| Message ID | 20240102015808.132373-24-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| State | Superseded |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: > Signed-off-by: Richard Henderson <richard.henderson@linaro.org> > --- > linux-user/mmap.c | 288 ++++++++++++++++++++++------------------------ > 1 file changed, 139 insertions(+), 149 deletions(-) > > diff --git a/linux-user/mmap.c b/linux-user/mmap.c > index 8b0a26e50d..552656edd4 100644 > --- a/linux-user/mmap.c > +++ b/linux-user/mmap.c > @@ -267,7 +267,16 @@ int target_mprotect(abi_ulong start, abi_ulong len, int target_prot) > return ret; > } > > -/* map an incomplete host page */ > +/* > + * Map an incomplete host page. > + * > + * Here be dragons. This case will not work if there is an existing > + * overlapping host page, which is file mapped, and for which the mapping > + * is beyond the end of the file. In that case, we will see SIGBUS when > + * trying to write a portion of this page. > + * > + * FIXME: Work around this with a temporary signal handler and longjmp. > + */ > static bool mmap_frag(abi_ulong real_start, abi_ulong start, abi_ulong last, > int prot, int flags, int fd, off_t offset) > { > @@ -674,19 +683,138 @@ static abi_long mmap_h_lt_g(abi_ulong start, abi_ulong len, int host_prot, > return mmap_end(start, last, start, pass_last, mmap_flags, page_flags); > } > > +/* > + * Special case host page size > target page size. > + * > + * The two special cases are address and file offsets that are valid > + * for the guest that cannot be directly represented by the host. > + */ > +static abi_long mmap_h_gt_g(abi_ulong start, abi_ulong len, > + int target_prot, int host_prot, > + int flags, int page_flags, int fd, > + off_t offset, int host_page_size) > +{ > + void *p, *want_p = g2h_untagged(start); > + off_t host_offset = offset & -host_page_size; > + abi_ulong last, real_start, real_last; > + bool misaligned_offset = false; > + size_t host_len; > + > + if (!(flags & (MAP_FIXED | MAP_FIXED_NOREPLACE))) { > + /* > + * Adjust the offset to something representable on the host. > + */ > + host_len = len + offset - host_offset; > + p = mmap(want_p, host_len, host_prot, flags, fd, host_offset); > + if (p == MAP_FAILED) { > + return -1; > + } > + > + /* Update start to the file position at offset. */ > + p += offset - host_offset; > + > + start = h2g(p); > + last = start + len - 1; > + return mmap_end(start, last, start, last, flags, page_flags); > + } > + > + if (!(flags & MAP_ANONYMOUS)) { > + misaligned_offset = (start ^ offset) & (host_page_size - 1); > + > + /* > + * The fallback for misalignment is a private mapping + read. > + * This carries none of semantics required of MAP_SHARED. > + */ > + if (misaligned_offset && (flags & MAP_TYPE) != MAP_PRIVATE) { > + errno = EINVAL; > + return -1; > + } > + } > + > + last = start + len - 1; > + real_start = start & -host_page_size; > + real_last = ROUND_UP(last, host_page_size) - 1; > + > + /* > + * Handle the start and end of the mapping. > + */ > + if (real_start < start) { > + abi_ulong real_page_last = real_start + host_page_size - 1; > + if (last <= real_page_last) { > + /* Entire allocation a subset of one host page. */ > + if (!mmap_frag(real_start, start, last, target_prot, > + flags, fd, offset)) { > + return -1; > + } > + return mmap_end(start, last, -1, 0, flags, page_flags); > + } > + > + if (!mmap_frag(real_start, start, real_page_last, target_prot, > + flags, fd, offset)) { > + return -1; > + } > + real_start = real_page_last + 1; > + } > + > + if (last < real_last) { > + abi_ulong real_page_start = real_last - host_page_size + 1; > + if (!mmap_frag(real_page_start, real_page_start, last, > + target_prot, flags, fd, > + offset + real_page_start - start)) { > + return -1; > + } > + real_last = real_page_start - 1; > + } > + > + if (real_start > real_last) { > + return mmap_end(start, last, -1, 0, flags, page_flags); > + } > + > + /* > + * Handle the middle of the mapping. > + */ > + > + host_len = real_last - real_start + 