When creating a new directory, we set the initial size to 1 block.
This meant that we were allocating a block up front, but the Inode's
internal block list cache was not populated with this block.
This broke write_bytes() on a new directory, since it assumed that
the block list cache would be up to date if the call to write_bytes()
would not change the directory's size.
This patch fixes the issue in two ways: First, we cache the initial
block list created for new directories.
Second, we now repopulate the block list cache in write_bytes() if it
is empty when we get there. This is basically just a safety fallback
to avoid having this kind of bug in the future.
We can't be calling the virtual FS::flush_writes() in order to flush
the disk cache from within the disk cache, since an FS subclass may
try to do cache stuff in its flush_writes() implementation.
Instead, separate out the implementation of DiskBackedFS's flushing
logic into a flush_writes_impl() and call that from the cache code.
Also cache the block group descriptor table in a KBuffer on file system
initialization, instead of on first access.
This reduces pressure on the kmalloc heap somewhat.
We were writing out the full block list whenever Ext2FSInode::resize()
was called, even if the old and new sizes were identical.
This patch makes it a no-op, which drastically improves "cp" speed
since we now take full advantage of the up-front call to ftruncate().
If there are not enough free blocks in the filesystem to accomodate
growing an Inode, we should fail with ENOSPC before even starting to
allocate blocks.
Add a simple cache to Ext2FS where we keep block bitmaps along with a
dirty bit. This allows us to coalesce bitmap flushes, giving us a nice
~3x improvement in disk_benchmark write speeds.
Store the cached super block as an ext2_super_block member instead of
caching it in a ByteBuffer and using a casting helper everywhere.
This patch also combines reading/writing of the super block into a
single disk device operation (instead of two.)
Add dedicated internal types for Int64 and UnsignedInt64. This makes it
a bit more straightforward to work with 64-bit numbers (instead of just
implicitly storing them as doubles.)
I have no idea why this was here. It makes no sense. If you're trying
to find out if something is a directory, why wouldn't you be allowed to
ask that about a FIFO? :^)
Thanks to Brandon for spotting this!
Also, while we're here, cache the directory state in a bool member so
we don't have to keep fetching inode metadata when checking this
repeatedly. This is important since sys$read() now calls it.
Previously, procfs$pid_fds would return nothing when called
for a process that had either no open files or a non-existent
handle. This could cause problems when a userspace program
expected a valid Json response.
Procfs$pid_fs now returns an empty array in the aforementioned
cases.
This reverts commit 1cca5142af.
This appears to be causing intermittent triple-faults and I don't know
why yet, so I'll just revert it to keep the tree in decent shape.
Background: DoubleBuffer is a handy buffer class in the kernel that
allows you to keep writing to it from the "outside" while the "inside"
reads from it. It's used for things like LocalSocket and PTY's.
Internally, it has a read buffer and a write buffer, but the two will
swap places when the read buffer is exhausted (by reading from it.)
Before this patch, it was internally implemented as two Vector<u8>
that we would swap between when the reader side had exhausted the data
in the read buffer. Now instead we preallocate a large KBuffer (64KB*2)
on DoubleBuffer construction and use that throughout its lifetime.
This removes all the kmalloc heap traffic caused by DoubleBuffers :^)
The Cache entries found in `DiskBackedFileSystem` are now stored in a
`KBuffer` object, instead of relying on `kmalloc_eternal`. The number
of entries was exceeding that of the number of bytes allocated to
`kmalloc_eternal`, which in turn caused `mount()` to fail epically
when called.
This patch adds three separate per-process fault counters:
- Inode faults
An inode fault happens when we've memory-mapped a file from disk
and we end up having to load 1 page (4KB) of the file into memory.
- Zero faults
Memory returned by mmap() is lazily zeroed out. Every time we have
to zero out 1 page, we count a zero fault.
- CoW faults
VM objects can be shared by multiple mappings that make their own
unique copy iff they want to modify it. The typical reason here is
memory shared between a parent and child process.
This patch overloads Inode::is_directory() with a faster version that
doesn't require instantiating the whole InodeMetadata.
If you have an Ext2FSInode&, calling is_directory() should be instant
since we can just look directly at the raw inode bits.
If we want to make a new entry in the disk cache when it's completely
full of dirty blocks, we'll now synchronously flush the writes at that
point. Maybe it's not ideal, but at least we can keep going.
This way clients are not required to have instantiated ByteBuffers
and can choose whatever memory scheme works best for them.
Also converted some of the Ext2FS code to use stack buffers instead.
Instead of having DiskBackedFS allocate a ByteBuffer, leave it to each
client to provide buffer space.
This is significantly faster in many cases where we can use a stack
buffer and avoid heap allocation entirely.
The hashmap cache was ridiculously slow and hurt us more than it helped
us. This patch replaces it with a flat memory cache that keeps up to
10'000 blocks in cache with a simple dirty bit.
The syncd task will wake up periodically and call flush_writes() on all
file systems, which now causes us to traverse the cache and write all
dirty blocks to disk.
There's a ton of room for improvement here, but this itself is already
drastically better when doing repeated GCC invocations.
1) Off-by-one in block allocation when block size != 1 KB
Due to a quirk in the on-disk layout of ext2, file systems with a block
size of 1 KB have block #1 as their first block, while all others start
on block #0.
2) We had no fallback mechanism when the preferred group was full
We now allocate blocks from the preferred block group as long as it's
possible, and fall back to a simple scan through all groups when the
preferred one is full.
This is a freelist allocator with static size classes that works as a
complement to the generic kmalloc(). It's a lot faster than kmalloc()
since allocation just means popping from the freelist.
It's also significantly more compact when there are a lot of objects
smaller than the minimum kmalloc chunk size (32 bytes.)
This patch enables it for the Region and PhysicalPage classes.
In the PhysicalPage (8 bytes) case, it's a huge improvement since we
no longer waste 75% of the storage allocated.
There are also a number of ways this can be improved, so let's keep
working on it going forward.
Also added some assertions to DirectoryEntry in case someone tries to
instantiate them with names that would overflow the name buffer.
DirectoryEntry is a crappy data structure, and the name buffer is also
crappy. Added a FIXME about replacing it with something nicer.
Before this patch, the DirectoryEntry::name buffer would overflow if
you did "touch extremely-long-file-name". Duh.
Fixes#538.
This was a workaround to be able to build on case-insensitive file
systems where it might get confused about <string.h> vs <String.h>.
Let's just not support building that way, so String.h can have an
objectively nicer name. :^)
- TmpFSInode::write_bytes() needs to allow non-zero offsets
- TmpFSInode::read_bytes() wasn't respecting the offset
GCC puts the temporary files generated during compilation in /tmp,
so this exposed some bugs in TmpFS.
