mmap: populate the mapping by default

Without pre-population, accessing all data through a mmap can result in many
pagefault, reducing performance significantly. If the mmap is prepopulated, the
performance can no longer get slower than a full read.
(See benchmark number below)

In some cases were very few data is read, prepopulating can be overkill and
slower than populating on access (through page fault). So that behavior can be
controlled when the caller can pre-determine the best behavior.
(See benchmark number below)

In addition, testing with populating in a secondary thread yield great result
combining the best of each approach. This might be implemented in later
changesets.

In all cases, using mmap has a great effect on memory usage when many processes
run in parallel on the same machine.

### Benchmarks

# What did I run

A couple of month back I ran a large benchmark campaign to assess the impact of
various approach for using mmap with the revlog (and other files), it
highlighted a few benchmarks that capture the impact of the changes well. So to
validate this change I checked the following:

- log command displaying various revisions
  (read the changelog index)
- log command displaying the patch of listed revisions
  (read the changelog index, the manifest index and a few files indexes)
- unbundling a few revisions
  (read and write changelog, manifest and few files indexes, and walk the graph
  to update some cache)
- pushing a few revisions
  (read and write changelog, manifest and few files indexes, walk the graph to
  update some cache, performs various accesses locally and remotely during
  discovery)

Benchmarks were run using the default module policy (c+py) and the rust one. No
significant difference were found between the two implementation, so we will
present result using the default policy (unless otherwise specified).

I ran them on a few repositories :

- mercurial: a "public changeset only" copy of mercurial from 2018-08-01 using
  zstd compression and sparse-revlog

- pypy: a copy of pypy from 2018-08-01 using zstd compression and sparse-revlog

- netbeans: a copy of netbeans from 2018-08-01 using zstd compression and
  sparse-revlog

- mozilla-try: a copy of mozilla-try from 2019-02-18 using zstd compression and
  sparse-revlog

- mozilla-try persistent-nodemap: Same as the above but with a persistent
  nodemap. Used for the log --patch benchmark only

# Results

For the smaller repositories (mercurial, pypy), the impact of mmap is almost
imperceptible, other cost dominating the operation. The impact of prepopulating
is undiscernible in the benchmark we ran.

For larger repositories the benchmark support explanation given above:

On netbeans, the log can be about 1% faster without repopulation (for a
difference < 100ms) but unbundle becomes a bit slower, even when small.

### data-env-vars.name                             = netbeans-2018-08-01-zstd-sparse-revlog
  # benchmark.name                                 = hg.command.unbundle
  # benchmark.variants.issue6528                   = disabled
  # benchmark.variants.reuse-external-delta-parent = yes
  # benchmark.variants.revs                        = any-1-extra-rev
  # benchmark.variants.source                      = unbundle
  # benchmark.variants.verbosity                   = quiet
with-populate: 0.240157
no-populate:   0.265087  (+10.38%, +0.02)
  # benchmark.variants.revs                        = any-100-extra-rev
with-populate: 1.459518
no-populate:   1.481290   (+1.49%, +0.02)
 ## benchmark.name                                 = hg.command.push
  # benchmark.variants.explicit-rev                = none
  # benchmark.variants.issue6528                   = disabled
  # benchmark.variants.protocol                    = ssh
  # benchmark.variants.reuse-external-delta-parent = yes
  # benchmark.variants.revs                        = any-1-extra-rev
with-populate: 0.771919
no-populate:   0.792025   (+2.60%, +0.02)
  # benchmark.variants.revs                        = any-100-extra-rev
with-populate: 1.459518
no-populate:   1.481290   (+1.49%, +0.02)


For mozilla-try, the "slow down" from pre-populate for small `hg log` is more
visible, but still small in absolute time. (using rust value for the persistent
nodemap value to be relevant).

### data-env-vars.name            = mozilla-try-2019-02-18-ds2-pnm
  # benchmark.name                = hg.command.log
  # bin-env-vars.hg.flavor        = rust
  # benchmark.variants.patch      = yes
  # benchmark.variants.limit-rev  = 1
with-populate: 0.237813
no-populate:   0.229452   (-3.52%, -0.01)
  # benchmark.variants.limit-rev  = 10
  # benchmark.variants.patch      = yes
with-populate: 1.213578
no-populate:   1.205189
### data-env-vars.name            = mozilla-try-2019-02-18-zstd-sparse-revlog
  # benchmark.variants.limit-rev  = 1000
  # benchmark.variants.patch      = no
  # benchmark.variants.rev        = tip
with-populate: 0.198607
no-populate:   0.195038   (-1.80%, -0.00)

However pre-populating provide a significant boost on more complex operations
like unbundle or push:

### data-env-vars.name                             = mozilla-try-2019-02-18-zstd-sparse-revlog
  # benchmark.name                                 = hg.command.push
  # benchmark.variants.explicit-rev                = none
  # benchmark.variants.issue6528                   = disabled
  # benchmark.variants.protocol                    = ssh
  # benchmark.variants.reuse-external-delta-parent = yes
  # benchmark.variants.revs                        = any-1-extra-rev
with-populate: 4.798632
no-populate:   4.953295   (+3.22%, +0.15)
  # benchmark.variants.revs                        = any-100-extra-rev
with-populate: 4.903618
no-populate:   5.014963   (+2.27%, +0.11)
 ## benchmark.name                                 = hg.command.unbundle
  # benchmark.variants.revs                        = any-1-extra-rev
with-populate: 1.423411
no-populate:   1.585365  (+11.38%, +0.16)
  # benchmark.variants.revs                        = any-100-extra-rev
with-populate: 1.537909
no-populate:   1.688489   (+9.79%, +0.15)

mercurial/util.py

@@ -440,13 +440,18 b' class bufferedinputpipe:'
         return data
 
 
-def mmapread(fp, size=None):
+def mmapread(fp, size=None, pre_populate=True):
     """Read a file content using mmap
 
     The responsability of checking the file system is mmap safe is the
-    responsability of the caller.
+    responsability of the caller (see `vfs.is_mmap_safe`).
 
     In some case, a normal string might be returned.
+
+    If `pre_populate` is True (the default), the mmapped data will be
+    pre-populated in memory if the system support this option, this slow down
+    the initial mmaping but avoid potentially crippling page fault on later
+    access. If this is not the desired behavior, set `pre_populate` to False.
     """
     if size == 0:
         # size of 0 to mmap.mmap() means "all data"
@@ -455,8 +460,12 b' def mmapread(fp, size=None):'
     elif size is None:
         size = 0
     fd = getattr(fp, 'fileno', lambda: fp)()
+    flags = mmap.MAP_PRIVATE
+    if pre_populate:
+        flags |= getattr(mmap, 'MAP_POPULATE', 0)
     try:
-        return mmap.mmap(fd, size, access=mmap.ACCESS_READ)
+        m = mmap.mmap(fd, size, flags=flags, prot=mmap.PROT_READ)
+        return m
     except ValueError:
         # Empty files cannot be mmapped, but mmapread should still work.  Check
         # if the file is empty, and if so, return an empty buffer.