Incorrect restoring of responces from chunked cache entries

[vankoven]

Applicable to saving and restoring responses from chunked cache entry. There is a situation when TfwCStr header can't be fit into the rest of current data chunk (TdbVRec). In this case TfwCStr will be saved into the next record. This condition is handled in tfw_cache_copy_hdr():

tempesta/tempesta_fw/cache.c

Line 686 in 65d9723

*p = tdb_entry_get_room(node_db(), trec, *p, n, *tot_len);

The tfw_cache_copy_hdr() has also an optimization in order not to split short TfwCStr:

tempesta/tempesta_fw/cache.c

Lines 675 to 678 in 65d9723

	/* Don't split short strings. */
	if (likely(!TFW_STR_DUP(src))
	&& sizeof(TfwCStr) + src->len <= L1_CACHE_BYTES)
	n += src->len;

But there is a couple of issues, when restoring responses from a cache entry. First of all optimization above is ignored. So if a response is written to cache with at least one such optimisation we can expect crash or serving corrupted response.

Secondly, crash will happen in tfw_cache_build_resp_hdr() if TfwCstr of the next header couldn't fit the rest of the current chunk and was placed into the next one:

tempesta/tempesta_fw/cache.c

Lines 1019 to 1020 in 65d9723

	*p += TFW_CSTR_HDRLEN;
	BUG_ON(*p > (*trec)->data + (*trec)->len);

To fix both issues we need to add a function to adjust current pointer in the current data chunk before reading TfwCStr similarly to tdb_entry_get_room() which is used before writing. And use the function any time we are reading a TfwCStr. To make function aware of optimization above, we need to write a padding TfwCStr header with flag SKIP_REC_TAIL set every time when the optimization is used. The flag will help us understand that the rest of data chunk doesn't contain any data and must be skipped. tdb_entry_get_room() must write this padding TfwCStr header.

[vankoven]

Note the #804 issue describing unit tests for this issue.

[krizhanovsky]

It's better to fix this after #515 which will affect caching a lot.

[const-t]

Fixed by #1831