3 * Reconsider command line arg style--confuses everyone.
7 * Add serial input support. Also, modify tests to redirect input from
8 /dev/null, to avoid stray keystrokes getting sent into the VM.
10 * Make pintos script read the serial output and kill the subprocess if
11 it panics (after waiting a few seconds) or triple-faults. Might
12 want it to be optional, so that interactive users don't get killed.
14 * Godmar: Introduce memory leak robustness tests - both for the
15 well-behaved as well as the mis-behaved case - that tests that the
16 kernel handles low-mem conditions well.
18 * Godmar: Another area is concurrency. I noticed that I had passed all
19 tests with bochs 2.2.1 (in reproducibility mode). Then I ran them
20 with qemu and hit two deadlocks (one of them in rox-*,
21 incidentally). After fixing those deadlocks, I upgraded to bochs
22 2.2.5 and hit yet another deadlock in reproducibility mode that
23 didn't show up in 2.2.1. All in all, a standard grading run would
24 have missed 3 deadlocks in my code. I'm not sure how to exploit
25 that for grading - either run with qemu n times (n=2 or 3), or run
26 it with bochs and a set of -j parameters. Some of which could be
27 known to the students, some not, depending on preference. (I ported
28 the -j patch to bochs 2.2.5 -
29 http://people.cs.vt.edu/~gback/pintos/bochs-2.2.5.jitter.patch but I
30 have to admit I never tried it so I don't know if it would have
31 uncovered the deadlocks that qemu and the switch to 2.2.5
34 * Godmar: There is also the option to require students to develop test
35 workloads themselves, for instance, to demonstrate the effectiveness
36 of a particular algorithm (page eviction & buffer cache replacement
37 come to mind.) This could involve a problem of the form: develop a
38 workload that you cover well, and develop a "worst-case" load where
39 you algorithm performs poorly, and show the results of your
40 quantitative evaluation in your report - this could then be part of
47 >> Describe a potential race in thread_set_priority() and explain how
48 >> your implementation avoids it. Can you use a lock to avoid this race?
50 I'm not sure what you're getting at here:
51 If changing the priority of a thread involves accessing the ready
52 list, then of course there's a race with interrupt handlers and locks
53 can't be used to resolve it.
55 Changing the priority however also involves a race with respect to
56 accessing a thread's "priority" field - this race is with respect to
57 other threads that attempt to donate priority to the thread that's
58 changing its priority. Since this is a thread-to-thread race, I would
59 tend to believe that locks could be used, although I'm not certain. [
60 I should point out, though, that lock_acquire currently disables
61 interrupts - the purpose of which I had doubted in an earlier email,
62 since sema_down() sufficiently establishes mutual exclusion. Taking
63 priority donation into account, disabling interrupts prevents the race
64 for the priority field, assuming the priority field of each thread is
65 always updated with interrupts disabled. ]
67 What answer are you looking for for this design document question?
71 >> Did any ambiguities in the scheduler specification make values in the
72 >> table uncertain? If so, what rule did you use to resolve them? Does
73 >> this match the behavior of your scheduler?
75 My guess is that you're referring to the fact the scheduler
76 specification does not prescribe any order in which the priorities of
77 all threads are updated, so if multiple threads end up with the same
78 priority, it doesn't say which one to pick. ("round-robin" order
79 would not apply here.)
85 One of my groups implemented priority donation with these data
86 structures in synch.cc:
90 struct list_elem elem; /* List element. */
91 int value; /* Item value. */
94 static struct value values[10];
95 static int start = 10;
96 static int numNest = 0;
98 In their implementation, the "elem" field in their "struct value" is
101 The sad part is that they've passed all tests that are currently in
102 the Pintos base with this implementation. (They do fail the additional
103 tests I added priority-donate-sema & priority-donate-multiple2.)
105 Another group managed to pass all tests with this construct:
109 struct thread *holder; /* Thread holding lock (for debugging). */
110 struct semaphore semaphore; /* Binary semaphore controlling access. */
111 //*************************************
113 int pri_delta; //Used for Priority Donation
114 /**************************************************/
117 where "pri_delta" keeps track of "priority deltas." They even pass
118 priority-donate-multiple2.
