| 1 | /* SPDX-License-Identifier: GPL-2.0-or-later */ |
| 2 | /* |
| 3 | * kmod dups - the kernel module autoloader duplicate suppressor |
| 4 | * |
| 5 | * Copyright (C) 2023 Luis Chamberlain <mcgrof@kernel.org> |
| 6 | */ |
| 7 | |
| 8 | #define pr_fmt(fmt) "module: " fmt |
| 9 | |
| 10 | #include <linux/module.h> |
| 11 | #include <linux/sched.h> |
| 12 | #include <linux/sched/task.h> |
| 13 | #include <linux/binfmts.h> |
| 14 | #include <linux/syscalls.h> |
| 15 | #include <linux/unistd.h> |
| 16 | #include <linux/kmod.h> |
| 17 | #include <linux/slab.h> |
| 18 | #include <linux/completion.h> |
| 19 | #include <linux/cred.h> |
| 20 | #include <linux/file.h> |
| 21 | #include <linux/workqueue.h> |
| 22 | #include <linux/security.h> |
| 23 | #include <linux/mount.h> |
| 24 | #include <linux/kernel.h> |
| 25 | #include <linux/init.h> |
| 26 | #include <linux/resource.h> |
| 27 | #include <linux/notifier.h> |
| 28 | #include <linux/suspend.h> |
| 29 | #include <linux/rwsem.h> |
| 30 | #include <linux/ptrace.h> |
| 31 | #include <linux/async.h> |
| 32 | #include <linux/uaccess.h> |
| 33 | |
| 34 | #include "internal.h" |
| 35 | |
| 36 | #undef MODULE_PARAM_PREFIX |
| 37 | #define MODULE_PARAM_PREFIX "module." |
| 38 | static bool enable_dups_trace = IS_ENABLED(CONFIG_MODULE_DEBUG_AUTOLOAD_DUPS_TRACE); |
| 39 | module_param(enable_dups_trace, bool_enable_only, 0644); |
| 40 | |
| 41 | /* |
| 42 | * Protects dup_kmod_reqs list, adds / removals with RCU. |
| 43 | */ |
| 44 | static DEFINE_MUTEX(kmod_dup_mutex); |
| 45 | static LIST_HEAD(dup_kmod_reqs); |
| 46 | |
| 47 | struct kmod_dup_req { |
| 48 | struct list_head list; |
| 49 | char name[MODULE_NAME_LEN]; |
| 50 | struct completion first_req_done; |
| 51 | struct work_struct complete_work; |
| 52 | struct delayed_work delete_work; |
| 53 | int dup_ret; |
| 54 | }; |
| 55 | |
| 56 | static struct kmod_dup_req *kmod_dup_request_lookup(char *module_name) |
| 57 | { |
| 58 | struct kmod_dup_req *kmod_req; |
| 59 | |
| 60 | list_for_each_entry_rcu(kmod_req, &dup_kmod_reqs, list, |
| 61 | lockdep_is_held(&kmod_dup_mutex)) { |
| 62 | if (strlen(kmod_req->name) == strlen(module_name) && |
| 63 | !memcmp(p: kmod_req->name, q: module_name, strlen(module_name))) { |
| 64 | return kmod_req; |
| 65 | } |
| 66 | } |
| 67 | |
| 68 | return NULL; |
| 69 | } |
| 70 | |
| 71 | static void kmod_dup_request_delete(struct work_struct *work) |
| 72 | { |
| 73 | struct kmod_dup_req *kmod_req; |
| 74 | kmod_req = container_of(to_delayed_work(work), struct kmod_dup_req, delete_work); |
| 75 | |
| 76 | /* |
| 77 | * The typical situation is a module successully loaded. In that |
| 78 | * situation the module will be present already in userspace. If |
| 79 | * new requests come in after that, userspace will already know the |
| 80 | * module is loaded so will just return 0 right away. There is still |
| 81 | * a small chance right after we delete this entry new request_module() |
| 82 | * calls may happen after that, they can happen. These heuristics |
| 83 | * are to protect finit_module() abuse for auto-loading, if modules |
| 84 | * are still tryign to auto-load even if a module is already loaded, |
| 85 | * that's on them, and those inneficiencies should not be fixed by |
