Skip to contentSkip to conversation

text_poke/ftrace/x86: Allow text_poke() to be called in early boot

RFC PATCH In review Steven Rostedt 3 participants 7 messages Oct 25, 2022

Subsystems (10)
x86 Library code MM core Tracing +6 more subsystems
lore
Share & tools
lore

Steven Rostedt

author maintainer
#0 lore raw

RFC PATCH 1/1 text_poke/ftrace/x86: Allow text_poke() to be called in early boot

From: "Steven Rostedt (Google)" <[email protected]>

Currently text_poke() just does a simple memcpy() on early boot because the kernel code is read writable at that time. But ftrace uses text_poke on the ftrace trampoline, which is not part of kernel text, and having non kernel text around that can be writable and executable causes several special cases where checks for system_state == SYSTEM_BOOTING needs to be done to ignore this special case. This is tricky and can lead to memory that can be kernel writable and executable after boot (due to bugs).

By moving poking_init() to mm_init() which is called before ftrace_init(), this will allow ftrace to create its trampoline as read only, and the text_poke() will do its normal thing.

This required some updates to fork and the maple_tree code to allow it to be called with enabling interrupts in the time when interrupts must remain disabled.

text_poke() will still use memcpy() on kernel core text during boot up as it keeps things fast for all static_branch()es and such as well as modifying the ftrace locations at boot up too.

This removes the special code added around ftrace trampolines in x86 to be writable and executable during boot up.

---

** Note this may break other architectures. **

 arch/x86/include/asm/ftrace.h |  6 ------
 arch/x86/kernel/alternative.c |  6 ++++--
 arch/x86/kernel/ftrace.c      | 29 +----------------------------
 arch/x86/mm/init_64.c         |  2 --
 init/main.c                   |  8 ++++----
 kernel/fork.c                 |  8 +++++++-
 lib/maple_tree.c              | 16 +++++++++++++++-
 7 files changed, 31 insertions(+), 44 deletions(-)
arch/x86/include/asm/ftrace.h +0 -6
index 908d99b127d3..b27cd4de3fb3 100644
@@ -85,12 +85,6 @@ struct dyn_arch_ftrace {
 
 #ifndef __ASSEMBLY__
 
-#if defined(CONFIG_FUNCTION_TRACER) && defined(CONFIG_DYNAMIC_FTRACE)
-extern void set_ftrace_ops_ro(void);
-#else
-static inline void set_ftrace_ops_ro(void) { }
-#endif
-
 #define ARCH_HAS_SYSCALL_MATCH_SYM_NAME
 static inline bool arch_syscall_match_sym_name(const char *sym, const char *name)
 {
arch/x86/kernel/alternative.c +4 -2
index 5cadcea035e0..ef30a6b78837 100644
@@ -1681,7 +1681,8 @@ void __ref text_poke_queue(void *addr, const void *opcode, size_t len, const voi
 {
 	struct text_poke_loc *tp;
 
-	if (unlikely(system_state == SYSTEM_BOOTING)) {
+	if (unlikely(system_state == SYSTEM_BOOTING &&
+		     core_kernel_text((unsigned long)addr))) {
 		text_poke_early(addr, opcode, len);
 		return;
 	}
@@ -1707,7 +1708,8 @@ void __ref text_poke_bp(void *addr, const void *opcode, size_t len, const void *
 {
 	struct text_poke_loc tp;
 
-	if (unlikely(system_state == SYSTEM_BOOTING)) {
+	if (unlikely(system_state == SYSTEM_BOOTING &&
+		     core_kernel_text((unsigned long)addr))) {
 		text_poke_early(addr, opcode, len);
 		return;
 	}
arch/x86/kernel/ftrace.c +1 -28
index bd165004776d..3aa4c02f63d2 100644
@@ -415,8 +415,7 @@ create_trampoline(struct ftrace_ops *ops, unsigned int *tramp_size)
 
 	set_vm_flush_reset_perms(trampoline);
 
-	if (likely(system_state != SYSTEM_BOOTING))
-		set_memory_ro((unsigned long)trampoline, npages);
+	set_memory_ro((unsigned long)trampoline, npages);
 	set_memory_x((unsigned long)trampoline, npages);
 	return (unsigned long)trampoline;
 fail:
@@ -424,32 +423,6 @@ create_trampoline(struct ftrace_ops *ops, unsigned int *tramp_size)
 	return 0;
 }
 
