1 files changed, 60 insertions, 0 deletions
diff --git a/results/classifier/gemma3:12b/performance/1740219 b/results/classifier/gemma3:12b/performance/1740219
new file mode 100644
index 00000000..84c708fe
--- /dev/null
+++ b/results/classifier/gemma3:12b/performance/1740219
@@ -0,0 +1,60 @@
+
+static linux-user ARM emulation has several-second startup time
+
+static linux-user emulation has several-second startup time
+
+My problem: I'm a Parabola packager, and I'm updating our
+qemu-user-static package from 2.8 to 2.11.  With my new
+statically-linked 2.11, running `qemu-arm /my/arm-chroot/bin/true`
+went from taking 0.006s to 3s!  This does not happen with the normal
+dynamically linked 2.11, or the old static 2.8.
+
+What happens is it gets stuck in
+`linux-user/elfload.c:init_guest_space()`.  What `init_guest_space`
+does is map 2 parts of the address space: `[base, base+guest_size]`
+and `[base+0xffff0000, base+0xffff0000+page_size]`; where it must find
+an acceptable `base`.  Its strategy is to `mmap(NULL, guest_size,
+...)` decide where the first range is, and then check if that
++0xffff0000 is also available.  If it isn't, then it starts trying
+`mmap(base, ...)` for the entire address space from low-address to
+high-address.
+
+"Normally," it finds an accaptable `base` within the first 2 tries.
+With a static 2.11, it's taking thousands of tries.
+
+----
+
+Now, from my understanding, there are 2 factors working together to
+cause that in static 2.11 but not the other builds:
+
+ - 2.11 increased the default `guest_size` from 0xf7000000 to 0xffff0000
+ - PIE (and thus ASLR) is disabled for static builds
+
+For some reason that I don't understand, with the smaller
+`guest_size` the initial `mmap(NULL, guest_size, ...)` usually
+returns an acceptable address range; but larger `guest_size` makes it
+consistently return a block of memory that butts right up against
+another already mapped chunk of memory.  This isn't just true on the
+older builds, it's true with the 2.11 builds if I use the `-R` flag to
+shrink the `guest_size` back down to 0xf7000000.  That is with
+linux-hardened 4.13.13 on x86-64.
+
+So then, it it falls back to crawling the entire address space; so it
+tries base=0x00001000.  With ASLR, that probably succeeds.  But with
+ASLR being disabled on static builds, the text segment is at
+0x60000000; which is does not leave room for the needed
+0xffff1000-size block before it.  So then it tries base=0x00002000.
+And so on, more than 6000 times until it finally gets to and passes
+the text segment; calling mmap more than 12000 times.
+
+----
+
+I'm not sure what the fix is.  Perhaps try to mmap a continuous chunk
+of size 0xffff1000, then munmap it and then mmap the 2 chunks that we
+actually need.  The disadvantage to that is that it does not support
+the sparse address space that the current algorithm supports for
+`guest_size < 0xffff0000`.  If `guest_size < 0xffff0000` *and* the big
+mmap fails, then it could fall back to a sparse search; though I'm not
+sure the current algorithm is a good choice for it, as we see in this
+bug.  Perhaps it should inspect /proc/self/maps to try to find a
+suitable range before ever calling mmap?
\ No newline at end of file