| 86 | * kmod. The inneficies there are a call to modprobe and modprobe |
| 87 | * just returning 0. |
| 88 | */ |
| 89 | mutex_lock(&kmod_dup_mutex); |
| 90 | list_del_rcu(entry: &kmod_req->list); |
| 91 | synchronize_rcu(); |
| 92 | mutex_unlock(lock: &kmod_dup_mutex); |
| 93 | kfree(objp: kmod_req); |
| 94 | } |
| 95 | |
| 96 | static void kmod_dup_request_complete(struct work_struct *work) |
| 97 | { |
| 98 | struct kmod_dup_req *kmod_req; |
| 99 | |
| 100 | kmod_req = container_of(work, struct kmod_dup_req, complete_work); |
| 101 | |
| 102 | /* |
| 103 | * This will ensure that the kernel will let all the waiters get |
| 104 | * informed its time to check the return value. It's time to |
| 105 | * go home. |
| 106 | */ |
| 107 | complete_all(&kmod_req->first_req_done); |
| 108 | |
| 109 | /* |
| 110 | * Now that we have allowed prior request_module() calls to go on |
| 111 | * with life, let's schedule deleting this entry. We don't have |
| 112 | * to do it right away, but we *eventually* want to do it so to not |
| 113 | * let this linger forever as this is just a boot optimization for |
| 114 | * possible abuses of vmalloc() incurred by finit_module() thrashing. |
| 115 | */ |
| 116 | queue_delayed_work(wq: system_wq, dwork: &kmod_req->delete_work, delay: 60 * HZ); |
| 117 | } |
| 118 | |
| 119 | bool kmod_dup_request_exists_wait(char *module_name, bool wait, int *dup_ret) |
| 120 | { |
| 121 | struct kmod_dup_req *kmod_req, *new_kmod_req; |
| 122 | int ret; |
| 123 | |
| 124 | /* |
| 125 | * Pre-allocate the entry in case we have to use it later |
| 126 | * to avoid contention with the mutex. |
| 127 | */ |
| 128 | new_kmod_req = kzalloc(sizeof(*new_kmod_req), GFP_KERNEL); |
| 129 | if (!new_kmod_req) |
| 130 | return false; |
| 131 | |
| 132 | memcpy(new_kmod_req->name, module_name, strlen(module_name)); |
| 133 | INIT_WORK(&new_kmod_req->complete_work, kmod_dup_request_complete); |
| 134 | INIT_DELAYED_WORK(&new_kmod_req->delete_work, kmod_dup_request_delete); |
| 135 | init_completion(x: &new_kmod_req->first_req_done); |
| 136 | |
| 137 | mutex_lock(&kmod_dup_mutex); |
| 138 | |
| 139 | kmod_req = kmod_dup_request_lookup(module_name); |
| 140 | if (!kmod_req) { |
| 141 | /* |
| 142 | * If the first request that came through for a module |
| 143 | * was with request_module_nowait() we cannot wait for it |
| 144 | * and share its return value with other users which may |
| 145 | * have used request_module() and need a proper return value |
| 146 | * so just skip using them as an anchor. |
| 147 | * |
| 148 | * If a prior request to this one came through with |
| 149 | * request_module() though, then a request_module_nowait() |
| 150 | * would benefit from duplicate detection. |
| 151 | */ |
| 152 | if (!wait) { |
| 153 | kfree(objp: new_kmod_req); |
| 154 | pr_debug("New request_module_nowait() for %s -- cannot track duplicates for this request\n" , module_name); |
| 155 | mutex_unlock(lock: &kmod_dup_mutex); |
| 156 | return false; |
| 157 | } |
| 158 | |
| 159 | /* |
| 160 | * There was no duplicate, just add the request so we can |
| 161 | * keep tab on duplicates later. |
| 162 | */ |
| 163 | pr_debug("New request_module() for %s\n" , module_name); |
| 164 | list_add_rcu(new: &new_kmod_req->list, head: &dup_kmod_reqs); |
