1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * Copyright (C) 2015 Thomas Meyer (thomas@m3y3r.de)
4  * Copyright (C) 2002 - 2007 Jeff Dike (jdike@{addtoit,linux.intel}.com)
5  */
6 
7 #include <linux/mm.h>
8 #include <linux/sched/signal.h>
9 #include <linux/slab.h>
10 
11 #include <shared/irq_kern.h>
12 #include <asm/pgalloc.h>
13 #include <asm/sections.h>
14 #include <asm/mmu_context.h>
15 #include <as-layout.h>
16 #include <os.h>
17 #include <skas.h>
18 #include <stub-data.h>
19 
20 /* Ensure the stub_data struct covers the allocated area */
21 static_assert(sizeof(struct stub_data) == STUB_DATA_PAGES * UM_KERN_PAGE_SIZE);
22 
23 spinlock_t mm_list_lock;
24 struct list_head mm_list;
25 
26 int init_new_context(struct task_struct *task, struct mm_struct *mm)
27 {
28 	struct mm_id *new_id = &mm->context.id;
29 	unsigned long stack = 0;
30 	int ret = -ENOMEM;
31 
32 	stack = __get_free_pages(GFP_KERNEL | __GFP_ZERO, ilog2(STUB_DATA_PAGES));
33 	if (stack == 0)
34 		goto out;
35 
36 	new_id->stack = stack;
37 
38 	scoped_guard(spinlock_irqsave, &mm_list_lock) {
39 		/* Insert into list, used for lookups when the child dies */
40 		list_add(&mm->context.list, &mm_list);
41 	}
42 
43 	ret = start_userspace(new_id);
44 	if (ret < 0)
45 		goto out_free;
46 
47 	/* Ensure the new MM is clean and nothing unwanted is mapped */
48 	unmap(new_id, 0, STUB_START);
49 
50 	return 0;
51 
52  out_free:
53 	free_pages(new_id->stack, ilog2(STUB_DATA_PAGES));
54  out:
55 	return ret;
56 }
57 
58 void destroy_context(struct mm_struct *mm)
59 {
60 	struct mm_context *mmu = &mm->context;
61 
62 	/*
63 	 * If init_new_context wasn't called, this will be
64 	 * zero, resulting in a kill(0), which will result in the
65 	 * whole UML suddenly dying.  Also, cover negative and
66 	 * 1 cases, since they shouldn't happen either.
67 	 *
68 	 * Negative cases happen if the child died unexpectedly.
69 	 */
70 	if (mmu->id.pid >= 0 && mmu->id.pid < 2) {
71 		printk(KERN_ERR "corrupt mm_context - pid = %d\n",
72 		       mmu->id.pid);
73 		return;
74 	}
75 
76 	if (mmu->id.pid > 0) {
77 		os_kill_ptraced_process(mmu->id.pid, 1);
78 		mmu->id.pid = -1;
79 	}
80 
81 	if (using_seccomp && mmu->id.sock)
82 		os_close_file(mmu->id.sock);
83 
84 	free_pages(mmu->id.stack, ilog2(STUB_DATA_PAGES));
85 
86 	guard(spinlock_irqsave)(&mm_list_lock);
87 
88 	list_del(&mm->context.list);
89 }
90 
91 static irqreturn_t mm_sigchld_irq(int irq, void* dev)
92 {
93 	struct mm_context *mm_context;
94 	pid_t pid;
95 
96 	guard(spinlock)(&mm_list_lock);
97 
98 	while ((pid = os_reap_child()) > 0) {
99 		/*
100 		* A child died, check if we have an MM with the PID. This is
101 		* only relevant in SECCOMP mode (as ptrace will fail anyway).
102 		*
103 		* See wait_stub_done_seccomp for more details.
104 		*/
105 		list_for_each_entry(mm_context, &mm_list, list) {
106 			if (mm_context->id.pid == pid) {
107 				struct stub_data *stub_data;
108 				printk("Unexpectedly lost MM child! Affected tasks will segfault.");
109 
110 				/* Marks the MM as dead */
111 				mm_context->id.pid = -1;
112 
113 				/*
114 				 * NOTE: If SMP is implemented, a futex_wake
115 				 * needs to be added here.
116 				 */
117 				stub_data = (void *)mm_context->id.stack;
118 				stub_data->futex = FUTEX_IN_KERN;
119 
120 				/*
121 				 * NOTE: Currently executing syscalls by
122 				 * affected tasks may finish normally.
123 				 */
124 				break;
125 			}
126 		}
127 	}
128 
129 	return IRQ_HANDLED;
130 }
131 
132 static int __init init_child_tracking(void)
133 {
134 	int err;
135 
136 	spin_lock_init(&mm_list_lock);
137 	INIT_LIST_HEAD(&mm_list);
138 
139 	err = request_irq(SIGCHLD_IRQ, mm_sigchld_irq, 0, "SIGCHLD", NULL);
140 	if (err < 0)
141 		panic("Failed to register SIGCHLD IRQ: %d", err);
142 
143 	return 0;
144 }
145 early_initcall(init_child_tracking)
146