| Message ID | 20210516040018.128105-1-madvenka@linux.microsoft.com (mailing list archive) |
|---|---|
| Headers | show |
| Series | arm64: Stack trace reliability checks in the unwinder | expand |
On Sat, May 15, 2021 at 11:00:16PM -0500, madvenka@linux.microsoft.com wrote: > Special cases > ============= > > Some special cases need to be mentioned: I think it'd be good if more of this cover letter, especially sections like this which cover the tricky bits, ended up in the code somehow - it's recorded here and will be in the list archive but that's not the most discoverable place so increases the maintainance burden. It'd be great to be able to compare the code directly with the reliable stacktrace requirements document and see everything getting ticked off, actually going all the way there might be too much and loose the code in the comments but I think we can get closer to it than we are. Given that a lot of this stuff rests on the denylist perhaps some comments just before it's called would be a good place to start? > - EL1 interrupt and exception handlers end up in sym_code_ranges[]. > So, all EL1 interrupt and exception stack traces will be considered > unreliable. This the correct behavior as interrupts and exceptions This stuff about exceptions and preemption is a big one, rejecting any exceptions makes a whole host of things easier (eg, Mark Rutland raised interactions between non-AAPCS code and PLTs as being an issue but if we're able to reliably reject stacks featuring any kind of preemption anyway that should sidestep the issue). > Performance > =========== > Currently, unwinder_blacklisted() does a linear search through > sym_code_functions[]. If reviewers prefer, I could sort the > sym_code_functions[] array and perform a binary search for better > performance. There are about 80 entries in the array. If people are trying to live patch a very busy/big system then this could be an issue, equally there's probably more people focused on getting boot times as fast as possible than live patching. Deferring the initialisation to first use would help boot times with or without sorting, without numbers I don't actually know that sorting is worth the effort or needs doing immediately - obvious correctness is also a benefit! My instinct is that for now it's probably OK leaving it as a linear scan and then revisiting if it's not adequately performant, but I'd defer to actual users there.
On 5/21/21 12:18 PM, Mark Brown wrote: > On Sat, May 15, 2021 at 11:00:16PM -0500, madvenka@linux.microsoft.com wrote: > >> Special cases >> ============= >> >> Some special cases need to be mentioned: > > I think it'd be good if more of this cover letter, especially sections > like this which cover the tricky bits, ended up in the code somehow - > it's recorded here and will be in the list archive but that's not the > most discoverable place so increases the maintainance burden. It'd be > great to be able to compare the code directly with the reliable > stacktrace requirements document and see everything getting ticked off, > actually going all the way there might be too much and loose the code in > the comments but I think we can get closer to it than we are. Given > that a lot of this stuff rests on the denylist perhaps some comments > just before it's called would be a good place to start? > I will add more comments in the code to make it clear. >> - EL1 interrupt and exception handlers end up in sym_code_ranges[]. >> So, all EL1 interrupt and exception stack traces will be considered >> unreliable. This the correct behavior as interrupts and exceptions > > This stuff about exceptions and preemption is a big one, rejecting any > exceptions makes a whole host of things easier (eg, Mark Rutland raised > interactions between non-AAPCS code and PLTs as being an issue but if > we're able to reliably reject stacks featuring any kind of preemption > anyway that should sidestep the issue). > Yes. I will include this in the code comments. >> Performance >> =========== > >> Currently, unwinder_blacklisted() does a linear search through >> sym_code_functions[]. If reviewers prefer, I could sort the >> sym_code_functions[] array and perform a binary search for better >> performance. There are about 80 entries in the array. > > If people are trying to live patch a very busy/big system then this > could be an issue, equally there's probably more people focused on > getting boot times as fast as possible than live patching. Deferring > the initialisation to first use would help boot times with or without > sorting, without numbers I don't actually know that sorting is worth the > effort or needs doing immediately - obvious correctness is also a > benefit! My instinct is that for now it's probably OK leaving it as a > linear scan and then revisiting if it's not adequately performant, but > I'd defer to actual users there. I have followed the example in the Kprobe deny list. I place the section in initdata so it can be unloaded during boot. This means that I need to copy the information before that in early_initcall(). If the initialization must be performed on first use, I probably have to move SYM_CODE_FUNCTIONS from initdata to some other place where it will be retained. If you prefer this, I could do it this way. Thanks! Madhavan
On Fri, May 21, 2021 at 12:32:52PM -0500, Madhavan T. Venkataraman wrote: > I have followed the example in the Kprobe deny list. I place the section > in initdata so it can be unloaded during boot. This means that I need to > copy the information before that in early_initcall(). > If the initialization must be performed on first use, I probably have to > move SYM_CODE_FUNCTIONS from initdata to some other place where it will > be retained. > If you prefer this, I could do it this way. No, I think if people are fine with this for kprobes they should be fine with it here too and if not we can always incrementally improve performance - let's just keep things simple and easy to understand for now.
