Merge pull request #25608 from tchaikov/bluestore_fbsd

os/bluestore: support for FreeBSD

Reviewed-by: Willem Jan Withagen <wjw@digiware.nl>
Reviewed-by: Sage Weil <sage@redhat.com>

CMakeLists.txt

@@ -258,8 +258,13 @@ endif()
 
 option(WITH_BLUESTORE "Bluestore OSD backend" ON)
 if(WITH_BLUESTORE)
-  find_package(aio)
-  set(HAVE_LIBAIO ${AIO_FOUND})
+  if(LINUX)
+    find_package(aio)
+    set(HAVE_LIBAIO ${AIO_FOUND})
+  elseif(FREEBSD)
+    # POSIX AIO is integrated into FreeBSD kernel, and exposed by libc.
+    set(HAVE_POSIXAIO ON)
+  endif()
 endif()
 
 if(CMAKE_SYSTEM_PROCESSOR MATCHES "i386|i686|amd64|x86_64|AMD64|aarch64")
@@ -285,7 +290,7 @@ if(WITH_PMEM)
 endif()
 
 if(WITH_BLUESTORE)
-  if(NOT AIO_FOUND AND NOT WITH_SPDK AND NOT WITH_PMEM)
+  if(NOT AIO_FOUND AND NOT HAVE_POSIXAIO AND NOT WITH_SPDK AND NOT WITH_PMEM)
     message(SEND_ERROR "WITH_BLUESTORE is ON, "
       "but none of the bluestore backends is enabled. "
       "Please install libaio, or enable WITH_SPDK or WITH_PMEM (experimental)")

do_freebsd.sh

@@ -44,7 +44,6 @@ fi
 	-D WITH_BABELTRACE=OFF \
 	-D WITH_SEASTAR=OFF \
 	-D WITH_BLKID=OFF \
-	-D WITH_BLUESTORE=OFF \
 	-D WITH_FUSE=ON \
 	-D WITH_KRBD=OFF \
 	-D WITH_XFS=OFF \

src/include/config-h.in.cmake

@@ -93,6 +93,9 @@
 /* Defined if you have libaio */
 #cmakedefine HAVE_LIBAIO
 
+/* Defind if you have POSIX AIO */
+#cmakedefine HAVE_POSIXAIO
+
 /* Defined if OpenLDAP enabled */
 #cmakedefine HAVE_OPENLDAP
 

src/os/CMakeLists.txt

@@ -35,7 +35,7 @@ if(WITH_BLUESTORE)
   )
 endif(WITH_BLUESTORE)
 
-if(HAVE_LIBAIO)
+if(HAVE_LIBAIO OR HAVE_POSIXAIO)
   list(APPEND libos_srcs
     bluestore/KernelDevice.cc
     bluestore/aio.cc)

src/os/bluestore/BlockDevice.cc

@@ -19,7 +19,7 @@
 
 #include "BlockDevice.h"
 
-#if defined(HAVE_LIBAIO)
+#if defined(HAVE_LIBAIO) || defined(HAVE_POSIXAIO)
 #include "KernelDevice.h"
 #endif
 
@@ -63,7 +63,7 @@ uint64_t IOContext::get_num_ios() const
   // a configurable (with different hdd and ssd defaults), and add
   // that to the bytes value.
   uint64_t ios = 0;
-#ifdef HAVE_LIBAIO
+#if defined(HAVE_LIBAIO) || defined(HAVE_POSIXAIO)
   for (auto& p : pending_aios) {
     ios += p.iov.size();
   }
@@ -77,7 +77,7 @@ uint64_t IOContext::get_num_ios() const
 void IOContext::release_running_aios()
 {
   ceph_assert(!num_running);
-#ifdef HAVE_LIBAIO
+#if defined(HAVE_LIBAIO) || defined(HAVE_POSIXAIO)
   // release aio contexts (including pinned buffers).
   running_aios.clear();
 #endif
@@ -120,7 +120,7 @@ BlockDevice *BlockDevice::create(CephContext* cct, const string& path,
     return new PMEMDevice(cct, cb, cbpriv);
   }
 #endif
-#if defined(HAVE_LIBAIO)
+#if defined(HAVE_LIBAIO) || defined(HAVE_POSIXAIO)
   if (type == "kernel") {
     return new KernelDevice(cct, cb, cbpriv, d_cb, d_cbpriv);
   }

src/os/bluestore/BlockDevice.h

@@ -29,8 +29,8 @@
 #include "acconfig.h"
 #include "common/ceph_mutex.h"
 