1; > + want_p += real_start - start; > + > + if (flags & MAP_ANONYMOUS) { > + p = mmap(want_p, host_len, host_prot, flags, -1, 0); > + } else if (!misaligned_offset) { > + p = mmap(want_p, host_len, host_prot, flags, fd, > + offset + real_start - start); > + } else { > + p = mmap(want_p, host_len, host_prot | PROT_WRITE, > + flags | MAP_ANONYMOUS, -1, 0); > + } > + if (p != want_p) { > + if (p != MAP_FAILED) { > + munmap(p, host_len); > + errno = EEXIST; > + } > + return -1; > + } > + > + if (misaligned_offset) { > + /* TODO: The read could be short. */ > + if (pread(fd, p, host_len, offset + real_start - start) != host_len) { > + munmap(p, host_len); > + return -1; > + } > + if (!(host_prot & PROT_WRITE)) { > + mprotect(p, host_len, host_prot); > + } > + } > + > + return mmap_end(start, last, -1, 0, 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) > { > int host_page_size = qemu_real_host_page_size(); > - abi_ulong ret, last, real_start, real_last, retaddr, host_len; > - abi_ulong passthrough_start = -1, passthrough_last = 0; > - off_t host_offset; > int host_prot; > > - real_start = start & -host_page_size; > - host_offset = offset & -host_page_size; > - > /* > * For reserved_va, we are in full control of the allocation. > * Find a suitible hole and convert to MAP_FIXED. > @@ -700,6 +828,8 @@ static abi_long target_mmap__locked(abi_ulong start, abi_ulong len, > } > flags = (flags & ~MAP_FIXED_NOREPLACE) | MAP_FIXED; > } else if (!(flags & MAP_FIXED)) { > + abi_ulong real_start = start & -host_page_size; > + off_t host_offset = offset & -host_page_size; > size_t real_len = len + offset - host_offset; > abi_ulong align = MAX(host_page_size, TARGET_PAGE_SIZE); > > @@ -721,150 +851,10 @@ static abi_long target_mmap__locked(abi_ulong start, abi_ulong len, > } 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))) { > - uintptr_t host_start; > - void *p; > - > - host_len = len + offset - host_offset; > - host_len = ROUND_UP(host_len, host_page_size); > - > - /* Note: we prefer to control the mapping address. */ > - p = mmap(g2h_untagged(start), host_len, host_prot, > - flags | MAP_FIXED | MAP_ANONYMOUS, -1, 0); > - if (p == MAP_FAILED) { > - return -1; > - } > - /* update start so that it points to the file position at 'offset' */ > - host_start = (uintptr_t)p; > - if (!(flags & MAP_ANONYMOUS)) { > - p = mmap(g2h_untagged(start), len, host_prot, > - flags | MAP_FIXED, fd, host_offset); > - if (p == MAP_FAILED) { > - munmap(g2h_untagged(start), host_len); > - return -1; > - } > - host_start += offset - host_offset; > - } > - start = h2g(host_start); > - last = start + len - 1; > - passthrough_start = start; > - passthrough_last = last; > } else { > - last = start + len - 1; > - real_last = ROUND_UP(last, host_page_size) - 1; > - > - if (flags & MAP_FIXED_NOREPLACE) { > - /* Validate that the chosen range is empty. */ > - if (!page_check_range_empty(start, last)) { > - errno = EEXIST; > - return -1; > - } > - > - /* > - * With reserved_va, the entire address space is mmaped in the > - * host to ensure it isn't accidentally used for something else. > - * We have just checked that the guest address is not mapped > - * within the guest, but need to replace the host reservation. > - * > - * Without reserved_va, despite the guest address check above, > - * keep MAP_FIXED_NOREPLACE so that the guest does not overwrite > - * any host address mappings. > - */ > - if (reserved_va) { > - flags = (flags & ~MAP_FIXED_NOREPLACE) | MAP_FIXED; > - } > - } > - > - /* > - * worst case: we cannot map the file because the offset is not > - * aligned, so we read it > - */ > - if (!(flags & MAP_ANONYMOUS) && > - (offset & (host_page_size - 1)) != (start & (host_page_size - 1))) { > - /* > - * msync() won't work here, so we return an error if write is > - * possible while it is a shared mapping > - */ > - if ((flags & MAP_TYPE) == MAP_SHARED > - && (target_prot & PROT_WRITE)) { > - errno = EINVAL; > - return -1; > - } > - retaddr = target_mmap(start, len, target_prot | PROT_WRITE, > - (flags & (MAP_FIXED | MAP_FIXED_NOREPLACE)) > - | MAP_PRIVATE | MAP_ANONYMOUS, > - -1, 0); > - if (retaddr == -1) { > - return -1; > - } > - if (pread(fd, g2h_untagged(start), len, offset) == -1) { > - return -1; > - } > - if (!