| Message ID | 20240102015808.132373-23-richard.henderson@linaro.org |
|---|---|
| State | Superseded |
| Headers | show |
| Series | linux-user: Improve host and guest page size handling | expand |
On 1/2/24 05:57, Richard Henderson wrote: > Work much harder to get alignment and mapping beyond the end > of the file correct. Both of which are excercised by our > test-mmap for alpha (8k pages) on any 4k page host. > > Signed-off-by: Richard Henderson <richard.henderson@linaro.org> > --- > linux-user/mmap.c | 156 +++++++++++++++++++++++++++++++++++++--------- > 1 file changed, 125 insertions(+), 31 deletions(-) > > diff --git a/linux-user/mmap.c b/linux-user/mmap.c > index 00003b8329..8b0a26e50d 100644 > --- a/linux-user/mmap.c > +++ b/linux-user/mmap.c > @@ -552,6 +552,128 @@ static abi_long mmap_h_eq_g(abi_ulong start, abi_ulong len, > return mmap_end(start, last, start, last, flags, page_flags); > } > > +/* > + * Special case host page size < target page size. > + * > + * The two special cases are increased guest alignment, and mapping > + * past the end of a file. > + * > + * When mapping files into a memory area larger than the file, > + * accesses to pages beyond the file size will cause a SIGBUS. > + * > + * For example, if mmaping a file of 100 bytes on a host with 4K > + * pages emulating a target with 8K pages, the target expects to > + * be able to access the first 8K. But the host will trap us on > + * any access beyond 4K. > + * > + * When emulating a target with a larger page-size than the hosts, > + * we may need to truncate file maps at EOF and add extra anonymous > + * pages up to the targets page boundary. > + * > + * This workaround only works for files that do not change. > + * If the file is later extended (e.g. ftruncate), the SIGBUS > + * vanishes and the proper behaviour is that changes within the > + * anon page should be reflected in the file. > + * > + * However, this case is rather common with executable images, > + * so the workaround is important for even trivial tests, whereas > + * the mmap of of a file being extended is less common. > + */ > +static abi_long mmap_h_lt_g(abi_ulong start, abi_ulong len, int host_prot, > + int mmap_flags, int page_flags, int fd, > + off_t offset, int host_page_size) > +{ > + void *p, *want_p = g2h_untagged(start); > + off_t fileend_adj = 0; > + int flags = mmap_flags; > + abi_ulong last, pass_last; > + > + if (!(flags & MAP_ANONYMOUS)) { > + struct stat sb; > + > + if (fstat(fd, &sb) == -1) { > + return -1; > + } > + if (offset >= sb.st_size) { > + /* > + * The entire map is beyond the end of the file. > + * Transform it to an anonymous mapping. > + */ > + flags |= MAP_ANONYMOUS; > + fd = -1; > + offset = 0; > + } else if (offset + len > sb.st_size) { > + /* > + * A portion of the map is beyond the end of the file. > + * Truncate the file portion of the allocation. > + */ > + fileend_adj = offset + len - sb.st_size; > + } > + } > + > + if (flags & (MAP_FIXED | MAP_FIXED_NOREPLACE)) { > + if (fileend_adj) { > + p = mmap(want_p, len, host_prot, flags | MAP_ANONYMOUS, -1, 0); > + } else { > + p = mmap(want_p, len, host_prot, flags, fd, offset); > + } > + if (p != want_p) { > + if (p != MAP_FAILED) { > + munmap(p, len); > + errno = EEXIST; > + } > + return -1; > + } > + > + if (fileend_adj) { > + void *t = mmap(p, len - fileend_adj, host_prot, > + (flags & ~MAP_FIXED_NOREPLACE) | MAP_FIXED, > + fd, offset); > + assert(t != MAP_FAILED); > + } > + } else { > + size_t host_len, part_len; > + > + /* > + * Take care to align the host memory. Perform a larger anonymous > + * allocation and extract the aligned portion. Remap the file on > + * top of that. > + */ > + host_len = len + TARGET_PAGE_SIZE - host_page_size; > + p = mmap(want_p, host_len, host_prot, flags | MAP_ANONYMOUS, -1, 0); > + if (p == MAP_FAILED) { > + return -1; > + } > + > + part_len = (uintptr_t)p & (TARGET_PAGE_SIZE - 1); > + if (part_len) { > + part_len = TARGET_PAGE_SIZE - part_len; > + munmap(p, part_len); > + p += part_len; > + host_len -= part_len; > + } > + if (len < host_len) { > + munmap(p + len, host_len - len); > + } > + > + if (!(flags & MAP_ANONYMOUS)) { > + void *t = mmap(p, len - fileend_adj, host_prot, > + flags | MAP_FIXED, fd, offset); > + assert(t != MAP_FAILED); > + } > + > + start = h2g(p); > + } > + > + last = start + len - 1; > + if (fileend_adj) { > + pass_last = ROUND_UP(last - fileend_adj, host_page_size) - 1; > + } else { > + pass_last = last; > + } > + return mmap_end(start, last, start, pass_last, mmap_flags, page_flags); > +} > + > static abi_long target_mmap__locked(abi_ulong start, abi_ulong len, > int target_prot, int flags, int page_flags, > int fd, off_t offset) > @@ -596,37 +718,9 @@ static abi_long target_mmap__locked(abi_ulong start, abi_ulong len, > if (host_page_size == TARGET_PAGE_SIZE) { > return mmap_h_eq_g(start, len, host_prot, flags, > page_flags, fd, offset); > - } > - > - /* > - * When mapping files into a memory area larger than the file, accesses > - * to pages beyond the file size will cause a SIGBUS. > - * > - * For example, if mmaping a file of 100 bytes on a host with 4K pages > - * emulating a target with 8K pages, the target expects to be able to > - * access the first 8K. But the host will trap us on any access beyond > - * 4K. > - * > - * When emulating a target with a larger page-size than the hosts, we > - * may need to truncate file maps at EOF and add extra anonymous pages > - * up to the targets page boundary. > - */ > - if (host_page_size < TARGET_PAGE_SIZE && !(flags & MAP_ANONYMOUS)) { > - struct stat sb; > - > - if (fstat(fd, &sb) == -1) { > - return -1; > - } > - > - /* Are we trying to create a map beyond EOF?. */ > - if (offset + len > sb.st_size) { > - /* > - * If so, truncate the file map at eof aligned with > - * the hosts real pagesize. Additional anonymous maps > - * will be created beyond EOF. > - */ > - len = ROUND_UP(sb.st_size - offset, host_page_size); > - } > + } else if (host_page_size < TARGET_PAGE_SIZE) { > + return mmap_h_lt_g(start, len, host_prot, flags, > + page_flags, fd, offset, host_page_size); > } > > if (!(flags & (MAP_FIXED | MAP_FIXED_NOREPLACE))) { Reviewed-by: Pierrick Bouvier <pierrick.bouvier@linaro.org>
On Tue, Jan 02, 2024 at 12:57:57PM +1100, Richard Henderson wrote: > Work much harder to get alignment and mapping beyond the end > of the file correct. Both of which are excercised by our > test-mmap for alpha (8k pages) on any 4k page host. > > Signed-off-by: Richard Henderson <richard.henderson@linaro.org> > --- > linux-user/mmap.c | 156 +++++++++++++++++++++++++++++++++++++--------- > 1 file changed, 125 insertions(+), 31 deletions(-) [...] > + if (fileend_adj) { > + void *t = mmap(p, len - fileend_adj, host_prot, > + (flags & ~MAP_FIXED_NOREPLACE) | MAP_FIXED, > + fd, offset); > + assert(t != MAP_FAILED); Is it possible to recover here? Of course, we are remapping the memory we've mapped a few lines earlier, but asserting the syscall result looks a bit odd. [...] > + if (!(flags & MAP_ANONYMOUS)) { > + void *t = mmap(p, len - fileend_adj, host_prot, > + flags | MAP_FIXED, fd, offset); > + assert(t != MAP_FAILED); Same here.
