diff options
Diffstat (limited to 'tools')
| -rw-r--r-- | tools/guest-diag/guest_pin_stream.py | 700 | ||||
| -rw-r--r-- | tools/guest-diag/pin_gr_stream.py | 134 |
2 files changed, 798 insertions, 36 deletions
diff --git a/tools/guest-diag/guest_pin_stream.py b/tools/guest-diag/guest_pin_stream.py index 3db61ba..ddc7098 100644 --- a/tools/guest-diag/guest_pin_stream.py +++ b/tools/guest-diag/guest_pin_stream.py @@ -211,7 +211,11 @@ def read_u32(kc, va): return struct.unpack("<I", d)[0] if d else 0xFFFFFFFF -DBG_BUDGET = int(os.environ.get("STREAM_DBG_BUDGET", "600")) +DBG_BUDGET = int(os.environ.get("STREAM_DBG_BUDGET", "2400")) +# SESSION 153-CONTINUED-65: minimum seconds between full-offset sweeps of +# one strict anchor (the demand file re-sends every ~2 s; the sweep can cost +# tens of seconds - see the throttle comment at the strict processing loop). +ANCHOR_RETRY_S = float(os.environ.get("STREAM_ANCHOR_RETRY_S", "20")) _dbg_count = [0] @@ -306,6 +310,12 @@ def _nv_vmap_areas(kc): # instead of walking the whole +/-16 neighbourhood. _area_off = {} +# SESSION 153-CONTINUED-55: per-GL-family learned {family_base: (area_lo, +# adj)}. Once one page of a GL push-buffer family is content-verified the +# whole family resolves through the learned area/offset (continuation and +# zero pages included); cleared when the area disappears (context rebuild). +_gl_fam = {} + def resolve_va_by_content(kc, va, out): """Demand pin: find the page carrying GPU VA 'va' by content match. @@ -320,6 +330,12 @@ def resolve_va_by_content(kc, va, out): lottery. Only the requested VA is pinned, so a wrong candidate cannot poison other VAs. + SESSION 153-CONTINUED-55: the same rule now covers the GL push-buffer + families (PB_GL_FAMILIES). The setup-push holds (va=0x4001xxx) and the + draw pushes (0x402xxxx) are nv_vm_map_pages allocations exactly like the + CE window; without this they were rejected here silently and the held + setup push was never pinned, so its 4K pages read zeros at the host. + 2026-10-16b: the offset search window was +/-1 page. Pushes of a client whose pushbuffer vmap carries a larger per-allocation header (observed with multi-MB CE staging: pushes at 0x120000f58+ never resolved and the @@ -328,11 +344,47 @@ def resolve_va_by_content(kc, va, out): vmap area, because consecutive pushes advance sequentially inside the same allocation and share the offset. """ - if not (PB_CE_LO <= va < PB_CE_HI): - return False - off = (va - PB_CE_LO) & 0xFFFFFF - k = off >> 12 + if PB_CE_LO <= va < PB_CE_HI: + k = ((va - PB_CE_LO) & 0xFFFFFF) >> 12 + gl_base = None + else: + gl_base = next((lo for lo, hi in PB_GL_FAMILIES if lo <= va < hi), + None) + if gl_base is None: + return False + k = (va - gl_base) >> 12 areas = _nv_vmap_areas(kc) + # SESSION 153-CONTINUED-55: a GL push-buffer family is one allocation + # whose vmap offset is IDENTITY with the family offset (verified: the + # setup push's bytes at va 0x4001004 read back byte-exact at the vmap + # page's +0x1004). Once one page of the family resolves, every page of + # the family is known: continuation pages of a stream do NOT start with + # a method header (the 60216-byte setup push crosses page boundaries + # mid-run, so pages 2/3/5/7 fail the prefix test) and unwritten pages + # read all zero (pages 8..15 of the same push) yet are real translations + # the host read still needs. Use the learned (area, adj) for the whole + # family; only the FIRST learn pays the content test. + learned_fam = _gl_fam.get(gl_base) if gl_base else None + if learned_fam is not None: + lo, adj = learned_fam + if any(a_lo == lo for (a_lo, _h, _p) in areas): + idx = k + adj + pages = next(p for (a_lo, _h, p) in areas if a_lo == lo) + if 0 <= idx < pages: + cpu = lo + idx * 0x1000 + pa = kc.va_to_pa(cpu) + d = kc.read(cpu, 64) if pa is not None else None + if pa is not None and d is not None: + if gl_base is not None or (_push_prefix_ok(d) and + any(d[:32])): + out.append("0x%x,0x%x,1\n" % (va & ~0xFFF, pa)) + dbg("demand-va: 0x%x -> pa=0x%x (gl family learned" + " area=0x%x adj=%+d)" + % (va & ~0xFFF, pa, lo, adj)) + return True + # The learned area is gone (context rebuilt): fall through to a + # fresh content match and re-learn. + _gl_fam.pop(gl_base, None) for (lo, _hi, pages) in areas: learned = _area_off.get(lo) if learned is not None: @@ -345,24 +397,514 @@ def resolve_va_by_content(kc, va, out): if idx < 0 or idx >= pages: continue cpu = lo + idx * 0x1000 - d = kc.read(cpu, 64) - if not d or not any(d[:32]): - continue - if not _push_prefix_ok(d): - continue pa = kc.va_to_pa(cpu) if pa is None: continue + # 0x200 (not 64): a GL family's page 0 can carry its first + # method header at offset ~0x68 behind a zero-filled prefix + # (observed on the 60216 B setup push), invisible to a 64-byte + # read. + d = kc.read(cpu, 0x200) + if not d or not any(d): + continue + if not _push_prefix_ok(d): + continue _area_off[lo] = adj + if gl_base is not None: + # First content-verified page of this GL family: the + # identity mapping (area, adj) now serves every page of + # the family, including continuation and zero pages. + _gl_fam[gl_base] = (lo, adj) out.append("0x%x,0x%x,1\n" % (va & ~0xFFF, pa)) - dbg("demand-va: 0x%x -> pa=0x%x (area off=%d adj=%+d)" - % (va & ~0xFFF, pa, idx, adj)) + dbg("demand-va: 0x%x -> pa=0x%x (area off=%d adj=%+d%s)" + % (va & ~0xFFF, pa, idx, adj, + " gl-learn" if gl_base is not None else "")) return True dbg("demand-va: 0x%x not found by content (%d areas)" % (va & ~0xFFF, len(areas))) return False +# SESSION 153-CONTINUED-64: candidate adjacencies for the strict window scan. +_STRICT_ADJ = (0, 1, -1, 2, -2, 3, -3, 4, -4, 6, -6, 8, -8, 14, -14, 16, -16) + + +def _stream_parse_full(buf): + """Exact port of the host's svc_stream_parse (gsp_sema_service.c), plus + the MME-setup content requirements the bare parse lacks. + + Zero words are NOP padding (skipped anywhere in the stream), ops 1/3/5 + are INC/NON_INC/ONE_INC runs with a 12-bit count (1..0xfff) and count + data words, ops 0/2/4/7 are header-only, anything else fails. The buffer + parses only when at least one method was seen and the walk lands exactly + on the end (trailing zero padding allowed). + + SESSION 153-CONTINUED-64b: the bare exact-end parse accepted a GMMU + page-table window - PTEs are {pfn<<12, flags} u32 pairs whose values sit + below 0x20000000, so every word reads as op 0 (header-only) and the walk + lands exactly on the end (r1122 pinned pa=0x11dbb0000: all 32 words of + every page op 0, zero data runs, host read `sema-bigrun hdr=0x1dbb7000`). + A window is now only accepted when it carries at least one data run AND + at least one run addressing an MME-setup method (0x114..0x124 / 0x234 / + 0x550..0x568 / 0x3400..0x34fc - the same windows the host's + svc_mme_setup_run_addr classifies with). A page table can never satisfy + either; the held setup push always carries both (the r1116 capture's + first run `20020047` = INC run @0x11c). + """ + off = 0 + methods = 0 + data_runs = 0 + mme_hits = 0 + code_words = 0 + n = len(buf) + while off + 4 <= n: + hdr = struct.unpack_from('<I', buf, off)[0] + if hdr == 0: + off += 4 + continue + op = (hdr >> 29) & 7 + count = (hdr >> 16) & 0xFFF + if op in (1, 3, 5): + if not count or off + 4 + count * 4 > n: + return False + data_runs += 1 + addr = (hdr & 0x1FFF) << 2 + if (0x114 <= addr <= 0x124 or addr == 0x234 or + 0x550 <= addr <= 0x568 or 0x3400 <= addr <= 0x34FC): + mme_hits += 1 + if addr == 0x114 and count >= 1: + code_words += count + off += 4 + count * 4 + elif op in (0, 2, 4, 7): + off += 4 + else: + return False + methods += 1 + if methods > 8192: + return False + while off < n and buf[off] == 0: + off += 1 + return (methods > 0 and off == n and data_runs > 0 and mme_hits > 0 + and code_words >= 8) + + +def _walk_measure(buf): + """Method-shape walk returning (end_ok, data_runs, mme_hits, code_words, + pair_runs, parsed_bytes). 'end_ok' mirrors _stream_parse_full's + exact-end + zero padding landing; 'pair_runs' counts runs addressing the + SA-pair method 0x11c (the client push signature); 'parsed_bytes' is how + far the walk got before a shape failure (or the window end).""" + off = 0 + methods = 0 + data_runs = 0 + mme_hits = 0 + code_words = 0 + pair_runs = 0 + n = len(buf) + end_ok = False + while off + 4 <= n: + hdr = struct.unpack_from('<I', buf, off)[0] + if hdr == 0: + off += 4 + continue + op = (hdr >> 29) & 7 + count = (hdr >> 16) & 0xFFF + if op in (1, 3, 5): + if not count or off + 4 + count * 4 > n: + break + data_runs += 1 + addr = (hdr & 0x1FFF) << 2 + if (0x114 <= addr <= 0x124 or addr == 0x234 or + 0x550 <= addr <= 0x568 or 0x3400 <= addr <= 