Guide to writing output filters

Filters and bucket brigades

Each time a filter is invoked, it is passed a bucket brigade, containing a sequence of buckets which represent both data content and metadata. Every bucket has a bucket type; a number of bucket types are defined and used by the httpd core modules (and the apr-util library which provides the bucket brigade interface), but modules are free to define their own types.

A filter can tell whether a bucket represents either data or metadata using the APR_BUCKET_IS_METADATA macro. Generally, all metadata buckets should be passed down the filter chain by an output filter. Filters may transform, delete, and insert data buckets as appropriate.

There are two metadata bucket types which all filters must pay attention to: the EOS bucket type, and the FLUSH bucket type. An EOS bucket indicates that the end of the response has been reached and no further buckets need be processed. A FLUSH bucket indicates that the filter should flush any buffered buckets (if applicable) down the filter chain immediately.

Filters can create FLUSH buckets and pass these down the filter chain if desired. Generating FLUSH buckets unnecessarily, or too frequently, can harm network utilisation since it may force large numbers of small packets to be sent, rather than a small number of larger packets. The section on Non-blocking bucket reads covers a case where filters are encouraged to generate FLUSH buckets.

This shows a bucket brigade which may be passed to a filter; it contains two metadata buckets (FLUSH and EOS), and two data buckets (HEAP and FILE).

Filter invocation

For any given request, an output filter might be invoked only once and be given a single brigade representing the entire response. It is also possible that the number of times a filter is invoked for a single response is proportional to the size of the content being filtered, with the filter being passed a brigade containing a single bucket each time. Filters must operate correctly in either case.

An output filter can distinguish the final invocation for a given response by the presence of an EOS bucket in the brigade. Any buckets in the brigade after an EOS should be ignored.

An output filter should never pass an empty brigade down the filter chain. To be defensive, filters should be prepared to accept an empty brigade, and should return success without passing this brigade on down the filter chain. The handling of an empty brigade should have no side effects (such as changing any state private to the filter).

apr_status_t dummy_filter(ap_filter_t *f, apr_bucket_brigade *bb)
{
    if (APR_BRIGADE_EMPTY(bb)) {
        return APR_SUCCESS;
    }
    ...

Brigade structure

A bucket brigade is a doubly-linked list of buckets. The list is terminated (at both ends) by a sentinel which can be distinguished from a normal bucket by comparing it with the pointer returned by APR_BRIGADE_SENTINEL. The list sentinel is in fact not a valid bucket structure; any attempt to call normal bucket functions (such as apr_bucket_read) on the sentinel will have undefined behaviour (i.e. will crash the process).

There are a variety of functions and macros for traversing and manipulating bucket brigades; see the apr_buckets.h header for complete coverage. Commonly used macros include:

APR_BRIGADE_FIRST(bb)

returns the first bucket in brigade bb

APR_BRIGADE_LAST(bb)

returns the last bucket in brigade bb

APR_BUCKET_NEXT(e)

gives the next bucket after bucket e

APR_BUCKET_PREV(e)

gives the bucket before bucket e

The apr_bucket_brigade structure itself is allocated out of a pool, so if a filter creates a new brigade, it must ensure that memory use is correctly bounded. A filter which allocates a new brigade out of the request pool (r->pool) on every invocation, for example, will fall foul of the warning above concerning memory use. Such a filter should instead create a brigade on the first invocation per request, and store that brigade in its state structure.

Processing buckets

When dealing with non-metadata buckets, it is important to understand that the "apr_bucket *" object is an abstract representation of data:

1. The amount of data represented by the bucket may or may not have a determinate length; for a bucket which represents data of indeterminate length, the ->length field is set to the value (apr_size_t)-1. For example, buckets of the PIPE bucket type have an indeterminate length; they represent the output from a pipe.
2. The data represented by a bucket may or may not be mapped into memory. The FILE bucket type, for example, represents data stored in a file on disk.

Filters read the data from a bucket using the apr_bucket_read function. When this function is invoked, the bucket may morph into a different bucket type, and may also insert a new bucket into the bucket brigade. This must happen for buckets which represent data not mapped into memory.

To give an example; consider a bucket brigade containing a single FILE bucket representing an entire file, 24 kilobytes in size:

When this bucket is read, it will read a block of data from the file, morph into a HEAP bucket to represent that data, and return the data to the caller. It also inserts a new FILE bucket representing the remainder of the file; after the apr_bucket_read call, the brigade looks like:

Filtering brigades

The basic function of any output filter will be to iterate through the passed-in brigade and transform (or simply examine) the content in some manner. The implementation of the iteration loop is critical to producing a well-behaved output filter.