(target_prot & PROT_WRITE)) { > - ret = target_mprotect(start, len, target_prot); > - assert(ret == 0); > - } > - return mmap_end(start, last, -1, 0, flags, page_flags); > - } > - > - /* handle the start of the mapping */ > - if (start > real_start) { > - if (real_last == real_start + host_page_size - 1) { > - /* one single host page */ > - if (!mmap_frag(real_start, start, last, > - target_prot, flags, fd, offset)) { > - return -1; > - } > - return mmap_end(start, last, -1, 0, flags, page_flags); > - } > - if (!mmap_frag(real_start, start, > - real_start + host_page_size - 1, > - target_prot, flags, fd, offset)) { > - return -1; > - } > - real_start += host_page_size; > - } > - /* handle the end of the mapping */ > - if (last < real_last) { > - abi_ulong real_page = real_last - host_page_size + 1; > - if (!mmap_frag(real_page, real_page, last, > - target_prot, flags, fd, > - offset + real_page - start)) { > - return -1; > - } > - real_last -= host_page_size; > - } > - > - /* map the middle (easier) */ > - if (real_start < real_last) { > - void *p, *want_p; > - off_t offset1; > - size_t len1; > - > - if (flags & MAP_ANONYMOUS) { > - offset1 = 0; > - } else { > - offset1 = offset + real_start - start; > - } > - len1 = real_last - real_start + 1; > - want_p = g2h_untagged(real_start); > - > - p = mmap(want_p, len1, host_prot, flags, fd, offset1); > - if (p != want_p) { > - if (p != MAP_FAILED) { > - munmap(p, len1); > - errno = EEXIST; > - } > - return -1; > - } > - passthrough_start = real_start; > - passthrough_last = real_last; > - } > + return mmap_h_gt_g(start, len, target_prot, host_prot, flags, > + page_flags, fd, offset, host_page_size); > } > - return mmap_end(start, last, passthrough_start, passthrough_last, > - flags, page_flags); > } > > /* NOTE: all the constants are the HOST ones */ Reviewed-by: Pierrick Bouvier <pierrick.bouvier@linaro.org>
On Tue, Jan 02, 2024 at 12:57:58PM +1100, Richard Henderson wrote: > Signed-off-by: Richard Henderson <richard.henderson@linaro.org> > --- > linux-user/mmap.c | 288 ++++++++++++++++++++++------------------------ > 1 file changed, 139 insertions(+), 149 deletions(-) > > diff --git a/linux-user/mmap.c b/linux-user/mmap.c > index 8b0a26e50d..552656edd4 100644 > --- a/linux-user/mmap.c > +++ b/linux-user/mmap.c [...] > + if (last < real_last) { > + abi_ulong real_page_start = real_last - host_page_size + 1; > + if (!mmap_frag(real_page_start, real_page_start, last, > + target_prot, flags, fd, > + offset + real_page_start - start)) { Should we unmap start here and in a few places below? [...]
diff --git a/linux-user/mmap.c b/linux-user/mmap.c index 8b0a26e50d..552656edd4 100644 --- a/linux-user/mmap.c +++ b/linux-user/mmap.c @@ -267,7 +267,16 @@ int target_mprotect(abi_ulong start, abi_ulong len, int target_prot) return ret; } -/* map an incomplete host page */ +/* + * Map an incomplete host page. + * + * Here be dragons. This case will not work if there is an existing + * overlapping host page, which is file mapped, and for which the mapping + * is beyond the end of the file. In that case, we will see SIGBUS when + * trying to write a portion of this page. + * + * FIXME: Work around this with a temporary signal handler and longjmp. + */ static bool mmap_frag(abi_ulong real_start, abi_ulong start, abi_ulong last, int prot, int flags, int fd, off_t offset) { @@ -674,19 +683,138 @@ static abi_long mmap_h_lt_g(abi_ulong start, abi_ulong len, int host_prot, return mmap_end(start, last, start, pass_last, mmap_flags, page_flags); } +/* + * Special case host page size > target page size. + * + * The two special cases are address and file offsets that are valid + * for the guest that cannot be directly represented by the host. + */ +static abi_long mmap_h_gt_g(abi_ulong start, abi_ulong len, + int target_prot, int host_prot, + int flags, int page_flags, int fd, + off_t offset, int host_page_size) +{ + void *p, *want_p = g2h_untagged(start); + off_t host_offset = offset & -host_page_size; + abi_ulong last, real_start, real_last; + bool misaligned_offset = false; + size_t host_len; + + if (!