On 5/21/21 12:47 PM, Mark Brown wrote: > On Fri, May 21, 2021 at 12:32:52PM -0500, Madhavan T. Venkataraman wrote: > >> I have followed the example in the Kprobe deny list. I place the section >> in initdata so it can be unloaded during boot. This means that I need to >> copy the information before that in early_initcall(). > >> If the initialization must be performed on first use, I probably have to >> move SYM_CODE_FUNCTIONS from initdata to some other place where it will >> be retained. > >> If you prefer this, I could do it this way. > > No, I think if people are fine with this for kprobes they should be fine > with it here too and if not we can always incrementally improve > performance - let's just keep things simple and easy to understand for > now. > OK. Madhavan
From: "Madhavan T. Venkataraman" <madvenka@linux.microsoft.com> There are a number of places in kernel code where the stack trace is not reliable. Enhance the unwinder to check for those cases and mark the stack trace as unreliable. Once all of the checks are in place, the unwinder can provide a reliable stack trace. But before this can be used for livepatch, some other entity needs to guarantee that the frame pointers are all set up correctly in kernel functions. objtool is currently being worked on to address that need. Return address check ==================== Check the return PC of every stack frame to make sure that it is a valid kernel text address (and not some generated code, for example). If it is not a valid kernel text address, mark the stack trace as unreliable. Assembly functions ================== There are a number of assembly functions in arm64. Except for a couple of them, these functions do not have a frame pointer prolog or epilog. Also, many of them manipulate low-level state such as registers. These functions are, by definition, unreliable from a stack unwinding perspective. That is, when these functions occur in a stack trace, the unwinder would not be able to unwind through them reliably. Assembly functions are defined as SYM_FUNC_*() functions or SYM_CODE_*() functions. objtool peforms static analysis of SYM_FUNC functions. It ignores SYM_CODE functions because they have low level code that is difficult to analyze. When objtool becomes ready eventually, SYM_FUNC functions will be analyzed and "fixed" as necessary. So, they are not "interesting" for the reliable unwinder. That leaves SYM_CODE functions. It is for the unwinder to deal with these for reliable stack trace. The unwinder needs to do the following: - Recognize SYM_CODE functions in a stack trace. - If a particular SYM_CODE function can be unwinded through using some special logic, then do it. E.g., the return trampoline for Function Graph Tracing. - Otherwise, mark the stack trace unreliable. Previous approach ================= Previously, I considered a per-section approach. The following sections mostly contain SYM_CODE functions: .entry.text .idmap.text .hyp.idmap.text .hyp.text .hibernate_exit.text .entry.tramp.text So, consider the whole sections unreliable. I created an array of code ranges, one for each section. The unwinder then checks each return PC against these sections. If there is a match, the stack trace is unreliable. Although this approach is reasonable, there are two problems: 1. There are SYM_FUNC functions in these sections. They also "become" unreliable in this approach. I think that we should be able to specify which functions are reliable and which ones are not in any section. 2. There are a few SYM_CODE functions in .text and .init.text. These sections contain tons of functions that are reliable. So, I considered moving the SYM_CODE functions into a separate section called ".code.text". I am not convinced that this approach is fine: - We should be able to place functions in .text and .init.text, if we wanted to. - Moving the hypervisor related functions caused reloc issues. - Moving things around just to satisfy the unwinder did not seem right. Current approach ================ In this version, I have taken a simpler approach to solve the issues mentioned before. Define an ARM64 version of SYM_CODE_END() like this: #define SYM_CODE_END(name) \ SYM_END(name, SYM_T_NONE) ;\ 