On Thu, 14 May 2015, Michal Hocko wrote:

On Wed 13-05-15 11:00:36, Eric B Munson wrote:

quoted

On Mon, 11 May 2015, Eric B Munson wrote:

quoted

On Fri, 08 May 2015, Andrew Morton wrote:

quoted

On Fri,  8 May 2015 15:33:43 -0400 Eric B Munson [off-list ref] wrote:

quoted

mlock() allows a user to control page out of program memory, but this
comes at the cost of faulting in the entire mapping when it is
allocated.  For large mappings where the entire area is not necessary
this is not ideal.

This series introduces new flags for mmap() and mlockall() that allow a
user to specify that the covered are should not be paged out, but only
after the memory has been used the first time.

Please tell us much much more about the value of these changes: the use
cases, the behavioural improvements and performance results which the
patchset brings to those use cases, etc.

To illustrate the proposed use case I wrote a quick program that mmaps
a 5GB file which is filled with random data and accesses 150,000 pages
from that mapping.  Setup and processing were timed separately to
illustrate the differences between the three tested approaches.  the
setup portion is simply the call to mmap, the processing is the
accessing of the various locations in  that mapping.  The following
values are in milliseconds and are the averages of 20 runs each with a
call to echo 3 > /proc/sys/vm/drop_caches between each run.

The first mapping was made with MAP_PRIVATE | MAP_LOCKED as a baseline:
Startup average:    9476.506
Processing average: 3.573

The second mapping was simply MAP_PRIVATE but each page was passed to
mlock() before being read:
Startup average:    0.051
Processing average: 721.859

The final mapping was MAP_PRIVATE | MAP_LOCKONFAULT:
Startup average:    0.084
Processing average: 42.125

Michal's suggestion of changing protections and locking in a signal
handler was better than the locking as needed, but still significantly
more work required than the LOCKONFAULT case.

Startup average:    0.047
Processing average: 86.431

Have you played with batching? Has it helped? Anyway it is to be
expected that the overhead will be higher than a single mmap call. The
question is whether you can live with it because adding a new semantic
to mlock sounds trickier and MAP_LOCKED is tricky enough already...

I reworked the experiment to better cover the batching solution.  The
same 5GB data file is used, however instead of 150,000 accesses at
regular intervals, the test program now does 15,000,000 accesses to
random pages in the mapping.  The rest of the setup remains the same.

mmap with MAP_LOCKED:
Setup avg:      11821.193
Processing avg: 3404.286

mmap with mlock() before each access:
Setup avg:      0.054
Processing avg: 34263.201

mmap with PROT_NONE and signal handler and batch size of 1 page:
With the default value in max_map_count, this gets ENOMEM as I attempt
to change the permissions, after upping the sysctl significantly I get:
Setup avg:      0.050
Processing avg: 67690.625

mmap with PROT_NONE and signal handler and batch size of 8 pages:
Setup avg:      0.098
Processing avg: 37344.197

mmap with PROT_NONE and signal handler and batch size of 16 pages:
Setup avg:      0.0548
Processing avg: 29295.669

mmap with MAP_LOCKONFAULT:
Setup avg:      0.073
Processing avg: 18392.136

The signal handler in the batch cases faulted in memory in two steps to
avoid having to know the start and end of the faulting mapping.  The
first step covers the page that caused the fault as we know that it will
be possible to lock.  The second step speculatively tries to mlock and
mprotect the batch size - 1 pages that follow.  There may be a clever
way to avoid this without having the program track each mapping to be
covered by this handeler in a globally accessible structure, but I could
not find it.

These results show that if the developer knows that a majority of the
mapping will be used, it is better to try and fault it in at once,
otherwise MAP_LOCKONFAULT is significantly faster.

Eric