On 1/29/24 05:26, Ilya Leoshkevich wrote: > On Tue, Jan 02, 2024 at 12:57:57PM +1100, Richard Henderson wrote: >> Work much harder to get alignment and mapping beyond the end >> of the file correct. Both of which are excercised by our >> test-mmap for alpha (8k pages) on any 4k page host. >> >> Signed-off-by: Richard Henderson <richard.henderson@linaro.org> >> --- >> linux-user/mmap.c | 156 +++++++++++++++++++++++++++++++++++++--------- >> 1 file changed, 125 insertions(+), 31 deletions(-) > > [...] > >> + if (fileend_adj) { >> + void *t = mmap(p, len - fileend_adj, host_prot, >> + (flags & ~MAP_FIXED_NOREPLACE) | MAP_FIXED, >> + fd, offset); >> + assert(t != MAP_FAILED); > > Is it possible to recover here? Of course, we are remapping the memory > we've mapped a few lines earlier, but asserting the syscall result > looks a bit odd. > Can you think of a failure mode? I couldn't. That's why I added the assert. I suppose there's the always present threat of running out of vmas... r~
On 1/29/24 05:26, Ilya Leoshkevich wrote: > On Tue, Jan 02, 2024 at 12:57:57PM +1100, Richard Henderson wrote: >> Work much harder to get alignment and mapping beyond the end >> of the file correct. Both of which are excercised by our >> test-mmap for alpha (8k pages) on any 4k page host. >> >> Signed-off-by: Richard Henderson <richard.henderson@linaro.org> >> --- >> linux-user/mmap.c | 156 +++++++++++++++++++++++++++++++++++++--------- >> 1 file changed, 125 insertions(+), 31 deletions(-) > > [...] > >> + if (fileend_adj) { >> + void *t = mmap(p, len - fileend_adj, host_prot, >> + (flags & ~MAP_FIXED_NOREPLACE) | MAP_FIXED, >> + fd, offset); >> + assert(t != MAP_FAILED); > > Is it possible to recover here? Of course, we are remapping the memory > we've mapped a few lines earlier, but asserting the syscall result > looks a bit odd. This first one we cannot recover from, because we've already (potentially) overwritten the previous memory mapping. >> + if (!(flags & MAP_ANONYMOUS)) { >> + void *t = mmap(p, len - fileend_adj, host_prot, >> + flags | MAP_FIXED, fd, offset); >> + assert(t != MAP_FAILED); > > Same here. This one we could, because the memory was previously unmapped. r~
On Tue, 2024-02-13 at 09:54 -1000, Richard Henderson wrote: > On 1/29/24 05:26, Ilya Leoshkevich wrote: > > On Tue, Jan 02, 2024 at 12:57:57PM +1100, Richard Henderson wrote: > > > Work much harder to get alignment and mapping beyond the end > > > of the file correct. Both of which are excercised by our > > > test-mmap for alpha (8k pages) on any 4k page host. > > > > > > Signed-off-by: Richard Henderson <richard.henderson@linaro.org> > > > --- > > > linux-user/mmap.c | 156 +++++++++++++++++++++++++++++++++++++-- > > > ------- > > > 1 file changed, 125 insertions(+), 31 deletions(-) > > > > [...] > > > > > + if (fileend_adj) { > > > + void *t = mmap(p, len - fileend_adj, host_prot, > > > + (flags & ~MAP_FIXED_NOREPLACE) | > > > MAP_FIXED, > > > + fd, offset); > > > + assert(t != MAP_FAILED); > > > > Is it possible to recover here? Of course, we are remapping the > > memory > > we've mapped a few lines earlier, but asserting the syscall result > > looks a bit odd. > > > > Can you think of a failure mode? I couldn't. > That's why I added the assert. > > I suppose there's the always present threat of running out of vmas... Right, and this should be easy to trigger by using ulimit -v. > > > r~ >
diff --git a/linux-user/mmap.c b/linux-user/mmap.c index 00003b8329..8b0a26e50d 100644 --- a/linux-user/mmap.c +++ b/linux-user/mmap.c @@ -552,6 +552,128 @@ static abi_long mmap_h_eq_g(abi_ulong start, abi_ulong len, return mmap_end(start, last, start, last, flags, page_flags); } +/* + * Special case host page size < target page size. + * + * The two special cases are increased guest alignment, and mapping + * past the end of a file. + * + * When mapping files into a memory area larger than the file, + * accesses to pages beyond the file size will cause a SIGBUS. + * + * For example, if mmaping a file of 100 bytes on a host with 4K + * pages emulating a target with 8K pages, the target expects to + * be able to access the first 8K. But the host will trap us on + * any access beyond 4K. + * + * When emulating a target with a larger page-size than the hosts, + * we may need to truncate file maps at EOF and add extra anonymous + * pages up to the targets page boundary. + * + * This workaround only works for files that do not change. + * If the file is later extended (e.g. ftruncate), the SIGBUS + * vanishes and the proper behaviour is that changes within the + * anon page should be reflected in the file. + * + * However, this case is rather common with executable images, + * so the workaround is important for even trivial tests, whereas + * the mmap of of a file being extended is less common. + */ +static abi_long mmap_h_lt_g(abi_ulong start, abi_ulong len, int host_prot, + int mmap_flags, int page_flags, int fd, + off_t offset, int host_page_size) +{ + void *p, *want_p = g2h_untagged(start); + off_t fileend_adj = 0; + int flags = mmap_flags; + abi_ulong last, pass_last; + + if (!