| 165 | mutex_unlock(lock: &kmod_dup_mutex); |
| 166 | return false; |
| 167 | } |
| 168 | mutex_unlock(lock: &kmod_dup_mutex); |
| 169 | |
| 170 | /* We are dealing with a duplicate request now */ |
| 171 | kfree(objp: new_kmod_req); |
| 172 | |
| 173 | /* |
| 174 | * To fix these try to use try_then_request_module() instead as that |
| 175 | * will check if the component you are looking for is present or not. |
| 176 | * You could also just queue a single request to load the module once, |
| 177 | * instead of having each and everything you need try to request for |
| 178 | * the module. |
| 179 | * |
| 180 | * Duplicate request_module() calls can cause quite a bit of wasted |
| 181 | * vmalloc() space when racing with userspace. |
| 182 | */ |
| 183 | if (enable_dups_trace) |
| 184 | WARN(1, "module-autoload: duplicate request for module %s\n" , module_name); |
| 185 | else |
| 186 | pr_warn("module-autoload: duplicate request for module %s\n" , module_name); |
| 187 | |
| 188 | if (!wait) { |
| 189 | /* |
| 190 | * If request_module_nowait() was used then the user just |
| 191 | * wanted to issue the request and if another module request |
| 192 | * was already its way with the same name we don't care for |
| 193 | * the return value either. Let duplicate request_module_nowait() |
| 194 | * calls bail out right away. |
| 195 | */ |
| 196 | *dup_ret = 0; |
| 197 | return true; |
| 198 | } |
| 199 | |
| 200 | /* |
| 201 | * If a duplicate request_module() was used they *may* care for |
| 202 | * the return value, so we have no other option but to wait for |
| 203 | * the first caller to complete. If the first caller used |
| 204 | * the request_module_nowait() call, subsquent callers will |
| 205 | * deal with the comprmise of getting a successful call with this |
| 206 | * optimization enabled ... |
| 207 | */ |
| 208 | ret = wait_for_completion_state(x: &kmod_req->first_req_done, |
| 209 | TASK_KILLABLE); |
| 210 | if (ret) { |
| 211 | *dup_ret = ret; |
| 212 | return true; |
| 213 | } |
| 214 | |
| 215 | /* Now the duplicate request has the same exact return value as the first request */ |
| 216 | *dup_ret = kmod_req->dup_ret; |
| 217 | |
| 218 | return true; |
| 219 | } |
| 220 | |
| 221 | void kmod_dup_request_announce(char *module_name, int ret) |
| 222 | { |
| 223 | struct kmod_dup_req *kmod_req; |
| 224 | |
| 225 | mutex_lock(&kmod_dup_mutex); |
| 226 | |
| 227 | kmod_req = kmod_dup_request_lookup(module_name); |
| 228 | if (!kmod_req) |
| 229 | goto out; |
| 230 | |
| 231 | kmod_req->dup_ret = ret; |
| 232 | |
| 233 | /* |
| 234 | * If we complete() here we may allow duplicate threads |
| 235 | * to continue before the first one that submitted the |
| 236 | * request. We're in no rush also, given that each and |
| 237 | * every bounce back to userspace is slow we avoid that |
| 238 | * with a slight delay here. So queueue up the completion |
| 239 | * and let duplicates suffer, just wait a tad bit longer. |
| 240 | * There is no rush. But we also don't want to hold the |
| 241 | * caller up forever or introduce any boot delays. |
| 242 | */ |
| 243 | queue_work(wq: system_wq, work: &kmod_req->complete_work); |
| 244 | |
| 245 | out: |
| 246 | mutex_unlock(lock: &kmod_dup_mutex); |
| 247 | } |
| 248 | |