120 I think we'll need a test where a larger number of threads & locks
121 simultaneously exercise priority donation to weed out those
124 It may also be a good idea to use non-constant deltas for the low,
125 medium, and high priority threads in the tests - otherwise, adding a
126 "priority delta" might give - by coincidence - the proper priority for
129 - Godmar: Another thing: one group passed all tests even though they
130 wake up all waiters on a lock_release(), rather than just
131 one. Since there's never more than one waiter in our tests, they
132 didn't fail anything. Another possible TODO item - this could be
133 part a series of "regression tests" that check that they didn't
134 break basic functionality in project 1. I don't think this would
135 be insulting to the students.
139 - Get rid of rox--causes more trouble than it's worth
141 - Extra credit: specifics on how to implement sbrk, malloc.
143 - Godmar: We're missing tests that pass arguments to system calls
144 that span multiple pages, where some are mapped and some are not.
145 An implementation that only checks the first page, rather than all
146 pages that can be touched during a call to read()/write() passes
149 - Godmar: Need some tests that test that illegal accesses lead to
150 process termination. I have written some, will add them. In P2,
151 obviously, this would require that the students break this
152 functionality since the page directory is initialized for them,
153 still it would be good to have.
155 - Godmar: There does not appear to be a test that checks that they
156 close all fd's on exit. Idea: add statistics & self-diagnostics
157 code to palloc.c and malloc.c. Self-diagnostics code could be
158 used for debugging. The statistics code would report how much
159 kernel memory is free. Add a system call
160 "get_kernel_memory_information". User programs could engage in a
161 variety of activities and notice leaks by checking the kernel
164 - Godmar: is there a test that tests that they properly kill a process that
165 attempts to access an invalid address in user code, e.g. *(void**)0 =
168 It seems all of the robustness tests deal with bad pointers passed to
169 system calls (at least judging from test/userprog/Rubric.robustness),
170 but none deals with bad accesses by user code, or I am missing
173 ps: I found tests/vm/pt-bad-addr, which is in project 3 only, though.
175 For completeness, we should probably check read/write/jump to unmapped
176 user virtual address and to mapped kernel address, for a total of 6
177 cases. I wrote up some tests, see
178 http://people.cs.vt.edu/~gback/pintos/bad-pointers/
180 - process_death test needs improvement
182 - Godmar: In the wait() tests, there's currently no test that tests
183 that a process can only wait for its own children. There's only
184 one test that tests that wait() on an invalid pid returns -1 (or
185 kills the process), but no test where a valid pid is used that is
186 not a child of the current process.
188 The current tests also do not ensure that both scenarios (parent waits
189 first vs. child exits first) are exercised. In this context, I'm
190 wondering if we should add a sleep() system call that would export
191 timer_sleep() to user processes; this would allow the construction of
192 such a test. It would also make it easier to construct a test for the
193 valid-pid, but not-a-child scenario.
195 As in Project 4, the baseline implementation of timer_sleep() should
196 suffice, so this would not necessarily require basing Project 2 on
197 Project 1. [ A related thought: IMO it would not be entirely
198 unreasonable to require timer_sleep() and priority scheduling sans
199 donation from Project 1 working for subsequent projects. ]
203 - Godmar: Get rid of mmap syscall, add sbrk.
205 - Godmar: page-linear, page-shuffle VM tests do not use enough
206 memory to force eviction. Should increase memory consumption.
208 - Godmar: fix the page* tests to require swapping
210 - Godmar: make sure the filesystem fails if not properly
211 concurrency-protected in project 3.
213 - Godmar: Another area in which tests could be created are for
214 project 3: tests that combine mmap with a paging workload to see
215 their kernel pages properly while mmapping pages - I don't think
216 the current tests test that, do they?
220 - Need a better way to measure performance improvement of buffer
221 cache. Some students reported that their system was slower with
222 cache--likely, Bochs doesn't simulate a disk with a realistic
225 (Perhaps we should count disk reads and writes, not time.)
227 - Do we check that non-empty directories cannot be removed?
229 - Need lots more tests.
231 - Add FS persistence test(s).
233 - Godmar: I'm not sure if I mentioned that already, but I passed all
234 tests for the filesys project without having implemented inode
235 deallocation. A test is needed that checks that blocks are
236 reclaimed when files are deleted.
