reuse a buffer for read/write-through-propolis-heap

this seems to have some effect, but not a lot; nothing when I/Os are all
the same size, on the order of 5% better throughput when I/Os vary
between 4 KiB, 32 KiB, 256 KiB, and in-between

Author: iximeow

lib/propolis/src/block/file.rs

@@ -59,6 +59,24 @@ impl SharedState {
     }
 
     fn processing_loop(&self, wctx: SyncWorkerCtx) {
+        // Conspire with gross hack in `preadv()` and `pwritev()`: there, we
+        // want to give the OS a buffer backed by Propolis heap, rather than a
+        // VMM segment, to avoid a slow fallback to lock pages in memory during
+        // the I/O. We could either allocate buffers on-demand sized for the
+        // I/O, or up front in the worker thread, here. On-demand allocation
+        // works, but when I/O sizes are mixed the high malloc/free rates
+        // (X00k/sec/process) get a bit busy on locks in libumem. So instead,
+        // retain a buffer across I/Os, which is passed into `process_request()`
+        // and on.
+        //
+        // This buffer is grown on-demand when handling an I/O because some
+        // guests' maximum I/O size seems to be smaller than maximums we
+        // indicate (for example, Linux seems to send NVMe I/Os 256 KiB or
+        // smaller, even though we report we'll tolerate up to 1 MiB). In those
+        // cases, sizing against an expected maximum will grossly oversize
+        // buffers and increase memory pressure for no good reason.
+        let mut scratch = Vec::new();
+
         while let Some(dreq) = wctx.block_for_req() {
             let req = dreq.req();
             if self.info.read_only && req.op.is_write() {
@@ -74,7 +92,8 @@ impl SharedState {
                 dreq.complete(block::Result::Failure);
                 continue;
             };
-            let res = match self.process_request(&req, &mem) {
+            let res = match self.process_request(&req, &mem, Some(&mut scratch))
+            {
                 Ok(_) => block::Result::Success,
                 Err(_) => block::Result::Failure,
             };
@@ -86,13 +105,14 @@ impl SharedState {
         &self,
         req: &block::Request,
         mem: &MemCtx,
+        scratch: Option<&mut Vec<u8>>,
     ) -> std::result::Result<(), &'static str> {
         match req.op {
             block::Operation::Read(off, len) => {
                 let maps = req.mappings(mem).ok_or("mapping unavailable")?;
 
                 let nbytes = maps
-                    .preadv(self.fp.as_raw_fd(), off as i64)
+                    .preadv(self.fp.as_raw_fd(), off as i64, scratch)
                     .map_err(|_| "io error")?;
                 if nbytes != len {
                     return Err("bad read length");
@@ -102,7 +122,7 @@ impl SharedState {
                 let maps = req.mappings(mem).ok_or("bad guest region")?;
 
                 let nbytes = maps
-                    .pwritev(self.fp.as_raw_fd(), off as i64)
+                    .pwritev(self.fp.as_raw_fd(), off as i64, scratch)
                     .map_err(|_| "io error")?;
                 if nbytes != len {
                     return Err("bad write length");

lib/propolis/src/vmm/mem.rs

@@ -781,10 +781,20 @@ unsafe impl Sync for SubMapping<'_> {}
 
 pub trait MappingExt {
     /// preadv from `file` into multiple mappings
-    fn preadv(&self, fd: RawFd, offset: i64) -> Result<usize>;
+    fn preadv(
+        &self,
+        fd: RawFd,
+        offset: i64,
+        scratch: Option<&mut Vec<u8>>,
+    ) -> Result<usize>;
 
     /// pwritev from multiple mappings to `file`
-    fn pwritev(&self, fd: RawFd, offset: i64) -> Result<usize>;
+    fn pwritev(
+        &self,
+        fd: RawFd,
+        offset: i64,
+        scratch: Option<&mut Vec<u8>>,
+    ) -> Result<usize>;
 }
 
 // Gross hack alert: since the mappings below are memory regions backed by
@@ -825,8 +835,45 @@ fn total_mapping_size(mappings: &[SubMapping<'_>]) -> Result<usize> {
         })
 }
 