(flags & (MAP_FIXED | MAP_FIXED_NOREPLACE))) { + /* + * Adjust the offset to something representable on the host. + */ + host_len = len + offset - host_offset; + p = mmap(want_p, host_len, host_prot, flags, fd, host_offset); + if (p == MAP_FAILED) { + return -1; + } + + /* Update start to the file position at offset. */ + p += offset - host_offset; + + start = h2g(p); + last = start + len - 1; + return mmap_end(start, last, start, last, flags, page_flags); + } + + if (!(flags & MAP_ANONYMOUS)) { + misaligned_offset = (start ^ offset) & (host_page_size - 1); + + /* + * The fallback for misalignment is a private mapping + read. + * This carries none of semantics required of MAP_SHARED. + */ + if (misaligned_offset && (flags & MAP_TYPE) != MAP_PRIVATE) { + errno = EINVAL; + return -1; + } + } + + last = start + len - 1; + real_start = start & -host_page_size; + real_last = ROUND_UP(last, host_page_size) - 1; + + /* + * Handle the start and end of the mapping. + */ + if (real_start < start) { + abi_ulong real_page_last = real_start + host_page_size - 1; + if (last <= real_page_last) { + /* Entire allocation a subset of one host page. */ + if (!mmap_frag(real_start, start, last, target_prot, + flags, fd, offset)) { + return -1; + } + return mmap_end(start, last, -1, 0, flags, page_flags); + } + + if (!mmap_frag(real_start, start, real_page_last, target_prot, + flags, fd, offset)) { + return -1; + } + real_start = real_page_last + 1; + } + + if (last < real_last) { + abi_ulong real_page_start = real_last - host_page_size + 1; + if (!mmap_frag(real_page_start, real_page_start, last, + target_prot, flags, fd, + offset + real_page_start - start)) { + return -1; + } + real_last = real_page_start - 1; + } + + if (real_start > real_last) { + return mmap_end(start, last, -1, 0, flags, page_flags); + } + + /* + * Handle the middle of the mapping. + */ + + host_len = real_last - real_start + 1; + want_p += real_start - start; + + if (flags & MAP_ANONYMOUS) { + p = mmap(want_p, host_len, host_prot, flags, -1, 0); + } else if (!misaligned_offset) { + p = mmap(want_p, host_len, host_prot, flags, fd, + offset + real_start - start); + } else { + p = mmap(want_p, host_len, host_prot | PROT_WRITE, + flags | MAP_ANONYMOUS, -1, 0); + } + if (p != want_p) { + if (p != MAP_FAILED) { + munmap(p, host_len); + errno = EEXIST; + } + return -1; + } + + if (misaligned_offset) { + /* TODO: The read could be short. */ + if (pread(fd, p, host_len, offset + real_start - start) != host_len) { + munmap(p, host_len); + return -1; + } + if (!(host_prot & PROT_WRITE)) { + mprotect(p, host_len, host_prot); + } + } + + return mmap_end(start, last, -1, 0, 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) { int host_page_size = qemu_real_host_page_size(); - abi_ulong ret, last, real_start, real_last, retaddr, host_len; - abi_ulong passthrough_start = -1, passthrough_last = 0; - off_t host_offset; int host_prot; - real_start = start & -host_page_size; - host_offset = offset & -host_page_size; - /* * For reserved_va, we are in full control of the allocation. * Find a suitible hole and convert to MAP_FIXED. @@ -700,6 +828,8 @@ static abi_long target_mmap__locked(abi_ulong start, abi_ulong len, } flags = (flags & ~MAP_FIXED_NOREPLACE) | MAP_FIXED; } else if (!(flags & MAP_FIXED)) { + abi_ulong real_start = start & -host_page_size; + off_t host_offset = offset & -host_page_size; size_t real_len = len + offset - host_offset; abi_ulong align = MAX(host_page_size, TARGET_PAGE_SIZE); @@ -721,150 +851,10 @@ static abi_long target_mmap__locked(abi_ulong start, abi_ulong len, } 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))) { - uintptr_t host_start; - void *p; - - host_len = len + offset - host_offset; - host_len = ROUND_UP(host_len, host_page_size); - - /* Note: we prefer to control the mapping address. */ - p = mmap(g2h_untagged(start), host_len, host_prot, - flags | MAP_FIXED | MAP_ANONYMOUS, -1, 0); - if (p == MAP_FAILED) { - return -1; - } - /* update start so that it points to the file position at 'offset' */ - host_start = (uintptr_t)p; - if (!