0x34FC): + mme_hits += 1 + if addr == 0x114: + code_words += count + if addr == 0x11c and count >= 2: + pair_runs += 1 + off += 4 + count * 4 + elif op in (0, 2, 4, 7): + off += 4 + else: + break + methods += 1 + if methods > 8192: + break + else: + o2 = off + while o2 < n and buf[o2] == 0: + o2 += 1 + end_ok = (methods > 0 and o2 == n) + return end_ok, data_runs, mme_hits, code_words, pair_runs, off + + +def _window_accept(buf, ln): + """Acceptance for a client-mapping candidate window. + + SESSION 153-CONTINUED-65: the full-offset sweep of the client-mem + resolver needs a false-positive gate that survives pointer/event data. + The probe of the r1125 boot found a 60216-byte window of pointer junk + that walked to an exact end with one misinterpreted 0x114 run + (mme=1): the old accept (data>0, mme>0, code>=8) would have pinned it. + Large windows therefore require the client push signature - at least + two MME-setup runs or two SA-pair (0x11c) runs - in addition to the + counts; small windows keep the historical strict rule. The relaxed + path additionally accepts a window whose tail is unmapped/unwritten + (walk parses >= 0x1000) when it carries >= 64 code words (the macro + upload signature). + + Returns (ok, mode). + """ + if ln < 0x8000: + return _stream_parse_full(buf), "strict" + end_ok, data, mme, code, pairs, parsed = _walk_measure(buf) + # The client-push signature: SA-pair (0x11c) runs or 0x114 code runs. + # The r1125 probe of the live client mappings measured 10k+ pointer/float + # windows with data/mme hits but pairs=0, code=0 - MME-setup-family hits + # alone (0x124/0x34xx) are NOT a discriminator; the pair/code runs are. + sig = (pairs >= 2 or code >= 8) + if end_ok and data >= 2 and sig: + return True, "strict" + if data >= 2 and sig and parsed >= 0x1000: + return True, "relaxed" + return False, None + + +def _prefix_walk_ok(buf, off): + """Cheap first-page pre-filter for a strict window candidate. + + Walks the method shapes from `off` to the end of `buf` WITHOUT requiring + the end to land on a boundary (runs may cross into the next page and a + run header near the page end declares data we cannot see yet). Fails on + the first header no method shape can produce; this kills arbitrary data + pages after a few words so the full parse only runs on survivors. + """ + n = len(buf) + while off + 4 <= n: + hdr = struct.unpack_from('<I', buf, off)[0] + if hdr == 0: + off += 4 + continue + op = (hdr >> 29) & 7 + count = (hdr >> 16) & 0xFFF + if op in (1, 3, 5): + if not count: + return False + off += 4 + count * 4 + elif op in (0, 2, 4, 7): + off += 4 + else: + return False + return True + + +def _read_window(kc, cpu, off, ln): + """Read [cpu+off, cpu+off+ln) page-by-page through kcore. + + A vmap area's pages are contiguous in the kernel VA but arbitrary in + physical memory, so the window is assembled one page at a time. + """ + out = b"" + a = cpu + off + left = ln + while left > 0: + n = min(left, 0x1000 - (a & 0xFFF)) + d = kc.read(a, n) + if not d: + return None + out += d + a += n + left -= n + return out + + +def _mme_run_in_page(d): + """True when a 4K page contains a run addressing an MME-setup method. + + The strict-window prefilter for client-mapping candidates (cheap: one + page read rejects the ~511/512 wrong shifts before the full parse). + """ + n = len(d) + off = 0 + while off + 4 <= n: + hdr = struct.unpack_from('<I', d, off)[0] + if hdr == 0: + off += 4 + continue + op = (hdr >> 29) & 7 + count = (hdr >> 16) & 0xFFF + if op in (1, 3, 5): + if not count: + off += 4 + continue + addr = (hdr & 0x1FFF) << 2 + if (0x114 <= addr <= 0x124 or addr == 0x234 or + 0x550 <= addr <= 0x568 or 0x3400 <= addr <= 0x34FC): + return True + off += 4 + count * 4 + else: + off += 4 + return False + + +def resolve_va_client_mem(pid, ent_va, off, ln, out): + """Strict pin for a setup push whose pages live in CLIENT sysmem. + + SESSION 153-CONTINUED-64c: when the setup push's allocation is UVM + sysmem it has NO kernel vmap (r1122: no nv_vm_map_pages area held the + `a0100045` 0x114 upload runs anywhere on the boot - the content exists + only in the client's `rw-s` /dev/nvidia mappings, and the host's + GPU-page-table walk answered an unrelated page-table buffer for the + held VA). The client's CPU mappings + /proc/<pid>/pagemap are the only + ground truth