Re: [Xen-devel] [RFC v3] xSplice design

[Martin Pohlack <mpohlack@xxxxxxxxxx>]

Hi,

I have a couple of high-level points here:

* Do you think it would be reasonable to have a minimal design without
  in-place patching, just using trampolines?  The examples discussed
  below suggest so.

* Regarding tboot integration: Do you plan to report hotpatching state
  to guests or do you want to keep that unchanged?

* If reporting:  Should we hash / sign the hotpatches in load order or
  should we work on an otherwise ordered set?

  In other words, do the sequences
  1: load + activate A, load + activate B
  2: load + activate B, load + activate A
  Result in the same system state and hash or not?

* Should we consider pairs of activation / deactivation events as
  no-ops?

  In other words, should hypervisor X and hypervisor X after activating
  and deactivating module A report the same system state and hash?

* What about auditing?  Currently the design discusses a method to
  query about the current state of affairs with regard to hotpatch
  modules.  Do we need something like a audit log for hotpatch
  operations?  We should at least report high-level operations that
  could impact the integrity on the console with a low threshold.

* There is a general (and mostly obscure) limitation on unloading
  hotpatches:

  In contrast to normal kernel modules where the module code adheres
  to specific conventions around resource allocation and locking,
  hotpatches typically contain code from any context.  That code is
  usually not aware that it can be unloaded.

  That code could leave behind in Xen references to itself, e.g., by
  installing a function pointer in a global data structure, without
  incrementing something like a usage count.  While most hotpatch code
  will probably be very simple and small, a similar effect could even
  be achieved by code called from the hotpatch in Xen, e.g., some code
  patch could dynamically generate a backtrace and later decide to
  inspect individual elements from the collected trace, later being a
  time, where the hotpatch has been unloaded again.

  One could approach that proplem from multiple angles: code
  inspection of generated hotpatches, testing, and by making unloading
  a very special and exceptional operation.


... and more inline comments below.

Regards,
Martin Pohlack

On 06.07.2015 22:26, Konrad Rzeszutek Wilk wrote:
> Since RFC v2 
> [http://lists.xen.org/archives/html/xen-devel/2015-05/msg02142.html]
>  - Ingested every review comment in.
> 
> For those who prefer an diff of what changed between v2 and this
> I am attaching an diff to help easy reviewing.
> 
> Please see inline the RFC v3 which in general:
>  - Ditches the attempt at defining an ELF payload using semi-Elf language
>    and just concentrates on structures.
>  - Expands on the preemption of the hypercalls
>  - Expands the implementation details with various topics that emerged
>    during v2 review
>  - Adds ASCII art (if you can call it that), and an example.
>  - state diagram the command hypercall.
> 
> # xSplice Design v1 (EXTERNAL RFC v3)
> 
> ## Rationale
> 
> A mechanism is required to binarily patch the running hypervisor with new
> opcodes that have come about due to primarily security updates.
> 
> This document describes the design of the API that would allow us to
> upload to the hypervisor binary patches.
> 
> The document is split in four sections:
>  - Detailed descriptions of the problem statement.
>  - Design of the data structures.
>  - Design of the hypercalls.
>  - Implementation notes that should be taken into consideration.
> 
> 
> ## Glossary
> 
>  * splice - patch in the binary code with new opcodes
>  * trampoline - a jump to a new instruction.
>  * payload - telemetries of the old code along with binary blob of the new
>    function (if needed).
>  * reloc - telemetries contained in the payload to construct proper 
> trampoline.
> 
> ## Multiple ways to patch
> 
> The mechanism needs to be flexible to patch the hypervisor in multiple ways
> and be as simple as possible. The compiled code is contiguous in memory with
> no gaps - so we have no luxury of 'moving' existing code and must either
> insert a trampoline to the new code to be executed - or only modify in-place
> the code if there is sufficient space. The placement of new code has to be 
> done
> by hypervisor and the virtual address for the new code is allocated 
> dynamically.
> 
> This implies that the hypervisor must compute the new offsets when splicing
> in the new trampoline code. Where the trampoline is added (inside
> the function we are patching or just the callers?) is also important.
> 
> To lessen the amount of code in hypervisor, the consumer of the API
> is responsible for identifying which mechanism to employ

The hypervisor at least needs to make sure that in-place patches fit in
the old place.

> and how many locations
> to patch. Combinations of modifying in-place code, adding trampoline, etc
> has to be supported. The API should allow read/write any memory within
> the hypervisor virtual address space.
> 
> We must also have a mechanism to query what has been applied and a mechanism
> to revert it if needed.

In-place modifications are more restrictive in when they can be applied
and need undo buffers in the size of the replacement code.  Trampolines
only need small fixed-size undo buffers (e.g., 5 bytes on x86).

If you have multiple levels of patching for a function, mixing the
different types can become tricky.  We need to think about where
in-place patches should go, always to the original function or also
potentially into replacement functions spliced via trampolines?

I feel that a pure trampoline mechanism could be enough and would result
in the smallest hypervisor code footprint.

In which scenarios would you prefer an in-place solution?

> We must also have a mechanism to: provide an copy of the old code - so that
> the hypervisor can verify it against the code in memory; the new code;
> the symbol name of the function to be patched; or offset from the symbol;
> or virtual address.

I am a bit confused by the "must" above:

If you consider my previous build ID suggestion as a hash over all
hypervisor code and you embed a designated build ID in each hotpatch
module together with a list of required previous modules, comparing the
two IDs and checking for required modules gives you a *safety* guarantee
of similar strength with a simpler implementation, right?

Providing a copy of the old code provides you with additional
flexibility as you can potentially load a specific hotpatch into more
Xen builds or you have more freedom regarding module load order.

We should think a bit about expected workflows of higher-level tools
that manage hotpatch stacks on production machines:

* The first obvious task is loading all available / suggested
  hotpatches around system start.
* Whenever new hotpatches are installed, they should be loaded too.
* One wants to query which modules have been loaded at runtime.
* If unloading is deemed safe (see comment far below), one may want to
  support a workflow where a specific hotpatch is marked as bad and
  unloaded.
* If we do no restrict module activation order and want to report tboot
state on sequences, we might have a complexity explosion problem, in
what system hashes should be considered acceptable.

> The complications that this design will encounter are explained later
> in this document.
> 
> ## Patching code
> 
> The first mechanism to patch that comes in mind is in-place replacement.
> That is replace the affected code with new code. Unfortunately the x86
> ISA is variable size which places limits on how much space we have available
> to replace the instructions.

I see no special limitation here for variable instruction size ISAs
except that you may have to fill the end of a smaller in-place patch
with nops.  You will have similar space trouble with all ISAs when the
replacement code is longer.

