Test FluidMem with LRUBuffer size 1

[annapanag]

I am facing an unexpected behavior when running a test with LRU size set to 1.

Here is the custom test that I run:

Expected behavior

FluidMem to capture 10 pagefaults in total. This is because the test case makes 10 accesses to 2 different pages in turn. Since the LRU buffer can only support 1 page, every access should result in a pagefault.

Actual results

From the ui tool, I get the following results:

Response received. size:357
LRU Buffer Size:	1
LRU Buffer Capacity:	1
Registered ufds:	0
Total Page Fault Count:	2
Zero Page Count:	2
Placed Data Page Count:	0
Zero Pages Encountered:	0
Page Eviction Count:	0
Cache Hit Count:	0
Cache Miss Count:	2
Cache Hit Percentage:	0
Writes Avoided:		0
Invalid Pages Dropped:	1
Page Fault Rate:	0.122843
Last Page Fault:	[2020-08-31T21:40:30.693]

Also, FluidMem monitor process produces this log message:

[2020-08-31T21:40:30.692] [11749] [DEBG] (/home/anna/fluidmem/include/buffer_allocator_array.h:21) get_local_tmp_page: local_tmp_page=0x7f8d03d4b000
[2020-08-31T21:40:30.692] [11749] [DEBG] (userfault.c:672) evict_page: src: 0x161e000, dst 0x7f8d03d4b000, ufd: 9
[2020-08-31T21:40:30.692] [11749] [DEBG] (userfault.c:948) evict_to_externram: skipping page for invalid fd 9, rc: 22, pageaddr: 0x161e000
[2020-08-31T21:40:30.692] [11749] [DEBG] (userfault.c:1076) evict_to_externram: munmap to 0x7f8d03d4b000
[2020-08-31T21:40:30.693] [11749] [DEBG] (userfault.c:450) evict_if_needed: eviction of page 0x161e000 skipped

Steps to reproduce

Resize LRU Buffer to size 1:

ui 127.0.0.1 r 1

Compile the test case linking with userfault library

Run the test case

Notes

Is it possible this is related to the asynchronous eviction mechanism? I have also observed in another custom test bus error to occur in a non-deterministic manner. Also, from the "skipping page" log of monitor process, I have the suspicion that the program has exited, thus the ufd file descriptor is invalid for the UFFDIO_REMAP ioctl call.