for those pages. Scan the client's readable+writable + mappings: for each candidate page shift, read the exact [va, va+len) + window via /proc/<pid>/mem, validate with the host's own full parse, + and emit pins for every page of the validated window from the + pagemap. The cheap per-page MME-run prefilter rejects wrong shifts. + """ + if not pid: + return False + try: + pfd = os.open('/proc/%d/pagemap' % pid, os.O_RDONLY) + mfd = os.open('/proc/%d/mem' % pid, os.O_RDONLY) + except OSError: + return False + try: + gl_base = next((lo for lo, hi in PB_GL_FAMILIES + if lo <= (ent_va & ~0xFFF) < hi), None) + k = ((ent_va & ~0xFFF) - gl_base) >> 12 if gl_base is not None else None + if ln <= 0 or ln > 0x40000: + return False + need = (off + ln + 0xFFF) // 0x1000 + if need > 64: + return False + try: + maps = open('/proc/%d/maps' % pid).read().splitlines() + except OSError: + return False + tried = 0 + for line in maps: + parts = line.split() + if len(parts) < 5 or 'rw' not in parts[1]: + continue + # SESSION 153-CONTINUED-64e: any rw mapping may carry the + # sysmem push (observed as pathless rw-s /dev/nvidia mappings; + # memfd-backed buffers cannot be excluded either - the strict + # parse + MME prefilter are the safety gates, not the path). + try: + lo, hi = [int(x, 16) for x in parts[0].split('-')] + except ValueError: + continue + if hi - lo < (need + 4) * 0x1000 or hi - lo > (64 << 20): + continue + pages = (hi - lo) >> 12 + if k is None or k == 0: + # Unknown family, or the FAMILY-BASE page (k == 0): neither + # has an identity page index to key on, so sweep every page + # start of the mapping (bounded). SESSION 153-CONTINUED-65: + # the 0x3200000 macro-upload push's window sits at an + # arbitrary page offset inside the client's mapping (the + # r1124 power-of-two neighbourhood never tried it; the + # window content measures code=712, exact-end, under the + # strict walk). + adjs = range(0, max(1, pages - need + 1)) + else: + # SESSION 153-CONTINUED-65: neighbourhood FIRST, then the + # FULL page sweep. The r1125 setup push (family pages k=1 + # and k=33, 60216 B) failed every power-of-two offset while + # a matching window could sit at an arbitrary mapping page; + # every candidate still passes _window_accept before any + # pin is emitted, so the widened sweep cannot pin a window + # the parse rejects. + near = [0, 1, -1, 2, -2, 4, -4, 8, -8, 16, -16, 32, -32, + 64, -64, 128, -128, 256, -256] + near_set = set(near) + rest = [a for a in range(-k, pages - need + 1 - k) + if a not in near_set] + adjs = near + rest + for adj in adjs: + idx = (k + adj) if k is not None else adj + if idx < 0 or idx + need > pages: + continue + win = lo + idx * 0x1000 + off + tried += 1 + if tried > 8192: + break + try: + first = os.pread(mfd, 0x1000 - off, win) + except OSError: + continue + if not first: + continue + # An all-zero first page is a legitimate push head (the + # method stream may start later); only a page with non-zero + # content and no MME-setup run is rejected outright. + if not _mme_run_in_page(first) and any(first): + continue + buf = b'' + left = ln + a = win + ok = True + while left > 0: + n = min(left, 0x1000 - (a & 0xFFF)) + try: + d = os.pread(mfd, n, a) + except OSError: + ok = False + break + if not d or len(d) != n: + ok = False + break + buf += d + a += n + left -= n + if not ok: + continue + acc, mode = _window_accept(buf, ln) + if not acc: + continue + pas = [] + for j in range(need): + vpg = (ent_va & ~0xFFF) + j * 0x1000 + cpg = lo + (idx + j) * 0x1000 + os.lseek(pfd, (cpg >> 12) * 8, 0) + e = int.from_bytes(os.read(pfd, 8), 'little') + if not (e >> 63) & 1: + pas = None + break + pas.append(((e & ((1 << 55) - 1)) << 12)) + if not pas: + continue + for j, pa in enumerate(pas): + out.append("0x%x,0x%x,1\n" + % ((ent_va & ~0xFFF) + j * 0x1000, pa)) + dbg("demand-va: 0x%x -> pa=0x%x (%s client-mem parse" + " %u B ok pages=%u map=0x%x idx=%d)" + % (ent_va & ~0xFFF, pas[0], mode, ln, need, lo, idx)) + return True + if tried > 8192: + break + return False + finally: + os.close(pfd) + os.close(mfd) + + +def resolve_va_strict(kc, ent_va, off, ln, out, pid=None): + """Demand pin validated by the host's exact full-parse of [ent_va, +ln). + + SESSION 153-CONTINUED-64: the content heuristic learned a GL family from + ONE page's weak 0x200-byte prefix test and then served every family page + from that (area, adj) with NO content check at all; r1120 pinned + 0x4001000 -> a page whose bytes (`00040000 00001000...