> The second mechanism is by replacing the call or jump to the
> old function with the address of the new function.
> 
> A third mechanism is to add a jump to the new function at the
> start of the old function.
> 
> ### Example of trampoline and in-place splicing
> 
> As example we will assume the hypervisor does not have XSA-132 (see
> *domctl/sysctl: don't leak hypervisor stack to toolstacks*
> 4ff3449f0e9d175ceb9551d3f2aecb59273f639d) and we would like to binary patch
> the hypervisor with it. The original code looks as so:
> 
> <pre>
>    48 89 e0                  mov    %rsp,%rax  
>    48 25 00 80 ff ff         and    $0xffffffffffff8000,%rax  
> </pre>
> 
> while the new patched hypervisor would be:
> 
> <pre>
>    48 c7 45 b8 00 00 00 00   movq   $0x0,-0x48(%rbp)  
>    48 c7 45 c0 00 00 00 00   movq   $0x0,-0x40(%rbp)  
>    48 c7 45 c8 00 00 00 00   movq   $0x0,-0x38(%rbp)  
>    48 89 e0                  mov    %rsp,%rax  
>    48 25 00 80 ff ff         and    $0xffffffffffff8000,%rax  
> </pre>
> 
> This is inside the arch_do_domctl. This new change adds 21 extra
> bytes of code which alters all the offsets inside the function. To alter
> these offsets and add the extra 21 bytes of code we might not have enough
> space in .text to squeze this in.

squeze -> squeeze

> As such we could simplify this problem by only patching the site
> which calls arch_do_domctl:
> 
> <pre>
> <do_domctl>:  
>  e8 4b b1 05 00          callq  ffff82d08015fbb9 <arch_do_domctl>  
> </pre>
> 
> with a new address for where the new `arch_do_domctl` would be (this
> area would be allocated dynamically).
> 
> Astute readers will wonder what we need to do if we were to patch `do_domctl`
> - which is not called directly by hypervisor but on behalf of the guests via
> the `compat_hypercall_table` and `hypercall_table`.
> Patching the offset in `hypercall_table` for `do_domctl:
> (ffff82d080103079 <do_domctl>:)
> <pre>
> 
>  ffff82d08024d490:   79 30  
>  ffff82d08024d492:   10 80 d0 82 ff ff   
> 
> </pre>
> with the new address where the new `do_domctl` is possible. The other
> place where it is used is in `hvm_hypercall64_table` which would need
> to be patched in a similar way. This would require an in-place splicing
> of the new virtual address of `arch_do_domctl`.
> 
> In summary this example patched the callee of the affected function by
>  * allocating memory for the new code to live in,
>  * changing the virtual address of all the functions which called the old

"of all" -> "in all" ?

>    code (computing the new offset, patching the callq with a new callq).
>  * changing the function pointer tables with the new virtual address of
>    the function (splicing in the new virtual address). Since this table
>    resides in the .rodata section we would need to temporarily change the
>    page table permissions during this part.
> 
> 
> However it has severe drawbacks - the safety checks which have to make sure
> the function is not on the stack - must also check every caller. For some
> patches this could if there were an sufficient large amount of callers

"could" -> "could mean"

> that we would never be able to apply the update.
> 
> ### Example of different trampoline patching.
> 
> An alternative mechanism exists where we can insert an trampoline in the
> existing function to be patched to jump directly to the new code. This
> lessens the locations to be patched to one but it puts pressure on the
> CPU branching logic (I-cache, but it is just one unconditional jump).
> 
> For this example we will assume that the hypervisor has not been compiled
> with fe2e079f642effb3d24a6e1a7096ef26e691d93e (XSA-125: *pre-fill structures
> for certain HYPERVISOR_xen_version sub-ops*) which mem-sets an structure
> in `xen_version` hypercall. This function is not called **anywhere** in
> the hypervisor (it is called by the guest) but referenced in the
> `compat_hypercall_table` and `hypercall_table` (and indirectly called
> from that). Patching the offset in `hypercall_table` for the old
> `do_xen_version` (ffff82d080112f9e <do_xen_version>)
> 
> </pre>
>  ffff82d08024b270 <hypercall_table>  
>  ...  
>  ffff82d08024b2f8:   9e 2f 11 80 d0 82 ff ff  
> 
> </pre>
> with the new address where the new `do_xen_version` is possible. The other
> place where it is used is in `hvm_hypercall64_table` which would need
> to be patched in a similar way. This would require an in-place splicing
> of the new virtual address of `do_xen_version`.
> 
> An alternative solution would be to patch insert an trampoline in the

"an trampoline" -> "a trampoline"

> old `do_xen_version' function to directly jump to the new `do_xen_version`.
> 
> <pre>
>  ffff82d080112f9e <do_xen_version>:  
>  ffff82d080112f9e:       48 c7 c0 da ff ff ff    mov    
> $0xffffffffffffffda,%rax  
>  ffff82d080112fa5:       83 ff 09                cmp    $0x9,%edi  
>  ffff82d080112fa8:       0f 87 24 05 00 00       ja     ffff82d0801134d2 
> <do_xen_version+0x534>  
> </pre>
> 
> with:
> 
> <pre>
>  ffff82d080112f9e <do_xen_version>:  
>  ffff82d080112f9e:       e9 XX YY ZZ QQ          jmpq   [new do_xen_version]  
> </pre>
> 
> which would lessen the amount of patching to just one location.
> 
> In summary this example patched the affected function to jump to the
> new replacement function which required:
>  * allocating memory for the new code to live in,
>  * inserting trampoline with new offset in the old function to point to the
>    new function.
>  * Optionally we can insert in the old function an trampoline jump to an 
> function

"an trampoline" -> "a trampoline"

>    providing an BUG_ON to catch errant code.

Where exactly in the old function would you put that second trampoline?

> The disadvantage of this are that the unconditional jump will consume a small
> I-cache penalty. However the simplicity of the patching of safety checks
> make this a worthwhile option.

I don't understand the "of safety checks".

> ### Security
> 
> With this method we can re-write the hypervisor - and as such we **MUST** be
> diligent in only allowing certain guests to perform this operation.
> 
> Furthermore with SecureBoot or tboot, we **MUST** also verify the signature
> of the payload to be certain it came from a trusted source.

+ "... and integrity was intact."

> 
> As such the hypercall **MUST** support an XSM policy to limit the what
> guest is allowed.

"to limit the what guest is allowed." -> "to limit what guest is allowed
to invoke it." ?

> If the system is booted with signature checking the
> signature checking will be enforced.

Are there any plans to extend signature reporting on hotpatch loading or
do you plan to hide that completely from guests?

What about a hotpatch loading and unloading sequence?  Do you consider
the hypervirsor to be in the same state with regard to its signature as
before the loading sequence and should it report the same signature
after such a pair of operations?

Do you plan for load order to be reflected in such updated signatures,
that is, do you plan the hash the set or sequence of hotpatches loaded?

> ## Design of payload format
> 
> The payload **MUST** contain enough data to allow us to apply the update
> and also safely reverse it. As such we **MUST** know:
> 
>  * What the old code is expected to be. We **MUST** be able verify it
>    against the runtime code.

See comment above about build IDs.