+/// Ensure the maybe-provided buffer is ready for I/O of `desired_size`.
+///
+/// Returns `Some` with a slice with length at least as large as `desired_size`,
+/// if the desired size is in the range we're willing to use the I/O buffer
+/// proxying workaround.
+///
+/// If the scratch buffer is provided, was too small, but the desired size is
+/// acceptable, this function will grow the scratch buffer.
+fn prepare_scratch_buffer(
+    scratch: Option<&mut Vec<u8>>,
+    desired_size: usize,
+) -> Option<&mut [u8]> {
+    scratch.and_then(|scratch| {
+        if desired_size > MAPPING_IO_LIMIT_BYTES {
+            // Even though the caller provided a buffer, we're not willing to
+            // grow it to handle this I/O.
+            return None;
+        }
+
+        if scratch.len() < desired_size {
+            // The provided buffer is smaller than necessary, but the needed
+            // size is acceptable, so grow the buffer.
+            //
+            // Buffer growth is done lazily because experience shows that guests
+            // may submit I/Os much smaller than the largest permitted by MDTS.
+            scratch.resize(desired_size, 0);
+        }
+
+        Some(scratch.as_mut_slice())
+    })
+}
+
 impl<'a, T: AsRef<[SubMapping<'a>]>> MappingExt for T {
-    fn preadv(&self, fd: RawFd, offset: i64) -> Result<usize> {
+    fn preadv(
+        &self,
+        fd: RawFd,
+        offset: i64,
+        scratch: Option<&mut Vec<u8>>,
+    ) -> Result<usize> {
         if !self
             .as_ref()
             .iter()
@@ -841,16 +888,11 @@ impl<'a, T: AsRef<[SubMapping<'a>]>> MappingExt for T {
         let total_capacity = total_mapping_size(self.as_ref())?;
 
         // Gross hack: see the comment on `MAPPING_IO_LIMIT_BYTES`.
-        if total_capacity <= MAPPING_IO_LIMIT_BYTES {
-            // If we're motivated to avoid the zero-fill via
-            // `Layout::with_size_align` + `GlobalAlloc::alloc`, we should
-            // probably avoid this gross hack entirely (see comment on
-            // MAPPING_IO_LIMIT_BYTES).
-            let mut buf = vec![0; total_capacity];
-
+        let scratch = prepare_scratch_buffer(scratch, total_capacity);
+        if let Some(buf) = scratch {
             let iov = [iovec {
                 iov_base: buf.as_mut_ptr() as *mut libc::c_void,
-                iov_len: buf.len(),
+                iov_len: total_capacity,
             }];
 
             let res = unsafe {
@@ -904,7 +946,12 @@ impl<'a, T: AsRef<[SubMapping<'a>]>> MappingExt for T {
         }
     }
 
-    fn pwritev(&self, fd: RawFd, offset: i64) -> Result<usize> {
+    fn pwritev(
+        &self,
+        fd: RawFd,
+        offset: i64,
+        scratch: Option<&mut Vec<u8>>,
+    ) -> Result<usize> {
         if !self
             .as_ref()
             .iter()
@@ -919,14 +966,9 @@ impl<'a, T: AsRef<[SubMapping<'a>]>> MappingExt for T {
         let total_capacity = total_mapping_size(self.as_ref())?;
 
         // Gross hack: see the comment on `MAPPING_IO_LIMIT_BYTES`.
-        let written = if total_capacity <= MAPPING_IO_LIMIT_BYTES {
-            // If we're motivated to avoid the zero-fill via
-            // `Layout::with_size_align` + `GlobalAlloc::alloc`, we should
-            // probably avoid this gross hack entirely (see comment on
-            // MAPPING_IO_LIMIT_BYTES).
-            let mut buf = vec![0; total_capacity];
-
-            let mut remaining = buf.as_mut_slice();
+        let scratch = prepare_scratch_buffer(scratch, total_capacity);
+        let written = if let Some(buf) = scratch {
+            let mut remaining = &mut buf[..];
             for mapping in self.as_ref().iter() {
                 // The original `buf` is at least as large as all mappings
                 // combined, so `remaining` is at least as large as this and all
@@ -940,7 +982,7 @@ impl<'a, T: AsRef<[SubMapping<'a>]>> MappingExt for T {
 
             let iovs = [iovec {
                 iov_base: buf.as_mut_ptr() as *mut libc::c_void,
-                iov_len: buf.len(),
+                iov_len: total_capacity,
             }];
 
             unsafe {