`) failed the + host's stream parse for the whole boot (`plausible=0 unreadable=1`), so + the setup push was never captured and the prelude gate deferred every + launch. This path runs the HOST'S OWN full parse on the exact bytes the + host will read before any pin is emitted, then pins EVERY page of the + validated window and learns the family from it. A family is never + served from an unvalidated learn. + """ + page = ent_va & ~0xFFF + gl_base = next((lo for lo, hi in PB_GL_FAMILIES if lo <= page < hi), None) + if ln <= 0 or ln > 0x40000: + return False + if gl_base is None: + # SESSION 153-CONTINUED-64e: an anchor outside the known families + # (r1124: the hold moved to a 153600-byte push at 0x3200000) has no + # identity page index; only the client-mapping sweep can resolve it. + try: + if resolve_va_client_mem(pid, ent_va, off, ln, out): + return True + except Exception: + pass + dbg("demand-va: 0x%x strict unknown-family client-mem failed" % page) + return False + gl_hi = next(hi for lo, hi in PB_GL_FAMILIES if lo <= page < hi) + k = (page - gl_base) >> 12 + need = (off + ln + 0xFFF) // 0x1000 + if need > 64 or page + need * 0x1000 > gl_hi: + return False + areas = _nv_vmap_areas(kc) + cands = [] + fam = _gl_fam.get(gl_base) + if fam is not None and any(a_lo == fam[0] for (a_lo, _h, _p) in areas): + cands.append(fam) + if k == 0: + # SESSION 153-CONTINUED-65: a family-BASE anchor (the 0x3200000 + # macro-upload push) has no identity page index to key on. The + # r1126 macro window lives in a KERNEL-mapped allocation (no + # process pagemap maps it: the whole-process reverse scan found + # none), so the kcore vmap sweep is the only resolver - and the + # power-of-two neighborhood missed the window's arbitrary page + # index. Sweep every page of every nv_vm_map_pages area; the + # _window_accept parse gates each candidate. + for (lo, _hi, pages) in areas: + for adj in range(0, max(1, pages - need + 1)): + c = (lo, adj) + if c not in cands: + cands.append(c) + else: + for (lo, _hi, pages) in areas: + for adj in _STRICT_ADJ: + c = (lo, adj) + if c not in cands: + cands.append(c) + tried = 0 + for (lo, adj) in cands: + pages = next(p for (a_lo, _h, p) in areas if a_lo == lo) + idx = k + adj + if idx < 0 or idx + need > pages: + continue + tried += 1 + if tried > (8192 if k == 0 else 128): + break + cpu = lo + idx * 0x1000 + first = kc.read(cpu, 0x1000) + if not first: + continue + # SESSION 153-CONTINUED-65: an all-zero first page is a legitimate + # push head - the r1126 macro-upload window (0x3200000, 153600 B) + # starts with a fully zero page (the method stream lands later in + # the window); the old `not any(first)` skip threw exactly that + # candidate away (the r1128 sweep tried 166 kernel-vmap windows and + # rejected the true one on this check). The prefix walk trivially + # passes an all-zero page; the full-window accept is the gate. + if any(first) and not _prefix_walk_ok(first, off): + continue + buf = _read_window(kc, cpu, off, ln) + if buf is None: + continue + acc, mode = _window_accept(buf, ln) + if not acc: + continue + pas = [] + for j in range(need): + vpa = kc.va_to_pa(cpu + j * 0x1000) + if vpa is None: + pas = None + break + pas.append(vpa) + if not pas: + continue + for j, vpa in enumerate(pas): + out.append("0x%x,0x%x,1\n" % (page + j * 0x1000, vpa)) + _area_off[lo] = adj + _gl_fam[gl_base] = (lo, adj) + dbg("demand-va: 0x%x -> pa=0x%x (%s parse %u B ok pages=%u" + " area=0x%x adj=%+d)" % (page, pas[0], mode, ln, need, lo, adj)) + return True + dbg("demand-va: 0x%x strict parse failed (%u B, %u windows tried)" + % (page, ln, tried)) + # SESSION 153-CONTINUED-64c: no kernel vmap held the content - the + # allocation is UVM sysmem visible only in the client's own mappings. + try: + if resolve_va_client_mem(pid, ent_va, off, ln, out): + return True + except Exception: + pass + dbg("demand-va: 0x%x strict client-mem parse failed too" % page) + return False + + def resolve_mem_pins(kc, mem, out, default_pages=16, va_lo=0, va_hi=0, page_off=0): """uvm_rm_mem kptr -> GPU-VA/GPA runs (direct map or page-table walk). @@ -810,6 +1352,19 @@ def resolve_pb_pins(kc, pb, out): # hold the ring-entry