>  * The locations in memory to be patched. This can be determined dynamically
>    via symbols or via virtual addresses.
>  * The new code (or data) to will be patched in.

to -> that

>  * Signature to verify the payload.
> 
> This binary format can be constructed using an custom binary format but
> there are severe disadvantages of it:
> 
>  * The format might need to be change and we need an mechanism to accommodate

change -> changed

>    that.
>  * It has to be platform agnostic.
>  * Easily constructed using existing tools.
> 
> As such having the payload in an ELF file is the sensible way. We would be
> carrying the various set of structures (and data) in the ELF sections under

set -> sets

> different names and with definitions. The prefix for the ELF section name
> would always be: *.xsplice* to match up to the names of the structures.
> 
> Note that every structure has padding. This is added so that the hypervisor
> can re-use those fields as it sees fit.
> 
> Earlier design attempted to ineptly explain the relations of the ELF sections
> to each other without using proper ELF mechanism (sh_info, sh_link, data
> structures using Elf_* types, etc). This design will explain in detail
> the structures and how they are used together and not dig in the ELF
> format - except mention that the section names should match the
> structure names.
> 
> ### ASCII art of structures.
> 
> The diagram below is omitting some entries to easy the relationship 
> explanation.

ommitting -> omitting

> 
> <pre>
>                                                                           
> /---------------------\  
>                                                                        +->| 
> xsplice_reloc_howto |  
>                                                                       /   
> \---------------------/  
>                                                 /---------------\ 1:1/  
>                                              +->| xsplice_reloc |   /  
>                                             /   | - howto       +--/  1:1 
> /----------------\  
>                                            /    | - symbol      +-------->| 
> xsplice_symbol |  
>                                      1:N  /     \---------------/       / 
> \----------------/  
> /----------\        /--------------\     /                             /  
> | xsplice  |  1:1   | xsplice_code |    /                          1:1/  
> | - new    +------->|  - relocs    +---/  1:N   /-----------------\  /  
> | - old    +------->|  - sections  +----------->| xsplice_section | /  
> \----------/        |  - patches   +--\         | - symbol        +/ 1:1   
> /----------------\  
>                     \--------------/   \        | - addr          +------->| 
> .text or .data |  
>                                         \       \----------------/         
> \----------------/  
>                                          \  
>                                       1:N \  
>                                            \    /----------------\  
>                                             +-->| xsplice_patch  |  1:1  
> /----------------\  
>                                                 | - content      +------>| 
> binary code or |  
>                                                 \----------------/       | 
> data           |  
>                                                                          
> \----------------/  
> 
> </pre>
> 
> ### xsplice structures
> 
> From the top (or left in the above diagram) the structures are:
> 
>  *  `xsplice`. The top most structure - contains the the name of the update,
>     the id to match against the hypervisor, the pointer to the metadata for
>     the new code and optionally the metadata for the old code.
> 
>  * `xsplice_code`. The structure that ties all of this together and defines
>    the payload. Contains arrays of `xsplice_reloc`, `xsplice_section`, and
>    `xsplice_patch`.
> 
>  * `xsplice_reloc` contains telemtry used for patching - which describes the

telemtry -> telemetry

>    targets to be patched and how to do it.
> 
>  * `xsplice_section` - the safety data for the code. Contains pointer to the
>    symbol (`xsplice_symbols`) and pointer to the code (`.text`) or data 
> (`.data`),
>    which are to be used during safety and dependency checking.
> 
>  * `xsplice_patch`: the description of the new function to be patched in
>    along with the binary code or data.
> 
>  * ` xsplice_reloc_howto`: the howto properly construct trampolines for an 
> patch.
>    We may have multiple locations for which we need to insert an trampoline 
> for a
>    payload and each location might require a different way of handling it.
> 
>  * `xsplice_symbols `.  The symbol that will be patched.
> 
> In short the *.xsplice* sections (with `xsplice` being the top) represent
> various structures to define the new code and safety checks for the old
> code (optional). The ELF provides the mechanism to glue it all together when
> loaded in memory.

I am wondering if it would be useful to provide special call-once hooks
for activation and deactivation for scenarios where once-off data
transformations should happen when activating or deactivating a patch,
for example, for small changes to data structures.  Or for debug-type
modules that allow to dump or change some runtime state (debug level).

> Note that a lot of these ideas are borrowed from kSplice which is
> available at: https://github.com/jirislaby/ksplice
> 
> ### struct xsplice
> 
> The top most structure is quite simple. It defines the name, the id
> of the hypervisor, pointer to the new code and an pointer to
> the old code (optional).
> 
> The new code uses all of the `xsplice_*` structures while the
> old code does not use the `xsplice_reloc` structures.
> 
> The sections defining the structures will explicitly state
> when they are not used.
> 
> <pre>
> struct xsplice {
>     const char *name; /* A sensible name for the patch. Up to 40 characters. 
> */  

A version for the module might come in handy, depending on the tooling
around it.  This information could also be encoded in the name.

>     const char *id; /* ID of the hypervisor this binary was built against. */ 
>  

Should we define an ID type here with a fixed length?  Otherwise you
would need to rely on 0 termination and store the ID as hex string
instead of binary.

>     struct xsplice_code *new; /* Pointer to the new code to be patched. */  
>     struct xsplice_code *old; /* Pointer to the old code to be checked 
> against. */  
>     uint8_t pad[32];  /* Must be zero. */  
> };
> </pre>
> 
> The size of this structure should be 64 bytes.
> 
> ### xsplice_code
> 
> The structure embedded within this section ties the other
> structures together. It has the pointers with an start and end
> address for each set of structures. This means that an update
> can be split in multiple changes - for example to accomodate
> an update that contains both code and data and will need patching
> in both .text and .data sections.
> 
> <pre>
> struct xsplice_code {  
>     struct xsplice_reloc *relocs, *relocs_end; /* How to patch it. */  
>     struct xsplice_section *sections, *sections_end; /* Safety data. */  
>     struct xsplice_patch *patches, *patches_end; /* Patch code and data */  
>     uint8_t pad[16]; /* Must be zero. */
> };
> </pre>
> 
> The size of this structure is 64 bytes.
> 
> There can be at most two of those structures in the payload.
> One for the new code and another for the old code (optional).

You mentioned several times that you also want to support patching data
sections but restrict yourself here to code.  Would there be a use-case
for having the old-data that you expect to be there before applying changes?

