Skip to content
Merged
Show file tree
Hide file tree
Changes from 5 commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
109 changes: 109 additions & 0 deletions src/liballoc/tests/vec.rs
Original file line number Diff line number Diff line change
Expand Up @@ -945,6 +945,115 @@ fn drain_filter_complex() {
}
}

#[test]
#[cfg(not(miri))] // Miri does not support catching panics
fn drain_filter_consumed_panic() {
use std::rc::Rc;
use std::sync::Mutex;

struct Check {
index: usize,
drop_counts: Rc<Mutex<Vec<usize>>>,
};

impl Drop for Check {
fn drop(&mut self) {
self.drop_counts.lock().unwrap()[self.index] += 1;
println!("drop: {}", self.index);
}
}

let check_count = 10;
let drop_counts = Rc::new(Mutex::new(vec![0_usize; check_count]));
let mut data: Vec<Check> = (0..check_count)
.map(|index| Check { index, drop_counts: Rc::clone(&drop_counts) })
.collect();

let _ = std::panic::catch_unwind(move || {
let filter = |c: &mut Check| {
if c.index == 2 {
panic!("panic at index: {}", c.index);
}
// Verify that if the filter could panic again on another element
// that it would not cause a double panic and all elements of the
// vec would still be dropped exactly once.
if c.index == 4 {
panic!("panic at index: {}", c.index);
}
c.index < 6
};
let drain = data.drain_filter(filter);

// NOTE: The DrainFilter is explictly consumed
drain.for_each(drop);
});

let drop_counts = drop_counts.lock().unwrap();
assert_eq!(check_count, drop_counts.len());

for (index, count) in drop_counts.iter().cloned().enumerate() {
assert_eq!(1, count, "unexpected drop count at index: {} (count: {})", index, count);
}
}

#[test]
#[cfg(not(miri))] // Miri does not support catching panics
fn drain_filter_unconsumed_panic() {
use std::rc::Rc;
use std::sync::Mutex;

struct Check {
index: usize,
drop_counts: Rc<Mutex<Vec<usize>>>,
};

impl Drop for Check {
fn drop(&mut self) {
self.drop_counts.lock().unwrap()[self.index] += 1;
println!("drop: {}", self.index);
}
}

let check_count = 10;
let drop_counts = Rc::new(Mutex::new(vec![0_usize; check_count]));
let mut data: Vec<Check> = (0..check_count)
.map(|index| Check { index, drop_counts: Rc::clone(&drop_counts) })
.collect();

let _ = std::panic::catch_unwind(move || {
let filter = |c: &mut Check| {
if c.index == 2 {
panic!("panic at index: {}", c.index);
}
// Verify that if the filter could panic again on another element
// that it would not cause a double panic and all elements of the
// vec would still be dropped exactly once.
if c.index == 4 {
panic!("panic at index: {}", c.index);
}
c.index < 6
};
let _drain = data.drain_filter(filter);

// NOTE: The DrainFilter is dropped without being consumed
});

let drop_counts = drop_counts.lock().unwrap();
assert_eq!(check_count, drop_counts.len());

for (index, count) in drop_counts.iter().cloned().enumerate() {
assert_eq!(1, count, "unexpected drop count at index: {} (count: {})", index, count);
}
}

#[test]
fn drain_filter_unconsumed() {
let mut vec = vec![1, 2, 3, 4];
let drain = vec.drain_filter(|&mut x| x % 2 != 0);
drop(drain);
assert_eq!(vec, [2, 4]);
}

#[test]
fn test_reserve_exact() {
// This is all the same as test_reserve
Expand Down
67 changes: 58 additions & 9 deletions src/liballoc/vec.rs
Original file line number Diff line number Diff line change
Expand Up @@ -2120,6 +2120,7 @@ impl<T> Vec<T> {
del: 0,
old_len,
pred: filter,
panic_flag: false,
}
}
}
Expand Down Expand Up @@ -2751,6 +2752,7 @@ pub struct DrainFilter<'a, T, F>
del: usize,
old_len: usize,
pred: F,
panic_flag: bool,
}
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

This has grown enough fields that they should really be documented. Either individually here, or with a high-level description of what we're trying to do in the implementation of next.

