Spelunking Linux – what is this auxv thing anyway

While spelunking Linux, trying to find an easier way to do this or that, I ran across this vDSO thing (virtual ELF Dynamic Shared Object). What?

Ok, it’s like this. I was implementing direct calling of syscalls on Linux, and reading up on how the C libraries do things. On Linux, when you wan to talk to the kernel, you typically go through syscalls, or ioctl calls, or netlinks. With syscall, it’s actually a fairly expensive process, switching from userspace to kernel space, issuing and responding to interrupts, etc. In some situations this could be a critical performance hit. So, to make things easier/faster, some of the system calls are implemented in this little ELF package (vDSO). This little elf package (a dynamic link library) is loaded into every application on Linux. Then, the C library can decide to make calls into that library, instead of syscalls, thus saving a lot of overhead and speeding things up. Not all systems have this capability, but many do.

Alright, so how does the C runtime know whether the capability is there or now, and where this little library is, and how to get at functions therewith? In steps our friend auxv. In the GNU C library, there is a lone call:

unsigned long getauxval(unsigned long);

What values can you get out of this thing? Well, the constants can be found in the elf.h file, and look like:

#define AT_PLATFORM 15
#define AT_PAGESZ 6

And about 30 others. How you use each of these depends on the type of the value that you are looking up. For instance, the AT_PLATFORM returns a pointer to a null terminated string. The AT_PAGESZ returns an integer which represents the memory page size of the machine you’re running on.

OK, so what’s the lua version?

ffi.cdef[[
static const int AT_NULL = 0;
static const int AT_IGNORE = 1;
static const int AT_EXECFD = 2;
static const int AT_PHDR = 3;
static const int AT_PHENT = 4;
static const int AT_PHNUM = 5;
static const int AT_PAGESZ = 6;
static const int AT_BASE = 7;
static const int AT_FLAGS = 8;
static const int AT_ENTRY = 9;
static const int AT_NOTELF = 10;
static const int AT_UID = 11;
static const int AT_EUID = 12;
static const int AT_GID = 13;
static const int AT_EGID = 14;
static const int AT_CLKTCK = 17;
static const int AT_PLATFORM = 15;
static const int AT_HWCAP = 16;
static const int AT_FPUCW = 18;
static const int AT_DCACHEBSIZE = 19;
static const int AT_ICACHEBSIZE = 20;
static const int AT_UCACHEBSIZE = 21;
static const int AT_IGNOREPPC = 22;
static const int AT_SECURE = 23;
static const int AT_BASE_PLATFORM = 24;
static const int AT_RANDOM = 25;
static const int AT_HWCAP2 = 26;
static const int AT_EXECFN = 31;
static const int AT_SYSINFO = 32;
static const int AT_SYSINFO_EHDR = 33;
static const int AT_L1I_CACHESHAPE = 34;
static const int AT_L1D_CACHESHAPE = 35;
static const int AT_L2_CACHESHAPE = 36;
]]

ffi.cdef[[
unsigned long getauxval(unsigned long);
]]

With this, I can then write code that looks like the following:

local function getStringAuxVal(atype)
	local res = libc.getauxval(atype)

	if res == 0 then 
		return false, "type not found"
	end

	local str = ffi.string(ffi.cast("char *", res));
	return str
end

local function getIntAuxValue(atype)
	local res = libc.getauxval(atype)

	if res == 0 then 
		return false, "type not found"
	end

	return tonumber(res);
end

local function getPtrAuxValue(atype)
	local res = libc.getauxval(atype)

	if res == 0 then 
		return false, "type not found"
	end

	return ffi.cast("intptr_t", res);
end


-- convenience functions
local function getExecPath()
	return getStringAuxVal(libc.AT_EXECFN);
end

local function getPlatform()
	return getStringAuxVal(libc.AT_PLATFORM);
end

local function getPageSize()
	return getIntAuxValue(libc.AT_PAGESZ);
end

local function getRandom()
	return getPtrAuxValue(libc.AT_RANDOM);
end

--[[
	Some test cases
--]]
print(" Platform: ", getPlatform());
print("Exec Path: ", getExecPath());
print("Page Size: ", getPageSize());
print("   Random: ", getRandom());


And so on and so forth, assuming you have the proper ‘libc’ luajit ffi binding, which gives you access to constants through the ‘libc.’ mechanism.

OK, fine, if I’m a C programmer, and I just want to port some code that’s already doing this sort of thing. By the way, the one value that we’re interested in is: AT_SYSINFO_EHDR. That contains a pointer to the beginning of our vDSO. Then you can call functions directly from there (there’s an API for that).