> If it is for the old code the relocs, and relocs_end values will be ignored.
> 
> 
> ### xsplice_reloc
> 
> The `xsplice_code` defines an array of these structures. As such
> an singular structure defines an singular point where to patch the
> hypervisor.
> 
> The structure contains the address of the hypervisor (if known),
> the symbol associated with this address, how the patching is to
> be done, and platform specific details.
> 
> The `isns_added` is an value to be used to compute the new offset
> due to the quirks of the operands of the instruction. For example
> to patch in an jump operation to the new code - the offset is relative
> to the program counter of the next instruction - hence the offset
> value has to be subtracted by four bytes - hence this would contain -4 .
> 
> The `isns_target` is the offset against the symbol.
> 
> The relation of this structure with `xsplice_patch` is 1:1, even
> for inline patches. See the section detailing the structure
> `xsplice_reloc_howto`.
> 
> The relation of this structure with `xsplice_section` is 1:1.
> 
> This structure is as follow:
> 
> <pre>
> struct xsplice_reloc {  
>     uint64_t addr; /* The address of the relocation (if known). */  
>     struct xsplice_symbol *symbol; /* Symbol for this relocation. */  
>     int64_t isns_target; /* rest of the ELF addend.  This is equal to the 
> offset against the symbol that the relocation refers to. */  
>     struct xsplice_reloc_howto  *howto; /* Pointer to the above structure. */ 
>  
>     int64_t isns_added; /* ELF addend resulting from quirks of instruction 
> one of whose operands is the relocation. For example, this is -4 on x86 
> pc-relative jumps. */  
>     uint8_t pad[24];  /* Must be zero. */  
> };  
> 
> </pre>
> 
> The size of this structure is 64 bytes.
> 
> ### xsplice_section
> 
> The structure defined in this section is used during pre-patching and
> during patching. Pre-patching it is used to verify that it is safe
> to update with the new changes - and contains safety data on the old code
> and what kind of matching we are to expect.
> 
> That is whether the address (either provided or resolved when payload is
> loaded by referencing the symbols) is:
> 
>  * in memory,
>  * correct size,
>  * in it's proper ELF section,
>  * has been already patched (or not),
>  * is expected not to be the CPU stack - (or it is OK for it be on the CPU 
> stack).

"the CPU stack" -> "on any CPU stack"?

> with what we expect it to be.
> 
> Some of the checks can be relaxed, as such the `flag` values
> can be or-ed together.
> 
> <pre>
> 
> #define XSPLICE_SECTION_TEXT   0x00000001 /* Section is in .text */  
> #define XSPLICE_SECTION_RODATA 0x00000002 /* Section is in .rodata */  
> #define XSPLICE_SECTION_DATA   0x00000004 /* Section is in .data */  
> #define XSPLICE_SECTION_STRING 0x00000008 /* Section is in .str */  
> 
> #define XSPLICE_SECTION_TEXT_INLINE 0x00000200 /* Change is to be inline. */  
>  
> #define XSPLICE_SECTION_MATCH_EXACT 0x00000400 /* Must match exactly. */  
> #define XSPLICE_SECTION_NO_STACKCHECK 0x00000800 /* Do not check the stack. 
> */  

Depending on the time when patching is done, stack checking might not be
required.

For stack checking, do you need to compile with framepointers?

> struct xsplice_section {  
>     struct xsplice_symbol *symbol; /* The symbol associated with this change. 
> */  
>     uint64_t address; /* The address of the section (if known). */  
>     uint32_t size; /* The size of the section. */  
>     uint32_t flags; /* Various XSPLICE_SECTION_* flags. */
>     uint8_t pad[12]; /* To be zero. */  
> };
> 
> </pre>
> 
> The size of this structure is 32 bytes.
> 
> ### xsplice_patch
> 
> This structure has the binary code (or data) to be patched. Depending on the
> type it can either an inline patch (data or text) or require an relocation
> change (which requires an trampoline). Naturally it also points to a blob
> of the binary data to patch in, and the size of the patch.
> 
> The `addr` is used when the patch is for inline change. If it is an relocation
> (requiring an trampoline), the `addr` should be zero.

Why?  Do you expect inline patches not to cover whole functions?

Could you not still address the area of the inline relative to a symbol?

> There must be an corresponding ` struct xsplice_reloc` and
> `struct xsplice_section` describing this patch.
> 
> <pre>
> #define XSPLICE_PATCH_INLINE_TEXT   0x1
> #define XSPLICE_PATCH_INLINE_DATA   0x2
> #define XSPLICE_PATCH_RELOC_TEXT    0x3
> 
> struct xsplice_patch {  
>     uint32_t type; /* XSPLICE_PATCH_* .*/  
>     uint32_t size; /* Size of patch. */  
>     uint64_t addr; /* The address of the inline new code (or data). */  
>     void *content; /* The bytes to be installed. */  
>     uint8_t pad[40]; /* Must be zero. */  
> };
> 
> </pre>
> 
> The size of this structure is 64 bytes.
> 
> ### xsplice_symbols
> 
> The structure contains an pointer to the name of the ELF symbol
> to be patched and as well an unique name for the symbol.
> 
> The `label` is used for diagnostic purposes - such as including the
> name and the offset.
> 
> The structure is as follow:
> 
> <pre>
> struct xsplice_symbol {  
>     const char *name; /* The ELF name of the symbol. */  
>     const char *label; /* A unique xSplice name for the symbol. */  
>     uint8_t pad[16]; /* Must be zero. */  
> };  
> </pre>
> 
> The size of this structure is 32 bytes.
> 
> 
> ### xsplice_reloc_howto
> 
> The howto defines in the detail the change. It contains the type,
> whether the relocation is relative, the size of the relocation,
> bitmask for which parts of the instruction or data are to be replaced,
> amount the final relocation is shifted by (to drop unwanted data), and
> whether the replacement should be interpreted as signed value.
> 
> The structure is as follow:
> 
> <pre>
> #define XSPLICE_HOWTO_RELOC_INLINE  0x1 /* It is an inline replacement. */  
> #define XSPLICE_HOWTO_RELOC_PATCH   0x2 /* Add an trampoline. */  

_RELOC_ ...

> #define XSPLICE_HOWTO_FLAG_PC_REL    0x1 /* Is PC relative. */  
> #define XSPLICE_HOWOT_FLAG_SIGN      0x2 /* Should the new value be treated 
> as signed value. */  
> 
> struct xsplice_reloc_howto {  
>     uint32_t    type; /* XSPLICE_HOWTO_* */  

... refers to type?  Rename to relationship clear?

>     uint32_t    flag; /* XSPLICE_HOWTO_FLAG_* */  
>     uint32_t    size; /* Size, in bytes, of the item to be relocated. */  
>     uint32_t    r_shift; /* The value the final relocation is shifted right 
> by; used to drop unwanted data from the relocation. */  
>     uint64_t    mask; /* Bitmask for which parts of the instruction or data 
> are replaced with the relocated value. */  
>     uint8_t     pad[8]; /* Must be zero. */  
> };  
> 
> </pre>
> 
> The size of this structure is 32 bytes.
> 
> ### Example
> 
> There is a wealth of information that the payload must have to define a simple
> patch.  For this example we will assume that the hypervisor has not been 
> compiled
> with fe2e079f642effb3d24a6e1a7096ef26e691d93e (XSA-125: *pre-fill structures
> for certain HYPERVISOR_xen_version sub-ops*) which mem-sets an structure
> in `xen_version` hypercall. This function is not called **anywhere** in
> the hypervisor (it is called by the guest) but referenced in the
> `compat_hypercall_table` and `hypercall_table` (and indirectly called
> from that). There are two ways to patch this:
> inline patch `hvm_hypercall64_table` and `hvm_hypercall` with a new
> address for the new `do_xen_version` , or insert
> trampoline in `do_xen_version` code. The example will focus on the later.
> 
> The `do_xen_version` code is located at virtual address ffff82d080112f9e.
> 
> <pre>
> struct xsplice_code xsplice_xsa125;  
> struct xsplice_reloc relocs[1];  
> struct xsplice_section sections[1];  
> struct xsplice_patch patches[1];  
> struct xsplice_symbol do_xen_version_symbol;  
> struct xsplice_reloc_howto do_xen_version_howto;  
> char do_xen_version_new_code[1728];  
> 
> #ifndef HYPERVISOR_ID  
> #define HYPERVISOR_ID "92dd05a61556c554155b1508c9cf67d993336d28"
> #endif  
> 
> struct xsplice xsa125 = {  
>     .name = "xsa125",  
>     .id = HYPERVISOR_ID,  
>     .old = NULL,  
>     .new = &xsplice_xsa125,  
> };  
> 
> struct xsplice_code xsplice_xsa125 = {  
>     .relocs = &relocs[0],  
>     .relocs_end = &relocs[0],  

Odd to have end point to the beginning of the last element.  An "int
n_relocs = 1" feels more readable and does not have to be relocated.

