feat(gpu): show VRAM in use as a bar with a percentage

Sum drm-resident-vram0 from DRM fdinfo, the same per-client counters
gputop reads, counting each drm-client-id once since a client repeats
its totals on every fd it holds. One grep over /proc/*/fdinfo, about
10ms warm, sampled through new_gate every ~2s.

The 12 GiB total is hardcoded: xe exposes no VRAM capacity in sysfs.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

4 files changed, 105 insertions, 12 deletions

diff --git a/README.md b/README.md
index 3de25f7..b137121 100644
--- a/README.md
+++ b/README.md
@@ -68,7 +68,8 @@ Seven cards cover the machine itself:
 - **System**: CPU load as a per-core equaliser, the CPU and motherboard model
   as two rows under the header, RAM used/total, and every temperature (CPU
   package, CCD die, NVMe and two board sensors).
-- **GPU**: Arc B580 package and VRAM temperatures, and fan speed.
+- **GPU**: Arc B580 package and VRAM temperatures, VRAM in use as a bar, and
+  fan speed.
 - **Disks**: `/`, `/home`, `/data` and one NFS mount per server, as usage
   rings, with free space and total size printed under each ring. Two servers
   are configured, each with the same export mounted twice; the card shows one
@@ -488,10 +489,15 @@ card that computed free space as `size - used` would overstate it by exactly
 that reserve.
 
 **The Arc B580 reports no utilisation through any stable interface.** The `xe`
-driver exposes no `gpu_busy_percent` (unlike `amdgpu`), `intel_gpu_top` refuses
-the device, and `gputop` prints per-process rows with ANSI escapes, which is not
-something to build a widget on. The GPU card therefore has no load bar; it shows
-what the hardware actually reports rather than inventing a number.
+driver exposes no `gpu_busy_percent` (unlike `amdgpu`) and `intel_gpu_top`
+refuses the device. The GPU card therefore has no load bar; it shows what the
+hardware actually reports rather than inventing a number.
+
+**VRAM in use comes from DRM fdinfo**, the per-client counters `gputop` reads:
+`drm-resident-vram0` in `/proc/<pid>/fdinfo/<fd>`, summed once per
+`drm-client-id`. Only processes readable by conky's user are seen, so a GPU
+client running as another user is not counted. The card's 12 GiB total is
+hardcoded: xe puts no VRAM capacity in sysfs.
 
 **`/data` has no temperature.** It sits on a spinning disk with no hwmon chip;
 reading its SMART temperature would need `smartctl` as root and would spin up a
diff --git a/lib/data.lua b/lib/data.lua
index 7ae96a7..d02d0f7 100644
--- a/lib/data.lua
+++ b/lib/data.lua
@@ -690,4 +690,45 @@ function M.nvtable_parse(text)
   return seen and out or nil
 end
 