Taking an example which loops through the entire brigade as follows:

apr_bucket *e = APR_BRIGADE_FIRST(bb);
const char *data;
apr_size_t length;

while (e != APR_BRIGADE_SENTINEL(bb)) {
    apr_bucket_read(e, &data, &length, APR_BLOCK_READ);
    e = APR_BUCKET_NEXT(e);
}

return ap_pass_brigade(bb);

The above implementation would consume memory proportional to content size. If passed a FILE bucket, for example, the entire file contents would be read into memory as each apr_bucket_read call morphed a FILE bucket into a HEAP bucket.

In contrast, the implementation below will consume a fixed amount of memory to filter any brigade; a temporary brigade is needed and must be allocated only once per response, see the Maintaining state section.

apr_bucket *e;
const char *data;
apr_size_t length;

while ((e = APR_BRIGADE_FIRST(bb)) != APR_BRIGADE_SENTINEL(bb)) {
    rv = apr_bucket_read(e, &data, &length, APR_BLOCK_READ);
    if (rv) ...;
    /* Remove bucket e from bb. */
    APR_BUCKET_REMOVE(e);
    /* Insert it into  temporary brigade. */
    APR_BRIGADE_INSERT_HEAD(tmpbb, e);
    /* Pass brigade downstream. */
    rv = ap_pass_brigade(f->next, tmpbb);
    if (rv) ...;
    apr_brigade_cleanup(tmpbb);
}

Maintaining state

A filter which needs to maintain state over multiple invocations per response can use the ->ctx field of its ap_filter_t structure. It is typical to store a temporary brigade in such a structure, to avoid having to allocate a new brigade per invocation as described in the Brigade structure section.

struct dummy_state {
    apr_bucket_brigade *tmpbb;
    int filter_state;
    ...
};

apr_status_t dummy_filter(ap_filter_t *f, apr_bucket_brigade *bb)
{
    struct dummy_state *state;

    state = f->ctx;
    if (state == NULL) {

        /* First invocation for this response: initialise state structure.
         */
        f->ctx = state = apr_palloc(f->r->pool, sizeof *state);

        state->tmpbb = apr_brigade_create(f->r->pool, f->c->bucket_alloc);
        state->filter_state = ...;
    }
    ...

Buffering buckets

If a filter decides to store buckets beyond the duration of a single filter function invocation (for example storing them in its ->ctx state structure), those buckets must be set aside. This is necessary because some bucket types provide buckets which represent temporary resources (such as stack memory) which will fall out of scope as soon as the filter chain completes processing the brigade.

To setaside a bucket, the apr_bucket_setaside function can be called. Not all bucket types can be setaside, but if successful, the bucket will have morphed to ensure it has a lifetime at least as long as the pool given as an argument to the apr_bucket_setaside function.

Alternatively, the ap_save_brigade function can be used, which will move all the buckets into a separate brigade containing buckets with a lifetime as long as the given pool argument. This function must be used with care, taking into account the following points:

1. On return, ap_save_brigade guarantees that all the buckets in the returned brigade will represent data mapped into memory. If given an input brigade containing, for example, a PIPE bucket, ap_save_brigade will consume an arbitrary amount of memory to store the entire output of the pipe.
2. When ap_save_brigade reads from buckets which cannot be setaside, it will always perform blocking reads, removing the opportunity to use Non-blocking bucket reads.
3. If ap_save_brigade is used without passing a non-NULL "saveto" (destination) brigade parameter, the function will create a new brigade, which may cause memory use to be proportional to content size as described in the Brigade structure section.

Non-blocking bucket reads

The apr_bucket_read function takes an apr_read_type_e argument which determines whether a blocking or non-blocking read will be performed from the data source. A good filter will first attempt to read from every data bucket using a non-blocking read; if that fails with APR_EAGAIN, then send a FLUSH bucket down the filter chain, and retry using a blocking read.

This mode of operation ensures that any filters further down the filter chain will flush any buffered buckets if a slow content source is being used.

A CGI script is an example of a slow content source which is implemented as a bucket type. mod_cgi will send PIPE buckets which represent the output from a CGI script; reading from such a bucket will block when waiting for the CGI script to produce more output.

apr_bucket *e;
apr_read_type_e mode = APR_NONBLOCK_READ;

while ((e = APR_BRIGADE_FIRST(bb)) != APR_BRIGADE_SENTINEL(bb)) {
    apr_status_t rv;

    rv = apr_bucket_read(e, &data, &length, mode);
    if (rv == APR_EAGAIN && mode == APR_NONBLOCK_READ) {

        /* Pass down a brigade containing a flush bucket: */
        APR_BRIGADE_INSERT_TAIL(tmpbb, apr_bucket_flush_create(...));
        rv = ap_pass_brigade(f->next, tmpbb);
        apr_brigade_cleanup(tmpbb);
        if (rv != APR_SUCCESS) return rv;