-void set_ftrace_ops_ro(void)
-{
-	struct ftrace_ops *ops;
-	unsigned long start_offset;
-	unsigned long end_offset;
-	unsigned long npages;
-	unsigned long size;
-
-	do_for_each_ftrace_op(ops, ftrace_ops_list) {
-		if (!(ops->flags & FTRACE_OPS_FL_ALLOC_TRAMP))
-			continue;
-
-		if (ops->flags & FTRACE_OPS_FL_SAVE_REGS) {
-			start_offset = (unsigned long)ftrace_regs_caller;
-			end_offset = (unsigned long)ftrace_regs_caller_end;
-		} else {
-			start_offset = (unsigned long)ftrace_caller;
-			end_offset = (unsigned long)ftrace_caller_end;
-		}
-		size = end_offset - start_offset;
-		size = size + RET_SIZE + sizeof(void *);
-		npages = DIV_ROUND_UP(size, PAGE_SIZE);
-		set_memory_ro((unsigned long)ops->trampoline, npages);
-	} while_for_each_ftrace_op(ops);
-}
-
 static unsigned long calc_trampoline_call_offset(bool save_regs)
 {
 	unsigned long start_offset;
arch/x86/mm/init_64.c +0 -2
index 3f040c6e5d13..03ac9f914f28 100644
@@ -1398,8 +1398,6 @@ void mark_rodata_ro(void)
 	all_end = roundup((unsigned long)_brk_end, PMD_SIZE);
 	set_memory_nx(text_end, (all_end - text_end) >> PAGE_SHIFT);
 
-	set_ftrace_ops_ro();
-
 #ifdef CONFIG_CPA_DEBUG
 	printk(KERN_INFO "Testing CPA: undo %lx-%lx\n", start, end);
 	set_memory_rw(start, (end-start) >> PAGE_SHIFT);
init/main.c +4 -4
index aa21add5f7c5..e5f4ae2d4cca 100644
@@ -860,6 +860,10 @@ static void __init mm_init(void)
 	/* Should be run after espfix64 is set up. */
 	pti_init();
 	kmsan_init_runtime();
+	proc_caches_init();
+	radix_tree_init();
+	maple_tree_init();
+	poking_init();
 }
 
 #ifdef CONFIG_RANDOMIZE_KSTACK_OFFSET
@@ -1011,8 +1015,6 @@ asmlinkage __visible void __init __no_sanitize_address start_kernel(void)
 	if (WARN(!irqs_disabled(),
 		 "Interrupts were enabled *very* early, fixing it\n"))
 		local_irq_disable();
-	radix_tree_init();
-	maple_tree_init();
 
 	/*
 	 * Set up housekeeping before setting up workqueues to allow the unbound
@@ -1117,7 +1119,6 @@ asmlinkage __visible void __init __no_sanitize_address start_kernel(void)
 	thread_stack_cache_init();
 	cred_init();
 	fork_init();
-	proc_caches_init();
 	uts_ns_init();
 	key_init();
 	security_init();
@@ -1134,7 +1135,6 @@ asmlinkage __visible void __init __no_sanitize_address start_kernel(void)
 	taskstats_init_early();
 	delayacct_init();
 
-	poking_init();
 	check_bugs();
 
 	acpi_subsystem_init();
kernel/fork.c +7 -1
index 08969f5aa38d..672967a9cbe9 100644
@@ -702,7 +702,13 @@ static __latent_entropy int dup_mmap(struct mm_struct *mm,
 	mas_destroy(&mas);
 out:
 	mmap_write_unlock(mm);
-	flush_tlb_mm(oldmm);
+	/*
+	 * poking_init() calls into here at early boot up.
+	 * At that time, there's no need to flush the tlb.
+	 * If we do, it will enable interrupts and cause a bug.
+	 */
+	if (likely(!early_boot_irqs_disabled))
+		flush_tlb_mm(oldmm);
 	mmap_write_unlock(oldmm);
 	dup_userfaultfd_complete(&uf);
 fail_uprobe_end:
lib/maple_tree.c +15 -1
index e1743803c851..e32206e840f6 100644
@@ -1253,7 +1253,21 @@ static inline void mas_alloc_nodes(struct ma_state *mas, gfp_t gfp)
 		}
 
 		max_req = min(requested, max_req);
-		count = mt_alloc_bulk(gfp, max_req, slots);
+
+		/*
+		 * text_poke() can be called very early, and it
+		 * calls dup_mm() which eventually leads down to here.
+		 * In that case, mt_alloc_bulk() will call kmem_cache_alloc_bulk()
+		 * which must be called with interrupts enabled. To avoid
+		 * doing that in early bootup, where interrupts must remain
+		 * disabled, just allocate a single slot.
+		 */
+		if (unlikely(early_boot_irqs_disabled)) {
+			slots[0] = mt_alloc_one(gfp | GFP_ATOMIC);
+			count = slots[0] ? 1 : 0;
+		} else {
+			count = mt_alloc_bulk(gfp, max_req, slots);
+		}
 		if (!count)
 			goto nomem_bulk;
 
Signature · 1 line
2.35.1
#1 lore raw
Quoted 4 lines · Steven Rostedt

This required some updates to fork and the maple_tree code to allow it to be called with enabling interrupts in the time when interrupts must remain disabled.