-#ifdef HAVE_LIBAIO
-#include "aio.h"
+#if defined(HAVE_LIBAIO) || defined(HAVE_POSIXAIO)
+#include "ceph_aio.h"
 #endif
 #include "include/ceph_assert.h"
 #include "include/buffer.h"
@@ -70,7 +70,7 @@ public:
   std::atomic_int total_nseg = {0};
 #endif
 
-#ifdef HAVE_LIBAIO
+#if defined(HAVE_LIBAIO) || defined(HAVE_POSIXAIO)
   std::list<aio_t> pending_aios;    ///< not yet submitted
   std::list<aio_t> running_aios;    ///< submitting or submitted
 #endif

src/os/bluestore/KernelDevice.cc

@@ -24,6 +24,9 @@
 #include "include/stringify.h"
 #include "common/blkdev.h"
 #include "common/errno.h"
+#if defined(__FreeBSD__)
+#include "bsm/audit_errno.h"
+#endif
 #include "common/debug.h"
 #include "common/align.h"
 
@@ -451,9 +454,14 @@ static bool is_expected_ioerr(const int r)
 {
   // https://lxr.missinglinkelectronics.com/linux+v4.15/block/blk-core.c#L135
   return (r == -EOPNOTSUPP || r == -ETIMEDOUT || r == -ENOSPC ||
-	  r == -ENOLINK || r == -EREMOTEIO || r == -EBADE ||
+	  r == -ENOLINK || r == -EREMOTEIO  || r == -EAGAIN || r == -EIO ||
 	  r == -ENODATA || r == -EILSEQ || r == -ENOMEM ||
-	  r == -EAGAIN || r == -EREMCHG || r == -EIO);
+#if defined(__linux__)
+	  r == -EREMCHG || r == -EBADE
+#elif defined(__FreeBSD__)
+	  r == - BSM_ERRNO_EREMCHG || r == -BSM_ERRNO_EBADE
+#endif
+	  );
 }
 
 void KernelDevice::_aio_thread()
@@ -729,6 +737,7 @@ int KernelDevice::_sync_write(uint64_t off, bufferlist &bl, bool buffered, int w
     derr << __func__ << " pwritev error: " << cpp_strerror(r) << dendl;
     return r;
   }
+#ifdef HAVE_SYNC_FILE_RANGE
   if (buffered) {
     // initiate IO (but do not wait)
     r = ::sync_file_range(fd_buffereds[WRITE_LIFE_NOT_SET], off, len, SYNC_FILE_RANGE_WRITE);
@@ -738,6 +747,7 @@ int KernelDevice::_sync_write(uint64_t off, bufferlist &bl, bool buffered, int w
       return r;
     }
   }
+#endif
 
   io_since_flush.store(true);
 

src/os/bluestore/KernelDevice.h

@@ -22,7 +22,7 @@
 #include "common/Thread.h"
 #include "include/utime.h"
 
-#include "aio.h"
+#include "ceph_aio.h"
 #include "BlockDevice.h"
 
 class KernelDevice : public BlockDevice {

src/os/bluestore/PMEMDevice.h

@@ -21,7 +21,7 @@
 
 #include "os/fs/FS.h"
 #include "include/interval_set.h"
-#include "aio.h"
+#include "ceph_aio.h"
 #include "BlockDevice.h"
 
 class PMEMDevice : public BlockDevice {

src/os/bluestore/aio.cc

@@ -2,7 +2,7 @@
 // vim: ts=8 sw=2 smarttab
 
 #include <algorithm>
-#include "aio.h"
+#include "ceph_aio.h"
 
