Re: [PATCH 1/2] lib: add fast lzo decompressor

[Andreas Robinson <andr345@gmail.com>]

On Fri, 2009-04-03 at 07:59 +1100, Nigel Cunningham wrote:

> static int toi_compress_write_page(unsigned long index,
>                 struct page *buffer_page, unsigned int buf_size)
> {
>         int ret, cpu = smp_processor_id();
>         struct cpu_context *ctx = &per_cpu(contexts, cpu);
> 
>         if (!ctx->transform)
>                 return next_driver->write_page(index, buffer_page, buf_size);
> 
>         ctx->buffer_start = kmap(buffer_page);
> 
>         ctx->len = buf_size;

The LZO compressor can produce more bytes than it consumes but here the
output buffer is the same size as the input.
This macro in linux/lzo.h defines how big the buffer needs to be:
#define lzo1x_worst_compress(x) ((x) + ((x) / 16) + 64 + 3)

If there are multiple threads perhaps they clobber each other's output
buffers?

>         ret = crypto_comp_compress(ctx->transform,
>                         ctx->buffer_start, buf_size,
>                         ctx->page_buffer, &ctx->len);
> 
>         kunmap(buffer_page);
> 
>         if (ret) {
>                 printk(KERN_INFO "Compression failed.\n");
>                 return ret;
>         }
> 
>         mutex_lock(&stats_lock);
>         toi_compress_bytes_in += buf_size;
>         toi_compress_bytes_out += ctx->len;
>         mutex_unlock(&stats_lock);
> 
>         if (ctx->len < buf_size) /* some compression */
>                 return next_driver->write_page(index,
>                                 virt_to_page(ctx->page_buffer),
>                                 ctx->len);
>         else
>                 return next_driver->write_page(index, buffer_page, buf_size);
> }