From b0b65ec59384a7b6fe473d052a05ce72c55cc71b Mon Sep 17 00:00:00 2001 From: Patrick Steinhardt Date: Sun, 27 Oct 2024 16:39:37 +0100 Subject: [PATCH 1/3] compat/mingw: share file handles created via `CreateFileW()` Unless told otherwise, Windows will keep other processes from reading, writing and deleting files when one has an open handle that was created via `CreateFileW()`. This behaviour can be altered via `FILE_SHARE_*` flags: - `FILE_SHARE_READ` allows a concurrent process to open the file for reading. - `FILE_SHARE_WRITE` allows a concurrent process to open the file for writing. - `FILE_SHARE_DELETE` allows a concurrent process to delete the file or to replace it via an atomic rename. This sharing mechanism is quite important in the context of Git, as we assume POSIX semantics all over the place. But there are two callsites where we don't pass all three of these flags: - We don't set `FILE_SHARE_DELETE` when creating a file for appending via `mingw_open_append()`. This makes it impossible to delete the file from another process or to replace it via an atomic rename. The function was introduced via d641097589 (mingw: enable atomic O_APPEND, 2018-08-13) and has been using `FILE_SHARE_READ | FILE_SHARE_WRITE` since the inception. There aren't any indicators that the omission of `FILE_SHARE_DELETE` was intentional. - We don't set any sharing flags in `mingw_utime()`, which changes the access and modification of a file. This makes it impossible to perform any kind of operation on this file at all from another process. While we only open the file for a short amount of time to update its timestamps, this still opens us up for a race condition with another process. `mingw_utime()` was originally implemented via `_wopen()`, which doesn't give you full control over the sharing mode. Instead, it calls `_wsopen()` with `_SH_DENYNO`, which ultimately translates to `FILE_SHARE_READ | FILE_SHARE_WRITE`. It was then refactored via 090a3085bc (t/helper/test-chmtime: update mingw to support chmtime on directories, 2022-03-02) to use `CreateFileW()`, but we stopped setting any sharing flags at all, which seems like an unintentional side effect. By restoring `FILE_SHARE_READ | FILE_SHARE_WRITE` we thus fix this and get back the old behaviour of `_wopen()`. The fact that we didn't set the equivalent of `FILE_SHARE_DELETE` can be explained, as well: neither `_wopen()` nor `_wsopen()` allow you to do so. So overall, it doesn't seem intentional that we didn't allow deletions here, either. Adapt both of these callsites to pass all three sharing flags. Signed-off-by: Patrick Steinhardt Signed-off-by: Taylor Blau --- compat/mingw.c | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/compat/mingw.c b/compat/mingw.c index 0e851ecae2..e326c6fcd2 100644 --- a/compat/mingw.c +++ b/compat/mingw.c @@ -502,7 +502,7 @@ static int mingw_open_append(wchar_t const *wfilename, int oflags, ...) * to append to the file. */ handle = CreateFileW(wfilename, FILE_APPEND_DATA, - FILE_SHARE_WRITE | FILE_SHARE_READ, + FILE_SHARE_WRITE | FILE_SHARE_READ | FILE_SHARE_DELETE, NULL, create, FILE_ATTRIBUTE_NORMAL, NULL); if (handle == INVALID_HANDLE_VALUE) { DWORD err = GetLastError(); @@ -1006,7 +1006,7 @@ int mingw_utime (const char *file_name, const struct utimbuf *times) osfilehandle = CreateFileW(wfilename, FILE_WRITE_ATTRIBUTES, - 0 /*FileShare.None*/, + FILE_SHARE_READ | FILE_SHARE_WRITE | FILE_SHARE_DELETE, NULL, OPEN_EXISTING, (attrs != INVALID_FILE_ATTRIBUTES && From a270cb1540a7f82d0c8c1de9e4103cb09e8642a8 Mon Sep 17 00:00:00 2001 From: Patrick Steinhardt Date: Sun, 27 Oct 2024 16:39:40 +0100 Subject: [PATCH 2/3] compat/mingw: allow deletion of most opened files On Windows, we emulate open(3p) via `mingw_open()`. This function implements handling of some platform-specific quirks that are required to make it behave as closely as possible like open(3p) would, but for most cases we just call the Windows-specific `_wopen()` function. This function has a major downside though: it does not allow us to specify the sharing mode. While there is `_wsopen()` that allows us to pass sharing flags, those sharing flags are not the same `FILE_SHARE_*` flags as `CreateFileW()` accepts. Instead, `_wsopen()` only allows concurrent read- and write-access, but does not allow for concurrent deletions. Unfortunately though, we have to allow concurrent deletions if we want to have POSIX-style atomic renames on top of an existing file that has open file handles. Implement a new function that emulates open(3p) for existing files via `CreateFileW()` such that we can set the required sharing flags. While we have the same issue when calling open(3p) with `O_CREAT`, implementing that mode would be more complex due to the required permission handling. Furthermore, atomic updates via renames typically write to exclusive lockfile and then perform the rename, and thus we don't have to handle the case where the locked path has been created with `O_CREATE`. So while it would be nice to have proper POSIX semantics in all paths, we instead aim for a minimum viable fix here. Signed-off-by: Patrick Steinhardt Signed-off-by: Taylor Blau --- compat/mingw.c | 66 ++++++++++++++++++++++++++++++++++++++++++++++++++ 1 file changed, 66 insertions(+) diff --git a/compat/mingw.c b/compat/mingw.c index e326c6fcd2..0d9600543c 100644 --- a/compat/mingw.c +++ b/compat/mingw.c @@ -532,6 +532,70 @@ static int mingw_open_append(wchar_t const *wfilename, int oflags, ...) return fd; } +/* + * Ideally, we'd use `_wopen()` to implement this functionality so that we + * don't have to reimplement it, but unfortunately we do not have tight control + * over the share mode there. And while `_wsopen()` and friends exist that give + * us _some_ control over the share mode, this family of functions doesn't give + * us the ability to enable FILE_SHARE_DELETE, either. But this is a strict + * requirement for us though so that we can unlink or rename over files that + * are held open by another process. + * + * We are thus forced to implement our own emulation of `open()`. To make our + * life simpler we only implement basic support for this, namely opening + * existing files for reading and/or writing. This means that newly created + * files won't have their sharing mode set up correctly, but for now I couldn't + * find any case where this matters. We may have to revisit that in the future + * though based on our needs. + */ +static int mingw_open_existing(const wchar_t *filename, int oflags, ...) +{ + SECURITY_ATTRIBUTES security_attributes = { + .nLength = sizeof(security_attributes), + .bInheritHandle = !(oflags & O_NOINHERIT), + }; + HANDLE handle; + DWORD access; + int fd; + + /* We only support basic flags. */ + if (oflags & ~(O_ACCMODE | O_NOINHERIT)) { + errno = ENOSYS; + return -1; + } + + switch (oflags & O_ACCMODE) { + case O_RDWR: + access = GENERIC_READ | GENERIC_WRITE; + break; + case O_WRONLY: + access = GENERIC_WRITE; + break; + default: + access = GENERIC_READ; + break; + } + + handle = CreateFileW(filename, access, + FILE_SHARE_WRITE | FILE_SHARE_READ | FILE_SHARE_DELETE, + &security_attributes, OPEN_EXISTING, FILE_ATTRIBUTE_NORMAL, NULL); + if (handle == INVALID_HANDLE_VALUE) { + DWORD err = GetLastError(); + + /* See `mingw_open_append()` for why we have this conversion. */ + if (err == ERROR_INVALID_PARAMETER) + err = ERROR_PATH_NOT_FOUND; + + errno = err_win_to_posix(err); + return -1; + } + + fd = _open_osfhandle((intptr_t)handle, oflags | O_BINARY); + if (fd < 0) + CloseHandle(handle); + return fd; +} + /* * Does the pathname map to the local named pipe filesystem? * That is, does it have a "//./pipe/" prefix? @@ -567,6 +631,8 @@ int mingw_open (const char *filename, int oflags, ...) if ((oflags & O_APPEND) && !is_local_named_pipe_path(filename)) open_fn = mingw_open_append; + else if (!