But, if I’m a lua programmer, I’ve come to expect more out of my environment, largely because I’m lazy and don’t like so much typing. Upon further examination, you can get this information yourself. If you’re hard core, you can look at the top of memory in your program, and map that location to a pointer you can fiddle with directly. Otherwise, you can actually get this information from a ‘file’ on Linux.

Turns out that you can get this info from ‘/proc/self/auxv’, if you’re running this command about your current process (which you most likely are). So, now what can I do with that? Well, the lua way would be the following:

-- auxv_iter.lua
local ffi = require("ffi")
local libc = require("libc")

local E = {}

-- This table maps the constant values for the various
-- AT_* types to their symbolic names.  This table is used
-- to both generate cdefs, as well and hand back symbolic names
-- for the keys.
local auxtbl = {
	[0] =  "AT_NULL";
	[1] =  "AT_IGNORE";
	[2] = "AT_EXECFD";
	[3] = "AT_PHDR";
	[4] = "AT_PHENT";
	[5] = "AT_PHNUM";
	[6] = "AT_PAGESZ";
	[7] = "AT_BASE";
	[8] = "AT_FLAGS";
	[9] = "AT_ENTRY";
	[10] = "AT_NOTELF";
	[11] = "AT_UID";
	[12] = "AT_EUID";
	[13] = "AT_GID";
	[14] = "AT_EGID";
	[17] = "AT_CLKTCK";
	[15] = "AT_PLATFORM";
	[16] = "AT_HWCAP";
	[18] = "AT_FPUCW";
	[19] = "AT_DCACHEBSIZE";
	[20] = "AT_ICACHEBSIZE";
	[21] = "AT_UCACHEBSIZE";
	[22] = "AT_IGNOREPPC";
	[23] = "AT_SECURE";
	[24] = "AT_BASE_PLATFORM";
	[25] = "AT_RANDOM";
	[26] = "AT_HWCAP2";
	[31] = "AT_EXECFN";
	[32] = "AT_SYSINFO";
	[33] = "AT_SYSINFO_EHDR";
	[34] = "AT_L1I_CACHESHAPE";
	[35] = "AT_L1D_CACHESHAPE";
	[36] = "AT_L2_CACHESHAPE";
}

-- Given a auxv key(type), and the value returned from reading
-- the file, turn the value into a lua specific type.
-- string pointers --> string
-- int values -> number
-- pointer values -> intptr_t

local function auxvaluefortype(atype, value)
	if atype == libc.AT_EXECFN or atype == libc.AT_PLATFORM then
		return ffi.string(ffi.cast("char *", value))
	end

	if atype == libc.AT_UID or atype == libc.AT_EUID or
		atype == libc.AT_GID or atype == libc.AT_EGID or 
		atype == libc.AT_FLAGS or atype == libc.AT_PAGESZ or
		atype == libc.AT_HWCAP or atype == libc.AT_CLKTCK or 
		atype == libc.AT_PHENT or atype == libc.AT_PHNUM then

		return tonumber(value)
	end

	if atype == libc.AT_SECURE then
		if value == 0 then 
			return false
		else
			return true;
		end
	end


	return ffi.cast("intptr_t", value);
end

-- iterate over the auxv values at the specified path
-- if no path is specified, use '/proc/self/auxv' to get
-- the values for the currently running program
local function auxviterator(path)
	path = path or "/proc/self/auxv"
	local fd = libc.open(path, libc.O_RDONLY);

	local params = {
		fd = fd;
		keybuff = ffi.new("intptr_t[1]");
		valuebuff = ffi.new("intptr_t[1]");
		buffsize = ffi.sizeof(ffi.typeof("intptr_t"));
	}


	local function gen_value(param, state)
		local res1 = libc.read(param.fd, param.keybuff, param.buffsize)
		local res2 = libc.read(param.fd, param.valuebuff, param.buffsize)
		if param.keybuff[0] == 0 then
			libc.close(param.fd);
			return nil;
		end

		local atype = tonumber(param.keybuff[0])
		return state, atype, auxvaluefortype(atype, param.valuebuff[0])
	end

	return gen_value, params, 0

end

-- generate ffi.cdef calls to turn the symbolic type names
-- into constant integer values
local cdefsGenerated = false;

local function gencdefs()
	for k,v in pairs(auxtbl) do		
		-- since we don't know if this is already defined, we wrap
		-- it in a pcall to catch the error
		pcall(function() ffi.cdef(string.format("static const int %s = %d;", v,k)) end)
	end
	cdefsGenerated = true;
end