 std::ostream& operator<<(std::ostream& os, const aio_t& aio)
 {
@@ -22,20 +22,37 @@ int aio_queue_t::submit_batch(aio_iter begin, aio_iter end,
   // 2^16 * 125us = ~8 seconds, so max sleep is ~16 seconds
   int attempts = 16;
   int delay = 125;
+  int r;
 
   aio_iter cur = begin;
-  struct iocb *piocb[aios_size];
+  struct aio_t *piocb[aios_size];
   int left = 0;
   while (cur != end) {
     cur->priv = priv;
-    *(piocb+left) = &cur->iocb;
+    *(piocb+left) = &(*cur);
     ++left;
     ++cur;
   }
   ceph_assert(aios_size >= left);
   int done = 0;
   while (left > 0) {
-    int r = io_submit(ctx, std::min(left, max_iodepth), piocb + done);
+#if defined(HAVE_LIBAIO)
+    r = io_submit(ctx, std::min(left, max_iodepth), (struct iocb**)(piocb + done));
+#elif defined(HAVE_POSIXAIO)
+    if (piocb[done]->n_aiocb == 1) {
+      // TODO: consider batching multiple reads together with lio_listio
+      piocb[done]->aio.aiocb.aio_sigevent.sigev_notify = SIGEV_KEVENT;
+      piocb[done]->aio.aiocb.aio_sigevent.sigev_notify_kqueue = ctx;
+      piocb[done]->aio.aiocb.aio_sigevent.sigev_value.sival_ptr = piocb[done];
+      r = aio_read(&piocb[done]->aio.aiocb);
+    } else {
+      struct sigevent sev;
+      sev.sigev_notify = SIGEV_KEVENT;
+      sev.sigev_notify_kqueue = ctx;
+      sev.sigev_value.sival_ptr = piocb[done];
+      r = lio_listio(LIO_NOWAIT, &piocb[done]->aio.aiocbp, piocb[done]->n_aiocb, &sev);
+    }
+#endif
     if (r < 0) {
       if (r == -EAGAIN && attempts-- > 0) {
 	usleep(delay);
@@ -56,7 +73,11 @@ int aio_queue_t::submit_batch(aio_iter begin, aio_iter end,
 
 int aio_queue_t::get_next_completed(int timeout_ms, aio_t **paio, int max)
 {
-  io_event event[max];
+#if defined(HAVE_LIBAIO)
+  io_event events[max];
+#elif defined(HAVE_POSIXAIO)
+  struct kevent events[max];
+#endif
   struct timespec t = {
     timeout_ms / 1000,
     (timeout_ms % 1000) * 1000 * 1000
@@ -64,12 +85,40 @@ int aio_queue_t::get_next_completed(int timeout_ms, aio_t **paio, int max)
 
   int r = 0;
   do {
-    r = io_getevents(ctx, 1, max, event, &t);
+#if defined(HAVE_LIBAIO)
+    r = io_getevents(ctx, 1, max, events, &t);
+#elif defined(HAVE_POSIXAIO)
+    r = kevent(ctx, NULL, 0, events, max, &t);
+    if (r < 0)
+      r = -errno;
+#endif
   } while (r == -EINTR);
 
   for (int i=0; i<r; ++i) {
-    paio[i] = (aio_t *)event[i].obj;
-    paio[i]->rval = event[i].res;
+#if defined(HAVE_LIBAIO)
+    paio[i] = (aio_t *)events[i].obj;
+    paio[i]->rval = events[i].res;
+#else
+    paio[i] = (aio_t*)events[i].udata;
+    if (paio[i]->n_aiocb == 1) {
+      paio[i]->rval = aio_return(&paio[i]->aio.aiocb);
+    } else {
+      // Emulate the return value of pwritev.  I can't find any documentation
+      // for what the value of io_event.res is supposed to be.  I'm going to
+      // assume that it's just like pwritev/preadv/pwrite/pread.
+      paio[i]->rval = 0;
+      for (int j = 0; j < paio[i]->n_aiocb; j++) {
+	int res = aio_return(&paio[i]->aio.aiocbp[j]);
+	if (res < 0) {
+	  paio[i]->rval = res;
+	  break;
+	} else {
+	  paio[i]->rval += res;
+	}
+      }
+      free(paio[i]->aio.aiocbp);
+    }
+#endif
   }
   return r;
 }

src/os/bluestore/ceph_aio.h

@@ -2,7 +2,15 @@
 // vim: ts=8 sw=2 smarttab
 
 #pragma once
-# include <libaio.h>
+
+#include "acconfig.h"
+
+#if defined(HAVE_LIBAIO)
+#include <libaio.h>
+#elif defined(HAVE_POSIXAIO)
+#include <aio.h>
+#include <sys/event.h>
+#endif
 
