adaptive fuse syscall switching (a1209bc3) · Commits · hpc / gekkofs

include/client/hooks.hpp

+14 −0

Original line number	Diff line number	Diff line
		@@ -47,6 +47,8 @@ extern "C" {
		#include <sys/syscall.h>
		}

		#include <unordered_map>

		#include <libsyscall_intercept_hook_point.h>

		#include <client/intercept.hpp>
		@@ -73,6 +75,18 @@ syscall_no_intercept_wrapper(long syscall_number, Args... args) {

		struct statfs;

		struct fd_state {
		off_t last_off = -1;
		size_t last_size = -1;
		int seq_count = 0;
		int random_count = 0;
		bool is_random = false;
		};
		static std::unordered_map<unsigned int, fd_state> fd_state_map;

		/// returns false to stay in syscall intercept, true to go to fuse
		static bool update_fd_state(unsigned int fd, off_t offset, size_t size);

		namespace gkfs::hook {

		int

src/client/hooks.cpp

+126 −8

Original line number	Diff line number	Diff line
		@@ -69,6 +69,56 @@ with_errno(T ret) {

		} // namespace

		static bool
		update_fd_state(unsigned int fd, off_t offset, size_t size) {
		auto it = fd_state_map.find(fd);
		if(it == fd_state_map.end()) {
		// No state for this FD, stay in syscall intercept
		return false;
		}
		auto& st = it->second;
		if(st.last_off == -1) {
		// first observation
		st.last_off = offset;
		st.last_size = size;
		return false;
		}

		const off_t expected = st.last_off + st.last_size;
		const off_t diff = offset - expected;

		// configurable thresholds
		constexpr off_t small_delta = 4096; // ≤4 KB jump = still sequential
		constexpr off_t big_jump_threshold =
		128 * 1024; // >128 KB = definitely random

		if(std::abs(diff) <= small_delta) {
		st.seq_count++;
		} else if(std::abs(diff) > big_jump_threshold) {
		st.random_count++;
		} else {
		// medium jump: weak indication of randomness
		st.random_count++;
		}

		if(st.random_count > st.seq_count * 2) {
		st.is_random = true;
		} else {
		st.is_random = false;
		}

		st.last_off = offset;
		st.last_size = size;
		return !st.is_random;
		}

		static size_t
		total_iovec_size(const struct iovec* iov, int iovcnt) {
		size_t total = 0;
		for(int i = 0; i < iovcnt; ++i) total += iov[i].iov_len;
		return total;
		}

		namespace gkfs::hook {

		int
		@@ -107,6 +157,7 @@ hook_openat(int dirfd, const char* cpath, int flags, mode_t mode) {
		CTX->file_map()->add(
		fd, std::make_shared<gkfs::filemap::OpenFile>(resolved,
		flags));
		fd_state_map[fd] = fd_state{};
		return with_errno(fd);
		}
		}
		@@ -122,6 +173,7 @@ hook_close(int fd) {
		LOG(DEBUG, "{}() called with fd: {}", __func__, fd);

		auto ret = gkfs::syscall::gkfs_close(fd);
		fd_state_map.erase(fd);

		if(ret < 0)
		LOG(DEBUG, "{}() close failed with fd: {}, errno {}", __func__, fd,
		@@ -257,7 +309,19 @@ hook_read(unsigned int fd, void* buf, size_t count) {
		fmt::ptr(buf), count);

		if(CTX->file_map()->exist(fd)) {
		auto gkfs_fd = CTX->file_map()->get(fd); // retrieve the current offset
		auto pos = gkfs_fd->pos();
		bool use_syscall = update_fd_state(fd, pos, count);
		if(use_syscall) {
		return with_errno(gkfs::syscall::gkfs_read(fd, buf, count));
		} else {
		int ret = syscall_no_intercept_wrapper(SYS_pread64, fd, buf, count,
		pos);
		if(ret > 0) {
		gkfs_fd->pos(pos + ret);
		}
		return ret;
		}
		}
		return syscall_no_intercept_wrapper(SYS_read, fd, buf, count);
		}
		@@ -269,8 +333,11 @@ hook_pread(unsigned int fd, char* buf, size_t count, loff_t pos) {
		fd, fmt::ptr(buf), count, pos);

		if(CTX->file_map()->exist(fd)) {
		bool use_syscall = update_fd_state(fd, pos, count);
		if(use_syscall) {
		return with_errno(gkfs::syscall::gkfs_pread(fd, buf, count, pos));
		}
		}
		/* Since kernel 2.6: pread() became pread64(), and pwrite() became
		* pwrite64(). */
		return syscall_no_intercept_wrapper(SYS_pread64, fd, buf, count, pos);
		@@ -283,7 +350,20 @@ hook_readv(unsigned long fd, const struct iovec* iov, unsigned long iovcnt) {
		fmt::ptr(iov), iovcnt);