-- get a single value for specified key.  A path can be specified
-- as well (default it '/proc/self/auxv')
-- this is most like the gnuC getauxval() function
local function getOne(key, path)
	-- iterate over the values, looking for the one we want
	for _, atype, value in auxviterator(path) do
		if atype == key then
			return value;
		end
	end

	return nil;
end

E.gencdefs = gencdefs;
E.keyvaluepairs = auxviterator;	
E.keynames = auxtbl;
E.getOne = getOne;

setmetatable(E, {
	-- we allow the user to specify one of the symbolic constants
	-- when doing a 'getOne()'.  This indexing allows for the creation
	-- and use of those constants if they haven't already been specified
	__index = function(self, key)
		if not cdefsGenerated then
			gencdefs();
		end

		local success, value = pcall(function() return ffi.C[key] end)
		if success then
			rawset(self, key, value);
			return value;
		end

		return nil;
	end,

})

return E

In a nutshell, this is all you need for all the lua based auxv goodness in your life. Here are a couple of examples of usage in action:

local init = require("test_setup")()
local auxv_util = require("auxv_iter")
local apairs = auxv_util.keyvaluepairs;
local keynames = auxv_util.keynames;
local auxvGetOne = auxv_util.getOne;


--auxv_util.gencdefs();
print("==== Iterate All ====")
local function printAll()
	for _, key, value in apairs(path) do
		io.write(string.format("%20s[%2d] : ", keynames[key], key))
		print(value);
	end
end

-- print all the entries
printAll();

-- try to get a specific one
print("==== Get Singles ====")
print(" Platform: ", auxvGetOne(auxv_util.AT_PLATFORM))
print("Page Size: ", auxvGetOne(auxv_util.AT_PAGESZ))

The output from printAll() might look like this:

==== Iterate All ====
     AT_SYSINFO_EHDR[33] : 140721446887424LL
            AT_HWCAP[16] : 3219913727
           AT_PAGESZ[ 6] : 4096
           AT_CLKTCK[17] : 100
             AT_PHDR[ 3] : 4194368LL
            AT_PHENT[ 4] : 56
            AT_PHNUM[ 5] : 10
             AT_BASE[ 7] : 140081410764800LL
            AT_FLAGS[ 8] : 0
            AT_ENTRY[ 9] : 4208720LL
              AT_UID[11] : 1000
             AT_EUID[12] : 1000
              AT_GID[13] : 1000
             AT_EGID[14] : 1000
           AT_SECURE[23] : false
           AT_RANDOM[25] : 140721446787113LL
           AT_EXECFN[31] : /usr/local/bin/luajit
         AT_PLATFORM[15] : x86_64

The printAll() function uses the auxv iteration function, which in turns reads the key/value pairs directly from the /proc/self/auxv file. No need for the GNU C lib function at all. It goes further and turns the raw ‘unsigned long’ values into the appropriate data type based on what kind of data the key specified represents. So, you get lua string, and not just a pointer to a C string.

In the second example, getting singles, the output is simply this:

 Platform: 	x86_64
Page Size: 	4096

The code for this goes into a bit of trickery that’s possible with luajit. first of all, notice the use of ‘auxv_util.AT_PAGESZ’. There is nothing in the auxv_iter.lua file that supports this value directly. There is the table of names, and then there’s that ‘setmetatable’ at the end of things. Here’s where the trickery happens. Basically, this function is called whenever you put a ‘.’ after a table to try and access something, and that something isn’t in the table. You get a chance to make something up and return it. Ini this case, we first call ‘gencdefs()’ if that hasn’t already been called. This will generate ‘static const int XXX’ for all the values in the table of names, so that we can then do a lookup of the values in the ‘ffi.C.’ namespace, using the name. If we find a name, then we add it to the table, so next time the lookup will succeed, and we won’t end up calling the __index function.

At any rate, we now have the requisite value to lookup. Then we just roll through the iterator, and return when we’ve got the value we were looking for. The conversion to the appropriate lua type is automatic.

And there you have it! From relative obscurity, to complete usability, in one iterator. Being able to actually get function pointers in the vDSO is the next step. That will require another API wrapper, or worst case, and all encompassing ELF parser…

Advertisements


Leave a Reply

Fill in your details below or click an icon to log in:

WordPress.com Logo

You are commenting using your WordPress.com account. Log Out / Change )

Twitter picture

You are commenting using your Twitter account. Log Out / Change )

Facebook photo

You are commenting using your Facebook account. Log Out / Change )

Google+ photo

You are commenting using your Google+ account. Log Out / Change )

Connecting to %s