 #include <boost/intrusive/list.hpp>
 #include <boost/container/small_vector.hpp>
@@ -11,7 +19,16 @@
 #include "include/types.h"
 
 struct aio_t {
+#if defined(HAVE_LIBAIO)
   struct iocb iocb{};  // must be first element; see shenanigans in aio_queue_t
+#elif defined(HAVE_POSIXAIO)
+  //  static long aio_listio_max = -1;
+  union {
+    struct aiocb aiocb;
+    struct aiocb *aiocbp;
+  } aio;
+  int n_aiocb;
+#endif
   void *priv;
   int fd;
   boost::container::small_vector<iovec,4> iov;
@@ -27,13 +44,34 @@ struct aio_t {
   void pwritev(uint64_t _offset, uint64_t len) {
     offset = _offset;
     length = len;
+#if defined(HAVE_LIBAIO)
     io_prep_pwritev(&iocb, fd, &iov[0], iov.size(), offset);
+#elif defined(HAVE_POSIXAIO)
+    n_aiocb = iov.size();
+    aio.aiocbp = (struct aiocb*)calloc(iov.size(), sizeof(struct aiocb));
+    for (int i = 0; i < iov.size(); i++) {
+      aio.aiocbp[i].aio_fildes = fd;
+      aio.aiocbp[i].aio_offset = offset;
+      aio.aiocbp[i].aio_buf = iov[i].iov_base;
+      aio.aiocbp[i].aio_nbytes = iov[i].iov_len;
+      aio.aiocbp[i].aio_lio_opcode = LIO_WRITE;
+      offset += iov[i].iov_len;
+    }
+#endif
   }
   void pread(uint64_t _offset, uint64_t len) {
     offset = _offset;
     length = len;
     bufferptr p = buffer::create_small_page_aligned(length);
+#if defined(HAVE_LIBAIO)
     io_prep_pread(&iocb, fd, p.c_str(), length, offset);
+#elif defined(HAVE_POSIXAIO)
+    n_aiocb = 1;
+    aio.aiocb.aio_fildes = fd;
+    aio.aiocb.aio_buf = p.c_str();
+    aio.aiocb.aio_nbytes = length;
+    aio.aiocb.aio_offset = offset;
+#endif
     bl.append(std::move(p));
   }
 
@@ -53,7 +91,11 @@ typedef boost::intrusive::list<
 
 struct aio_queue_t {
   int max_iodepth;
+#if defined(HAVE_LIBAIO)
   io_context_t ctx;
+#elif defined(HAVE_POSIXAIO)
+  int ctx;
+#endif
 
   typedef list<aio_t>::iterator aio_iter;
 
@@ -67,6 +109,7 @@ struct aio_queue_t {
 
   int init() {
     ceph_assert(ctx == 0);
+#if defined(HAVE_LIBAIO)
     int r = io_setup(max_iodepth, &ctx);
     if (r < 0) {
       if (ctx) {
@@ -75,10 +118,21 @@ struct aio_queue_t {
       }
     }
     return r;
+#elif defined(HAVE_POSIXAIO)
+    ctx = kqueue();
+    if (ctx < 0)
+      return -errno;
+    else
+      return 0;
+#endif
   }
   void shutdown() {
     if (ctx) {
+#if defined(HAVE_LIBAIO)
       int r = io_destroy(ctx);
+#elif defined(HAVE_POSIXAIO)
+      int r = close(ctx);
+#endif
       ceph_assert(r == 0);
       ctx = 0;
     }