99: ;\ .pushsection "sym_code_functions", "aw" ;\ .quad name ;\ .quad 99b ;\ .popsection NOTE: I use the numeric label 99 as the end address of a function. It does not conflict with anything right now. Please let me know if this is OK. I can add a comment that the label 99 is reserved and should not be used by assembly code. The above macro does the usual SYM_END(). In addition, it records the function's address range in a special section called "sym_code_functions". This way, all SYM_CODE functions get recorded in the section automatically. Implement an early_initcall() called init_sym_code_functions() that allocates an array called sym_code_functions[] and copies the function ranges from the section to the array. Implement an unwinder_blacklisted() function that compares a return PC to the ranges. If there is a match, return true. Else, return false. Call unwinder_blacklisted() on every return PC from unwind_frame(). If there is a match, then mark the stack trace as unreliable. Last stack frame ================ If a SYM_CODE function occurs in the very last frame in the stack trace, then the stack trace is not considered unreliable. This is because there is no more unwinding to do. Examples: - EL0 exception stack traces end in the top level EL0 exception handlers. - All kernel thread stack traces end in ret_from_fork(). Special SYM_CODE functions ========================== The return trampolines of the Function Graph Tracer and Kretprobe can be recognized by the unwinder. If the return PCs can be translated to the original PCs, then, the unwinder should perform that translation before checking for reliability. The final PC that we end up with after all the translations is the one we need to check for reliability. Accordingly, I have moved the reliability checks to after the logic that handles the Function Graph Tracer. So, the approach is - all SYM_CODE functions are unreliable. If a SYM_CODE function is "fixed" to make it reliable, then it should become a SYM_FUNC function. Or, if the unwinder has special logic to unwind through a SYM_CODE function, then that can be done. Previously, I had some extra code to handle the Function Graph Trace return trampoline case. I have removed it. Special casing is orthogonal to this work and can be done separately. Special cases ============= Some special cases need to be mentioned: - EL1 interrupt and exception handlers end up in sym_code_ranges[]. So, all EL1 interrupt and exception stack traces will be considered unreliable. This the correct behavior as interrupts and exceptions can happen on any instruction including ones in the frame pointer prolog and epilog. Unless objtool generates metadata so the unwinder can unwind through these special cases, such stack traces will be considered unreliable. - A task can get preempted at the end of an interrupt. Stack traces of preempted tasks will show the interrupt frame in the stack trace and will be considered unreliable. - Breakpoints are exceptions. So, all stack traces in the break point handler (including probes) will be considered unreliable. - All of the ftrace trampolines end up in sym_code_functions[]. All stack traces taken from tracer functions will be considered unreliable. Performance =========== Currently, unwinder_blacklisted() does a linear search through sym_code_functions[]. If reviewers prefer, I could sort the sym_code_functions[] array and perform a binary search for better performance. There are about 80 entries in the array. --- Changelog: v4: From Mark Brown: - I was checking the return PC with __kernel_text_address() before the Function Graph trace handling. Mark Brown felt that all the reliability checks should be performed on the original return PC once that is obtained. So, I have moved all the reliability checks to after the Function Graph Trace handling code in the unwinder. Basically, the unwinder should perform PC translations first (for rhe return trampoline for Function Graph Tracing, Kretprobes, etc). Then, the reliability checks should be applied to the resulting PC. - Mark said to improve the naming of the new functions so they don't collide with existing ones. I have used a prefix "unwinder_" for all the new functions. From