prov/shm: incorrect iteration in fi_av_insert()

[tatarintsevsv]

Looks like smr_av_insert() has invalid addresses iteration
Docs says:

When using the FI_ADDR_STR format, the addr parameter should reference an array of strings (char **).

But in smr_av_insert():
for (i = 0; i < count; i++, addr = (char *) addr + strlen(addr) + 1)
istead of iteration with
for (i = 0; i < count; i++)
So I got segfault when trying to pass char** addr
If this is really incorrect, I could fix it and make a PR

[aingerson]

Hi @tatarintsevsv and thanks for opening up an issue!
This is just poorly phrased in the doc. It is an array of strings in the sense that the strings live contiguously within the allocation (ie string1, string2, string3 back to back in memory). The problem is that a string is a char array so it ends up being a char ** but really you could think of it more like a char[]*. So it's a pointer to a list of characters. Because shm addrlen is non-constant (so not all names are the same length), we can't iterate normally and have to shift by the length of the name.
I think you're thinking of it more like an array of string pointers, not an array of strings themselves.

What you're trying to do:

addr = calloc (3, char *);
addr[0] = calloc(1, sizeof(name0));
addr[1] = calloc(1, sizeof(name1));
addr[2] = calloc(1, sizeof(name2));

//set all addrs

fi_av_insert(addr, 3);

what's expected:

addr = calloc(1, sizeof(name0)+sizeof(name1)+sizeof(name2));

//set all addrs

fi_av_insert(addr, 3);

You can see that the second way (what's expected) just simplifies the calloc and also acts like a traditional array which is how applications manage other address formats as well.

We can definitely update the language to help clarify the usage and if you would like to do that, you are more than welcome to! Otherwise I can and I will get you to review it to make sure it makes sense to a user.

[tatarintsevsv]

So shm works as intended. Then the problem is different..
Let me introduce the problem with an example

1. We use prov/shm

char *addr = "shm1\0shm2"; // string-by-string

fi_av_insert(av, addr, 2, fi_addr, 0, NULL); // it's ok

2. In case of prov/psm3

char *addr_array[2] = {"fi_addr_psmx3://0:0:0:0", "fi_addr_psmx3://0:0:0:0"}; // array of strings
char *addr = "fi_addr_psmx3://0:0:0:0\0fi_addr_psmx3://0:0:0:0"; // string-by-string (like shm)
	
fi_av_insert(av, addr_array, 2, fi_addr, 0, NULL); // ok
fi_av_insert(av, addr, 2, fi_addr, 0, NULL); // got segfault

i.e. different behavior for different providers. I mean the problem is not in the documentation

[j-xiong]

With a further look, I think the documentation is correct. We need to revisit how the shm provider handles addr.

[tatarintsevsv]

I created a PR for the fix.
Also fixed fi_pingpong for FI_ADDR_STR usage cases

[j-xiong]

Just for bookkeeping, the related PR is #11336