		if(CTX->file_map()->exist(fd)) {
		auto gkfs_fd = CTX->file_map()->get(fd); // retrieve the current offset
		auto pos = gkfs_fd->pos();
		bool use_syscall =
		update_fd_state(fd, pos, total_iovec_size(iov, iovcnt));
		if(use_syscall) {
		return with_errno(gkfs::syscall::gkfs_readv(fd, iov, iovcnt));
		} else {
		int ret = syscall_no_intercept_wrapper(SYS_preadv, fd, iov, iovcnt,
		pos);
		if(ret > 0) {
		gkfs_fd->pos(pos + ret);
		}
		return ret;
		}
		}
		return syscall_no_intercept_wrapper(SYS_readv, fd, iov, iovcnt);
		}
		@@ -299,7 +379,12 @@ hook_preadv(unsigned long fd, const struct iovec* iov, unsigned long iovcnt,
		__func__, fd, fmt::ptr(iov), iovcnt, pos_l, pos_h);

		if(CTX->file_map()->exist(fd)) {
		return with_errno(gkfs::syscall::gkfs_preadv(fd, iov, iovcnt, pos_l));
		bool use_syscall =
		update_fd_state(fd, pos_l, total_iovec_size(iov, iovcnt));
		if(use_syscall) {
		return with_errno(
		gkfs::syscall::gkfs_preadv(fd, iov, iovcnt, pos_l));
		}
		}
		return syscall_no_intercept_wrapper(SYS_preadv, fd, iov, iovcnt, pos_l);
		}
		@@ -311,7 +396,19 @@ hook_write(unsigned int fd, const char* buf, size_t count) {
		fmt::ptr(buf), count);

		if(CTX->file_map()->exist(fd)) {
		auto gkfs_fd = CTX->file_map()->get(fd); // retrieve the current offset
		auto pos = gkfs_fd->pos();
		bool use_syscall = update_fd_state(fd, pos, count);
		if(use_syscall) {
		return with_errno(gkfs::syscall::gkfs_write(fd, buf, count));
		} else {
		int ret = syscall_no_intercept_wrapper(SYS_pwrite64, fd, buf, count,
		pos);
		if(ret > 0) {
		gkfs_fd->pos(pos + ret);
		}
		return ret;
		}
		}
		return syscall_no_intercept_wrapper(SYS_write, fd, buf, count);
		}
		@@ -323,8 +420,11 @@ hook_pwrite(unsigned int fd, const char* buf, size_t count, loff_t pos) {
		fd, fmt::ptr(buf), count, pos);

		if(CTX->file_map()->exist(fd)) {
		bool use_syscall = update_fd_state(fd, pos, count);
		if(use_syscall) {
		return with_errno(gkfs::syscall::gkfs_pwrite(fd, buf, count, pos));
		}
		}
		/* Since kernel 2.6: pread() became pread64(), and pwrite() became
		* pwrite64(). */
		return syscall_no_intercept_wrapper(SYS_pwrite64, fd, buf, count, pos);
		@@ -337,7 +437,20 @@ hook_writev(unsigned long fd, const struct iovec* iov, unsigned long iovcnt) {
		fmt::ptr(iov), iovcnt);

		if(CTX->file_map()->exist(fd)) {
		auto gkfs_fd = CTX->file_map()->get(fd); // retrieve the current offset
		auto pos = gkfs_fd->pos();
		bool use_syscall =
		update_fd_state(fd, pos, total_iovec_size(iov, iovcnt));
		if(use_syscall) {
		return with_errno(gkfs::syscall::gkfs_writev(fd, iov, iovcnt));
		} else {
		int ret = syscall_no_intercept_wrapper(SYS_pwritev, fd, iov, iovcnt,
		pos);
		if(ret > 0) {
		gkfs_fd->pos(pos + ret);
		}
		return ret;
		}
		}
		return syscall_no_intercept_wrapper(SYS_writev, fd, iov, iovcnt);
		}
		@@ -353,7 +466,12 @@ hook_pwritev(unsigned long fd, const struct iovec* iov, unsigned long iovcnt,
		__func__, fd, fmt::ptr(iov), iovcnt, pos_l, pos_h);

		if(CTX->file_map()->exist(fd)) {
		return with_errno(gkfs::syscall::gkfs_pwritev(fd, iov, iovcnt, pos_l));
		bool use_syscall =
		update_fd_state(fd, pos_l, total_iovec_size(iov, iovcnt));
		if(use_syscall) {
		return with_errno(
		gkfs::syscall::gkfs_pwritev(fd, iov, iovcnt, pos_l));
		}
		}
		return syscall_no_intercept_wrapper(SYS_pwritev, fd, iov, iovcnt, pos_l);
		}