Josh Poimboeuf: - In the error scenarios in the unwinder, the reliable flag in the stack frame should be set. Implemented this. - Some of the other comments are not relevant to the new code as I have taken a different approach in the new code. That is why I have not made those changes. E.g., Ard wanted me to add the "const" keyword to the global section array. That array does not exist in v4. Similarly, Mark Brown said to use ARRAY_SIZE() for the same array in a for loop. Other changes: - Add a new definition for SYM_CODE_END() that adds the address range of the function to a special section called "sym_code_functions". - Include the new section under initdata in vmlinux.lds.S. - Define an early_initcall() to copy the contents of the "sym_code_functions" section to an array by the same name. - Define a function unwinder_blacklisted() that compares a return PC against sym_code_sections[]. If there is a match, mark the stack trace unreliable. Call this from unwind_frame(). v3: - Implemented a sym_code_ranges[] array to contains sections bounds for text sections that contain SYM_CODE_*() functions. The unwinder checks each return PC against the sections. If it falls in any of the sections, the stack trace is marked unreliable. - Moved SYM_CODE functions from .text and .init.text into a new text section called ".code.text". Added this section to vmlinux.lds.S and sym_code_ranges[]. - Fixed the logic in the unwinder that handles Function Graph Tracer return trampoline. - Removed all the previous code that handles: - ftrace entry code for traced function - special_functions[] array that lists individual functions - kretprobe_trampoline() special case v2 - Removed the terminating entry { 0, 0 } in special_functions[] and replaced it with the idiom { /* sentinel */ }. - Change the ftrace trampoline entry ftrace_graph_call in special_functions[] to ftrace_call + 4 and added explanatory comments. - Unnested #ifdefs in special_functions[] for FTRACE. v1 - Define a bool field in struct stackframe. This will indicate if a stack trace is reliable. - Implement a special_functions[] array that will be populated with special functions in which the stack trace is considered unreliable. - Using kallsyms_lookup(), get the address ranges for the special functions and record them. - Implement an is_reliable_function(pc). This function will check if a given return PC falls in any of the special functions. If it does, the stack trace is unreliable. - Implement check_reliability() function that will check if a stack frame is reliable. Call is_reliable_function() from check_reliability(). - Before a return PC is checked against special_funtions[], it must be validates as a proper kernel text address. Call __kernel_text_address() from check_reliability(). - Finally, call check_reliability() from unwind_frame() for each stack frame. - Add EL1 exception handlers to special_functions[]. el1_sync(); el1_irq(); el1_error(); el1_sync_invalid(); el1_irq_invalid(); el1_fiq_invalid(); el1_error_invalid(); - The above functions are currently defined as LOCAL symbols. Make them global so that they can be referenced from the unwinder code. - Add FTRACE trampolines to special_functions[]: ftrace_graph_call() ftrace_graph_caller() return_to_handler() - Add the kretprobe trampoline to special functions[]: kretprobe_trampoline() Previous versions and discussion ================================ v3: https://lore.kernel.org/linux-arm-kernel/20210503173615.21576-1-madvenka@linux.microsoft.com/ v2: https://lore.kernel.org/linux-arm-kernel/20210405204313.21346-1-madvenka@linux.microsoft.com/ v1: https://lore.kernel.org/linux-arm-kernel/20210330190955.13707-1-madvenka@linux.microsoft.com/ Madhavan T. Venkataraman (2): arm64: Introduce stack trace reliability checks in the unwinder arm64: Create a list of SYM_CODE functions, blacklist them in the unwinder arch/arm64/include/asm/linkage.h | 12 ++++ arch/arm64/include/asm/sections.h | 1 + arch/arm64/include/asm/stacktrace.h | 4 ++ arch/arm64/kernel/stacktrace.c | 94 +++++++++++++++++++++++++++-- arch/arm64/kernel/vmlinux.lds.S | 7 +++ 5 files changed, 113 insertions(+), 5 deletions(-) base-commit: bf05bf16c76bb44ab5156223e1e58e26dfe30a88