238 - Godmar: I'm in the middle of project 4, I've started by
239 implementing a buffer cache and plugging it into the existing
240 filesystem. Along the way I was wondering how we could test the
243 Maybe one could adopt a similar testing strategy as in project 1
244 for the MLQFS scheduler: add a function that reads
245 "get_cache_accesses()" and a function "get_cache_hits()". Then
246 create a version of pintos that creates access traces for a
247 to-be-determined workload. Run an off-line analysis that would
248 determine how many hits a perfect cache would have (MAX), and how
249 much say an LRU strategy would give (MIN). Then add a fudge
250 factor to account for different index strategies and test that the
251 reported number of cache hits/accesses is within (MIN, MAX) +/-
254 (As an aside - I am curious why you chose to use a clock-style
255 algorithm rather than the more straightforward LRU for your buffer
256 cache implementation in your sample solution. Is there a reason
257 for that? I was curious to see if it made a difference, so I
258 implemented LRU for your cache implementation and ran the test
259 workload of project 4 and printed cache hits/accesses. I found
260 that for that workload, the clock-based algorithm performs almost
261 identical to LRU (within about 1%, but I ran nondeterministally
262 with QEMU). I then reduced the cache size to 32 blocks and found
263 again the same performance, which raises the suspicion that the
264 test workload might not force any cache replacement, so the
265 eviction strategy doesn't matter.)
267 - Godmar: I haven't analyzed the tests for project 4 yet, but I'm
268 wondering if the fairness requirements your specification has for
269 readers/writers are covered in the tests or not.
274 - Add "Digging Deeper" sections that describe the nitty-gritty x86
275 details for the benefit of those interested.
277 - Add explanations of what "real" OSes do to give students some
280 * To add partition support:
282 - Find four partition types that are more or less unused and choose
283 to use them for Pintos. (This is implemented.)
285 - Bootloader reads partition tables of all BIOS devices to find the
286 first that has the "Pintos kernel" partition type. (This is
287 implemented.) Ideally the bootloader would make sure there is
288 exactly one such partition, but I didn't implement that yet.
290 - Bootloader reads kernel into memory at 1 MB using BIOS calls.
291 (This is implemented.)
293 - Kernel arguments have to go into a separate sector because the
294 bootloader is otherwise too big to fit now? (I don't recall if I
295 did anything about this.)
297 - Kernel at boot also scans partition tables of all the disks it can
298 find to find the ones with the four Pintos partition types
299 (perhaps not all exist). After that, it makes them available to
300 the rest of the kernel (and doesn't allow access to other devices,
303 - "pintos" and "pintos-mkdisk" need to write a partition table to
304 the disks that they create. "pintos-mkdisk" will need to take a
305 new parameter specifying the type. (I might have partially
306 implemented this, don't remember.)
308 - "pintos" should insist on finding a partition header on disks
309 handed to it, for safety.
311 - Need some way for "pintos" to assemble multiple disks or
312 partitions into a single image that can be copied directly to a
313 USB block device. (I don't know whether I came up with a good
314 solution yet or not, or whether I implemented any of it.)
316 * To add USB support:
318 - Needs to be able to scan PCI bus for UHCI controller. (I
319 implemented this partially.)
321 - May want to be able to initialize USB controllers over CardBus
322 bridges. I don't know whether this requires additional work or
323 if it's useful enough to warrant extra work. (It's of special
324 interest for me because I have a laptop that only has USB via
327 - There are many protocol layers involved: SCSI over USB-Mass
328 Storage over USB over UHCI over PCI. (I may be forgetting one.)
329 I don't know yet whether it's best to separate the layers or to
330 merge (some of) them. I think that a simple and clean
331 organization should be a priority.
333 - VMware can likely be used for testing because it can expose host
334 USB devices as guest USB devices. This is safer and more
335 convenient than using real hardware for testing.
337 - Should test with a variety of USB keychain devices because there
338 seems to be wide variation among them, especially in the SCSI
339 protocols they support. Should try to use a "lowest-common
340 denominator" SCSI protocol if any such thing really exists.
342 - Might want to add a feature whereby kernel arguments can be
343 given interactively, rather than passed on-disk. Needs some