+-- VRAM in use on an xe device, in bytes, from DRM fdinfo: the same source
+-- gputop reads. Input is `grep -H` output, one 'path:key:<tab>value' line per
+-- field, so fields are grouped by the fd file they came from.
+--
+-- One client (drm-client-id) is usually open on several fds, each repeating
+-- the client's full totals, so clients are counted once, not fds. Buffers
+-- shared between clients are still counted in each, as gputop does.
+local UNIT = { B = 1, KiB = 1024, MiB = 1024 ^ 2, GiB = 1024 ^ 3 }
+
+function M.drm_vram_parse(text)
+  if type(text) ~= 'string' then return nil end
+  local fds = {}
+  for path, key, val in text:gmatch('([^\n]-):(drm%-[%w%-]+):%s*([^\n]*)') do
+    local fd = fds[path] or {}
+    fds[path] = fd
+    fd[key] = val
+  end
+  local seen, total = {}, nil
+  for _, fd in pairs(fds) do
+    local id = fd['drm-client-id']
+    local n, unit = (fd['drm-resident-vram0'] or ''):match('^(%d+)%s*(%a*)')
+    if fd['drm-driver'] == 'xe' and id and n and not seen[id] then
+      seen[id] = true
+      total = (total or 0) + tonumber(n) * (UNIT[unit ~= '' and unit or 'B'] or 1)
+    end
+  end
+  return total
+end
+
+-- ponytail: a popen per call, ~10ms warm; callers gate it. Only fds of
+-- processes readable by conky's user are seen, so another user's GPU client
+-- (a system service) is not counted.
+function M.drm_vram()
+  local p = io.popen("grep -sH -e '^drm-driver' -e '^drm-client-id'"
+    .. " -e '^drm-resident-vram0' /proc/[0-9]*/fdinfo/* 2>/dev/null")
+  if not p then return nil end
+  local out = p:read('*a')
+  p:close()
+  return M.drm_vram_parse(out)
+end
+
 return M
diff --git a/test/test_data.lua b/test/test_data.lua
index 2d4d5c7..ec2b6ed 100644
--- a/test/test_data.lua
+++ b/test/test_data.lua
@@ -479,6 +479,26 @@ assert(up.upgrade == 3 and up.bump == 1 and up.current == 46 and up.missing == 1
   'updates summary')
 assert(data.nvtable_parse('') == nil, 'updates: empty cache is nil')
 
+-- drm_vram_parse: xe clients counted once across their fds, other drivers
+-- skipped, units honoured.
+local fdinfo = table.concat({
+  '/proc/1/fdinfo/5:drm-driver:\txe',
+  '/proc/1/fdinfo/5:drm-client-id:\t7',
+  '/proc/1/fdinfo/5:drm-resident-vram0:\t1024 KiB',
+  '/proc/1/fdinfo/6:drm-driver:\txe',
+  '/proc/1/fdinfo/6:drm-client-id:\t7',
+  '/proc/1/fdinfo/6:drm-resident-vram0:\t1024 KiB',
+  '/proc/2/fdinfo/9:drm-driver:\txe',
+  '/proc/2/fdinfo/9:drm-client-id:\t8',
+  '/proc/2/fdinfo/9:drm-resident-vram0:\t2 MiB',
+  '/proc/3/fdinfo/4:drm-driver:\tamdgpu',
+  '/proc/3/fdinfo/4:drm-client-id:\t9',
+  '/proc/3/fdinfo/4:drm-resident-vram0:\t5 GiB',
+}, '\n')
+assert(data.drm_vram_parse(fdinfo) == 3 * 1024 ^ 2,
+  'drm vram: got ' .. tostring(data.drm_vram_parse(fdinfo)))
+assert(data.drm_vram_parse('') == nil, 'drm vram: no clients is nil')
+
 -- new_gate: opens at most once per interval on the clock it is given.
 local gate = data.new_gate(1.9)
 assert(gate(100.0) == true, 'gate: first call opens')
diff --git a/widgets/gpu.lua b/widgets/gpu.lua
index 7d5f2ed..abac856 100644
--- a/widgets/gpu.lua
+++ b/widgets/gpu.lua
@@ -1,10 +1,9 @@
--- GPU: Arc B580 temperatures and fan speed.
+-- GPU: Arc B580 temperatures, VRAM in use and fan speed.
 --
 -- No utilisation figure, deliberately. The xe driver exposes no
--- gpu_busy_percent, intel_gpu_top refuses the device outright ("Detected Xe
--- device which is not supported"), and gputop prints per-process rows with
--- ANSI escapes, which is not an interface to build a widget on. The card shows
--- what the hardware actually reports rather than inventing a number.
+-- gpu_busy_percent and intel_gpu_top refuses the device outright ("Detected Xe
+-- device which is not supported"). VRAM comes from DRM fdinfo, the per-client
+-- counters gputop itself reads (see data.drm_vram).
 --
 -- The integrated AMD GPU does expose gpu_busy_percent and is deliberately not
 -- shown: the discrete card is the one in use.
@@ -14,6 +13,15 @@ local data = require 'lib.data'
 
