[libvirt RFC v2] virfile: set pipe size in virFileWrapperFdNew to improve throughput

[Claudio Fontana]

On 3/25/22 11:41 AM, Daniel P. Berrangé wrote:
> On Mon, Mar 21, 2022 at 09:13:20AM +0100, Claudio Fontana wrote:
>> virsh save is very slow with a default pipe size, so set a larger one.
>>
>> This change improves throughput by ~400% on fast nvme or ramdisk,
>> for the current only user of virFileWrapperFdNew: the qemu driver.
>>
>> Best value currently measured is 1MB, which happens to be also
>> the kernel default for the pipe-max-size.
>>
>> We do not try to use a pipe buffer larger than what the setting
>> of /proc/sys/fs/pipe-max-size currently allows.
>>
>> Signed-off-by: Claudio Fontana <cfontana at suse.de>
>> ---
>>  src/util/virfile.c | 69 ++++++++++++++++++++++++++++++++++++++++++++++
>>  1 file changed, 69 insertions(+)
>>
>> see v1 at
>> https://listman.redhat.com/archives/libvir-list/2022-March/229252.html
>>
>> Changes v1 -> v2:
>>
>> * removed VIR_FILE_WRAPPER_BIG_PIPE, made the new pipe resizing
>>   unconditional (Michal)
>>
>> * moved code to separate functions (Michal)
>>
>> * removed ternary op, disliked in libvirt (Michal)
>>
>> * added #ifdef __linux__ (Ani Sinha)
>>
>> * try smallest value between currently best measured value (1MB)
>>   and the pipe-max-size setting. If pipe-max-size cannot be read,
>>   try kernel default max (1MB). (Daniel)
>>
>>
>>
>> diff --git a/src/util/virfile.c b/src/util/virfile.c
>> index a04f888e06..13bdd42c68 100644
>> --- a/src/util/virfile.c
>> +++ b/src/util/virfile.c
>> @@ -201,6 +201,71 @@ struct _virFileWrapperFd {
>>  };
>>  
>>  #ifndef WIN32
>> +
>> +#ifdef __linux__
>> +/**
>> + * virFileWrapperGetBestPipeSize:
>> + *
>> + * get the best pipe size to use with virFileWrapper.
>> + *
>> + * We first check the maximum we are allowed by the system pipe-max-size,
>> + * and then use the minimum between that and our tested best value.
>> + * This is because a request beyond pipe-max-size may fail with EPERM.
>> + * If we are unable to read pipe-max-size, use the kernel default (1MB).
>> + *
>> + * Return value is the pipe size to use.
>> + */
>> +
>> +static int virFileWrapperGetBestPipeSize(void)
>> +{
>> +    const char path[] = "/proc/sys/fs/pipe-max-size";
>> +    int best_sz = 1024 * 1024; /* good virsh save results with this size */
>> +    int max_sz;
>> +
>> +    if (virFileReadValueInt(&max_sz, path) < 0) {
>> +        max_sz = 1024 * 1024; /* this is the kernel default pipe-max-size */
>> +        VIR_WARN("failed to read %s, trying default %d", path, max_sz);
>> +    } else if (max_sz > best_sz) {
>> +        max_sz = best_sz;
>> +    }
>> +    return max_sz;
>> +}
>> +
>> +/**
>> + * virFileWrapperSetPipeSize:
>> + * @fd: the fd of the pipe
>> + *
>> + * Set best pipe size on the passed file descriptor for bulk transfers of data.
>> + *
>> + * default pipe size (usually 64K) is generally not suited for large transfers
>> + * to fast devices. This has been measured to improve virsh save by 400%
>> + * in ideal conditions.
>> + *
>> + * Return value is 0 on success, -1 and errno set on error.
>> + * OS note: only for linux, on other OS this is a no-op.
>> + */
>> +static int
>> +virFileWrapperSetPipeSize(int fd)
>> +{
>> +    int pipe_sz = virFileWrapperGetBestPipeSize();
> 
> I wonder if we shouldn't just ignore the proc setting and instead
> 
>    for (sz = 1024 * 1024 ; sz >= 64 * 1024; sz /= 2) {
>       int rv = fcntl(fd, F_SETPIPE_SZ, sz);
>       if (rv < 0 && errno == EPERM) {
>         continue;
>       }
>       if (rv < 0) {
>          virReportError(...)
> 	 return -1;
>       }
>       
>       VIR_INFO("fd %d pipe size adjusted to %d", fd, sz);
>       return 0;
>    }
> 
> 
> We'll only have 1 loop iteration in the default case, and 4 iterations
> in the worst case, and gracefully leave it on the default if the last
> ieratino fails
> 
> With regards,
> Daniel
> 

Yes, seems better to me,

Claudio