pushes is chosen (content anchor, _mem_content_ok). PB_CE_LO = 0x120000000 PB_CE_HI = 0x121000000 +# SESSION 153-CONTINUED-55: GL push-buffer VA families. The GL client's +# setup/MME-prelude push buffers (observed hold va=0x4001004 len=60216) and +# the draw pushes (observed va=0x4021004) are separate nv_vm_map_pages +# allocations in two fixed windows; the same family-relative page index +# rule applies: k = (va - base) >> 12. +# SESSION 153-CONTINUED-63: the r1117 boot's setup push lives at +# va=0x4200110 len=12288 - a THIRD window (0x4200000..0x4210000) the two +# original families did not cover, so the content resolver answered +# nothing and the read lottery could never be won for it (the +# 0x4020000-family hold re-armed for its whole 20 s window). +PB_GL_FAMILIES = ((0x3200000, 0x3300000), (0x4000000, 0x4100000), + (0x4200000, 0x4210000), + (0x42000000, 0x43000000)) # m165 (D-F1a): UVM semaphore-pool page range. The pool is a 4 KB sysmem # uvm_rm_mem placed by the UVM arena right after the 16 MB CE push buffer; # the client's tracker words (0x121000000 + 4*idx) live in its first page. @@ -1674,7 +2229,8 @@ def main(): # deinit), so the set deliberately persists across client switches. pool_pages = {} pool_reemit_next = 0.0 - _want_state = {"mtime": None, "done": set()} + _want_state = {"mtime": None, "done": set(), "strict_pending": {}, + "strict_try": {}} loop = 0 while True: pid = cdt_pid() @@ -1788,17 +2344,119 @@ def main(): # main(), stopping ALL new pins (the observed # pool-pin latency). Extract every 0x token instead; # a bad token is skipped, never fatal. - wvas = [] - for t in re.findall(r"0x[0-9a-fA-F]+", - fw.read()): + text = fw.read() + _want_state["mtime"] = st_d.st_mtime + # SESSION 153-CONTINUED-64: strict demand anchors first. + # A line `S 0x<entry_va> 0x<off> <len>` (written by + # pin_gr_stream from the service's anchored debugfs + # demand record) asks for the exact [entry_va, +len) + # read window to be validated with the host's own full + # stream parse before a pin is injected. Loose lines + # (plain `0x<va>`) keep the content/pagemap paths, but + # a GL family with a pending anchor is deferred so its + # learn can only come from the strict validation. + strict = [] + loose_tokens = [] + for line in text.splitlines(): + ms = re.match( + r"\s*S\s+0x([0-9a-fA-F]+)\s+0x([0-9a-fA-F]+)" + r"\s+(\d+)\s*$", line) + if ms: try: - wvas.append(int(t, 16)) + strict.append((int(ms.group(1), 16), + int(ms.group(2), 16), + int(ms.group(3)))) except ValueError: pass - _want_state["mtime"] = st_d.st_mtime - for wv in wvas[-64:]: + continue + loose_tokens += re.findall(r"0x[0-9a-fA-F]+", line) + now_d = time.time() + for epage in list(_want_state["strict_pending"]): + if now_d - _want_state["strict_pending"][epage] > 300.0: + del _want_state["strict_pending"][epage] + for (ev, eoff, eln) in reversed(strict): + epage = ev & ~0xFFF + if epage in _want_state["done"]: + continue + # SESSION 153-CONTINUED-65: throttle anchor + # re-resolution. The full-offset sweep can read + # thousands of windows (tens of seconds of CPU); + # the demand file re-sends the anchor every ~2 s, + # so without this throttle the scanner starves the + # client (both measured at ~91% CPU at r1125). One + # attempt per anchor per window is enough - content + # that lands late is picked up on the next window. + lt = _want_state["strict_try"].get(epage, 0.0) + if now_d - lt < ANCHOR_RETRY_S: + continue + _want_state["strict_try"][epage] = now_d + got = [] + try: + ok = resolve_va_strict(kc, ev, eoff, eln, got, + pid) + except Exception: + ok = False + got = [] + if not ok and eoff: + # The demand's off is the entry va's in-page + # offset; the push content can be page-aligned + # instead (r1123 vs r1125 parse behaviour). A + # second try at page alignment is cheap next to + # the completed sweep and the same pins result. + got2 = [] + try: + ok = resolve_va_strict(kc, ev & ~0xFFF, 0, + eln, got2, pid) + except Exception: + ok = False + if ok: + got = got2 + if ok and got: + lines.extend(got) + for g in got: + try: + _want_state["done"].add( + int(g.split(",", 1)[0], 16)) + except ValueError: + pass + _want_state["done"].add(epage) + _want_state["strict_pending"].pop(epage, None) + else: + _want_state["strict_pending"].setdefault( + epage, now_d) + # Loose VAs in a family whose