Similar for all _end pointers below.

>     .sections = &sections[0],  
>     .sections_end = &sections[0],  
>     .patches = &patches[0],  
>     .patches_end = &patches[0],   
> };
> 
> struct xsplice_reloc relocs[1] = {  
>     {  
>         .addr = 0xffff82d080112f9e,  
>         .symbol = &do_xen_version_symbol,  
>         .isns_target = 0,  
>         .howto = &do_xen_version_howto,  
>         .isns_added = -4,  
>     },  
> };  
> 
> struct xsplice_symbol do_xen_version_symbol = {  
>     .name = "do_xen_version",  
>     .label = "do_xen_version+<0x0>",  
> };  
> 
> struct xsplice_reloc_howto do_xen_version_howto = {  
>     .type = XSPLICE_HOWTO_RELOC_PATCH,  
>     .flag = XSPLICE_HOWTO_FLAG_PC_REL,  
>     .r_shift = 0,  
>     .mask = (-1ULL),  
> };  
> 
> 
> struct xsplice_section sections[1] = {  
>     {  
>         .symbol = &do_xen_version_symbol,  
>         .address = 0xffff82d080112f9e,  
>         .size = 1728,  
>         .flags = XSPLICE_SECTION_TEXT,  
>     },  
> };  
> 
> struct xsplice_patch patches[1] = {  
>     {  
>         .type = XSPLICE_PATCH_RELOC_TEXT,  
>         .size = 1728,  
>         .addr = 0,  
>         .content = &do_xen_version_new_code,  
>     },  
> };  
> 
> char do_xen_version_new_code[1728] = { 0x83, 0xff, 0x09, /* And more code. 
> */};  
> </pre>

Nice example.

> ## Signature checking requirements.
> 
> The signature checking requires that the layout of the data in memory
> **MUST** be same for signature to be verified. This means that the payload
> data layout in ELF format **MUST** match what the hypervisor would be
> expecting such that it can properly do signature verification.
> 
> The signature is based on the all of the payloads continuously laid out
> in memory. The signature is to be appended at the end of the ELF payload
> prefixed with the string '~Module signature appended~\n", followed by

Unbalanced quotes break my syntax highlighting :-/

> an signature header then followed by the signature, key identifier, and 
> signers
> name.
> 
> Specifically the signature header would be:
> 
> <pre>
> #define PKEY_ALGO_DSA       0  
> #define PKEY_ALGO_RSA       1  
> 
> #define PKEY_ID_PGP         0 /* OpenPGP generated key ID */  
> #define PKEY_ID_X509        1 /* X.509 arbitrary subjectKeyIdentifier */  
> 
> #define HASH_ALGO_MD4          0  
> #define HASH_ALGO_MD5          1  
> #define HASH_ALGO_SHA1         2  
> #define HASH_ALGO_RIPE_MD_160  3  
> #define HASH_ALGO_SHA256       4  
> #define HASH_ALGO_SHA384       5  
> #define HASH_ALGO_SHA512       6  
> #define HASH_ALGO_SHA224       7  
> #define HASH_ALGO_RIPE_MD_128  8  
> #define HASH_ALGO_RIPE_MD_256  9  
> #define HASH_ALGO_RIPE_MD_320 10  
> #define HASH_ALGO_WP_256      11  
> #define HASH_ALGO_WP_384      12  
> #define HASH_ALGO_WP_512      13  
> #define HASH_ALGO_TGR_128     14  
> #define HASH_ALGO_TGR_160     15  
> #define HASH_ALGO_TGR_192     16  
> 
> 
> struct elf_payload_signature {  
>       u8      algo;           /* Public-key crypto algorithm PKEY_ALGO_*. */  
>       u8      hash;           /* Digest algorithm: HASH_ALGO_*. */  
>       u8      id_type;        /* Key identifier type PKEY_ID*. */  
>       u8      signer_len;     /* Length of signer's name */  
>       u8      key_id_len;     /* Length of key identifier */  
>       u8      __pad[3];  
>       __be32  sig_len;        /* Length of signature data */  
> };
> 
> </pre>
> (Note that this has been borrowed from Linux module signature code.).
> 
> 
> ## Hypercalls
> 
> We will employ the sub operations of the system management hypercall (sysctl).
> There are to be four sub-operations:
> 
>  * upload the payloads.
>  * listing of payloads summary uploaded and their state.
>  * getting an particular payload summary and its state.
>  * command to apply, delete, or revert the payload.

Runtime querying of the hypervisor ID of the currently active hypervisor
could also be part of this design.

> The actions are asynchronous therefore the caller is responsible

"listing" and "getting payload summary" are probably not asynchronous.

> to verify that it has been applied properly by retrieving the summary of it
> and verifying that there are no error codes associated with the payload.
> 
> We **MUST** make it asynchronous due to the nature of patching: it requires
> every physical CPU to be lock-step with each other. The patching mechanism
> while an implementation detail, is not an short operation and as such
> the design **MUST** assume it will be an long-running operation.

I disagree with the strong wording here, but we had some discussions
about this before.  The specific quiescencing mechanism plays a role here.

> The sub-operations will spell out how preemption is to be handled (if at all).
> 
> Furthermore it is possible to have multiple different payloads for the same
> function. As such an unique id has to be visible to allow proper manipulation.

Unique ID per hotpatch?

> The hypercall is part of the `xen_sysctl`. The top level structure contains
> one uint32_t to determine the sub-operations:
> 
> <pre>
> struct xen_sysctl_xsplice_op {  
>     uint32_t cmd;  
>       union {  
>           ... see below ...  
>         } u;  
> };  
> 
> </pre>
> while the rest of hypercall specific structures are part of the this 
> structure.
> 
> ### XEN_SYSCTL_XSPLICE_UPLOAD (0)
> 
> Upload a payload to the hypervisor. The payload is verified and if there
> are any issues the proper return code will be returned. The payload is
> not applied at this time - that is controlled by *XEN_SYSCTL_XSPLICE_ACTION*.
> 
> The caller provides:
> 
>  * `id` unique id.

