Cross-repo refilling is caused by cached fs

[WaberZhuang]

What happened in your environment?

ICacheStore *ICachePool::open(std::string_view filename, int flags, mode_t mode) {
    // ...
    auto len = this->fn_trans_func(filename, store_name, sizeof(store_name));
    std::string_view store_sv = len ? std::string_view(store_name, len) : filename;
    auto ctor = [&]() -> ICacheStore * {
        // ...
        cache_store->set_src_name(filename);
        // ...
    };
    auto store = cast(m_stores)->acquire(store_sv, ctor);
    if (store) {
        auto cnt = store->ref_.fetch_add(1, std::memory_order_relaxed);
        if (cnt)
            cast(m_stores)->release(store_sv);
    }
    return store;
}

If we call cached_fs->open("/foo/sha256:111") and cached_fs->open("/bar/sha256:111") in sequence, both files will refill from /foo/sha256:111. Which means if repository foo is deleted at some point, all requests to the layer sha256:111 will fail. Furthermore, if retries occur frequently, failures may persist indefinitely. (m_stores keeps for 10 seconds).

What did you expect to happen?

The request to /bar/sha256:111 should succeed or succeed after a retry period.

How can we reproduce it?

Add following lines at this unit test, which can be used to verify this issue.

TEST(CachedFS, fn_trans_func) {
    // ...
    EXPECT_STREQ("/path_aaa/sha256:test", cs1->get_src_name().data());
    EXPECT_STREQ("/path_bbb/sha256:test", cs2->get_src_name().data());
}

What is the version of your Overlaybd?

v1.0.16

What is your OS environment?

any

Are you willing to submit PRs to fix it?

• Yes, I am willing to fix it.