diff --git a/lib/libspl/include/umem.h b/lib/libspl/include/umem.h index 0ed55ae5af..87db1f451e 100644 --- a/lib/libspl/include/umem.h +++ b/lib/libspl/include/umem.h @@ -37,6 +37,7 @@ */ #include +#include #ifdef __cplusplus extern "C" { @@ -82,9 +83,30 @@ umem_alloc(size_t size, int flags) { void *ptr; - ptr = malloc(size); - while (ptr == NULL && (flags & UMEM_NOFAIL)) + do { ptr = malloc(size); + } while (ptr == NULL && (flags & UMEM_NOFAIL)); + + return ptr; +} + +static inline void * +umem_alloc_aligned(size_t size, size_t align, int flags) +{ + void *ptr; + int rc; + + do { + rc = posix_memalign(&ptr, align, size); + } while (rc == ENOMEM && (flags & UMEM_NOFAIL)); + + if (rc == EINVAL) { + fprintf(stderr, "%s: invalid memory alignment (%zd)\n", + __func__, align); + if (flags & UMEM_NOFAIL) + abort(); + return NULL; + } return ptr; } @@ -146,7 +168,11 @@ umem_cache_alloc(umem_cache_t *cp, int flags) { void *ptr; - ptr = umem_alloc(cp->cache_bufsize, flags); + if (cp->cache_align != 0) + ptr = umem_alloc_aligned(cp->cache_bufsize, cp->cache_align, flags); + else + ptr = umem_alloc(cp->cache_bufsize, flags); + if (ptr && cp->cache_constructor) cp->cache_constructor(ptr, cp->cache_private, UMEM_DEFAULT); diff --git a/lib/libzpool/kernel.c b/lib/libzpool/kernel.c index 5f767e5949..d65c90a731 100644 --- a/lib/libzpool/kernel.c +++ b/lib/libzpool/kernel.c @@ -37,6 +37,7 @@ #include #include #include +#include /* for BLKGETSIZE64 */ #include /* @@ -552,13 +553,13 @@ vn_open(char *path, int x1, int flags, int mode, vnode_t **vpp, int x2, int x3) return (errno); } -#ifdef __linux__ if (!(flags & FCREAT) && S_ISBLK(st.st_mode)) { +#ifdef __linux__ flags |= O_DIRECT; - if (flags & FWRITE) - flags |= O_EXCL; - } #endif + /* We shouldn't be writing to block devices in userspace */ + VERIFY(!(flags & FWRITE)); + } if (flags & FCREAT) old_umask = umask(0); @@ -581,6 +582,16 @@ vn_open(char *path, int x1, int flags, int mode, vnode_t **vpp, int x2, int x3) return (err); } +#ifdef __linux__ + /* In Linux, use an ioctl to get the size of a block device. */ + if (S_ISBLK(st.st_mode)) { + if (ioctl(fd, BLKGETSIZE64, &st.st_size) != 0) { + err = errno; + close(fd); + return (err); + } + } +#endif (void) fcntl(fd, F_SETFD, FD_CLOEXEC); *vpp = vp = umem_zalloc(sizeof (vnode_t), UMEM_NOFAIL); @@ -634,6 +645,16 @@ vn_rdwr(int uio, vnode_t *vp, void *addr, ssize_t len, offset_t offset, } } +#ifdef __linux__ + if (rc == -1 && errno == EINVAL) { + /* + * Under Linux, this most likely means an alignment issue + * (memory or disk) due to O_DIRECT, so we abort() in order to + * catch the offender. + */ + abort(); + } +#endif if (rc == -1) return (errno);