The hypervisor could return a unique ID as return value of the operation
for future reference.  How can the caller generate a unique ID?

After some thinking about this:

* You probably have something like a UUID in mind that stays fixed
  forever for a given hotpatch module.  Maybe a hash-like build ID of
  the module itself.  It could be embedded into the module at creation
  time.

  It could also be used for specifying hotpatch module
  interdependencies.  That would be nice.

* I had a transient ID in mind, that the hypervisor could simply create
  by counting up a 64-bit value and that would only be used for
  addressing specific payloads at runtime, but not necessarily uniquely
  over time and accross different machines.

>  * `payload` the virtual address of where the ELF payload is.
> 
> The return value is zero if the payload was succesfully uploaded and the
> signature was verified. Otherwise an EXX return value is provided.

That sounds synchronous.

> Duplicate `id` are not supported.
> 
> The `payload` is the ELF payload as mentioned in the `Payload format` section.
> 
> This operation can be preempted by the hypercall returning EAGAIN.

Would the client be expected to submit the request again or should he
spin a bit and query for the results?

> This is due to the nature of signature verification - which may require
> SecureBoot firmware calls which are unbounded.
> 
> The structure is as follow:
> 
> <pre>
> struct xen_sysctl_xsplice_upload {  
>     char id[40];  /* IN, name of the patch. */

Does this have to be unique?  Earlier there was a distinction between
the patch name and a request ID.

I was a bit confused about the different IDs at play here, I suggest to
introduce a term (e.g., payload ID) for the IDs referring to a specific
payload.  It seems you expect the ID for a given payload to remain
constant, ideally over time?

Where does the size 40 come from?

Inspecting the common ld --buil-id options, sha1 with 160 bits is the
largest and can be represented with 40 characters in ASCII
representation or 20 bytes binary.

sha-256 would also fit in binary representation with 32 bytes.

Is there more behind this?

>     uint64_t size; /* IN, size of the ELF file. */
>     XEN_GUEST_HANDLE_64(uint8) payload; /* ELF file. */  
> }; 
> </pre>
> 
> ### XEN_SYSCTL_XSPLICE_GET (1)
> 
> Retrieve an summary of an specific payload. This caller provides:
> 
>  * `id` the unique id.
>  * `status` *MUST* be set to zero.
> 
> The `summary` structure contains an summary of payload which includes:
> 
>  * `id` the unique id.
>  * `status` - whether it has been:
>  1. *XSPLICE_STATUS_LOADED* (0) has been loaded.
>  2. *XSPLICE_STATUS_PROGRESS* (1) acting on the **XEN_SYSCTL_XSPLICE_ACTION** 
> command.
>  3. *XSPLICE_STATUS_CHECKED*  (2) the ELF payload safety checks passed.
>  4. *XSPLICE_STATUS_APPLIED* (3) loaded, checked, and applied.
>  5. *XSPLICE_STATUS_REVERTED* (4) loaded, checked, applied and then also 
> reverted.
>  6. Negative values is an error. The error would be of EXX format.
> 
> The return value is zero on success and EXX on failure. This operation
> is synchronous and does not require preemption.
> 
> The structure is as follow:
> 
> <pre>
> #define XSPLICE_STATUS_LOADED    0  
> #define XSPLICE_STATUS_PROGRESS  1  
> #define XSPLICE_STATUS_CHECKED   2  
> #define XSPLICE_STATUS_APPLIED   3  
> #define XSPLICE_STATUS_REVERTED  4  
> 
> struct xen_sysctl_xsplice_summary {  
>     char id[40];  /* IN/OUT, name of the patch. */  

ID or name (comment)?

>     int32_t status;   /* OUT */  
> }; 
> </pre>
> 
> ### XEN_SYSCTL_XSPLICE_LIST (2)
> 
> Retrieve an array of abbreviated summary of payloads that are loaded in the
> hypervisor.
> 
> The caller provides:
> 
>  * `version`. Initially it *MUST* be zero.
>  * `idx` index iterator. Initially it *MUST* be zero.
>  * `count` the max number of entries to populate.
>  * `summary` virtual address of where to write payload summaries.
> 
> The hypercall returns zero on success and updates the `idx` (index) iterator
> with the number of payloads returned, `count` to the number of remaining
> payloads, and `summary` with an number of payload summaries. The `version`
> is updated on every hypercall

It remains a bit unclear what is updated when.  I am guessing:

  Each operation in Xen that mutates relevant payload state updates the
  single global payload-version field in Xen.

  Each invocation of XEN_SYSCTL_XSPLICE_LIST will also return a current
  copy thereoff.  An acquired snapshot in userland is only consistent
  if all individual operations returned the same version.

> - if it varies from one hypercall to another
> the data is stale and further calls could fail.

... or result in an inconsistent snapshot even if returning succesfully?

> If the hypercall returns E2BIG the `count` is too big and should be
> lowered.
> 
> Note that due to the asynchronous nature of hypercalls the domain might have
> added or removed the number of payloads making this information stale. It is
> the responsibility of the toolstack to use the `version` field to check
> between each invocation.
> 
> This operation is synchronous and does not require preemption.
> 
> The `summary` structure contains an summary of payload which includes:
> 
>  * `version` version of the data.
>  * `id` unique id.

Per payload?

>  * `status` - whether it has been:
>  1. *XSPLICE_STATUS_LOADED* (0) has been loaded.
>  2. *XSPLICE_STATUS_PROGRESS* (1) acting on the **XEN_SYSCTL_XSPLICE_ACTION** 
> command.
>  3. *XSPLICE_STATUS_CHECKED*  (2) the ELF payload safety checks passed.
>  4. *XSPLICE_STATUS_APPLIED* (3) loaded, checked, and applied.
>  5. *XSPLICE_STATUS_REVERTED* (4) loaded, checked, applied and then also 
> reverted.
>  6. Any negative values means there has been error. The value is in EXX 
> format.
> 
> The structure is as follow:
> 
> <pre>
> struct xen_sysctl_xsplice_list {  
>     uint32_t version; /* OUT */
>     uint32_t idx;  /* IN/OUT */  
>     uint32_t count;  /* IN/OUT */
>     XEN_GUEST_HANDLE_64(xen_sysctl_xsplice_summary) summary;  /* OUT */  
> };  
> 
> struct xen_sysctl_xsplice_summary {  
>     char id[40];  /* OUT, name of the patch. */  

Again, name or module ID?

>     int32_t status;  /* OUT */  
> };  
> 
> </pre>
> ### XEN_SYSCTL_XSPLICE_ACTION (3)
> 
> Perform an operation on the payload structure referenced by the `id` field.
> The operation request is asynchronous and the status should be retrieved
> by using either **XEN_SYSCTL_XSPLICE_GET** or **XEN_SYSCTL_XSPLICE_LIST** 
> hypercall.
> 
> There are two ways about doing preemption. Either via returning back EBUSY
> or the mechanism outlined here.

What do you mean by preemption?  Are you referring to quiescing Xen?

> Doing it in userland would remove any tracking of states in
> the hypervisor - except the simple commands apply, unload, and revert.

?

