mirrors/zig
1
0
Fork 0
mirror of https://codeberg.org/ziglang/zig.git synced 2025-12-06 22:04:21 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions 1

Fix ublk uring usage due to SQE32 vs 128

Browse source 
When IORING_SETUP_SQE128 is set, the kernel uses 128-byte spacing
instead of 64, and with IORING_SETUP_CQE32 it uses 32-byte CQE
spacing instead of 16 bytes.

This was causing problems when attempting to use zig for a linux ublk
server.
This commit is contained in:
Brent Graveland 2025-11-20 15:17:47 +01:00
parent a9568ed296
commit ab78e489fe
No known key found for this signature in database
1 changed files with 29 additions and 2 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

31
lib/std/os/linux/IoUring.zig
View file

@ -139,7 +139,8 @@ pub fn get_sqe(self: *IoUring) !*linux.io_uring_sqe {
// We must therefore use wrapping addition and subtraction to avoid a runtime crash.
const next = self.sq.sqe_tail +% 1;
if (next -% head > self.sq.sqes.len) return error.SubmissionQueueFull;
const sqe = &self.sq.sqes[self.sq.sqe_tail & self.sq.mask];
// Use shift-based indexing for SQE128 to match liburing
const sqe = &self.sq.sqes[(self.sq.sqe_tail & self.sq.mask) << self.sq.sqe_shift];
self.sq.sqe_tail = next;
return sqe;
}
@ -284,6 +285,14 @@ fn copy_cqes_ready(self: *IoUring, cqes: []linux.io_uring_cqe) u32 {
const count = @min(cqes.len, ready);
const head = self.cq.head.* & self.cq.mask;
// Copy CQEs using shift-based indexing to match liburing
// When CQE32 is enabled, shift=1 doubles the index: cqes[(head & mask) << shift]
for (0..count) |i| {
const cqe_index = (head + @as(u32, @intCast(i))) & self.cq.mask;
const array_index = cqe_index << self.cq.cqe_shift;
cqes[i] = self.cq.cqes[array_index];
}
// before wrapping
const n = @min(self.cq.cqes.len - head, count);
@memcpy(cqes[0..n], self.cq.cqes[head..][0..n]);
@ -1512,6 +1521,7 @@ pub const SubmissionQueue = struct {
// This allows us to amortize the cost of the @atomicStore to `tail` across multiple SQEs.
sqe_head: u32 = 0,
sqe_tail: u32 = 0,
sqe_shift: u1, // Index shift for SQE128: 0 for normal, 1 for SQE128 (doubles index)
pub fn init(fd: linux.fd_t, p: linux.io_uring_params) !SubmissionQueue {
assert(fd >= 0);
@ -1533,7 +1543,12 @@ pub const SubmissionQueue = struct {
// The motivation for the `sqes` and `array` indirection is to make it possible for the
// application to preallocate static linux.io_uring_sqe entries and then replay them when needed.
const size_sqes = p.sq_entries * @sizeOf(linux.io_uring_sqe);
// With SQE128, each SQE is 128 bytes instead of 64 (kernel uses double spacing)
const sqe_size: usize = if ((p.flags & linux.IORING_SETUP_SQE128) != 0)
@sizeOf(linux.io_uring_sqe) * 2
else
@sizeOf(linux.io_uring_sqe);
const size_sqes = p.sq_entries * sqe_size;
const mmap_sqes = try posix.mmap(
null,
size_sqes,
@ -1545,6 +1560,10 @@ pub const SubmissionQueue = struct {
errdefer posix.munmap(mmap_sqes);
assert(mmap_sqes.len == size_sqes);
// Determine SQE shift: 1 if SQE128 is set (doubles array index for 128-byte stride)
// Matches liburing: sqe = &sqes[(sqe_tail & mask) << shift]
const sqe_shift: u1 = if ((p.flags & linux.IORING_SETUP_SQE128) != 0) 1 else 0;
const array: [*]u32 = @ptrCast(@alignCast(&mmap[p.sq_off.array]));
const sqes: [*]linux.io_uring_sqe = @ptrCast(@alignCast(&mmap_sqes[0]));
// We expect the kernel copies p.sq_entries to the u32 pointed to by p.sq_off.ring_entries,
@ -1560,6 +1579,7 @@ pub const SubmissionQueue = struct {
.sqes = sqes[0..p.sq_entries],
.mmap = mmap,
.mmap_sqes = mmap_sqes,
.sqe_shift = sqe_shift,
};
}
@ -1575,6 +1595,7 @@ pub const CompletionQueue = struct {
mask: u32,
overflow: *u32,
cqes: []linux.io_uring_cqe,
cqe_shift: u1, // Index shift for CQE32: 0 for normal, 1 for CQE32 (doubles index)
pub fn init(fd: linux.fd_t, p: linux.io_uring_params, sq: SubmissionQueue) !CompletionQueue {
assert(fd >= 0);
@ -1582,12 +1603,18 @@ pub const CompletionQueue = struct {
const mmap = sq.mmap;
const cqes: [*]linux.io_uring_cqe = @ptrCast(@alignCast(&mmap[p.cq_off.cqes]));
assert(p.cq_entries == @as(*u32, @ptrCast(@alignCast(&mmap[p.cq_off.ring_entries]))).*);
// Determine CQE shift: 1 if CQE32 flag is set (doubles array index for 32-byte stride)
// Matches liburing's implementation: cqe = &cqes[(head & mask) << shift]
const cqe_shift: u1 = if ((p.flags & linux.IORING_SETUP_CQE32) != 0) 1 else 0;
return CompletionQueue{
.head = @ptrCast(@alignCast(&mmap[p.cq_off.head])),
.tail = @ptrCast(@alignCast(&mmap[p.cq_off.tail])),
.mask = @as(*u32, @ptrCast(@alignCast(&mmap[p.cq_off.ring_mask]))).*,
.overflow = @ptrCast(@alignCast(&mmap[p.cq_off.overflow])),
.cqes = cqes[0..p.cq_entries],
.cqe_shift = cqe_shift,
};
}