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I added docs to each field in the latest commit.


#[unstable(feature = "drain_filter", reason = "recently added", issue = "43244")]
Expand All @@ -2761,20 +2763,23 @@ impl<T, F> Iterator for DrainFilter<'_, T, F>

fn next(&mut self) -> Option<T> {
unsafe {
while self.idx != self.old_len {
while self.idx < self.old_len {
let i = self.idx;
self.idx += 1;
let v = slice::from_raw_parts_mut(self.vec.as_mut_ptr(), self.old_len);
if (self.pred)(&mut v[i]) {
self.panic_flag = true;
let drained = (self.pred)(&mut v[i]);
self.panic_flag = false;
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I have a terrible compulsion to try to encode this state in some magic combination of old_len/idx/del, but this is probably clearest, and easiest for llvm to evaporate when it sees pred can't unwind.

Copy link
Contributor Author

@aloucks aloucks Jul 7, 2019

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I considered that route initially but came to the conclusion that even if it's possible, the simple flag would be much easier to understand and maintain.

// Update the index *after* the predicate is called. If the index
// is updated prior and the predicate panics, the element at this
// index would be leaked.
self.idx += 1;
if drained {
self.del += 1;
return Some(ptr::read(&v[i]));
} else if self.del > 0 {
let del = self.del;
let src: *const T = &v[i];
let dst: *mut T = &mut v[i - del];
// This is safe because self.vec has length 0
// thus its elements will not have Drop::drop
// called on them in the event of a panic.
Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Is this comment no longer true? Why is it safe instead then?

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I believe the intent was to leak instead of double drop, but it didn't quite work. I've done some minor refactoring and added additional comments.

It's safe now because there are additional checks in DrainFilter::drop that perform the cleanup in the event of a panic in the filter predicate.

ptr::copy_nonoverlapping(src, dst, 1);
}
}
Expand All @@ -2792,9 +2797,53 @@ impl<T, F> Drop for DrainFilter<'_, T, F>
where F: FnMut(&mut T) -> bool,
{
fn drop(&mut self) {
self.for_each(drop);
unsafe {
self.vec.set_len(self.old_len - self.del);
struct BackshiftOnDrop<'a, 'b, T, F>
where
F: FnMut(&mut T) -> bool,
{
drain: &'b mut DrainFilter<'a, T, F>,
}

impl<'a, 'b, T, F> Drop for BackshiftOnDrop<'a, 'b, T, F>
where
F: FnMut(&mut T) -> bool
{
fn drop(&mut self) {
unsafe {
// Backshift any unprocessed elements, preventing double-drop
// of any element that *should* have been previously overwritten
// but was not due to a panic in the filter predicate. This is
// implemented via drop so that it's guaranteed to run even in
// the event of a panic while consuming the remainder of the
// DrainFilter.
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

(currently checking with the lang team that we guarantee that this works)

Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

cc @rust-lang/lang @matthewjasper @arielb1

Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

specifically: if we panic in a destructor and we weren't yet panicking:

  • we unwind the drop impl, dropping its locals
  • then drop_in_place self's fields as normal
  • then proceed into a normal panic, unwinding the stack

(the middle point I'm not sure is necessary here, but is just my understanding of what is supposed to happen.)