(flags & MAP_ANONYMOUS)) { - p = mmap(g2h_untagged(start), len, host_prot, - flags | MAP_FIXED, fd, host_offset); - if (p == MAP_FAILED) { - munmap(g2h_untagged(start), host_len); - return -1; - } - host_start += offset - host_offset; - } - start = h2g(host_start); - last = start + len - 1; - passthrough_start = start; - passthrough_last = last; } else { - last = start + len - 1; - real_last = ROUND_UP(last, host_page_size) - 1; - - if (flags & MAP_FIXED_NOREPLACE) { - /* Validate that the chosen range is empty. */ - if (!page_check_range_empty(start, last)) { - errno = EEXIST; - return -1; - } - - /* - * With reserved_va, the entire address space is mmaped in the - * host to ensure it isn't accidentally used for something else. - * We have just checked that the guest address is not mapped - * within the guest, but need to replace the host reservation. - * - * Without reserved_va, despite the guest address check above, - * keep MAP_FIXED_NOREPLACE so that the guest does not overwrite - * any host address mappings. - */ - if (reserved_va) { - flags = (flags & ~MAP_FIXED_NOREPLACE) | MAP_FIXED; - } - } - - /* - * worst case: we cannot map the file because the offset is not - * aligned, so we read it - */ - if (!(flags & MAP_ANONYMOUS) && - (offset & (host_page_size - 1)) != (start & (host_page_size - 1))) { - /* - * msync() won't work here, so we return an error if write is - * possible while it is a shared mapping - */ - if ((flags & MAP_TYPE) == MAP_SHARED - && (target_prot & PROT_WRITE)) { - errno = EINVAL; - return -1; - } - retaddr = target_mmap(start, len, target_prot | PROT_WRITE, - (flags & (MAP_FIXED | MAP_FIXED_NOREPLACE)) - | MAP_PRIVATE | MAP_ANONYMOUS, - -1, 0); - if (retaddr == -1) { - return -1; - } - if (pread(fd, g2h_untagged(start), len, offset) == -1) { - return -1; - } - if (!(target_prot & PROT_WRITE)) { - ret = target_mprotect(start, len, target_prot); - assert(ret == 0); - } - return mmap_end(start, last, -1, 0, flags, page_flags); - } - - /* handle the start of the mapping */ - if (start > real_start) { - if (real_last == real_start + host_page_size - 1) { - /* one single host page */ - if (!mmap_frag(real_start, start, last, - target_prot, flags, fd, offset)) { - return -1; - } - return mmap_end(start, last, -1, 0, flags, page_flags); - } - if (!mmap_frag(real_start, start, - real_start + host_page_size - 1, - target_prot, flags, fd, offset)) { - return -1; - } - real_start += host_page_size; - } - /* handle the end of the mapping */ - if (last < real_last) { - abi_ulong real_page = real_last - host_page_size + 1; - if (!mmap_frag(real_page, real_page, last, - target_prot, flags, fd, - offset + real_page - start)) { - return -1; - } - real_last -= host_page_size; - } - - /* map the middle (easier) */ - if (real_start < real_last) { - void *p, *want_p; - off_t offset1; - size_t len1; - - if (flags & MAP_ANONYMOUS) { - offset1 = 0; - } else { - offset1 = offset + real_start - start; - } - len1 = real_last - real_start + 1; - want_p = g2h_untagged(real_start); - - p = mmap(want_p, len1, host_prot, flags, fd, offset1); - if (p != want_p) { - if (p != MAP_FAILED) { - munmap(p, len1); - errno = EEXIST; - } - return -1; - } - passthrough_start = real_start; - passthrough_last = real_last; - } + return mmap_h_gt_g(start, len, target_prot, host_prot, flags, + page_flags, fd, offset, host_page_size); } - return mmap_end(start, last, passthrough_start, passthrough_last, - flags, page_flags); } /* NOTE: all the constants are the HOST ones */
Signed-off-by: Richard Henderson <richard.henderson@linaro.org> --- linux-user/mmap.c | 288 ++++++++++++++++++++++------------------------ 1 file changed, 139 insertions(+), 149 deletions(-)