> However we would not be able to patch all the code that is invoked while
> this hypercall is in progress. That is - the do_domctl, the spinlocks,
> anything put on the stack, etc.
> 
> The disadvantage of the mechanism outlined here is that the hypervisor
> code has to keep the state atomic and have an upper bound of time
> on actions. If within the time the operation does not succeed the
> operation would go in error state.
> 
>  * `id` the unique id.
>  * `time` the upper bound of time the cmd should take. Zero means infinite.
>  * `cmd` the command requested:
>   1. *XSPLICE_ACTION_CHECK* (1) check that the payload will apply properly.
>   2. *XSPLICE_ACTION_UNLOAD* (2) unload the payload.
>    Any further hypercalls against the `id` will result in failure unless
>    **XEN_SYSCTL_XSPLICE_UPLOAD** hypercall is perfomed with same `id`.
>   3. *XSPLICE_ACTION_REVERT* (3) revert the payload. If the operation takes
>   more time than the upper bound of time the `status` will EBUSY.
>   4. *XSPLICE_ACTION_APPLY* (4) apply the payload. If the operation takes
>   more time than the upper bound of time the `status` will be EBUSY.
>   5. *XSPLICE_ACTION_LOADED* is an initial state and cannot be requested.  
> 
> The return value will be zero unless the provided fields are incorrect.
> 
> The structure is as follow:
> 
> <pre>
> #define XSPLICE_ACTION_LOADED 0  
> #define XSPLICE_ACTION_CHECK  1  
> #define XSPLICE_ACTION_UNLOAD 2  
> #define XSPLICE_ACTION_REVERT 3  
> #define XSPLICE_ACTION_APPLY  4  
> 
> struct xen_sysctl_xsplice_action {  
>     char id[40];  /* IN, name of the patch. */  
>     uint64_t time; /* IN, upper bound of time (ms) for the operation to take. 
> */  
>     uint32_t cmd; /* IN */  
> };  
> 
> </pre>
> 
> ## State diagrams of XSPLICE_ACTION values.
> 
> There is a strict ordering state of what the commands can be.
> The XSPLICE_ACTION prefix has been dropped to easy reading:
> 
> <pre>
>                         /->\  
>                         \  /  
>              /-------< CHECK <--------\  
>              |          |             |  
>              |           +            /  
>              |    +--->UNLOAD<--\    /  
>              |   /               \  /  
>              | /                  \/  
>       /-> APPLY -----------> REVERT --\  
>       |                               |  
>       \-------------------------------/  
> 
> </pre>
> Or an state transition table of valid states:
> <pre>
> +-------+-------+--------+--------+---------+-------+------------------+  
> | CHECK | APPLY | REVERT | UNLOAD | Current | Next  | Result           |  
> +-------+-------+--------+--------+---------+-------+------------------+  
> |   x   |       |        |        | LOADED  | CHECK | Check payload.   |  
> +-------+-------+--------+--------+---------+-------+------------------+  
> |   x   |       |        |        | CHECK   | CHECK | Check payload.   |  
> +-------+-------+--------+--------+---------+-------+------------------+  
> |       |   x   |        |        | CHECK   | APPLY | Apply payload.   |  
> +-------+-------+--------+--------+---------+-------+------------------+  
> |       |       |        |   x    | CHECK   | UNLOAD| Unload payload.  |  
> +-------+-------+--------+--------+---------+-------+------------------+  
> |       |       |   x    |        | APPLY   | REVERT| Revert payload.  |  
> +-------+-------+--------+--------+---------+-------+------------------+  
> |       |       |        |   x    | APPLY   | UNLOAD| unload payload.  |  
> +-------+-------+--------+--------+---------+-------+------------------+  
> |       |   x   |        |        | REVERT  | APPLY | Apply payload.   |  
> +-------+-------+--------+--------+---------+-------+------------------+  
> </pre>
> All the other states are invalid.

"states" -> "state transistions" ?

> ## Sequence of events.
> 
> The normal sequence of events is to:
> 
>  1. *XEN_SYSCTL_XSPLICE_UPLOAD* to upload the payload. If there are errors 
> *STOP* here.
>  2. *XEN_SYSCTL_XSPLICE_GET* to check the `->status`. If in 
> *XSPLICE_STATUS_PROGRESS* spin. If in *XSPLICE_STATUS_LOADED* go to next step.
>  3. *XEN_SYSCTL_XSPLICE_ACTION* with *XSPLICE_ACTION_CHECK* command to verify 
> that the payload can be succesfully applied.
>  4. *XEN_SYSCTL_XSPLICE_GET* to check the `->status`. If in 
> *XSPLICE_STATUS_PROGRESS* spin. If in *XSPLICE_STATUS_CHECKED* go to next 
> step.
>  5. *XEN_SYSCTL_XSPLICE_ACTION* with *XSPLICE_ACTION_APPLY* to apply the 
> patch.
>  6. *XEN_SYSCTL_XSPLICE_GET* to check the `->status`. If in 
> *XSPLICE_STATUS_PROGRESS* spin. If in *XSPLICE_STATUS_APPLIED* exit with 
> success.
> 
> ## Addendum
> 
> Implementation quirks should not be discussed in a design document.
> 
> However these observations can provide aid when developing against this
> document.
> 
> 
> ### Alternative assembler
> 
> Alternative assembler is a mechanism to use different instructions depending
> on what the CPU supports. This is done by providing multiple streams of code
> that can be patched in - or if the CPU does not support it - padded with
> `nop` operations. The alternative assembler macros cause the compiler to
> expand the code to place a most generic code in place - emit a special
> ELF .section header to tag this location. During run-time the hypervisor
> can leave the areas alone or patch them with an better suited opcodes.
> 
> However these sections are part of .init. and as such can't reasonably be
> subject to patching.
> 
> ### .rodata sections
> 
> The patching might require strings to be updated as well. As such we must be
> also able to patch the strings as needed. This sounds simple - but the 
> compiler
> has a habit of coalescing strings that are the same - which means if we 
> in-place
> alter the strings - other users will be inadvertently affected as well.
> 
> This is also where pointers to functions live - and we may need to patch this
> as well.

switch-style jump tables sometimes too.

> To guard against that we must be prepared to do patching similar to
> trampoline patching or in-line depending on the flavour. If we can
> do in-line patching we would need to:
> 
>  * alter `.rodata` to be writeable.
>  * inline patch.
>  * alter `.rodata` to be read-only.
> 
> If are doing trampoline patching we would need to:
> 
>  * allocate a new memory location for the string.
>  * all locations which use this string will have to be updated to use the
>    offset to the string.
>  * mark the region RO when we are done.
> 
> ### .bss and .data sections.
> 
> Patching writable data is not suitable as it is unclear what should be done
> depending on the current state of data. As such it should not be attempted.
> 
> 
> ### Patching code which is in the stack.
> 
> We should not patch the code which is on the stack. That can lead
> to corruption.

To clarify:
* Are you referreing to code that is generated on actual stacks (where
  is that done in Xen?), or
* Code which is target to return addresses that are live in specific
  CPU stacks?

