curl/docs/internals/BUFQ.md

bufq

This is an internal module for managing I/O buffers. A bufq can be written to and read from. It manages read and write positions and has a maximum size.

read/write

Its basic read/write functions have a similar signature and return code handling as many internal Curl read and write ones.

ssize_t Curl_bufq_write(struct bufq *q, const unsigned char *buf, size_t len, CURLcode *err);

- returns the length written into `q` or -1 on error.
- writing to a full `q` returns -1 and set *err to CURLE_AGAIN

ssize_t Curl_bufq_read(struct bufq *q, unsigned char *buf, size_t len, CURLcode *err);

- returns the length read from `q` or -1 on error.
- reading from an empty `q` returns -1 and set *err to CURLE_AGAIN

To pass data into a bufq without an extra copy, read callbacks can be used.

typedef ssize_t Curl_bufq_reader(void *reader_ctx, unsigned char *buf, size_t len,
                                 CURLcode *err);

ssize_t Curl_bufq_slurp(struct bufq *q, Curl_bufq_reader *reader, void *reader_ctx,
                        CURLcode *err);

Curl_bufq_slurp() invokes the given reader callback, passing it its own internal buffer memory to write to. It may invoke the reader several times, as long as it has space and while the reader always returns the length that was requested. There are variations of slurp that call the reader at most once or only read in a maximum amount of bytes.

The analog mechanism for write out buffer data is:

typedef ssize_t Curl_bufq_writer(void *writer_ctx, const unsigned char *buf, size_t len,
                                 CURLcode *err);

ssize_t Curl_bufq_pass(struct bufq *q, Curl_bufq_writer *writer, void *writer_ctx,
                       CURLcode *err);

Curl_bufq_pass() invokes the writer, passing its internal memory and remove the amount that writer reports.

peek and skip

It is possible to get access to the memory of data stored in a bufq with:

bool Curl_bufq_peek(const struct bufq *q, const unsigned char **pbuf, size_t *plen);

On returning TRUE, pbuf points to internal memory with plen bytes that one may read. This is only valid until another operation on bufq is performed.

Instead of reading bufq data, one may simply skip it:

void Curl_bufq_skip(struct bufq *q, size_t amount);

This removes amount number of bytes from the bufq.

unwrite

It is possible to undo writes by calling:

CURLcode Curl_bufq_unwrite(struct bufq *q, size_t len);

This will remove len bytes from the end of the bufq again. When removing more bytes than are present, CURLE_AGAIN is returned and the bufq will be empty.

lifetime

bufq is initialized and freed similar to the dynbuf module. Code using bufq holds a struct bufq somewhere. Before it uses it, it invokes:

void Curl_bufq_init(struct bufq *q, size_t chunk_size, size_t max_chunks);

The bufq is told how many "chunks" of data it shall hold at maximum and how large those "chunks" should be. There are some variants of this, allowing for more options. How "chunks" are handled in a bufq is presented in the section about memory management.

The user of the bufq has the responsibility to call:

void Curl_bufq_free(struct bufq *q);

to free all resources held by q. It is possible to reset a bufq to empty via:

void Curl_bufq_reset(struct bufq *q);

memory management

Internally, a bufq uses allocation of fixed size, e.g. the "chunk_size", up to a maximum number, e.g. "max_chunks". These chunks are allocated on demand, therefore writing to a bufq may return CURLE_OUT_OF_MEMORY. Once the max number of chunks are used, the bufq reports that it is "full".

Each chunks has a read and write index. A bufq keeps its chunks in a list. Reading happens always at the head chunk, writing always goes to the tail chunk. When the head chunk becomes empty, it is removed. When the tail chunk becomes full, another chunk is added to the end of the list, becoming the new tail.

Chunks that are no longer used are returned to a spare list by default. If the bufq is created with option BUFQ_OPT_NO_SPARES those chunks are freed right away.

If a bufq is created with a bufc_pool, the no longer used chunks are returned to the pool. Also bufq asks the pool for a chunk when it needs one. More in section "pools".

empty, full and overflow

One can ask about the state of a bufq with methods such as Curl_bufq_is_empty(q), Curl_bufq_is_full(q), etc. The amount of data held by a bufq is the sum of the data in all its chunks. This is what is reported by Curl_bufq_len(q).

Note that a bufq length and it being "full" are only loosely related. A simple example:

• create a bufq with chunk_size=1000 and max_chunks=4.
• write 4000 bytes to it, it reports "full"
• read 1 bytes from it, it still reports "full"
• read 999 more bytes from it, and it is no longer "full"

The reason for this is that full really means: bufq uses max_chunks and the last one cannot be written to.

When you read 1 byte from the head chunk in the example above, the head still hold 999 unread bytes. Only when those are also read, can the head chunk be removed and a new tail be added.

There is another variation to this. If you initialized a bufq with option BUFQ_OPT_SOFT_LIMIT, it allows writes beyond the max_chunks. It reports full, but one can still write. This option is necessary, if partial writes need to be avoided. It means that you need other checks to keep the bufq from growing ever larger and larger.

pools

A struct bufc_pool may be used to create chunks for a bufq and keep spare ones around. It is initialized and used via:

void Curl_bufcp_init(struct bufc_pool *pool, size_t chunk_size, size_t spare_max);

void Curl_bufq_initp(struct bufq *q, struct bufc_pool *pool, size_t max_chunks, int opts);

The pool gets the size and the mount of spares to keep. The bufq gets the pool and the max_chunks. It no longer needs to know the chunk sizes, as those are managed by the pool.

A pool can be shared between many bufqs, as long as all of them operate in the same thread. In curl that would be true for all transfers using the same multi handle. The advantages of a pool are:

• when all bufqs are empty, only memory for max_spare chunks in the pool is used. Empty bufqs holds no memory.
• the latest spare chunk is the first to be handed out again, no matter which bufq needs it. This keeps the footprint of "recently used" memory smaller.