        /* Retry, using a blocking read. */
        mode = APR_BLOCK_READ;
        continue;
    }
    else if (rv != APR_SUCCESS) {
        /* handle errors */
    }

    /* Next time, try a non-blocking read first. */
    mode = APR_NONBLOCK_READ;
    ...
}

Ten rules for output filters

In summary, here is a set of rules for all output filters to follow:

1. Output filters should not pass empty brigades down the filter chain, but should be tolerant of being passed empty brigades.
2. Output filters must pass all metadata buckets down the filter chain; FLUSH buckets should be respected by passing any pending or buffered buckets down the filter chain.
3. Output filters should ignore any buckets following an EOS bucket.
4. Output filters must process a fixed amount of data at a time, to ensure that memory consumption is not proportional to the size of the content being filtered.
5. Output filters should be agnostic with respect to bucket types, and must be able to process buckets of unfamiliar type.
6. After calling ap_pass_brigade to pass a brigade down the filter chain, output filters should call apr_brigade_cleanup to ensure the brigade is empty before reusing that brigade structure; output filters should never use apr_brigade_destroy to "destroy" brigades.
7. Output filters must setaside any buckets which are preserved beyond the duration of the filter function.
8. Output filters must not ignore the return value of ap_pass_brigade, and must return appropriate errors back up the filter chain.
9. Output filters must only create a fixed number of bucket brigades for each response, rather than one per invocation.
10. Output filters should first attempt non-blocking reads from each data bucket, and send a FLUSH bucket down the filter chain if the read blocks, before retrying with a blocking read.

Use case: buffering in mod_ratelimit

The r1833875 change is a good example to show what buffering and keeping state means in the context of an output filter. In this use case, a user asked on the users' mailing list a interesting question about why mod_ratelimit seemed not to honor its setting with proxied content (either rate limiting at a different speed or simply not doing it at all). Before diving deep into the solution, it is better to explain on a high level how mod_ratelimit works. The trick is really simple: take the rate limit settings and calculate a chunk size of data to flush every 200ms to the client. For example, let's imagine that to set rate-limit 60 in our config, these are the high level steps to find the chunk size:

/* milliseconds to wait between each flush of data */
RATE_INTERVAL_MS = 200;
/* rate limit speed in b/s */
speed = 60 * 1024;
/* final chunk size is 12228 bytes */
chunk_size = (speed / (1000 / RATE_INTERVAL_MS));

If we apply this calculation to a bucket brigade carrying 38400 bytes, it means that the filter will try to do the following:

1. Split the 38400 bytes in chunks of maximum 12228 bytes each.
2. Flush the first 12228 chunk of bytes and sleep 200ms.
3. Flush the second 12228 chunk of bytes and sleep 200ms.
4. Flush the third 12228 chunk of bytes and sleep 200ms.
5. Flush the remaining 1716 bytes.

The above pseudo code works fine if the output filter handles only one brigade for each response, but it might happen that it needs to be called multiple times with different brigade sizes as well. The former use case is for example when httpd directly serves some content, like a static file: the bucket brigade abstraction takes care of handling the whole content, and rate limiting works nicely. But if the same static content is served via mod_proxy_http (for example a backend is serving it rather than httpd) then the content generator (in this case mod_proxy_http) may use a maximum buffer size and then send data as bucket brigades to the output filters chain regularly, triggering of course multiple calls to mod_ratelimit. If the reader tries to execute the pseudo code assuming multiple calls to the output filter, each one requiring to process a bucket brigade of 38400 bytes, then it is easy to spot some anomalies:

1. Between the last flush of a brigade and the first one of the next, there is no sleep.
2. Even if the sleep was forced after the last flush, then that chunk size would not be the ideal size (1716 bytes instead of 12228) and the final client's speed would quickly become different than what set in the httpd's config.

In this case, two things might help:

1. Use the ctx internal data structure, initialized by mod_ratelimit for each response handling cycle, to "remember" when the last sleep was performed across multiple invocations, and act accordingly.
2. If a bucket brigade is not splittable into a finite number of chunk_size blocks, store the remaining bytes (located in the tail of the bucket brigade) in a temporary holding area (namely another bucket brigade) and then use ap_save_brigade to set them aside. These bytes will be prepended to the next bucket brigade that will be handled in the subsequent invocation.
3. Avoid the previous logic if the bucket brigade that is currently being processed contains the end of stream bucket (EOS). There is no need to sleep or buffering data if the end of stream is reached.

The commit linked in the beginning of the section contains also a bit of code refactoring so it is not trivial to read during the first pass, but the overall idea is basically what written up to now. The goal of this section is not to cause a headache to the reader trying to read C code, but to put him/her into the right mindset needed to use efficiently the tools offered by the httpd's filter chain toolset.