, arch_addr_t addr,                    \
+            uint##SIZE##_t *lp)                                        \
+       {                                                               \
+               union {                                                 \
+                       uint##SIZE##_t dst;                             \
+                       char buf[0];                                    \
+               } u;                                                    \
+               if (umovebytes(proc, addr, &u.buf, sizeof(u.dst))       \
+                   != sizeof(u.dst))                                   \
+                       return -1;                                      \
+               *lp = u.dst;                                            \
+               return 0;                                               \
+       }
+
+DEF_READER(proc_read_8, 8)
+DEF_READER(proc_read_16, 16)
+DEF_READER(proc_read_32, 32)
+DEF_READER(proc_read_64, 64)
+
+#undef DEF_READER