(oflags & ~(O_ACCMODE | O_NOINHERIT))) + open_fn = mingw_open_existing; else open_fn = _wopen; From 391bceae4350136a05d977573caeaa07059f2136 Mon Sep 17 00:00:00 2001 From: Patrick Steinhardt Date: Sun, 27 Oct 2024 16:39:43 +0100 Subject: [PATCH 3/3] compat/mingw: support POSIX semantics for atomic renames By default, Windows restricts access to files when those files have been opened by another process. As explained in the preceding commits, these restrictions can be loosened such that reads, writes and/or deletes of files with open handles _are_ allowed. While we set up those sharing flags in most relevant code paths now, we still don't properly handle POSIX-style atomic renames in case the target path is open. This is failure demonstrated by t0610, where one of our tests spawns concurrent writes in a reftable-enabled repository and expects all of them to succeed. This test fails most of the time because the process that has acquired the "tables.list" lock is unable to rename it into place while other processes are busy reading that file. Windows 10 has introduced the `FILE_RENAME_FLAG_POSIX_SEMANTICS` flag that allows us to fix this usecase [1]. When set, it is possible to rename a file over a preexisting file even when the target file still has handles open. Those handles must have been opened with the `FILE_SHARE_DELETE` flag, which we have ensured in the preceding commits. Careful readers might have noticed that [1] does not mention the above flag, but instead mentions `FILE_RENAME_POSIX_SEMANTICS`. This flag is not for use with `SetFileInformationByHandle()` though, which is what we use. And while the `FILE_RENAME_FLAG_POSIX_SEMANTICS` flag exists, it is not documented on [2] or anywhere else as far as I can tell. Unfortunately, we still support Windows systems older than Windows 10 that do not yet have this new flag. Our `_WIN32_WINNT` SDK version still targets 0x0600, which is Windows Vista and later. And even though that Windows version is out-of-support, bumping the SDK version all the way to 0x0A00, which is Windows 10 and later, is not an option as it would make it impossible to compile on Windows 8.1, which is still supported. Instead, we have to manually declare the relevant infrastructure to make this feature available and have fallback logic in place in case we run on a Windows version that does not yet have this flag. On another note: `mingw_rename()` has a retry loop that is used in case deleting a file failed because it's still open in another process. One might be pressed to not use this loop anymore when we can use POSIX semantics. But unfortunately, we have to keep it around due to our dependence on the `FILE_SHARE_DELETE` flag. While we know to set that sharing flag now, other applications may not do so and may thus still cause sharing violations when we try to rename a file. This fixes concurrent writes in the reftable backend as demonstrated in t0610, but may also end up fixing other usecases where Git wants to perform renames. [1]: https://learn.microsoft.com/en-us/windows-hardware/drivers/ddi/ntifs/ns-ntifs-_file_rename_information [2]: https://learn.microsoft.com/en-us/windows/win32/api/winbase/ns-winbase-file_rename_info Signed-off-by: Patrick Steinhardt Signed-off-by: Taylor Blau Reviewed-by: Johannes Sixt Signed-off-by: Junio C Hamano --- compat/mingw.c | 87 ++++++++++++++++++++++++++++++++++++-- t/t0610-reftable-basics.sh | 8 ++-- 2 files changed, 88 insertions(+), 7 deletions(-) diff --git a/compat/mingw.c b/compat/mingw.c index 0d9600543c..c4320769db 100644 --- a/compat/mingw.c +++ b/compat/mingw.c @@ -2217,10 +2217,16 @@ int mingw_accept(int sockfd1, struct sockaddr *sa, socklen_t *sz) #undef rename int mingw_rename(const char *pold, const char *pnew) { + static int supports_file_rename_info_ex = 1; DWORD attrs, gle; int tries = 0; wchar_t wpold[MAX_PATH], wpnew[MAX_PATH]; - if (xutftowcs_path(wpold, pold) < 0 || xutftowcs_path(wpnew, pnew) < 0) + int wpnew_len; + + if (xutftowcs_path(wpold, pold) < 0) + return -1; + wpnew_len = xutftowcs_path(wpnew, pnew); + if (wpnew_len < 0) return -1; /* @@ -2231,11 +2237,84 @@ int mingw_rename(const char *pold, const char *pnew) return 0; if (errno != EEXIST) return -1; + repeat: - if (MoveFileExW(wpold, wpnew, MOVEFILE_REPLACE_EXISTING)) - return 0; + if (supports_file_rename_info_ex) { + /* + * Our minimum required Windows version is still set to Windows + * Vista. We thus have to declare required infrastructure for + * FileRenameInfoEx ourselves until we bump _WIN32_WINNT to + * 0x0A00. Furthermore, we have to handle cases where the + * FileRenameInfoEx call isn't supported yet. + */ +#define FILE_RENAME_FLAG_REPLACE_IF_EXISTS 0x00000001 +#define FILE_RENAME_FLAG_POSIX_SEMANTICS 0x00000002 + FILE_INFO_BY_HANDLE_CLASS FileRenameInfoEx = 22; + struct { + /* + * This is usually an unnamed union, but that is not + * part of ISO C99. We thus inline the field, as we + * really only care for the Flags field anyway. + */ + DWORD Flags; + HANDLE RootDirectory; + DWORD FileNameLength; + /* + * The actual structure is defined with a single-character + * flex array so that the structure has to be allocated on + * the heap. As we declare this structure ourselves though + * we can avoid the allocation and define FileName to have + * MAX_PATH bytes. + */ + WCHAR FileName[MAX_PATH]; + } rename_info = { 0 }; + HANDLE old_handle = INVALID_HANDLE_VALUE; + BOOL success; + + old_handle = CreateFileW(wpold, DELETE, + FILE_SHARE_WRITE | FILE_SHARE_READ | FILE_SHARE_DELETE, + NULL, OPEN_EXISTING, FILE_ATTRIBUTE_NORMAL, NULL); + if (old_handle == INVALID_HANDLE_VALUE) { + errno = err_win_to_posix(GetLastError()); + return -1; + } + + rename_info.Flags = FILE_RENAME_FLAG_REPLACE_IF_EXISTS | + FILE_RENAME_FLAG_POSIX_SEMANTICS; + rename_info.FileNameLength = wpnew_len * sizeof(WCHAR); + memcpy(rename_info.FileName, wpnew, wpnew_len * sizeof(WCHAR)); + + success = SetFileInformationByHandle(old_handle, FileRenameInfoEx, + &rename_info, sizeof(rename_info)); + gle = GetLastError(); + CloseHandle(old_handle); + if (success) + return 0; + + /* + * When we see ERROR_INVALID_PARAMETER we can assume that the + * current system doesn't support FileRenameInfoEx. Keep us + * from using it in future calls and retry. + */ + if (gle == ERROR_INVALID_PARAMETER) { + supports_file_rename_info_ex = 0; + goto repeat; + } + + /* + * In theory, we shouldn't get ERROR_ACCESS_DENIED because we + * always open files with FILE_SHARE_DELETE But in practice we + * cannot assume that Git is the only one accessing files, and + * other applications may not set FILE_SHARE_DELETE. So we have + * to retry. + */ + } else { + if (MoveFileExW(wpold, wpnew, MOVEFILE_REPLACE_EXISTING)) + return 0; + gle = GetLastError(); + } + /* TODO: translate more errors */ - gle = GetLastError(); if (gle == ERROR_ACCESS_DENIED && (attrs = GetFileAttributesW(wpnew)) != INVALID_FILE_ATTRIBUTES) { if (attrs & FILE_ATTRIBUTE_DIRECTORY) { diff --git a/t/t0610-reftable-basics.sh b/t/t0610-reftable-basics.sh index babec7993e..eaf6fab6d2 100755 --- a/t/t0610-reftable-basics.sh +++ b/t/t0610-reftable-basics.sh @@ -450,10 +450,12 @@ test_expect_success 'ref transaction: retry acquiring tables.list lock' ' ) ' -# This test fails most of the time on Windows systems. The root cause is +# This test fails most of the time on Cygwin systems. The root cause is # that Windows does not allow us to rename the "tables.list.lock" file into -# place when "tables.list" is open for reading by a concurrent process. -test_expect_success !WINDOWS 'ref transaction: many concurrent writers' ' +# place when "tables.list" is open for reading by a concurrent process. We have +# worked around that in our MinGW-based rename emulation, but the Cygwin +# emulation seems to be insufficient. +test_expect_success !CYGWIN 'ref transaction: many concurrent writers' ' test_when_finished "rm -rf repo" && git init repo && (