(flags & MAP_ANONYMOUS)) { + struct stat sb; + + if (fstat(fd, &sb) == -1) { + return -1; + } + if (offset >= sb.st_size) { + /* + * The entire map is beyond the end of the file. + * Transform it to an anonymous mapping. + */ + flags |= MAP_ANONYMOUS; + fd = -1; + offset = 0; + } else if (offset + len > sb.st_size) { + /* + * A portion of the map is beyond the end of the file. + * Truncate the file portion of the allocation. + */ + fileend_adj = offset + len - sb.st_size; + } + } + + if (flags & (MAP_FIXED | MAP_FIXED_NOREPLACE)) { + if (fileend_adj) { + p = mmap(want_p, len, host_prot, flags | MAP_ANONYMOUS, -1, 0); + } else { + p = mmap(want_p, len, host_prot, flags, fd, offset); + } + if (p != want_p) { + if (p != MAP_FAILED) { + munmap(p, len); + errno = EEXIST; + } + return -1; + } + + if (fileend_adj) { + void *t = mmap(p, len - fileend_adj, host_prot, + (flags & ~MAP_FIXED_NOREPLACE) | MAP_FIXED, + fd, offset); + assert(t != MAP_FAILED); + } + } else { + size_t host_len, part_len; + + /* + * Take care to align the host memory. Perform a larger anonymous + * allocation and extract the aligned portion. Remap the file on + * top of that. + */ + host_len = len + TARGET_PAGE_SIZE - host_page_size; + p = mmap(want_p, host_len, host_prot, flags | MAP_ANONYMOUS, -1, 0); + if (p == MAP_FAILED) { + return -1; + } + + part_len = (uintptr_t)p & (TARGET_PAGE_SIZE - 1); + if (part_len) { + part_len = TARGET_PAGE_SIZE - part_len; + munmap(p, part_len); + p += part_len; + host_len -= part_len; + } + if (len < host_len) { + munmap(p + len, host_len - len); + } + + if (!(flags & MAP_ANONYMOUS)) { + void *t = mmap(p, len - fileend_adj, host_prot, + flags | MAP_FIXED, fd, offset); + assert(t != MAP_FAILED); + } + + start = h2g(p); + } + + last = start + len - 1; + if (fileend_adj) { + pass_last = ROUND_UP(last - fileend_adj, host_page_size) - 1; + } else { + pass_last = last; + } + return mmap_end(start, last, start, pass_last, mmap_flags, page_flags); +} + static abi_long target_mmap__locked(abi_ulong start, abi_ulong len, int target_prot, int flags, int page_flags, int fd, off_t offset) @@ -596,37 +718,9 @@ static abi_long target_mmap__locked(abi_ulong start, abi_ulong len, if (host_page_size == TARGET_PAGE_SIZE) { return mmap_h_eq_g(start, len, host_prot, flags, page_flags, fd, offset); - } - - /* - * When mapping files into a memory area larger than the file, accesses - * to pages beyond the file size will cause a SIGBUS. - * - * For example, if mmaping a file of 100 bytes on a host with 4K pages - * emulating a target with 8K pages, the target expects to be able to - * access the first 8K. But the host will trap us on any access beyond - * 4K. - * - * When emulating a target with a larger page-size than the hosts, we - * may need to truncate file maps at EOF and add extra anonymous pages - * up to the targets page boundary. - */ - if (host_page_size < TARGET_PAGE_SIZE && !(flags & MAP_ANONYMOUS)) { - struct stat sb; - - if (fstat(fd, &sb) == -1) { - return -1; - } - - /* Are we trying to create a map beyond EOF?. */ - if (offset + len > sb.st_size) { - /* - * If so, truncate the file map at eof aligned with - * the hosts real pagesize. Additional anonymous maps - * will be created beyond EOF. - */ - len = ROUND_UP(sb.st_size - offset, host_page_size); - } + } else if (host_page_size < TARGET_PAGE_SIZE) { + return mmap_h_lt_g(start, len, host_prot, flags, + page_flags, fd, offset, host_page_size); } if (!(flags & (MAP_FIXED | MAP_FIXED_NOREPLACE))) {
Work much harder to get alignment and mapping beyond the end of the file correct. Both of which are excercised by our test-mmap for alpha (8k pages) on any 4k page host. Signed-off-by: Richard Henderson <richard.henderson@linaro.org> --- linux-user/mmap.c | 156 +++++++++++++++++++++++++++++++++++++--------- 1 file changed, 125 insertions(+), 31 deletions(-)