I have some packed structs which I will be writing to a memory mapped file. They are all POD.
To accommodate some generic programming I'm doing, I want to be able to write a std::tuple of several packed structs.
I'm worried that writing the members of a std::tuple to my mapped region's address, and then later casting that address back to a std::tuple is going to break.
I've written a small examplar program, and it does seem to work, but I'm worried that I have undefined behaviour.
Here are my structs:
struct Foo
{
char c;
uint8_t pad[3];
int i;
double d;
} __attribute__((packed));
struct Bar
{
int i;
char c;
uint8_t pad[3];
double d;
} __attribute__((packed));
I define a std::tuple of these structs:
using Tup = std::tuple<Foo, Bar>;
To simulate the memory mapped file I have created a small object with some inline storage and a size:
When adding a tuple it uses placement new to construct the tuple in the inline storage.
struct Storage
{
Tup& push_back(Tup&& t)
{
Tup* p = reinterpret_cast<Tup*>(buf) + size;
new (p) Tup(std::move(t));
size += 1;
return *p;
}
const Tup& get(std::size_t i) const
{
const Tup* p = reinterpret_cast<const Tup*>(buf) + i;
return *p;
}
std::size_t size = 0;
std::uint8_t buf[100];
};
To simulate writing to a file and then reading it again I create one Storage object, populate it, copy it, and then let the original go out of scope.
Storage s2;
// scope of s1
{
Storage s1;
Tup t1 = { Foo { 'a', 1, 2.3 }, Bar { 2, 'b', 3.4 } };
Tup t2 = { Foo { 'c', 3, 5.6 }, Bar { 4, 'd', 7.8 } };
Tup& s1t1 = s1.push_back(std::move(t1));
Tup& s1t2 = s1.push_back(std::move(t2));
std::get<0>(s1t1).c = 'x';
std::get<1>(s1t2).c = 'z';
s2 = s1;
}
I then read my tuples using Storage::get which just does a reinterpret_cast<Tup&> of the inline storage.
const Tup& s2t1 = s2.get(0);
When I access the structs within the tuple they have the correct values.
In addition, running through valgrind doesn't throw up any errors.
reinterpret_cast from my inline storage to std::tuple if the tuple was originally placement newed there (into a file which will be closed and then later remapped and reread)?Memory mapped file:
The actual storage I use is a struct cast onto a boost::mapped_region.
The struct is:
struct Storage
{
std::size_t size;
std::uint8_t buf[1]; // address of buf is beginning of Tup array
};
I cast it as follows:
boost::mapped_region region_ = ...;
Storage* storage = reinterpret_cast<Storage*>(region_.get_address());
Will the alignment issues mentioned in answers below be a problem?
Full example below:
#include <cassert>
#include <cstdint>
#include <tuple>
struct Foo
{
char c;
uint8_t pad[3];
int i;
double d;
} __attribute__((packed));
struct Bar
{
int i;
char c;
uint8_t pad[3];
double d;
} __attribute__((packed));
using Tup = std::tuple<Foo, Bar>;
struct Storage
{
Tup& push_back(Tup&& t)
{
Tup* p = reinterpret_cast<Tup*>(buf) + size;
new (p) Tup(std::move(t));
size += 1;
return *p;
}
const Tup& get(std::size_t i) const
{
const Tup* p = reinterpret_cast<const Tup*>(buf) + i;
return *p;
}
std::size_t size = 0;
std::uint8_t buf[100];
};
int main ()
{
Storage s2;
// scope of s1
{
Storage s1;
Tup t1 = { Foo { 'a', 1, 2.3 }, Bar { 2, 'b', 3.4 } };
Tup t2 = { Foo { 'c', 3, 5.6 }, Bar { 4, 'd', 7.8 } };
Tup& s1t1 = s1.push_back(std::move(t1));
Tup& s1t2 = s1.push_back(std::move(t2));
std::get<0>(s1t1).c = 'x';
std::get<1>(s1t2).c = 'z';
s2 = s1;
}
const Tup& s2t1 = s2.get(0);
const Tup& s2t2 = s2.get(1);
const Foo& f1 = std::get<0>(s2t1);
const Bar& b1 = std::get<1>(s2t1);
const Foo& f2 = std::get<0>(s2t2);
const Bar& b2 = std::get<1>(s2t2);
assert(f1.c == 'x');
assert(f1.i == 1);
assert(f1.d == 2.3);
assert(b1.i == 2);
assert(b1.c == 'b');
assert(b1.d == 3.4);
assert(f2.c == 'c');
assert(f2.i == 3);
assert(f2.d == 5.6);
assert(b2.i == 4);
assert(b2.c == 'z');
assert(b2.d == 7.8);
return 0;
}
You may like to align std::uint8_t buf[100] storage because unaligned access is undefined behaviour:
aligned_storage<sizeof(Tup) * 100, alignof(Tup)>::type buf;
(originally you had 100 bytes, this is for 100 Tups).
When you map pages they start on at least 4k boundary on x86. If your storage starts on a page start then that storage is suitably aligned for any power-2 alignment up to 4k.
I'm worried that writing the members of a
std::tupleto my mapped region's address, and then later casting that address back to astd::tupleis going to break.
As long as the applications communicating through mapped memory use the same ABI, that works as expected.