Yeah, moving special cases from one place to another doesn't really help. Particularly to something as core as dup_mm().

All of this comes from "poking_init()" being a steaming pile of bovine excrement, doing random odd things, and having that special "copy_init_mm()" helper that just makes things even worse. Nothing else uses that, and it shouldn't have called "dup_mm()" in the first place.

At this point, there is no actual user VM to even copy, so 99% of everything that duip_mm() does is not just pointless, but actively wrong, like the mmap_write_lock_nested() when we're in early boot.

I'm not even sure why "poking_mm" exists at all, and why it has created a whole new copy of "init_mm", and why this code isn't just using '&init_mm' like everything else that wants to just walk the kernel page tables.

Yes, I see that commit 4fc19708b165 ("x86/alternatives: Initialize temporary mm for patching"), and no, none of that makes any sense to me. It seems just (mis-)designed to fail.

Linus

author maintainer
#2 lore raw
Quoted 15 lines · On Mon, 24 Oct 2022 17:11:13 -0700 Linus Torvalds…
Quoted 4 lines · Steven Rostedt

This required some updates to fork and the maple_tree code to allow it to be called with enabling interrupts in the time when interrupts must remain disabled.

Yeah, moving special cases from one place to another doesn't really help. Particularly to something as core as dup_mm().

All of this comes from "poking_init()" being a steaming pile of bovine excrement, doing random odd things, and having that special "copy_init_mm()" helper that just makes things even worse. Nothing else uses that, and it shouldn't have called "dup_mm()" in the first place.

Quoted 8 lines

At this point, there is no actual user VM to even copy, so 99% of everything that duip_mm() does is not just pointless, but actively wrong, like the mmap_write_lock_nested() when we're in early boot.

I'm not even sure why "poking_mm" exists at all, and why it has created a whole new copy of "init_mm", and why this code isn't just using '&init_mm' like everything else that wants to just walk the kernel page tables.

It's not just walking the page tables, it's creating one that nobody else is using. Since we want to keep all executable code read only, the way text_poke() works is to create a new memory mapping where the pages it has isn't visible by anyone else (which is why it doesn't use init_mm). And then makes a mapping to the executable address as non executable and writable. Makes the update, and then removes the mapping.

Quoted 5 lines

Yes, I see that commit 4fc19708b165 ("x86/alternatives: Initialize temporary mm for patching"), and no, none of that makes any sense to me. It seems just (mis-)designed to fail.

It's all about updating read only pages that are executable with a shadow mm.

-- Steve

#3 lore raw
Quoted 2 lines · Steven Rostedt

It's all about updating read only pages that are executable with a shadow mm.

Right. And it doesn't actually need the mm at all, all it wants is the kernel page tables. Which is why all the "dup_mmap()" stuff seems so wrong.

I suspect mm_alloc() does everything that VM actually needs.

IOW, it shouldn't have used the fork() helper, it should have used the execve() helper that actually starts out from a clean slate. Because a clean slate is exactly what that code wants.

No?

Linus

author maintainer
#4 lore raw
Quoted 8 lines · On Mon, 24 Oct 2022 18:02:32 -0700 Linus Torvalds…
Quoted 2 lines · Steven Rostedt

It's all about updating read only pages that are executable with a shadow mm.

Right. And it doesn't actually need the mm at all, all it wants is the kernel page tables. Which is why all the "dup_mmap()" stuff seems so wrong.

Quoted 8 lines

I suspect mm_alloc() does everything that VM actually needs.

IOW, it shouldn't have used the fork() helper, it should have used the execve() helper that actually starts out from a clean slate. Because a clean slate is exactly what that code wants.

No?

Something to look into. But I'm guessing that's best for the next merge window, and not for the -rc releases?

-- Steve

#5 lore raw
Quoted 3 lines · Steven Rostedt

Something to look into. But I'm guessing that's best for the next merge window, and not for the -rc releases?

Yes., I just applied your v1 patch.

Linus

maintainer
#6 lore raw
In reply to Linus Torvalds
Quoted 5 lines · Linus Torvalds

All of this comes from "poking_init()" being a steaming pile of bovine excrement, doing random odd things, and having that special "copy_init_mm()" helper that just makes things even worse. Nothing else uses that, and it shouldn't have called "dup_mm()" in the first place.

Agreed; dup_mm() makes no sense and it is easily removed, see my earlier patch. Perhaps it can be simplified further to:

__poking_mm = init_mm

omitting the mm_init() I retained, but I need to stare harder at all that.

Quoted 4 lines

I'm not even sure why "poking_mm" exists at all, and why it has created a whole new copy of "init_mm", and why this code isn't just using '&init_mm' like everything else that wants to just walk the kernel page tables.

Because it instantiates user-space page-tables in it, you really don't want those in init_mm.

The whole (and sole) purpose of poking_mm is to contain the writable aliases. Only the CPU that has the poking_mm active has access to them.