Re: [PATCH v3 22/33] linux-user: Split out mmap_h_lt_g
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 > > > --- > > > 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~ >
Re: [PATCH v3 22/33] linux-user: Split out mmap_h_lt_g
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 --- 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~
Re: [PATCH v3 22/33] linux-user: Split out mmap_h_lt_g
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 --- 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~
Re: [PATCH v3 22/33] linux-user: Split out mmap_h_lt_g
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 > --- > 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.
Re: [PATCH v3 22/33] linux-user: Split out mmap_h_lt_g
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 --- 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 3b8329..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, ) == -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,
[PATCH v3 22/33] linux-user: Split out mmap_h_lt_g
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 --- 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 3b8329..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, ) == -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) @@