Copy link
Member

@RalfJung RalfJung Jul 5, 2019

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

* then drop_in_place `self`'s fields as normal

This is indeed what we're currently doing:

// START rustc.ptr-real_drop_in_place.std__vec__Vec_i32_.AddMovesForPackedDrops.before.mir
// bb0: {
// goto -> bb7;
// }
// bb1: {
// return;
// }
// bb2 (cleanup): {
// resume;
// }
// bb3: {
// goto -> bb1;
// }
// bb4 (cleanup): {
// goto -> bb2;
// }
// bb5 (cleanup): {
// drop(((*_1).0: alloc::raw_vec::RawVec<i32>)) -> bb4;
// }
// bb6: {
// drop(((*_1).0: alloc::raw_vec::RawVec<i32>)) -> [return: bb3, unwind: bb4];
// }
// bb7: {
// _2 = &mut (*_1);
// _3 = const <std::vec::Vec<i32> as std::ops::Drop>::drop(move _2) -> [return: bb6, unwind: bb5];
// }
// END rustc.ptr-real_drop_in_place.std__vec__Vec_i32_.AddMovesForPackedDrops.before.mir

Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Copy link
Contributor

@gnzlbg gnzlbg Jul 7, 2019

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

@gankro your understanding is correct. If drop panics:

  • we unwind the drop impl, dropping its locals,
  • then drop_in_place self's fields as normal (where normal means in field declaration order - first field is dropped before second field is dropped),
  • then we continue unwinding Drop::drop's caller stack

If dropping a value in Drop::drop stack while its unwinding panics, or if dropping one of the type's fields panics, then we have a double-drop, and the process is guarantee to abort for all double drops.

Note also that the panic doesn't need to originate in the Drop::impl, this one can succeed, but then the fields are dropped in declaration order, and one of them can panic. When this happens, the remaining fields are still dropped in place, so if a second one of them panics, then you have a double drop. If the Drop::impl does not panic, but dropping one of the fields panics, and you don't get a double drop, you are guaranteed that the field that panicked is the only one that wasn't properly dropped.

EDIT: for all practical purposes, if dropping a value panics, you should treat that value as dropped, since "most" of the value will have been dropped in the process (at most one field wasn't dropped).

while self.drain.idx < self.drain.old_len {
let i = self.drain.idx;
self.drain.idx += 1;
let v = slice::from_raw_parts_mut(
self.drain.vec.as_mut_ptr(),
self.drain.old_len,
);
if self.drain.del > 0 {
let del = self.drain.del;
let src: *const T = &v[i];
let dst: *mut T = &mut v[i - del];
ptr::copy_nonoverlapping(src, dst, 1);
}
}
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I don't understand why this while loop exists. Surely this should just be:

if self.drain.idx < self.drain.old_len {
  // It looks like `pred` panicked, so we didn't process all the elements.
  // This is a pretty messed up state, and there isn't really an obviously right
  // thing to do (and we don't want to keep trying to execute `pred`). So we
  // just backshift all the unprocessed elements and tell the vec that they still
  // exist, hoping that doesn't mess up anyone further along in the panic.
  let idx = self.drain.idx;
  let num_deleted = self.drain.del;
  let tail_len = self.drain.old_len - idx;
  let ptr = self.drain.vec.as_mut_ptr();
  if num_deleted > 0 {
    ptr.add(idx).copy_to(ptr.add(idx - num_deleted), tail_len);
  }
}

self.drain.vec.set_len(self.drain.old_len - self.drain.del);

(Here I modernized the code a bit to use the newer raw pointer APIs, and some clearer names. It would be a nice cleanup to this code if you also did that to the Iterator's fields and its next impl as well, although not a blocker.)

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I incorporated this in the latest commit. I also consolidated the the self.drain.del check into the parent if and made the src and dst pointers explicit while still using the more modern pointer APIs.

self.drain.vec.set_len(self.drain.old_len - self.drain.del);
}
}
}

let backshift = BackshiftOnDrop {
drain: self
};

// Attempt to consume any remaining elements if the filter predicate
// has not yet panicked. We'll backshift any remaining elements
// whether we've already panicked or if the consumption here panics.
if !backshift.drain.panic_flag {
backshift.drain.for_each(drop);
}
}
}