strict anchor is still + # pending are deferred to the strict path only. + pend_bases = set() + for epage in _want_state["strict_pending"]: + for lo, hi in PB_GL_FAMILIES: + if lo <= epage < hi: + pend_bases.add((lo, hi)) + # SESSION 153-CONTINUED-55: process the NEWEST VAs first, + # not the deduped head. The file is append-only and grows + # to hundreds of lines (every stalled-VA row is appended + # each demand cycle, ~48/4 s under load), so both the + # literal tail (`wvas[-64:]`) and a first-occurrence + # dedupe sorted older pending VAs out of the window (the + # 16-page GL setup-push demand fell out between its + # appends and the next mtime change). Walk the raw tail + # backwards, newest instruction first, skipping resolved + # VAs, up to a bounded batch. + wvas = [] + for t in reversed(loose_tokens): + try: + v = int(t, 16) + except ValueError: + continue + if v not in _want_state["done"] and v not in wvas: + wvas.append(v) + if len(wvas) >= 256: + break + for wv in wvas: if wv in _want_state["done"]: continue + if any(lo <= (wv & ~0xFFF) < hi + for lo, hi in pend_bases): + continue got = [] if POOL_SEM_LO <= (wv & ~0xFFF) < POOL_SEM_HI: # m165 (D-F1a): pool/waiter VAs have no method @@ -1833,8 +2491,8 @@ def main(): # page: resolve it through the client's own # pagemap. Without this the demand for those # VAs was a silent no-op (resolve_va_by_content - # rejects anything outside PB_CE_LO..PB_CE_HI) - # and the fills kept dropping them. + # rejects anything outside its windows) and the + # fills kept dropping them. if not ok and not got and pid: try: pfd = os.open('/proc/%d/pagemap' % pid, diff --git a/tools/guest-diag/pin_gr_stream.py b/tools/guest-diag/pin_gr_stream.py index 776b440..35c2ea3 100644 --- a/tools/guest-diag/pin_gr_stream.py +++ b/tools/guest-diag/pin_gr_stream.py @@ -131,11 +131,29 @@ def _demand_send(vas): pass -def _start_cmd(): +def _push_guest_script(): + """Push the scanner into the guest in argv-safe chunks. + + SESSION 153-CONTINUED-64: the one-shot `echo <b64> | base64 -d >` form + put the whole ~135 KB base64 of the (grown) scanner on vgpu_vm.py's + argv, and the host's MAX_ARG_STRLEN (128 KB) rejected the exec before + it ever reached the guest ('OSError: Argument list too long'), so no + scanner started and the boot streamed zero pins. Push 60 KB chunks and + print the guest-side md5 for verification. + """ with open(GUEST_SCRIPT) as f: b64 = subprocess.run(["base64", "-w0"], stdin=f, capture_output=True, text=True, - timeout=10).stdout + timeout=10).stdout.strip() + vexec("rm -f /tmp/gps.b64", 30) + for off in range(0, len(b64), 60000): + vexec("printf %%s '%s' >> /tmp/gps.b64" % b64[off:off + 60000], 30) + return vexec("base64 -d /tmp/gps.b64 > /tmp/guest_pin_stream.py && " + "rm -f /tmp/gps.b64 && md5sum /tmp/guest_pin_stream.py " + "&& echo PUSH_OK", 60) or "" + + +def _start_cmd(): prefixes = os.environ.get("CLIENT_PREFIXES", "cdt,va,victory,Xorg,X,glxgears,gzdoom") vma_every = os.environ.get("STREAM_VMA_EVERY", "1") @@ -151,14 +169,13 @@ def _start_cmd(): # failed silently (r149). return ("P=$(printf '^python3 /tmp/guest_pin_%%s.py' stream); " "sudo pkill -f \"$P\" 2>/dev/null; sleep 1; " - "echo '%s' | base64 -d > /tmp/guest_pin_stream.py; " "sudo rm -f /tmp/guest_pin_stream.pid; " "sudo setsid nohup env CLIENT_PREFIXES='%s' " "STREAM_VMA_EVERY='%s' " "python3 /tmp/guest_pin_stream.py %d %d %s %s %s " ">/tmp/guest_pin_stream.log 2>&1 & sleep 2; " "sudo cat /tmp/guest_pin_stream.pid 2>/dev/null" - % (b64, prefixes, vma_every, 2, PORT, LO, HI, PAMAX)) + % (prefixes, vma_every, 2, PORT, LO, HI, PAMAX)) def install_and_start(): @@ -167,6 +184,14 @@ def install_and_start(): # a boot without the scanner streams no pins and the client then hangs. for attempt in range(5): _heartbeat() + try: + out_push = _push_guest_script() + if "PUSH_OK" not in (out_push or ""): + log("WARN: scanner push attempt %d incomplete: %s" + % (attempt + 1, (out_push or "")[-120:])) + except Exception as e: + log("WARN: scanner push attempt %d failed: %s" + % (attempt + 1, e)) out = vexec(_start_cmd(), 60) if any(ch.isdigit() for ch in out): return True @@ -230,13 +255,31 @@ def _demand_loop(): expiry) as the stalled-VA class. """ track = {} # va -> [first_seen, last_sent] + # SESSION 153-CONTINUED-64: strict demand anchors (entry va -> [first, + # last_sent]) from