> ### Trampoline (e9 opcode)
> 
> The e9 opcode used for jmpq uses a 32-bit signed displacement. That means
> we are limited to up to 2GB of virtual address to place the new code
> from the old code. That should not be a problem since Xen hypervisor has
> a very small footprint.
> However if we need - we can always add two trampolines. One at the 2GB
> limit that calls the next trampoline.

There is a small limitation for trampolines in function entries: The
target function (+ trailing padding) must be able to accomodate the
trampoline.  On x86 with +-2 GB relative jumps, this means 5 bytes are
required.

Depending on compiler settings, there are several functions in Xen that
are smaller (without inter-function padding).

readelf -sW xen-syms | grep " FUNC " | \
    awk '{ if ($3 < 5) print $3, $4, $5, $8 }'

...
3 FUNC LOCAL wbinvd_ipi
3 FUNC LOCAL shadow_l1_index
...

A compile-time check for, e.g., a minimum alignment of functions or a
runtime check that verifies symbol size (+ padding to next symbols) for
that in the hypervisor seems advised.

> ### When to patch
> 
> During the discussion on the design two candidates bubbled where
> the call stack for each CPU would be deterministic. This would
> minimize the chance of the patch not being applied due to safety
> checks failing.
> 
> #### Rendezvous code instead of stop_machine for patching
> 
> The hypervisor's time rendezvous code runs synchronously across all CPUs
> every second. Using the stop_machine to patch can stall the time rendezvous
> code and result in NMI. As such having the patching be done at the tail
> of rendezvous code should avoid this problem.

Could you indicate what functions you would expect to be live at this
point in each CPU?  Can the rendezvous code interrupt other hypervisor
operations or is that entered via the softirq mechanism close to return
from hypervisor code?

> #### Before entering the guest code.
> 
> Before we call VMXResume we check whether any soft IRQs need to be executed.
> This is a good spot because all Xen stacks are effectively empty at
> that point.
> 
> To randezvous all the CPUs an barrier with an maximum timeout (which
> could be adjusted), combined with forcing all other CPUs through the
> hypervisor with IPIs, can be utilized to have all the CPUs be lockstep.
> 
> The approach is similar in concept to stop_machine and the time rendezvous
> but is time-bound.
> 
> ### Compiling the hypervisor code
> 
> Hotpatch generation often requires support for compiling the target
> with -ffunction-sections / -fdata-sections.  Changes would have to
> be done to the linker scripts to support this.
> 
> 
> ### Generation of xSplice ELF payloads
> 
> The design of that is not discussed in this design.
> 
> The author of this design envisions objdump and objcopy along
> with special GCC parameters (see above) to create .o.xsplice files
> which can be used to splice an ELF with the new payload.

I would expect ksplice userland code to be a good source of inspiration
at least.

> ### Exception tables and symbol tables growth
> 
> We may need support for adapting or augmenting exception tables if
> patching such code.  Hotpatches may need to bring their own small
> exception tables (similar to how Linux modules support this).
> 
> If supporting hotpatches that introduce additional exception-locations
> is not important, one could also change the exception table in-place
> and reorder it afterwards.
> 
> 
> ### xSplice interdependencies
> 
> xSplice patches interdependencies are tricky.
> 
> There are the ways this can be addressed:
>  * A single large patch that subsumes and replaces all previous ones.
>    Over the life-time of patching the hypervisor this large patch
>    grows to accumulate all the code changes.
>  * Hotpatch stack - where an mechanism exists that loads the hotpatches
>    in the same order they were built in. We would need an build-id
>    of the hypevisor to make sure the hot-patches are build against the
>    correct build.
>  * Payload containing the old code to check against that. That allows
>    the hotpatches to be loaded indepedently (if they don't overlap) - or
>    if the old code also containst previously patched code - even if they
>    overlap.
> 
> The disadvantage of the first large patch is that it can grow over
> time and not provide an bisection mechanism to identify faulty patches.
> 
> The hot-patch stack puts stricts requirements on the order of the patches
> being loaded and requires an hypervisor build-id to match against.
> 
> The old code allows much more flexibility and an additional guard,
> but is more complex to implement.
> 
> ### Hypervisor ID (buid-id)
> 
> The build-id can help with:
> 
>   * Prevent loading of wrong hotpatches (intended for other builds)
> 
>   * Allow to identify suitable hotpatches on disk and help with runtime
>     tooling (if laid out using build ID)
> 
> The build-id (aka hypervisor id) can be easily obtained by utilizing
> the ld --build-id operatin which (copied from ld):
> 
> <pre>
> --build-id  
>     --build-id=style  
>         Request creation of ".note.gnu.build-id" ELF note section.  The 
> contents of the note are unique bits identifying this  
>         linked file.  style can be "uuid" to use 128 random bits, "sha1" to 
> use a 160-bit SHA1 hash on the normative parts of the  
>         output contents, "md5" to use a 128-bit MD5 hash on the normative 
> parts of the output contents, or "0xhexstring" to use a  
>         chosen bit string specified as an even number of hexadecimal digits 
> ("-" and ":" characters between digit pairs are  
>         ignored).  If style is omitted, "sha1" is used.  
> 
>         The "md5" and "sha1" styles produces an identifier that is always the 
> same in an identical output file, but will be  
>         unique among all nonidentical output files.  It is not intended to be 
> compared as a checksum for the file's contents.  A  
>         linked file may be changed later by other tools, but the build ID bit 
> string identifying the original linked file does  
>         not change.  
> 
>         Passing "none" for style disables the setting from any "--build-id" 
> options earlier on the command line.  
> 
> </pre>
> 
> ### Symbol names
> 
> 
> Xen as it is now, has a couple of non-unique symbol names which will
> make runtime symbol identification hard.  Sometimes, static symbols
> simply have the same name in C files, sometimes such symbols get
> included via header files, and some C files are also compiled
> multiple times and linked under different names (guest_walk.c).
> 
> As such we need to modify the linker to make sure that the symbol
> table qualifies also symbols by their source file name.

The convention for file-type symbols (that would allow to map many
symbols to their compilation unit) says that only the basename (i.e.,
without directories) is embedded.  This creates another layer of
confusion for duplicate file names in the build tree.

> find . -name \*.c -print0 | xargs -0 -n1 basename | sort | uniq -c | sort -n 
> | tail -n10
      3 shutdown.c
      3 sysctl.c
      3 time.c
      3 xenoprof.c
      4 gdbstub.c
      4 irq.c
      5 domain.c
      5 mm.c
      5 pci.c
      5 traps.c

> For the awkward situations in which C-files are compiled multiple
> times patches we would need to some modification in the Xen code.
> 
> ### Security
> 
> Only the privileged domain should be allowed to do this operation.
> 

Amazon Development Center Germany GmbH
Krausenstr. 38
10117 Berlin
Geschaeftsfuehrer: Dr. Ralf Herbrich, Christian Schlaeger
Ust-ID: DE289237879
Eingetragen am Amtsgericht Charlottenburg HRB 149173 B


_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxx
http://lists.xen.org/xen-devel