 local M = {}
 
+-- The B580's VRAM. fdinfo reports what clients hold, not the card's size, and
+-- xe puts no capacity in sysfs. ponytail: hardcoded like the 'ARC B580' label.
+local VRAM_TOTAL = 12 * 1024 ^ 3
+
+-- fdinfo is a popen over every fd on the system, so sample it every ~2s, not
+-- every frame.
+local gate = data.new_gate(1.9)
+local vram
+
 -- chip, file, label, warn, crit. Package (temp2) is the headline above, so it
 -- is not repeated here; these are the rows drawn under it.
 local TEMPS = {
@@ -24,6 +32,8 @@ function M.draw(cr, rect, colors)
   local inner = card.card(cr, rect, colors)
   local function clamp(v, lo, hi) return math.max(lo, math.min(hi, v)) end
 
+  if gate() then vram = data.drm_vram() end
+
   local pkg = data.sensor('xe', 'temp2_input')
   local big_txt = pkg and string.format('%d\u{00B0}', pkg) or '--'
 
@@ -32,8 +42,8 @@ function M.draw(cr, rect, colors)
   -- only where the header or a row would overrun the card. So a tall cell
   -- gets larger type instead of empty space, and nothing ever collides.
   local LABEL_F, ROW_F = 0.32, 0.34
-  local row_n = #TEMPS + 1                            -- VRAM + FAN
-  local S_h = inner.h * 0.94 / (2.088 + row_n * 0.646)
+  local row_n = #TEMPS + 2                            -- VRAM + MEM + FAN
+  local S_h = inner.h * 0.94 / (2.088 + row_n * 0.646 + 0.3)  -- + MEM bar
   local groups = {
     { { card.title_str(card.ICON.gpu, 'GPU'), card.FONT_MONO, LABEL_F },
       { big_txt, card.FONT_HEAVY, 1.0 } },
@@ -43,6 +53,8 @@ function M.draw(cr, rect, colors)
     groups[#groups + 1] = { { t[3], card.FONT_MONO, ROW_F },
                             { '888\u{00B0}', card.FONT_MONO, ROW_F } }
   end
+  groups[#groups + 1] = { { 'MEM', card.FONT_MONO, ROW_F },
+                          { '100%', card.FONT_MONO, ROW_F } }
   groups[#groups + 1] = { { 'FAN', card.FONT_MONO, ROW_F },
                           { '8888 RPM', card.FONT_MONO, ROW_F } }
   local S = clamp(math.min(S_h, card.fit_unit(cr, inner.w * 0.96, groups, 100)), 10, 72)
@@ -80,6 +92,20 @@ function M.draw(cr, rect, colors)
     ey = ey + step
   end
 
+  -- VRAM as a fraction of the card's, a bar under the row like RAM in system.
+  local bar_h = S * 0.12
+  if ey + step + bar_h + S * 0.1 <= inner.y + inner.h then
+    local frac = vram and vram / VRAM_TOTAL
+    local colour = frac and card.threshold(frac * 100, 25, 75, colors) or colors.label
+    card.rgba(cr, colors.label)
+    card.text(cr, inner.x, ey, 'MEM')
+    card.rgba(cr, colour)
+    card.text_right(cr, inner.x + inner.w, ey,
+      frac and string.format('%d%%', math.floor(frac * 100 + 0.5)) or '--')
+    card.bar(cr, inner.x, ey + row_size * 0.6, inner.w, bar_h, frac, colour, colors)
+    ey = ey + step + bar_h + S * 0.1
+  end
+
   -- Fan RPM is not a temperature, so it carries no threshold colour: a fast
   -- fan is the cooling working, not a fault.
   if ey + step <= inner.y + inner.h then