the anchored debugfs lines. The guest validates the + # exact [entry_va, +len) read window with the host's full stream parse + # before pinning, so the setup-push pin can never land on a page whose + # bytes the host's read will reject. Resent fast (2 s) while the hold + # is up; the loose cadence (20 s) is too slow for the client's churn. + strict_track = {} + it = 0 while True: - time.sleep(3) - try: - r = subprocess.run(["sudo", "-n", "dmesg"], - capture_output=True, text=True, timeout=20) - except Exception: - continue + # SESSION 153-CONTINUED-57: the 3 s pass was the losing half of the + # r1096 race - the beef setup-hold's demanded pages resolved in the + # guest ~2-4 s after the arm, while the pump re-read the entry + # every ~110 ms and moved on. Poll fast (0.75 s) so a newly armed + # demand goes to the guest within a second; the expensive dmesg + # harvest stays on a slower cadence (every 4th pass) because the + # debugfs setup-demand set is the time-critical source. + time.sleep(0.75) + it += 1 + r = None + if it % 4 == 1: + try: + r = subprocess.run(["sudo", "-n", "dmesg"], + capture_output=True, text=True, timeout=20) + except Exception: + continue now = time.time() for m in re.finditer( r"(?:stalled VA #\d+[:=]\s*" @@ -258,24 +301,85 @@ def _demand_loop(): # cannot, the guest content resolver is the last source - # harvest the VA so it resolves and injects the real page. r"|sema-pushzero: h=0x[0-9a-f]+ va=(0x[0-9a-f]+)", - (r.stdout or "")[-200000:]): + (r.stdout or "")[-2000000:] if r else ""): va = int(m.group(1) or m.group(2) or m.group(3), 16) track.setdefault(va, [now, 0.0]) + # SESSION 153-CONTINUED-51: the pushzero line now carries the + # push length; a zero-head WINDOW push spans 1-2 pages, and the + # demand of a single page leaves the capture zero - expand the + # range and demand every page so the guest pagemap resolves the + # whole push at once (bulk demand instead of one page per cycle). + for m in re.finditer( + r"sema-pushzero: h=0x[0-9a-f]+ va=(0x[0-9a-f]+) len=(\d+)", + (r.stdout or "")[-2000000:] if r else ""): + zva = int(m.group(1), 16) + zlen = int(m.group(2)) + for zoff in range(0, min(zlen, 0x8000), 0x1000): + track.setdefault(zva + zoff, [now, 0.0]) + # SESSION 153-CONTINUED-56: the armed setup-hold's demand VA set is + # exposed by the service as STATE (debugfs), not as a dmesg row, so + # it cannot race the ring rotation under vgsp-sema spam (r1090 lost + # the setup-push pages that way). Read it every pass and feed the + # same track/resend machinery: new VAs are sent this pass (last=0), + # existing ones keep the 20 s resend / 300 s expiry cadence. + try: + rr = subprocess.run( + ["sudo", "-n", "cat", + "/sys/kernel/debug/vgsp/sema_setup_demand"], + capture_output=True, text=True, timeout=10) + if rr.returncode == 0: + for line in (rr.stdout or "").splitlines(): + ma = re.match( + r"\s*0x([0-9a-f]+)\s+0x([0-9a-f]+)\s+(\d+)\s*$", + line) + if ma: + ev = int(ma.group(1), 16) + off = int(ma.group(2), 16) + ln = int(ma.group(3)) + if ev and 0 < ln <= 0x40000: + strict_track.setdefault(ev, [now, 0.0, + off, ln]) + continue + for m in re.finditer(r"0x[0-9a-f]+", line): + sva = int(m.group(0), 16) & ~0xfff + if sva: + track.setdefault(sva, [now, 0.0]) + except Exception: + pass for va in list(track): first, last = track[va] if now - first > 300.0: del track[va] + for ev in list(strict_track): + first, last, off, ln = strict_track[ev] + if now - first > 300.0: + del strict_track[ev] + cmds = [] todo = [va for va, (f, l) in track.items() if now - l > 20.0] - if not todo: + if todo: + todo = todo[:48] + cmds.append(_demand_write_cmd(todo)) + stodo = [ev for ev, (f, l, o, ln) in strict_track.items() + if now - l > 2.0] + if stodo: + stodo = stodo[:16] + cmds.append("; ".join( + "echo S 0x%x 0x%x %d >> /tmp/guest_want_vas.txt" + % (ev, strict_track[ev][2], strict_track[ev][3]) + for ev in stodo)) + if not cmds: continue - cmd = _demand_write_cmd(todo) try: - vexec(cmd, 30) + vexec("; ".join(cmds), 30) except Exception: continue for va in todo: track[va][1] = now - log("demand: sent %d stalled VA(s) to guest" % len(todo)) + for ev in stodo: + strict_track[ev][1] = now + log("demand: sent %d stalled VA(s)%s to guest" % + (len(todo), + " + %d strict anchor(s)" % len(stodo) if stodo else "")) _DEMAND_STARTED = False |